hacking

9 files changed, 386 insertions, 29 deletions

diff --git a/configure.ac b/configure.ac
index 48c83e3..c8a7d44 100644
--- a/configure.ac
+++ b/configure.ac
@@ -621,9 +621,9 @@ if test $enable_orc = yes ; then
       ORCC=orcc
    fi
    AC_SUBST(ORCC)
-   AC_DEFINE(USE_ORC, 1, [use Orc])
+   AC_DEFINE(HAVE_ORC, 1, [use Orc])
 fi
-AM_CONDITIONAL(USE_ORC, [test "x$enable_orc" = xyes])
+AM_CONDITIONAL(HAVE_ORC, [test "x$enable_orc" = xyes])
 
 DEP_CFLAGS="$DEP_CFLAGS $ORC_CFLAGS"
 DEP_LIBS="$DEP_LIBS $ORC_LIBS"
diff --git a/orc.mak b/orc.mak
new file mode 100644
index 0000000..b592f3f
--- /dev/null
+++ b/orc.mak
@@ -0,0 +1,73 @@
+#
+# This is a makefile.am fragment to build Orc code.
+#
+# Define ORC_SOURCE and then include this file, such as:
+#
+#  ORC_SOURCE=gstadderorc
+#  include $(top_srcdir)/common/orc.mak
+#
+# This fragment will create tmp-orc.c and gstadderorc.h from
+# gstadderorc.orc.
+#
+# When 'make dist' is run at the top level, or 'make orc-update'
+# in a directory including this fragment, the generated source 
+# files will be copied to $(ORC_SOURCE)-dist.[ch].  These files
+# should be checked in to git, since they are used if Orc is
+# disabled.
+# 
+# Note that this file defines BUILT_SOURCES, so any later usage
+# of BUILT_SOURCES in the Makefile.am that includes this file
+# must use '+='.
+#
+
+
+EXTRA_DIST = $(ORC_SOURCE).orc
+
+ORC_NODIST_SOURCES = tmp-orc.c $(ORC_SOURCE).h
+BUILT_SOURCES = tmp-orc.c $(ORC_SOURCE).h
+
+
+orc-update: tmp-orc.c $(ORC_SOURCE).h
+	$(top_srcdir)/common/gst-indent tmp-orc.c
+	cp tmp-orc.c $(srcdir)/$(ORC_SOURCE)-dist.c
+	cp $(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE)-dist.h
+
+orcc_v_gen = $(orcc_v_gen_$(V))
+orcc_v_gen_ = $(orcc_v_gen_$(AM_DEFAULT_VERBOSITY))
+orcc_v_gen_0 = @echo "  ORCC   $@";
+
+cp_v_gen = $(cp_v_gen_$(V))
+cp_v_gen_ = $(cp_v_gen_$(AM_DEFAULT_VERBOSITY))
+cp_v_gen_0 = @echo "  CP     $@";
+
+if HAVE_ORC
+tmp-orc.c: $(srcdir)/$(ORC_SOURCE).orc
+	$(orcc_v_gen)$(ORCC) --implementation -o tmp-orc.c $(srcdir)/$(ORC_SOURCE).orc
+
+$(ORC_SOURCE).h: $(srcdir)/$(ORC_SOURCE).orc
+	$(orcc_v_gen)$(ORCC) --header -o $(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE).orc
+else
+tmp-orc.c: $(srcdir)/$(ORC_SOURCE).orc
+	$(cp_v_gen)cp $(srcdir)/$(ORC_SOURCE)-dist.c tmp-orc.c
+
+$(ORC_SOURCE).h: $(srcdir)/$(ORC_SOURCE).orc
+	$(cp_v_gen)cp $(srcdir)/$(ORC_SOURCE)-dist.h $(ORC_SOURCE).h
+endif
+
+clean-local: clean-orc
+.PHONY: clean-orc
+clean-orc:
+	rm -f tmp-orc.c $(ORC_SOURCE).h
+
+dist-hook: dist-hook-orc
+.PHONY: dist-hook-orc
+dist-hook-orc: tmp-orc.c $(ORC_SOURCE).h
+	$(top_srcdir)/common/gst-indent tmp-orc.c
+	rm -f tmp-orc.c~
+	cmp -s tmp-orc.c $(srcdir)/$(ORC_SOURCE)-dist.c || \
+	  cp tmp-orc.c $(srcdir)/$(ORC_SOURCE)-dist.c
+	cmp -s $(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE)-dist.h || \
+	  cp $(ORC_SOURCE).h $(srcdir)/$(ORC_SOURCE)-dist.h
+	cp -p $(srcdir)/$(ORC_SOURCE)-dist.c $(distdir)/
+	cp -p $(srcdir)/$(ORC_SOURCE)-dist.h $(distdir)/
+
diff --git a/pixman/Makefile.am b/pixman/Makefile.am
index 3443003..03e5958 100644
--- a/pixman/Makefile.am
+++ b/pixman/Makefile.am
@@ -39,7 +39,10 @@ libpixmanincludedir = $(includedir)/pixman-1
 libpixmaninclude_HEADERS = pixman.h pixman-version.h
 noinst_LTLIBRARIES = 
 
