mirror of
https://gitlab.freedesktop.org/gstreamer/gstreamer.git
synced 2024-11-03 16:09:39 +00:00
15257 lines
398 KiB
C
15257 lines
398 KiB
C
|
|
/* autogenerated from gstcogorc.orc */
|
|
|
|
#ifdef HAVE_CONFIG_H
|
|
#include "config.h"
|
|
#endif
|
|
#include <glib.h>
|
|
|
|
#ifndef _ORC_INTEGER_TYPEDEFS_
|
|
#define _ORC_INTEGER_TYPEDEFS_
|
|
#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
|
|
#include <stdint.h>
|
|
typedef int8_t orc_int8;
|
|
typedef int16_t orc_int16;
|
|
typedef int32_t orc_int32;
|
|
typedef int64_t orc_int64;
|
|
typedef uint8_t orc_uint8;
|
|
typedef uint16_t orc_uint16;
|
|
typedef uint32_t orc_uint32;
|
|
typedef uint64_t orc_uint64;
|
|
#define ORC_UINT64_C(x) UINT64_C(x)
|
|
#elif defined(_MSC_VER)
|
|
typedef signed __int8 orc_int8;
|
|
typedef signed __int16 orc_int16;
|
|
typedef signed __int32 orc_int32;
|
|
typedef signed __int64 orc_int64;
|
|
typedef unsigned __int8 orc_uint8;
|
|
typedef unsigned __int16 orc_uint16;
|
|
typedef unsigned __int32 orc_uint32;
|
|
typedef unsigned __int64 orc_uint64;
|
|
#define ORC_UINT64_C(x) (x##Ui64)
|
|
#define inline __inline
|
|
#else
|
|
#include <limits.h>
|
|
typedef signed char orc_int8;
|
|
typedef short orc_int16;
|
|
typedef int orc_int32;
|
|
typedef unsigned char orc_uint8;
|
|
typedef unsigned short orc_uint16;
|
|
typedef unsigned int orc_uint32;
|
|
#if INT_MAX == LONG_MAX
|
|
typedef long long orc_int64;
|
|
typedef unsigned long long orc_uint64;
|
|
#define ORC_UINT64_C(x) (x##ULL)
|
|
#else
|
|
typedef long orc_int64;
|
|
typedef unsigned long orc_uint64;
|
|
#define ORC_UINT64_C(x) (x##UL)
|
|
#endif
|
|
#endif
|
|
typedef union
|
|
{
|
|
orc_int16 i;
|
|
orc_int8 x2[2];
|
|
} orc_union16;
|
|
typedef union
|
|
{
|
|
orc_int32 i;
|
|
float f;
|
|
orc_int16 x2[2];
|
|
orc_int8 x4[4];
|
|
} orc_union32;
|
|
typedef union
|
|
{
|
|
orc_int64 i;
|
|
double f;
|
|
orc_int32 x2[2];
|
|
float x2f[2];
|
|
orc_int16 x4[4];
|
|
} orc_union64;
|
|
#endif
|
|
#ifndef ORC_RESTRICT
|
|
#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
|
|
#define ORC_RESTRICT restrict
|
|
#elif defined(__GNUC__) && __GNUC__ >= 4
|
|
#define ORC_RESTRICT __restrict__
|
|
#else
|
|
#define ORC_RESTRICT
|
|
#endif
|
|
#endif
|
|
|
|
#ifndef DISABLE_ORC
|
|
#include <orc/orc.h>
|
|
#endif
|
|
void cogorc_memcpy_2d (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_downsample_horiz_cosite_1tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n);
|
|
void cogorc_downsample_horiz_cosite_3tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
|
|
int n);
|
|
void cogorc_downsample_420_jpeg (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
|
|
int n);
|
|
void cogorc_downsample_vert_halfsite_2tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
int n);
|
|
void cogorc_downsample_vert_cosite_3tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n);
|
|
void cogorc_downsample_vert_halfsite_4tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4,
|
|
int n);
|
|
void cogorc_upsample_horiz_cosite_1tap (guint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int n);
|
|
void cogorc_upsample_horiz_cosite (guint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
int n);
|
|
void cogorc_upsample_vert_avgub (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
int n);
|
|
void orc_unpack_yuyv_y (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n);
|
|
void orc_unpack_yuyv_u (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void orc_unpack_yuyv_v (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void orc_pack_yuyv (orc_uint32 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n);
|
|
void orc_unpack_uyvy_y (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n);
|
|
void orc_unpack_uyvy_u (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void orc_unpack_uyvy_v (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void orc_pack_uyvy (orc_uint32 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n);
|
|
void orc_addc_convert_u8_s16 (orc_uint8 * ORC_RESTRICT d1,
|
|
const gint16 * ORC_RESTRICT s1, int n);
|
|
void orc_subc_convert_s16_u8 (gint16 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int n);
|
|
void orc_splat_u8_ns (orc_uint8 * ORC_RESTRICT d1, int p1, int n);
|
|
void orc_splat_s16_ns (gint16 * ORC_RESTRICT d1, int p1, int n);
|
|
void orc_matrix2_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, int p1, int p2, int p3, int n);
|
|
void orc_matrix2_11_u8 (guint8 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2, int p1,
|
|
int p2, int n);
|
|
void orc_matrix2_12_u8 (guint8 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2, int p1,
|
|
int p2, int n);
|
|
void orc_matrix3_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1,
|
|
int p2, int p3, int p4, int n);
|
|
void orc_matrix3_100_u8 (guint8 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
|
|
const guint8 * ORC_RESTRICT s3, int p1, int p2, int p3, int n);
|
|
void orc_matrix3_100_offset_u8 (guint8 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
|
|
const guint8 * ORC_RESTRICT s3, int p1, int p2, int p3, int p4, int p5,
|
|
int n);
|
|
void orc_matrix3_000_u8 (guint8 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
|
|
const guint8 * ORC_RESTRICT s3, int p1, int p2, int p3, int p4, int p5,
|
|
int n);
|
|
void orc_pack_123x (guint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int p1, int n);
|
|
void orc_pack_x123 (guint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int p1, int n);
|
|
void cogorc_combine2_u8 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
int p1, int p2, int n);
|
|
void cogorc_combine4_u8 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4,
|
|
int p1, int p2, int p3, int p4, int n);
|
|
void cogorc_unpack_axyz_0 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void cogorc_unpack_axyz_1 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void cogorc_unpack_axyz_2 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void cogorc_unpack_axyz_3 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n);
|
|
void cogorc_resample_horiz_1tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int p1, int p2, int n);
|
|
void cogorc_resample_horiz_2tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int p1, int p2, int n);
|
|
void cogorc_convert_I420_UYVY (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint16 * ORC_RESTRICT s1,
|
|
const orc_uint16 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n);
|
|
void cogorc_convert_I420_YUY2 (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint16 * ORC_RESTRICT s1,
|
|
const orc_uint16 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n);
|
|
void cogorc_convert_I420_AYUV (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n);
|
|
void cogorc_convert_YUY2_I420 (orc_uint16 * ORC_RESTRICT d1,
|
|
orc_uint16 * ORC_RESTRICT d2, orc_uint8 * ORC_RESTRICT d3,
|
|
orc_uint8 * ORC_RESTRICT d4, const orc_uint32 * ORC_RESTRICT s1,
|
|
const orc_uint32 * ORC_RESTRICT s2, int n);
|
|
void cogorc_convert_UYVY_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_planar_chroma_420_422 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_planar_chroma_420_444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_planar_chroma_422_444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_planar_chroma_444_422 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_planar_chroma_444_420 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride, int n, int m);
|
|
void cogorc_planar_chroma_422_420 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride, int n, int m);
|
|
void cogorc_convert_YUY2_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_UYVY_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_YUY2_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m);
|
|
void cogorc_convert_UYVY_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m);
|
|
void cogorc_convert_YUY2_Y444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint16 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m);
|
|
void cogorc_convert_UYVY_Y444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint16 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m);
|
|
void cogorc_convert_UYVY_I420 (orc_uint16 * ORC_RESTRICT d1,
|
|
orc_uint16 * ORC_RESTRICT d2, orc_uint8 * ORC_RESTRICT d3,
|
|
orc_uint8 * ORC_RESTRICT d4, const orc_uint32 * ORC_RESTRICT s1,
|
|
const orc_uint32 * ORC_RESTRICT s2, int n);
|
|
void cogorc_convert_AYUV_I420 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, orc_uint8 * ORC_RESTRICT d4, int d4_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s2, int s2_stride, int n, int m);
|
|
void cogorc_convert_AYUV_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_AYUV_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_AYUV_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m);
|
|
void cogorc_convert_AYUV_Y444 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m);
|
|
void cogorc_convert_Y42B_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m);
|
|
void cogorc_convert_Y42B_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m);
|
|
void cogorc_convert_Y42B_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m);
|
|
void cogorc_convert_Y444_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint16 * ORC_RESTRICT s3, int s3_stride, int n, int m);
|
|
void cogorc_convert_Y444_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint16 * ORC_RESTRICT s3, int s3_stride, int n, int m);
|
|
void cogorc_convert_Y444_AYUV (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m);
|
|
void cogorc_convert_AYUV_ARGB (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_AYUV_BGRA (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_AYUV_ABGR (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_AYUV_RGBA (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m);
|
|
void cogorc_convert_I420_BGRA (orc_uint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n);
|
|
void cogorc_convert_I420_BGRA_avg (orc_uint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4,
|
|
const orc_uint8 * ORC_RESTRICT s5, int n);
|
|
|
|
|
|
/* begin Orc C target preamble */
|
|
#define ORC_CLAMP(x,a,b) ((x)<(a) ? (a) : ((x)>(b) ? (b) : (x)))
|
|
#define ORC_ABS(a) ((a)<0 ? -(a) : (a))
|
|
#define ORC_MIN(a,b) ((a)<(b) ? (a) : (b))
|
|
#define ORC_MAX(a,b) ((a)>(b) ? (a) : (b))
|
|
#define ORC_SB_MAX 127
|
|
#define ORC_SB_MIN (-1-ORC_SB_MAX)
|
|
#define ORC_UB_MAX 255
|
|
#define ORC_UB_MIN 0
|
|
#define ORC_SW_MAX 32767
|
|
#define ORC_SW_MIN (-1-ORC_SW_MAX)
|
|
#define ORC_UW_MAX 65535
|
|
#define ORC_UW_MIN 0
|
|
#define ORC_SL_MAX 2147483647
|
|
#define ORC_SL_MIN (-1-ORC_SL_MAX)
|
|
#define ORC_UL_MAX 4294967295U
|
|
#define ORC_UL_MIN 0
|
|
#define ORC_CLAMP_SB(x) ORC_CLAMP(x,ORC_SB_MIN,ORC_SB_MAX)
|
|
#define ORC_CLAMP_UB(x) ORC_CLAMP(x,ORC_UB_MIN,ORC_UB_MAX)
|
|
#define ORC_CLAMP_SW(x) ORC_CLAMP(x,ORC_SW_MIN,ORC_SW_MAX)
|
|
#define ORC_CLAMP_UW(x) ORC_CLAMP(x,ORC_UW_MIN,ORC_UW_MAX)
|
|
#define ORC_CLAMP_SL(x) ORC_CLAMP(x,ORC_SL_MIN,ORC_SL_MAX)
|
|
#define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX)
|
|
#define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))
|
|
#define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))
|
|
#define ORC_SWAP_Q(x) ((((x)&ORC_UINT64_C(0xff))<<56) | (((x)&ORC_UINT64_C(0xff00))<<40) | (((x)&ORC_UINT64_C(0xff0000))<<24) | (((x)&ORC_UINT64_C(0xff000000))<<8) | (((x)&ORC_UINT64_C(0xff00000000))>>8) | (((x)&ORC_UINT64_C(0xff0000000000))>>24) | (((x)&ORC_UINT64_C(0xff000000000000))>>40) | (((x)&ORC_UINT64_C(0xff00000000000000))>>56))
|
|
#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))
|
|
#define ORC_DENORMAL(x) ((x) & ((((x)&0x7f800000) == 0) ? 0xff800000 : 0xffffffff))
|
|
#define ORC_ISNAN(x) ((((x)&0x7f800000) == 0x7f800000) && (((x)&0x007fffff) != 0))
|
|
#define ORC_DENORMAL_DOUBLE(x) ((x) & ((((x)&ORC_UINT64_C(0x7ff0000000000000)) == 0) ? ORC_UINT64_C(0xfff0000000000000) : ORC_UINT64_C(0xffffffffffffffff)))
|
|
#define ORC_ISNAN_DOUBLE(x) ((((x)&ORC_UINT64_C(0x7ff0000000000000)) == ORC_UINT64_C(0x7ff0000000000000)) && (((x)&ORC_UINT64_C(0x000fffffffffffff)) != 0))
|
|
#ifndef ORC_RESTRICT
|
|
#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
|
|
#define ORC_RESTRICT restrict
|
|
#elif defined(__GNUC__) && __GNUC__ >= 4
|
|
#define ORC_RESTRICT __restrict__
|
|
#else
|
|
#define ORC_RESTRICT
|
|
#endif
|
|
#endif
|
|
/* end Orc C target preamble */
|
|
|
|
|
|
|
|
/* cogorc_memcpy_2d */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_memcpy_2d (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: copyb */
|
|
var33 = var32;
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_memcpy_2d (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: copyb */
|
|
var33 = var32;
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_memcpy_2d (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_memcpy_2d");
|
|
orc_program_set_backup_function (p, _backup_cogorc_memcpy_2d);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
|
|
orc_program_append_2 (p, "copyb", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_downsample_horiz_cosite_1tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_downsample_horiz_cosite_1tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var32 = ptr4[i];
|
|
/* 1: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var32.i;
|
|
var33 = _src.x2[0];
|
|
}
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_downsample_horiz_cosite_1tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var32 = ptr4[i];
|
|
/* 1: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var32.i;
|
|
var33 = _src.x2[0];
|
|
}
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_downsample_horiz_cosite_1tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_downsample_horiz_cosite_1tap");
|
|
orc_program_set_backup_function (p,
|
|
_backup_cogorc_downsample_horiz_cosite_1tap);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_downsample_horiz_cosite_3tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_downsample_horiz_cosite_3tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
|
|
int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_union16 var43;
|
|
orc_int8 var44;
|
|
orc_int8 var45;
|
|
orc_int8 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
ptr5 = (orc_union16 *) s2;
|
|
|
|
/* 9: loadpw */
|
|
var40.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
/* 13: loadpw */
|
|
var41.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var38 = ptr4[i];
|
|
/* 1: copyw */
|
|
var43.i = var38.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var44 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var45 = _src.x2[1];
|
|
}
|
|
/* 4: loadw */
|
|
var39 = ptr5[i];
|
|
/* 5: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var46 = _src.x2[0];
|
|
}
|
|
/* 6: convubw */
|
|
var47.i = (orc_uint8) var44;
|
|
/* 7: convubw */
|
|
var48.i = (orc_uint8) var45;
|
|
/* 8: convubw */
|
|
var49.i = (orc_uint8) var46;
|
|
/* 10: mullw */
|
|
var50.i = (var48.i * var40.i) & 0xffff;
|
|
/* 11: addw */
|
|
var51.i = var47.i + var49.i;
|
|
/* 12: addw */
|
|
var52.i = var51.i + var50.i;
|
|
/* 14: addw */
|
|
var53.i = var52.i + var41.i;
|
|
/* 15: shrsw */
|
|
var54.i = var53.i >> 2;
|
|
/* 16: convsuswb */
|
|
var42 = ORC_CLAMP_UB (var54.i);
|
|
/* 17: storeb */
|
|
ptr0[i] = var42;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_downsample_horiz_cosite_3tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_union16 var43;
|
|
orc_int8 var44;
|
|
orc_int8 var45;
|
|
orc_int8 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
ptr5 = (orc_union16 *) ex->arrays[5];
|
|
|
|
/* 9: loadpw */
|
|
var40.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
/* 13: loadpw */
|
|
var41.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var38 = ptr4[i];
|
|
/* 1: copyw */
|
|
var43.i = var38.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var44 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var45 = _src.x2[1];
|
|
}
|
|
/* 4: loadw */
|
|
var39 = ptr5[i];
|
|
/* 5: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var46 = _src.x2[0];
|
|
}
|
|
/* 6: convubw */
|
|
var47.i = (orc_uint8) var44;
|
|
/* 7: convubw */
|
|
var48.i = (orc_uint8) var45;
|
|
/* 8: convubw */
|
|
var49.i = (orc_uint8) var46;
|
|
/* 10: mullw */
|
|
var50.i = (var48.i * var40.i) & 0xffff;
|
|
/* 11: addw */
|
|
var51.i = var47.i + var49.i;
|
|
/* 12: addw */
|
|
var52.i = var51.i + var50.i;
|
|
/* 14: addw */
|
|
var53.i = var52.i + var41.i;
|
|
/* 15: shrsw */
|
|
var54.i = var53.i >> 2;
|
|
/* 16: convsuswb */
|
|
var42 = ORC_CLAMP_UB (var54.i);
|
|
/* 17: storeb */
|
|
ptr0[i] = var42;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_downsample_horiz_cosite_3tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
|
|
int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_downsample_horiz_cosite_3tap");
|
|
orc_program_set_backup_function (p,
|
|
_backup_cogorc_downsample_horiz_cosite_3tap);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 2, "s2");
|
|
orc_program_add_constant (p, 4, 0x00000002, "c1");
|
|
orc_program_add_temporary (p, 1, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
orc_program_add_temporary (p, 2, "t5");
|
|
orc_program_add_temporary (p, 2, "t6");
|
|
|
|
orc_program_append_2 (p, "copyw", 0, ORC_VAR_T4, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T1, ORC_VAR_T4,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T2, ORC_VAR_T4,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T3, ORC_VAR_S2,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T4, ORC_VAR_T1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T5, ORC_VAR_T2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T6, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T4,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_downsample_420_jpeg */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_downsample_420_jpeg (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
|
|
int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_int8 var45;
|
|
orc_int8 var46;
|
|
orc_int8 var47;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
ptr5 = (orc_union16 *) s2;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var37 = ptr4[i];
|
|
/* 1: copyw */
|
|
var40.i = var37.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.i;
|
|
var41 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.i;
|
|
var42 = _src.x2[1];
|
|
}
|
|
/* 4: avgub */
|
|
var43 = ((orc_uint8) var41 + (orc_uint8) var42 + 1) >> 1;
|
|
/* 5: loadw */
|
|
var38 = ptr5[i];
|
|
/* 6: copyw */
|
|
var44.i = var38.i;
|
|
/* 7: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var44.i;
|
|
var45 = _src.x2[0];
|
|
}
|
|
/* 8: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var44.i;
|
|
var46 = _src.x2[1];
|
|
}
|
|
/* 9: avgub */
|
|
var47 = ((orc_uint8) var45 + (orc_uint8) var46 + 1) >> 1;
|
|
/* 10: avgub */
|
|
var39 = ((orc_uint8) var43 + (orc_uint8) var47 + 1) >> 1;
|
|
/* 11: storeb */
|
|
ptr0[i] = var39;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_downsample_420_jpeg (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_int8 var45;
|
|
orc_int8 var46;
|
|
orc_int8 var47;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
ptr5 = (orc_union16 *) ex->arrays[5];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var37 = ptr4[i];
|
|
/* 1: copyw */
|
|
var40.i = var37.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.i;
|
|
var41 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.i;
|
|
var42 = _src.x2[1];
|
|
}
|
|
/* 4: avgub */
|
|
var43 = ((orc_uint8) var41 + (orc_uint8) var42 + 1) >> 1;
|
|
/* 5: loadw */
|
|
var38 = ptr5[i];
|
|
/* 6: copyw */
|
|
var44.i = var38.i;
|
|
/* 7: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var44.i;
|
|
var45 = _src.x2[0];
|
|
}
|
|
/* 8: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var44.i;
|
|
var46 = _src.x2[1];
|
|
}
|
|
/* 9: avgub */
|
|
var47 = ((orc_uint8) var45 + (orc_uint8) var46 + 1) >> 1;
|
|
/* 10: avgub */
|
|
var39 = ((orc_uint8) var43 + (orc_uint8) var47 + 1) >> 1;
|
|
/* 11: storeb */
|
|
ptr0[i] = var39;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_downsample_420_jpeg (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, const orc_uint16 * ORC_RESTRICT s2,
|
|
int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_downsample_420_jpeg");
|
|
orc_program_set_backup_function (p, _backup_cogorc_downsample_420_jpeg);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 2, "s2");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 1, "t4");
|
|
orc_program_add_temporary (p, 1, "t5");
|
|
|
|
orc_program_append_2 (p, "copyw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T3, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "copyw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T4, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T5, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_downsample_vert_halfsite_2tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_downsample_vert_halfsite_2tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: loadb */
|
|
var33 = ptr5[i];
|
|
/* 2: avgub */
|
|
var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_downsample_vert_halfsite_2tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: loadb */
|
|
var33 = ptr5[i];
|
|
/* 2: avgub */
|
|
var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_downsample_vert_halfsite_2tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_downsample_vert_halfsite_2tap");
|
|
orc_program_set_backup_function (p,
|
|
_backup_cogorc_downsample_vert_halfsite_2tap);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_downsample_vert_cosite_3tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_downsample_vert_cosite_3tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 6: loadpw */
|
|
var38.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
/* 10: loadpw */
|
|
var39.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var35 = ptr4[i];
|
|
/* 1: convubw */
|
|
var41.i = (orc_uint8) var35;
|
|
/* 2: loadb */
|
|
var36 = ptr5[i];
|
|
/* 3: convubw */
|
|
var42.i = (orc_uint8) var36;
|
|
/* 4: loadb */
|
|
var37 = ptr6[i];
|
|
/* 5: convubw */
|
|
var43.i = (orc_uint8) var37;
|
|
/* 7: mullw */
|
|
var44.i = (var42.i * var38.i) & 0xffff;
|
|
/* 8: addw */
|
|
var45.i = var41.i + var43.i;
|
|
/* 9: addw */
|
|
var46.i = var45.i + var44.i;
|
|
/* 11: addw */
|
|
var47.i = var46.i + var39.i;
|
|
/* 12: shrsw */
|
|
var48.i = var47.i >> 2;
|
|
/* 13: convsuswb */
|
|
var40 = ORC_CLAMP_UB (var48.i);
|
|
/* 14: storeb */
|
|
ptr0[i] = var40;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_downsample_vert_cosite_3tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 6: loadpw */
|
|
var38.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
/* 10: loadpw */
|
|
var39.i = (int) 0x00000002; /* 2 or 9.88131e-324f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var35 = ptr4[i];
|
|
/* 1: convubw */
|
|
var41.i = (orc_uint8) var35;
|
|
/* 2: loadb */
|
|
var36 = ptr5[i];
|
|
/* 3: convubw */
|
|
var42.i = (orc_uint8) var36;
|
|
/* 4: loadb */
|
|
var37 = ptr6[i];
|
|
/* 5: convubw */
|
|
var43.i = (orc_uint8) var37;
|
|
/* 7: mullw */
|
|
var44.i = (var42.i * var38.i) & 0xffff;
|
|
/* 8: addw */
|
|
