cog: autogenerate code, colorspace improvements

configure.ac

@@ -681,6 +681,11 @@ AG_GST_CHECK_FEATURE(COG, [Cog plugin], cog, [
   ])
   AC_SUBST(COG_CFLAGS)
   AC_SUBST(COG_LIBS)
+  ORCC=`$PKG_CONFIG --variable=orcc orc-0.4`
+  if test "x$ORCC" = x ; then
+    ORCC=orcc
+  fi
+  AC_SUBST(ORCC)
 ])
 
 dnl *** dc1394  ***

ext/cog/Makefile.am

@@ -18,8 +18,6 @@ libgstcog_la_SOURCES = \
 	cog.h \
 	cogframe.c \
 	cogframe.h \
-	cogorc.c \
-	cogorc.h \
 	cogutils.h \
 	cogvirtframe.c \
 	cogvirtframe.h \
@@ -34,3 +32,14 @@ libgstcog_la_SOURCES = \
 	gstcolorconvert.c \
 	gstlogoinsert.c
 
+nodist_libgstcog_la_SOURCES = cogorc.c cogorc.h
+CLEANFILES = cogorc.c cogorc.h
+
+EXTRA_DIST=cog.orc
+
+cogorc.c: cog.orc
+	$(ORCC) --implementation -o cogorc.c cog.orc
+
+cogorc.h: cog.orc
+	$(ORCC) --header -o cogorc.h cog.orc
+

