r300/compiler: Add support for inline literals

On R500 chips, shader instructions can take 7-bit (3-bit mantissa, 4-bit
exponent) floating point values as inputs in place of registers.

12 files changed, 192 insertions, 9 deletions

diff --git a/src/gallium/drivers/r300/Makefile.sources b/src/gallium/drivers/r300/Makefile.sources
index 90105d6e694..e27b14e5702 100644
--- a/src/gallium/drivers/r300/Makefile.sources
+++ b/src/gallium/drivers/r300/Makefile.sources
@@ -28,6 +28,7 @@ C_SOURCES := \
 	compiler/radeon_compiler_util.c \
 	compiler/radeon_emulate_branches.c \
 	compiler/radeon_emulate_loops.c \
+	compiler/radeon_inline_literals.c \
 	compiler/radeon_program.c \
 	compiler/radeon_program_print.c \
 	compiler/radeon_opcodes.c \
diff --git a/src/gallium/drivers/r300/compiler/r3xx_fragprog.c b/src/gallium/drivers/r300/compiler/r3xx_fragprog.c
index 8fea4db94ea..8ef2d24fc99 100644
--- a/src/gallium/drivers/r300/compiler/r3xx_fragprog.c
+++ b/src/gallium/drivers/r300/compiler/r3xx_fragprog.c
@@ -125,6 +125,7 @@ void r3xx_compile_fragment_program(struct r300_fragment_program_compiler* c)
 		{"emulate loops",		1, !is_r500,	rc_emulate_loops,		NULL},
 		{"register rename",		1, !is_r500 || opt,		rc_rename_regs,			NULL},
 		{"dataflow optimize",		1, opt,		rc_optimize,			NULL},
+		{"inline literals",		1, is_r500 && opt,		rc_inline_literals,			NULL},
 		{"dataflow swizzles",		1, 1,		rc_dataflow_swizzles,		NULL},
 		{"dead constants",		1, 1,		rc_remove_unused_constants,	&c->code->constants_remap_table},
 		{"pair translate",		1, 1,		rc_pair_translate,		NULL},
diff --git a/src/gallium/drivers/r300/compiler/r500_fragprog.c b/src/gallium/drivers/r300/compiler/r500_fragprog.c
index cf99f5e4538..499aa92423b 100644
--- a/src/gallium/drivers/r300/compiler/r500_fragprog.c
+++ b/src/gallium/drivers/r300/compiler/r500_fragprog.c
@@ -218,6 +218,8 @@ static int r500_swizzle_is_native(rc_opcode opcode, struct rc_src_register reg)
 			return 1;
 
 		return 0;
+	} else if (reg.File == RC_FILE_INLINE) {
+		return 1;
 	} else {
 		/* ALU instructions support almost everything */
 		relevant = 0;
diff --git a/src/gallium/drivers/r300/compiler/r500_fragprog_emit.c b/src/gallium/drivers/r300/compiler/r500_fragprog_emit.c
index 87b96d15079..f6b6c0f9c06 100644
--- a/src/gallium/drivers/r300/compiler/r500_fragprog_emit.c
+++ b/src/gallium/drivers/r300/compiler/r500_fragprog_emit.c
@@ -210,6 +210,8 @@ static unsigned int use_source(struct r500_fragment_program_code* code, struct r
 	} else if (src.File == RC_FILE_TEMPORARY || src.File == RC_FILE_INPUT) {
 		use_temporary(code, src.Index);
 		return src.Index;
+	} else if (src.File == RC_FILE_INLINE) {
+		return src.Index | (1 << 7);
 	}
 
 	return 0;
diff --git a/src/gallium/drivers/r300/compiler/radeon_compiler.c b/src/gallium/drivers/r300/compiler/radeon_compiler.c
index 986e3b77403..4d4eb645841 100644
--- a/src/gallium/drivers/r300/compiler/radeon_compiler.c
+++ b/src/gallium/drivers/r300/compiler/radeon_compiler.c
@@ -357,21 +357,22 @@ void rc_transform_fragment_face(struct radeon_compiler *c, unsigned face)
 static void reg_count_callback(void * userdata, struct rc_instruction * inst,
 		rc_register_file file, unsigned int index, unsigned int mask)
 {
-	int *max_reg = userdata;
+	struct rc_program_stats *s = userdata;
 	if (file == RC_FILE_TEMPORARY)
-		(int)index > *max_reg ? *max_reg = index : 0;
+		(int)index > s->num_temp_regs ? s->num_temp_regs = index : 0;
+	if (file == RC_FILE_INLINE)
+		s->num_inline_literals++;
 }
 