var45.i = var41.i + var43.i;
|
|
/* 9: addw */
|
|
var46.i = var45.i + var44.i;
|
|
/* 11: addw */
|
|
var47.i = var46.i + var39.i;
|
|
/* 12: shrsw */
|
|
var48.i = var47.i >> 2;
|
|
/* 13: convsuswb */
|
|
var40 = ORC_CLAMP_UB (var48.i);
|
|
/* 14: storeb */
|
|
ptr0[i] = var40;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_downsample_vert_cosite_3tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_downsample_vert_cosite_3tap");
|
|
orc_program_set_backup_function (p,
|
|
_backup_cogorc_downsample_vert_cosite_3tap);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_constant (p, 4, 0x00000002, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T3, ORC_VAR_S3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_downsample_vert_halfsite_4tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_downsample_vert_halfsite_4tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
ptr7 = (orc_int8 *) s4;
|
|
|
|
/* 9: loadpw */
|
|
var40.i = (int) 0x0000001a; /* 26 or 1.28457e-322f */
|
|
/* 12: loadpw */
|
|
var41.i = (int) 0x00000006; /* 6 or 2.96439e-323f */
|
|
/* 15: loadpw */
|
|
var42.i = (int) 0x00000020; /* 32 or 1.58101e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 2: loadb */
|
|
var37 = ptr5[i];
|
|
/* 3: convubw */
|
|
var45.i = (orc_uint8) var37;
|
|
/* 4: loadb */
|
|
var38 = ptr6[i];
|
|
/* 5: convubw */
|
|
var46.i = (orc_uint8) var38;
|
|
/* 6: loadb */
|
|
var39 = ptr7[i];
|
|
/* 7: convubw */
|
|
var47.i = (orc_uint8) var39;
|
|
/* 8: addw */
|
|
var48.i = var45.i + var46.i;
|
|
/* 10: mullw */
|
|
var49.i = (var48.i * var40.i) & 0xffff;
|
|
/* 11: addw */
|
|
var50.i = var44.i + var47.i;
|
|
/* 13: mullw */
|
|
var51.i = (var50.i * var41.i) & 0xffff;
|
|
/* 14: addw */
|
|
var52.i = var49.i + var51.i;
|
|
/* 16: addw */
|
|
var53.i = var52.i + var42.i;
|
|
/* 17: shrsw */
|
|
var54.i = var53.i >> 6;
|
|
/* 18: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var54.i);
|
|
/* 19: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_downsample_vert_halfsite_4tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
ptr7 = (orc_int8 *) ex->arrays[7];
|
|
|
|
/* 9: loadpw */
|
|
var40.i = (int) 0x0000001a; /* 26 or 1.28457e-322f */
|
|
/* 12: loadpw */
|
|
var41.i = (int) 0x00000006; /* 6 or 2.96439e-323f */
|
|
/* 15: loadpw */
|
|
var42.i = (int) 0x00000020; /* 32 or 1.58101e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 2: loadb */
|
|
var37 = ptr5[i];
|
|
/* 3: convubw */
|
|
var45.i = (orc_uint8) var37;
|
|
/* 4: loadb */
|
|
var38 = ptr6[i];
|
|
/* 5: convubw */
|
|
var46.i = (orc_uint8) var38;
|
|
/* 6: loadb */
|
|
var39 = ptr7[i];
|
|
/* 7: convubw */
|
|
var47.i = (orc_uint8) var39;
|
|
/* 8: addw */
|
|
var48.i = var45.i + var46.i;
|
|
/* 10: mullw */
|
|
var49.i = (var48.i * var40.i) & 0xffff;
|
|
/* 11: addw */
|
|
var50.i = var44.i + var47.i;
|
|
/* 13: mullw */
|
|
var51.i = (var50.i * var41.i) & 0xffff;
|
|
/* 14: addw */
|
|
var52.i = var49.i + var51.i;
|
|
/* 16: addw */
|
|
var53.i = var52.i + var42.i;
|
|
/* 17: shrsw */
|
|
var54.i = var53.i >> 6;
|
|
/* 18: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var54.i);
|
|
/* 19: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_downsample_vert_halfsite_4tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_downsample_vert_halfsite_4tap");
|
|
orc_program_set_backup_function (p,
|
|
_backup_cogorc_downsample_vert_halfsite_4tap);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_source (p, 1, "s4");
|
|
orc_program_add_constant (p, 4, 0x0000001a, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000006, "c2");
|
|
orc_program_add_constant (p, 4, 0x00000020, "c3");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T3, ORC_VAR_S3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T4, ORC_VAR_S4, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T2,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->arrays[ORC_VAR_S4] = (void *) s4;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_upsample_horiz_cosite_1tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_upsample_horiz_cosite_1tap (guint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
|
|
ptr0 = (orc_union16 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: copyb */
|
|
var35 = var33;
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35;
|
|
_dest.x2[1] = var35;
|
|
var34.i = _dest.i;
|
|
}
|
|
/* 3: storew */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_upsample_horiz_cosite_1tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
|
|
ptr0 = (orc_union16 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: copyb */
|
|
var35 = var33;
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35;
|
|
_dest.x2[1] = var35;
|
|
var34.i = _dest.i;
|
|
}
|
|
/* 3: storew */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_upsample_horiz_cosite_1tap (guint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_upsample_horiz_cosite_1tap");
|
|
orc_program_set_backup_function (p,
|
|
_backup_cogorc_upsample_horiz_cosite_1tap);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_temporary (p, 1, "t1");
|
|
|
|
orc_program_append_2 (p, "copyb", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_upsample_horiz_cosite */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_upsample_horiz_cosite (guint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2, int n)
|
|
{
|
|
int i;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
|
|
ptr0 = (orc_union16 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: copyb */
|
|
var37 = var34;
|
|
/* 2: loadb */
|
|
var35 = ptr5[i];
|
|
/* 3: avgub */
|
|
var38 = ((orc_uint8) var37 + (orc_uint8) var35 + 1) >> 1;
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37;
|
|
_dest.x2[1] = var38;
|
|
var36.i = _dest.i;
|
|
}
|
|
/* 5: storew */
|
|
ptr0[i] = var36;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_upsample_horiz_cosite (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
|
|
ptr0 = (orc_union16 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: copyb */
|
|
var37 = var34;
|
|
/* 2: loadb */
|
|
var35 = ptr5[i];
|
|
/* 3: avgub */
|
|
var38 = ((orc_uint8) var37 + (orc_uint8) var35 + 1) >> 1;
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37;
|
|
_dest.x2[1] = var38;
|
|
var36.i = _dest.i;
|
|
}
|
|
/* 5: storew */
|
|
ptr0[i] = var36;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_upsample_horiz_cosite (guint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_upsample_horiz_cosite");
|
|
orc_program_set_backup_function (p, _backup_cogorc_upsample_horiz_cosite);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_temporary (p, 1, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
|
|
orc_program_append_2 (p, "copyb", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_upsample_vert_avgub */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_upsample_vert_avgub (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: loadb */
|
|
var33 = ptr5[i];
|
|
/* 2: avgub */
|
|
var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_upsample_vert_avgub (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: loadb */
|
|
var33 = ptr5[i];
|
|
/* 2: avgub */
|
|
var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_upsample_vert_avgub (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_upsample_vert_avgub");
|
|
orc_program_set_backup_function (p, _backup_cogorc_upsample_vert_avgub);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_unpack_yuyv_y */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_unpack_yuyv_y (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var32 = ptr4[i];
|
|
/* 1: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var32.i;
|
|
var33 = _src.x2[0];
|
|
}
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_unpack_yuyv_y (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var32 = ptr4[i];
|
|
/* 1: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var32.i;
|
|
var33 = _src.x2[0];
|
|
}
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_unpack_yuyv_y (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_unpack_yuyv_y");
|
|
orc_program_set_backup_function (p, _backup_orc_unpack_yuyv_y);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_unpack_yuyv_u */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_unpack_yuyv_u (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_unpack_yuyv_u (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_unpack_yuyv_u (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_unpack_yuyv_u");
|
|
orc_program_set_backup_function (p, _backup_orc_unpack_yuyv_u);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_unpack_yuyv_v */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_unpack_yuyv_v (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_unpack_yuyv_v (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_unpack_yuyv_v (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_unpack_yuyv_v");
|
|
orc_program_set_backup_function (p, _backup_orc_unpack_yuyv_v);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_pack_yuyv */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_pack_yuyv (orc_uint32 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_int8 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var37 = ptr4[i];
|
|
/* 1: copyw */
|
|
var41.i = var37.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var42 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var43 = _src.x2[1];
|
|
}
|
|
/* 4: loadb */
|
|
var38 = ptr5[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var42;
|
|
_dest.x2[1] = var38;
|
|
var44.i = _dest.i;
|
|
}
|
|
/* 6: loadb */
|
|
var39 = ptr6[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43;
|
|
_dest.x2[1] = var39;
|
|
var45.i = _dest.i;
|
|
}
|
|
/* 8: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var44.i;
|
|
_dest.x2[1] = var45.i;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 9: storel */
|
|
ptr0[i] = var40;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_pack_yuyv (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_int8 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var37 = ptr4[i];
|
|
/* 1: copyw */
|
|
var41.i = var37.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var42 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var43 = _src.x2[1];
|
|
}
|
|
/* 4: loadb */
|
|
var38 = ptr5[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var42;
|
|
_dest.x2[1] = var38;
|
|
var44.i = _dest.i;
|
|
}
|
|
/* 6: loadb */
|
|
var39 = ptr6[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43;
|
|
_dest.x2[1] = var39;
|
|
var45.i = _dest.i;
|
|
}
|
|
/* 8: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var44.i;
|
|
_dest.x2[1] = var45.i;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 9: storel */
|
|
ptr0[i] = var40;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_pack_yuyv (orc_uint32 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_pack_yuyv");
|
|
orc_program_set_backup_function (p, _backup_orc_pack_yuyv);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_temporary (p, 1, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
orc_program_add_temporary (p, 2, "t5");
|
|
|
|
orc_program_append_2 (p, "copyw", 0, ORC_VAR_T5, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T1, ORC_VAR_T5,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T2, ORC_VAR_T5,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_unpack_uyvy_y */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_unpack_uyvy_y (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var32 = ptr4[i];
|
|
/* 1: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var32.i;
|
|
var33 = _src.x2[1];
|
|
}
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_unpack_uyvy_y (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var32 = ptr4[i];
|
|
/* 1: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var32.i;
|
|
var33 = _src.x2[1];
|
|
}
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_unpack_uyvy_y (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_unpack_uyvy_y");
|
|
orc_program_set_backup_function (p, _backup_orc_unpack_uyvy_y);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_unpack_uyvy_u */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_unpack_uyvy_u (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_unpack_uyvy_u (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_unpack_uyvy_u (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_unpack_uyvy_u");
|
|
orc_program_set_backup_function (p, _backup_orc_unpack_uyvy_u);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_unpack_uyvy_v */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_unpack_uyvy_v (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_unpack_uyvy_v (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_unpack_uyvy_v (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_unpack_uyvy_v");
|
|
orc_program_set_backup_function (p, _backup_orc_unpack_uyvy_v);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_pack_uyvy */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_pack_uyvy (orc_uint32 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_int8 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var37 = ptr4[i];
|
|
/* 1: copyw */
|
|
var41.i = var37.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var42 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var43 = _src.x2[1];
|
|
}
|
|
/* 4: loadb */
|
|
var38 = ptr5[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38;
|
|
_dest.x2[1] = var42;
|
|
var44.i = _dest.i;
|
|
}
|
|
/* 6: loadb */
|
|
var39 = ptr6[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39;
|
|
_dest.x2[1] = var43;
|
|
var45.i = _dest.i;
|
|
}
|
|
/* 8: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var44.i;
|
|
_dest.x2[1] = var45.i;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 9: storel */
|
|
ptr0[i] = var40;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_pack_uyvy (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_int8 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var37 = ptr4[i];
|
|
/* 1: copyw */
|
|
var41.i = var37.i;
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var42 = _src.x2[0];
|
|
}
|
|
/* 3: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.i;
|
|
var43 = _src.x2[1];
|
|
}
|
|
/* 4: loadb */
|
|
var38 = ptr5[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38;
|
|
_dest.x2[1] = var42;
|
|
var44.i = _dest.i;
|
|
}
|
|
/* 6: loadb */
|
|
var39 = ptr6[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39;
|
|
_dest.x2[1] = var43;
|
|
var45.i = _dest.i;
|
|
}
|
|
/* 8: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var44.i;
|
|
_dest.x2[1] = var45.i;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 9: storel */
|
|
ptr0[i] = var40;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_pack_uyvy (orc_uint32 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_pack_uyvy");
|
|
orc_program_set_backup_function (p, _backup_orc_pack_uyvy);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_temporary (p, 1, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
orc_program_add_temporary (p, 2, "t5");
|
|
|
|
orc_program_append_2 (p, "copyw", 0, ORC_VAR_T5, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_T1, ORC_VAR_T5,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_T2, ORC_VAR_T5,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T3, ORC_VAR_S2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T4, ORC_VAR_S3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_addc_convert_u8_s16 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_addc_convert_u8_s16 (orc_uint8 * ORC_RESTRICT d1,
|
|
const gint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union16 *) s1;
|
|
|
|
/* 1: loadpw */
|
|
var34.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var33 = ptr4[i];
|
|
/* 2: addw */
|
|
var36.i = var33.i + var34.i;
|
|
/* 3: convsuswb */
|
|
var35 = ORC_CLAMP_UB (var36.i);
|
|
/* 4: storeb */
|
|
ptr0[i] = var35;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_addc_convert_u8_s16 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
|
|
/* 1: loadpw */
|
|
var34.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var33 = ptr4[i];
|
|
/* 2: addw */
|
|
var36.i = var33.i + var34.i;
|
|
/* 3: convsuswb */
|
|
var35 = ORC_CLAMP_UB (var36.i);
|
|
/* 4: storeb */
|
|
ptr0[i] = var35;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_addc_convert_u8_s16 (orc_uint8 * ORC_RESTRICT d1,
|
|
const gint16 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_addc_convert_u8_s16");
|
|
orc_program_set_backup_function (p, _backup_orc_addc_convert_u8_s16);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_subc_convert_s16_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_subc_convert_s16_u8 (gint16 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
orc_union16 var35;
|
|
orc_union16 var36;
|
|
|
|
ptr0 = (orc_union16 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
|
|
/* 2: loadpw */
|
|
var34.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: convubw */
|
|
var36.i = (orc_uint8) var33;
|
|
/* 3: subw */
|
|
var35.i = var36.i - var34.i;
|
|
/* 4: storew */
|
|
ptr0[i] = var35;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_subc_convert_s16_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
orc_union16 var35;
|
|
orc_union16 var36;
|
|
|
|
ptr0 = (orc_union16 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
|
|
/* 2: loadpw */
|
|
var34.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: convubw */
|
|
var36.i = (orc_uint8) var33;
|
|
/* 3: subw */
|
|
var35.i = var36.i - var34.i;
|
|
/* 4: storew */
|
|
ptr0[i] = var35;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_subc_convert_s16_u8 (gint16 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_subc_convert_s16_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_subc_convert_s16_u8);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_splat_u8_ns */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_splat_u8_ns (orc_uint8 * ORC_RESTRICT d1, int p1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
|
|
/* 0: loadpb */
|
|
var32 = p1;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 1: copyb */
|
|
var33 = var32;
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_splat_u8_ns (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
|
|
/* 0: loadpb */
|
|
var32 = ex->params[24];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 1: copyb */
|
|
var33 = var32;
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_splat_u8_ns (orc_uint8 * ORC_RESTRICT d1, int p1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_splat_u8_ns");
|
|
orc_program_set_backup_function (p, _backup_orc_splat_u8_ns);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_parameter (p, 1, "p1");
|
|
|
|
orc_program_append_2 (p, "copyb", 0, ORC_VAR_D1, ORC_VAR_P1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_splat_s16_ns */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_splat_s16_ns (gint16 * ORC_RESTRICT d1, int p1, int n)
|
|
{
|
|
int i;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 var32;
|
|
orc_union16 var33;
|
|
|
|
ptr0 = (orc_union16 *) d1;
|
|
|
|
/* 0: loadpw */
|
|
var32.i = p1;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 1: copyw */
|
|
var33.i = var32.i;
|
|
/* 2: storew */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_splat_s16_ns (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 var32;
|
|
orc_union16 var33;
|
|
|
|
ptr0 = (orc_union16 *) ex->arrays[0];
|
|
|
|
/* 0: loadpw */
|
|
var32.i = ex->params[24];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 1: copyw */
|
|
var33.i = var32.i;
|
|
/* 2: storew */
|
|
ptr0[i] = var33;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_splat_s16_ns (gint16 * ORC_RESTRICT d1, int p1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_splat_s16_ns");
|
|
orc_program_set_backup_function (p, _backup_orc_splat_s16_ns);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
|
|
orc_program_append_2 (p, "copyw", 0, ORC_VAR_D1, ORC_VAR_P1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_matrix2_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_matrix2_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, int p1, int p2, int p3, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
|
|
/* 2: loadpw */
|
|
var35.i = p1;
|
|
/* 6: loadpw */
|
|
var37.i = p2;
|
|
/* 9: loadpw */
|
|
var38.i = p3;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var40.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var41.i = (var40.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var42.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var43.i = (var42.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var44.i = var41.i + var43.i;
|
|
/* 10: addw */
|
|
var45.i = var44.i + var38.i;
|
|
/* 11: shrsw */
|
|
var46.i = var45.i >> 6;
|
|
/* 12: convsuswb */
|
|
var39 = ORC_CLAMP_UB (var46.i);
|
|
/* 13: storeb */
|
|
ptr0[i] = var39;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_matrix2_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
|
|
/* 2: loadpw */
|
|
var35.i = ex->params[24];
|
|
/* 6: loadpw */
|
|
var37.i = ex->params[25];
|
|
/* 9: loadpw */
|
|
var38.i = ex->params[26];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var40.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var41.i = (var40.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var42.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var43.i = (var42.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var44.i = var41.i + var43.i;
|
|
/* 10: addw */
|
|
var45.i = var44.i + var38.i;
|
|
/* 11: shrsw */
|
|
var46.i = var45.i >> 6;
|
|
/* 12: convsuswb */
|
|
var39 = ORC_CLAMP_UB (var46.i);
|
|
/* 13: storeb */
|
|
ptr0[i] = var39;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_matrix2_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, int p1, int p2, int p3, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_matrix2_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_matrix2_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_constant (p, 4, 0x00000006, "c1");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_parameter (p, 2, "p3");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
ex->params[ORC_VAR_P3] = p3;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_matrix2_11_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_matrix2_11_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, int p1, int p2, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
|
|
/* 2: loadpw */
|
|
var37.i = (int) 0x00000010; /* 16 or 7.90505e-323f */
|
|
/* 4: loadpw */
|
|
var38.i = p1;
|
|
/* 8: loadpw */
|
|
var40.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 10: loadpw */
|
|
var41.i = p2;
|
|
/* 13: loadpw */
|
|
var42.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 3: subw */
|
|
var45.i = var44.i - var37.i;
|
|
/* 5: mullw */
|
|
var46.i = (var45.i * var38.i) & 0xffff;
|
|
/* 6: loadb */
|
|
var39 = ptr5[i];
|
|
/* 7: convubw */
|
|
var47.i = (orc_uint8) var39;
|
|
/* 9: subw */
|
|
var48.i = var47.i - var40.i;
|
|
/* 11: mullw */
|
|
var49.i = (var48.i * var41.i) & 0xffff;
|
|
/* 12: addw */
|
|
var50.i = var46.i + var49.i;
|
|
/* 14: addw */
|
|
var51.i = var50.i + var42.i;
|
|
/* 15: shrsw */
|
|
