mirrors/gstreamer
mirrors/gstreamer
1
1
Fork 0
mirror of https://gitlab.freedesktop.org/gstreamer/gstreamer.git synced 2024-12-22 16:26:39 +00:00
Code Issues 1 Projects Releases Wiki Activity
gstreamer/gst/videomixer/blendorc-dist.c
David Schleef e9d4710cc8 Update generated orc code
2010-10-15 14:02:57 -07:00

1147 lines
34 KiB
C
Raw Blame History

/* autogenerated from blendorc.orc */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#ifndef DISABLE_ORC
#include <orc/orc.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)
#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
void orc_splat_u32 (guint32 * d1, int p1, int n);
void orc_memcpy_u32 (guint32 * d1, const guint32 * s1, int n);
void orc_blend_u8 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, int n, int m);
void orc_blend_argb (guint8 * d1, int d1_stride, const guint8 * s1,
int s1_stride, int p1, int n, int m);
void orc_blend_bgra (guint8 * d1, int d1_stride, const guint8 * s1,
int s1_stride, int p1, int n, int m);
/* 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))
#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
/* end Orc C target preamble */
/* orc_splat_u32 */
#ifdef DISABLE_ORC
void
orc_splat_u32 (guint32 * d1, int p1, int n)
{
int i;
orc_union32 *ORC_RESTRICT ptr0;
orc_union32 var32;
orc_union32 var33;
ptr0 = (orc_union32 *) d1;
/* 0: loadpl */
var32.i = p1;
for (i = 0; i < n; i++) {
/* 1: copyl */
var33.i = var32.i;
/* 2: storel */
ptr0[i] = var33;
}
}
#else
static void
_backup_orc_splat_u32 (OrcExecutor * ORC_RESTRICT ex)
{
int i;
int n = ex->n;
orc_union32 *ORC_RESTRICT ptr0;
orc_union32 var32;
orc_union32 var33;
ptr0 = (orc_union32 *) ex->arrays[0];
/* 0: loadpl */
var32.i = ex->params[24];
for (i = 0; i < n; i++) {
/* 1: copyl */
var33.i = var32.i;
/* 2: storel */
ptr0[i] = var33;
}
}
void
orc_splat_u32 (guint32 * 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) {
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_name (p, "orc_splat_u32");
orc_program_set_backup_function (p, _backup_orc_splat_u32);
orc_program_add_destination (p, 4, "d1");
orc_program_add_parameter (p, 4, "p1");
orc_program_append_2 (p, "copyl", 0, ORC_VAR_D1, ORC_VAR_P1, ORC_VAR_D1,
ORC_VAR_D1);
result = 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_memcpy_u32 */
#ifdef DISABLE_ORC
void
orc_memcpy_u32 (guint32 * d1, const guint32 * s1, int n)
{
int i;
orc_union32 *ORC_RESTRICT ptr0;
const orc_union32 *ORC_RESTRICT ptr4;
orc_union32 var32;
orc_union32 var33;
ptr0 = (orc_union32 *) d1;
ptr4 = (orc_union32 *) s1;
for (i = 0; i < n; i++) {
/* 0: loadl */
var32 = ptr4[i];
/* 1: copyl */
var33.i = var32.i;
/* 2: storel */
ptr0[i] = var33;
}
}
#else
static void
_backup_orc_memcpy_u32 (OrcExecutor * ORC_RESTRICT ex)
{
int i;
int n = ex->n;
orc_union32 *ORC_RESTRICT ptr0;
const orc_union32 *ORC_RESTRICT ptr4;
orc_union32 var32;
orc_union32 var33;
ptr0 = (orc_union32 *) ex->arrays[0];
ptr4 = (orc_union32 *) ex->arrays[4];
for (i = 0; i < n; i++) {
/* 0: loadl */
var32 = ptr4[i];
/* 1: copyl */
var33.i = var32.i;
/* 2: storel */
ptr0[i] = var33;
}
}
void
