some experimentsorc-0.5-bits

4 files changed, 121 insertions, 23 deletions

diff --git a/gst-libs/gst/audio/gstaudiopack.orc b/gst-libs/gst/audio/gstaudiopack.orc
index c13a688ce..e7bf3aa79 100644
--- a/gst-libs/gst/audio/gstaudiopack.orc
+++ b/gst-libs/gst/audio/gstaudiopack.orc
@@ -423,3 +423,59 @@ divd d1, t1, 2147483648.0L
 muld t1, s1, 2147483648.0L
 convdl d1, t1
 
+.function audio_orc_v_linear_s16
+.dest 2 d gint16
+.source 2 s1 gint16
+.source 2 s2 gint16
+.param 2 p1
+.param 2 p2
+.temp 2 t1
+.temp 4 t2
+.temp 4 t3
+
+mulswl t1, s1, p1
+mulswl t2, s2, p2
+addl t1, t1, t2
+addl t1, t1, 16384
+shrsl t1, t1, 15
+convlw d, t1
+
+.function audio_orc_v_linear_s32
+.dest 4 d gint32
+.source 4 s1 gint32
+.source 4 s2 gint32
+.param 4 p1
+.param 4 p2
+.temp 8 t1
+.temp 8 t2
+
+mulslq t1, s1, p1
+mulslq t2, s2, p2
+addq t1, t1, t2
+addq t1, t1, 1073741824
+shrsq t1, t1, 15
+convql d, t1
+
+
+.function audio_orc_v_linear_f32
+.dest 4 d gfloat
+.source 4 s1 gfloat
+.source 4 s2 gfloat
+.param 4 p1
+.temp 4 t1
+
+subf t1, s1, s2
+mulf t1, t1, p1
+addf d, t1, s2
+
+.function audio_orc_v_linear_f64
+.dest 8 d gdouble
+.source 8 s1 gdouble
+.source 8 s2 gdouble
+.param 8 p1
+.temp 8 t1
+
+subd t1, s1, s2
+muld t1, t1, p1
+addd d, t1, s2
+
diff --git a/gst-libs/gst/video/video-converter.c b/gst-libs/gst/video/video-converter.c
index f29278d02..9f036ce01 100644
--- a/gst-libs/gst/video/video-converter.c
+++ b/gst-libs/gst/video/video-converter.c
@@ -937,17 +937,17 @@ videoconvert_convert_init_tables (MatrixData * data)
     gint64 r = 0, g = 0, b = 0;
 
     for (j = 0; j < 3; j++) {
-      r = (r << 16) + data->im[j][0] * i;
-      g = (g << 16) + data->im[j][1] * i;
-      b = (b << 16) + data->im[j][2] * i;
+      r = (r << 16) + data->im[2 - j][0] * i;
+      g = (g << 16) + data->im[2 - j][1] * i;
+      b = (b << 16) + data->im[2 - j][2] * i;
     }
-    data->t_r[i] = r;
-    data->t_g[i] = g;
-    data->t_b[i] = b;
+    data->t_r[i] = r << 16;
+    data->t_g[i] = g << 16;
+    data->t_b[i] = b << 16;
   }
-  data->t_c = ((gint64) data->im[0][3] << 32)
-      + ((gint64) data->im[1][3] << 16)
-      + ((gint64) data->im[2][3] << 0);
+  data->t_c = ((gint64) data->im[2][3] << 48)
+      + ((gint64) data->im[1][3] << 32)
+      + ((gint64) data->im[0][3] << 16);
 }
 
 void
@@ -1000,6 +1000,10 @@ video_converter_matrix8 (MatrixData * data, gpointer pixels)
 static void
 video_converter_matrix8_table (MatrixData * data, gpointer pixels)
 {
+#if 1
+  video_orc_matrix8_table (pixels, pixels, data->t_r, data->t_g, data->t_b,
+      data->t_c, data->width);
+#else
   gint i, width = data->width * 4;
   guint8 r, g, b;
   gint64 c = data->t_c;
@@ -1013,10 +1017,11 @@ video_converter_matrix8_table (MatrixData * data, gpointer pixels)
 
     x = data->t_r[r] + data->t_g[g] + data->t_b[b] + c;
 
-    p[i + 1] = x >> (32 + SCALE);
-    p[i + 2] = x >> (16 + SCALE);
-    p[i + 3] = x >> (0 + SCALE);
+    p[i + 1] = x >> (16 + SCALE);
+    p[i + 2] = x >> (32 + SCALE);
+    p[i + 3] = x >> (48 + SCALE);
   }
+#endif
 }
 
 static void
diff --git a/gst-libs/gst/video/video-orc.orc b/gst-libs/gst/video/video-orc.orc
index 60d7c95d1..185e0403b 100644
--- a/gst-libs/gst/video/video-orc.orc
+++ b/gst-libs/gst/video/video-orc.orc
@@ -908,7 +908,7 @@ ldreslinl d1, s1, p1, p2
 .dest 1 d1
 .source 1 s1
 .source 1 s2
-.param 1 p1
+.param 2 p1
 .temp 2 t1
 .temp 2 t2
 .temp 1 a
@@ -1663,6 +1663,7 @@ x4 addb argb, x, c128
 .temp 1 b
 .temp 4 x
 .const 1 c128 128
+.const 1 c1284 128
 
 subb r, y, c128
 splatbw wy, r
@@ -1693,7 +1694,7 @@ convssswb g, wg
 
 mergebw wb, b, g
 mergewl x, wb, wr
-x4 addb argb, x, c128
+x4 addb argb, x, c1284
 