 void rc_get_stats(struct radeon_compiler *c, struct rc_program_stats *s)
 {
-	int max_reg = -1;
 	struct rc_instruction * tmp;
 	memset(s, 0, sizeof(*s));
 
 	for(tmp = c->Program.Instructions.Next; tmp != &c->Program.Instructions;
 							tmp = tmp->Next){
 		const struct rc_opcode_info * info;
-		rc_for_all_reads_mask(tmp, reg_count_callback, &max_reg);
+		rc_for_all_reads_mask(tmp, reg_count_callback, s);
 		if (tmp->Type == RC_INSTRUCTION_NORMAL) {
 			info = rc_get_opcode_info(tmp->U.I.Opcode);
 			if (info->Opcode == RC_OPCODE_BEGIN_TEX)
@@ -405,7 +406,9 @@ void rc_get_stats(struct radeon_compiler *c, struct rc_program_stats *s)
 			s->num_tex_insts++;
 		s->num_insts++;
 	}
-	s->num_temp_regs = max_reg + 1;
+	/* Increment here because the reg_count_callback store the max
+	 * temporary reg index in s->nun_temp_regs. */
+	s->num_temp_regs++;
 }
 
 static void print_stats(struct radeon_compiler * c)
@@ -437,10 +440,11 @@ static void print_stats(struct radeon_compiler * c)
 			       "~%4u Presub Operations\n"
 			       "~%4u OMOD Operations\n"
 			       "~%4u Temporary Registers\n"
+			       "~%4u Inline Literals\n"
 			       "~~~~~~~~~~~~~~ END ~~~~~~~~~~~~~~\n",
 			       s.num_insts, s.num_rgb_insts, s.num_alpha_insts,
 			       s.num_fc_insts, s.num_tex_insts, s.num_presub_ops,
-			       s.num_omod_ops, s.num_temp_regs);
+			       s.num_omod_ops, s.num_temp_regs, s.num_inline_literals);
 		break;
 	default:
 		assert(0);
diff --git a/src/gallium/drivers/r300/compiler/radeon_compiler.h b/src/gallium/drivers/r300/compiler/radeon_compiler.h
index ac9691c816f..e7ccbb732d1 100644
--- a/src/gallium/drivers/r300/compiler/radeon_compiler.h
+++ b/src/gallium/drivers/r300/compiler/radeon_compiler.h
@@ -161,6 +161,7 @@ struct rc_program_stats {
 	unsigned num_presub_ops;
 	unsigned num_temp_regs;
 	unsigned num_omod_ops;
+	unsigned num_inline_literals;
 };
 
 void rc_get_stats(struct radeon_compiler *c, struct rc_program_stats *s);
diff --git a/src/gallium/drivers/r300/compiler/radeon_dataflow.h b/src/gallium/drivers/r300/compiler/radeon_dataflow.h
index d8a627258ea..bb8d48206e2 100644
--- a/src/gallium/drivers/r300/compiler/radeon_dataflow.h
+++ b/src/gallium/drivers/r300/compiler/radeon_dataflow.h
@@ -130,5 +130,6 @@ void rc_dataflow_swizzles(struct radeon_compiler * c, void *user);
 /*@}*/
 
 void rc_optimize(struct radeon_compiler * c, void *user);
+void rc_inline_literals(struct radeon_compiler *c, void *user);
 