orc_memcpy_u32 (guint32 * d1, const guint32 * 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) {
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_name (p, "orc_memcpy_u32");
orc_program_set_backup_function (p, _backup_orc_memcpy_u32);
orc_program_add_destination (p, 4, "d1");
orc_program_add_source (p, 4, "s1");
orc_program_append_2 (p, "copyl", 0, ORC_VAR_D1, ORC_VAR_S1, ORC_VAR_D1,
ORC_VAR_D1);
result = 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_blend_u8 */
#ifdef DISABLE_ORC
void
orc_blend_u8 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, int n, int m)
{
int i;
int j;
orc_int8 *ORC_RESTRICT ptr0;
const orc_int8 *ORC_RESTRICT ptr4;
orc_int8 var34;
orc_int8 var35;
orc_union16 var36;
orc_int8 var37;
orc_union16 var38;
orc_union16 var39;
orc_union16 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);
/* 5: loadpw */
var36.i = p1;
for (i = 0; i < n; i++) {
/* 0: loadb */
var34 = ptr0[i];
/* 1: convubw */
var38.i = (orc_uint8) var34;
/* 2: loadb */
var35 = ptr4[i];
/* 3: convubw */
var39.i = (orc_uint8) var35;
/* 4: subw */
var40.i = var39.i - var38.i;
/* 6: mullw */
var41.i = (var40.i * var36.i) & 0xffff;
/* 7: shlw */
var42.i = var38.i << 8;
/* 8: addw */
var43.i = var42.i + var41.i;
/* 9: shruw */
var44.i = ((orc_uint16) var43.i) >> 8;
/* 10: convsuswb */
var37 = ORC_CLAMP_UB (var44.i);
/* 11: storeb */
ptr0[i] = var37;
}
}
}
#else
static void
_backup_orc_blend_u8 (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 var34;
orc_int8 var35;
orc_union16 var36;
orc_int8 var37;
orc_union16 var38;
orc_union16 var39;
orc_union16 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);
/* 5: loadpw */
var36.i = ex->params[24];
for (i = 0; i < n; i++) {
/* 0: loadb */
var34 = ptr0[i];
/* 1: convubw */
var38.i = (orc_uint8) var34;
/* 2: loadb */
var35 = ptr4[i];
/* 3: convubw */
var39.i = (orc_uint8) var35;
/* 4: subw */
var40.i = var39.i - var38.i;
/* 6: mullw */
var41.i = (var40.i * var36.i) & 0xffff;
/* 7: shlw */
var42.i = var38.i << 8;
/* 8: addw */
var43.i = var42.i + var41.i;
/* 9: shruw */
var44.i = ((orc_uint16) var43.i) >> 8;
/* 10: convsuswb */
var37 = ORC_CLAMP_UB (var44.i);
/* 11: storeb */
ptr0[i] = var37;
}
}
}
void
orc_blend_u8 (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, 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) {
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_2d (p);
orc_program_set_name (p, "orc_blend_u8");
orc_program_set_backup_function (p, _backup_orc_blend_u8);
orc_program_add_destination (p, 1, "d1");
orc_program_add_source (p, 1, "s1");
orc_program_add_constant (p, 1, 0x00000008, "c1");
orc_program_add_parameter (p, 2, "p1");
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_D1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convubw", 0, ORC_VAR_T2, ORC_VAR_S1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "subw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_T1,
ORC_VAR_D1);
orc_program_append_2 (p, "mullw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_P1,
ORC_VAR_D1);
orc_program_append_2 (p, "shlw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
