Implement lower level composite opcodes

1 files changed, 78 insertions, 2 deletions

diff --git a/orc/orcopcodes.c b/orc/orcopcodes.c
index 200f4c9..3be150a 100644
--- a/orc/orcopcodes.c
+++ b/orc/orcopcodes.c
@@ -693,10 +693,80 @@ convlf (OrcOpcodeExecutor *ex, void *user)
 #define ORC_ARGB_G(x) (((x)>>8)&0xff)
 #define ORC_ARGB_B(x) (((x)>>0)&0xff)
 
+#define ORC_ARGB16_A(x) (((x)>>48)&0xffff)
+#define ORC_ARGB16_R(x) (((x)>>32)&0xffff)
+#define ORC_ARGB16_G(x) (((x)>>16)&0xffff)
+#define ORC_ARGB16_B(x) (((x)>>0)&0xffff)
+
 #define ORC_ARGB(a,r,g,b) \
   ((ORC_CLAMP((b),0,255)<<0)|(ORC_CLAMP((g),0,255)<<8)| \
    (ORC_CLAMP((r),0,255)<<16)|(ORC_CLAMP((a),0,255)<<24))
 
+#define ORC_ARGB16(a,r,g,b) \
+  (((uint64_t)ORC_CLAMP((b),0,65535)<<0)|((uint64_t)ORC_CLAMP((g),0,65535)<<16)| \
+   ((uint64_t)ORC_CLAMP((r),0,65535)<<32)|((uint64_t)ORC_CLAMP((a),0,65535)<<48))
+
+
+static void
+convubw4 (OrcOpcodeExecutor *ex, void *user)
+{
+  unsigned int src = ex->src_values[0];
+
+  ex->dest_values[0] = ORC_ARGB16(
+      ORC_ARGB_A(src),
+      ORC_ARGB_R(src),
+      ORC_ARGB_G(src),
+      ORC_ARGB_B(src));
+}
+
+static void
+convwb4 (OrcOpcodeExecutor *ex, void *user)
+{
+  uint64_t src = ex->src_values[0];
+
+  ex->dest_values[0] = ORC_ARGB(
+      ORC_ARGB16_A(src),
+      ORC_ARGB16_R(src),
+      ORC_ARGB16_G(src),
+      ORC_ARGB16_B(src));
+}
+
+static void
+div255w4 (OrcOpcodeExecutor *ex, void *user)
+{
+  uint64_t src = ex->src_values[0];
+
+  ex->dest_values[0] = ORC_ARGB16(
+      ORC_DIVIDE_255(ORC_ARGB16_A(src)),
+      ORC_DIVIDE_255(ORC_ARGB16_R(src)),
+      ORC_DIVIDE_255(ORC_ARGB16_G(src)),
+      ORC_DIVIDE_255(ORC_ARGB16_B(src)));
+}
+
+static void
+splat0w4 (OrcOpcodeExecutor *ex, void *user)
+{
+  uint64_t src = ex->src_values[0];
+
+  ex->dest_values[0] = ORC_ARGB16(
+      ORC_ARGB_A(src),
+      ORC_ARGB_A(src),
+      ORC_ARGB_A(src),
+      ORC_ARGB_A(src));
+}
+
+static void
+mullw4 (OrcOpcodeExecutor *ex, void *user)
+{
+  uint64_t src1 = ex->src_values[0];
+  uint64_t src2 = ex->src_values[1];
+
+  ex->dest_values[0] = ORC_ARGB16(
+      ORC_ARGB16_A(src1)*ORC_ARGB16_A(src2),
+      ORC_ARGB16_R(src1)*ORC_ARGB16_R(src2),
+      ORC_ARGB16_G(src1)*ORC_ARGB16_G(src2),
+      ORC_ARGB16_B(src1)*ORC_ARGB16_B(src2));
+}
 
 static void
 compin (OrcOpcodeExecutor *ex, void *user)
@@ -746,7 +816,7 @@ compovera (OrcOpcodeExecutor *ex, void *user)
 }
 
 static void
-compadd (OrcOpcodeExecutor *ex, void *user)
+addusb4 (OrcOpcodeExecutor *ex, void *user)
 {
   unsigned int src1 = ex->src_values[0];
   unsigned int src2 = ex->src_values[1];
@@ -959,11 +1029,17 @@ static OrcStaticOpcode opcodes[] = {
   { "convlf", convlf, NULL, ORC_STATIC_OPCODE_FLOAT_DEST, { 4 }, { 4 } },
 
   /* pixel ops */
+  { "convubw4", convubw4, NULL, 0, { 8 }, { 4 } },
+  { "convwb4", convwb4, NULL, 0, { 4 }, { 8 } },
+  { "div255w4", div255w4, NULL, 0, { 8 }, { 8 } },
+  { "splat0w4", splat0w4, NULL, 0, { 8 }, { 8 } },
+  { "mullw4", mullw4, NULL, 0, { 8 }, { 8, 8 } },
+
   { "compin", compin, NULL, 0, { 4 }, { 4, 4 } },
   { "compina", compina, NULL, 0, { 1 }, { 1, 1 } },
   { "compover", compover, NULL, 0, { 4 }, { 4, 4 } },
   { "compovera", compovera, NULL, 0, { 1 }, { 1, 1 } },
-  { "compadd", compadd, NULL, 0, { 4 }, { 4, 4 } },
+  { "addusb4", addusb4, NULL, 0, { 4 }, { 4, 4 } },
   { "compout", compout, NULL, 0, { 4 }, { 4, 4 } },
   { "compatop", compatop, NULL, 0, { 4 }, { 4, 4 } },
   { "compxor", compxor, NULL, 0, { 4 }, { 4, 4 } },