video-converter: add more fastpaths for I420 -> RGB

Use the I420->BGRA and a new I420->ARGB to speed up any I420 to RGB
operation.

4 files changed, 648 insertions, 2 deletions

diff --git a/gst-libs/gst/video/video-converter.c b/gst-libs/gst/video/video-converter.c
index f29278d02..9aba42925 100644
--- a/gst-libs/gst/video/video-converter.c
+++ b/gst-libs/gst/video/video-converter.c
@@ -3599,9 +3599,93 @@ convert_I420_BGRA (GstVideoConverter * convert, const GstVideoFrame * src,
     sv = FRAME_GET_V_LINE (src, (i + convert->in_y) >> 1);
     sv += (convert->in_x >> 1);
 
+#if G_BYTE_ORDER == G_LITTLE_ENDIAN
     video_orc_convert_I420_BGRA (d, sy, su, sv,
         data->im[0][0], data->im[0][2],
         data->im[2][1], data->im[1][1], data->im[1][2], width);
+#else
+    video_orc_convert_I420_ARGB (d, sy, su, sv,
+        data->im[0][0], data->im[0][2],
+        data->im[2][1], data->im[1][1], data->im[1][2], width);
+#endif
+  }
+  convert_fill_border (convert, dest);
+}
+
+static void
+convert_I420_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;
+
+  for (i = 0; i < height; i++) {
+    guint8 *sy, *su, *sv, *d;
+
+    d = FRAME_GET_LINE (dest, i + convert->out_y);
+    d += (convert->out_x * 4);
+    sy = FRAME_GET_Y_LINE (src, i + convert->in_y);
+    sy += convert->in_x;
+    su = FRAME_GET_U_LINE (src, (i + convert->in_y) >> 1);
+    su += (convert->in_x >> 1);
+    sv = FRAME_GET_V_LINE (src, (i + convert->in_y) >> 1);
+    sv += (convert->in_x >> 1);
+
+#if G_BYTE_ORDER == G_LITTLE_ENDIAN
+    video_orc_convert_I420_ARGB (d, sy, su, sv,
+        data->im[0][0], data->im[0][2],
+        data->im[2][1], data->im[1][1], data->im[1][2], width);
+#else
+    video_orc_convert_I420_BGRA (d, sy, su, sv,
+        data->im[0][0], data->im[0][2],
+        data->im[2][1], data->im[1][1], data->im[1][2], width);
+#endif
+  }
+  convert_fill_border (convert, dest);
+}
+
+static void
+convert_I420_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;
+  gpointer tmp = convert->tmpline;
+  gpointer d[GST_VIDEO_MAX_PLANES];
+  gint pstride = GST_VIDEO_FORMAT_INFO_PSTRIDE (dest->info.finfo, 0);
+
+  d[0] = FRAME_GET_LINE (dest, 0);
+  d[0] = (guint8 *) d[0] + convert->out_x * pstride;
+
+  for (i = 0; i < height; i++) {
+    guint8 *sy, *su, *sv;
+
+    sy = FRAME_GET_Y_LINE (src, i + convert->in_y);
+    sy += convert->in_x;
+    su = FRAME_GET_U_LINE (src, (i + convert->in_y) >> 1);
+    su += (convert->in_x >> 1);
+    sv = FRAME_GET_V_LINE (src, (i + convert->in_y) >> 1);
+    sv += (convert->in_x >> 1);
+
+#if G_BYTE_ORDER == G_LITTLE_ENDIAN
+    video_orc_convert_I420_ARGB (tmp, sy, su, sv,
+        data->im[0][0], data->im[0][2],
+        data->im[2][1], data->im[1][1], data->im[1][2], width);
+#else
+    video_orc_convert_I420_BGRA (tmp, sy, su, sv,
+        data->im[0][0], data->im[0][2],
+        data->im[2][1], data->im[1][1], data->im[1][2], width);
+#endif
+    dest->info.finfo->pack_func (dest->info.finfo,
+        (GST_VIDEO_FRAME_IS_INTERLACED (dest) ?
+            GST_VIDEO_PACK_FLAG_INTERLACED :
+            GST_VIDEO_PACK_FLAG_NONE),
+        tmp, 0, d, dest->info.stride,
+        dest->info.chroma_site, i + convert->out_y, width);
   }
   convert_fill_border (convert, dest);
 }
@@ -4609,6 +4693,7 @@ static const VideoTransform transforms[] = {
       FALSE, FALSE, FALSE, 0, 0, convert_AYUV_ABGR},    /* alias */
   {GST_VIDEO_FORMAT_AYUV, GST_VIDEO_FORMAT_RGBx, TRUE, TRUE, TRUE, TRUE, TRUE,
       FALSE, FALSE, FALSE, 0, 0, convert_AYUV_RGBA},    /* alias */
+#endif
 