ORC_VAR_D1);
orc_program_append_2 (p, "addw", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_T2,
ORC_VAR_D1);
orc_program_append_2 (p, "shruw", 0, ORC_VAR_T2, ORC_VAR_T2, ORC_VAR_C1,
ORC_VAR_D1);
orc_program_append_2 (p, "convsuswb", 0, ORC_VAR_D1, ORC_VAR_T2,
ORC_VAR_D1, ORC_VAR_D1);
result = 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->params[ORC_VAR_P1] = p1;
func = p->code_exec;
func (ex);
}
#endif
/* orc_blend_argb */
#ifdef DISABLE_ORC
void
orc_blend_argb (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, int n, int m)
{
int i;
int j;
orc_union32 *ORC_RESTRICT ptr0;
const orc_union32 *ORC_RESTRICT ptr4;
orc_union64 var39;
orc_union32 var40;
orc_union32 var41;
orc_union16 var42;
orc_int8 var43;
orc_union32 var44;
orc_union64 var45;
orc_union64 var46;
orc_union64 var47;
orc_union64 var48;
orc_union32 var49;
orc_union64 var50;
orc_union64 var51;
orc_union64 var52;
orc_union64 var53;
orc_union64 var54;
orc_union32 var55;
orc_union32 var56;
for (j = 0; j < m; j++) {
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
/* 5: loadpw */
var39.x4[0] = p1;
var39.x4[1] = p1;
var39.x4[2] = p1;
var39.x4[3] = p1;
/* 16: loadpl */
var40.i = 0x000000ff; /* 255 or 1.25987e-321f */
for (i = 0; i < n; i++) {
/* 0: loadl */
var41 = ptr4[i];
/* 1: convlw */
var42.i = var41.i;
/* 2: convwb */
var43 = var42.i;
/* 3: splatbl */
var44.i =
((var43 & 0xff) << 24) | ((var43 & 0xff) << 16) | ((var43 & 0xff) <<
8) | (var43 & 0xff);
/* 4: convubw */
var45.x4[0] = (orc_uint8) var44.x4[0];
var45.x4[1] = (orc_uint8) var44.x4[1];
var45.x4[2] = (orc_uint8) var44.x4[2];
var45.x4[3] = (orc_uint8) var44.x4[3];
/* 6: mullw */
var46.x4[0] = (var45.x4[0] * var39.x4[0]) & 0xffff;
var46.x4[1] = (var45.x4[1] * var39.x4[1]) & 0xffff;
var46.x4[2] = (var45.x4[2] * var39.x4[2]) & 0xffff;
var46.x4[3] = (var45.x4[3] * var39.x4[3]) & 0xffff;
/* 7: shruw */
var47.x4[0] = ((orc_uint16) var46.x4[0]) >> 8;
var47.x4[1] = ((orc_uint16) var46.x4[1]) >> 8;
var47.x4[2] = ((orc_uint16) var46.x4[2]) >> 8;
var47.x4[3] = ((orc_uint16) var46.x4[3]) >> 8;
/* 8: convubw */
var48.x4[0] = (orc_uint8) var41.x4[0];
var48.x4[1] = (orc_uint8) var41.x4[1];
var48.x4[2] = (orc_uint8) var41.x4[2];
var48.x4[3] = (orc_uint8) var41.x4[3];
/* 9: loadl */
var49 = ptr0[i];
/* 10: convubw */
var50.x4[0] = (orc_uint8) var49.x4[0];
var50.x4[1] = (orc_uint8) var49.x4[1];
var50.x4[2] = (orc_uint8) var49.x4[2];
var50.x4[3] = (orc_uint8) var49.x4[3];
/* 11: subw */
var51.x4[0] = var48.x4[0] - var50.x4[0];
var51.x4[1] = var48.x4[1] - var50.x4[1];
var51.x4[2] = var48.x4[2] - var50.x4[2];
var51.x4[3] = var48.x4[3] - var50.x4[3];
/* 12: mullw */
var52.x4[0] = (var51.x4[0] * var47.x4[0]) & 0xffff;
var52.x4[1] = (var51.x4[1] * var47.x4[1]) & 0xffff;
var52.x4[2] = (var51.x4[2] * var47.x4[2]) & 0xffff;
var52.x4[3] = (var51.x4[3] * var47.x4[3]) & 0xffff;
/* 13: div255w */
var53.x4[0] =
((uint16_t) (((orc_uint16) (var52.x4[0] + 128)) +
(((orc_uint16) (var52.x4[0] + 128)) >> 8))) >> 8;
var53.x4[1] =
((uint16_t) (((orc_uint16) (var52.x4[1] + 128)) +
(((orc_uint16) (var52.x4[1] + 128)) >> 8))) >> 8;
