cog: implement chroma-site

Determine chroma site position from the caps and use it when
upsampling/downsampling chroma.

ext/cog/cog.orc

@@ -32,6 +32,69 @@ shrsw t4, t4, 2
 convsuswb d1, t4
 
 
+.function cogorc_downsample_420_jpeg
+.dest 1 d1
+.source 2 s1
+.source 2 s2
+.temp 2 t1
+.temp 1 t2
+.temp 1 t3
+.temp 1 t4
+.temp 1 t5
+
+copyw t1, s1
+select0wb t2, t1
+select1wb t3, t1
+avgub t2, t2, t3
+copyw t1, s2
+select0wb t4, t1
+select1wb t5, t1
+avgub t4, t4, t5
+avgub d1, t2, t4
+
+
+.function cogorc_downsample_420_mpeg2
+.dest 1 d1
+.source 2 s1
+.source 2 s2
+.source 2 s3
+.source 2 s4
+.temp 1 t1
+.temp 1 t2
+.temp 1 t3
+.temp 2 t4
+.temp 2 t5
+.temp 2 t6
+.temp 2 t7
+
+copyw t4, s1
+select0wb t1, t4
+select1wb t2, t4
+select0wb t3, s2
+convubw t4, t1
+convubw t5, t2
+convubw t6, t3
+mullw t5, t5, 2
+addw t4, t4, t6
+addw t7, t4, t5
+copyw t4, s3
+
+select0wb t1, t4
+select1wb t2, t4
+select0wb t3, s4
+convubw t4, t1
+convubw t5, t2
+convubw t6, t3
+mullw t5, t5, 2
+addw t4, t4, t6
+addw t4, t4, t5
+
+addw t7, t7, t4
+addw t7, t7, 4
+shrsw t7, t7, 3
+convsuswb d1, t7
+
+
 .function cogorc_downsample_vert_halfsite_2tap
 .dest 1 d1
 .source 1 s1

ext/cog/cogframe.c

@@ -864,7 +864,8 @@ cog_frame_convert (CogFrame * dest, CogFrame * src)
   }
 
   if ((dest_format & 3) != (frame->format & 3)) {
-    frame = cog_virt_frame_new_subsample (frame, dest_format, 2);
+    frame = cog_virt_frame_new_subsample (frame, dest_format,
+        COG_CHROMA_SITE_MPEG2, 2);
     GST_DEBUG ("subsample %p", frame);
   }
 

ext/cog/cogframe.h

@@ -19,6 +19,12 @@ typedef enum _CogColorMatrix {
   COG_COLOR_MATRIX_SDTV
 } CogColorMatrix;
 
+typedef enum _CogChromaSite {
+  COG_CHROMA_SITE_UNKNOWN = 0,
+  COG_CHROMA_SITE_MPEG2 = 1,
+  COG_CHROMA_SITE_JPEG
+} CogChromaSite;
+
 /* bit pattern:
  *  0x100 - 0: normal, 1: indirect (packed)
  *  0x001 - horizontal chroma subsampling: 0: 1, 1: 2

ext/cog/cogvirtframe.c

@@ -1524,6 +1524,9 @@ cog_virt_frame_new_color_matrix_YCbCr_to_YCbCr (CogFrame * vf,
 {
   CogFrame *virt_frame;
 
+  if (in_color_matrix == out_color_matrix)
+    return vf;
+
   virt_frame = cog_frame_new_virtual (NULL, COG_FRAME_FORMAT_U8_444,
       vf->width, vf->height);
   virt_frame->virt_frame1 = vf;
@@ -1593,8 +1596,132 @@ convert_422_420 (CogFrame * frame, void *_dest, int component, int i)
   }
 }
 