ext/cog/cog.orc

@@ -0,0 +1,376 @@
+
+.function cogorc_downsample_horiz_cosite_1tap
+.dest 1 d1
+.source 2 s1
+
+select0wb d1, s1
+
+
+.function cogorc_downsample_horiz_cosite_3tap
+.dest 1 d1
+.source 2 s1
+.source 2 s2
+.temp 1 t1
+.temp 1 t2
+.temp 1 t3
+.temp 2 t4
+.temp 2 t5
+.temp 2 t6
+
+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 t4, t4, t5
+addw t4, t4, 2
+shrsw t4, t4, 2
+convsuswb d1, t4
+
+
+.function cogorc_downsample_vert_halfsite_2tap
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+
+avgub d1, s1, s2
+
+
+.function cogorc_downsample_vert_cosite_3tap
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+.source 1 s3
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+
+convubw t1, s1
+convubw t2, s2
+convubw t3, s3
+mullw t2, t2, 2
+addw t1, t1, t3
+addw t1, t1, t2
+addw t1, t1, 2
+shrsw t1, t1, 2
+convsuswb d1, t1
+
+
+
+.function cogorc_downsample_vert_halfsite_4tap
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+.source 1 s3
+.source 1 s4
+.temp 2 t1
+.temp 2 t2
+.temp 2 t3
+.temp 2 t4
+
+convubw t1, s1
+convubw t2, s2
+convubw t3, s3
+convubw t4, s4
+addw t2, t2, t3
+mullw t2, t2, 26
+addw t1, t1, t4
+mullw t1, t1, 6
+addw t2, t2, t1
+addw t2, t2, 32
+shrsw t2, t2, 6
+convsuswb d1, t2
+
+
+.function cogorc_upsample_horiz_cosite_1tap
+.dest 2 d1 uint8_t
+.source 1 s1
+.temp 1 t1
+
+copyb t1, s1
+mergebw d1, t1, t1
+
+
+.function cogorc_upsample_horiz_cosite
+.dest 2 d1 uint8_t
+.source 1 s1
+.source 1 s2
+.temp 1 t1
+.temp 1 t2
+
+copyb t1, s1
+avgub t2, t1, s2
+mergebw d1, t1, t2
+
+
+.function cogorc_upsample_vert_avgub
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+
+avgub d1, s1, s2
+
+
+
+
+.function orc_unpack_yuyv_y
+.dest 1 d1
+.source 2 s1
+
+select0wb d1, s1
+
+
+.function orc_unpack_yuyv_u
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select0lw t1, s1
+select1wb d1, t1
+
+
+.function orc_unpack_yuyv_v
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select1lw t1, s1
+select1wb d1, t1
+
+
+.function orc_pack_yuyv
+.dest 4 d1
+.source 2 s1 uint8_t
+.source 1 s2
+.source 1 s3
+.temp 1 t1
+.temp 1 t2
+.temp 2 t3
+.temp 2 t4
+.temp 2 t5
+
+copyw t5, s1
+select0wb t1, t5
+select1wb t2, t5
+mergebw t3, t1, s2
+mergebw t4, t2, s3
+mergewl d1, t3, t4
+
+
+.function orc_unpack_uyvy_y
+.dest 1 d1
+.source 2 s1
+
+select1wb d1, s1
+
+
+.function orc_unpack_uyvy_u
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select0lw t1, s1
+select0wb d1, t1
+
+
+.function orc_unpack_uyvy_v
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select1lw t1, s1
+select0wb d1, t1
+
+
+.function orc_pack_uyvy
+.dest 4 d1
+.source 2 s1 uint8_t
+.source 1 s2
+.source 1 s3
+.temp 1 t1
+.temp 1 t2
+.temp 2 t3
+.temp 2 t4
+.temp 2 t5
+
+copyw t5, s1
+select0wb t1, t5
+select1wb t2, t5
+mergebw t3, s2, t1
+mergebw t4, s3, t2
+mergewl d1, t3, t4
+
+
+.function orc_memcpy
+.dest 1 d1 void
+.source 1 s1 void
+
+copyb d1, s1
+
+
+.function orc_addc_convert_u8_s16
+.dest 1 d1
+.source 2 s1 int16_t
+.temp 2 t1
+
+addw t1, s1, 128
+convsuswb d1, t1
+
+
+.function orc_subc_convert_s16_u8
+.dest 2 d1 int16_t
+.source 1 s1
+.temp 2 t1
+
+convubw t1, s1
+subw d1, t1, 128
+
+
+.function orc_splat_u8_ns
+.dest 1 d1
+.param 1 p1
+
+copyb d1, p1
+
+
+.function orc_splat_s16_ns
+.dest 2 d1 int16_t
+.param 2 p1
+
+copyw d1, p1
+
+
+.function orc_matrix2_u8
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t1, t1, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+addw t1, t1, p3
+shrsw t1, t1, 6
+convsuswb d1, t1
+
+
+.function orc_matrix3_u8
+.dest 1 d1 uint8_t
+.source 1 s1 uint8_t
+.source 1 s2 uint8_t
+.source 1 s3 uint8_t
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t1, t1, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+convubw t2, s3
+mullw t2, t2, p3
+addw t1, t1, t2
+addw t1, t1, p4
+shrsw t1, t1, 6
+convsuswb d1, t1
+
+
+
+.function orc_pack_123x
+.dest 4 d1 uint32_t
+.source 1 s1
+.source 1 s2
+.source 1 s3
+.param 1 p1
+.temp 2 t1
+.temp 2 t2
+
+mergebw t1, s1, s2
+mergebw t2, s3, p1
+mergewl d1, t1, t2
+
+
+.function orc_pack_x123
+.dest 4 d1 uint32_t
+.source 1 s1
+.source 1 s2
+.source 1 s3
+.param 1 p1
+.temp 2 t1
+.temp 2 t2
+
+mergebw t1, p1, s1
+mergebw t2, s2, s3
+mergewl d1, t1, t2
+
+
+.function cogorc_combine4_u8
+.dest 1 d1
+.source 1 s1
+.source 1 s2
+.source 1 s3
+.source 1 s4
+.param 2 p1
+.param 2 p2
+.param 2 p3
+.param 2 p4
+.temp 2 t1
+.temp 2 t2
+
+convubw t1, s1
+mullw t1, t1, p1
+convubw t2, s2
+mullw t2, t2, p2
+addw t1, t1, t2
+convubw t2, s3
+mullw t2, t2, p3
+addw t1, t1, t2
+convubw t2, s4
+mullw t2, t2, p4
+addw t1, t1, t2
+addw t1, t1, 32
+shrsw t1, t1, 6
+convsuswb d1, t1
+
+
+.function cogorc_unpack_ayuv_y
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select0lw t1, d1
+select1wb d1, s1
+
+
+.function cogorc_unpack_ayuv_u
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select1lw t1, d1
+select0wb d1, s1
+
+
+.function cogorc_unpack_ayuv_v
+.dest 1 d1
+.source 4 s1
+.temp 2 t1
+
+select1lw t1, d1
+select1wb d1, s1
+
+