   {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_BGRA, FALSE, TRUE, TRUE, TRUE,
       TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_BGRA},
@@ -4618,7 +4703,57 @@ static const VideoTransform transforms[] = {
       TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_BGRA},
   {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_BGRx, FALSE, TRUE, TRUE, TRUE,
       TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_BGRA},
-#endif
+
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_ARGB, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_xRGB, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_ARGB, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_xRGB, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_ARGB},
+
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_ABGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_xBGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_RGBA, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_RGBx, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_RGB, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_BGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_RGB15, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_BGR15, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_RGB16, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_I420, GST_VIDEO_FORMAT_BGR16, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_ABGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_xBGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_RGBA, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_RGBx, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_RGB, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_BGR, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_RGB15, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_BGR15, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, GST_VIDEO_FORMAT_RGB16, FALSE, TRUE, TRUE, TRUE,
+      TRUE, FALSE, FALSE, FALSE, 0, 0, convert_I420_pack_ARGB},
+  {GST_VIDEO_FORMAT_YV12, 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,
diff --git a/gst-libs/gst/video/video-orc-dist.c b/gst-libs/gst/video/video-orc-dist.c
index fa57537d7..e59237b8a 100644
--- a/gst-libs/gst/video/video-orc-dist.c
+++ b/gst-libs/gst/video/video-orc-dist.c
@@ -333,6 +333,10 @@ 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_matrix8 (guint8 * ORC_RESTRICT d1,
     const guint8 * ORC_RESTRICT s1, orc_int64 p1, orc_int64 p2, orc_int64 p3,
     orc_int64 p4, int n);
@@ -22026,6 +22030,457 @@ video_orc_convert_I420_BGRA (guint8 * ORC_RESTRICT d1,
 #endif
 