 #endif /* RADEON_DATAFLOW_H */
diff --git a/src/gallium/drivers/r300/compiler/radeon_inline_literals.c b/src/gallium/drivers/r300/compiler/radeon_inline_literals.c
new file mode 100644
index 00000000000..568a3d60e14
--- /dev/null
+++ b/src/gallium/drivers/r300/compiler/radeon_inline_literals.c
@@ -0,0 +1,140 @@
+
+#include "radeon_compiler.h"
+#include "radeon_compiler_util.h"
+#include "radeon_dataflow.h"
+#include "radeon_program.h"
+#include "radeon_program_constants.h"
+#include <stdio.h>
+
+#define VERBOSE 0
+
+#define DBG(...) do { if (VERBOSE) fprintf(stderr, __VA_ARGS__); } while(0)
+
+/* IEEE-754:
+ * 22:0 mantissa
+ * 30:23 exponent
+ * 31 sign
+ *
+ * R300:
+ * 0:2 mantissa
+ * 3:6 exponent (bias 7)
+ */
+static int ieee_754_to_r300_float(float f, unsigned char *r300_float_out)
+{
+	unsigned float_bits = *((unsigned *)&f);
+	/* XXX: Handle big-endian */
+	unsigned mantissa = float_bits &         0x007fffff;
+	unsigned biased_exponent = (float_bits & 0x7f800000) >> 23;
+	unsigned negate = !!(float_bits &         0x80000000);
+	int exponent = biased_exponent - 127;
+	unsigned mantissa_mask = 0xff8fffff;
+	unsigned r300_exponent, r300_mantissa;
+
+	DBG("Converting %f (0x%x) to 7-bit:\n", f, float_bits);
+	DBG("Raw exponent = %d\n", exponent);
+
+	if (exponent < -7 || exponent > 8) {
+		DBG("Failed exponent out of range\n\n");
+		return 0;
+	}
+
+	if (mantissa & mantissa_mask) {
+		DBG("Failed mantisa has too many bits:\n"
+			"manitssa=0x%x mantissa_mask=0x%x, and=0x%x\n\n",
+			mantissa, mantissa_mask,
+			mantissa & mantissa_mask);
+		return 0;
+	}
+
+	r300_exponent = exponent + 7;
+	r300_mantissa = (mantissa & ~mantissa_mask) >> 20;
+	*r300_float_out = r300_mantissa | (r300_exponent << 3);
+
+	DBG("Success! r300_float = 0x%x\n\n", *r300_float_out);
+
+	if (negate)
+		return -1;
+	else
+		return 1;
+}
+
+void rc_inline_literals(struct radeon_compiler *c, void *user)
+{
+	struct rc_instruction * inst;
+
+	for(inst = c->Program.Instructions.Next;
+					inst != &c->Program.Instructions;
+					inst = inst->Next) {
+		const struct rc_opcode_info * info =
+					rc_get_opcode_info(inst->U.I.Opcode);
+
+		unsigned src_idx;
+		struct rc_constant * constant;
+		float float_value;
+		unsigned char r300_float;
+		int ret;
+
+		/* XXX: Handle presub */
+
+		/* We aren't using rc_for_all_reads_src here, because presub
+		 * sources need to be handled differently. */
+		for (src_idx = 0; src_idx < info->NumSrcRegs; src_idx++) {
+			unsigned new_swizzle;
+			unsigned use_literal = 0;
+			unsigned negate_mask = 0;
+			unsigned swz, chan;
+			struct rc_src_register * src_reg =
+						&inst->U.I.SrcReg[src_idx];
+			swz = RC_SWIZZLE_UNUSED;
+			if (src_reg->File != RC_FILE_CONSTANT) {
+				continue;
+			}
+			constant =
+				&c->Program.Constants.Constants[src_reg->Index];
+			if (constant->Type != RC_CONSTANT_IMMEDIATE) {
+				continue;
+			}
+			new_swizzle = rc_init_swizzle(RC_SWIZZLE_UNUSED, 0);
+			for (chan = 0; chan < 4; chan++) {
+				unsigned char r300_float_tmp;
+				swz = GET_SWZ(src_reg->Swizzle, chan);
+				if (swz == RC_SWIZZLE_UNUSED) {
+					continue;
+				}
+				float_value = constant->u.Immediate[swz];
+				ret = ieee_754_to_r300_float(float_value,
+								&r300_float_tmp);
+				if (!ret || (use_literal &&
+						r300_float != r300_float_tmp)) {
+					use_literal = 0;
+					break;
+				}
+
+				if (ret == -1 && src_reg->Abs) {
+					use_literal = 0;
+					break;
+				}
+
+				if (!use_literal) {
+					r300_float = r300_float_tmp;
+					use_literal = 1;
+				}
+
+				/* Use RC_SWIZZLE_W for the inline constant, so
+				 * it will become one of the alpha sources. */
+				SET_SWZ(new_swizzle, chan, RC_SWIZZLE_W);
+				if (ret == -1) {
+					negate_mask |= (1 << chan);
+				}
+			}
+
+			if (!use_literal) {
+				continue;
+			}
+			src_reg->File = RC_FILE_INLINE;
+			src_reg->Index = r300_float;
+			src_reg->Swizzle = new_swizzle;
+			src_reg->Negate = src_reg->Negate ^ negate_mask;
+		}
+	}
+}
diff --git a/src/gallium/drivers/r300/compiler/radeon_pair_translate.c b/src/gallium/drivers/r300/compiler/radeon_pair_translate.c
index 7d9c8d1fab6..c6050bdf4dd 100644
--- a/src/gallium/drivers/r300/compiler/radeon_pair_translate.c
+++ b/src/gallium/drivers/r300/compiler/radeon_pair_translate.c
@@ -268,7 +268,15 @@ static void set_pair_instruction(struct r300_fragment_program_compiler *c,
 			pair->Alpha.Arg[i].Source = source;
 			pair->Alpha.Arg[i].Swizzle = rc_init_swizzle(swz, 1);
 			pair->Alpha.Arg[i].Abs = inst->SrcReg[i].Abs;
-			pair->Alpha.Arg[i].Negate = !!(inst->SrcReg[i].Negate & RC_MASK_W);
+
+			if (istranscendent) {
+				pair->Alpha.Arg[i].Negate =
+					!!(inst->SrcReg[i].Negate &
+							inst->DstReg.WriteMask);
+			} else {
+				pair->Alpha.Arg[i].Negate =
+					!!(inst->SrcReg[i].Negate & RC_MASK_W);
+			}
 		}
 	}
 