var53.x4[2] =
((uint16_t) (((orc_uint16) (var52.x4[2] + 128)) +
(((orc_uint16) (var52.x4[2] + 128)) >> 8))) >> 8;
var53.x4[3] =
((uint16_t) (((orc_uint16) (var52.x4[3] + 128)) +
(((orc_uint16) (var52.x4[3] + 128)) >> 8))) >> 8;
/* 14: addw */
var54.x4[0] = var50.x4[0] + var53.x4[0];
var54.x4[1] = var50.x4[1] + var53.x4[1];
var54.x4[2] = var50.x4[2] + var53.x4[2];
var54.x4[3] = var50.x4[3] + var53.x4[3];
/* 15: convwb */
var55.x4[0] = var54.x4[0];
var55.x4[1] = var54.x4[1];
var55.x4[2] = var54.x4[2];
var55.x4[3] = var54.x4[3];
/* 17: orl */
var56.i = var55.i | var40.i;
/* 18: storel */
ptr0[i] = var56;
}
}
}
#else
static void
_backup_orc_blend_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_union64 var39;
orc_union32 var40;
orc_union32 var41;
orc_union16 var42;
orc_int8 var43;
orc_union32 var44;
orc_union64 var45;
orc_union64 var46;
orc_union64 var47;
orc_union64 var48;
orc_union32 var49;
orc_union64 var50;
orc_union64 var51;
orc_union64 var52;
orc_union64 var53;
orc_union64 var54;
orc_union32 var55;
orc_union32 var56;
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);
/* 5: loadpw */
var39.x4[0] = ex->params[24];
var39.x4[1] = ex->params[24];
var39.x4[2] = ex->params[24];
var39.x4[3] = ex->params[24];
/* 16: loadpl */
var40.i = 0x000000ff; /* 255 or 1.25987e-321f */
for (i = 0; i < n; i++) {
/* 0: loadl */
var41 = ptr4[i];
/* 1: convlw */
var42.i = var41.i;
/* 2: convwb */
var43 = var42.i;
/* 3: splatbl */
var44.i =
((var43 & 0xff) << 24) | ((var43 & 0xff) << 16) | ((var43 & 0xff) <<
8) | (var43 & 0xff);
/* 4: convubw */
var45.x4[0] = (orc_uint8) var44.x4[0];
var45.x4[1] = (orc_uint8) var44.x4[1];
var45.x4[2] = (orc_uint8) var44.x4[2];
var45.x4[3] = (orc_uint8) var44.x4[3];
/* 6: mullw */
var46.x4[0] = (var45.x4[0] * var39.x4[0]) & 0xffff;
var46.x4[1] = (var45.x4[1] * var39.x4[1]) & 0xffff;
var46.x4[2] = (var45.x4[2] * var39.x4[2]) & 0xffff;
var46.x4[3] = (var45.x4[3] * var39.x4[3]) & 0xffff;
/* 7: shruw */
var47.x4[0] = ((orc_uint16) var46.x4[0]) >> 8;
var47.x4[1] = ((orc_uint16) var46.x4[1]) >> 8;
var47.x4[2] = ((orc_uint16) var46.x4[2]) >> 8;
var47.x4[3] = ((orc_uint16) var46.x4[3]) >> 8;
/* 8: convubw */
var48.x4[0] = (orc_uint8) var41.x4[0];
var48.x4[1] = (orc_uint8) var41.x4[1];
var48.x4[2] = (orc_uint8) var41.x4[2];
var48.x4[3] = (orc_uint8) var41.x4[3];
/* 9: loadl */
var49 = ptr0[i];
/* 10: convubw */
var50.x4[0] = (orc_uint8) var49.x4[0];
var50.x4[1] = (orc_uint8) var49.x4[1];
var50.x4[2] = (orc_uint8) var49.x4[2];
var50.x4[3] = (orc_uint8) var49.x4[3];
/* 11: subw */
var51.x4[0] = var48.x4[0] - var50.x4[0];
var51.x4[1] = var48.x4[1] - var50.x4[1];
var51.x4[2] = var48.x4[2] - var50.x4[2];
var51.x4[3] = var48.x4[3] - var50.x4[3];
/* 12: mullw */
var52.x4[0] = (var51.x4[0] * var47.x4[0]) & 0xffff;
var52.x4[1] = (var51.x4[1] * var47.x4[1]) & 0xffff;
var52.x4[2] = (var51.x4[2] * var47.x4[2]) & 0xffff;
var52.x4[3] = (var51.x4[3] * var47.x4[3]) & 0xffff;
/* 13: div255w */
var53.x4[0] =
((uint16_t) (((orc_uint16) (var52.x4[0] + 128)) +
(((orc_uint16) (var52.x4[0] + 128)) >> 8))) >> 8;
var53.x4[1] =