-BUILT_SOURCES = pixman-combine32.h pixman-combine32.c pixman-combine64.h pixman-combine64.c
+ORC_SOURCE=pixman-orccode
+include $(top_srcdir)/orc.mak
+
+BUILT_SOURCES += pixman-combine32.h pixman-combine32.c pixman-combine64.h pixman-combine64.c
 
 pixman-combine32.c : pixman-combine.c.template pixman-combine32.h make-combine.pl
 	$(PERL) $(srcdir)/make-combine.pl 8 < $(srcdir)/pixman-combine.c.template > $@ || ($(RM) $@; exit 1)
@@ -51,7 +54,7 @@ pixman-combine64.c : pixman-combine.c.template pixman-combine64.h make-combine.p
 pixman-combine64.h : pixman-combine.h.template make-combine.pl
 	$(PERL) $(srcdir)/make-combine.pl 16 < $(srcdir)/pixman-combine.h.template > $@ || ($(RM) $@; exit 1)
 
-EXTRA_DIST = Makefile.win32 pixman-combine.c.template make-combine.pl pixman-region.c \
+EXTRA_DIST += Makefile.win32 pixman-combine.c.template make-combine.pl pixman-region.c \
 	pixman-combine.h.template solaris-hwcap.mapfile
 CLEANFILES = pixman-combine32.c pixman-combine64.c pixman-combine32.h pixman-combine64.h
 
@@ -125,29 +128,17 @@ ASM_CFLAGS_arm_neon=
 endif
 
 # orc code
-if USE_ORC
+if HAVE_ORC
+
 noinst_LTLIBRARIES += libpixman-orc.la
 libpixman_orc_la_SOURCES = \
 	pixman-orc.c
+nodist_libpixman_orc_la_SOURCES = $(ORC_NODIST_SOURCES)
 libpixman_orc_la_CFLAGS = $(DEP_CFLAGS) $(ORC_CFLAGS)
 libpixman_orc_la_LIBADD = $(DEP_LIBS)
 libpixman_1_la_LDFLAGS += $(ORC_LDFLAGS)
 libpixman_1_la_LIBADD += libpixman-orc.la
 
-$(libpixman_orc_la_SOURCES:.c=.s) : ASM_CFLAGS=
-
-nodist_libpixman_orc_la_SOURCES = pixman-orccode.c pixman-orccode.h
-BUILT_SOURCES += pixman-orccode.c pixman-orccode.h
-CLEANFILES += pixman-orccode.c pixman-orccode.h
-
-EXTRA_DIST += pixman-orccode.orc
-
-pixman-orccode.c: $(srcdir)/pixman-orccode.orc
-	$(ORCC) --implementation -o pixman-orccode.c $(srcdir)/pixman-orccode.orc
-
-pixman-orccode.h: $(srcdir)/pixman-orccode.orc
-	$(ORCC) --header -o pixman-orccode.h $(srcdir)/pixman-orccode.orc
-
 endif
 
