radeon_program_pair.c

Mesa is an open-source implementation of the OpenGL specification - a system for rendering interacti

	if (inst->DstReg.File == PROGRAM_TEMPORARY) {
		struct pair_register_translation *t = &s->Temps[inst->DstReg.Index];
		deref_hw_reg(s, t->HwIndex);

		int i;
		for(i = 0; i < 4; ++i) {
			if (!GET_BIT(inst->DstReg.WriteMask, i))
				continue;

			t->Value[i] = pairinst->Values[i];
			if (t->Value[i]->NumReaders) {
				struct reg_value_reader *r;
				for(r = pairinst->Values[i]->Readers; r; r = r->Next)
					decrement_dependencies(s, r->IP);
			} else if (t->Value[i]->Next) {
				/* This happens when the only reader writes
				 * the register at the same time */
				decrement_dependencies(s, t->Value[i]->Next->IP);
			}
		}
	}

	int nsrc = _mesa_num_inst_src_regs(inst->Opcode);
	int i;
	for(i = 0; i < nsrc; i++) {
		struct pair_register_translation *t = get_register(s, inst->SrcReg[i].File, inst->SrcReg[i].Index);
		if (!t)
			continue;

		deref_hw_reg(s, get_hw_reg(s, inst->SrcReg[i].File, inst->SrcReg[i].Index));

		if (inst->SrcReg[i].File != PROGRAM_TEMPORARY)
			continue;

		int j;
		for(j = 0; j < 4; ++j) {
			GLuint swz = GET_SWZ(inst->SrcReg[i].Swizzle, j);
			if (swz >= 4)
				continue;
			if (!t->Value[swz])
				continue;

			/* Do not free a dependency if this instruction
			 * also rewrites the value. See scan_instructions. */
			if (inst->DstReg.File == PROGRAM_TEMPORARY &&
			    inst->DstReg.Index == inst->SrcReg[i].Index &&
			    GET_BIT(inst->DstReg.WriteMask, swz))
				continue;

			if (!--t->Value[swz]->NumReaders) {
				if (t->Value[swz]->Next)
					decrement_dependencies(s, t->Value[swz]->Next->IP);
			}
		}
	}
}


/**
 * Emit all ready texture instructions in a single block.
 *
 * Emit as a single block to (hopefully) sample many textures in parallel,
 * and to avoid hardware indirections on R300.
 *
 * In R500, we don't really know when the result of a texture instruction
 * arrives. So allocate all destinations first, to make sure they do not
 * arrive early and overwrite a texture coordinate we're going to use later
 * in the block.
 */
static void emit_all_tex(struct pair_state *s)
{
	struct pair_state_instruction *readytex;
	struct pair_state_instruction *pairinst;

	ASSERT(s->ReadyTEX);

	// Don't let the ready list change under us!
	readytex = s->ReadyTEX;
	s->ReadyTEX = 0;

	// Allocate destination hardware registers in one block to avoid conflicts.
	for(pairinst = readytex; pairinst; pairinst = pairinst->NextReady) {
		int ip = pairinst - s->Instructions;
		struct prog_instruction *inst = s->Program->Instructions + ip;
		if (inst->Opcode != OPCODE_KIL)
			get_hw_reg(s, inst->DstReg.File, inst->DstReg.Index);
	}

	if (s->Debug)
		_mesa_printf(" BEGIN_TEX\n");

	if (s->Handler->BeginTexBlock)
		s->Error = s->Error || !s->Handler->BeginTexBlock(s->UserData);

	for(pairinst = readytex; pairinst; pairinst = pairinst->NextReady) {
		int ip = pairinst - s->Instructions;
		struct prog_instruction *inst = s->Program->Instructions + ip;
		commit_instruction(s, ip);

		if (inst->Opcode != OPCODE_KIL)
			inst->DstReg.Index = get_hw_reg(s, inst->DstReg.File, inst->DstReg.Index);
		inst->SrcReg[0].Index = get_hw_reg(s, inst->SrcReg[0].File, inst->SrcReg[0].Index);

		if (s->Debug) {
			_mesa_printf("   ");
			_mesa_print_instruction(inst);
		}
		s->Error = s->Error || !s->Handler->EmitTex(s->UserData, inst);
	}

	if (s->Debug)
		_mesa_printf(" END_TEX\n");
}


static int alloc_pair_source(struct pair_state *s, struct radeon_pair_instruction *pair,
	struct prog_src_register src, GLboolean rgb, GLboolean alpha)
{
	int candidate = -1;
	int candidate_quality = -1;
	int i;

	if (!rgb && !alpha)
		return 0;

