video-converter: Add support for A420 to RGB fast path

Add fast path for A420 -> RGB format conversion

Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-base/-/merge_requests/1245>

4 files changed, 1228 insertions, 4 deletions

diff --git a/gst-libs/gst/video/video-converter.c b/gst-libs/gst/video/video-converter.c
index 990168eef..7d5c2297d 100644
--- a/gst-libs/gst/video/video-converter.c
+++ b/gst-libs/gst/video/video-converter.c
@@ -5919,6 +5919,175 @@ convert_I420_pack_ARGB (GstVideoConverter * convert, const GstVideoFrame * src,
 }
 
 static void
+convert_A420_pack_ARGB_task (FConvertTask * task)
+{
+  gint i;
+  gpointer d[GST_VIDEO_MAX_PLANES];
+
+  d[0] = FRAME_GET_LINE (task->dest, 0);
+  d[0] =
+      (guint8 *) d[0] +
+      task->out_x * GST_VIDEO_FORMAT_INFO_PSTRIDE (task->dest->info.finfo, 0);
+
+  for (i = task->height_0; i < task->height_1; i++) {
+    guint8 *sy, *su, *sv, *sa;
+
+    sy = FRAME_GET_Y_LINE (task->src, i + task->in_y);
+    sy += task->in_x;
+    su = FRAME_GET_U_LINE (task->src, (i + task->in_y) >> 1);
+    su += (task->in_x >> 1);
+    sv = FRAME_GET_V_LINE (task->src, (i + task->in_y) >> 1);
+    sv += (task->in_x >> 1);
+    sa = FRAME_GET_A_LINE (task->src, i + task->in_y);
+    sa += task->in_x;
+
+#if G_BYTE_ORDER == G_LITTLE_ENDIAN
+    video_orc_convert_A420_ARGB (task->tmpline, sy, su, sv, sa,
+        task->data->im[0][0], task->data->im[0][2],
+        task->data->im[2][1], task->data->im[1][1], task->data->im[1][2],
+        task->width);
+#else
+    video_orc_convert_A420_BGRA (task->tmpline, sy, su, sv, sa,
+        task->data->im[0][0], task->data->im[0][2],
+        task->data->im[2][1], task->data->im[1][1], task->data->im[1][2],
+        task->width);
+#endif
+
+    task->dest->info.finfo->pack_func (task->dest->info.finfo,
+        (GST_VIDEO_FRAME_IS_INTERLACED (task->dest) ?
+            GST_VIDEO_PACK_FLAG_INTERLACED :
+            GST_VIDEO_PACK_FLAG_NONE),
+        task->tmpline, 0, d, task->dest->info.stride,
+        task->dest->info.chroma_site, i + task->out_y, task->width);
+  }
+}
+
+static void
+convert_A420_pack_ARGB (GstVideoConverter * convert, const GstVideoFrame * src,
+    GstVideoFrame * dest)
+{
+  int i;
+  gint width = convert->in_width;
+  gint height = convert->in_height;
+  MatrixData *data = &convert->convert_matrix;
+  FConvertTask *tasks;
+  FConvertTask **tasks_p;
+  gint n_threads;
+  gint lines_per_thread;
+
+  n_threads = convert->conversion_runner->n_threads;
+  tasks = convert->tasks[0] =
+      g_renew (FConvertTask, convert->tasks[0], n_threads);
+  tasks_p = convert->tasks_p[0] =
+      g_renew (FConvertTask *, convert->tasks_p[0], n_threads);
+
+  lines_per_thread = (height + n_threads - 1) / n_threads;
+
+  for (i = 0; i < n_threads; i++) {
+    tasks[i].src = src;
+    tasks[i].dest = dest;
+
+    tasks[i].width = width;
+    tasks[i].data = data;
+    tasks[i].in_x = convert->in_x;
+    tasks[i].in_y = convert->in_y;
+    tasks[i].out_x = convert->out_x;
+    tasks[i].out_y = convert->out_y;
+    tasks[i].tmpline = convert->tmpline[i];
+
+    tasks[i].height_0 = i * lines_per_thread;
+    tasks[i].height_1 = tasks[i].height_0 + lines_per_thread;
+    tasks[i].height_1 = MIN (height, tasks[i].height_1);
+
+    tasks_p[i] = &tasks[i];
+  }
+
+  gst_parallelized_task_runner_run (convert->conversion_runner,
+      (GstParallelizedTaskFunc) convert_A420_pack_ARGB_task,
+      (gpointer) tasks_p);
+
+  convert_fill_border (convert, dest);
+}
+
+static void
+convert_A420_BGRA_task (FConvertTask * task)
+{
+  gint i;
+
+  for (i = task->height_0; i < task->height_1; i++) {
+    guint8 *sy, *su, *sv, *sa, *d;
+
+    d = FRAME_GET_LINE (task->dest, i + task->out_y);
+    d += (task->out_x * 4);
+    sy = FRAME_GET_Y_LINE (task->src, i + task->in_y);
+    sy += task->in_x;
+    su = FRAME_GET_U_LINE (task->src, (i + task->in_y) >> 1);
+    su += (task->in_x >> 1);
+    sv = FRAME_GET_V_LINE (task->src, (i + task->in_y) >> 1);
+    sv += (task->in_x >> 1);
+    sa = FRAME_GET_A_LINE (task->src, i + task->in_y);
+    sa += task->in_x;
+
+#if G_BYTE_ORDER == G_LITTLE_ENDIAN
+    video_orc_convert_A420_BGRA (d, sy, su, sv, sa,
+        task->data->im[0][0], task->data->im[0][2],
+        task->data->im[2][1], task->data->im[1][1], task->data->im[1][2],
+        task->width);
+#else