+static void
+convert_444_420_mpeg2 (CogFrame * frame, void *_dest, int component, int i)
+{
+  uint8_t *dest = _dest;
+  uint8_t *src;
+
+  if (component == 0) {
+    src = cog_virt_frame_get_line (frame->virt_frame1, component, i);
+    orc_memcpy (dest, src, frame->components[component].width);
+  } else {
+    uint8_t *dest = _dest;
+    uint8_t *src1;
+    uint8_t *src2;
+    int n_src;
+
+    n_src = frame->virt_frame1->components[component].height;
+    src1 = cog_virt_frame_get_line (frame->virt_frame1, component,
+        CLAMP (i * 2 + 0, 0, n_src - 1));
+    src2 = cog_virt_frame_get_line (frame->virt_frame1, component,
+        CLAMP (i * 2 + 1, 0, n_src - 1));
+
+    cogorc_downsample_420_mpeg2 (dest + 1,
+        (uint16_t *) src1, (uint16_t *) (src1 + 2),
+        (uint16_t *) src2, (uint16_t *) (src2 + 2),
+        frame->components[component].width - 1);
+    {
+      int j;
+      int x;
+
+      j = 0;
+      x = 1 * src1[CLAMP (j * 2 - 1, 0, n_src - 1)];
+      x += 2 * src1[CLAMP (j * 2 + 0, 0, n_src - 1)];
+      x += 1 * src1[CLAMP (j * 2 + 1, 0, n_src - 1)];
+      x += 1 * src2[CLAMP (j * 2 - 1, 0, n_src - 1)];
+      x += 2 * src2[CLAMP (j * 2 + 0, 0, n_src - 1)];
+      x += 1 * src2[CLAMP (j * 2 + 1, 0, n_src - 1)];
+      dest[j] = CLAMP ((x + 4) >> 3, 0, 255);
+    }
+  }
+}
+
+static void
+convert_444_420_jpeg (CogFrame * frame, void *_dest, int component, int i)
+{
+  uint8_t *dest = _dest;
+  uint8_t *src;
+
+  if (component == 0) {
+    src = cog_virt_frame_get_line (frame->virt_frame1, component, i);
+    orc_memcpy (dest, src, frame->components[component].width);
+  } else {
+    uint8_t *dest = _dest;
+    uint8_t *src1;
+    uint8_t *src2;
+    int n_src;
+
+    n_src = frame->virt_frame1->components[component].height;
+    src1 = cog_virt_frame_get_line (frame->virt_frame1, component,
+        CLAMP (i * 2 + 0, 0, n_src - 1));
+    src2 = cog_virt_frame_get_line (frame->virt_frame1, component,
+        CLAMP (i * 2 + 1, 0, n_src - 1));
+
+    cogorc_downsample_420_jpeg (dest,
+        (uint16_t *) src1, (uint16_t *) src2,
+        frame->components[component].width);
+  }
+}
+
 /* up */
 