	GLuint constant;
	GLuint index;

	if (src.File == PROGRAM_TEMPORARY || src.File == PROGRAM_INPUT) {
		constant = 0;
		index = get_hw_reg(s, src.File, src.Index);
	} else {
		constant = 1;
		s->Error |= !s->Handler->EmitConst(s->UserData, src.File, src.Index, &index);
	}

	for(i = 0; i < 3; ++i) {
		int q = 0;
		if (rgb) {
			if (pair->RGB.Src[i].Used) {
				if (pair->RGB.Src[i].Constant != constant ||
				    pair->RGB.Src[i].Index != index)
					continue;
				q++;
			}
		}
		if (alpha) {
			if (pair->Alpha.Src[i].Used) {
				if (pair->Alpha.Src[i].Constant != constant ||
				    pair->Alpha.Src[i].Index != index)
					continue;
				q++;
			}
		}
		if (q > candidate_quality) {
			candidate_quality = q;
			candidate = i;
		}
	}

	if (candidate >= 0) {
		if (rgb) {
			pair->RGB.Src[candidate].Used = 1;
			pair->RGB.Src[candidate].Constant = constant;
			pair->RGB.Src[candidate].Index = index;
		}
		if (alpha) {
			pair->Alpha.Src[candidate].Used = 1;
			pair->Alpha.Src[candidate].Constant = constant;
			pair->Alpha.Src[candidate].Index = index;
		}
	}

	return candidate;
}

/**
 * Fill the given ALU instruction's opcodes and source operands into the given pair,
 * if possible.
 */
static GLboolean fill_instruction_into_pair(struct pair_state *s, struct radeon_pair_instruction *pair, int ip)
{
	struct pair_state_instruction *pairinst = s->Instructions + ip;
	struct prog_instruction *inst = s->Program->Instructions + ip;

	ASSERT(!pairinst->NeedRGB || pair->RGB.Opcode == OPCODE_NOP);
	ASSERT(!pairinst->NeedAlpha || pair->Alpha.Opcode == OPCODE_NOP);

	if (pairinst->NeedRGB) {
		if (pairinst->IsTranscendent)
			pair->RGB.Opcode = OPCODE_REPL_ALPHA;
		else
			pair->RGB.Opcode = inst->Opcode;
		if (inst->SaturateMode == SATURATE_ZERO_ONE)
			pair->RGB.Saturate = 1;
	}
	if (pairinst->NeedAlpha) {
		pair->Alpha.Opcode = inst->Opcode;
		if (inst->SaturateMode == SATURATE_ZERO_ONE)
			pair->Alpha.Saturate = 1;
	}

	int nargs = _mesa_num_inst_src_regs(inst->Opcode);
	int i;

	/* Special case for DDX/DDY (MDH/MDV). */
	if (inst->Opcode == OPCODE_DDX || inst->Opcode == OPCODE_DDY) {
		if (pair->RGB.Src[0].Used || pair->Alpha.Src[0].Used)
			return GL_FALSE;
		else
			nargs++;
	}

	for(i = 0; i < nargs; ++i) {
		int source;
		if (pairinst->NeedRGB && !pairinst->IsTranscendent) {
			GLboolean srcrgb = GL_FALSE;
			GLboolean srcalpha = GL_FALSE;
			GLuint negatebase = 0;
			int j;
			for(j = 0; j < 3; ++j) {
				GLuint swz = GET_SWZ(inst->SrcReg[i].Swizzle, j);
				if (swz < 3)
					srcrgb = GL_TRUE;
				else if (swz < 4)
					srcalpha = GL_TRUE;
				if (swz != SWIZZLE_NIL && GET_BIT(inst->SrcReg[i].NegateBase, j))
					negatebase = 1;
			}
			source = alloc_pair_source(s, pair, inst->SrcReg[i], srcrgb, srcalpha);
			if (source < 0)
				return GL_FALSE;
			pair->RGB.Arg[i].Source = source;
			pair->RGB.Arg[i].Swizzle = inst->SrcReg[i].Swizzle & 0x1ff;
			pair->RGB.Arg[i].Abs = inst->SrcReg[i].Abs;
			pair->RGB.Arg[i].Negate = (negatebase & ~pair->RGB.Arg[i].Abs) ^ inst->SrcReg[i].NegateAbs;
		}
		if (pairinst->NeedAlpha) {
			GLboolean srcrgb = GL_FALSE;
			GLboolean srcalpha = GL_FALSE;
			GLuint negatebase = GET_BIT(inst->SrcReg[i].NegateBase, pairinst->IsTranscendent ? 0 : 3);
			GLuint swz = GET_SWZ(inst->SrcReg[i].Swizzle, pairinst->IsTranscendent ? 0 : 3);
			if (swz < 3)
				srcrgb = GL_TRUE;
			else if (swz < 4)
				srcalpha = GL_TRUE;
			source = alloc_pair_source(s, pair, inst->SrcReg[i], srcrgb, srcalpha);
			if (source < 0)
				return GL_FALSE;
			pair->Alpha.Arg[i].Source = source;
			pair->Alpha.Arg[i].Swizzle = swz;
			pair->Alpha.Arg[i].Abs = inst->SrcReg[i].Abs;
			pair->Alpha.Arg[i].Negate = (negatebase & ~pair->RGB.Arg[i].Abs) ^ inst->SrcReg[i].NegateAbs;
		}
	}