var52.i = var51.i >> 8;
|
|
/* 16: addw */
|
|
var53.i = var52.i + var45.i;
|
|
/* 17: addw */
|
|
var54.i = var53.i + var48.i;
|
|
/* 18: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var54.i);
|
|
/* 19: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_matrix2_11_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
|
|
/* 2: loadpw */
|
|
var37.i = (int) 0x00000010; /* 16 or 7.90505e-323f */
|
|
/* 4: loadpw */
|
|
var38.i = ex->params[24];
|
|
/* 8: loadpw */
|
|
var40.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 10: loadpw */
|
|
var41.i = ex->params[25];
|
|
/* 13: loadpw */
|
|
var42.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 3: subw */
|
|
var45.i = var44.i - var37.i;
|
|
/* 5: mullw */
|
|
var46.i = (var45.i * var38.i) & 0xffff;
|
|
/* 6: loadb */
|
|
var39 = ptr5[i];
|
|
/* 7: convubw */
|
|
var47.i = (orc_uint8) var39;
|
|
/* 9: subw */
|
|
var48.i = var47.i - var40.i;
|
|
/* 11: mullw */
|
|
var49.i = (var48.i * var41.i) & 0xffff;
|
|
/* 12: addw */
|
|
var50.i = var46.i + var49.i;
|
|
/* 14: addw */
|
|
var51.i = var50.i + var42.i;
|
|
/* 15: shrsw */
|
|
var52.i = var51.i >> 8;
|
|
/* 16: addw */
|
|
var53.i = var52.i + var45.i;
|
|
/* 17: addw */
|
|
var54.i = var53.i + var48.i;
|
|
/* 18: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var54.i);
|
|
/* 19: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_matrix2_11_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, int p1, int p2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_matrix2_11_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_matrix2_11_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_constant (p, 4, 0x00000010, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x00000008, "c3");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T3,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_matrix2_12_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_matrix2_12_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, int p1, int p2, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union16 var55;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
|
|
/* 2: loadpw */
|
|
var37.i = (int) 0x00000010; /* 16 or 7.90505e-323f */
|
|
/* 4: loadpw */
|
|
var38.i = p1;
|
|
/* 8: loadpw */
|
|
var40.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 10: loadpw */
|
|
var41.i = p2;
|
|
/* 13: loadpw */
|
|
var42.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 3: subw */
|
|
var45.i = var44.i - var37.i;
|
|
/* 5: mullw */
|
|
var46.i = (var45.i * var38.i) & 0xffff;
|
|
/* 6: loadb */
|
|
var39 = ptr5[i];
|
|
/* 7: convubw */
|
|
var47.i = (orc_uint8) var39;
|
|
/* 9: subw */
|
|
var48.i = var47.i - var40.i;
|
|
/* 11: mullw */
|
|
var49.i = (var48.i * var41.i) & 0xffff;
|
|
/* 12: addw */
|
|
var50.i = var46.i + var49.i;
|
|
/* 14: addw */
|
|
var51.i = var50.i + var42.i;
|
|
/* 15: shrsw */
|
|
var52.i = var51.i >> 8;
|
|
/* 16: addw */
|
|
var53.i = var52.i + var45.i;
|
|
/* 17: addw */
|
|
var54.i = var53.i + var48.i;
|
|
/* 18: addw */
|
|
var55.i = var54.i + var48.i;
|
|
/* 19: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var55.i);
|
|
/* 20: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_matrix2_12_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union16 var55;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
|
|
/* 2: loadpw */
|
|
var37.i = (int) 0x00000010; /* 16 or 7.90505e-323f */
|
|
/* 4: loadpw */
|
|
var38.i = ex->params[24];
|
|
/* 8: loadpw */
|
|
var40.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 10: loadpw */
|
|
var41.i = ex->params[25];
|
|
/* 13: loadpw */
|
|
var42.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 3: subw */
|
|
var45.i = var44.i - var37.i;
|
|
/* 5: mullw */
|
|
var46.i = (var45.i * var38.i) & 0xffff;
|
|
/* 6: loadb */
|
|
var39 = ptr5[i];
|
|
/* 7: convubw */
|
|
var47.i = (orc_uint8) var39;
|
|
/* 9: subw */
|
|
var48.i = var47.i - var40.i;
|
|
/* 11: mullw */
|
|
var49.i = (var48.i * var41.i) & 0xffff;
|
|
/* 12: addw */
|
|
var50.i = var46.i + var49.i;
|
|
/* 14: addw */
|
|
var51.i = var50.i + var42.i;
|
|
/* 15: shrsw */
|
|
var52.i = var51.i >> 8;
|
|
/* 16: addw */
|
|
var53.i = var52.i + var45.i;
|
|
/* 17: addw */
|
|
var54.i = var53.i + var48.i;
|
|
/* 18: addw */
|
|
var55.i = var54.i + var48.i;
|
|
/* 19: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var55.i);
|
|
/* 20: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_matrix2_12_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, int p1, int p2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_matrix2_12_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_matrix2_12_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_constant (p, 4, 0x00000010, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x00000008, "c3");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T3,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_matrix3_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_matrix3_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1,
|
|
int p2, int p3, int p4, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 2: loadpw */
|
|
var35.i = p1;
|
|
/* 6: loadpw */
|
|
var37.i = p2;
|
|
/* 11: loadpw */
|
|
var39.i = p3;
|
|
/* 14: loadpw */
|
|
var40.i = p4;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var42.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var43.i = (var42.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var45.i = (var44.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var46.i = var43.i + var45.i;
|
|
/* 9: loadb */
|
|
var38 = ptr6[i];
|
|
/* 10: convubw */
|
|
var47.i = (orc_uint8) var38;
|
|
/* 12: mullw */
|
|
var48.i = (var47.i * var39.i) & 0xffff;
|
|
/* 13: addw */
|
|
var49.i = var46.i + var48.i;
|
|
/* 15: addw */
|
|
var50.i = var49.i + var40.i;
|
|
/* 16: shrsw */
|
|
var51.i = var50.i >> 6;
|
|
/* 17: convsuswb */
|
|
var41 = ORC_CLAMP_UB (var51.i);
|
|
/* 18: storeb */
|
|
ptr0[i] = var41;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_matrix3_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 2: loadpw */
|
|
var35.i = ex->params[24];
|
|
/* 6: loadpw */
|
|
var37.i = ex->params[25];
|
|
/* 11: loadpw */
|
|
var39.i = ex->params[26];
|
|
/* 14: loadpw */
|
|
var40.i = ex->params[27];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var42.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var43.i = (var42.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var45.i = (var44.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var46.i = var43.i + var45.i;
|
|
/* 9: loadb */
|
|
var38 = ptr6[i];
|
|
/* 10: convubw */
|
|
var47.i = (orc_uint8) var38;
|
|
/* 12: mullw */
|
|
var48.i = (var47.i * var39.i) & 0xffff;
|
|
/* 13: addw */
|
|
var49.i = var46.i + var48.i;
|
|
/* 15: addw */
|
|
var50.i = var49.i + var40.i;
|
|
/* 16: shrsw */
|
|
var51.i = var50.i >> 6;
|
|
/* 17: convsuswb */
|
|
var41 = ORC_CLAMP_UB (var51.i);
|
|
/* 18: storeb */
|
|
ptr0[i] = var41;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_matrix3_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1,
|
|
int p2, int p3, int p4, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_matrix3_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_matrix3_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_constant (p, 4, 0x00000006, "c1");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_parameter (p, 2, "p3");
|
|
orc_program_add_parameter (p, 2, "p4");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
ex->params[ORC_VAR_P3] = p3;
|
|
ex->params[ORC_VAR_P4] = p4;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_matrix3_100_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_matrix3_100_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1,
|
|
int p2, int p3, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_int8 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union16 var55;
|
|
orc_union16 var56;
|
|
orc_union16 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 2: loadpw */
|
|
var36.i = (int) 0x00000010; /* 16 or 7.90505e-323f */
|
|
/* 4: loadpw */
|
|
var37.i = p1;
|
|
/* 8: loadpw */
|
|
var39.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 10: loadpw */
|
|
var40.i = p2;
|
|
/* 15: loadpw */
|
|
var42.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 17: loadpw */
|
|
var43.i = p3;
|
|
/* 20: loadpw */
|
|
var44.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var35 = ptr4[i];
|
|
/* 1: convubw */
|
|
var46.i = (orc_uint8) var35;
|
|
/* 3: subw */
|
|
var47.i = var46.i - var36.i;
|
|
/* 5: mullw */
|
|
var48.i = (var47.i * var37.i) & 0xffff;
|
|
/* 6: loadb */
|
|
var38 = ptr5[i];
|
|
/* 7: convubw */
|
|
var49.i = (orc_uint8) var38;
|
|
/* 9: subw */
|
|
var50.i = var49.i - var39.i;
|
|
/* 11: mullw */
|
|
var51.i = (var50.i * var40.i) & 0xffff;
|
|
/* 12: addw */
|
|
var52.i = var48.i + var51.i;
|
|
/* 13: loadb */
|
|
var41 = ptr6[i];
|
|
/* 14: convubw */
|
|
var53.i = (orc_uint8) var41;
|
|
/* 16: subw */
|
|
var54.i = var53.i - var42.i;
|
|
/* 18: mullw */
|
|
var55.i = (var54.i * var43.i) & 0xffff;
|
|
/* 19: addw */
|
|
var56.i = var52.i + var55.i;
|
|
/* 21: addw */
|
|
var57.i = var56.i + var44.i;
|
|
/* 22: shrsw */
|
|
var58.i = var57.i >> 8;
|
|
/* 23: addw */
|
|
var59.i = var58.i + var47.i;
|
|
/* 24: convsuswb */
|
|
var45 = ORC_CLAMP_UB (var59.i);
|
|
/* 25: storeb */
|
|
ptr0[i] = var45;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_matrix3_100_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_int8 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union16 var55;
|
|
orc_union16 var56;
|
|
orc_union16 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 2: loadpw */
|
|
var36.i = (int) 0x00000010; /* 16 or 7.90505e-323f */
|
|
/* 4: loadpw */
|
|
var37.i = ex->params[24];
|
|
/* 8: loadpw */
|
|
var39.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 10: loadpw */
|
|
var40.i = ex->params[25];
|
|
/* 15: loadpw */
|
|
var42.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 17: loadpw */
|
|
var43.i = ex->params[26];
|
|
/* 20: loadpw */
|
|
var44.i = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var35 = ptr4[i];
|
|
/* 1: convubw */
|
|
var46.i = (orc_uint8) var35;
|
|
/* 3: subw */
|
|
var47.i = var46.i - var36.i;
|
|
/* 5: mullw */
|
|
var48.i = (var47.i * var37.i) & 0xffff;
|
|
/* 6: loadb */
|
|
var38 = ptr5[i];
|
|
/* 7: convubw */
|
|
var49.i = (orc_uint8) var38;
|
|
/* 9: subw */
|
|
var50.i = var49.i - var39.i;
|
|
/* 11: mullw */
|
|
var51.i = (var50.i * var40.i) & 0xffff;
|
|
/* 12: addw */
|
|
var52.i = var48.i + var51.i;
|
|
/* 13: loadb */
|
|
var41 = ptr6[i];
|
|
/* 14: convubw */
|
|
var53.i = (orc_uint8) var41;
|
|
/* 16: subw */
|
|
var54.i = var53.i - var42.i;
|
|
/* 18: mullw */
|
|
var55.i = (var54.i * var43.i) & 0xffff;
|
|
/* 19: addw */
|
|
var56.i = var52.i + var55.i;
|
|
/* 21: addw */
|
|
var57.i = var56.i + var44.i;
|
|
/* 22: shrsw */
|
|
var58.i = var57.i >> 8;
|
|
/* 23: addw */
|
|
var59.i = var58.i + var47.i;
|
|
/* 24: convsuswb */
|
|
var45 = ORC_CLAMP_UB (var59.i);
|
|
/* 25: storeb */
|
|
ptr0[i] = var45;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_matrix3_100_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1,
|
|
int p2, int p3, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_matrix3_100_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_matrix3_100_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_constant (p, 4, 0x00000010, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x00000008, "c3");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_parameter (p, 2, "p3");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T3,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
ex->params[ORC_VAR_P3] = p3;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_matrix3_100_offset_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_matrix3_100_offset_u8 (guint8 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
|
|
const guint8 * ORC_RESTRICT s3, int p1, int p2, int p3, int p4, int p5,
|
|
int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 2: loadpw */
|
|
var36.i = p1;
|
|
/* 6: loadpw */
|
|
var38.i = p2;
|
|
/* 11: loadpw */
|
|
var40.i = p3;
|
|
/* 14: loadpw */
|
|
var41.i = p4;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var35 = ptr4[i];
|
|
/* 1: convubw */
|
|
var43.i = (orc_uint8) var35;
|
|
/* 3: mullw */
|
|
var44.i = (var43.i * var36.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var37 = ptr5[i];
|
|
/* 5: convubw */
|
|
var45.i = (orc_uint8) var37;
|
|
/* 7: mullw */
|
|
var46.i = (var45.i * var38.i) & 0xffff;
|
|
/* 8: addw */
|
|
var47.i = var44.i + var46.i;
|
|
/* 9: loadb */
|
|
var39 = ptr6[i];
|
|
/* 10: convubw */
|
|
var48.i = (orc_uint8) var39;
|
|
/* 12: mullw */
|
|
var49.i = (var48.i * var40.i) & 0xffff;
|
|
/* 13: addw */
|
|
var50.i = var47.i + var49.i;
|
|
/* 15: addw */
|
|
var51.i = var50.i + var41.i;
|
|
/* 16: shrsw */
|
|
var52.i = var51.i >> p5;
|
|
/* 17: addw */
|
|
var53.i = var52.i + var43.i;
|
|
/* 18: convsuswb */
|
|
var42 = ORC_CLAMP_UB (var53.i);
|
|
/* 19: storeb */
|
|
ptr0[i] = var42;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_matrix3_100_offset_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var35;
|
|
orc_union16 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_int8 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 2: loadpw */
|
|
var36.i = ex->params[24];
|
|
/* 6: loadpw */
|
|
var38.i = ex->params[25];
|
|
/* 11: loadpw */
|
|
var40.i = ex->params[26];
|
|
/* 14: loadpw */
|
|
var41.i = ex->params[27];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var35 = ptr4[i];
|
|
/* 1: convubw */
|
|
var43.i = (orc_uint8) var35;
|
|
/* 3: mullw */
|
|
var44.i = (var43.i * var36.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var37 = ptr5[i];
|
|
/* 5: convubw */
|
|
var45.i = (orc_uint8) var37;
|
|
/* 7: mullw */
|
|
var46.i = (var45.i * var38.i) & 0xffff;
|
|
/* 8: addw */
|
|
var47.i = var44.i + var46.i;
|
|
/* 9: loadb */
|
|
var39 = ptr6[i];
|
|
/* 10: convubw */
|
|
var48.i = (orc_uint8) var39;
|
|
/* 12: mullw */
|
|
var49.i = (var48.i * var40.i) & 0xffff;
|
|
/* 13: addw */
|
|
var50.i = var47.i + var49.i;
|
|
/* 15: addw */
|
|
var51.i = var50.i + var41.i;
|
|
/* 16: shrsw */
|
|
var52.i = var51.i >> ex->params[28];
|
|
/* 17: addw */
|
|
var53.i = var52.i + var43.i;
|
|
/* 18: convsuswb */
|
|
var42 = ORC_CLAMP_UB (var53.i);
|
|
/* 19: storeb */
|
|
ptr0[i] = var42;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_matrix3_100_offset_u8 (guint8 * ORC_RESTRICT d1,
|
|
const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
|
|
const guint8 * ORC_RESTRICT s3, int p1, int p2, int p3, int p4, int p5,
|
|
int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_matrix3_100_offset_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_matrix3_100_offset_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_parameter (p, 2, "p3");
|
|
orc_program_add_parameter (p, 2, "p4");
|
|
orc_program_add_parameter (p, 2, "p5");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T3, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
ex->params[ORC_VAR_P3] = p3;
|
|
ex->params[ORC_VAR_P4] = p4;
|
|
ex->params[ORC_VAR_P5] = p5;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_matrix3_000_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_matrix3_000_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1,
|
|
int p2, int p3, int p4, int p5, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 2: loadpw */
|
|
var35.i = p1;
|
|
/* 6: loadpw */
|
|
var37.i = p2;
|
|
/* 11: loadpw */
|
|
var39.i = p3;
|
|
/* 14: loadpw */
|
|
var40.i = p4;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var42.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var43.i = (var42.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var45.i = (var44.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var46.i = var43.i + var45.i;
|
|
/* 9: loadb */
|
|
var38 = ptr6[i];
|
|
/* 10: convubw */
|
|
var47.i = (orc_uint8) var38;
|
|
/* 12: mullw */
|
|
var48.i = (var47.i * var39.i) & 0xffff;
|
|
/* 13: addw */
|
|
var49.i = var46.i + var48.i;
|
|
/* 15: addw */
|
|
var50.i = var49.i + var40.i;
|
|
/* 16: shrsw */
|
|
var51.i = var50.i >> p5;
|
|
/* 17: convwb */
|
|
var41 = var51.i;
|
|
/* 18: storeb */
|
|
ptr0[i] = var41;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_matrix3_000_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_int8 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 2: loadpw */
|
|
var35.i = ex->params[24];
|
|
/* 6: loadpw */
|
|
var37.i = ex->params[25];
|
|
/* 11: loadpw */
|
|
var39.i = ex->params[26];
|
|
/* 14: loadpw */
|
|
var40.i = ex->params[27];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var42.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var43.i = (var42.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var44.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var45.i = (var44.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var46.i = var43.i + var45.i;
|
|
/* 9: loadb */
|
|
var38 = ptr6[i];
|
|
/* 10: convubw */
|
|
var47.i = (orc_uint8) var38;
|
|
/* 12: mullw */
|
|
var48.i = (var47.i * var39.i) & 0xffff;
|
|
/* 13: addw */
|
|
var49.i = var46.i + var48.i;
|
|
/* 15: addw */
|
|
var50.i = var49.i + var40.i;
|
|
/* 16: shrsw */
|
|
var51.i = var50.i >> ex->params[28];
|
|
/* 17: convwb */
|
|
var41 = var51.i;
|
|
/* 18: storeb */
|
|
ptr0[i] = var41;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_matrix3_000_u8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1,
|
|
const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, int p1,
|
|
int p2, int p3, int p4, int p5, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_matrix3_000_u8");
|
|
orc_program_set_backup_function (p, _backup_orc_matrix3_000_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_parameter (p, 2, "p3");
|
|
orc_program_add_parameter (p, 2, "p4");
|
|
orc_program_add_parameter (p, 2, "p5");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convwb", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
ex->params[ORC_VAR_P3] = p3;
|
|
ex->params[ORC_VAR_P4] = p4;
|
|
ex->params[ORC_VAR_P5] = p5;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_pack_123x */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_pack_123x (guint32 * ORC_RESTRICT d1, const orc_uint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
int p1, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 4: loadpb */
|
|
var37 = p1;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: loadb */
|
|
var35 = ptr5[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var34;
|
|
_dest.x2[1] = var35;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadb */
|
|
var36 = ptr6[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var39.i;
|
|
_dest.x2[1] = var40.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_pack_123x (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 4: loadpb */
|
|
var37 = ex->params[24];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: loadb */
|
|
var35 = ptr5[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var34;
|
|
_dest.x2[1] = var35;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadb */
|
|
var36 = ptr6[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var39.i;
|
|
_dest.x2[1] = var40.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_pack_123x (guint32 * ORC_RESTRICT d1, const orc_uint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
int p1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_pack_123x");
|
|
orc_program_set_backup_function (p, _backup_orc_pack_123x);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_parameter (p, 1, "p1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* orc_pack_x123 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
orc_pack_x123 (guint32 * ORC_RESTRICT d1, const orc_uint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
int p1, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 0: loadpb */
|
|
var34 = p1;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 1: loadb */
|
|
var35 = ptr4[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var34;
|
|
_dest.x2[1] = var35;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadb */
|
|
var36 = ptr5[i];
|
|
/* 4: loadb */
|
|
var37 = ptr6[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var39.i;
|
|
_dest.x2[1] = var40.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_orc_pack_x123 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 0: loadpb */
|
|
var34 = ex->params[24];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 1: loadb */
|
|
var35 = ptr4[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var34;
|
|
_dest.x2[1] = var35;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadb */
|
|
var36 = ptr5[i];
|
|
/* 4: loadb */
|
|
var37 = ptr6[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var39.i;
|
|
_dest.x2[1] = var40.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
orc_pack_x123 (guint32 * ORC_RESTRICT d1, const orc_uint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
int p1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "orc_pack_x123");
|
|
orc_program_set_backup_function (p, _backup_orc_pack_x123);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_parameter (p, 1, "p1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_P1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_combine2_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_combine2_u8 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
int p1, int p2, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
|
|
/* 2: loadpw */
|
|
var35.i = p1;
|
|
/* 6: loadpw */
|
|
var37.i = p2;
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var39.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var40.i = (var39.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var41.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var42.i = (var41.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var43.i = var40.i + var42.i;
|
|
/* 9: shruw */
|
|
var44.i = ((orc_uint16) var43.i) >> 8;
|
|
/* 10: convsuswb */
|
|
var38 = ORC_CLAMP_UB (var44.i);
|
|
/* 11: storeb */
|
|
ptr0[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_combine2_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
|
|
/* 2: loadpw */
|
|
var35.i = ex->params[24];
|
|
/* 6: loadpw */
|
|
var37.i = ex->params[25];
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var39.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var40.i = (var39.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var41.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var42.i = (var41.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var43.i = var40.i + var42.i;
|
|
/* 9: shruw */
|
|
var44.i = ((orc_uint16) var43.i) >> 8;
|
|
/* 10: convsuswb */
|
|
var38 = ORC_CLAMP_UB (var44.i);
|
|
/* 11: storeb */
|
|
ptr0[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_combine2_u8 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
int p1, int p2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_combine2_u8");
|
|
orc_program_set_backup_function (p, _backup_cogorc_combine2_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_constant (p, 4, 0x00000008, "c1");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shruw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_combine4_u8 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_combine4_u8 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4,