 .c.s : $(libpixmaninclude_HEADERS) $(BUILT_SOURCES)
diff --git a/pixman/pixman-cpu.c b/pixman/pixman-cpu.c
index 17060b1..9112278 100644
--- a/pixman/pixman-cpu.c
+++ b/pixman/pixman-cpu.c
@@ -581,7 +581,7 @@ _pixman_choose_implementation (void)
     imp = _pixman_implementation_create_general();
     imp = _pixman_implementation_create_fast_path (imp);
     
-#ifdef USE_ORC
+#ifdef HAVE_ORC
     return _pixman_implementation_create_orc ();
 #endif
 #ifdef USE_MMX
diff --git a/pixman/pixman-orc.c b/pixman/pixman-orc.c
index 56c2042..58fa058 100644
--- a/pixman/pixman-orc.c
+++ b/pixman/pixman-orc.c
@@ -127,6 +127,36 @@ orc_combine_xor_u (pixman_implementation_t *imp,
   }
 }
 
+static void
+orc_combine_add_ca (pixman_implementation_t *imp,
+                    pixman_op_t              op,
+                    uint32_t *               dst,
+                    const uint32_t *         src,
+                    const uint32_t *         mask,
+                    int                      width)
+{
+  if (mask) {
+    orc_code_combine_add_ca (dst, src, mask, width);
+  } else {
+    orc_code_combine_add_ca_n (dst, src, width);
+  }
+}
+
+static void
+orc_combine_over_ca (pixman_implementation_t *imp,
+                    pixman_op_t              op,
+                    uint32_t *               dst,
+                    const uint32_t *         src,
+                    const uint32_t *         mask,
+                    int                      width)
+{
+  if (mask) {
+    orc_code_combine_over_ca (dst, src, mask, width);
+  } else {
+    orc_code_combine_over_ca_n (dst, src, width);
+  }
+}
+
 
 pixman_bool_t
 pixman_fill_orc (uint32_t *bits,
@@ -329,6 +359,46 @@ orc_composite_add_n_8_8 (pixman_implementation_t *imp,
     }
 }
 
+static void
+orc_composite_over_8888_8_8888 (pixman_implementation_t *imp,
+                                 pixman_op_t              op,
+                                 pixman_image_t *         src_image,
+                                 pixman_image_t *         mask_image,
+                                 pixman_image_t *         dst_image,
+                                 int32_t                  src_x,
+                                 int32_t                  src_y,
+                                 int32_t                  mask_x,
+                                 int32_t                  mask_y,
+                                 int32_t                  dest_x,
+                                 int32_t                  dest_y,
+                                 int32_t                  width,
+                                 int32_t                  height)
+{
+    uint32_t    *src, *src_line;
+    uint32_t    *dst, *dst_line;
+    uint8_t     *mask, *mask_line;
+    int src_stride, dst_stride, mask_stride;
+
+    PIXMAN_IMAGE_GET_LINE (
+	dst_image, dest_x, dest_y, uint32_t, dst_stride, dst_line, 1);
+    PIXMAN_IMAGE_GET_LINE (
+	src_image, src_x, src_y, uint32_t, src_stride, src_line, 1);
+    PIXMAN_IMAGE_GET_LINE (
+	mask_image, mask_x, mask_y, uint8_t, mask_stride, mask_line, 1);
+
+    while (height--) {
+        src = src_line;
+        src_line += src_stride;
+        dst = dst_line;
+        dst_line += dst_stride;
+        mask = mask_line;
+        mask_line += mask_stride;
+
+        orc_composite_over_8888_8_8888_line (dst, src, mask, width);
+    }
+}
+
+
 