+    video_orc_convert_A420_ARGB (d, sy, su, sv, sa,
+        task->data->im[0][0], task->data->im[0][2],
+        task->data->im[2][1], task->data->im[1][1], task->data->im[1][2],
+        task->width);
+#endif
+  }
+}
+
+static void
+convert_A420_BGRA (GstVideoConverter * convert, const GstVideoFrame * src,
+    GstVideoFrame * dest)
+{
+  int i;
+  gint width = convert->in_width;
+  gint height = convert->in_height;
+  MatrixData *data = &convert->convert_matrix;
+  FConvertTask *tasks;
+  FConvertTask **tasks_p;
+  gint n_threads;
+  gint lines_per_thread;
+
+  n_threads = convert->conversion_runner->n_threads;
+  tasks = convert->tasks[0] =
+      g_renew (FConvertTask, convert->tasks[0], n_threads);
+  tasks_p = convert->tasks_p[0] =
+      g_renew (FConvertTask *, convert->tasks_p[0], n_threads);
+
+  lines_per_thread = (height + n_threads - 1) / n_threads;
+
+  for (i = 0; i < n_threads; i++) {
+    tasks[i].src = src;
+    tasks[i].dest = dest;
+
+    tasks[i].width = width;
+    tasks[i].data = data;
+    tasks[i].in_x = convert->in_x;
+    tasks[i].in_y = convert->in_y;
+    tasks[i].out_x = convert->out_x;
+    tasks[i].out_y = convert->out_y;
+
+    tasks[i].height_0 = i * lines_per_thread;
+    tasks[i].height_1 = tasks[i].height_0 + lines_per_thread;
+    tasks[i].height_1 = MIN (height, tasks[i].height_1);
+
+    tasks_p[i] = &tasks[i];
+  }
+
+  gst_parallelized_task_runner_run (convert->conversion_runner,
+      (GstParallelizedTaskFunc) convert_A420_BGRA_task, (gpointer) tasks_p);
+
+  convert_fill_border (convert, dest);
+}
+
+static void
 memset_u24 (guint8 * data, guint8 col[3], unsigned int n)
 {
   unsigned int i;
@@ -7372,6 +7541,28 @@ static const VideoTransform transforms[] = {
   {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_BGR16, FALSE, TRUE, TRUE, TRUE,
       TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
 
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_ABGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, TRUE, FALSE, FALSE, 0, 0, convert_A420_pack_ARGB},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_RGBA, FALSE, TRUE, TRUE, TRUE,
+      TRUE, TRUE, FALSE, FALSE, 0, 0, convert_A420_pack_ARGB},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_BGRA, FALSE, TRUE, TRUE, TRUE,
+      TRUE, TRUE, FALSE, FALSE, 0, 0, convert_A420_BGRA},
+  /* A420 to non-alpha RGB formats, reuse I420_* method */
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_BGRx, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_BGRA},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_xBGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_RGBx, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_RGB, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_BGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_RGB15, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_A420, GST_VIDEO_FORMAT_BGR16, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+
   /* scalers */
   {GST_VIDEO_FORMAT_GBR, GST_VIDEO_FORMAT_GBR, TRUE, FALSE, FALSE, TRUE,
       TRUE, FALSE, FALSE, FALSE, 0, 0, convert_scale_planes},
diff --git a/gst-libs/gst/video/video-orc-dist.c b/gst-libs/gst/video/video-orc-dist.c
index 809637f1d..f8ca3f5f3 100644
--- a/gst-libs/gst/video/video-orc-dist.c
+++ b/gst-libs/gst/video/video-orc-dist.c
@@ -360,6 +360,14 @@ void video_orc_convert_I420_ARGB (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 video_orc_convert_A420_ARGB (guint8 * ORC_RESTRICT d1,
+    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
+    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
+    int p2, int p3, int p4, int p5, int n);
+void video_orc_convert_A420_BGRA (guint8 * ORC_RESTRICT d1,
+    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
+    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
+    int p2, int p3, int p4, int p5, int n);
 void video_orc_matrix8 (guint8 * ORC_RESTRICT d1,
     const guint8 * ORC_RESTRICT s1, orc_int64 p1, orc_int64 p2, orc_int64 p3,
     orc_int64 p4, int n);
@@ -24295,6 +24303,920 @@ video_orc_convert_I420_ARGB (guint8 * ORC_RESTRICT d1,
 #endif
 
 
+/* video_orc_convert_A420_ARGB */
+#ifdef DISABLE_ORC
+void
+video_orc_convert_A420_ARGB (guint8 * ORC_RESTRICT d1,
+    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
+    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