|
|
int p1, int p2, int p3, int p4, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union16 var55;
|
|
orc_union16 var56;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
ptr7 = (orc_int8 *) s4;
|
|
|
|
/* 2: loadpw */
|
|
var35.i = p1;
|
|
/* 6: loadpw */
|
|
var37.i = p2;
|
|
/* 11: loadpw */
|
|
var39.i = p3;
|
|
/* 16: loadpw */
|
|
var41.i = p4;
|
|
/* 19: loadpw */
|
|
var42.i = (int) 0x00000020; /* 32 or 1.58101e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var45.i = (var44.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var46.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var47.i = (var46.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var48.i = var45.i + var47.i;
|
|
/* 9: loadb */
|
|
var38 = ptr6[i];
|
|
/* 10: convubw */
|
|
var49.i = (orc_uint8) var38;
|
|
/* 12: mullw */
|
|
var50.i = (var49.i * var39.i) & 0xffff;
|
|
/* 13: addw */
|
|
var51.i = var48.i + var50.i;
|
|
/* 14: loadb */
|
|
var40 = ptr7[i];
|
|
/* 15: convubw */
|
|
var52.i = (orc_uint8) var40;
|
|
/* 17: mullw */
|
|
var53.i = (var52.i * var41.i) & 0xffff;
|
|
/* 18: addw */
|
|
var54.i = var51.i + var53.i;
|
|
/* 20: addw */
|
|
var55.i = var54.i + var42.i;
|
|
/* 21: shrsw */
|
|
var56.i = var55.i >> 6;
|
|
/* 22: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var56.i);
|
|
/* 23: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_combine4_u8 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
orc_union16 var47;
|
|
orc_union16 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union16 var55;
|
|
orc_union16 var56;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
ptr7 = (orc_int8 *) ex->arrays[7];
|
|
|
|
/* 2: loadpw */
|
|
var35.i = ex->params[24];
|
|
/* 6: loadpw */
|
|
var37.i = ex->params[25];
|
|
/* 11: loadpw */
|
|
var39.i = ex->params[26];
|
|
/* 16: loadpw */
|
|
var41.i = ex->params[27];
|
|
/* 19: loadpw */
|
|
var42.i = (int) 0x00000020; /* 32 or 1.58101e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr4[i];
|
|
/* 1: convubw */
|
|
var44.i = (orc_uint8) var34;
|
|
/* 3: mullw */
|
|
var45.i = (var44.i * var35.i) & 0xffff;
|
|
/* 4: loadb */
|
|
var36 = ptr5[i];
|
|
/* 5: convubw */
|
|
var46.i = (orc_uint8) var36;
|
|
/* 7: mullw */
|
|
var47.i = (var46.i * var37.i) & 0xffff;
|
|
/* 8: addw */
|
|
var48.i = var45.i + var47.i;
|
|
/* 9: loadb */
|
|
var38 = ptr6[i];
|
|
/* 10: convubw */
|
|
var49.i = (orc_uint8) var38;
|
|
/* 12: mullw */
|
|
var50.i = (var49.i * var39.i) & 0xffff;
|
|
/* 13: addw */
|
|
var51.i = var48.i + var50.i;
|
|
/* 14: loadb */
|
|
var40 = ptr7[i];
|
|
/* 15: convubw */
|
|
var52.i = (orc_uint8) var40;
|
|
/* 17: mullw */
|
|
var53.i = (var52.i * var41.i) & 0xffff;
|
|
/* 18: addw */
|
|
var54.i = var51.i + var53.i;
|
|
/* 20: addw */
|
|
var55.i = var54.i + var42.i;
|
|
/* 21: shrsw */
|
|
var56.i = var55.i >> 6;
|
|
/* 22: convsuswb */
|
|
var43 = ORC_CLAMP_UB (var56.i);
|
|
/* 23: storeb */
|
|
ptr0[i] = var43;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_combine4_u8 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4,
|
|
int p1, int p2, int p3, int p4, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_combine4_u8");
|
|
orc_program_set_backup_function (p, _backup_cogorc_combine4_u8);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_source (p, 1, "s4");
|
|
orc_program_add_constant (p, 4, 0x00000020, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000006, "c2");
|
|
orc_program_add_parameter (p, 2, "p1");
|
|
orc_program_add_parameter (p, 2, "p2");
|
|
orc_program_add_parameter (p, 2, "p3");
|
|
orc_program_add_parameter (p, 2, "p4");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S4, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->arrays[ORC_VAR_S4] = (void *) s4;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
ex->params[ORC_VAR_P3] = p3;
|
|
ex->params[ORC_VAR_P4] = p4;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_unpack_axyz_0 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_unpack_axyz_0 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_unpack_axyz_0 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_unpack_axyz_0 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_unpack_axyz_0");
|
|
orc_program_set_backup_function (p, _backup_cogorc_unpack_axyz_0);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_unpack_axyz_1 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_unpack_axyz_1 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_unpack_axyz_1 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select0lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[0];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_unpack_axyz_1 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_unpack_axyz_1");
|
|
orc_program_set_backup_function (p, _backup_cogorc_unpack_axyz_1);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select0lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_unpack_axyz_2 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_unpack_axyz_2 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_unpack_axyz_2 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select0wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_unpack_axyz_2 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_unpack_axyz_2");
|
|
orc_program_set_backup_function (p, _backup_cogorc_unpack_axyz_2);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select0wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_unpack_axyz_3 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_unpack_axyz_3 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_union32 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_unpack_axyz_3 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: select1lw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var33.i;
|
|
var35.i = _src.x2[1];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.i;
|
|
var34 = _src.x2[1];
|
|
}
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_unpack_axyz_3 (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint32 * ORC_RESTRICT s1, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_unpack_axyz_3");
|
|
orc_program_set_backup_function (p, _backup_cogorc_unpack_axyz_3);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "select1lw", 0, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_resample_horiz_1tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_resample_horiz_1tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int p1, int p2, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: ldresnearb */
|
|
var32 = ptr4[(p1 + i * p2) >> 16];
|
|
/* 1: storeb */
|
|
ptr0[i] = var32;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_resample_horiz_1tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: ldresnearb */
|
|
var32 = ptr4[(ex->params[24] + i * ex->params[25]) >> 16];
|
|
/* 1: storeb */
|
|
ptr0[i] = var32;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_resample_horiz_1tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int p1, int p2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_resample_horiz_1tap");
|
|
orc_program_set_backup_function (p, _backup_cogorc_resample_horiz_1tap);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_parameter (p, 4, "p1");
|
|
orc_program_add_parameter (p, 4, "p2");
|
|
|
|
orc_program_append_2 (p, "ldresnearb", 0, ORC_VAR_D1, ORC_VAR_S1,
|
|
ORC_VAR_P1, ORC_VAR_P2);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_resample_horiz_2tap */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_resample_horiz_2tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int p1, int p2, int n)
|
|
{
|
|
int i;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
|
|
ptr0 = (orc_int8 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: ldreslinb */
|
|
{
|
|
int tmp = p1 + i * p2;
|
|
var32 =
|
|
((orc_uint8) ptr4[tmp >> 16] * (256 - ((tmp >> 8) & 0xff)) +
|
|
(orc_uint8) ptr4[(tmp >> 16) + 1] * ((tmp >> 8) & 0xff)) >> 8;
|
|
}
|
|
/* 1: storeb */
|
|
ptr0[i] = var32;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_resample_horiz_2tap (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
|
|
ptr0 = (orc_int8 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: ldreslinb */
|
|
{
|
|
int tmp = ex->params[24] + i * ex->params[25];
|
|
var32 =
|
|
((orc_uint8) ptr4[tmp >> 16] * (256 - ((tmp >> 8) & 0xff)) +
|
|
(orc_uint8) ptr4[(tmp >> 16) + 1] * ((tmp >> 8) & 0xff)) >> 8;
|
|
}
|
|
/* 1: storeb */
|
|
ptr0[i] = var32;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_resample_horiz_2tap (orc_uint8 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, int p1, int p2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_resample_horiz_2tap");
|
|
orc_program_set_backup_function (p, _backup_cogorc_resample_horiz_2tap);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_parameter (p, 4, "p1");
|
|
orc_program_add_parameter (p, 4, "p2");
|
|
|
|
orc_program_append_2 (p, "ldreslinb", 0, ORC_VAR_D1, ORC_VAR_S1,
|
|
ORC_VAR_P1, ORC_VAR_P2);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_P1] = p1;
|
|
ex->params[ORC_VAR_P2] = p2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_I420_UYVY */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_I420_UYVY (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint16 * ORC_RESTRICT s1,
|
|
const orc_uint16 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
orc_union32 *ORC_RESTRICT ptr1;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr1 = (orc_union32 *) d2;
|
|
ptr4 = (orc_union16 *) s1;
|
|
ptr5 = (orc_union16 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
ptr7 = (orc_int8 *) s4;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr6[i];
|
|
/* 1: loadb */
|
|
var34 = ptr7[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[0];
|
|
_dest.x2[1] = var35.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[1];
|
|
_dest.x2[1] = var35.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
/* 6: loadw */
|
|
var37 = ptr5[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 8: storel */
|
|
ptr1[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_I420_UYVY (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
orc_union32 *ORC_RESTRICT ptr1;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr1 = (orc_union32 *) ex->arrays[1];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
ptr5 = (orc_union16 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
ptr7 = (orc_int8 *) ex->arrays[7];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr6[i];
|
|
/* 1: loadb */
|
|
var34 = ptr7[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[0];
|
|
_dest.x2[1] = var35.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[1];
|
|
_dest.x2[1] = var35.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
/* 6: loadw */
|
|
var37 = ptr5[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 8: storel */
|
|
ptr1[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_I420_UYVY (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint16 * ORC_RESTRICT s1,
|
|
const orc_uint16 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_convert_I420_UYVY");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_I420_UYVY);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_destination (p, 4, "d2");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 2, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_source (p, 1, "s4");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S3, ORC_VAR_S4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D2, ORC_VAR_T1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->arrays[ORC_VAR_S4] = (void *) s4;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_I420_YUY2 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_I420_YUY2 (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint16 * ORC_RESTRICT s1,
|
|
const orc_uint16 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
orc_union32 *ORC_RESTRICT ptr1;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr1 = (orc_union32 *) d2;
|
|
ptr4 = (orc_union16 *) s1;
|
|
ptr5 = (orc_union16 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
ptr7 = (orc_int8 *) s4;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr6[i];
|
|
/* 1: loadb */
|
|
var34 = ptr7[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
/* 6: loadw */
|
|
var37 = ptr5[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 8: storel */
|
|
ptr1[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_I420_YUY2 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
orc_union32 *ORC_RESTRICT ptr1;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr1 = (orc_union32 *) ex->arrays[1];
|
|
ptr4 = (orc_union16 *) ex->arrays[4];
|
|
ptr5 = (orc_union16 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
ptr7 = (orc_int8 *) ex->arrays[7];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr6[i];
|
|
/* 1: loadb */
|
|
var34 = ptr7[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
/* 6: loadw */
|
|
var37 = ptr5[i];
|
|
/* 7: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 8: storel */
|
|
ptr1[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_I420_YUY2 (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint16 * ORC_RESTRICT s1,
|
|
const orc_uint16 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_convert_I420_YUY2");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_I420_YUY2);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_destination (p, 4, "d2");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 2, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_source (p, 1, "s4");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S3, ORC_VAR_S4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D2, ORC_VAR_S2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->arrays[ORC_VAR_S4] = (void *) s4;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_I420_AYUV */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_I420_AYUV (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
orc_union32 *ORC_RESTRICT ptr1;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_int8 var39;
|
|
orc_int8 var40;
|
|
orc_union32 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr1 = (orc_union32 *) d2;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
ptr7 = (orc_int8 *) s4;
|
|
|
|
/* 3: loadpb */
|
|
var36 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
/* 8: loadpb */
|
|
var39 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadupdb */
|
|
var42 = ptr6[i >> 1];
|
|
/* 1: loadupdb */
|
|
var43 = ptr7[i >> 1];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var42;
|
|
_dest.x2[1] = var43;
|
|
var44.i = _dest.i;
|
|
}
|
|
/* 4: loadb */
|
|
var37 = ptr4[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var45.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var45.i;
|
|
_dest.x2[1] = var44.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
/* 9: loadb */
|
|
var40 = ptr5[i];
|
|
/* 10: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39;
|
|
_dest.x2[1] = var40;
|
|
var46.i = _dest.i;
|
|
}
|
|
/* 11: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var46.i;
|
|
_dest.x2[1] = var44.i;
|
|
var41.i = _dest.i;
|
|
}
|
|
/* 12: storel */
|
|
ptr1[i] = var41;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_I420_AYUV (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
orc_union32 *ORC_RESTRICT ptr1;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_int8 var39;
|
|
orc_int8 var40;
|
|
orc_union32 var41;
|
|
orc_int8 var42;
|
|
orc_int8 var43;
|
|
orc_union16 var44;
|
|
orc_union16 var45;
|
|
orc_union16 var46;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr1 = (orc_union32 *) ex->arrays[1];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
ptr7 = (orc_int8 *) ex->arrays[7];
|
|
|
|
/* 3: loadpb */
|
|
var36 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
/* 8: loadpb */
|
|
var39 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadupdb */
|
|
var42 = ptr6[i >> 1];
|
|
/* 1: loadupdb */
|
|
var43 = ptr7[i >> 1];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var42;
|
|
_dest.x2[1] = var43;
|
|
var44.i = _dest.i;
|
|
}
|
|
/* 4: loadb */
|
|
var37 = ptr4[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var45.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var45.i;
|
|
_dest.x2[1] = var44.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
/* 9: loadb */
|
|
var40 = ptr5[i];
|
|
/* 10: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var39;
|
|
_dest.x2[1] = var40;
|
|
var46.i = _dest.i;
|
|
}
|
|
/* 11: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var46.i;
|
|
_dest.x2[1] = var44.i;
|
|
var41.i = _dest.i;
|
|
}
|
|
/* 12: storel */
|
|
ptr1[i] = var41;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_I420_AYUV (orc_uint32 * ORC_RESTRICT d1,
|
|
orc_uint32 * ORC_RESTRICT d2, const orc_uint8 * ORC_RESTRICT s1,
|
|
const orc_uint8 * ORC_RESTRICT s2, const orc_uint8 * ORC_RESTRICT s3,
|
|
const orc_uint8 * ORC_RESTRICT s4, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_convert_I420_AYUV");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_I420_AYUV);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_destination (p, 4, "d2");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_source (p, 1, "s4");
|
|
orc_program_add_constant (p, 1, 0x000000ff, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 1, "t4");
|
|
|
|
orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T3, ORC_VAR_S3,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T4, ORC_VAR_S4,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D2, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->arrays[ORC_VAR_S4] = (void *) s4;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_YUY2_I420 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_YUY2_I420 (orc_uint16 * ORC_RESTRICT d1,
|
|
orc_uint16 * ORC_RESTRICT d2, orc_uint8 * ORC_RESTRICT d3,
|
|
orc_uint8 * ORC_RESTRICT d4, const orc_uint32 * ORC_RESTRICT s1,
|
|
const orc_uint32 * ORC_RESTRICT s2, int n)
|
|
{
|
|
int i;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
orc_int8 *ORC_RESTRICT ptr3;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
const orc_union32 *ORC_RESTRICT ptr5;
|
|
orc_union32 var35;
|
|
orc_union32 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
ptr0 = (orc_union16 *) d1;
|
|
ptr1 = (orc_union16 *) d2;
|
|
ptr2 = (orc_int8 *) d3;
|
|
ptr3 = (orc_int8 *) d4;
|
|
ptr4 = (orc_union32 *) s1;
|
|
ptr5 = (orc_union32 *) s2;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var40;
|
|
/* 3: loadl */
|
|
var36 = ptr5[i];
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var41.x2[0] = _src.x2[1];
|
|
var42.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var41.x2[1] = _src.x2[1];
|
|
var42.x2[1] = _src.x2[0];
|
|
}
|
|
/* 5: storew */
|
|
ptr1[i] = var42;
|
|
/* 6: avgub */
|
|
var43.x2[0] = ((orc_uint8) var39.x2[0] + (orc_uint8) var41.x2[0] + 1) >> 1;
|
|
var43.x2[1] = ((orc_uint8) var39.x2[1] + (orc_uint8) var41.x2[1] + 1) >> 1;
|
|
/* 7: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var37 = _src.x2[1];
|
|
var38 = _src.x2[0];
|
|
}
|
|
/* 8: storeb */
|
|
ptr3[i] = var37;
|
|
/* 9: storeb */
|
|
ptr2[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_YUY2_I420 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
orc_int8 *ORC_RESTRICT ptr3;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
const orc_union32 *ORC_RESTRICT ptr5;
|
|
orc_union32 var35;
|
|
orc_union32 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
ptr0 = (orc_union16 *) ex->arrays[0];
|
|
ptr1 = (orc_union16 *) ex->arrays[1];
|
|
ptr2 = (orc_int8 *) ex->arrays[2];
|
|
ptr3 = (orc_int8 *) ex->arrays[3];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
ptr5 = (orc_union32 *) ex->arrays[5];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var40;
|
|
/* 3: loadl */
|
|
var36 = ptr5[i];
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var41.x2[0] = _src.x2[1];
|
|
var42.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var41.x2[1] = _src.x2[1];
|
|
var42.x2[1] = _src.x2[0];
|
|
}
|
|
/* 5: storew */
|
|
ptr1[i] = var42;
|
|
/* 6: avgub */
|
|
var43.x2[0] = ((orc_uint8) var39.x2[0] + (orc_uint8) var41.x2[0] + 1) >> 1;
|
|
var43.x2[1] = ((orc_uint8) var39.x2[1] + (orc_uint8) var41.x2[1] + 1) >> 1;
|
|
/* 7: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var37 = _src.x2[1];
|
|
var38 = _src.x2[0];
|
|
}
|
|
/* 8: storeb */
|
|
ptr3[i] = var37;
|
|
/* 9: storeb */
|
|
ptr2[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_YUY2_I420 (orc_uint16 * ORC_RESTRICT d1,
|
|
orc_uint16 * ORC_RESTRICT d2, orc_uint8 * ORC_RESTRICT d3,
|
|
orc_uint8 * ORC_RESTRICT d4, const orc_uint32 * ORC_RESTRICT s1,
|
|
const orc_uint32 * ORC_RESTRICT s2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_convert_YUY2_I420");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_I420);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 2, "d2");
|
|
orc_program_add_destination (p, 1, "d3");
|
|
orc_program_add_destination (p, 1, "d4");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_source (p, 4, "s2");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "storew", 0, ORC_VAR_D2, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D4, ORC_VAR_D3, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->arrays[ORC_VAR_D4] = d4;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_UYVY_YUY2 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_UYVY_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var32;
|
|
orc_union32 var33;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var32 = ptr4[i];
|
|
/* 1: swapw */
|
|
var33.x2[0] = ORC_SWAP_W (var32.x2[0]);
|
|
var33.x2[1] = ORC_SWAP_W (var32.x2[1]);
|
|
/* 2: storel */
|
|
ptr0[i] = var33;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_UYVY_YUY2 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var32;
|
|
orc_union32 var33;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var32 = ptr4[i];
|
|
/* 1: swapw */
|
|
var33.x2[0] = ORC_SWAP_W (var32.x2[0]);
|
|
var33.x2[1] = ORC_SWAP_W (var32.x2[1]);
|
|
/* 2: storel */
|
|
ptr0[i] = var33;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_UYVY_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_UYVY_YUY2");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_YUY2);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
|
|
orc_program_append_2 (p, "swapw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_planar_chroma_420_422 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_planar_chroma_420_422 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: copyb */
|
|
var33 = var32;
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
/* 3: loadb */
|
|
var34 = ptr4[i];
|
|
/* 4: copyb */
|
|
var35 = var34;
|
|
/* 5: storeb */
|
|
ptr1[i] = var35;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_planar_chroma_420_422 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: copyb */
|
|
var33 = var32;
|
|
/* 2: storeb */
|
|
ptr0[i] = var33;
|
|
/* 3: loadb */
|
|
var34 = ptr4[i];
|
|
/* 4: copyb */
|
|
var35 = var34;
|
|
/* 5: storeb */
|
|