 static const pixman_fast_path_t orc_fast_paths[] =
 {
@@ -349,7 +419,9 @@ static const pixman_fast_path_t orc_fast_paths[] =
     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_a8b8g8r8, orc_composite_over_n_8_8888       },
     { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_a8,       PIXMAN_x8b8g8r8, orc_composite_over_n_8_8888       },
     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_a8,       PIXMAN_x8r8g8b8, orc_composite_over_8888_8_8888    },
+#endif
     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_a8,       PIXMAN_a8r8g8b8, orc_composite_over_8888_8_8888    },
+#if 0
     { PIXMAN_OP_OVER, PIXMAN_a8b8g8r8, PIXMAN_a8,       PIXMAN_x8b8g8r8, orc_composite_over_8888_8_8888    },
     { PIXMAN_OP_OVER, PIXMAN_a8b8g8r8, PIXMAN_a8,       PIXMAN_a8b8g8r8, orc_composite_over_8888_8_8888    },
     { PIXMAN_OP_OVER, PIXMAN_x8r8g8b8, PIXMAN_a8,       PIXMAN_x8r8g8b8, orc_composite_over_x888_8_8888    },
@@ -520,7 +592,7 @@ _pixman_implementation_create_orc (void)
     //imp->combine_32[PIXMAN_OP_SATURATE] = orc_combine_saturate_u;
 
     //imp->combine_32_ca[PIXMAN_OP_SRC] = orc_combine_src_ca;
-    //imp->combine_32_ca[PIXMAN_OP_OVER] = orc_combine_over_ca;
+    imp->combine_32_ca[PIXMAN_OP_OVER] = orc_combine_over_ca;
     //imp->combine_32_ca[PIXMAN_OP_OVER_REVERSE] = orc_combine_over_reverse_ca;
     //imp->combine_32_ca[PIXMAN_OP_IN] = orc_combine_in_ca;
     //imp->combine_32_ca[PIXMAN_OP_IN_REVERSE] = orc_combine_in_reverse_ca;
@@ -529,7 +601,7 @@ _pixman_implementation_create_orc (void)
     //imp->combine_32_ca[PIXMAN_OP_ATOP] = orc_combine_atop_ca;
     //imp->combine_32_ca[PIXMAN_OP_ATOP_REVERSE] = orc_combine_atop_reverse_ca;
     //imp->combine_32_ca[PIXMAN_OP_XOR] = orc_combine_xor_ca;
-    //imp->combine_32_ca[PIXMAN_OP_ADD] = orc_combine_add_ca;
+    imp->combine_32_ca[PIXMAN_OP_ADD] = orc_combine_add_ca;
 
     imp->composite = orc_composite;
     imp->blt = orc_blt;
diff --git a/pixman/pixman-orccode.orc b/pixman/pixman-orccode.orc
index 0d4d850..d39c9a8 100644
--- a/pixman/pixman-orccode.orc
+++ b/pixman/pixman-orccode.orc
@@ -370,3 +370,110 @@ x4 addusb d1, t4, t5
 
 
 