 
+/* video_orc_convert_I420_ARGB */
+#ifdef DISABLE_ORC
+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)
+{
+  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 var42;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var43;
+#else
+  orc_int8 var43;
+#endif
+  orc_union16 var44;
+  orc_union16 var45;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var46;
+#else
+  orc_int8 var46;
+#endif
+  orc_union16 var47;
+  orc_union16 var48;
+  orc_union16 var49;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_union32 var50;
+#else
+  orc_union32 var50;
+#endif
+  orc_union32 var51;
+  orc_int8 var52;
+  orc_union16 var53;
+  orc_int8 var54;
+  orc_int8 var55;
+  orc_union16 var56;
+  orc_int8 var57;
+  orc_int8 var58;
+  orc_union16 var59;
+  orc_union16 var60;
+  orc_union16 var61;
+  orc_union16 var62;
+  orc_int8 var63;
+  orc_union16 var64;
+  orc_union16 var65;
+  orc_union16 var66;
+  orc_int8 var67;
+  orc_union16 var68;
+  orc_union16 var69;
+  orc_union16 var70;
+  orc_union16 var71;
+  orc_int8 var72;
+  orc_union16 var73;
+  orc_union32 var74;
+
+  ptr0 = (orc_union32 *) d1;
+  ptr4 = (orc_int8 *) s1;
+  ptr5 = (orc_int8 *) s2;
+  ptr6 = (orc_int8 *) s3;
+
+  /* 1: loadpb */
+  var43 = (int) 0x00000080;     /* 128 or 6.32404e-322f */
+  /* 10: loadpw */
+  var44.i = p1;
+  /* 12: loadpw */
+  var45.i = p2;
+  /* 16: loadpb */
+  var46 = (int) 0x0000007f;     /* 127 or 6.27463e-322f */
+  /* 18: loadpw */
+  var47.i = p3;
+  /* 22: loadpw */
+  var48.i = p4;
+  /* 25: loadpw */
+  var49.i = p5;
+  /* 31: loadpb */
+  var50.x4[0] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+  var50.x4[1] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+  var50.x4[2] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+  var50.x4[3] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var42 = ptr4[i];
+    /* 2: subb */
+    var52 = var42 - var43;
+    /* 3: splatbw */
+    var53.i = ((var52 & 0xff) << 8) | (var52 & 0xff);
+    /* 4: loadupdb */
+    var54 = ptr5[i >> 1];
+    /* 5: subb */
+    var55 = var54 - var43;
+    /* 6: splatbw */
+    var56.i = ((var55 & 0xff) << 8) | (var55 & 0xff);
+    /* 7: loadupdb */
+    var57 = ptr6[i >> 1];
+    /* 8: subb */
+    var58 = var57 - var43;
+    /* 9: splatbw */
+    var59.i = ((var58 & 0xff) << 8) | (var58 & 0xff);
+    /* 11: mulhsw */
+    var60.i = (var53.i * var44.i) >> 16;
+    /* 13: mulhsw */
+    var61.i = (var59.i * var45.i) >> 16;
+    /* 14: addw */
+    var62.i = var60.i + var61.i;
+    /* 15: convssswb */
+    var63 = ORC_CLAMP_SB (var62.i);
+    /* 17: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var46;
+      _dest.x2[1] = var63;
+      var64.i = _dest.i;
+    }
+    /* 19: mulhsw */
+    var65.i = (var56.i * var47.i) >> 16;
+    /* 20: addw */
+    var66.i = var60.i + var65.i;
+    /* 21: convssswb */
+    var67 = ORC_CLAMP_SB (var66.i);
+    /* 23: mulhsw */
+    var68.i = (var56.i * var48.i) >> 16;
+    /* 24: addw */
+    var69.i = var60.i + var68.i;
+    /* 26: mulhsw */
+    var70.i = (var59.i * var49.i) >> 16;
+    /* 27: addw */
+    var71.i = var69.i + var70.i;
+    /* 28: convssswb */
+    var72 = ORC_CLAMP_SB (var71.i);
+    /* 29: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var72;
+      _dest.x2[1] = var67;
+      var73.i = _dest.i;
+    }
+    /* 30: mergewl */
+    {
+      orc_union32 _dest;
+      _dest.x2[0] = var64.i;
+      _dest.x2[1] = var73.i;
+      var74.i = _dest.i;
+    }
+    /* 32: addb */
+    var51.x4[0] = var74.x4[0] + var50.x4[0];
+    var51.x4[1] = var74.x4[1] + var50.x4[1];
+    var51.x4[2] = var74.x4[2] + var50.x4[2];
+    var51.x4[3] = var74.x4[3] + var50.x4[3];
+    /* 33: storel */
+    ptr0[i] = var51;
+  }
+
+}
+
+#else
+static void
+_backup_video_orc_convert_I420_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;
+  orc_int8 var42;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var43;
+#else
+  orc_int8 var43;
+#endif
+  orc_union16 var44;
+  orc_union16 var45;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_int8 var46;
+#else
+  orc_int8 var46;
+#endif
+  orc_union16 var47;
+  orc_union16 var48;
+  orc_union16 var49;
+#if defined(__APPLE__) && __GNUC__ == 4 && __GNUC_MINOR__ == 2 && defined (__i386__)
+  volatile orc_union32 var50;
+#else
+  orc_union32 var50;
+#endif
+  orc_union32 var51;
+  orc_int8 var52;
+  orc_union16 var53;
+  orc_int8 var54;
+  orc_int8 var55;
+  orc_union16 var56;
+  orc_int8 var57;
+  orc_int8 var58;
+  orc_union16 var59;
+  orc_union16 var60;
+  orc_union16 var61;
+  orc_union16 var62;
+  orc_int8 var63;
+  orc_union16 var64;
+  orc_union16 var65;
+  orc_union16 var66;
+  orc_int8 var67;
+  orc_union16 var68;
+  orc_union16 var69;
+  orc_union16 var70;
+  orc_union16 var71;
+  orc_int8 var72;
+  orc_union16 var73;
+  orc_union32 var74;
+
+  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 */
+  var43 = (int) 0x00000080;     /* 128 or 6.32404e-322f */
+  /* 10: loadpw */
+  var44.i = ex->params[24];
+  /* 12: loadpw */
+  var45.i = ex->params[25];
+  /* 16: loadpb */
+  var46 = (int) 0x0000007f;     /* 127 or 6.27463e-322f */
+  /* 18: loadpw */
+  var47.i = ex->params[26];
+  /* 22: loadpw */