((uint16_t) (((orc_uint16) (var52.x4[1] + 128)) +
(((orc_uint16) (var52.x4[1] + 128)) >> 8))) >> 8;
var53.x4[2] =
((uint16_t) (((orc_uint16) (var52.x4[2] + 128)) +
(((orc_uint16) (var52.x4[2] + 128)) >> 8))) >> 8;
var53.x4[3] =
((uint16_t) (((orc_uint16) (var52.x4[3] + 128)) +
(((orc_uint16) (var52.x4[3] + 128)) >> 8))) >> 8;
/* 14: addw */
var54.x4[0] = var50.x4[0] + var53.x4[0];
var54.x4[1] = var50.x4[1] + var53.x4[1];
var54.x4[2] = var50.x4[2] + var53.x4[2];
var54.x4[3] = var50.x4[3] + var53.x4[3];
/* 15: convwb */
var55.x4[0] = var54.x4[0];
var55.x4[1] = var54.x4[1];
var55.x4[2] = var54.x4[2];
var55.x4[3] = var54.x4[3];
/* 17: orl */
var56.i = var55.i | var40.i;
/* 18: storel */
ptr0[i] = var56;
}
}
}
void
orc_blend_argb (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, 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) {
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_2d (p);
orc_program_set_name (p, "orc_blend_argb");
orc_program_set_backup_function (p, _backup_orc_blend_argb);
orc_program_add_destination (p, 4, "d1");
orc_program_add_source (p, 4, "s1");
orc_program_add_constant (p, 4, 0x000000ff, "c1");
orc_program_add_constant (p, 4, 0x00000008, "c2");
orc_program_add_parameter (p, 2, "p1");
orc_program_add_temporary (p, 4, "t1");
orc_program_add_temporary (p, 2, "t2");
orc_program_add_temporary (p, 1, "t3");
orc_program_add_temporary (p, 4, "t4");
orc_program_add_temporary (p, 8, "t5");
orc_program_add_temporary (p, 8, "t6");
orc_program_add_temporary (p, 8, "t7");
orc_program_append_2 (p, "loadl", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convlw", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convwb", 0, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "splatbl", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convubw", 2, ORC_VAR_T7, ORC_VAR_T4, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "mullw", 2, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_P1,
ORC_VAR_D1);
orc_program_append_2 (p, "shruw", 2, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_C2,
ORC_VAR_D1);
orc_program_append_2 (p, "convubw", 2, ORC_VAR_T6, ORC_VAR_T1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "loadl", 0, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convubw", 2, ORC_VAR_T5, ORC_VAR_T1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "subw", 2, ORC_VAR_T6, ORC_VAR_T6, ORC_VAR_T5,
ORC_VAR_D1);
orc_program_append_2 (p, "mullw", 2, ORC_VAR_T6, ORC_VAR_T6, ORC_VAR_T7,
ORC_VAR_D1);
orc_program_append_2 (p, "div255w", 2, ORC_VAR_T6, ORC_VAR_T6, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "addw", 2, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_T6,
ORC_VAR_D1);
orc_program_append_2 (p, "convwb", 2, ORC_VAR_T1, ORC_VAR_T5, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "orl", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
ORC_VAR_D1);
orc_program_append_2 (p, "storel", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1,
ORC_VAR_D1);
result = 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->params[ORC_VAR_P1] = p1;
func = p->code_exec;
func (ex);