ptr1[i] = var35;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_planar_chroma_420_422 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_planar_chroma_420_422");
|
|
orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_420_422);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_destination (p, 1, "d2");
|
|
orc_program_add_source (p, 1, "s1");
|
|
|
|
orc_program_append_2 (p, "copyb", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "copyb", 0, ORC_VAR_D2, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_planar_chroma_420_444 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_planar_chroma_420_444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: splatbw */
|
|
var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff);
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
/* 3: storew */
|
|
ptr1[i] = var34;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_planar_chroma_420_444 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: splatbw */
|
|
var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff);
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
/* 3: storew */
|
|
ptr1[i] = var34;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_planar_chroma_420_444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_planar_chroma_420_444");
|
|
orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_420_444);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 2, "d2");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "storew", 0, ORC_VAR_D2, ORC_VAR_T1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_planar_chroma_422_444 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_planar_chroma_422_444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: splatbw */
|
|
var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff);
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_planar_chroma_422_444 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
orc_int8 var33;
|
|
orc_union16 var34;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr4[i];
|
|
/* 1: splatbw */
|
|
var34.i = ((var33 & 0xff) << 8) | (var33 & 0xff);
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_planar_chroma_422_444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_planar_chroma_422_444");
|
|
orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_422_444);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_planar_chroma_444_422 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_planar_chroma_444_422 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var34 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var34.i;
|
|
var36 = _src.x2[1];
|
|
var37 = _src.x2[0];
|
|
}
|
|
/* 2: avgub */
|
|
var35 = ((orc_uint8) var36 + (orc_uint8) var37 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var35;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_planar_chroma_444_422 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var34 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var34.i;
|
|
var36 = _src.x2[1];
|
|
var37 = _src.x2[0];
|
|
}
|
|
/* 2: avgub */
|
|
var35 = ((orc_uint8) var36 + (orc_uint8) var37 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var35;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_planar_chroma_444_422 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_planar_chroma_444_422");
|
|
orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_444_422);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_temporary (p, 1, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_planar_chroma_444_420 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_planar_chroma_444_420 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
orc_union16 var35;
|
|
orc_union16 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_int8 var40;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var35 = ptr4[i];
|
|
/* 1: loadw */
|
|
var36 = ptr5[i];
|
|
/* 2: avgub */
|
|
var38.x2[0] =
|
|
((orc_uint8) var35.x2[0] + (orc_uint8) var36.x2[0] + 1) >> 1;
|
|
var38.x2[1] =
|
|
((orc_uint8) var35.x2[1] + (orc_uint8) var36.x2[1] + 1) >> 1;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var38.i;
|
|
var39 = _src.x2[1];
|
|
var40 = _src.x2[0];
|
|
}
|
|
/* 4: avgub */
|
|
var37 = ((orc_uint8) var39 + (orc_uint8) var40 + 1) >> 1;
|
|
/* 5: storeb */
|
|
ptr0[i] = var37;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_planar_chroma_444_420 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
orc_union16 var35;
|
|
orc_union16 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_int8 var39;
|
|
orc_int8 var40;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var35 = ptr4[i];
|
|
/* 1: loadw */
|
|
var36 = ptr5[i];
|
|
/* 2: avgub */
|
|
var38.x2[0] =
|
|
((orc_uint8) var35.x2[0] + (orc_uint8) var36.x2[0] + 1) >> 1;
|
|
var38.x2[1] =
|
|
((orc_uint8) var35.x2[1] + (orc_uint8) var36.x2[1] + 1) >> 1;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var38.i;
|
|
var39 = _src.x2[1];
|
|
var40 = _src.x2[0];
|
|
}
|
|
/* 4: avgub */
|
|
var37 = ((orc_uint8) var39 + (orc_uint8) var40 + 1) >> 1;
|
|
/* 5: storeb */
|
|
ptr0[i] = var37;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_planar_chroma_444_420 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_planar_chroma_444_420");
|
|
orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_444_420);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 2, "s2");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_planar_chroma_422_420 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_planar_chroma_422_420 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: loadb */
|
|
var33 = ptr5[i];
|
|
/* 2: avgub */
|
|
var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_planar_chroma_422_420 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
orc_int8 var32;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var32 = ptr4[i];
|
|
/* 1: loadb */
|
|
var33 = ptr5[i];
|
|
/* 2: avgub */
|
|
var34 = ((orc_uint8) var32 + (orc_uint8) var33 + 1) >> 1;
|
|
/* 3: storeb */
|
|
ptr0[i] = var34;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_planar_chroma_422_420 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_planar_chroma_422_420");
|
|
orc_program_set_backup_function (p, _backup_cogorc_planar_chroma_422_420);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_YUY2_AYUV */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_YUY2_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union64 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union64 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union32 var41;
|
|
orc_union32 var42;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
/* 2: loadpb */
|
|
var37.x2[0] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
var37.x2[1] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var36 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 3: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var40.x2[0];
|
|
var41.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var40.x2[1];
|
|
var41.x2[1] = _dest.i;
|
|
}
|
|
/* 4: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var39.i;
|
|
_dest.x2[1] = var39.i;
|
|
var42.i = _dest.i;
|
|
}
|
|
/* 5: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[0];
|
|
_dest.x2[1] = var42.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[1];
|
|
_dest.x2[1] = var42.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storeq */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_YUY2_AYUV (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union64 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union64 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union32 var41;
|
|
orc_union32 var42;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
/* 2: loadpb */
|
|
var37.x2[0] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
var37.x2[1] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var36 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 3: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var40.x2[0];
|
|
var41.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var40.x2[1];
|
|
var41.x2[1] = _dest.i;
|
|
}
|
|
/* 4: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var39.i;
|
|
_dest.x2[1] = var39.i;
|
|
var42.i = _dest.i;
|
|
}
|
|
/* 5: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[0];
|
|
_dest.x2[1] = var42.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[1];
|
|
_dest.x2[1] = var42.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storeq */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_YUY2_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_YUY2_AYUV");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_AYUV);
|
|
orc_program_add_destination (p, 8, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_constant (p, 2, 0x000000ff, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 4, "t3");
|
|
orc_program_add_temporary (p, 4, "t4");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T3, ORC_VAR_C1, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_UYVY_AYUV */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_UYVY_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union64 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union64 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union32 var41;
|
|
orc_union32 var42;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
/* 2: loadpb */
|
|
var37.x2[0] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
var37.x2[1] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var36 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 3: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var41.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var41.x2[1] = _dest.i;
|
|
}
|
|
/* 4: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var40.i;
|
|
_dest.x2[1] = var40.i;
|
|
var42.i = _dest.i;
|
|
}
|
|
/* 5: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[0];
|
|
_dest.x2[1] = var42.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[1];
|
|
_dest.x2[1] = var42.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storeq */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_UYVY_AYUV (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union64 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
orc_union64 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union32 var41;
|
|
orc_union32 var42;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
/* 2: loadpb */
|
|
var37.x2[0] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
var37.x2[1] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var36 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 3: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var41.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var41.x2[1] = _dest.i;
|
|
}
|
|
/* 4: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var40.i;
|
|
_dest.x2[1] = var40.i;
|
|
var42.i = _dest.i;
|
|
}
|
|
/* 5: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[0];
|
|
_dest.x2[1] = var42.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.x2[1];
|
|
_dest.x2[1] = var42.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storeq */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_UYVY_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_UYVY_AYUV");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_AYUV);
|
|
orc_program_add_destination (p, 8, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_constant (p, 2, 0x000000ff, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 4, "t3");
|
|
orc_program_add_temporary (p, 4, "t4");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_T2, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T3, ORC_VAR_C1, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_YUY2_Y42B */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_YUY2_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[0];
|
|
var37.x2[0] = _src.x2[1];
|
|
var34.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[1];
|
|
var37.x2[1] = _src.x2[1];
|
|
var34.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var37.i;
|
|
var35 = _src.x2[1];
|
|
var36 = _src.x2[0];
|
|
}
|
|
/* 4: storeb */
|
|
ptr2[i] = var35;
|
|
/* 5: storeb */
|
|
ptr1[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_YUY2_Y42B (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[0];
|
|
var37.x2[0] = _src.x2[1];
|
|
var34.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[1];
|
|
var37.x2[1] = _src.x2[1];
|
|
var34.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var37.i;
|
|
var35 = _src.x2[1];
|
|
var36 = _src.x2[0];
|
|
}
|
|
/* 4: storeb */
|
|
ptr2[i] = var35;
|
|
/* 5: storeb */
|
|
ptr1[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_YUY2_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_YUY2_Y42B");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_Y42B);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 1, "d2");
|
|
orc_program_add_destination (p, 1, "d3");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->params[ORC_VAR_D3] = d3_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_UYVY_Y42B */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_UYVY_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[0];
|
|
var34.x2[0] = _src.x2[1];
|
|
var37.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[1];
|
|
var34.x2[1] = _src.x2[1];
|
|
var37.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var37.i;
|
|
var35 = _src.x2[1];
|
|
var36 = _src.x2[0];
|
|
}
|
|
/* 4: storeb */
|
|
ptr2[i] = var35;
|
|
/* 5: storeb */
|
|
ptr1[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_UYVY_Y42B (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var33;
|
|
orc_union16 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var33 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[0];
|
|
var34.x2[0] = _src.x2[1];
|
|
var37.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var33.x2[1];
|
|
var34.x2[1] = _src.x2[1];
|
|
var37.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var34;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var37.i;
|
|
var35 = _src.x2[1];
|
|
var36 = _src.x2[0];
|
|
}
|
|
/* 4: storeb */
|
|
ptr2[i] = var35;
|
|
/* 5: storeb */
|
|
ptr1[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_UYVY_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_UYVY_Y42B");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_Y42B);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 1, "d2");
|
|
orc_program_add_destination (p, 1, "d3");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->params[ORC_VAR_D3] = d3_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_YUY2_Y444 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_YUY2_Y444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint16 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_union16 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var35;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_int8 var41;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var36.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var36.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var36;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var40 = _src.x2[1];
|
|
var41 = _src.x2[0];
|
|
}
|
|
/* 4: splatbw */
|
|
var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff);
|
|
/* 5: storew */
|
|
ptr1[i] = var37;
|
|
/* 6: splatbw */
|
|
var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff);
|
|
/* 7: storew */
|
|
ptr2[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_YUY2_Y444 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_union16 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var35;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_int8 var41;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var36.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var36.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var36;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var40 = _src.x2[1];
|
|
var41 = _src.x2[0];
|
|
}
|
|
/* 4: splatbw */
|
|
var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff);
|
|
/* 5: storew */
|
|
ptr1[i] = var37;
|
|
/* 6: splatbw */
|
|
var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff);
|
|
/* 7: storew */
|
|
ptr2[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_YUY2_Y444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint16 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_YUY2_Y444");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_YUY2_Y444);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 2, "d2");
|
|
orc_program_add_destination (p, 2, "d3");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D2, ORC_VAR_T2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D3, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->params[ORC_VAR_D3] = d3_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_UYVY_Y444 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_UYVY_Y444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint16 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_union16 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var35;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_int8 var41;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var36.x2[0] = _src.x2[1];
|
|
var39.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var36.x2[1] = _src.x2[1];
|
|
var39.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var36;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var40 = _src.x2[1];
|
|
var41 = _src.x2[0];
|
|
}
|
|
/* 4: splatbw */
|
|
var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff);
|
|
/* 5: storew */
|
|
ptr1[i] = var37;
|
|
/* 6: splatbw */
|
|
var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff);
|
|
/* 7: storew */
|
|
ptr2[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_UYVY_Y444 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_union16 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var35;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_int8 var40;
|
|
orc_int8 var41;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var36.x2[0] = _src.x2[1];
|
|
var39.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var36.x2[1] = _src.x2[1];
|
|
var39.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var36;
|
|
/* 3: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var40 = _src.x2[1];
|
|
var41 = _src.x2[0];
|
|
}
|
|
/* 4: splatbw */
|
|
var37.i = ((var41 & 0xff) << 8) | (var41 & 0xff);
|
|
/* 5: storew */
|
|
ptr1[i] = var37;
|
|
/* 6: splatbw */
|
|
var38.i = ((var40 & 0xff) << 8) | (var40 & 0xff);
|
|
/* 7: storew */
|
|
ptr2[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_UYVY_Y444 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint16 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_UYVY_Y444");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_Y444);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 2, "d2");
|
|
orc_program_add_destination (p, 2, "d3");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 1, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D2, ORC_VAR_T2, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splatbw", 0, ORC_VAR_D3, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->params[ORC_VAR_D3] = d3_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_UYVY_I420 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_UYVY_I420 (orc_uint16 * ORC_RESTRICT d1,
|
|
orc_uint16 * ORC_RESTRICT d2, orc_uint8 * ORC_RESTRICT d3,
|
|
orc_uint8 * ORC_RESTRICT d4, const orc_uint32 * ORC_RESTRICT s1,
|
|
const orc_uint32 * ORC_RESTRICT s2, int n)
|
|
{
|
|
int i;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
orc_int8 *ORC_RESTRICT ptr3;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
const orc_union32 *ORC_RESTRICT ptr5;
|
|
orc_union32 var35;
|
|
orc_union32 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
ptr0 = (orc_union16 *) d1;
|
|
ptr1 = (orc_union16 *) d2;
|
|
ptr2 = (orc_int8 *) d3;
|
|
ptr3 = (orc_int8 *) d4;
|
|
ptr4 = (orc_union32 *) s1;
|
|
ptr5 = (orc_union32 *) s2;
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var39;
|
|
/* 3: loadl */
|
|
var36 = ptr5[i];
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var41.x2[0] = _src.x2[1];
|
|
var42.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var41.x2[1] = _src.x2[1];
|
|
var42.x2[1] = _src.x2[0];
|
|
}
|
|
/* 5: storew */
|
|
ptr1[i] = var41;
|
|
/* 6: avgub */
|
|
var43.x2[0] = ((orc_uint8) var40.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] = ((orc_uint8) var40.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 7: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var37 = _src.x2[1];
|
|
var38 = _src.x2[0];
|
|
}
|
|
/* 8: storeb */
|
|
ptr3[i] = var37;
|
|
/* 9: storeb */
|
|
ptr2[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_UYVY_I420 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
orc_int8 *ORC_RESTRICT ptr3;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
const orc_union32 *ORC_RESTRICT ptr5;
|
|
orc_union32 var35;
|
|
orc_union32 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
ptr0 = (orc_union16 *) ex->arrays[0];
|
|
ptr1 = (orc_union16 *) ex->arrays[1];
|
|
ptr2 = (orc_int8 *) ex->arrays[2];
|
|
ptr3 = (orc_int8 *) ex->arrays[3];
|
|
ptr4 = (orc_union32 *) ex->arrays[4];
|
|
ptr5 = (orc_union32 *) ex->arrays[5];
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var35 = ptr4[i];
|
|
/* 1: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var35.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: storew */
|
|
ptr0[i] = var39;
|
|
/* 3: loadl */
|
|
var36 = ptr5[i];
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[0];
|
|
var41.x2[0] = _src.x2[1];
|
|
var42.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var36.x2[1];
|
|
var41.x2[1] = _src.x2[1];
|
|
var42.x2[1] = _src.x2[0];
|
|
}
|
|
/* 5: storew */
|
|
ptr1[i] = var41;
|
|
/* 6: avgub */
|
|
var43.x2[0] = ((orc_uint8) var40.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] = ((orc_uint8) var40.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 7: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var43.i;
|
|
var37 = _src.x2[1];
|
|
var38 = _src.x2[0];
|
|
}
|
|
/* 8: storeb */
|
|
ptr3[i] = var37;
|
|
/* 9: storeb */
|
|
ptr2[i] = var38;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_UYVY_I420 (orc_uint16 * ORC_RESTRICT d1,
|
|
orc_uint16 * ORC_RESTRICT d2, orc_uint8 * ORC_RESTRICT d3,
|
|
orc_uint8 * ORC_RESTRICT d4, const orc_uint32 * ORC_RESTRICT s1,
|
|
const orc_uint32 * ORC_RESTRICT s2, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_convert_UYVY_I420");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_UYVY_I420);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 2, "d2");
|
|
orc_program_add_destination (p, 1, "d3");
|
|
orc_program_add_destination (p, 1, "d4");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_source (p, 4, "s2");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "storew", 0, ORC_VAR_D1, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "storew", 0, ORC_VAR_D2, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D4, ORC_VAR_D3, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->arrays[ORC_VAR_D4] = d4;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_I420 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_I420 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, orc_uint8 * ORC_RESTRICT d4, int d4_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s2, int s2_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