+    int p2, int p3, int p4, int p5, 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;
+  orc_int8 var43;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var44;
+#else
+  orc_int8 var44;
+#endif
+  orc_int8 var45;
+  orc_union16 var46;
+  orc_union16 var47;
+  orc_union16 var48;
+  orc_union16 var49;
+  orc_union16 var50;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_union32 var51;
+#else
+  orc_union32 var51;
+#endif
+  orc_union32 var52;
+  orc_int8 var53;
+  orc_union16 var54;
+  orc_int8 var55;
+  orc_int8 var56;
+  orc_union16 var57;
+  orc_int8 var58;
+  orc_int8 var59;
+  orc_union16 var60;
+  orc_int8 var61;
+  orc_union16 var62;
+  orc_union16 var63;
+  orc_union16 var64;
+  orc_int8 var65;
+  orc_union16 var66;
+  orc_union16 var67;
+  orc_union16 var68;
+  orc_int8 var69;
+  orc_union16 var70;
+  orc_union16 var71;
+  orc_union16 var72;
+  orc_union16 var73;
+  orc_int8 var74;
+  orc_union16 var75;
+  orc_union32 var76;
+
+  ptr0 = (orc_union32 *) d1;
+  ptr4 = (orc_int8 *) s1;
+  ptr5 = (orc_int8 *) s2;
+  ptr6 = (orc_int8 *) s3;
+  ptr7 = (orc_int8 *) s4;
+
+  /* 1: loadpb */
+  var44 = 0x00000080;           /* 128 or 6.32404e-322f */
+  /* 12: loadpw */
+  var46.i = p1;
+  /* 14: loadpw */
+  var47.i = p2;
+  /* 19: loadpw */
+  var48.i = p3;
+  /* 23: loadpw */
+  var49.i = p4;
+  /* 26: loadpw */
+  var50.i = p5;
+  /* 32: loadpb */
+  var51.x4[0] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[1] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[2] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[3] = 0x00000080;     /* 128 or 6.32404e-322f */
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var43 = ptr4[i];
+    /* 2: subb */
+    var53 = var43 - var44;
+    /* 3: splatbw */
+    var54.i = ((var53 & 0xff) << 8) | (var53 & 0xff);
+    /* 4: loadupdb */
+    var55 = ptr5[i >> 1];
+    /* 5: subb */
+    var56 = var55 - var44;
+    /* 6: splatbw */
+    var57.i = ((var56 & 0xff) << 8) | (var56 & 0xff);
+    /* 7: loadupdb */
+    var58 = ptr6[i >> 1];
+    /* 8: subb */
+    var59 = var58 - var44;
+    /* 9: splatbw */
+    var60.i = ((var59 & 0xff) << 8) | (var59 & 0xff);
+    /* 10: loadb */
+    var45 = ptr7[i];
+    /* 11: subb */
+    var61 = var45 - var44;
+    /* 13: mulhsw */
+    var62.i = (var54.i * var46.i) >> 16;
+    /* 15: mulhsw */
+    var63.i = (var60.i * var47.i) >> 16;
+    /* 16: addw */
+    var64.i = var62.i + var63.i;
+    /* 17: convssswb */
+    var65 = ORC_CLAMP_SB (var64.i);
+    /* 18: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var61;
+      _dest.x2[1] = var65;
+      var66.i = _dest.i;
+    }
+    /* 20: mulhsw */
+    var67.i = (var57.i * var48.i) >> 16;
+    /* 21: addw */
+    var68.i = var62.i + var67.i;
+    /* 22: convssswb */
+    var69 = ORC_CLAMP_SB (var68.i);
+    /* 24: mulhsw */
+    var70.i = (var57.i * var49.i) >> 16;
+    /* 25: addw */
+    var71.i = var62.i + var70.i;
+    /* 27: mulhsw */
+    var72.i = (var60.i * var50.i) >> 16;
+    /* 28: addw */
+    var73.i = var71.i + var72.i;
+    /* 29: convssswb */
+    var74 = ORC_CLAMP_SB (var73.i);
+    /* 30: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var74;
+      _dest.x2[1] = var69;
+      var75.i = _dest.i;
+    }
+    /* 31: mergewl */
+    {
+      orc_union32 _dest;
+      _dest.x2[0] = var66.i;
+      _dest.x2[1] = var75.i;
+      var76.i = _dest.i;
+    }
+    /* 33: addb */
+    var52.x4[0] = var76.x4[0] + var51.x4[0];
+    var52.x4[1] = var76.x4[1] + var51.x4[1];
+    var52.x4[2] = var76.x4[2] + var51.x4[2];
+    var52.x4[3] = var76.x4[3] + var51.x4[3];
+    /* 34: storel */
+    ptr0[i] = var52;
+  }
+
+}
+
+#else
+static void
+_backup_video_orc_convert_A420_ARGB (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;
+  orc_int8 var43;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var44;
+#else
+  orc_int8 var44;
+#endif
+  orc_int8 var45;
+  orc_union16 var46;
+  orc_union16 var47;
+  orc_union16 var48;
+  orc_union16 var49;
+  orc_union16 var50;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_union32 var51;
+#else
+  orc_union32 var51;
+#endif
+  orc_union32 var52;
+  orc_int8 var53;
+  orc_union16 var54;
+  orc_int8 var55;
+  orc_int8 var56;
+  orc_union16 var57;
+  orc_int8 var58;
+  orc_int8 var59;
+  orc_union16 var60;
+  orc_int8 var61;
+  orc_union16 var62;
+  orc_union16 var63;
+  orc_union16 var64;
+  orc_int8 var65;
+  orc_union16 var66;
