scenechange: use orc to compute score

Add an orc implementation for SAD operation.

Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-bad/-/merge_requests/1024>

gst/videofilters/gstscenechange.c

@@ -82,6 +82,7 @@
 #include <gst/video/gstvideofilter.h>
 #include <string.h>
 #include "gstscenechange.h"
+#include "gstscenechangeorc.h"
 
 GST_DEBUG_CATEGORY_STATIC (gst_scene_change_debug_category);
 #define GST_CAT_DEFAULT gst_scene_change_debug_category
@@ -143,23 +144,14 @@ gst_scene_change_init (GstSceneChange * scenechange)
 static double
 get_frame_score (GstVideoFrame * f1, GstVideoFrame * f2)
 {
-  int i;
+  guint32 score = 0;
-  int j;
-  int score = 0;
   int width, height;
-  guint8 *s1;
-  guint8 *s2;
 
   width = f1->info.width;
   height = f1->info.height;
 
-  for (j = 0; j < height; j++) {
+  orc_sad_nxm_u8 (&score, f1->data[0], f1->info.stride[0],
-    s1 = (guint8 *) f1->data[0] + f1->info.stride[0] * j;
+      f2->data[0], f2->info.stride[0], width, height);
-    s2 = (guint8 *) f2->data[0] + f2->info.stride[0] * j;
-    for (i = 0; i < width; i++) {
-      score += ABS (s1[i] - s2[i]);
-    }
-  }
 
   return ((double) score) / (width * height);
 }

gst/videofilters/gstscenechangeorc-dist.c

@@ -0,0 +1,273 @@
+
+/* autogenerated from gstscenechangeorc.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 ORC_INTERNAL
+#if defined(__SUNPRO_C) && (__SUNPRO_C >= 0x590)
+#define ORC_INTERNAL __attribute__((visibility("hidden")))
+#elif defined(__SUNPRO_C) && (__SUNPRO_C >= 0x550)
+#define ORC_INTERNAL __hidden
+#elif defined (__GNUC__)
+#define ORC_INTERNAL __attribute__((visibility("hidden")))
+#else
+#define ORC_INTERNAL
+#endif
+#endif
+
+
+#ifndef DISABLE_ORC
+#include <orc/orc.h>
+#endif
+void orc_sad_nxm_u8 (orc_uint32 * ORC_RESTRICT a1,
+    const orc_uint8 * ORC_RESTRICT s1, int s1_stride,
+    const orc_uint8 * ORC_RESTRICT s2, int s2_stride, 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 (orc_uint8) 255
+#define ORC_UB_MIN 0
+#define ORC_SW_MAX 32767
+#define ORC_SW_MIN (-1-ORC_SW_MAX)
+#define ORC_UW_MAX (orc_uint16)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)&0xffU)<<8) | (((x)&0xff00U)>>8))
+#define ORC_SWAP_L(x) ((((x)&0xffU)<<24) | (((x)&0xff00U)<<8) | (((x)&0xff0000U)>>8) | (((x)&0xff000000U)>>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 */
+
+
+
+/* orc_sad_nxm_u8 */
+#ifdef DISABLE_ORC
+void
+orc_sad_nxm_u8 (orc_uint32 * ORC_RESTRICT a1, 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;
+  const orc_int8 *ORC_RESTRICT ptr4;
+  const orc_int8 *ORC_RESTRICT ptr5;
+  orc_union32 var12 = { 0 };
+  orc_int8 var32;
+  orc_int8 var33;
+
+  for (j = 0; j < m; 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: accsadubl */
+      var12.i =
+          var12.i + ORC_ABS ((orc_int32) (orc_uint8) var32 -
+          (orc_int32) (orc_uint8) var33);
+    }
+  }
+  *a1 = var12.i;
+
+}
+
+#else
+static void
+_backup_orc_sad_nxm_u8 (OrcExecutor * ORC_RESTRICT ex)
+{
+  int i;
+  int j;
+  int n = ex->n;
+  int m = ex->params[ORC_VAR_A1];
+  const orc_int8 *ORC_RESTRICT ptr4;
+  const orc_int8 *ORC_RESTRICT ptr5;
+  orc_union32 var12 = { 0 };
+  orc_int8 var32;
+  orc_int8 var33;
+
+  for (j = 0; j < m; 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: accsadubl */
+      var12.i =
+          var12.i + ORC_ABS ((orc_int32) (orc_uint8) var32 -
+          (orc_int32) (orc_uint8) var33);
+    }
+  }
+  ex->accumulators[0] = var12.i;
+
+}
+
+void
+orc_sad_nxm_u8 (orc_uint32 * ORC_RESTRICT a1, 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 volatile int p_inited = 0;
+  static OrcCode *c = 0;
+  void (*func) (OrcExecutor *);
+
+  if (!p_inited) {
+    orc_once_mutex_lock ();
+    if (!p_inited) {
+      OrcProgram *p;
+
+#if 1
+      static const orc_uint8 bc[] = {
+        1, 7, 9, 14, 111, 114, 99, 95, 115, 97, 100, 95, 110, 120, 109, 95,
+        117, 56, 12, 1, 1, 12, 1, 1, 13, 4, 182, 12, 4, 5, 2, 0,
+
+      };
+      p = orc_program_new_from_static_bytecode (bc);
+      orc_program_set_backup_function (p, _backup_orc_sad_nxm_u8);
+#else
+      p = orc_program_new ();
+      orc_program_set_2d (p);
+      orc_program_set_name (p, "orc_sad_nxm_u8");
+      orc_program_set_backup_function (p, _backup_orc_sad_nxm_u8);
+      orc_program_add_source (p, 1, "s1");
+      orc_program_add_source (p, 1, "s2");
+      orc_program_add_accumulator (p, 4, "a1");
+
+      orc_program_append_2 (p, "accsadubl", 0, ORC_VAR_A1, ORC_VAR_S1,
+          ORC_VAR_S2, ORC_VAR_D1);
+#endif
+
+      orc_program_compile (p);
+      c = orc_program_take_code (p);
+      orc_program_free (p);
+    }
+    p_inited = TRUE;
+    orc_once_mutex_unlock ();
+  }
+  ex->arrays[ORC_VAR_A2] = c;
+  ex->program = 0;
+
+  ex->n = n;
+  ORC_EXECUTOR_M (ex) = m;
+  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 = c->exec;
+  func (ex);
+  *a1 = orc_executor_get_accumulator (ex, ORC_VAR_A1);
+}
+#endif