}
#endif
/* orc_blend_bgra */
#ifdef DISABLE_ORC
void
orc_blend_bgra (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, int n, int m)
{
int i;
int j;
orc_union32 *ORC_RESTRICT ptr0;
const orc_union32 *ORC_RESTRICT ptr4;
orc_union64 var40;
orc_union32 var41;
orc_union32 var42;
orc_union32 var43;
orc_union16 var44;
orc_int8 var45;
orc_union32 var46;
orc_union64 var47;
orc_union64 var48;
orc_union64 var49;
orc_union64 var50;
orc_union32 var51;
orc_union64 var52;
orc_union64 var53;
orc_union64 var54;
orc_union64 var55;
orc_union64 var56;
orc_union32 var57;
orc_union32 var58;
for (j = 0; j < m; j++) {
ptr0 = ORC_PTR_OFFSET (d1, d1_stride * j);
ptr4 = ORC_PTR_OFFSET (s1, s1_stride * j);
/* 6: loadpw */
var40.x4[0] = p1;
var40.x4[1] = p1;
var40.x4[2] = p1;
var40.x4[3] = p1;
/* 17: loadpl */
var41.i = 0xff000000; /* -16777216 or 2.11371e-314f */
for (i = 0; i < n; i++) {
/* 0: loadl */
var42 = ptr4[i];
/* 1: shrul */
var43.i = ((orc_uint32) var42.i) >> 24;
/* 2: convlw */
var44.i = var43.i;
/* 3: convwb */
var45 = var44.i;
/* 4: splatbl */
var46.i =
((var45 & 0xff) << 24) | ((var45 & 0xff) << 16) | ((var45 & 0xff) <<
8) | (var45 & 0xff);
/* 5: convubw */
var47.x4[0] = (orc_uint8) var46.x4[0];
var47.x4[1] = (orc_uint8) var46.x4[1];
var47.x4[2] = (orc_uint8) var46.x4[2];
var47.x4[3] = (orc_uint8) var46.x4[3];
/* 7: mullw */
var48.x4[0] = (var47.x4[0] * var40.x4[0]) & 0xffff;
var48.x4[1] = (var47.x4[1] * var40.x4[1]) & 0xffff;
var48.x4[2] = (var47.x4[2] * var40.x4[2]) & 0xffff;
var48.x4[3] = (var47.x4[3] * var40.x4[3]) & 0xffff;
/* 8: shruw */
var49.x4[0] = ((orc_uint16) var48.x4[0]) >> 8;
var49.x4[1] = ((orc_uint16) var48.x4[1]) >> 8;
var49.x4[2] = ((orc_uint16) var48.x4[2]) >> 8;
var49.x4[3] = ((orc_uint16) var48.x4[3]) >> 8;
/* 9: convubw */
var50.x4[0] = (orc_uint8) var42.x4[0];
var50.x4[1] = (orc_uint8) var42.x4[1];
var50.x4[2] = (orc_uint8) var42.x4[2];
var50.x4[3] = (orc_uint8) var42.x4[3];
/* 10: loadl */
var51 = ptr0[i];
/* 11: convubw */
var52.x4[0] = (orc_uint8) var51.x4[0];
var52.x4[1] = (orc_uint8) var51.x4[1];
var52.x4[2] = (orc_uint8) var51.x4[2];
var52.x4[3] = (orc_uint8) var51.x4[3];
/* 12: subw */
var53.x4[0] = var50.x4[0] - var52.x4[0];
var53.x4[1] = var50.x4[1] - var52.x4[1];
var53.x4[2] = var50.x4[2] - var52.x4[2];
var53.x4[3] = var50.x4[3] - var52.x4[3];
/* 13: mullw */
var54.x4[0] = (var53.x4[0] * var49.x4[0]) & 0xffff;
var54.x4[1] = (var53.x4[1] * var49.x4[1]) & 0xffff;
var54.x4[2] = (var53.x4[2] * var49.x4[2]) & 0xffff;
var54.x4[3] = (var53.x4[3] * var49.x4[3]) & 0xffff;
/* 14: div255w */
var55.x4[0] =
((uint16_t) (((orc_uint16) (var54.x4[0] + 128)) +
(((orc_uint16) (var54.x4[0] + 128)) >> 8))) >> 8;
var55.x4[1] =
((uint16_t) (((orc_uint16) (var54.x4[1] + 128)) +
(((orc_uint16) (var54.x4[1] + 128)) >> 8))) >> 8;
var55.x4[2] =
((uint16_t) (((orc_uint16) (var54.x4[2] + 128)) +
(((orc_uint16) (var54.x4[2] + 128)) >> 8))) >> 8;
var55.x4[3] =
((uint16_t) (((orc_uint16) (var54.x4[3] + 128)) +
(((orc_uint16) (var54.x4[3] + 128)) >> 8))) >> 8;
/* 15: addw */
var56.x4[0] = var52.x4[0] + var55.x4[0];