+  orc_union16 var67;
+  orc_union16 var68;
+  orc_int8 var69;
+  orc_union16 var70;
+  orc_union16 var71;
+  orc_union16 var72;
+  orc_union16 var73;
+  orc_int8 var74;
+  orc_union16 var75;
+  orc_union32 var76;
+
+  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];
+
+  /* 1: loadpb */
+  var44 = 0x00000080;           /* 128 or 6.32404e-322f */
+  /* 12: loadpw */
+  var46.i = ex->params[24];
+  /* 14: loadpw */
+  var47.i = ex->params[25];
+  /* 19: loadpw */
+  var48.i = ex->params[26];
+  /* 23: loadpw */
+  var49.i = ex->params[27];
+  /* 26: loadpw */
+  var50.i = ex->params[28];
+  /* 32: loadpb */
+  var51.x4[0] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[1] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[2] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[3] = 0x00000080;     /* 128 or 6.32404e-322f */
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var43 = ptr4[i];
+    /* 2: subb */
+    var53 = var43 - var44;
+    /* 3: splatbw */
+    var54.i = ((var53 & 0xff) << 8) | (var53 & 0xff);
+    /* 4: loadupdb */
+    var55 = ptr5[i >> 1];
+    /* 5: subb */
+    var56 = var55 - var44;
+    /* 6: splatbw */
+    var57.i = ((var56 & 0xff) << 8) | (var56 & 0xff);
+    /* 7: loadupdb */
+    var58 = ptr6[i >> 1];
+    /* 8: subb */
+    var59 = var58 - var44;
+    /* 9: splatbw */
+    var60.i = ((var59 & 0xff) << 8) | (var59 & 0xff);
+    /* 10: loadb */
+    var45 = ptr7[i];
+    /* 11: subb */
+    var61 = var45 - var44;
+    /* 13: mulhsw */
+    var62.i = (var54.i * var46.i) >> 16;
+    /* 15: mulhsw */
+    var63.i = (var60.i * var47.i) >> 16;
+    /* 16: addw */
+    var64.i = var62.i + var63.i;
+    /* 17: convssswb */
+    var65 = ORC_CLAMP_SB (var64.i);
+    /* 18: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var61;
+      _dest.x2[1] = var65;
+      var66.i = _dest.i;
+    }
+    /* 20: mulhsw */
+    var67.i = (var57.i * var48.i) >> 16;
+    /* 21: addw */
+    var68.i = var62.i + var67.i;
+    /* 22: convssswb */
+    var69 = ORC_CLAMP_SB (var68.i);
+    /* 24: mulhsw */
+    var70.i = (var57.i * var49.i) >> 16;
+    /* 25: addw */
+    var71.i = var62.i + var70.i;
+    /* 27: mulhsw */
+    var72.i = (var60.i * var50.i) >> 16;
+    /* 28: addw */
+    var73.i = var71.i + var72.i;
+    /* 29: convssswb */
+    var74 = ORC_CLAMP_SB (var73.i);
+    /* 30: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var74;
+      _dest.x2[1] = var69;
+      var75.i = _dest.i;
+    }
+    /* 31: mergewl */
+    {
+      orc_union32 _dest;
+      _dest.x2[0] = var66.i;
+      _dest.x2[1] = var75.i;
+      var76.i = _dest.i;
+    }
+    /* 33: addb */
+    var52.x4[0] = var76.x4[0] + var51.x4[0];
+    var52.x4[1] = var76.x4[1] + var51.x4[1];
+    var52.x4[2] = var76.x4[2] + var51.x4[2];
+    var52.x4[3] = var76.x4[3] + var51.x4[3];
+    /* 34: storel */
+    ptr0[i] = var52;
+  }
+
+}
+
+void
+video_orc_convert_A420_ARGB (guint8 * ORC_RESTRICT d1,
+    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
+    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
+    int p2, int p3, int p4, int p5, int n)
+{
+  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, 9, 27, 118, 105, 100, 101, 111, 95, 111, 114, 99, 95, 99, 111, 110,
+        118, 101, 114, 116, 95, 65, 52, 50, 48, 95, 65, 82, 71, 66, 11, 4,
+        4, 12, 1, 1, 12, 1, 1, 12, 1, 1, 12, 1, 1, 14, 1, 128,
+        0, 0, 0, 16, 2, 16, 2, 16, 2, 16, 2, 16, 2, 20, 2, 20,
+        2, 20, 2, 20, 2, 20, 2, 20, 2, 20, 1, 20, 1, 20, 1, 20,
+        1, 20, 4, 65, 38, 4, 16, 151, 32, 38, 45, 38, 5, 65, 38, 38,
+        16, 151, 33, 38, 45, 38, 6, 65, 38, 38, 16, 151, 34, 38, 65, 41,
+        7, 16, 90, 32, 32, 24, 90, 35, 34, 25, 70, 35, 32, 35, 159, 38,
+        35, 196, 35, 41, 38, 90, 37, 33, 26, 70, 37, 32, 37, 159, 40, 37,
+        90, 36, 33, 27, 70, 36, 32, 36, 90, 32, 34, 28, 70, 36, 36, 32,
+        159, 39, 36, 196, 37, 39, 40, 195, 42, 35, 37, 21, 2, 33, 0, 42,
+        16, 2, 0,
+      };
+      p = orc_program_new_from_static_bytecode (bc);
+      orc_program_set_backup_function (p, _backup_video_orc_convert_A420_ARGB);
+#else
+      p = orc_program_new ();
+      orc_program_set_name (p, "video_orc_convert_A420_ARGB");
+      orc_program_set_backup_function (p, _backup_video_orc_convert_A420_ARGB);