ext/cog/cogframe.c

@@ -7,7 +7,7 @@
 #include <cog/cog.h>
 #include <cog/cogframe.h>
 #include <cog/cogvirtframe.h>
-#include <cog/cogorc.h>
+#include "cogorc.h"
 #include <gst/gst.h>
 
 #include <stdlib.h>

ext/cog/cogorc.h

@@ -1,36 +0,0 @@
-
-/* autogenerated from cog.orc */
-
-#ifndef _ORC_OUT_H_
-#define _ORC_OUT_H_
-
-void cogorc_downsample_horiz_cosite_3tap (uint8_t * d1, uint16_t * s1, uint16_t * s2, int n);
-void cogorc_downsample_vert_halfsite_2tap (uint8_t * d1, uint8_t * s1, uint8_t * s2, int n);
-void cogorc_downsample_vert_halfsite_3tap (uint8_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, int n);
-void cogorc_downsample_vert_halfsite_4tap (uint8_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, uint8_t * s4, int n);
-void cogorc_upsample_horiz_cosite (uint8_t * d1, uint8_t * s1, uint8_t * s2, int n);
-void cogorc_upsample_vert_avgub (uint8_t * d1, uint8_t * s1, uint8_t * s2, int n);
-void orc_unpack_yuyv_y (uint8_t * d1, uint16_t * s1, int n);
-void orc_unpack_yuyv_u (uint8_t * d1, uint32_t * s1, int n);
-void orc_unpack_yuyv_v (uint8_t * d1, uint32_t * s1, int n);
-void orc_pack_yuyv (uint32_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, int n);
-void orc_unpack_uyvy_y (uint8_t * d1, uint16_t * s1, int n);
-void orc_unpack_uyvy_u (uint8_t * d1, uint32_t * s1, int n);
-void orc_unpack_uyvy_v (uint8_t * d1, uint32_t * s1, int n);
-void orc_pack_uyvy (uint32_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, int n);
-void orc_memcpy (void * d1, void * s1, int n);
-void orc_addc_convert_u8_s16 (uint8_t * d1, int16_t * s1, int n);
-void orc_subc_convert_s16_u8 (int16_t * d1, uint8_t * s1, int n);
-void orc_splat_u8_ns (uint8_t * d1, int p1, int n);
-void orc_splat_s16_ns (int16_t * d1, int p1, int n);
-void orc_matrix2_u8 (uint8_t * d1, uint8_t * s1, uint8_t * s2, int p1, int p2, int p3, int n);
-void orc_matrix3_u8 (uint8_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, int p1, int p2, int p3, int p4, int n);
-void orc_pack_123x (uint32_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, int p1, int n);
-void orc_pack_x123 (uint32_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, int p1, int n);
-void cogorc_combine4_u8 (uint8_t * d1, uint8_t * s1, uint8_t * s2, uint8_t * s3, uint8_t * s4, int p1, int p2, int p3, int p4, int n);
-void cogorc_unpack_ayuv_y (uint8_t * d1, uint32_t * s1, int n);
-void cogorc_unpack_ayuv_u (uint8_t * d1, uint32_t * s1, int n);
-void cogorc_unpack_ayuv_v (uint8_t * d1, uint32_t * s1, int n);
-
-#endif
-

ext/cog/cogvirtframe.c

@@ -6,13 +6,14 @@
 #define COG_ENABLE_UNSTABLE_API 1
 
 #include <cog/cogvirtframe.h>
-#include <cog/cogorc.h>
 #include <cog/cog.h>
 #include <string.h>
 #include <math.h>
 #include <orc/orc.h>
 #include <gst/gst.h>
 
