update for current orc

2 files changed, 302 insertions, 78 deletions

diff --git a/pixman/pixman-orccode.orc b/pixman/pixman-orccode.orc
index 969a334..0d4d850 100644
--- a/pixman/pixman-orccode.orc
+++ b/pixman/pixman-orccode.orc
@@ -50,7 +50,7 @@ copyl d1, s1
 .dest 4 d1
 .source 4 s1
 
-addusb4 d1, d1, s1
+x4 addusb d1, d1, s1
 
 
 .function orc_composite_add_8_8_line
@@ -64,133 +64,309 @@ addusb d1, d1, s1
 .dest 1 d1
 .source 1 s1
 .param 1 p1
-.temp 1 t1
+.temp 2 t1
+.temp 1 t2
 
-compina t1, p1, s1
-addusb d1, d1, t1
+#compina t1, p1, s1
+convubw t1, s1
+mullw t1, t1, p1
+div255w t1, t1
+convwb t2, t1
+addusb d1, d1, t2
 
 
 .function orc_code_combine_add_u
 .dest 4 d1
 .source 4 s1
 .source 4 s2
-.temp 4 t1
+.temp 8 t1
+.temp 8 t2
+.temp 4 t3
 
-compin t1, s1, s2
-addusb4 d1, d1, t1
+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_u_n
 .dest 4 d1
 .source 4 s1
 
-addusb4 d1, d1, s1
+x4 addusb d1, d1, s1
 
 
 .function orc_code_combine_over_u
 .dest 4 d1
 .source 4 s1
 .source 4 s2
-.temp 4 t1
+.temp 8 t1
+.temp 8 t2
+.temp 4 t3
+.temp 4 d
+.temp 8 d_wide
 
-compin t1, s1, s2
-compover d1, d1, t1
+#compin t1, s1, s2
+#compover d1, d1, t1
+x4 convubw t1, s1
+x4 convubw t2, s2
+splatw0q t2, t2
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+x4 convwb t3, t1
+# ((d) + (s) - ORC_MULDIV_255((d),(m)))
+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
 
 
 .function orc_code_combine_over_u_n
 .dest 4 d1
 .source 4 s1
-
-compover d1, d1, s1
+.temp 8 t1
+.temp 8 t2
+.temp 4 d
+.temp 4 s
+.temp 8 d_wide
+
+loadl s, s1
+x4 convubw t1, s
+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, s
 
 
 .function orc_code_combine_in_u
 .dest 4 d1
 .source 4 s1
 .source 4 s2
-#.temp 4 t1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
+.temp 8 t1
 .temp 8 t2
-.temp 8 t3
 
-#compin t1, s1, s2
-convubw4 t2, s1
-convubw4 t3, s2
-splat0w4 t3, t3
-mullw4 t2, t2, t3
-div255w4 t2, t2
-#convwb4 t1, t2
-#compin d1, t1, d1
-#convubw4 t2, t1
-convubw4 t3, d1
-splat0w4 t3, t3
-mullw4 t2, t2, t3
-div255w4 t2, t2
-convwb4 d1, t2
+x4 convubw t1, s1
+x4 convubw t2, s2
+splatw0q t2, t2
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+# ORC_MULDIV_255((s),(m)), m is from dest
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+x4 convwb d1, t1
+
 
 
 .function orc_code_combine_in_u_n
 .dest 4 d1
 .source 4 s1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
 .temp 8 t1
 .temp 8 t2
 
-convubw4 t1, s1
-convubw4 t2, d1
-splat0w4 t2, t2
-mullw4 t1, t1, t2
-div255w4 t1, t1
-convwb4 d1, t1
-#compin d1, s1, d1
+x4 convubw t1, s1
+# ORC_MULDIV_255((s),(m)), m is from dest
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+x4 convwb d1, t1
 
 
 .function orc_code_combine_out_u
 .dest 4 d1
 .source 4 s1
 .source 4 s2
-.temp 4 t1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
+.temp 8 t1
+.temp 8 t2
+
+x4 convubw t1, s1
+x4 convubw t2, s2
+splatw0q t2, t2
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+# ORC_MULDIV_255((s),(m)), m is from dest
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 xorw t2, t2, 0x00ff
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+x4 convwb d1, t1
 
-compin t1, s1, s2
-compout d1, d1, t1
 
 
 .function orc_code_combine_out_u_n
 .dest 4 d1
 .source 4 s1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
+.temp 8 t1
+.temp 8 t2
+
+x4 convubw t1, s1
+# ORC_MULDIV_255((s),(m)), m is from dest
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 xorw t2, t2, 0x00ff
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+x4 convwb d1, t1
 
