Implement lower level composite opcodes

3 files changed, 172 insertions, 6 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 } },
diff --git a/orc/orcprogram-c.c b/orc/orcprogram-c.c
index 6ec010e..817ef7b 100644
--- a/orc/orcprogram-c.c
+++ b/orc/orcprogram-c.c
@@ -120,6 +120,13 @@ orc_target_c_get_asm_preamble (void)
     "#define ORC_ARGB(a,r,g,b) \\\n"
     "  ((ORC_CLAMP((b),0,255)<<0)|(ORC_CLAMP((g),0,255)<<8)| \\\n"
     "   (ORC_CLAMP((r),0,255)<<16)|(ORC_CLAMP((a),0,255)<<24))\n"
+    "#define ORC_ARGB16_A(x) (((x)>>48)&0xffff)\n"
+    "#define ORC_ARGB16_R(x) (((x)>>32)&0xffff)\n"
+    "#define ORC_ARGB16_G(x) (((x)>>16)&0xffff)\n"
+    "#define ORC_ARGB16_B(x) (((x)>>0)&0xffff)\n"
+    "#define ORC_ARGB16(a,r,g,b) \\\n"
+    "  (((uint64_t)ORC_CLAMP((b),0,65535)<<0)|((uint64_t)ORC_CLAMP((g),0,65535)<<16)| \\\n"
+    "   ((uint64_t)ORC_CLAMP((r),0,65535)<<32)|((uint64_t)ORC_CLAMP((a),0,65535)<<48))\n"
     "#ifdef __GNUC_PREREQ\n"
     "#if __GNUC_PREREQ(4,2)\n"
     "#pragma GCC diagnostic ignored \"-Wstrict-aliasing\"\n"
@@ -716,6 +723,82 @@ c_rule_splitwb (OrcCompiler *p, void *user, OrcInstruction *insn)
 }
 
 static void
+c_rule_convubw4 (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src1[20];
+
+  c_get_name (dest, p, insn->dest_args[0]);
+  c_get_name (src1, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,"    %s = ORC_ARGB16(\n", dest);
+  ORC_ASM_CODE(p,"        ORC_ARGB_A(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB_R(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB_G(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB_B(%s));\n", src1);
+}
+
+static void
+c_rule_convwb4 (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src1[20];
+
+  c_get_name (dest, p, insn->dest_args[0]);
+  c_get_name (src1, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,"    %s = ORC_ARGB(\n", dest);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_A(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_R(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_G(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_B(%s));\n", src1);
+}
+
+static void
+c_rule_div255w4 (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src1[20];
+
+  c_get_name (dest, p, insn->dest_args[0]);
+  c_get_name (src1, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,"    %s = ORC_ARGB16(\n", dest);
+  ORC_ASM_CODE(p,"        ORC_DIVIDE_255(ORC_ARGB16_A(%s)),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_DIVIDE_255(ORC_ARGB16_R(%s)),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_DIVIDE_255(ORC_ARGB16_G(%s)),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_DIVIDE_255(ORC_ARGB16_B(%s)));\n", src1);
+}
+
+static void
+c_rule_splat0w4 (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src1[20];
+
+  c_get_name (dest, p, insn->dest_args[0]);
+  c_get_name (src1, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,"    %s = ORC_ARGB16(\n", dest);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_A(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_A(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_A(%s),\n", src1);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_A(%s));\n", src1);
+}
+
+static void
+c_rule_mullw4 (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src1[20], src2[20];
+
+  c_get_name (dest, p, insn->dest_args[0]);
+  c_get_name (src1, p, insn->src_args[0]);
+  c_get_name (src2, p, insn->src_args[1]);
+
+  ORC_ASM_CODE(p,"    %s = ORC_ARGB16(\n", dest);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_A(%s)*ORC_ARGB16_A(%s),\n", src1, src2);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_R(%s)*ORC_ARGB16_R(%s),\n", src1, src2);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_G(%s)*ORC_ARGB16_G(%s),\n", src1, src2);
+  ORC_ASM_CODE(p,"        ORC_ARGB16_B(%s)*ORC_ARGB16_B(%s));\n", src1, src2);
+}
+
+static void
 c_rule_compin (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   char dest[20], src1[20], src2[20];
@@ -772,7 +855,7 @@ c_rule_compovera (OrcCompiler *p, void *user, OrcInstruction *insn)
 }
 
 static void
-c_rule_compadd (OrcCompiler *p, void *user, OrcInstruction *insn)
+c_rule_addusb4 (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   char dest[20], src1[20], src2[20];
 
@@ -891,11 +974,18 @@ orc_c_init (void)
   orc_rule_register (rule_set, "accsadubl", c_rule_accsadubl, NULL);
   orc_rule_register (rule_set, "splitlw", c_rule_splitlw, NULL);
   orc_rule_register (rule_set, "splitwb", c_rule_splitwb, NULL);
+
+  orc_rule_register (rule_set, "convubw4", c_rule_convubw4, NULL);
+  orc_rule_register (rule_set, "convwb4", c_rule_convwb4, NULL);
+  orc_rule_register (rule_set, "div255w4", c_rule_div255w4, NULL);
+  orc_rule_register (rule_set, "splat0w4", c_rule_splat0w4, NULL);
+  orc_rule_register (rule_set, "mullw4", c_rule_mullw4, NULL);
+
   orc_rule_register (rule_set, "compin", c_rule_compin, NULL);
   orc_rule_register (rule_set, "compina", c_rule_compina, NULL);
   orc_rule_register (rule_set, "compover", c_rule_compover, NULL);
   orc_rule_register (rule_set, "compovera", c_rule_compovera, NULL);
-  orc_rule_register (rule_set, "compadd", c_rule_compadd, NULL);
+  orc_rule_register (rule_set, "addusb4", c_rule_addusb4, NULL);
   orc_rule_register (rule_set, "compout", c_rule_compout, NULL);
   orc_rule_register (rule_set, "compatop", c_rule_compatop, NULL);
   orc_rule_register (rule_set, "compxor", c_rule_compxor, NULL);
diff --git a/orc/orcprogram.h b/orc/orcprogram.h
index c24948f..9e3e9ba 100644
--- a/orc/orcprogram.h
+++ b/orc/orcprogram.h
@@ -438,8 +438,8 @@ struct _OrcCompiler {
  */
 struct _OrcOpcodeExecutor {
   /*< private >*/
-  int src_values[ORC_STATIC_OPCODE_N_SRC];
-  int dest_values[ORC_STATIC_OPCODE_N_DEST];
+  int64_t src_values[ORC_STATIC_OPCODE_N_SRC];
+  int64_t dest_values[ORC_STATIC_OPCODE_N_DEST];
 };
 
 /**