+      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_constant (p, 1, 0x00000080, "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_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_add_temporary (p, 2, "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_add_temporary (p, 1, "t9");
+      orc_program_add_temporary (p, 1, "t10");
+      orc_program_add_temporary (p, 4, "t11");
+
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T7, ORC_VAR_S1, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_D1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T7, ORC_VAR_S2,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T2, ORC_VAR_T7, ORC_VAR_D1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T7, ORC_VAR_S3,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T3, ORC_VAR_T7, ORC_VAR_D1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T10, ORC_VAR_S4, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_P2,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T1, ORC_VAR_T4,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T7, ORC_VAR_T4,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T4, ORC_VAR_T10,
+          ORC_VAR_T7, ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T6, ORC_VAR_T2, ORC_VAR_P3,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T6, ORC_VAR_T1, ORC_VAR_T6,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T9, ORC_VAR_T6,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T5, ORC_VAR_T2, ORC_VAR_P4,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T5, ORC_VAR_T1, ORC_VAR_T5,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_P5,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_T1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T8, ORC_VAR_T5,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T6, ORC_VAR_T8, ORC_VAR_T9,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T11, ORC_VAR_T4,
+          ORC_VAR_T6, ORC_VAR_D1);
+      orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T11, ORC_VAR_C1,
+          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;
+  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;
+  ex->params[ORC_VAR_P5] = p5;
+
+  func = c->exec;
+  func (ex);
+}
+#endif
+
+
+/* video_orc_convert_A420_BGRA */
+#ifdef DISABLE_ORC
+void
+video_orc_convert_A420_BGRA (guint8 * ORC_RESTRICT d1,
+    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
+    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
+    int p2, int p3, int p4, int p5, 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;
+  orc_int8 var43;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var44;
+#else
+  orc_int8 var44;
+#endif
+  orc_int8 var45;
+  orc_union16 var46;
+  orc_union16 var47;
+  orc_union16 var48;
+  orc_union16 var49;
+  orc_union16 var50;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_union32 var51;
+#else
+  orc_union32 var51;
+#endif
+  orc_union32 var52;
+  orc_int8 var53;
+  orc_union16 var54;
+  orc_int8 var55;
+  orc_int8 var56;
+  orc_union16 var57;
+  orc_int8 var58;
+  orc_int8 var59;
+  orc_union16 var60;
+  orc_int8 var61;
+  orc_union16 var62;
+  orc_union16 var63;
+  orc_union16 var64;
+  orc_int8 var65;
+  orc_union16 var66;
+  orc_union16 var67;
+  orc_union16 var68;
+  orc_int8 var69;
+  orc_union16 var70;
+  orc_union16 var71;
+  orc_union16 var72;
+  orc_union16 var73;
+  orc_int8 var74;
+  orc_union16 var75;
+  orc_union32 var76;
+
+  ptr0 = (orc_union32 *) d1;
+  ptr4 = (orc_int8 *) s1;
+  ptr5 = (orc_int8 *) s2;
+  ptr6 = (orc_int8 *) s3;
+  ptr7 = (orc_int8 *) s4;
+
+  /* 1: loadpb */
+  var44 = 0x00000080;           /* 128 or 6.32404e-322f */
+  /* 12: loadpw */
+  var46.i = p1;
+  /* 14: loadpw */
+  var47.i = p2;
+  /* 19: loadpw */
+  var48.i = p3;
+  /* 23: loadpw */
+  var49.i = p4;
+  /* 26: loadpw */
+  var50.i = p5;
+  /* 32: loadpb */
+  var51.x4[0] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[1] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[2] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[3] = 0x00000080;     /* 128 or 6.32404e-322f */
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var43 = ptr4[i];
+    /* 2: subb */
+    var53 = var43 - var44;
+    /* 3: splatbw */
+    var54.i = ((var53 & 0xff) << 8) | (var53 & 0xff);
+    /* 4: loadupdb */
+    var55 = ptr5[i >> 1];
+    /* 5: subb */
+    var56 = var55 - var44;
+    /* 6: splatbw */
+    var57.i = ((var56 & 0xff) << 8) | (var56 & 0xff);
+    /* 7: loadupdb */