orc_int8 *ORC_RESTRICT ptr3;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
const orc_union64 *ORC_RESTRICT ptr5;
|
|
orc_union64 var40;
|
|
orc_union16 var41;
|
|
orc_union64 var42;
|
|
orc_union16 var43;
|
|
orc_int8 var44;
|
|
orc_int8 var45;
|
|
orc_union32 var46;
|
|
orc_union32 var47;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union32 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_int8 var53;
|
|
orc_int8 var54;
|
|
orc_int8 var55;
|
|
orc_int8 var56;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j);
|
|
ptr3 = ORC_PTR_OFFSET (d4, d4_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var40 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.x2[0];
|
|
var46.x2[0] = _src.x2[1];
|
|
var47.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.x2[1];
|
|
var46.x2[1] = _src.x2[1];
|
|
var47.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var47.x2[0];
|
|
var41.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var47.x2[1];
|
|
var41.x2[1] = _src.x2[1];
|
|
}
|
|
/* 3: storew */
|
|
ptr0[i] = var41;
|
|
/* 4: loadq */
|
|
var42 = ptr5[i];
|
|
/* 5: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var42.x2[0];
|
|
var48.x2[0] = _src.x2[1];
|
|
var49.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var42.x2[1];
|
|
var48.x2[1] = _src.x2[1];
|
|
var49.x2[1] = _src.x2[0];
|
|
}
|
|
/* 6: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var49.x2[0];
|
|
var43.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var49.x2[1];
|
|
var43.x2[1] = _src.x2[1];
|
|
}
|
|
/* 7: storew */
|
|
ptr1[i] = var43;
|
|
/* 8: avgub */
|
|
var50.x4[0] =
|
|
((orc_uint8) var46.x4[0] + (orc_uint8) var48.x4[0] + 1) >> 1;
|
|
var50.x4[1] =
|
|
((orc_uint8) var46.x4[1] + (orc_uint8) var48.x4[1] + 1) >> 1;
|
|
var50.x4[2] =
|
|
((orc_uint8) var46.x4[2] + (orc_uint8) var48.x4[2] + 1) >> 1;
|
|
var50.x4[3] =
|
|
((orc_uint8) var46.x4[3] + (orc_uint8) var48.x4[3] + 1) >> 1;
|
|
/* 9: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var50.x2[0];
|
|
var51.x2[0] = _src.x2[1];
|
|
var52.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var50.x2[1];
|
|
var51.x2[1] = _src.x2[1];
|
|
var52.x2[1] = _src.x2[0];
|
|
}
|
|
/* 10: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var52.i;
|
|
var53 = _src.x2[1];
|
|
var54 = _src.x2[0];
|
|
}
|
|
/* 11: avgub */
|
|
var44 = ((orc_uint8) var53 + (orc_uint8) var54 + 1) >> 1;
|
|
/* 12: storeb */
|
|
ptr2[i] = var44;
|
|
/* 13: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var51.i;
|
|
var55 = _src.x2[1];
|
|
var56 = _src.x2[0];
|
|
}
|
|
/* 14: avgub */
|
|
var45 = ((orc_uint8) var55 + (orc_uint8) var56 + 1) >> 1;
|
|
/* 15: storeb */
|
|
ptr3[i] = var45;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_I420 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_union16 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
orc_int8 *ORC_RESTRICT ptr3;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
const orc_union64 *ORC_RESTRICT ptr5;
|
|
orc_union64 var40;
|
|
orc_union16 var41;
|
|
orc_union64 var42;
|
|
orc_union16 var43;
|
|
orc_int8 var44;
|
|
orc_int8 var45;
|
|
orc_union32 var46;
|
|
orc_union32 var47;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union32 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_int8 var53;
|
|
orc_int8 var54;
|
|
orc_int8 var55;
|
|
orc_int8 var56;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j);
|
|
ptr3 = ORC_PTR_OFFSET (ex->arrays[3], ex->params[3] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var40 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.x2[0];
|
|
var46.x2[0] = _src.x2[1];
|
|
var47.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.x2[1];
|
|
var46.x2[1] = _src.x2[1];
|
|
var47.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var47.x2[0];
|
|
var41.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var47.x2[1];
|
|
var41.x2[1] = _src.x2[1];
|
|
}
|
|
/* 3: storew */
|
|
ptr0[i] = var41;
|
|
/* 4: loadq */
|
|
var42 = ptr5[i];
|
|
/* 5: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var42.x2[0];
|
|
var48.x2[0] = _src.x2[1];
|
|
var49.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var42.x2[1];
|
|
var48.x2[1] = _src.x2[1];
|
|
var49.x2[1] = _src.x2[0];
|
|
}
|
|
/* 6: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var49.x2[0];
|
|
var43.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var49.x2[1];
|
|
var43.x2[1] = _src.x2[1];
|
|
}
|
|
/* 7: storew */
|
|
ptr1[i] = var43;
|
|
/* 8: avgub */
|
|
var50.x4[0] =
|
|
((orc_uint8) var46.x4[0] + (orc_uint8) var48.x4[0] + 1) >> 1;
|
|
var50.x4[1] =
|
|
((orc_uint8) var46.x4[1] + (orc_uint8) var48.x4[1] + 1) >> 1;
|
|
var50.x4[2] =
|
|
((orc_uint8) var46.x4[2] + (orc_uint8) var48.x4[2] + 1) >> 1;
|
|
var50.x4[3] =
|
|
((orc_uint8) var46.x4[3] + (orc_uint8) var48.x4[3] + 1) >> 1;
|
|
/* 9: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var50.x2[0];
|
|
var51.x2[0] = _src.x2[1];
|
|
var52.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var50.x2[1];
|
|
var51.x2[1] = _src.x2[1];
|
|
var52.x2[1] = _src.x2[0];
|
|
}
|
|
/* 10: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var52.i;
|
|
var53 = _src.x2[1];
|
|
var54 = _src.x2[0];
|
|
}
|
|
/* 11: avgub */
|
|
var44 = ((orc_uint8) var53 + (orc_uint8) var54 + 1) >> 1;
|
|
/* 12: storeb */
|
|
ptr2[i] = var44;
|
|
/* 13: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var51.i;
|
|
var55 = _src.x2[1];
|
|
var56 = _src.x2[0];
|
|
}
|
|
/* 14: avgub */
|
|
var45 = ((orc_uint8) var55 + (orc_uint8) var56 + 1) >> 1;
|
|
/* 15: storeb */
|
|
ptr3[i] = var45;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_I420 (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint16 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, orc_uint8 * ORC_RESTRICT d4, int d4_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s2, int s2_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_I420");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_I420);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 2, "d2");
|
|
orc_program_add_destination (p, 1, "d3");
|
|
orc_program_add_destination (p, 1, "d4");
|
|
orc_program_add_source (p, 8, "s1");
|
|
orc_program_add_source (p, 8, "s2");
|
|
orc_program_add_temporary (p, 4, "t1");
|
|
orc_program_add_temporary (p, 4, "t2");
|
|
orc_program_add_temporary (p, 4, "t3");
|
|
orc_program_add_temporary (p, 4, "t4");
|
|
orc_program_add_temporary (p, 2, "t5");
|
|
orc_program_add_temporary (p, 2, "t6");
|
|
orc_program_add_temporary (p, 1, "t7");
|
|
orc_program_add_temporary (p, 1, "t8");
|
|
|
|
orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_S2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D2, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 2, ORC_VAR_T4, ORC_VAR_T2, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 1, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T7, ORC_VAR_T8, ORC_VAR_T5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D3, ORC_VAR_T7, ORC_VAR_T8,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T7, ORC_VAR_T8, ORC_VAR_T6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_D4, ORC_VAR_T7, ORC_VAR_T8,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->params[ORC_VAR_D3] = d3_stride;
|
|
ex->arrays[ORC_VAR_D4] = d4;
|
|
ex->params[ORC_VAR_D4] = d4_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_YUY2 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
orc_union64 var37;
|
|
orc_union32 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var37 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var39.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 3: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 4: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[0];
|
|
var44.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[1];
|
|
var44.x2[1] = _src.x2[1];
|
|
}
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var44.x2[0];
|
|
_dest.x2[1] = var43.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var44.x2[1];
|
|
_dest.x2[1] = var43.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_YUY2 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
orc_union64 var37;
|
|
orc_union32 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var37 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var39.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 3: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 4: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[0];
|
|
var44.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[1];
|
|
var44.x2[1] = _src.x2[1];
|
|
}
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var44.x2[0];
|
|
_dest.x2[1] = var43.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var44.x2[1];
|
|
_dest.x2[1] = var43.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_YUY2");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_YUY2);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 8, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 4, "t4");
|
|
orc_program_add_temporary (p, 4, "t5");
|
|
|
|
orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 1, ORC_VAR_T1, ORC_VAR_T4,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_UYVY */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
orc_union64 var37;
|
|
orc_union32 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var37 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var39.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 3: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 4: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[0];
|
|
var44.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[1];
|
|
var44.x2[1] = _src.x2[1];
|
|
}
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[0];
|
|
_dest.x2[1] = var44.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[1];
|
|
_dest.x2[1] = var44.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_UYVY (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
orc_union64 var37;
|
|
orc_union32 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var37 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
var40.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var37.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
var40.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var39.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 3: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 4: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[0];
|
|
var44.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var40.x2[1];
|
|
var44.x2[1] = _src.x2[1];
|
|
}
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[0];
|
|
_dest.x2[1] = var44.x2[0];
|
|
var38.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[1];
|
|
_dest.x2[1] = var44.x2[1];
|
|
var38.x2[1] = _dest.i;
|
|
}
|
|
/* 6: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_UYVY");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_UYVY);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 8, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 4, "t4");
|
|
orc_program_add_temporary (p, 4, "t5");
|
|
|
|
orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T3, ORC_VAR_T5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 1, ORC_VAR_T1, ORC_VAR_T4,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_Y42B */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
orc_union64 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union32 var40;
|
|
orc_union32 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var36 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var36.x2[0];
|
|
var40.x2[0] = _src.x2[1];
|
|
var41.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var36.x2[1];
|
|
var40.x2[1] = _src.x2[1];
|
|
var41.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.i;
|
|
var42.i = _src.x2[1];
|
|
var43.i = _src.x2[0];
|
|
}
|
|
/* 3: avgub */
|
|
var44.x2[0] =
|
|
((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1;
|
|
var44.x2[1] =
|
|
((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1;
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var44.i;
|
|
var37 = _src.x2[1];
|
|
var38 = _src.x2[0];
|
|
}
|
|
/* 5: storeb */
|
|
ptr2[i] = var37;
|
|
/* 6: storeb */
|
|
ptr1[i] = var38;
|
|
/* 7: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
}
|
|
/* 8: storew */
|
|
ptr0[i] = var39;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_Y42B (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union16 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union64 *ORC_RESTRICT ptr4;
|
|
orc_union64 var36;
|
|
orc_int8 var37;
|
|
orc_int8 var38;
|
|
orc_union16 var39;
|
|
orc_union32 var40;
|
|
orc_union32 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
orc_union16 var44;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadq */
|
|
var36 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var36.x2[0];
|
|
var40.x2[0] = _src.x2[1];
|
|
var41.x2[0] = _src.x2[0];
|
|
}
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var36.x2[1];
|
|
var40.x2[1] = _src.x2[1];
|
|
var41.x2[1] = _src.x2[0];
|
|
}
|
|
/* 2: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.i;
|
|
var42.i = _src.x2[1];
|
|
var43.i = _src.x2[0];
|
|
}
|
|
/* 3: avgub */
|
|
var44.x2[0] =
|
|
((orc_uint8) var42.x2[0] + (orc_uint8) var43.x2[0] + 1) >> 1;
|
|
var44.x2[1] =
|
|
((orc_uint8) var42.x2[1] + (orc_uint8) var43.x2[1] + 1) >> 1;
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var44.i;
|
|
var37 = _src.x2[1];
|
|
var38 = _src.x2[0];
|
|
}
|
|
/* 5: storeb */
|
|
ptr2[i] = var37;
|
|
/* 6: storeb */
|
|
ptr1[i] = var38;
|
|
/* 7: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.x2[0];
|
|
var39.x2[0] = _src.x2[1];
|
|
}
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var41.x2[1];
|
|
var39.x2[1] = _src.x2[1];
|
|
}
|
|
/* 8: storew */
|
|
ptr0[i] = var39;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_Y42B (orc_uint16 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint64 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_Y42B");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_Y42B);
|
|
orc_program_add_destination (p, 2, "d1");
|
|
orc_program_add_destination (p, 1, "d2");
|
|
orc_program_add_destination (p, 1, "d3");
|
|
orc_program_add_source (p, 8, "s1");
|
|
orc_program_add_temporary (p, 4, "t1");
|
|
orc_program_add_temporary (p, 4, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
|
|
orc_program_append_2 (p, "splitlw", 1, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 1, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->params[ORC_VAR_D3] = d3_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_Y444 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_Y444 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr1 = ORC_PTR_OFFSET (d2, d2_stride * j);
|
|
ptr2 = ORC_PTR_OFFSET (d3, d3_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var34 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var34.i;
|
|
var38.i = _src.x2[1];
|
|
var39.i = _src.x2[0];
|
|
}
|
|
/* 2: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var38.i;
|
|
var35 = _src.x2[1];
|
|
var36 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr2[i] = var35;
|
|
/* 4: storeb */
|
|
ptr1[i] = var36;
|
|
/* 5: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var37 = _src.x2[1];
|
|
}
|
|
/* 6: storeb */
|
|
ptr0[i] = var37;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_Y444 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_int8 *ORC_RESTRICT ptr0;
|
|
orc_int8 *ORC_RESTRICT ptr1;
|
|
orc_int8 *ORC_RESTRICT ptr2;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr1 = ORC_PTR_OFFSET (ex->arrays[1], ex->params[1] * j);
|
|
ptr2 = ORC_PTR_OFFSET (ex->arrays[2], ex->params[2] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var34 = ptr4[i];
|
|
/* 1: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var34.i;
|
|
var38.i = _src.x2[1];
|
|
var39.i = _src.x2[0];
|
|
}
|
|
/* 2: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var38.i;
|
|
var35 = _src.x2[1];
|
|
var36 = _src.x2[0];
|
|
}
|
|
/* 3: storeb */
|
|
ptr2[i] = var35;
|
|
/* 4: storeb */
|
|
ptr1[i] = var36;
|
|
/* 5: select1wb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var39.i;
|
|
var37 = _src.x2[1];
|
|
}
|
|
/* 6: storeb */
|
|
ptr0[i] = var37;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_Y444 (orc_uint8 * ORC_RESTRICT d1, int d1_stride,
|
|
orc_uint8 * ORC_RESTRICT d2, int d2_stride, orc_uint8 * ORC_RESTRICT d3,
|
|
int d3_stride, const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n,
|
|
int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_Y444");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_Y444);
|
|
orc_program_add_destination (p, 1, "d1");
|
|
orc_program_add_destination (p, 1, "d2");
|
|
orc_program_add_destination (p, 1, "d3");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_D3, ORC_VAR_D2, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "select1wb", 0, ORC_VAR_D1, ORC_VAR_T1,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_D2] = d2;
|
|
ex->params[ORC_VAR_D2] = d2_stride;
|
|
ex->arrays[ORC_VAR_D3] = d3;
|
|
ex->params[ORC_VAR_D3] = d3_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_Y42B_YUY2 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_Y42B_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr5[i];
|
|
/* 1: loadb */
|
|
var34 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var37.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_Y42B_YUY2 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr5[i];
|
|
/* 1: loadb */
|
|
var34 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var37.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var35.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_Y42B_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_Y42B_YUY2");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_Y42B_YUY2);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_S3] = s3_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_Y42B_UYVY */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_Y42B_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr5[i];
|
|
/* 1: loadb */
|
|
var34 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var37.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var35.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var35.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_Y42B_UYVY (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var33;
|
|
orc_int8 var34;
|
|
orc_union16 var35;
|
|
orc_union32 var36;
|
|
orc_union16 var37;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var33 = ptr5[i];
|
|
/* 1: loadb */
|
|
var34 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var33;
|
|
_dest.x2[1] = var34;
|
|
var37.i = _dest.i;
|
|
}
|
|
/* 3: loadw */
|
|
var35 = ptr4[i];
|
|
/* 4: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[0];
|
|
_dest.x2[1] = var35.x2[0];
|
|
var36.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var37.x2[1];
|
|
_dest.x2[1] = var35.x2[1];
|
|
var36.x2[1] = _dest.i;
|
|
}
|
|
/* 5: storel */
|
|
ptr0[i] = var36;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_Y42B_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_Y42B_UYVY");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_Y42B_UYVY);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_S3] = s3_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_Y42B_AYUV */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_Y42B_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union64 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_union64 var40;
|
|
orc_union16 var41;
|
|
orc_union32 var42;
|
|
orc_union32 var43;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j);
|
|
|
|
/* 3: loadpb */
|
|
var38.x2[0] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
var38.x2[1] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr5[i];
|
|
/* 1: loadb */
|
|
var37 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var41.i = _dest.i;
|
|
}
|
|
/* 4: loadw */
|
|
var39 = ptr4[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var42.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var42.x2[1] = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.i;
|
|
_dest.x2[1] = var41.i;
|
|
var43.i = _dest.i;
|
|
}
|
|
/* 7: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var42.x2[0];
|
|
_dest.x2[1] = var43.x2[0];
|
|
var40.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var42.x2[1];
|
|
_dest.x2[1] = var43.x2[1];
|
|
var40.x2[1] = _dest.i;
|
|
}
|
|
/* 8: storeq */
|
|
ptr0[i] = var40;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_Y42B_AYUV (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union64 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union16 var38;
|
|
orc_union16 var39;
|
|
orc_union64 var40;
|
|
orc_union16 var41;
|
|
orc_union32 var42;
|
|
orc_union32 var43;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j);
|
|
|
|
/* 3: loadpb */
|
|
var38.x2[0] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
var38.x2[1] = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var36 = ptr5[i];
|
|
/* 1: loadb */
|
|
var37 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var41.i = _dest.i;
|
|
}
|
|
/* 4: loadw */
|
|
var39 = ptr4[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[0];
|
|
_dest.x2[1] = var39.x2[0];
|
|
var42.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[1];
|
|
_dest.x2[1] = var39.x2[1];
|
|
var42.x2[1] = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var41.i;
|
|
_dest.x2[1] = var41.i;
|
|
var43.i = _dest.i;
|
|
}
|
|
/* 7: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var42.x2[0];
|
|
_dest.x2[1] = var43.x2[0];
|
|
var40.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var42.x2[1];
|
|
_dest.x2[1] = var43.x2[1];
|
|
var40.x2[1] = _dest.i;
|
|
}
|
|
/* 8: storeq */
|
|
ptr0[i] = var40;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_Y42B_AYUV (orc_uint64 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_Y42B_AYUV");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_Y42B_AYUV);
|
|
orc_program_add_destination (p, 8, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_constant (p, 1, 0x000000ff, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 4, "t3");
|
|
orc_program_add_temporary (p, 4, "t4");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T4, ORC_VAR_C1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T3, ORC_VAR_T1, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 1, ORC_VAR_D1, ORC_VAR_T4, ORC_VAR_T3,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_S3] = s3_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_Y444_YUY2 */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_Y444_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint16 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_union16 *ORC_RESTRICT ptr6;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var36 = ptr5[i];
|
|
/* 1: loadw */
|
|
var37 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var40.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var40.x2[1] = _dest.i;
|
|
}
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 4: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 5: loadw */
|
|
var38 = ptr4[i];
|
|
/* 6: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[0];