	return GL_TRUE;
}


/**
 * Fill in the destination register information.
 *
 * This is split from filling in source registers because we want
 * to avoid allocating hardware temporaries for destinations until
 * we are absolutely certain that we're going to emit a certain
 * instruction pairing.
 */
static void fill_dest_into_pair(struct pair_state *s, struct radeon_pair_instruction *pair, int ip)
{
	struct pair_state_instruction *pairinst = s->Instructions + ip;
	struct prog_instruction *inst = s->Program->Instructions + ip;

	if (inst->DstReg.File == PROGRAM_OUTPUT) {
		if (inst->DstReg.Index == FRAG_RESULT_COLR) {
			pair->RGB.OutputWriteMask |= inst->DstReg.WriteMask & WRITEMASK_XYZ;
			pair->Alpha.OutputWriteMask |= GET_BIT(inst->DstReg.WriteMask, 3);
		} else if (inst->DstReg.Index == FRAG_RESULT_DEPR) {
			pair->Alpha.DepthWriteMask |= GET_BIT(inst->DstReg.WriteMask, 3);
		}
	} else {
		GLuint hwindex = get_hw_reg(s, inst->DstReg.File, inst->DstReg.Index);
		if (pairinst->NeedRGB) {
			pair->RGB.DestIndex = hwindex;
			pair->RGB.WriteMask |= inst->DstReg.WriteMask & WRITEMASK_XYZ;
		}
		if (pairinst->NeedAlpha) {
			pair->Alpha.DestIndex = hwindex;
			pair->Alpha.WriteMask |= GET_BIT(inst->DstReg.WriteMask, 3);
		}
	}
}


/**
 * Find a good ALU instruction or pair of ALU instruction and emit it.
 *
 * Prefer emitting full ALU instructions, so that when we reach a point
 * where no full ALU instruction can be emitted, we have more candidates
 * for RGB/Alpha pairing.
 */
static void emit_alu(struct pair_state *s)
{
	struct radeon_pair_instruction pair;

	if (s->ReadyFullALU || !(s->ReadyRGB && s->ReadyAlpha)) {
		int ip;
		if (s->ReadyFullALU) {
			ip = s->ReadyFullALU - s->Instructions;
			s->ReadyFullALU = s->ReadyFullALU->NextReady;
		} else if (s->ReadyRGB) {
			ip = s->ReadyRGB - s->Instructions;
			s->ReadyRGB = s->ReadyRGB->NextReady;
		} else {
			ip = s->ReadyAlpha - s->Instructions;
			s->ReadyAlpha = s->ReadyAlpha->NextReady;
		}

		_mesa_bzero(&pair, sizeof(pair));
		fill_instruction_into_pair(s, &pair, ip);
		fill_dest_into_pair(s, &pair, ip);
		commit_instruction(s, ip);
	} else {
		struct pair_state_instruction **prgb;
		struct pair_state_instruction **palpha;