gst/videofilters/gstscenechangeorc-dist.h

@@ -0,0 +1,87 @@
+
+/* autogenerated from gstscenechangeorc.orc */
+
+#pragma once
+
+#include <glib.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+
+#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 ORC_INTERNAL
+#if defined(__SUNPRO_C) && (__SUNPRO_C >= 0x590)
+#define ORC_INTERNAL __attribute__((visibility("hidden")))
+#elif defined(__SUNPRO_C) && (__SUNPRO_C >= 0x550)
+#define ORC_INTERNAL __hidden
+#elif defined (__GNUC__)
+#define ORC_INTERNAL __attribute__((visibility("hidden")))
+#else
+#define ORC_INTERNAL
+#endif
+#endif
+
+void orc_sad_nxm_u8 (orc_uint32 * ORC_RESTRICT a1, const orc_uint8 * ORC_RESTRICT s1, int s1_stride, const orc_uint8 * ORC_RESTRICT s2, int s2_stride, int n, int m);
+
+#ifdef __cplusplus
+}
+#endif
+

gst/videofilters/gstscenechangeorc.orc

@@ -0,0 +1,7 @@
+.function orc_sad_nxm_u8
+.flags 2d
+.accumulator 4 a1 orc_uint32
+.source 1 s1 orc_uint8
+.source 1 s2 orc_uint8
+
+accsadubl a1, s1, s2

gst/videofilters/meson.build

@@ -5,8 +5,28 @@ vfilt_sources = [
   'gstvideofiltersbad.c',
 ]
 
+orcsrc = 'gstscenechangeorc'
+if have_orcc
+  orc_h = custom_target(orcsrc + '.h',
+    input : orcsrc + '.orc',
+    output : orcsrc + '.h',
+    command : orcc_args + ['--header', '-o', '@OUTPUT@', '@INPUT@'])
+  orc_c = custom_target(orcsrc + '.c',
+    input : orcsrc + '.orc',
+    output : orcsrc + '.c',
+    command : orcc_args + ['--implementation', '-o', '@OUTPUT@', '@INPUT@'])
+  orc_targets += {'name': orcsrc, 'orc-source': files(orcsrc + '.orc'), 'header': orc_h, 'source': orc_c}
+else
+  orc_h = configure_file(input : orcsrc + '-dist.h',
+    output : orcsrc + '.h',
+    copy : true)
+  orc_c = configure_file(input : orcsrc + '-dist.c',
+    output : orcsrc + '.c',
+    copy : true)
+endif
+
 gstvideofiltersbad = library('gstvideofiltersbad',
-  vfilt_sources,
+  vfilt_sources, orc_c, orc_h,
   c_args : gst_plugins_bad_args,
   include_directories : [configinc],
   dependencies : [gstvideo_dep, gstbase_dep, orc_dep, libm],