|
|
_dest.x2[1] = var43.x2[0];
|
|
var39.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[1];
|
|
_dest.x2[1] = var43.x2[1];
|
|
var39.x2[1] = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var39;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_Y444_YUY2 (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_union16 *ORC_RESTRICT ptr6;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var36 = ptr5[i];
|
|
/* 1: loadw */
|
|
var37 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var40.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var40.x2[1] = _dest.i;
|
|
}
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 4: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 5: loadw */
|
|
var38 = ptr4[i];
|
|
/* 6: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[0];
|
|
_dest.x2[1] = var43.x2[0];
|
|
var39.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var38.x2[1];
|
|
_dest.x2[1] = var43.x2[1];
|
|
var39.x2[1] = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var39;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_Y444_YUY2 (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint16 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_Y444_YUY2");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_Y444_YUY2);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 2, "s2");
|
|
orc_program_add_source (p, 2, "s3");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 4, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_S3] = s3_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_Y444_UYVY */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_Y444_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint16 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_union16 *ORC_RESTRICT ptr6;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var36 = ptr5[i];
|
|
/* 1: loadw */
|
|
var37 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var40.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var40.x2[1] = _dest.i;
|
|
}
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 4: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 5: loadw */
|
|
var38 = ptr4[i];
|
|
/* 6: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[0];
|
|
_dest.x2[1] = var38.x2[0];
|
|
var39.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[1];
|
|
_dest.x2[1] = var38.x2[1];
|
|
var39.x2[1] = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var39;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_Y444_UYVY (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union16 *ORC_RESTRICT ptr4;
|
|
const orc_union16 *ORC_RESTRICT ptr5;
|
|
const orc_union16 *ORC_RESTRICT ptr6;
|
|
orc_union16 var36;
|
|
orc_union16 var37;
|
|
orc_union16 var38;
|
|
orc_union32 var39;
|
|
orc_union32 var40;
|
|
orc_union16 var41;
|
|
orc_union16 var42;
|
|
orc_union16 var43;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j);
|
|
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadw */
|
|
var36 = ptr5[i];
|
|
/* 1: loadw */
|
|
var37 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[0];
|
|
_dest.x2[1] = var37.x2[0];
|
|
var40.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36.x2[1];
|
|
_dest.x2[1] = var37.x2[1];
|
|
var40.x2[1] = _dest.i;
|
|
}
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var40.i;
|
|
var41.i = _src.x2[1];
|
|
var42.i = _src.x2[0];
|
|
}
|
|
/* 4: avgub */
|
|
var43.x2[0] =
|
|
((orc_uint8) var41.x2[0] + (orc_uint8) var42.x2[0] + 1) >> 1;
|
|
var43.x2[1] =
|
|
((orc_uint8) var41.x2[1] + (orc_uint8) var42.x2[1] + 1) >> 1;
|
|
/* 5: loadw */
|
|
var38 = ptr4[i];
|
|
/* 6: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[0];
|
|
_dest.x2[1] = var38.x2[0];
|
|
var39.x2[0] = _dest.i;
|
|
}
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var43.x2[1];
|
|
_dest.x2[1] = var38.x2[1];
|
|
var39.x2[1] = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var39;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_Y444_UYVY (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint16 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint16 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_Y444_UYVY");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_Y444_UYVY);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 2, "s1");
|
|
orc_program_add_source (p, 2, "s2");
|
|
orc_program_add_source (p, 2, "s3");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 4, "t2");
|
|
orc_program_add_temporary (p, 2, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_T2, ORC_VAR_S2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T3, ORC_VAR_T4, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 1, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 1, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_S3] = s3_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_Y444_AYUV */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_Y444_AYUV (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
ptr5 = ORC_PTR_OFFSET (s2, s2_stride * j);
|
|
ptr6 = ORC_PTR_OFFSET (s3, s3_stride * j);
|
|
|
|
/* 3: loadpb */
|
|
var36 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr5[i];
|
|
/* 1: loadb */
|
|
var35 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var34;
|
|
_dest.x2[1] = var35;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 4: loadb */
|
|
var37 = ptr4[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var40.i;
|
|
_dest.x2[1] = var39.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_Y444_AYUV (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var34;
|
|
orc_int8 var35;
|
|
orc_int8 var36;
|
|
orc_int8 var37;
|
|
orc_union32 var38;
|
|
orc_union16 var39;
|
|
orc_union16 var40;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
ptr5 = ORC_PTR_OFFSET (ex->arrays[5], ex->params[5] * j);
|
|
ptr6 = ORC_PTR_OFFSET (ex->arrays[6], ex->params[6] * j);
|
|
|
|
/* 3: loadpb */
|
|
var36 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var34 = ptr5[i];
|
|
/* 1: loadb */
|
|
var35 = ptr6[i];
|
|
/* 2: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var34;
|
|
_dest.x2[1] = var35;
|
|
var39.i = _dest.i;
|
|
}
|
|
/* 4: loadb */
|
|
var37 = ptr4[i];
|
|
/* 5: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var36;
|
|
_dest.x2[1] = var37;
|
|
var40.i = _dest.i;
|
|
}
|
|
/* 6: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var40.i;
|
|
_dest.x2[1] = var39.i;
|
|
var38.i = _dest.i;
|
|
}
|
|
/* 7: storel */
|
|
ptr0[i] = var38;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_Y444_AYUV (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s1, int s1_stride,
|
|
const orc_uint8 * ORC_RESTRICT s2, int s2_stride,
|
|
const orc_uint8 * ORC_RESTRICT s3, int s3_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_Y444_AYUV");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_Y444_AYUV);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_constant (p, 1, 0x000000ff, "c1");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_S2, ORC_VAR_S3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_C1, ORC_VAR_S1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_D1, ORC_VAR_T2, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->params[ORC_VAR_S2] = s2_stride;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->params[ORC_VAR_S3] = s3_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_ARGB */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_ARGB (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var61;
|
|
_dest.x2[1] = var87;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var88;
|
|
_dest.x2[1] = var89;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_ARGB (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var61;
|
|
_dest.x2[1] = var87;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var88;
|
|
_dest.x2[1] = var89;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_ARGB (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_ARGB");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_ARGB);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_constant (p, 1, 0x00000008, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x0000002a, "c3");
|
|
orc_program_add_constant (p, 4, 0x00000067, "c4");
|
|
orc_program_add_constant (p, 4, 0x00000004, "c5");
|
|
orc_program_add_constant (p, 4, 0x00000064, "c6");
|
|
orc_program_add_constant (p, 4, 0x00000068, "c7");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 1, "t4");
|
|
orc_program_add_temporary (p, 1, "t5");
|
|
orc_program_add_temporary (p, 1, "t6");
|
|
orc_program_add_temporary (p, 2, "t7");
|
|
orc_program_add_temporary (p, 2, "t8");
|
|
orc_program_add_temporary (p, 2, "t9");
|
|
orc_program_add_temporary (p, 2, "t10");
|
|
orc_program_add_temporary (p, 2, "t11");
|
|
orc_program_add_temporary (p, 2, "t12");
|
|
orc_program_add_temporary (p, 1, "t13");
|
|
orc_program_add_temporary (p, 1, "t14");
|
|
orc_program_add_temporary (p, 1, "t15");
|
|
orc_program_add_temporary (p, 4, "t16");
|
|
|
|
orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T9, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T8, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T3,
|
|
ORC_VAR_T13, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T14,
|
|
ORC_VAR_T15, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_BGRA */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_BGRA (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var89;
|
|
_dest.x2[1] = var88;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var87;
|
|
_dest.x2[1] = var61;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_BGRA (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var89;
|
|
_dest.x2[1] = var88;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var87;
|
|
_dest.x2[1] = var61;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_BGRA (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_BGRA");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_BGRA);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_constant (p, 1, 0x00000008, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x0000002a, "c3");
|
|
orc_program_add_constant (p, 4, 0x00000067, "c4");
|
|
orc_program_add_constant (p, 4, 0x00000004, "c5");
|
|
orc_program_add_constant (p, 4, 0x00000064, "c6");
|
|
orc_program_add_constant (p, 4, 0x00000068, "c7");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 1, "t4");
|
|
orc_program_add_temporary (p, 1, "t5");
|
|
orc_program_add_temporary (p, 1, "t6");
|
|
orc_program_add_temporary (p, 2, "t7");
|
|
orc_program_add_temporary (p, 2, "t8");
|
|
orc_program_add_temporary (p, 2, "t9");
|
|
orc_program_add_temporary (p, 2, "t10");
|
|
orc_program_add_temporary (p, 2, "t11");
|
|
orc_program_add_temporary (p, 2, "t12");
|
|
orc_program_add_temporary (p, 1, "t13");
|
|
orc_program_add_temporary (p, 1, "t14");
|
|
orc_program_add_temporary (p, 1, "t15");
|
|
orc_program_add_temporary (p, 4, "t16");
|
|
|
|
orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T9, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T8, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T15,
|
|
ORC_VAR_T14, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T13,
|
|
ORC_VAR_T3, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_ABGR */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_ABGR (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var61;
|
|
_dest.x2[1] = var89;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var88;
|
|
_dest.x2[1] = var87;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_ABGR (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var61;
|
|
_dest.x2[1] = var89;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var88;
|
|
_dest.x2[1] = var87;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_ABGR (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_ABGR");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_ABGR);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_constant (p, 1, 0x00000008, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x0000002a, "c3");
|
|
orc_program_add_constant (p, 4, 0x00000067, "c4");
|
|
orc_program_add_constant (p, 4, 0x00000004, "c5");
|
|
orc_program_add_constant (p, 4, 0x00000064, "c6");
|
|
orc_program_add_constant (p, 4, 0x00000068, "c7");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 1, "t4");
|
|
orc_program_add_temporary (p, 1, "t5");
|
|
orc_program_add_temporary (p, 1, "t6");
|
|
orc_program_add_temporary (p, 2, "t7");
|
|
orc_program_add_temporary (p, 2, "t8");
|
|
orc_program_add_temporary (p, 2, "t9");
|
|
orc_program_add_temporary (p, 2, "t10");
|
|
orc_program_add_temporary (p, 2, "t11");
|
|
orc_program_add_temporary (p, 2, "t12");
|
|
orc_program_add_temporary (p, 1, "t13");
|
|
orc_program_add_temporary (p, 1, "t14");
|
|
orc_program_add_temporary (p, 1, "t15");
|
|
orc_program_add_temporary (p, 4, "t16");
|
|
|
|
orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T9, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T8, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T3,
|
|
ORC_VAR_T15, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T14,
|
|
ORC_VAR_T13, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_AYUV_RGBA */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_AYUV_RGBA (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
int i;
|
|
int j;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
|
|
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var87;
|
|
_dest.x2[1] = var88;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var89;
|
|
_dest.x2[1] = var61;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_AYUV_RGBA (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int j;
|
|
int n = ex->n;
|
|
int m = ex->params[ORC_VAR_A1];
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_union32 *ORC_RESTRICT ptr4;
|
|
orc_union32 var48;
|
|
orc_union32 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_union16 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_int8 var87;
|
|
orc_int8 var88;
|
|
orc_int8 var89;
|
|
orc_union16 var90;
|
|
orc_union16 var91;
|
|
orc_union32 var92;
|
|
|
|
for (j = 0; j < m; j++) {
|
|
ptr0 = ORC_PTR_OFFSET (ex->arrays[0], ex->params[0] * j);
|
|
ptr4 = ORC_PTR_OFFSET (ex->arrays[4], ex->params[4] * j);
|
|
|
|
/* 1: loadpb */
|
|
var49.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var49.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 14: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 21: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 25: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 29: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 40: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadl */
|
|
var48 = ptr4[i];
|
|
/* 2: subb */
|
|
var57.x4[0] = var48.x4[0] - var49.x4[0];
|
|
var57.x4[1] = var48.x4[1] - var49.x4[1];
|
|
var57.x4[2] = var48.x4[2] - var49.x4[2];
|
|
var57.x4[3] = var48.x4[3] - var49.x4[3];
|
|
/* 3: splitlw */
|
|
{
|
|
orc_union32 _src;
|
|
_src.i = var57.i;
|
|
var58.i = _src.x2[1];
|
|
var59.i = _src.x2[0];
|
|
}
|
|
/* 4: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var59.i;
|
|
var60 = _src.x2[1];
|
|
var61 = _src.x2[0];
|
|
}
|
|
/* 5: splitwb */
|
|
{
|
|
orc_union16 _src;
|
|
_src.i = var58.i;
|
|
var62 = _src.x2[1];
|
|
var63 = _src.x2[0];
|
|
}
|
|
/* 6: convsbw */
|
|
var64.i = var60;
|
|
/* 7: convsbw */
|
|
var65.i = var63;
|
|
/* 8: convsbw */
|
|
var66.i = var62;
|
|
/* 10: mullw */
|
|
var67.i = (var64.i * var50.i) & 0xffff;
|
|
/* 11: shrsw */
|
|
var68.i = var67.i >> 8;
|
|
/* 12: addssw */
|
|
var69.i = ORC_CLAMP_SW (var64.i + var68.i);
|
|
/* 13: addssw */
|
|
var70.i = ORC_CLAMP_SW (var69.i + var66.i);
|
|
/* 15: mullw */
|
|
var71.i = (var66.i * var51.i) & 0xffff;
|
|
/* 16: shrsw */
|
|
var72.i = var71.i >> 8;
|
|
/* 17: subssw */
|
|
var73.i = ORC_CLAMP_SW (var70.i - var72.i);
|
|
/* 18: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var66.i);
|
|
/* 19: addssw */
|
|
var75.i = ORC_CLAMP_SW (var69.i + var65.i);
|
|
/* 20: addssw */
|
|
var76.i = ORC_CLAMP_SW (var75.i + var65.i);
|
|
/* 22: mullw */
|
|
var77.i = (var65.i * var52.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var78.i = var77.i >> 8;
|
|
/* 24: addssw */
|
|
var79.i = ORC_CLAMP_SW (var76.i + var78.i);
|
|
/* 26: mullw */
|
|
var80.i = (var65.i * var53.i) & 0xffff;
|
|
/* 27: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 28: subssw */
|
|
var82.i = ORC_CLAMP_SW (var69.i - var81.i);
|
|
/* 30: mullw */
|
|
var83.i = (var66.i * var54.i) & 0xffff;
|
|
/* 31: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 32: subssw */
|
|
var85.i = ORC_CLAMP_SW (var82.i - var84.i);
|
|
/* 33: subssw */
|
|
var86.i = ORC_CLAMP_SW (var85.i - var84.i);
|
|
/* 34: convssswb */
|
|
var87 = ORC_CLAMP_SB (var74.i);
|
|
/* 35: convssswb */
|
|
var88 = ORC_CLAMP_SB (var86.i);
|
|
/* 36: convssswb */
|
|
var89 = ORC_CLAMP_SB (var79.i);
|
|
/* 37: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var87;
|
|
_dest.x2[1] = var88;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 38: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var89;
|
|
_dest.x2[1] = var61;
|
|
var91.i = _dest.i;
|
|
}
|
|
/* 39: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var90.i;
|
|
_dest.x2[1] = var91.i;
|
|
var92.i = _dest.i;
|
|
}
|
|
/* 41: addb */
|
|
var56.x4[0] = var92.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var92.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var92.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var92.x4[3] + var55.x4[3];
|
|
/* 42: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_AYUV_RGBA (orc_uint32 * ORC_RESTRICT d1, int d1_stride,
|
|
const orc_uint32 * ORC_RESTRICT s1, int s1_stride, int n, int m)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_2d (p);
|
|
orc_program_set_name (p, "cogorc_convert_AYUV_RGBA");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_AYUV_RGBA);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 4, "s1");
|
|
orc_program_add_constant (p, 1, 0x00000008, "c1");
|
|
orc_program_add_constant (p, 4, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x0000002a, "c3");
|
|
orc_program_add_constant (p, 4, 0x00000067, "c4");
|
|
orc_program_add_constant (p, 4, 0x00000004, "c5");
|
|
orc_program_add_constant (p, 4, 0x00000064, "c6");
|
|
orc_program_add_constant (p, 4, 0x00000068, "c7");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 1, "t4");
|
|
orc_program_add_temporary (p, 1, "t5");
|
|
orc_program_add_temporary (p, 1, "t6");
|
|
orc_program_add_temporary (p, 2, "t7");
|
|
orc_program_add_temporary (p, 2, "t8");
|
|
orc_program_add_temporary (p, 2, "t9");
|
|
orc_program_add_temporary (p, 2, "t10");
|
|
orc_program_add_temporary (p, 2, "t11");
|
|
orc_program_add_temporary (p, 2, "t12");
|
|
orc_program_add_temporary (p, 1, "t13");
|
|
orc_program_add_temporary (p, 1, "t14");
|
|
orc_program_add_temporary (p, 1, "t15");
|
|
orc_program_add_temporary (p, 4, "t16");
|
|
|
|
orc_program_append_2 (p, "subb", 2, 47, ORC_VAR_S1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitlw", 0, ORC_VAR_T1, ORC_VAR_T2, 47,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "splitwb", 0, ORC_VAR_T6, ORC_VAR_T5, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T9, ORC_VAR_T6, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T7, ORC_VAR_T9,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T9, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T7, ORC_VAR_T8,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T8, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T12, ORC_VAR_T12,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T8, ORC_VAR_C6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T9, ORC_VAR_C7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T11, ORC_VAR_T11,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T14, ORC_VAR_T11,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T15, ORC_VAR_T12,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T13,
|
|
ORC_VAR_T14, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T15,
|
|
ORC_VAR_T3, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, 47, ORC_VAR_T1, ORC_VAR_T2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, 47, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ORC_EXECUTOR_M (ex) = m;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->params[ORC_VAR_D1] = d1_stride;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->params[ORC_VAR_S1] = s1_stride;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_I420_BGRA */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_I420_BGRA (orc_uint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var45;
|
|
orc_int8 var46;
|
|
orc_int8 var47;
|
|
orc_int8 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_int8 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_int8 var57;
|
|
orc_union16 var58;
|
|
orc_int8 var59;
|
|
orc_int8 var60;
|
|
orc_union16 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_int8 var85;
|
|
orc_int8 var86;
|
|
orc_int8 var87;
|
|
orc_union16 var88;
|
|
orc_union16 var89;
|
|
orc_union32 var90;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
|
|
/* 1: loadpb */
|
|
var46 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 5: loadpb */
|
|
var47 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpb */
|
|
var48 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 12: loadpw */
|
|
var49.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 17: loadpw */
|
|
var50.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 24: loadpw */
|
|
var51.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 28: loadpw */
|
|
var52.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 32: loadpw */
|
|
var53.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 41: loadpb */
|
|
var54 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
/* 44: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var45 = ptr4[i];
|
|
/* 2: subb */
|
|
var57 = var45 - var46;
|
|
/* 3: convsbw */
|
|
var58.i = var57;
|
|
/* 4: loadupib */
|
|
var59 =
|
|
(i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] +
|
|
1) >> 1 : ptr5[i >> 1];
|
|
/* 6: subb */
|
|
var60 = var59 - var47;
|
|
/* 7: convsbw */
|
|
var61.i = var60;
|
|
/* 8: loadupib */
|
|
var62 =
|
|
(i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] +
|
|
1) >> 1 : ptr6[i >> 1];
|
|
/* 10: subb */
|
|
var63 = var62 - var48;
|
|
/* 11: convsbw */
|
|
var64.i = var63;
|
|
/* 13: mullw */
|
|
var65.i = (var58.i * var49.i) & 0xffff;
|
|
/* 14: shrsw */
|
|
var66.i = var65.i >> 8;
|
|
/* 15: addssw */
|
|
var67.i = ORC_CLAMP_SW (var58.i + var66.i);
|
|
/* 16: addssw */
|
|
var68.i = ORC_CLAMP_SW (var67.i + var64.i);
|
|
/* 18: mullw */
|
|
var69.i = (var64.i * var50.i) & 0xffff;
|
|
/* 19: shrsw */
|
|
var70.i = var69.i >> 8;
|
|
/* 20: subssw */
|
|
var71.i = ORC_CLAMP_SW (var68.i - var70.i);
|
|
/* 21: addssw */
|
|
var72.i = ORC_CLAMP_SW (var71.i + var64.i);
|
|
/* 22: addssw */
|
|
var73.i = ORC_CLAMP_SW (var67.i + var61.i);
|
|
/* 23: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var61.i);
|
|
/* 25: mullw */
|
|
var75.i = (var61.i * var51.i) & 0xffff;
|
|
/* 26: shrsw */
|
|
var76.i = var75.i >> 8;
|
|
/* 27: addssw */
|
|
var77.i = ORC_CLAMP_SW (var74.i + var76.i);
|
|
/* 29: mullw */
|
|
var78.i = (var61.i * var52.i) & 0xffff;
|
|
/* 30: shrsw */