+static void
+convert_420_444_mpeg2 (CogFrame * frame, void *_dest, int component, int i)
+{
+  uint8_t *dest = _dest;
+  uint8_t *src;
+  int n_taps = frame->param1;
+
+  if (component == 0) {
+    src = cog_virt_frame_get_line (frame->virt_frame1, component, i);
+
+    orc_memcpy (dest, src, frame->width);
+  } else {
+    src = cog_virt_frame_get_line (frame->virt_frame1, component, i / 2);
+
+    switch (n_taps) {
+      default:
+      case 2:
+        cogorc_upsample_horiz_cosite (dest, src, src + 1,
+            frame->components[component].width / 2 - 1);
+        break;
+    }
+    dest[frame->components[component].width - 2] =
+        src[frame->components[component].width / 2 - 1];
+    dest[frame->components[component].width - 1] =
+        src[frame->components[component].width / 2 - 1];
+  }
+}
+
+static void
+convert_420_444_jpeg (CogFrame * frame, void *_dest, int component, int i)
+{
+  uint8_t *dest = _dest;
+  uint8_t *src;
+  int n_taps = frame->param1;
+
+  if (component == 0) {
+    src = cog_virt_frame_get_line (frame->virt_frame1, component, i);
+
+    orc_memcpy (dest, src, frame->width);
+  } else {
+    src = cog_virt_frame_get_line (frame->virt_frame1, component, i / 2);
+
+    switch (n_taps) {
+      default:
+      case 1:
+        cogorc_upsample_horiz_cosite_1tap (dest, src,
+            frame->components[component].width / 2 - 1);
+        break;
+    }
+    dest[frame->components[component].width - 2] =
+        src[frame->components[component].width / 2 - 1];
+    dest[frame->components[component].width - 1] =
+        src[frame->components[component].width / 2 - 1];
+  }
+}
+
 static void
 convert_422_444 (CogFrame * frame, void *_dest, int component, int i)
 {
@@ -1609,10 +1736,6 @@ convert_422_444 (CogFrame * frame, void *_dest, int component, int i)
   } else {
     switch (n_taps) {
       default:
-      case 1:
-        cogorc_upsample_horiz_cosite_1tap (dest, src,
-            frame->components[component].width / 2 - 1);
-        break;
       case 2:
         cogorc_upsample_horiz_cosite (dest, src, src + 1,
             frame->components[component].width / 2 - 1);
@@ -1661,7 +1784,8 @@ convert_420_422 (CogFrame * frame, void *_dest, int component, int i)
 }
 
 CogFrame *
-cog_virt_frame_new_subsample (CogFrame * vf, CogFrameFormat format, int n_taps)
+cog_virt_frame_new_subsample (CogFrame * vf, CogFrameFormat format,
+    CogChromaSite chroma_site, int n_taps)
 {
   CogFrame *virt_frame;
   CogFrameRenderFunc render_line;
@@ -1674,14 +1798,11 @@ cog_virt_frame_new_subsample (CogFrame * vf, CogFrameFormat format, int n_taps)
     render_line = convert_422_420;
   } else if (vf->format == COG_FRAME_FORMAT_U8_444 &&
       format == COG_FRAME_FORMAT_U8_420) {
-    virt_frame = cog_frame_new_virtual (NULL, COG_FRAME_FORMAT_U8_422,
+    if (chroma_site == COG_CHROMA_SITE_MPEG2) {
-        vf->width, vf->height);
+      render_line = convert_444_420_mpeg2;
-    virt_frame->virt_frame1 = vf;
+    } else {
-    virt_frame->render_line = convert_444_422;
+      render_line = convert_444_420_jpeg;
-    virt_frame->param1 = n_taps;
+    }
-    vf = virt_frame;
-
-    render_line = convert_422_420;
   } else if (vf->format == COG_FRAME_FORMAT_U8_444 &&
       format == COG_FRAME_FORMAT_U8_422) {
     render_line = convert_444_422;
@@ -1690,14 +1811,11 @@ cog_virt_frame_new_subsample (CogFrame * vf, CogFrameFormat format, int n_taps)
     render_line = convert_420_422;
   } else if (vf->format == COG_FRAME_FORMAT_U8_420 &&
       format == COG_FRAME_FORMAT_U8_444) {
-    virt_frame = cog_frame_new_virtual (NULL, COG_FRAME_FORMAT_U8_422,
+    if (chroma_site == COG_CHROMA_SITE_MPEG2) {
-        vf->width, vf->height);
+      render_line = convert_420_444_mpeg2;
-    virt_frame->virt_frame1 = vf;
+    } else {
-    virt_frame->render_line = convert_420_422;
+      render_line = convert_420_444_jpeg;
-    virt_frame->param1 = n_taps;
+    }
-    vf = virt_frame;
-
-    render_line = convert_422_444;
   } else if (vf->format == COG_FRAME_FORMAT_U8_422 &&
       format == COG_FRAME_FORMAT_U8_444) {
     render_line = convert_422_444;