+.function orc_code_combine_add_ca
+.dest 4 d1
+.source 4 s1
+.source 4 s2
+.temp 8 t1
+.temp 8 t2
+.temp 4 t3
+
+x4 convubw t1, s1
+x4 convubw t2, s2
+#splatw0q t2, t2
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+x4 convwb t3, t1
+x4 addusb d1, d1, t3
+
+
+.function orc_code_combine_add_ca_n
+.dest 4 d1
+.source 4 s1
+
+x4 addusb d1, d1, s1
+
+
+.function orc_code_combine_over_ca
+.dest 4 d1
+.source 4 s1
+.source 4 s2
+.temp 8 t1
+.temp 8 t2
+.temp 4 d
+.temp 8 d_wide
+.temp 8 m_wide
+.temp 8 s_wide
+.temp 8 xa
+.temp 4 s
+
+x4 convubw s_wide, s1
+x4 convubw m_wide, s2
+splatw0q xa, s_wide
+x4 mullw s_wide, s_wide, m_wide
+x4 div255w s_wide, s_wide
+x4 convwb s, s_wide
+x4 mullw m_wide, m_wide, xa
+x4 div255w m_wide, m_wide
+loadl d, d1
+x4 convubw d_wide, d
+x4 xorw m_wide, m_wide, 0x00ff
+x4 mullw t1, d_wide, m_wide
+x4 div255w t1, t1
+x4 convwb d, t1
+x4 addusb d1, d, s
+
+
+
+.function orc_code_combine_over_ca_n
+.dest 4 d1
+.source 4 s1
+.temp 8 t1
+.temp 8 t2
+.temp 4 d
+.temp 8 d_wide
+.temp 8 m_wide
+.temp 8 s_wide
+.temp 8 xa
+.temp 4 s
+
+x4 convubw s_wide, s1
+splatw0q xa, s_wide
+x4 convwb s, s_wide
+x4 copyw m_wide, xa
+loadl d, d1
+x4 convubw d_wide, d
+x4 xorw m_wide, m_wide, 0x00ff
+x4 mullw t1, d_wide, m_wide
+x4 div255w t1, t1
+x4 convwb d, t1
+x4 addusb d1, d, s
+
+
+
+.function orc_composite_over_8888_8_8888_line
+.dest 4 d1
+.source 4 s1
+.source 1 s2
+.temp 8 t1
+.temp 8 t2
+.temp 4 t3
+.temp 4 d
+.temp 4 mask
+.temp 8 d_wide
+
+x4 convubw t1, s1
+splatbl mask, s2
+x4 convubw t2, mask
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+x4 convwb t3, t1
+loadl d, d1
+x4 convubw d_wide, d
+x4 xorw t1, t1, 0x00ff
+splatw0q t2, t1
+x4 mullw t1, d_wide, t2
+x4 div255w t1, t1
+x4 convwb d, t1
+x4 addusb d1, d, t3
+
diff --git a/pixman/pixman-private.h b/pixman/pixman-private.h
index ed24d20..9887646 100644
--- a/pixman/pixman-private.h
+++ b/pixman/pixman-private.h
@@ -548,7 +548,7 @@ pixman_implementation_t *
 _pixman_implementation_create_vmx (pixman_implementation_t *fallback);
 #endif
 
-#ifdef USE_ORC
+#ifdef HAVE_ORC
 pixman_implementation_t *
 _pixman_implementation_create_orc (void);
 #endif
diff --git a/test/Makefile.am b/test/Makefile.am
index 6134920..80f78d7 100644
--- a/test/Makefile.am
+++ b/test/Makefile.am
@@ -30,7 +30,7 @@ blitters_test_SOURCES = blitters-test.c utils.c utils.h
 composite_traps_test_SOURCES = composite-traps-test.c utils.c utils.h
 
 ds_LDADD = $(TEST_LDADD)
-ds_SOURCES = ds.c ../pixman/pixman-orccode.c
+ds_SOURCES = ds.c ../pixman/tmp-orc.c
 
 scaling_test_SOURCES = scaling-test.c utils.c utils.h
 affine_test_SOURCES = affine-test.c utils.c utils.h
diff --git a/test/ds.c b/test/ds.c
index d16d98b..4fa3be3 100644
--- a/test/ds.c
+++ b/test/ds.c
@@ -164,6 +164,113 @@ combine_xor_u (pixman_implementation_t *imp,
     }
 }
 