-compout d1, d1, s1
 
+# atop: (ORC_DIVIDE_255((s)*(da))+ORC_DIVIDE_255((d)*(255-(sa))))
 
 .function orc_code_combine_atop_u
 .dest 4 d1
 .source 4 s1
 .source 4 s2
-.temp 4 t1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
+.temp 8 t1
+.temp 8 t2
+.temp 8 t3
+.temp 4 t4
+.temp 4 t5
+
+x4 convubw t1, s1
+x4 convubw t2, s2
+splatw0q t2, t2
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
+
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 mullw t3, t1, t2
+x4 div255w t3, t3
+x4 convwb t4, t3
 
-compin t1, s1, s2
-compatop d1, d1, t1
+x4 convubw d_wide, d1
+splatw0q t2, t1
+x4 xorw t2, t2, 0x00ff
+x4 mullw t1, d_wide, t2
+x4 div255w t1, t1
+x4 convwb t5, t1
+
+x4 addusb d1, t4, t5
 
 
 .function orc_code_combine_atop_u_n
 .dest 4 d1
 .source 4 s1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
+.temp 8 t1
+.temp 8 t2
+.temp 8 t3
+.temp 4 t4
+.temp 4 t5
+
+x4 convubw t1, s1
+
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 mullw t3, t1, t2
+x4 div255w t3, t3
+x4 convwb t4, t3
+
+x4 convubw d_wide, d1
+splatw0q t2, t1
+x4 xorw t2, t2, 0x00ff
+x4 mullw t1, d_wide, t2
+x4 div255w t1, t1
+x4 convwb t5, t1
+
+x4 addusb d1, t4, t5
 
-compatop d1, d1, s1
 
 
 .function orc_code_combine_xor_u
 .dest 4 d1
 .source 4 s1
 .source 4 s2
-.temp 4 t1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
+.temp 8 t1
+.temp 8 t2
+.temp 8 t3
+.temp 4 t4
+.temp 4 t5
+
+x4 convubw t1, s1
+x4 convubw t2, s2
+splatw0q t2, t2
+x4 mullw t1, t1, t2
+x4 div255w t1, t1
 
-compin t1, s1, s2
-compxor d1, d1, t1
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 xorw t2, t2, 0x00ff
+x4 mullw t3, t1, t2
+x4 div255w t3, t3
+x4 convwb t4, t3
+
+x4 convubw d_wide, d1
+splatw0q t2, t1
+x4 xorw t2, t2, 0x00ff
+x4 mullw t1, d_wide, t2
+x4 div255w t1, t1
+x4 convwb t5, t1
+
+x4 addusb d1, t4, t5
 
 
 .function orc_code_combine_xor_u_n
 .dest 4 d1
 .source 4 s1
+.temp 8 d_wide
+.temp 8 s_wide
+.temp 8 m_wide
+.temp 8 t1
+.temp 8 t2
+.temp 8 t3
+.temp 4 t4
+.temp 4 t5
+
+x4 convubw t1, s1
+x4 convubw d_wide, d1
+splatw0q t2, d_wide
+x4 xorw t2, t2, 0x00ff
+x4 mullw t3, t1, t2
+x4 div255w t3, t3
+x4 convwb t4, t3
+
+x4 convubw d_wide, d1
+splatw0q t2, t1
+x4 xorw t2, t2, 0x00ff
+x4 mullw t1, d_wide, t2
+x4 div255w t1, t1
+x4 convwb t5, t1
+
+x4 addusb d1, t4, t5
 
-compxor d1, d1, s1
 
 
diff --git a/test/ds.c b/test/ds.c
index 24a0ee9..d16d98b 100644
--- a/test/ds.c
+++ b/test/ds.c
@@ -38,6 +38,8 @@
 
 #define randu8() (random()&0xff)
 
+#define static
+
 static uint32_t
 combine_mask (const uint32_t *src, const uint32_t *mask, int i)
 {
@@ -59,8 +61,46 @@ combine_mask (const uint32_t *src, const uint32_t *mask, int i)
     return s;
 }
 