var56.x4[1] = var52.x4[1] + var55.x4[1];
var56.x4[2] = var52.x4[2] + var55.x4[2];
var56.x4[3] = var52.x4[3] + var55.x4[3];
/* 16: convwb */
var57.x4[0] = var56.x4[0];
var57.x4[1] = var56.x4[1];
var57.x4[2] = var56.x4[2];
var57.x4[3] = var56.x4[3];
/* 18: orl */
var58.i = var57.i | var41.i;
/* 19: storel */
ptr0[i] = var58;
}
}
}
#else
static void
_backup_orc_blend_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_union64 var40;
orc_union32 var41;
orc_union32 var42;
orc_union32 var43;
orc_union16 var44;
orc_int8 var45;
orc_union32 var46;
orc_union64 var47;
orc_union64 var48;
orc_union64 var49;
orc_union64 var50;
orc_union32 var51;
orc_union64 var52;
orc_union64 var53;
orc_union64 var54;
orc_union64 var55;
orc_union64 var56;
orc_union32 var57;
orc_union32 var58;
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);
/* 6: loadpw */
var40.x4[0] = ex->params[24];
var40.x4[1] = ex->params[24];
var40.x4[2] = ex->params[24];
var40.x4[3] = ex->params[24];
/* 17: loadpl */
var41.i = 0xff000000; /* -16777216 or 2.11371e-314f */
for (i = 0; i < n; i++) {
/* 0: loadl */
var42 = ptr4[i];
/* 1: shrul */
var43.i = ((orc_uint32) var42.i) >> 24;
/* 2: convlw */
var44.i = var43.i;
/* 3: convwb */
var45 = var44.i;
/* 4: splatbl */
var46.i =
((var45 & 0xff) << 24) | ((var45 & 0xff) << 16) | ((var45 & 0xff) <<
8) | (var45 & 0xff);
/* 5: convubw */
var47.x4[0] = (orc_uint8) var46.x4[0];
var47.x4[1] = (orc_uint8) var46.x4[1];
var47.x4[2] = (orc_uint8) var46.x4[2];
var47.x4[3] = (orc_uint8) var46.x4[3];
/* 7: mullw */
var48.x4[0] = (var47.x4[0] * var40.x4[0]) & 0xffff;
var48.x4[1] = (var47.x4[1] * var40.x4[1]) & 0xffff;
var48.x4[2] = (var47.x4[2] * var40.x4[2]) & 0xffff;
var48.x4[3] = (var47.x4[3] * var40.x4[3]) & 0xffff;
/* 8: shruw */
var49.x4[0] = ((orc_uint16) var48.x4[0]) >> 8;
var49.x4[1] = ((orc_uint16) var48.x4[1]) >> 8;
var49.x4[2] = ((orc_uint16) var48.x4[2]) >> 8;
var49.x4[3] = ((orc_uint16) var48.x4[3]) >> 8;
/* 9: convubw */
var50.x4[0] = (orc_uint8) var42.x4[0];
var50.x4[1] = (orc_uint8) var42.x4[1];
var50.x4[2] = (orc_uint8) var42.x4[2];
var50.x4[3] = (orc_uint8) var42.x4[3];
/* 10: loadl */
var51 = ptr0[i];
/* 11: convubw */
var52.x4[0] = (orc_uint8) var51.x4[0];
var52.x4[1] = (orc_uint8) var51.x4[1];
var52.x4[2] = (orc_uint8) var51.x4[2];
var52.x4[3] = (orc_uint8) var51.x4[3];
/* 12: subw */
var53.x4[0] = var50.x4[0] - var52.x4[0];
var53.x4[1] = var50.x4[1] - var52.x4[1];
var53.x4[2] = var50.x4[2] - var52.x4[2];
var53.x4[3] = var50.x4[3] - var52.x4[3];
/* 13: mullw */
var54.x4[0] = (var53.x4[0] * var49.x4[0]) & 0xffff;
var54.x4[1] = (var53.x4[1] * var49.x4[1]) & 0xffff;
var54.x4[2] = (var53.x4[2] * var49.x4[2]) & 0xffff;
var54.x4[3] = (var53.x4[3] * var49.x4[3]) & 0xffff;
/* 14: div255w */
var55.x4[0] =
((uint16_t) (((orc_uint16) (var54.x4[0] + 128)) +
(((orc_uint16) (var54.x4[0] + 128)) >> 8))) >> 8;
var55.x4[1] =
((uint16_t) (((orc_uint16) (var54.x4[1] + 128)) +
(((orc_uint16) (var54.x4[1] + 128)) >> 8))) >> 8;
var55.x4[2] =
((uint16_t) (((orc_uint16) (var54.x4[2] + 128)) +
(((orc_uint16) (var54.x4[2] + 128)) >> 8))) >> 8;
var55.x4[3] =