+static void
+combine_mask_ca (uint32_t *src, uint32_t *mask)
+{
+    uint32_t a = *mask;
+
+    uint32_t x;
+    uint16_t xa;
+
+    if (!a)
+    {
+        *(src) = 0;
+        return;
+    }
+
+    x = *(src);
+    if (a == ~0)
+    {
+        x = x >> A_SHIFT;
+        x |= x << G_SHIFT;
+        x |= x << R_SHIFT;
+        *(mask) = x;
+        return;
+    }
+
+    xa = x >> A_SHIFT;
+    UN8x4_MUL_UN8x4 (x, a);
+    *(src) = x;
+
+    UN8x4_MUL_UN8 (a, xa);
+    *(mask) = a;
+}
+
+static void
+combine_mask_value_ca (uint32_t *src, const uint32_t *mask)
+{
+    uint32_t a = *mask;
+    uint32_t x;
+
+    if (!a)
+    {
+        *(src) = 0;
+        return;
+    }
+
+    if (a == ~0)
+        return;
+
+    x = *(src);
+    UN8x4_MUL_UN8x4 (x, a);
+    *(src) = x;
+}
+
+static void
+combine_add_ca (pixman_implementation_t *imp,
+                pixman_op_t              op,
+                uint32_t *                dest,
+                const uint32_t *          src,
+                const uint32_t *          mask,
+                int                      width)
+{
+    int i;
+
+    for (i = 0; i < width; ++i)
+    {
+        uint32_t s = *(src + i);
+        uint32_t m = *(mask + i);
+        uint32_t d = *(dest + i);
+
+        combine_mask_value_ca (&s, &m);
+
+        UN8x4_ADD_UN8x4 (d, s);
+
+        *(dest + i) = d;
+    }
+}
+
+static void
+combine_over_ca (pixman_implementation_t *imp,
+                 pixman_op_t              op,
+                 uint32_t *                dest,
+                 const uint32_t *          src,
+                 const uint32_t *          mask,
+                 int                      width)
+{
+    int i;
+
+    for (i = 0; i < width; ++i)
+    {
+        uint32_t s = *(src + i);
+        uint32_t m = *(mask + i);
+        uint32_t a;
+
+        combine_mask_ca (&s, &m);
+
+        a = ~m;
+        if (a)
+        {
+            uint32_t d = *(dest + i);
+            UN8x4_MUL_UN8x4_ADD_UN8x4 (d, a, s);
+            s = d;
+        }
+
+        *(dest + i) = s;
+    }
+}
+
+
 
 uint32_t dest_orig[256];
 uint32_t dest[256];
@@ -202,11 +309,15 @@ int main (int argc, char *argv[])
 
   for(i=0;i<256;i++){
     //src[i]=ORC_ARGB(i,i,0,0);
-    src[i]=rand_argb();
+    //src[i]=rand_argb();
+    src[i]=ORC_ARGB(i,i,i,i);
   }
 
   for(i=0;i<256;i++){
-    mask[i]=ORC_ARGB(i,0,0,0);
+    //mask[i]=ORC_ARGB(i,0,0,0);
+    //mask[i]=ORC_ARGB(i,i,i,i);
+    mask[i]=rand_argb();
+    //mask[i]=ORC_ARGB(255,255,255,255);
   }
 
   //combine_add_u (NULL, 0, dest_ref, src, mask, 256);
@@ -215,8 +326,8 @@ int main (int argc, char *argv[])
   //combine_over_u (NULL, 0, dest_ref, src, mask, 256);
   //orc_code_combine_over_u (dest, src, mask, 256);
 
-  combine_xor_u (NULL, 0, dest_ref, src, mask, 256);
-  orc_code_combine_xor_u (dest, src, mask, 256);
+  combine_over_ca (NULL, 0, dest_ref, src, mask, 256);
+  orc_code_combine_over_ca (dest, src, mask, 256);
 
 #if 0
   for(i=0;i<256;i++){
@@ -244,7 +355,7 @@ int main (int argc, char *argv[])
 #endif
 
   for(i=0;i<256;i++){
-    printf("%02x %02x %02x %02x  %02x %02x %02x %02x  %02x -> "
+    printf("%02x %02x %02x %02x  %02x %02x %02x %02x  %02x %02x %02x %02x -> "
         "%02x %02x %02x %02x  %02x %02x %02x %02x %s\n",
         ORC_ARGB_A(dest_orig[i]),
         ORC_ARGB_R(dest_orig[i]),
@@ -255,6 +366,9 @@ int main (int argc, char *argv[])
         ORC_ARGB_G(src[i]),
         ORC_ARGB_B(src[i]),
         ORC_ARGB_A(mask[i]),
+        ORC_ARGB_R(mask[i]),
+        ORC_ARGB_G(mask[i]),
+        ORC_ARGB_B(mask[i]),
         ORC_ARGB_A(dest_ref[i]),
         ORC_ARGB_R(dest_ref[i]),
         ORC_ARGB_G(dest_ref[i]),