+    var58 = ptr6[i >> 1];
+    /* 8: subb */
+    var59 = var58 - var44;
+    /* 9: splatbw */
+    var60.i = ((var59 & 0xff) << 8) | (var59 & 0xff);
+    /* 10: loadb */
+    var45 = ptr7[i];
+    /* 11: subb */
+    var61 = var45 - var44;
+    /* 13: mulhsw */
+    var62.i = (var54.i * var46.i) >> 16;
+    /* 15: mulhsw */
+    var63.i = (var60.i * var47.i) >> 16;
+    /* 16: addw */
+    var64.i = var62.i + var63.i;
+    /* 17: convssswb */
+    var65 = ORC_CLAMP_SB (var64.i);
+    /* 18: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var65;
+      _dest.x2[1] = var61;
+      var66.i = _dest.i;
+    }
+    /* 20: mulhsw */
+    var67.i = (var57.i * var48.i) >> 16;
+    /* 21: addw */
+    var68.i = var62.i + var67.i;
+    /* 22: convssswb */
+    var69 = ORC_CLAMP_SB (var68.i);
+    /* 24: mulhsw */
+    var70.i = (var57.i * var49.i) >> 16;
+    /* 25: addw */
+    var71.i = var62.i + var70.i;
+    /* 27: mulhsw */
+    var72.i = (var60.i * var50.i) >> 16;
+    /* 28: addw */
+    var73.i = var71.i + var72.i;
+    /* 29: convssswb */
+    var74 = ORC_CLAMP_SB (var73.i);
+    /* 30: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var69;
+      _dest.x2[1] = var74;
+      var75.i = _dest.i;
+    }
+    /* 31: mergewl */
+    {
+      orc_union32 _dest;
+      _dest.x2[0] = var75.i;
+      _dest.x2[1] = var66.i;
+      var76.i = _dest.i;
+    }
+    /* 33: addb */
+    var52.x4[0] = var76.x4[0] + var51.x4[0];
+    var52.x4[1] = var76.x4[1] + var51.x4[1];
+    var52.x4[2] = var76.x4[2] + var51.x4[2];
+    var52.x4[3] = var76.x4[3] + var51.x4[3];
+    /* 34: storel */
+    ptr0[i] = var52;
+  }
+
+}
+
+#else
+static void
+_backup_video_orc_convert_A420_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;
+  const orc_int8 *ORC_RESTRICT ptr7;
+  orc_int8 var43;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var44;
+#else
+  orc_int8 var44;
+#endif
+  orc_int8 var45;
+  orc_union16 var46;
+  orc_union16 var47;
+  orc_union16 var48;
+  orc_union16 var49;
+  orc_union16 var50;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_union32 var51;
+#else
+  orc_union32 var51;
+#endif
+  orc_union32 var52;
+  orc_int8 var53;
+  orc_union16 var54;
+  orc_int8 var55;
+  orc_int8 var56;
+  orc_union16 var57;
+  orc_int8 var58;
+  orc_int8 var59;
+  orc_union16 var60;
+  orc_int8 var61;
+  orc_union16 var62;
+  orc_union16 var63;
+  orc_union16 var64;
+  orc_int8 var65;
+  orc_union16 var66;
+  orc_union16 var67;
+  orc_union16 var68;
+  orc_int8 var69;
+  orc_union16 var70;
+  orc_union16 var71;
+  orc_union16 var72;
+  orc_union16 var73;
+  orc_int8 var74;
+  orc_union16 var75;
+  orc_union32 var76;
+
+  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];
+
+  /* 1: loadpb */
+  var44 = 0x00000080;           /* 128 or 6.32404e-322f */
+  /* 12: loadpw */
+  var46.i = ex->params[24];
+  /* 14: loadpw */
+  var47.i = ex->params[25];
+  /* 19: loadpw */
+  var48.i = ex->params[26];
+  /* 23: loadpw */
+  var49.i = ex->params[27];
+  /* 26: loadpw */
+  var50.i = ex->params[28];
+  /* 32: loadpb */
+  var51.x4[0] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[1] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[2] = 0x00000080;     /* 128 or 6.32404e-322f */
+  var51.x4[3] = 0x00000080;     /* 128 or 6.32404e-322f */
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var43 = ptr4[i];
+    /* 2: subb */
+    var53 = var43 - var44;
+    /* 3: splatbw */
+    var54.i = ((var53 & 0xff) << 8) | (var53 & 0xff);
+    /* 4: loadupdb */
+    var55 = ptr5[i >> 1];
+    /* 5: subb */
+    var56 = var55 - var44;
+    /* 6: splatbw */
+    var57.i = ((var56 & 0xff) << 8) | (var56 & 0xff);
+    /* 7: loadupdb */
+    var58 = ptr6[i >> 1];
+    /* 8: subb */
+    var59 = var58 - var44;
+    /* 9: splatbw */
+    var60.i = ((var59 & 0xff) << 8) | (var59 & 0xff);
+    /* 10: loadb */
+    var45 = ptr7[i];
+    /* 11: subb */
+    var61 = var45 - var44;
+    /* 13: mulhsw */
+    var62.i = (var54.i * var46.i) >> 16;
+    /* 15: mulhsw */
+    var63.i = (var60.i * var47.i) >> 16;
+    /* 16: addw */
+    var64.i = var62.i + var63.i;
+    /* 17: convssswb */
+    var65 = ORC_CLAMP_SB (var64.i);
+    /* 18: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var65;
+      _dest.x2[1] = var61;
+      var66.i = _dest.i;
+    }
+    /* 20: mulhsw */
+    var67.i = (var57.i * var48.i) >> 16;