((uint16_t) (((orc_uint16) (var54.x4[3] + 128)) +
(((orc_uint16) (var54.x4[3] + 128)) >> 8))) >> 8;
/* 15: addw */
var56.x4[0] = var52.x4[0] + var55.x4[0];
var56.x4[1] = var52.x4[1] + var55.x4[1];
var56.x4[2] = var52.x4[2] + var55.x4[2];
var56.x4[3] = var52.x4[3] + var55.x4[3];
/* 16: convwb */
var57.x4[0] = var56.x4[0];
var57.x4[1] = var56.x4[1];
var57.x4[2] = var56.x4[2];
var57.x4[3] = var56.x4[3];
/* 18: orl */
var58.i = var57.i | var41.i;
/* 19: storel */
ptr0[i] = var58;
}
}
}
void
orc_blend_bgra (guint8 * d1, int d1_stride, const guint8 * s1, int s1_stride,
int p1, 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) {
OrcCompileResult result;
p = orc_program_new ();
orc_program_set_2d (p);
orc_program_set_name (p, "orc_blend_bgra");
orc_program_set_backup_function (p, _backup_orc_blend_bgra);
orc_program_add_destination (p, 4, "d1");
orc_program_add_source (p, 4, "s1");
orc_program_add_constant (p, 4, 0xff000000, "c1");
orc_program_add_constant (p, 4, 0x00000018, "c2");
orc_program_add_constant (p, 4, 0x00000008, "c3");
orc_program_add_parameter (p, 2, "p1");
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, 1, "t4");
orc_program_add_temporary (p, 4, "t5");
orc_program_add_temporary (p, 8, "t6");
orc_program_add_temporary (p, 8, "t7");
orc_program_add_temporary (p, 8, "t8");
orc_program_append_2 (p, "loadl", 0, ORC_VAR_T1, ORC_VAR_S1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "shrul", 0, ORC_VAR_T2, ORC_VAR_T1, ORC_VAR_C2,
ORC_VAR_D1);
orc_program_append_2 (p, "convlw", 0, ORC_VAR_T3, ORC_VAR_T2, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convwb", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "splatbl", 0, ORC_VAR_T5, ORC_VAR_T4, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convubw", 2, ORC_VAR_T8, ORC_VAR_T5, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "mullw", 2, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_P1,
ORC_VAR_D1);
orc_program_append_2 (p, "shruw", 2, ORC_VAR_T8, ORC_VAR_T8, ORC_VAR_C3,
ORC_VAR_D1);
orc_program_append_2 (p, "convubw", 2, ORC_VAR_T7, ORC_VAR_T1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "loadl", 0, ORC_VAR_T1, ORC_VAR_D1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "convubw", 2, ORC_VAR_T6, ORC_VAR_T1, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "subw", 2, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T6,
ORC_VAR_D1);
orc_program_append_2 (p, "mullw", 2, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_T8,
ORC_VAR_D1);
orc_program_append_2 (p, "div255w", 2, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "addw", 2, ORC_VAR_T6, ORC_VAR_T6, ORC_VAR_T7,
ORC_VAR_D1);
orc_program_append_2 (p, "convwb", 2, ORC_VAR_T1, ORC_VAR_T6, ORC_VAR_D1,
ORC_VAR_D1);
orc_program_append_2 (p, "orl", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_C1,
ORC_VAR_D1);
orc_program_append_2 (p, "storel", 0, ORC_VAR_D1, ORC_VAR_T1, ORC_VAR_D1,
ORC_VAR_D1);
result = 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->params[ORC_VAR_P1] = p1;
func = p->code_exec;
func (ex);
}
#endif
Reference in a new issue View git blame Copy permalink