 .function video_orc_matrix8
 .backup _custom_video_orc_matrix8
@@ -1751,6 +1752,41 @@ x4 addw aq, aq, q1
 x4 convssswb ayuv2, aq
 x4 addb ayuv, ayuv2, c128
 
+.function video_orc_matrix8_table
+.dest 4 d1
+.source 4 s1
+.source 8 t_r gint64
+.source 8 t_g gint64
+.source 8 t_b gint64
+.param 8 c
+.temp 2 tw1
+.temp 2 tw2
+.temp 1 tba
+.temp 1 tbr
+.temp 1 tbg
+.temp 1 tbb
+.temp 8 tq1
+.temp 8 tq2
+.temp 8 tq3
+.temp 4 tla
+.temp 4 tl1
+
+andl tla, s1, 0xff
+select0lw tw1, s1
+select1wb tbr, tw1
+select1lw tw2, s1
+select0wb tbg, tw2
+select1wb tbb, tw2
+transbq tq1, tbr, t_r
+transbq tq2, tbg, t_g
+transbq tq3, tbb, t_b
+addq tq1, tq1, tq2
+addq tq1, tq1, tq3
+addq tq1, tq1, c
+x4 convhwb tl1, tq1
+orl d1, tl1, tla
+
+
 #.function video_orc_resample_h_near_u32
 #.source 4 src guint32
 #.source 4 idx
@@ -1782,7 +1818,8 @@ ldreslinb d1, s1, p1, p2
 .param 4 p1
 .param 4 p2
 
-ldreslinl d1, s1, p1, p2
+#ldreslinl d1, s1, p1, p2
+x4 ldreslinb d1, s1, p1, p2
 
 .function video_orc_resample_h_2tap_u8_lq
 .source 1 s1 guint8
@@ -2154,7 +2191,7 @@ addl d, d, t1
 
 .function video_orc_resample_v_multaps_u16
 .source 2 s guint16
-.param 2 t gint16
+.param 4 t gint32
 .dest 4 d gint32
 .temp 4 l1
 
diff --git a/tests/check/Makefile.am b/tests/check/Makefile.am
index ed1d02654..245555480 100644
--- a/tests/check/Makefile.am
+++ b/tests/check/Makefile.am
@@ -658,25 +658,25 @@ pipelines_simple_launch_lines_LDADD = \
 	$(GST_BASE_LIBS) \
 	$(LDADD)
 
-orc_video_CFLAGS = $(ORC_CFLAGS)
-orc_video_LDADD = $(ORC_LIBS) -lorc-test-0.4
+orc_video_CFLAGS = $(ORC_CFLAGS) $(GLIB_CFLAGS)
+orc_video_LDADD = $(ORC_LIBS) -lorc-test-0.5
 nodist_orc_video_SOURCES = orc/video.c
 orc_audio_CFLAGS = $(ORC_CFLAGS)
-orc_audio_LDADD = $(ORC_LIBS) -lorc-test-0.4
+orc_audio_LDADD = $(ORC_LIBS) -lorc-test-0.5
 nodist_orc_audio_SOURCES = orc/audio.c
 orc_adder_CFLAGS = $(ORC_CFLAGS)
-orc_adder_LDADD = $(ORC_LIBS) -lorc-test-0.4
+orc_adder_LDADD = $(ORC_LIBS) -lorc-test-0.5
 nodist_orc_adder_SOURCES = orc/adder.c
 orc_volume_CFLAGS = $(ORC_CFLAGS)
-orc_volume_LDADD = $(ORC_LIBS) -lorc-test-0.4
+orc_volume_LDADD = $(ORC_LIBS) -lorc-test-0.5
 nodist_orc_volume_SOURCES = orc/volume.c
 orc_videotestsrc_CFLAGS = $(ORC_CFLAGS)
-orc_videotestsrc_LDADD = $(ORC_LIBS) -lorc-test-0.4
+orc_videotestsrc_LDADD = $(ORC_LIBS) -lorc-test-0.5
 nodist_orc_videotestsrc_SOURCES = orc/videotestsrc.c
 
 orc/video.c: $(top_srcdir)/gst-libs/gst/video/video-orc.orc
 	$(MKDIR_P) orc/
-	$(ORCC) --test -o $@ $<
+	$(ORCC) --test --include glib.h -o $@ $<
 
 orc/audio.c: $(top_srcdir)/gst-libs/gst/audio/gstaudiopack.orc
 	$(MKDIR_P) orc/