+    /* 21: addw */
+    var68.i = var62.i + var67.i;
+    /* 22: convssswb */
+    var69 = ORC_CLAMP_SB (var68.i);
+    /* 24: mulhsw */
+    var70.i = (var57.i * var49.i) >> 16;
+    /* 25: addw */
+    var71.i = var62.i + var70.i;
+    /* 27: mulhsw */
+    var72.i = (var60.i * var50.i) >> 16;
+    /* 28: addw */
+    var73.i = var71.i + var72.i;
+    /* 29: convssswb */
+    var74 = ORC_CLAMP_SB (var73.i);
+    /* 30: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var69;
+      _dest.x2[1] = var74;
+      var75.i = _dest.i;
+    }
+    /* 31: mergewl */
+    {
+      orc_union32 _dest;
+      _dest.x2[0] = var75.i;
+      _dest.x2[1] = var66.i;
+      var76.i = _dest.i;
+    }
+    /* 33: addb */
+    var52.x4[0] = var76.x4[0] + var51.x4[0];
+    var52.x4[1] = var76.x4[1] + var51.x4[1];
+    var52.x4[2] = var76.x4[2] + var51.x4[2];
+    var52.x4[3] = var76.x4[3] + var51.x4[3];
+    /* 34: storel */
+    ptr0[i] = var52;
+  }
+
+}
+
+void
+video_orc_convert_A420_BGRA (guint8 * ORC_RESTRICT d1,
+    const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2,
+    const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1,
+    int p2, int p3, int p4, int p5, int n)
+{
+  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, 9, 27, 118, 105, 100, 101, 111, 95, 111, 114, 99, 95, 99, 111, 110,
+        118, 101, 114, 116, 95, 65, 52, 50, 48, 95, 66, 71, 82, 65, 11, 4,
+        4, 12, 1, 1, 12, 1, 1, 12, 1, 1, 12, 1, 1, 14, 1, 128,
+        0, 0, 0, 16, 2, 16, 2, 16, 2, 16, 2, 16, 2, 20, 2, 20,
+        2, 20, 2, 20, 2, 20, 2, 20, 2, 20, 1, 20, 1, 20, 1, 20,
+        1, 20, 4, 65, 38, 4, 16, 151, 32, 38, 45, 38, 5, 65, 38, 38,
+        16, 151, 33, 38, 45, 38, 6, 65, 38, 38, 16, 151, 34, 38, 65, 41,
+        7, 16, 90, 32, 32, 24, 90, 35, 34, 25, 70, 35, 32, 35, 159, 38,
+        35, 196, 35, 38, 41, 90, 37, 33, 26, 70, 37, 32, 37, 159, 40, 37,
+        90, 36, 33, 27, 70, 36, 32, 36, 90, 32, 34, 28, 70, 36, 36, 32,
+        159, 39, 36, 196, 37, 40, 39, 195, 42, 37, 35, 21, 2, 33, 0, 42,
+        16, 2, 0,
+      };
+      p = orc_program_new_from_static_bytecode (bc);
+      orc_program_set_backup_function (p, _backup_video_orc_convert_A420_BGRA);
+#else
+      p = orc_program_new ();
+      orc_program_set_name (p, "video_orc_convert_A420_BGRA");
+      orc_program_set_backup_function (p, _backup_video_orc_convert_A420_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_source (p, 1, "s4");
+      orc_program_add_constant (p, 1, 0x00000080, "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_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_add_temporary (p, 2, "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_add_temporary (p, 1, "t9");
+      orc_program_add_temporary (p, 1, "t10");
+      orc_program_add_temporary (p, 4, "t11");
+
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T7, ORC_VAR_S1, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T1, ORC_VAR_T7, ORC_VAR_D1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T7, ORC_VAR_S2,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T2, ORC_VAR_T7, ORC_VAR_D1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "loadupdb", 0, ORC_VAR_T7, ORC_VAR_S3,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T7, ORC_VAR_T7, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "splatbw", 0, ORC_VAR_T3, ORC_VAR_T7, ORC_VAR_D1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "subb", 0, ORC_VAR_T10, ORC_VAR_S4, ORC_VAR_C1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T1, ORC_VAR_T1, ORC_VAR_P1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T4, ORC_VAR_T3, ORC_VAR_P2,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T4, ORC_VAR_T1, ORC_VAR_T4,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T7, ORC_VAR_T4,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T4, ORC_VAR_T7,
+          ORC_VAR_T10, ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T6, ORC_VAR_T2, ORC_VAR_P3,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T6, ORC_VAR_T1, ORC_VAR_T6,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T9, ORC_VAR_T6,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T5, ORC_VAR_T2, ORC_VAR_P4,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T5, ORC_VAR_T1, ORC_VAR_T5,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mulhsw", 0, ORC_VAR_T1, ORC_VAR_T3, ORC_VAR_P5,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "addw", 0, ORC_VAR_T5, ORC_VAR_T5, ORC_VAR_T1,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "convssswb", 0, ORC_VAR_T8, ORC_VAR_T5,