ext/cog/cogvirtframe.h

@@ -32,7 +32,8 @@ CogFrame * cog_virt_frame_new_color_matrix_RGB_to_YCbCr (CogFrame * vf, CogColor
 CogFrame * cog_virt_frame_new_color_matrix_YCbCr_to_YCbCr (CogFrame * vf,
     CogColorMatrix in_color_matrix, CogColorMatrix out_color_matrix,
     int bits);
-CogFrame *cog_virt_frame_new_subsample (CogFrame *vf, CogFrameFormat format, int n_taps);
+CogFrame *cog_virt_frame_new_subsample (CogFrame *vf, CogFrameFormat format,
+    CogChromaSite site, int n_taps);
 
 CogFrame * cog_virt_frame_new_convert_u8 (CogFrame *vf);
 CogFrame * cog_virt_frame_new_convert_s16 (CogFrame *vf);

ext/cog/gstcogcolorspace.c

@@ -275,6 +275,7 @@ gst_cogcolorspace_caps_remove_format_info (GstCaps * caps)
     gst_structure_remove_field (structure, "alpha_mask");
     gst_structure_remove_field (structure, "palette_data");
     gst_structure_remove_field (structure, "color-matrix");
+    gst_structure_remove_field (structure, "chroma-site");
   }
 
   gst_caps_do_simplify (caps);
@@ -365,6 +366,25 @@ gst_cogcolorspace_caps_get_color_matrix (GstCaps * caps)
   return COG_COLOR_MATRIX_SDTV;
 }
 
+static CogChromaSite
+gst_cogcolorspace_caps_get_chroma_site (GstCaps * caps)
+{
+  const char *s;
+
+  s = gst_video_parse_caps_chroma_site (caps);
+
+  if (s == NULL)
+    return COG_COLOR_MATRIX_SDTV;
+
+  if (strcmp (s, "jpeg") == 0) {
+    return COG_CHROMA_SITE_JPEG;
+  } else if (strcmp (s, "mpeg2") == 0) {
+    return COG_CHROMA_SITE_MPEG2;
+  }
+
+  return COG_CHROMA_SITE_MPEG2;
+}
+
 static GstFlowReturn
 gst_cogcolorspace_transform (GstBaseTransform * base_transform,
     GstBuffer * inbuf, GstBuffer * outbuf)
@@ -379,6 +399,8 @@ gst_cogcolorspace_transform (GstBaseTransform * base_transform,
   gboolean ret;
   CogColorMatrix in_color_matrix;
   CogColorMatrix out_color_matrix;
+  CogChromaSite in_chroma_site;
+  CogChromaSite out_chroma_site;
 
   g_return_val_if_fail (GST_IS_COGCOLORSPACE (base_transform), GST_FLOW_ERROR);
   compress = GST_COGCOLORSPACE (base_transform);
@@ -393,6 +415,9 @@ gst_cogcolorspace_transform (GstBaseTransform * base_transform,
   in_color_matrix = gst_cogcolorspace_caps_get_color_matrix (inbuf->caps);
   out_color_matrix = gst_cogcolorspace_caps_get_color_matrix (outbuf->caps);
 
+  in_chroma_site = gst_cogcolorspace_caps_get_chroma_site (inbuf->caps);
+  out_chroma_site = gst_cogcolorspace_caps_get_chroma_site (outbuf->caps);
+
   frame = gst_cog_buffer_wrap (gst_buffer_ref (inbuf),
       in_format, width, height);
   out_frame = gst_cog_buffer_wrap (gst_buffer_ref (outbuf),
@@ -431,22 +456,24 @@ gst_cogcolorspace_transform (GstBaseTransform * base_transform,
       gst_video_format_is_rgb (in_format)) {
     frame = cog_virt_frame_new_color_matrix_RGB_to_YCbCr (frame,
         out_color_matrix, 8);
+    frame = cog_virt_frame_new_subsample (frame, new_subsample,
+        out_chroma_site, (compress->quality >= 3) ? 2 : 1);
   }
 