+#include "cogorc.h"
+
 
 CogFrame *
 cog_frame_new_virtual (CogMemoryDomain * domain, CogFrameFormat format,
@@ -354,7 +355,7 @@ cog_virt_frame_render_downsample_vert_cosite (CogFrame * frame,
   src3 = cog_virt_frame_get_line (frame->virt_frame1, component,
       CLAMP (i * 2 + 1, 0, n_src - 1));
 
-  cogorc_downsample_vert_halfsite_3tap (dest, src1, src2, src3,
+  cogorc_downsample_vert_cosite_3tap (dest, src1, src2, src3,
       frame->components[component].width);
 }
 
@@ -1196,9 +1197,8 @@ convert_444_422 (CogFrame * frame, void *_dest, int component, int i)
   if (component == 0) {
     orc_memcpy (dest, src, frame->width);
   } else {
-    cogorc_downsample_horiz_cosite_3tap (dest + 1,
+    cogorc_downsample_horiz_cosite_1tap (dest + 1,
-        (uint16_t *) (src + 1), (uint16_t *) (src + 3),
+        (uint16_t *) (src + 2), frame->components[component].width - 1);
-        frame->components[component].width - 1);
 
     {
       int j;
@@ -1226,18 +1226,15 @@ convert_422_420 (CogFrame * frame, void *_dest, int component, int i)
     uint8_t *dest = _dest;
     uint8_t *src1;
     uint8_t *src2;
-    uint8_t *src3;
     int n_src;
 
     n_src = frame->virt_frame1->components[component].height;
     src1 = cog_virt_frame_get_line (frame->virt_frame1, component,
-        CLAMP (i * 2 - 1, 0, n_src - 1));
-    src2 = cog_virt_frame_get_line (frame->virt_frame1, component,
         CLAMP (i * 2 + 0, 0, n_src - 1));
-    src3 = cog_virt_frame_get_line (frame->virt_frame1, component,
+    src2 = cog_virt_frame_get_line (frame->virt_frame1, component,
         CLAMP (i * 2 + 1, 0, n_src - 1));
 
-    cogorc_downsample_vert_halfsite_3tap (dest, src1, src2, src3,
+    cogorc_downsample_vert_halfsite_2tap (dest, src1, src2,
         frame->components[component].width);
   }
 }
@@ -1255,8 +1252,12 @@ convert_422_444 (CogFrame * frame, void *_dest, int component, int i)
   if (component == 0) {
     orc_memcpy (dest, src, frame->width);
   } else {
+    cogorc_upsample_horiz_cosite_1tap (dest, src,
+        frame->components[component].width / 2 - 1);
+#if 0
     cogorc_upsample_horiz_cosite (dest, src, src + 1,
         frame->components[component].width / 2 - 1);
+#endif
     dest[frame->components[component].width - 2] =
         src[frame->components[component].width / 2 - 1];
     dest[frame->components[component].width - 1] =
@@ -1274,6 +1275,7 @@ convert_420_422 (CogFrame * frame, void *_dest, int component, int i)
     src = cog_virt_frame_get_line (frame->virt_frame1, component, i);
     orc_memcpy (dest, src, frame->components[component].width);
   } else {
+#if 0
     if ((i & 1) && i < frame->components[component].height - 1) {
       uint8_t *src2;
 
@@ -1286,6 +1288,10 @@ convert_420_422 (CogFrame * frame, void *_dest, int component, int i)
       src = cog_virt_frame_get_line (frame->virt_frame1, component, i >> 1);
       orc_memcpy (dest, src, frame->components[component].width);
     }
+#else
+    src = cog_virt_frame_get_line (frame->virt_frame1, component, i >> 1);
+    orc_memcpy (dest, src, frame->components[component].width);
+#endif
   }
 }
 

ext/cog/gstcogcolorspace.c

@@ -189,6 +189,8 @@ gst_cogcolorspace_class_init (gpointer g_class, gpointer class_data)
   base_transform_class->transform = gst_cogcolorspace_transform;
   base_transform_class->transform_caps = gst_cogcolorspace_transform_caps;
   base_transform_class->get_unit_size = gst_cogcolorspace_get_unit_size;
+
+  base_transform_class->passthrough_on_same_caps = TRUE;
 }
 
 static void