+  var48.i = ex->params[27];
+  /* 25: loadpw */
+  var49.i = ex->params[28];
+  /* 31: loadpb */
+  var50.x4[0] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+  var50.x4[1] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+  var50.x4[2] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+  var50.x4[3] = (int) 0x00000080;       /* 128 or 6.32404e-322f */
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var42 = ptr4[i];
+    /* 2: subb */
+    var52 = var42 - var43;
+    /* 3: splatbw */
+    var53.i = ((var52 & 0xff) << 8) | (var52 & 0xff);
+    /* 4: loadupdb */
+    var54 = ptr5[i >> 1];
+    /* 5: subb */
+    var55 = var54 - var43;
+    /* 6: splatbw */
+    var56.i = ((var55 & 0xff) << 8) | (var55 & 0xff);
+    /* 7: loadupdb */
+    var57 = ptr6[i >> 1];
+    /* 8: subb */
+    var58 = var57 - var43;
+    /* 9: splatbw */
+    var59.i = ((var58 & 0xff) << 8) | (var58 & 0xff);
+    /* 11: mulhsw */
+    var60.i = (var53.i * var44.i) >> 16;
+    /* 13: mulhsw */
+    var61.i = (var59.i * var45.i) >> 16;
+    /* 14: addw */
+    var62.i = var60.i + var61.i;
+    /* 15: convssswb */
+    var63 = ORC_CLAMP_SB (var62.i);
+    /* 17: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var46;
+      _dest.x2[1] = var63;
+      var64.i = _dest.i;
+    }
+    /* 19: mulhsw */
+    var65.i = (var56.i * var47.i) >> 16;
+    /* 20: addw */
+    var66.i = var60.i + var65.i;
+    /* 21: convssswb */
+    var67 = ORC_CLAMP_SB (var66.i);
+    /* 23: mulhsw */
+    var68.i = (var56.i * var48.i) >> 16;
+    /* 24: addw */
+    var69.i = var60.i + var68.i;
+    /* 26: mulhsw */
+    var70.i = (var59.i * var49.i) >> 16;
+    /* 27: addw */
+    var71.i = var69.i + var70.i;
+    /* 28: convssswb */
+    var72 = ORC_CLAMP_SB (var71.i);
+    /* 29: mergebw */
+    {
+      orc_union16 _dest;
+      _dest.x2[0] = var72;
+      _dest.x2[1] = var67;
+      var73.i = _dest.i;
+    }
+    /* 30: mergewl */
+    {
+      orc_union32 _dest;
+      _dest.x2[0] = var64.i;
+      _dest.x2[1] = var73.i;
+      var74.i = _dest.i;
+    }
+    /* 32: addb */
+    var51.x4[0] = var74.x4[0] + var50.x4[0];
+    var51.x4[1] = var74.x4[1] + var50.x4[1];
+    var51.x4[2] = var74.x4[2] + var50.x4[2];
+    var51.x4[3] = var74.x4[3] + var50.x4[3];
+    /* 33: storel */
+    ptr0[i] = var51;
+  }
+
+}
+
+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)
+{
+  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, 73, 52, 50, 48, 95, 65, 82, 71, 66, 11, 4,
+        4, 12, 1, 1, 12, 1, 1, 12, 1, 1, 14, 1, 128, 0, 0, 0,
+        14, 1, 127, 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, 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, 90,
+        32, 32, 24, 90, 35, 34, 25, 70, 35, 32, 35, 159, 38, 35, 196, 35,
+        17, 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, 41, 35, 37, 21, 2, 33, 0, 41, 16, 2, 0,
+
+      };
+      p = orc_program_new_from_static_bytecode (bc);
+      orc_program_set_backup_function (p, _backup_video_orc_convert_I420_ARGB);
+#else
+      p = orc_program_new ();
+      orc_program_set_name (p, "video_orc_convert_I420_ARGB");
+      orc_program_set_backup_function (p, _backup_video_orc_convert_I420_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_constant (p, 1, 0x00000080, "c1");
+      orc_program_add_constant (p, 1, 0x0000007f, "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_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, 4, "t10");
+
+      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, "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_C2, 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_T10, ORC_VAR_T4,
+          ORC_VAR_T6, ORC_VAR_D1);
+      orc_program_append_2 (p, "addb", 2, ORC_VAR_D1, ORC_VAR_T10, 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->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 c2fef7d7e..d6a2aa664 100644
--- a/gst-libs/gst/video/video-orc-dist.h
+++ b/gst-libs/gst/video/video-orc-dist.h
@@ -177,6 +177,7 @@ void video_orc_convert_AYUV_BGRA (guint8 * ORC_RESTRICT d1, int d1_stride, const
 void video_orc_convert_AYUV_ABGR (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_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_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);
diff --git a/gst-libs/gst/video/video-orc.orc b/gst-libs/gst/video/video-orc.orc
index 60d7c95d1..5af6c90ab 100644
--- a/gst-libs/gst/video/video-orc.orc
+++ b/gst-libs/gst/video/video-orc.orc
@@ -1663,6 +1663,7 @@ x4 addb argb, x, c128
 .temp 1 b
 .temp 4 x
 .const 1 c128 128
+.const 4 c4128 128
 
 subb r, y, c128
 splatbw wy, r
@@ -1693,7 +1694,61 @@ convssswb g, wg
 
 mergebw wb, b, g
 mergewl x, wb, wr
-x4 addb argb, x, c128
+x4 addb argb, x, c4128
+
+.function video_orc_convert_I420_ARGB
+.dest 4 argb guint8
+.source 1 y guint8
+.source 1 u guint8
+.source 1 v 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 4 x
+.const 1 c128 128
+.const 4 c4128 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
+
+mulhsw wy, wy, p1
+
+mulhsw wr, wv, p2
+addw wr, wy, wr
+convssswb r, wr
+mergebw wr, 127, 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, c4128
 
 .function video_orc_matrix8
 .backup _custom_video_orc_matrix8