   if (gst_video_format_is_yuv (out_format) &&
       gst_video_format_is_yuv (in_format) &&
-      in_color_matrix != out_color_matrix) {
+      (in_color_matrix != out_color_matrix ||
+          in_chroma_site != out_chroma_site)) {
     frame = cog_virt_frame_new_subsample (frame, COG_FRAME_FORMAT_U8_444,
-        (compress->quality >= 5) ? 8 : 6);
+        in_chroma_site, (compress->quality >= 5) ? 8 : 6);
     frame = cog_virt_frame_new_color_matrix_YCbCr_to_YCbCr (frame,
         in_color_matrix, out_color_matrix, 8);
   }
 
-  frame = cog_virt_frame_new_subsample (frame, new_subsample,
-      (compress->quality >= 3) ? 2 : 1);
-
   if (gst_video_format_is_rgb (out_format) &&
       gst_video_format_is_yuv (in_format)) {
+    frame = cog_virt_frame_new_subsample (frame, new_subsample,
+        in_chroma_site, (compress->quality >= 3) ? 2 : 1);
     frame = cog_virt_frame_new_color_matrix_YCbCr_to_RGB (frame,
         in_color_matrix, (compress->quality >= 5) ? 8 : 6);
   }

ext/cog/gstcolorconvert.c

@@ -238,18 +238,22 @@ gst_colorconvert_transform_ip (GstBaseTransform * base_transform,
       li->format, li->width, li->height);
 
   vf = cog_virt_frame_new_unpack (cog_frame_ref (frame));
-  vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_444, 2);
+  vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_444,
+      COG_CHROMA_SITE_MPEG2, 2);
   vf = cog_virt_frame_new_color_transform (vf);
   if (frame->format == COG_FRAME_FORMAT_YUYV) {
-    vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_422, 2);
+    vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_422,
+        COG_CHROMA_SITE_MPEG2, 2);
     vf = cog_virt_frame_new_pack_YUY2 (vf);
   } else if (frame->format == COG_FRAME_FORMAT_UYVY) {
-    vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_422, 2);
+    vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_422,
+        COG_CHROMA_SITE_MPEG2, 2);
     vf = cog_virt_frame_new_pack_UYVY (vf);
   } else if (frame->format == COG_FRAME_FORMAT_AYUV) {
     vf = cog_virt_frame_new_pack_AYUV (vf);
   } else if (frame->format == COG_FRAME_FORMAT_U8_420) {
-    vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_420, 2);
+    vf = cog_virt_frame_new_subsample (vf, COG_FRAME_FORMAT_U8_420,
+        COG_CHROMA_SITE_MPEG2, 2);
   } else {
     g_assert_not_reached ();
   }

ext/cog/gstlogoinsert.c

@@ -261,13 +261,15 @@ gst_logoinsert_transform_ip (GstBaseTransform * base_transform, GstBuffer * buf)
     CogFrame *f;
 
     f = cog_virt_frame_extract_alpha (cog_frame_ref (li->argb_frame));
-    f = cog_virt_frame_new_subsample (f, frame->format, 2);
+    f = cog_virt_frame_new_subsample (f, frame->format,
+        COG_CHROMA_SITE_MPEG2, 2);
     li->alpha_frame = cog_frame_realize (f);
 
     f = cog_virt_frame_new_unpack (cog_frame_ref (li->argb_frame));
     f = cog_virt_frame_new_color_matrix_RGB_to_YCbCr (f, COG_COLOR_MATRIX_SDTV,
         8);
-    f = cog_virt_frame_new_subsample (f, frame->format, 2);
+    f = cog_virt_frame_new_subsample (f, frame->format,
+        COG_CHROMA_SITE_MPEG2, 2);
     li->overlay_frame = cog_frame_realize (f);
   }