|
|
var79.i = var78.i >> 8;
|
|
/* 31: subssw */
|
|
var80.i = ORC_CLAMP_SW (var67.i - var79.i);
|
|
/* 33: mullw */
|
|
var81.i = (var64.i * var53.i) & 0xffff;
|
|
/* 34: shrsw */
|
|
var82.i = var81.i >> 8;
|
|
/* 35: subssw */
|
|
var83.i = ORC_CLAMP_SW (var80.i - var82.i);
|
|
/* 36: subssw */
|
|
var84.i = ORC_CLAMP_SW (var83.i - var82.i);
|
|
/* 37: convssswb */
|
|
var85 = ORC_CLAMP_SB (var72.i);
|
|
/* 38: convssswb */
|
|
var86 = ORC_CLAMP_SB (var84.i);
|
|
/* 39: convssswb */
|
|
var87 = ORC_CLAMP_SB (var77.i);
|
|
/* 40: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var87;
|
|
_dest.x2[1] = var86;
|
|
var88.i = _dest.i;
|
|
}
|
|
/* 42: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var85;
|
|
_dest.x2[1] = var54;
|
|
var89.i = _dest.i;
|
|
}
|
|
/* 43: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var88.i;
|
|
_dest.x2[1] = var89.i;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 45: addb */
|
|
var56.x4[0] = var90.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var90.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var90.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var90.x4[3] + var55.x4[3];
|
|
/* 46: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_I420_BGRA (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
orc_int8 var45;
|
|
orc_int8 var46;
|
|
orc_int8 var47;
|
|
orc_int8 var48;
|
|
orc_union16 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_int8 var54;
|
|
orc_union32 var55;
|
|
orc_union32 var56;
|
|
orc_int8 var57;
|
|
orc_union16 var58;
|
|
orc_int8 var59;
|
|
orc_int8 var60;
|
|
orc_union16 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_union16 var65;
|
|
orc_union16 var66;
|
|
orc_union16 var67;
|
|
orc_union16 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_int8 var85;
|
|
orc_int8 var86;
|
|
orc_int8 var87;
|
|
orc_union16 var88;
|
|
orc_union16 var89;
|
|
orc_union32 var90;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
|
|
/* 1: loadpb */
|
|
var46 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 5: loadpb */
|
|
var47 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 9: loadpb */
|
|
var48 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 12: loadpw */
|
|
var49.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 17: loadpw */
|
|
var50.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 24: loadpw */
|
|
var51.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 28: loadpw */
|
|
var52.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 32: loadpw */
|
|
var53.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 41: loadpb */
|
|
var54 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
/* 44: loadpb */
|
|
var55.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var55.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var45 = ptr4[i];
|
|
/* 2: subb */
|
|
var57 = var45 - var46;
|
|
/* 3: convsbw */
|
|
var58.i = var57;
|
|
/* 4: loadupib */
|
|
var59 =
|
|
(i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] +
|
|
1) >> 1 : ptr5[i >> 1];
|
|
/* 6: subb */
|
|
var60 = var59 - var47;
|
|
/* 7: convsbw */
|
|
var61.i = var60;
|
|
/* 8: loadupib */
|
|
var62 =
|
|
(i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] +
|
|
1) >> 1 : ptr6[i >> 1];
|
|
/* 10: subb */
|
|
var63 = var62 - var48;
|
|
/* 11: convsbw */
|
|
var64.i = var63;
|
|
/* 13: mullw */
|
|
var65.i = (var58.i * var49.i) & 0xffff;
|
|
/* 14: shrsw */
|
|
var66.i = var65.i >> 8;
|
|
/* 15: addssw */
|
|
var67.i = ORC_CLAMP_SW (var58.i + var66.i);
|
|
/* 16: addssw */
|
|
var68.i = ORC_CLAMP_SW (var67.i + var64.i);
|
|
/* 18: mullw */
|
|
var69.i = (var64.i * var50.i) & 0xffff;
|
|
/* 19: shrsw */
|
|
var70.i = var69.i >> 8;
|
|
/* 20: subssw */
|
|
var71.i = ORC_CLAMP_SW (var68.i - var70.i);
|
|
/* 21: addssw */
|
|
var72.i = ORC_CLAMP_SW (var71.i + var64.i);
|
|
/* 22: addssw */
|
|
var73.i = ORC_CLAMP_SW (var67.i + var61.i);
|
|
/* 23: addssw */
|
|
var74.i = ORC_CLAMP_SW (var73.i + var61.i);
|
|
/* 25: mullw */
|
|
var75.i = (var61.i * var51.i) & 0xffff;
|
|
/* 26: shrsw */
|
|
var76.i = var75.i >> 8;
|
|
/* 27: addssw */
|
|
var77.i = ORC_CLAMP_SW (var74.i + var76.i);
|
|
/* 29: mullw */
|
|
var78.i = (var61.i * var52.i) & 0xffff;
|
|
/* 30: shrsw */
|
|
var79.i = var78.i >> 8;
|
|
/* 31: subssw */
|
|
var80.i = ORC_CLAMP_SW (var67.i - var79.i);
|
|
/* 33: mullw */
|
|
var81.i = (var64.i * var53.i) & 0xffff;
|
|
/* 34: shrsw */
|
|
var82.i = var81.i >> 8;
|
|
/* 35: subssw */
|
|
var83.i = ORC_CLAMP_SW (var80.i - var82.i);
|
|
/* 36: subssw */
|
|
var84.i = ORC_CLAMP_SW (var83.i - var82.i);
|
|
/* 37: convssswb */
|
|
var85 = ORC_CLAMP_SB (var72.i);
|
|
/* 38: convssswb */
|
|
var86 = ORC_CLAMP_SB (var84.i);
|
|
/* 39: convssswb */
|
|
var87 = ORC_CLAMP_SB (var77.i);
|
|
/* 40: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var87;
|
|
_dest.x2[1] = var86;
|
|
var88.i = _dest.i;
|
|
}
|
|
/* 42: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var85;
|
|
_dest.x2[1] = var54;
|
|
var89.i = _dest.i;
|
|
}
|
|
/* 43: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var88.i;
|
|
_dest.x2[1] = var89.i;
|
|
var90.i = _dest.i;
|
|
}
|
|
/* 45: addb */
|
|
var56.x4[0] = var90.x4[0] + var55.x4[0];
|
|
var56.x4[1] = var90.x4[1] + var55.x4[1];
|
|
var56.x4[2] = var90.x4[2] + var55.x4[2];
|
|
var56.x4[3] = var90.x4[3] + var55.x4[3];
|
|
/* 46: storel */
|
|
ptr0[i] = var56;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_I420_BGRA (orc_uint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_convert_I420_BGRA");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_I420_BGRA);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_constant (p, 1, 0x00000008, "c1");
|
|
orc_program_add_constant (p, 1, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x0000002a, "c3");
|
|
orc_program_add_constant (p, 4, 0x00000067, "c4");
|
|
orc_program_add_constant (p, 4, 0x00000004, "c5");
|
|
orc_program_add_constant (p, 4, 0x00000064, "c6");
|
|
orc_program_add_constant (p, 4, 0x00000068, "c7");
|
|
orc_program_add_constant (p, 4, 0x000000ff, "c8");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 2, "t4");
|
|
orc_program_add_temporary (p, 2, "t5");
|
|
orc_program_add_temporary (p, 2, "t6");
|
|
orc_program_add_temporary (p, 2, "t7");
|
|
orc_program_add_temporary (p, 2, "t8");
|
|
orc_program_add_temporary (p, 2, "t9");
|
|
orc_program_add_temporary (p, 1, "t10");
|
|
orc_program_add_temporary (p, 1, "t11");
|
|
orc_program_add_temporary (p, 1, "t12");
|
|
orc_program_add_temporary (p, 4, "t13");
|
|
|
|
orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_S1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S2,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T5, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S3,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T6, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T4, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T4, ORC_VAR_T4, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_T6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T9, ORC_VAR_T4, ORC_VAR_T5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T5, ORC_VAR_C5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T5, ORC_VAR_C6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T4, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T10, ORC_VAR_T7,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T11, ORC_VAR_T8,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T12, ORC_VAR_T9,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T12,
|
|
ORC_VAR_T11, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T10,
|
|
ORC_VAR_C8, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T13, ORC_VAR_T1,
|
|
ORC_VAR_T2, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T13, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|
|
|
|
|
|
/* cogorc_convert_I420_BGRA_avg */
|
|
#ifdef DISABLE_ORC
|
|
void
|
|
cogorc_convert_I420_BGRA_avg (orc_uint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4,
|
|
const orc_uint8 * ORC_RESTRICT s5, int n)
|
|
{
|
|
int i;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
const orc_int8 *ORC_RESTRICT ptr8;
|
|
orc_int8 var46;
|
|
orc_int8 var47;
|
|
orc_int8 var48;
|
|
orc_int8 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_int8 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_int8 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_int8 var65;
|
|
orc_int8 var66;
|
|
orc_int8 var67;
|
|
orc_int8 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_union16 var87;
|
|
orc_union16 var88;
|
|
orc_union16 var89;
|
|
orc_int8 var90;
|
|
orc_int8 var91;
|
|
orc_int8 var92;
|
|
orc_union16 var93;
|
|
orc_union16 var94;
|
|
orc_union32 var95;
|
|
|
|
ptr0 = (orc_union32 *) d1;
|
|
ptr4 = (orc_int8 *) s1;
|
|
ptr5 = (orc_int8 *) s2;
|
|
ptr6 = (orc_int8 *) s3;
|
|
ptr7 = (orc_int8 *) s4;
|
|
ptr8 = (orc_int8 *) s5;
|
|
|
|
/* 1: loadpb */
|
|
var47 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 7: loadpb */
|
|
var48 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 13: loadpb */
|
|
var49 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 16: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 21: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 28: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 32: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 36: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 45: loadpb */
|
|
var55 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
/* 48: loadpb */
|
|
var56.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var56.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var56.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var56.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var46 = ptr4[i];
|
|
/* 2: subb */
|
|
var58 = var46 - var47;
|
|
/* 3: convsbw */
|
|
var59.i = var58;
|
|
/* 4: loadupib */
|
|
var60 =
|
|
(i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] +
|
|
1) >> 1 : ptr5[i >> 1];
|
|
/* 5: loadupib */
|
|
var61 =
|
|
(i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] +
|
|
1) >> 1 : ptr6[i >> 1];
|
|
/* 6: avgub */
|
|
var62 = ((orc_uint8) var60 + (orc_uint8) var61 + 1) >> 1;
|
|
/* 8: subb */
|
|
var63 = var62 - var48;
|
|
/* 9: convsbw */
|
|
var64.i = var63;
|
|
/* 10: loadupib */
|
|
var65 =
|
|
(i & 1) ? ((orc_uint8) ptr7[i >> 1] + (orc_uint8) ptr7[(i >> 1) + 1] +
|
|
1) >> 1 : ptr7[i >> 1];
|
|
/* 11: loadupib */
|
|
var66 =
|
|
(i & 1) ? ((orc_uint8) ptr8[i >> 1] + (orc_uint8) ptr8[(i >> 1) + 1] +
|
|
1) >> 1 : ptr8[i >> 1];
|
|
/* 12: avgub */
|
|
var67 = ((orc_uint8) var65 + (orc_uint8) var66 + 1) >> 1;
|
|
/* 14: subb */
|
|
var68 = var67 - var49;
|
|
/* 15: convsbw */
|
|
var69.i = var68;
|
|
/* 17: mullw */
|
|
var70.i = (var59.i * var50.i) & 0xffff;
|
|
/* 18: shrsw */
|
|
var71.i = var70.i >> 8;
|
|
/* 19: addssw */
|
|
var72.i = ORC_CLAMP_SW (var59.i + var71.i);
|
|
/* 20: addssw */
|
|
var73.i = ORC_CLAMP_SW (var72.i + var69.i);
|
|
/* 22: mullw */
|
|
var74.i = (var69.i * var51.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var75.i = var74.i >> 8;
|
|
/* 24: subssw */
|
|
var76.i = ORC_CLAMP_SW (var73.i - var75.i);
|
|
/* 25: addssw */
|
|
var77.i = ORC_CLAMP_SW (var76.i + var69.i);
|
|
/* 26: addssw */
|
|
var78.i = ORC_CLAMP_SW (var72.i + var64.i);
|
|
/* 27: addssw */
|
|
var79.i = ORC_CLAMP_SW (var78.i + var64.i);
|
|
/* 29: mullw */
|
|
var80.i = (var64.i * var52.i) & 0xffff;
|
|
/* 30: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 31: addssw */
|
|
var82.i = ORC_CLAMP_SW (var79.i + var81.i);
|
|
/* 33: mullw */
|
|
var83.i = (var64.i * var53.i) & 0xffff;
|
|
/* 34: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 35: subssw */
|
|
var85.i = ORC_CLAMP_SW (var72.i - var84.i);
|
|
/* 37: mullw */
|
|
var86.i = (var69.i * var54.i) & 0xffff;
|
|
/* 38: shrsw */
|
|
var87.i = var86.i >> 8;
|
|
/* 39: subssw */
|
|
var88.i = ORC_CLAMP_SW (var85.i - var87.i);
|
|
/* 40: subssw */
|
|
var89.i = ORC_CLAMP_SW (var88.i - var87.i);
|
|
/* 41: convssswb */
|
|
var90 = ORC_CLAMP_SB (var77.i);
|
|
/* 42: convssswb */
|
|
var91 = ORC_CLAMP_SB (var89.i);
|
|
/* 43: convssswb */
|
|
var92 = ORC_CLAMP_SB (var82.i);
|
|
/* 44: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var92;
|
|
_dest.x2[1] = var91;
|
|
var93.i = _dest.i;
|
|
}
|
|
/* 46: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var90;
|
|
_dest.x2[1] = var55;
|
|
var94.i = _dest.i;
|
|
}
|
|
/* 47: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var93.i;
|
|
_dest.x2[1] = var94.i;
|
|
var95.i = _dest.i;
|
|
}
|
|
/* 49: addb */
|
|
var57.x4[0] = var95.x4[0] + var56.x4[0];
|
|
var57.x4[1] = var95.x4[1] + var56.x4[1];
|
|
var57.x4[2] = var95.x4[2] + var56.x4[2];
|
|
var57.x4[3] = var95.x4[3] + var56.x4[3];
|
|
/* 50: storel */
|
|
ptr0[i] = var57;
|
|
}
|
|
|
|
}
|
|
|
|
#else
|
|
static void
|
|
_backup_cogorc_convert_I420_BGRA_avg (OrcExecutor * ORC_RESTRICT ex)
|
|
{
|
|
int i;
|
|
int n = ex->n;
|
|
orc_union32 *ORC_RESTRICT ptr0;
|
|
const orc_int8 *ORC_RESTRICT ptr4;
|
|
const orc_int8 *ORC_RESTRICT ptr5;
|
|
const orc_int8 *ORC_RESTRICT ptr6;
|
|
const orc_int8 *ORC_RESTRICT ptr7;
|
|
const orc_int8 *ORC_RESTRICT ptr8;
|
|
orc_int8 var46;
|
|
orc_int8 var47;
|
|
orc_int8 var48;
|
|
orc_int8 var49;
|
|
orc_union16 var50;
|
|
orc_union16 var51;
|
|
orc_union16 var52;
|
|
orc_union16 var53;
|
|
orc_union16 var54;
|
|
orc_int8 var55;
|
|
orc_union32 var56;
|
|
orc_union32 var57;
|
|
orc_int8 var58;
|
|
orc_union16 var59;
|
|
orc_int8 var60;
|
|
orc_int8 var61;
|
|
orc_int8 var62;
|
|
orc_int8 var63;
|
|
orc_union16 var64;
|
|
orc_int8 var65;
|
|
orc_int8 var66;
|
|
orc_int8 var67;
|
|
orc_int8 var68;
|
|
orc_union16 var69;
|
|
orc_union16 var70;
|
|
orc_union16 var71;
|
|
orc_union16 var72;
|
|
orc_union16 var73;
|
|
orc_union16 var74;
|
|
orc_union16 var75;
|
|
orc_union16 var76;
|
|
orc_union16 var77;
|
|
orc_union16 var78;
|
|
orc_union16 var79;
|
|
orc_union16 var80;
|
|
orc_union16 var81;
|
|
orc_union16 var82;
|
|
orc_union16 var83;
|
|
orc_union16 var84;
|
|
orc_union16 var85;
|
|
orc_union16 var86;
|
|
orc_union16 var87;
|
|
orc_union16 var88;
|
|
orc_union16 var89;
|
|
orc_int8 var90;
|
|
orc_int8 var91;
|
|
orc_int8 var92;
|
|
orc_union16 var93;
|
|
orc_union16 var94;
|
|
orc_union32 var95;
|
|
|
|
ptr0 = (orc_union32 *) ex->arrays[0];
|
|
ptr4 = (orc_int8 *) ex->arrays[4];
|
|
ptr5 = (orc_int8 *) ex->arrays[5];
|
|
ptr6 = (orc_int8 *) ex->arrays[6];
|
|
ptr7 = (orc_int8 *) ex->arrays[7];
|
|
ptr8 = (orc_int8 *) ex->arrays[8];
|
|
|
|
/* 1: loadpb */
|
|
var47 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 7: loadpb */
|
|
var48 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 13: loadpb */
|
|
var49 = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
/* 16: loadpw */
|
|
var50.i = (int) 0x0000002a; /* 42 or 2.07508e-322f */
|
|
/* 21: loadpw */
|
|
var51.i = (int) 0x00000067; /* 103 or 5.08888e-322f */
|
|
/* 28: loadpw */
|
|
var52.i = (int) 0x00000004; /* 4 or 1.97626e-323f */
|
|
/* 32: loadpw */
|
|
var53.i = (int) 0x00000064; /* 100 or 4.94066e-322f */
|
|
/* 36: loadpw */
|
|
var54.i = (int) 0x00000068; /* 104 or 5.13828e-322f */
|
|
/* 45: loadpb */
|
|
var55 = (int) 0x000000ff; /* 255 or 1.25987e-321f */
|
|
/* 48: loadpb */
|
|
var56.x4[0] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var56.x4[1] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var56.x4[2] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
var56.x4[3] = (int) 0x00000080; /* 128 or 6.32404e-322f */
|
|
|
|
for (i = 0; i < n; i++) {
|
|
/* 0: loadb */
|
|
var46 = ptr4[i];
|
|
/* 2: subb */
|
|
var58 = var46 - var47;
|
|
/* 3: convsbw */
|
|
var59.i = var58;
|
|
/* 4: loadupib */
|
|
var60 =
|
|
(i & 1) ? ((orc_uint8) ptr5[i >> 1] + (orc_uint8) ptr5[(i >> 1) + 1] +
|
|
1) >> 1 : ptr5[i >> 1];
|
|
/* 5: loadupib */
|
|
var61 =
|
|
(i & 1) ? ((orc_uint8) ptr6[i >> 1] + (orc_uint8) ptr6[(i >> 1) + 1] +
|
|
1) >> 1 : ptr6[i >> 1];
|
|
/* 6: avgub */
|
|
var62 = ((orc_uint8) var60 + (orc_uint8) var61 + 1) >> 1;
|
|
/* 8: subb */
|
|
var63 = var62 - var48;
|
|
/* 9: convsbw */
|
|
var64.i = var63;
|
|
/* 10: loadupib */
|
|
var65 =
|
|
(i & 1) ? ((orc_uint8) ptr7[i >> 1] + (orc_uint8) ptr7[(i >> 1) + 1] +
|
|
1) >> 1 : ptr7[i >> 1];
|
|
/* 11: loadupib */
|
|
var66 =
|
|
(i & 1) ? ((orc_uint8) ptr8[i >> 1] + (orc_uint8) ptr8[(i >> 1) + 1] +
|
|
1) >> 1 : ptr8[i >> 1];
|
|
/* 12: avgub */
|
|
var67 = ((orc_uint8) var65 + (orc_uint8) var66 + 1) >> 1;
|
|
/* 14: subb */
|
|
var68 = var67 - var49;
|
|
/* 15: convsbw */
|
|
var69.i = var68;
|
|
/* 17: mullw */
|
|
var70.i = (var59.i * var50.i) & 0xffff;
|
|
/* 18: shrsw */
|
|
var71.i = var70.i >> 8;
|
|
/* 19: addssw */
|
|
var72.i = ORC_CLAMP_SW (var59.i + var71.i);
|
|
/* 20: addssw */
|
|
var73.i = ORC_CLAMP_SW (var72.i + var69.i);
|
|
/* 22: mullw */
|
|
var74.i = (var69.i * var51.i) & 0xffff;
|
|
/* 23: shrsw */
|
|
var75.i = var74.i >> 8;
|
|
/* 24: subssw */
|
|
var76.i = ORC_CLAMP_SW (var73.i - var75.i);
|
|
/* 25: addssw */
|
|
var77.i = ORC_CLAMP_SW (var76.i + var69.i);
|
|
/* 26: addssw */
|
|
var78.i = ORC_CLAMP_SW (var72.i + var64.i);
|
|
/* 27: addssw */
|
|
var79.i = ORC_CLAMP_SW (var78.i + var64.i);
|
|
/* 29: mullw */
|
|
var80.i = (var64.i * var52.i) & 0xffff;
|
|
/* 30: shrsw */
|
|
var81.i = var80.i >> 8;
|
|
/* 31: addssw */
|
|
var82.i = ORC_CLAMP_SW (var79.i + var81.i);
|
|
/* 33: mullw */
|
|
var83.i = (var64.i * var53.i) & 0xffff;
|
|
/* 34: shrsw */
|
|
var84.i = var83.i >> 8;
|
|
/* 35: subssw */
|
|
var85.i = ORC_CLAMP_SW (var72.i - var84.i);
|
|
/* 37: mullw */
|
|
var86.i = (var69.i * var54.i) & 0xffff;
|
|
/* 38: shrsw */
|
|
var87.i = var86.i >> 8;
|
|
/* 39: subssw */
|
|
var88.i = ORC_CLAMP_SW (var85.i - var87.i);
|
|
/* 40: subssw */
|
|
var89.i = ORC_CLAMP_SW (var88.i - var87.i);
|
|
/* 41: convssswb */
|
|
var90 = ORC_CLAMP_SB (var77.i);
|
|
/* 42: convssswb */
|
|
var91 = ORC_CLAMP_SB (var89.i);
|
|
/* 43: convssswb */
|
|
var92 = ORC_CLAMP_SB (var82.i);
|
|
/* 44: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var92;
|
|
_dest.x2[1] = var91;
|
|
var93.i = _dest.i;
|
|
}
|
|
/* 46: mergebw */
|
|
{
|
|
orc_union16 _dest;
|
|
_dest.x2[0] = var90;
|
|
_dest.x2[1] = var55;
|
|
var94.i = _dest.i;
|
|
}
|
|
/* 47: mergewl */
|
|
{
|
|
orc_union32 _dest;
|
|
_dest.x2[0] = var93.i;
|
|
_dest.x2[1] = var94.i;
|
|
var95.i = _dest.i;
|
|
}
|
|
/* 49: addb */
|
|
var57.x4[0] = var95.x4[0] + var56.x4[0];
|
|
var57.x4[1] = var95.x4[1] + var56.x4[1];
|
|
var57.x4[2] = var95.x4[2] + var56.x4[2];
|
|
var57.x4[3] = var95.x4[3] + var56.x4[3];
|
|
/* 50: storel */
|
|
ptr0[i] = var57;
|
|
}
|
|
|
|
}
|
|
|
|
void
|
|
cogorc_convert_I420_BGRA_avg (orc_uint32 * ORC_RESTRICT d1,
|
|
const orc_uint8 * ORC_RESTRICT s1, const orc_uint8 * ORC_RESTRICT s2,
|
|
const orc_uint8 * ORC_RESTRICT s3, const orc_uint8 * ORC_RESTRICT s4,
|
|
const orc_uint8 * ORC_RESTRICT s5, int n)
|
|
{
|
|
OrcExecutor _ex, *ex = &_ex;
|
|
static int p_inited = 0;
|
|
static OrcProgram *p = 0;
|
|
void (*func) (OrcExecutor *);
|
|
|
|
if (!p_inited) {
|
|
orc_once_mutex_lock ();
|
|
if (!p_inited) {
|
|
|
|
p = orc_program_new ();
|
|
orc_program_set_name (p, "cogorc_convert_I420_BGRA_avg");
|
|
orc_program_set_backup_function (p, _backup_cogorc_convert_I420_BGRA_avg);
|
|
orc_program_add_destination (p, 4, "d1");
|
|
orc_program_add_source (p, 1, "s1");
|
|
orc_program_add_source (p, 1, "s2");
|
|
orc_program_add_source (p, 1, "s3");
|
|
orc_program_add_source (p, 1, "s4");
|
|
orc_program_add_source (p, 1, "s5");
|
|
orc_program_add_constant (p, 1, 0x00000008, "c1");
|
|
orc_program_add_constant (p, 1, 0x00000080, "c2");
|
|
orc_program_add_constant (p, 4, 0x0000002a, "c3");
|
|
orc_program_add_constant (p, 4, 0x00000067, "c4");
|
|
orc_program_add_constant (p, 4, 0x00000004, "c5");
|
|
orc_program_add_constant (p, 4, 0x00000064, "c6");
|
|
orc_program_add_constant (p, 4, 0x00000068, "c7");
|
|
orc_program_add_constant (p, 4, 0x000000ff, "c8");
|
|
orc_program_add_temporary (p, 2, "t1");
|
|
orc_program_add_temporary (p, 2, "t2");
|
|
orc_program_add_temporary (p, 1, "t3");
|
|
orc_program_add_temporary (p, 1, "t4");
|
|
orc_program_add_temporary (p, 2, "t5");
|
|
orc_program_add_temporary (p, 2, "t6");
|
|
orc_program_add_temporary (p, 2, "t7");
|
|
orc_program_add_temporary (p, 2, "t8");
|
|
orc_program_add_temporary (p, 2, "t9");
|
|
orc_program_add_temporary (p, 2, "t10");
|
|
orc_program_add_temporary (p, 1, "t11");
|
|
orc_program_add_temporary (p, 1, "t12");
|
|
orc_program_add_temporary (p, 1, "t13");
|
|
orc_program_add_temporary (p, 4, "t14");
|
|
|
|
orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_S1, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T5, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S2,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T4, ORC_VAR_S3,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T6, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T3, ORC_VAR_S4,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "loadupib", 0, ORC_VAR_T4, ORC_VAR_S5,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "avgub", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_T4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subb", 0, ORC_VAR_T3, ORC_VAR_T3, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convsbw", 0, ORC_VAR_T7, ORC_VAR_T3, ORC_VAR_D1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T5, ORC_VAR_C3,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_T7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C4,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_T7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T5, ORC_VAR_T6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T6, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C5,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addssw", 0, ORC_VAR_T10, ORC_VAR_T10,
|
|
ORC_VAR_T1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_C6,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T9, ORC_VAR_T5, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_C7,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "shrsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "subssw", 0, ORC_VAR_T9, ORC_VAR_T9, ORC_VAR_T1,
|
|
ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T11, ORC_VAR_T8,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T12, ORC_VAR_T9,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T13, ORC_VAR_T10,
|
|
ORC_VAR_D1, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T1, ORC_VAR_T13,
|
|
ORC_VAR_T12, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T2, ORC_VAR_T11,
|
|
ORC_VAR_C8, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T14, ORC_VAR_T1,
|
|
ORC_VAR_T2, ORC_VAR_D1);
|
|
orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T14, ORC_VAR_C2,
|
|
ORC_VAR_D1);
|
|
|
|
orc_program_compile (p);
|
|
}
|
|
p_inited = TRUE;
|
|
orc_once_mutex_unlock ();
|
|
}
|
|
ex->program = p;
|
|
|
|
ex->n = n;
|
|
ex->arrays[ORC_VAR_D1] = d1;
|
|
ex->arrays[ORC_VAR_S1] = (void *) s1;
|
|
ex->arrays[ORC_VAR_S2] = (void *) s2;
|
|
ex->arrays[ORC_VAR_S3] = (void *) s3;
|
|
ex->arrays[ORC_VAR_S4] = (void *) s4;
|
|
ex->arrays[ORC_VAR_S5] = (void *) s5;
|
|
|
|
func = p->code_exec;
|
|
func (ex);
|
|
}
|
|
#endif
|