diff --git a/src/gallium/drivers/r300/compiler/radeon_program_constants.h b/src/gallium/drivers/r300/compiler/radeon_program_constants.h
index 4f59c47815e..c07c492b0c9 100644
--- a/src/gallium/drivers/r300/compiler/radeon_program_constants.h
+++ b/src/gallium/drivers/r300/compiler/radeon_program_constants.h
@@ -85,7 +85,12 @@ typedef enum {
 	 * Indicates this register should use the result of the presubtract
 	 * operation.
 	 */
-	RC_FILE_PRESUB
+	RC_FILE_PRESUB,
+
+	/**
+	 * Indicates that the source index has been encoded as a 7-bit float.
+	 */
+	RC_FILE_INLINE
 } rc_register_file;
 
 enum {
diff --git a/src/gallium/drivers/r300/compiler/radeon_program_pair.h b/src/gallium/drivers/r300/compiler/radeon_program_pair.h
index b6eb0ebe47d..085ff994501 100644
--- a/src/gallium/drivers/r300/compiler/radeon_program_pair.h
+++ b/src/gallium/drivers/r300/compiler/radeon_program_pair.h
@@ -57,7 +57,7 @@ struct radeon_compiler;
 
 struct rc_pair_instruction_source {
 	unsigned int Used:1;
-	unsigned int File:3;
+	unsigned int File:4;
 	unsigned int Index:RC_REGISTER_INDEX_BITS;
 };
 
diff --git a/src/gallium/drivers/r300/compiler/radeon_program_print.c b/src/gallium/drivers/r300/compiler/radeon_program_print.c
index dc40d7fd4f2..e3d2104b250 100644
--- a/src/gallium/drivers/r300/compiler/radeon_program_print.c
+++ b/src/gallium/drivers/r300/compiler/radeon_program_print.c
@@ -109,6 +109,22 @@ static void rc_print_comparefunc(FILE * f, const char * lhs, rc_compare_func fun
 	}
 }
 
+static void rc_print_inline_float(FILE * f, int index)
+{
+	int r300_exponent = (index >> 3) & 0xf;
+	unsigned r300_mantissa = index & 0x7;
+	unsigned float_exponent;
+	unsigned real_float;
+	float * print_float = (float*) &real_float;
+
+	r300_exponent -= 7;
+	float_exponent = r300_exponent + 127;
+	real_float = (r300_mantissa << 20) | (float_exponent << 23);
+
+	fprintf(f, "%f (0x%x)", *print_float, index);
+
+}
+
 static void rc_print_register(FILE * f, rc_register_file file, int index, unsigned int reladdr)
 {
 	if (file == RC_FILE_NONE) {
@@ -118,6 +134,8 @@ static void rc_print_register(FILE * f, rc_register_file file, int index, unsign
 		case RC_SPECIAL_ALU_RESULT: fprintf(f, "aluresult"); break;
 		default: fprintf(f, "special[%i]", index); break;
 		}
+	} else if (file == RC_FILE_INLINE) {
+		rc_print_inline_float(f, index);
 	} else {
 		const char * filename;
 		switch(file) {