+static void
+combine_add_u (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 = combine_mask (src, mask, i);
+        uint32_t d = *(dest + i);
+        UN8x4_ADD_UN8x4 (d, s);
+        *(dest + i) = d;
+    }
+}
+
 void
-combine_xor_u (pixman_implementation_t *imp,
+combine_in_u (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 = combine_mask (src, mask, i);
+        uint32_t a = ALPHA_8 (*(dest + i));
+        UN8x4_MUL_UN8 (s, a);
+        *(dest + i) = s;
+    }
+}
+
+static void
+combine_over_u (pixman_implementation_t *imp,
                 pixman_op_t              op,
                 uint32_t *                dest,
                 const uint32_t *          src,
@@ -73,22 +113,20 @@ combine_xor_u (pixman_implementation_t *imp,
     {
         uint32_t s = combine_mask (src, mask, i);
         uint32_t d = *(dest + i);
-        uint32_t src_ia = ALPHA_8 (~s);
-        uint32_t dest_ia = ALPHA_8 (~d);
+        uint32_t ia = ALPHA_8 (~s);
 
-        UN8x4_MUL_UN8_ADD_UN8x4_MUL_UN8 (s, dest_ia, d, src_ia);
-        *(dest + i) = s;
+        UN8x4_MUL_UN8_ADD_UN8x4 (d, ia, s);
+        *(dest + i) = d;
     }
 }
 
-#if 0
 static void
-combine_add_u (pixman_implementation_t *imp,
-               pixman_op_t              op,
-               uint32_t *                dest,
-               const uint32_t *          src,
-               const uint32_t *          mask,
-               int                      width)
+combine_atop_u (pixman_implementation_t *imp,
+                pixman_op_t              op,
+                uint32_t *                dest,
+                const uint32_t *          src,
+                const uint32_t *          mask,
+                int                      width)
 {
     int i;
 
@@ -96,27 +134,32 @@ combine_add_u (pixman_implementation_t *imp,
     {
         uint32_t s = combine_mask (src, mask, i);
         uint32_t d = *(dest + i);
-        UN8x4_ADD_UN8x4 (d, s);
-        *(dest + i) = d;
+        uint32_t dest_a = ALPHA_8 (d);
+        uint32_t src_ia = ALPHA_8 (~s);
+
+        UN8x4_MUL_UN8_ADD_UN8x4_MUL_UN8 (s, dest_a, d, src_ia);
+        *(dest + i) = s;
     }
 }
-#endif
 
 static void
-combine_in_u (pixman_implementation_t *imp,
-              pixman_op_t              op,
-              uint32_t *                dest,
-              const uint32_t *          src,
-              const uint32_t *          mask,
-              int                      width)
+combine_xor_u (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 = combine_mask (src, mask, i);
-        uint32_t a = ALPHA_8 (*(dest + i));
-        UN8x4_MUL_UN8 (s, a);
+        uint32_t d = *(dest + i);
+        uint32_t src_ia = ALPHA_8 (~s);
+        uint32_t dest_ia = ALPHA_8 (~d);
+
+        UN8x4_MUL_UN8_ADD_UN8x4_MUL_UN8 (s, dest_ia, d, src_ia);
         *(dest + i) = s;
     }
 }
@@ -151,8 +194,8 @@ int main (int argc, char *argv[])
   orc_init();
 
   for(i=0;i<256;i++){
-    //dest_orig[i]=ORC_ARGB(255,0,0,255);
     dest_orig[i]=rand_argb();
+    //dest_orig[i]=ORC_ARGB(0,0,0,255);
     dest[i]=dest_orig[i];
     dest_ref[i]=dest_orig[i];
   }
@@ -163,13 +206,17 @@ int main (int argc, char *argv[])
   }
 
   for(i=0;i<256;i++){
-    mask[i]=ORC_ARGB(255,0,0,0);
+    mask[i]=ORC_ARGB(i,0,0,0);
   }
 
-  //combine_xor_u (NULL, 0, dest, src, mask, 256);
-  //combine_xor_u (NULL, 0, dest_ref, src, NULL, 256);
-  combine_in_u (NULL, 0, dest_ref, src, NULL, 256);
-  orc_code_combine_in_u_n (dest, src, 256);
+  //combine_add_u (NULL, 0, dest_ref, src, mask, 256);
+  //orc_code_combine_add_u (dest, src, mask, 256);
+
+  //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);
 
 #if 0
   for(i=0;i<256;i++){
@@ -197,7 +244,7 @@ int main (int argc, char *argv[])
 #endif
 
   for(i=0;i<256;i++){
-    printf("%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 %s\n",
         ORC_ARGB_A(dest_orig[i]),
         ORC_ARGB_R(dest_orig[i]),
@@ -207,6 +254,7 @@ int main (int argc, char *argv[])
         ORC_ARGB_R(src[i]),
         ORC_ARGB_G(src[i]),
         ORC_ARGB_B(src[i]),
+        ORC_ARGB_A(mask[i]),
         ORC_ARGB_A(dest_ref[i]),
         ORC_ARGB_R(dest_ref[i]),
         ORC_ARGB_G(dest_ref[i]),