+          ORC_VAR_D1, ORC_VAR_D1);
+      orc_program_append_2 (p, "mergebw", 0, ORC_VAR_T6, ORC_VAR_T9, ORC_VAR_T8,
+          ORC_VAR_D1);
+      orc_program_append_2 (p, "mergewl", 0, ORC_VAR_T11, ORC_VAR_T6,
+          ORC_VAR_T4, ORC_VAR_D1);
+      orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T11, ORC_VAR_C1,
+          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;
+  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;
+  ex->params[ORC_VAR_P5] = p5;
+
+  func = c->exec;
+  func (ex);
+}
+#endif
+
+
 /* video_orc_matrix8 */
 #ifdef DISABLE_ORC
 void
diff --git a/gst-libs/gst/video/video-orc-dist.h b/gst-libs/gst/video/video-orc-dist.h
index 19fb7aae5..4f060205b 100644
--- a/gst-libs/gst/video/video-orc-dist.h
+++ b/gst-libs/gst/video/video-orc-dist.h
@@ -1,8 +1,7 @@
 
 /* autogenerated from video-orc.orc */
 
-#ifndef _VIDEO_ORC_H_
-#define _VIDEO_ORC_H_
+#pragma once
 
 #include <glib.h>
 
@@ -189,6 +188,8 @@ void video_orc_convert_AYUV_ABGR (guint8 * ORC_RESTRICT d1, int d1_stride, const
 void video_orc_convert_AYUV_RGBA (guint8 * ORC_RESTRICT d1, int d1_stride, const guint8 * ORC_RESTRICT s1, int s1_stride, int p1, int p2, int p3, int p4, int p5, int n, int m);
 void video_orc_convert_I420_BGRA (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 video_orc_convert_I420_ARGB (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 video_orc_convert_A420_ARGB (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1, int p2, int p3, int p4, int p5, int n);
+void video_orc_convert_A420_BGRA (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1, const guint8 * ORC_RESTRICT s2, const guint8 * ORC_RESTRICT s3, const guint8 * ORC_RESTRICT s4, int p1, int p2, int p3, int p4, int p5, int n);
 void video_orc_matrix8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1, orc_int64 p1, orc_int64 p2, orc_int64 p3, orc_int64 p4, int n);
 void _custom_video_orc_matrix8 (guint8 * ORC_RESTRICT d1, const guint8 * ORC_RESTRICT s1, orc_int64 p1, orc_int64 p2, orc_int64 p3, orc_int64 p4, int n);
 void video_orc_resample_h_near_u32_lq (guint32 * ORC_RESTRICT d1, const guint32 * ORC_RESTRICT s1, int p1, int p2, int n);
@@ -242,5 +243,3 @@ void video_orc_convert_UYVY_GRAY8 (guint8 * ORC_RESTRICT d1, int d1_stride, cons
 }
 #endif
 
-#endif
-
diff --git a/gst-libs/gst/video/video-orc.orc b/gst-libs/gst/video/video-orc.orc
index 03a302f1e..4d88844ba 100644
--- a/gst-libs/gst/video/video-orc.orc
+++ b/gst-libs/gst/video/video-orc.orc
@@ -1876,6 +1876,118 @@ mergebw wb, g, b
 mergewl x, wr, wb
 x4 addb argb, x, c4128
 
+.function video_orc_convert_A420_ARGB
+.dest 4 argb guint8
+.source 1 y guint8
+.source 1 u guint8
+.source 1 v guint8
+.source 1 a guint8
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.param 2 p5
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 1 da
+.temp 4 x
+.const 1 c128 128
+
+subb r, y, c128
+splatbw wy, r
+loadupdb r, u
+subb r, r, c128
+splatbw wu, r
+loadupdb r, v
+subb r, r, c128
+splatbw wv, r
+subb da, a, c128
+
+mulhsw wy, wy, p1
+
+mulhsw wr, wv, p2
+addw wr, wy, wr
+convssswb r, wr
+mergebw wr, da, r
+
+mulhsw wb, wu, p3
+addw wb, wy, wb
+convssswb b, wb
+
+mulhsw wg, wu, p4
+addw wg, wy, wg
+mulhsw wy, wv, p5
+addw wg, wg, wy
+
+convssswb g, wg
+
+mergebw wb, g, b
+mergewl x, wr, wb
+x4 addb argb, x, c128
+
+.function video_orc_convert_A420_BGRA
+.dest 4 argb guint8
+.source 1 y guint8
+.source 1 u guint8
+.source 1 v guint8
+.source 1 a guint8
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.param 2 p5
+.temp 2 wy
+.temp 2 wu
+.temp 2 wv
+.temp 2 wr
+.temp 2 wg
+.temp 2 wb
+.temp 1 r
+.temp 1 g
+.temp 1 b
+.temp 1 da
+.temp 4 x
+.const 1 c128 128
+
+subb r, y, c128
+splatbw wy, r
+loadupdb r, u
+subb r, r, c128
+splatbw wu, r
+loadupdb r, v
+subb r, r, c128
+splatbw wv, r
+subb da, a, c128
+
+mulhsw wy, wy, p1
+
+mulhsw wr, wv, p2
+addw wr, wy, wr
+convssswb r, wr
+mergebw wr, r, da
+
+mulhsw wb, wu, p3
+addw wb, wy, wb
+convssswb b, wb
+
+mulhsw wg, wu, p4
+addw wg, wy, wg
+mulhsw wy, wv, p5
+addw wg, wg, wy
+
+convssswb g, wg
+
+mergebw wb, b, g
+mergewl x, wb, wr
+x4 addb argb, x, c128
+
 .function video_orc_matrix8
 .backup _custom_video_orc_matrix8
 .source 4 argb guint8