		/* Some pairings might fail because they require too
		 * many source slots; try all possible pairings if necessary */
		for(prgb = &s->ReadyRGB; *prgb; prgb = &(*prgb)->NextReady) {
			for(palpha = &s->ReadyAlpha; *palpha; palpha = &(*palpha)->NextReady) {
				int rgbip = *prgb - s->Instructions;
				int alphaip = *palpha - s->Instructions;
				_mesa_bzero(&pair, sizeof(pair));
				fill_instruction_into_pair(s, &pair, rgbip);
				if (!fill_instruction_into_pair(s, &pair, alphaip))
					continue;
				*prgb = (*prgb)->NextReady;
				*palpha = (*palpha)->NextReady;
				fill_dest_into_pair(s, &pair, rgbip);
				fill_dest_into_pair(s, &pair, alphaip);
				commit_instruction(s, rgbip);
				commit_instruction(s, alphaip);
				goto success;
			}
		}

		/* No success in pairing; just take the first RGB instruction */
		int ip = s->ReadyRGB - s->Instructions;
		s->ReadyRGB = s->ReadyRGB->NextReady;
		_mesa_bzero(&pair, sizeof(pair));
		fill_instruction_into_pair(s, &pair, ip);
		fill_dest_into_pair(s, &pair, ip);
		commit_instruction(s, ip);
	success: ;
	}

	if (s->Debug)
		radeonPrintPairInstruction(&pair);

	s->Error = s->Error || !s->Handler->EmitPaired(s->UserData, &pair);
}


GLboolean radeonPairProgram(GLcontext *ctx, struct gl_program *program,
	const struct radeon_pair_handler* handler, void *userdata)
{
	struct pair_state s;

	_mesa_bzero(&s, sizeof(s));
	s.Ctx = ctx;
	s.Program = program;
	s.Handler = handler;
	s.UserData = userdata;
	s.Debug = (RADEON_DEBUG & DEBUG_PIXEL) ? GL_TRUE : GL_FALSE;
	s.Verbose = GL_FALSE && s.Debug;

	s.Instructions = (struct pair_state_instruction*)_mesa_calloc(
		sizeof(struct pair_state_instruction)*s.Program->NumInstructions);
	s.ValuePool = (struct reg_value*)_mesa_calloc(sizeof(struct reg_value)*s.Program->NumInstructions*4);
	s.ReaderPool = (struct reg_value_reader*)_mesa_calloc(
		sizeof(struct reg_value_reader)*s.Program->NumInstructions*12);

	if (s.Debug)
		_mesa_printf("Emit paired program\n");

	scan_instructions(&s);
	allocate_input_registers(&s);

	while(!s.Error &&
	      (s.ReadyTEX || s.ReadyRGB || s.ReadyAlpha || s.ReadyFullALU)) {
		if (s.ReadyTEX)
			emit_all_tex(&s);

		while(s.ReadyFullALU || s.ReadyRGB || s.ReadyAlpha)
			emit_alu(&s);
	}

	if (s.Debug)
		_mesa_printf(" END\n");

	_mesa_free(s.Instructions);
	_mesa_free(s.ValuePool);
	_mesa_free(s.ReaderPool);

	return !s.Error;
}


static void print_pair_src(int i, struct radeon_pair_instruction_source* src)
{
	_mesa_printf("  Src%i = %s[%i]", i, src->Constant ? "CNST" : "TEMP", src->Index);
}

static const char* opcode_string(GLuint opcode)
{
	if (opcode == OPCODE_REPL_ALPHA)
		return "SOP";
	else
		return _mesa_opcode_string(opcode);
}

static int num_pairinst_args(GLuint opcode)
{
	if (opcode == OPCODE_REPL_ALPHA)
		return 0;
	else
		return _mesa_num_inst_src_regs(opcode);
}

static char swizzle_char(GLuint swz)
{
	switch(swz) {
	case SWIZZLE_X: return 'x';
	case SWIZZLE_Y: return 'y';
	case SWIZZLE_Z: return 'z';
	case SWIZZLE_W: return 'w';
	case SWIZZLE_ZERO: return '0';
	case SWIZZLE_ONE: return '1';
	case SWIZZLE_NIL: return '_';
	default: return '?';
	}
}

void radeonPrintPairInstruction(struct radeon_pair_instruction *inst)
{
	int nargs;
	int i;

	_mesa_printf("       RGB:  ");
	for(i = 0; i < 3; ++i) {
		if (inst->RGB.Src[i].Used)
			print_pair_src(i, inst->RGB.Src + i);
	}
	_mesa_printf("\n");
	_mesa_printf("       Alpha:");
	for(i = 0; i < 3; ++i) {
		if (inst->Alpha.Src[i].Used)
			print_pair_src(i, inst->Alpha.Src + i);
	}
	_mesa_printf("\n");

	_mesa_printf("  %s%s", opcode_string(inst->RGB.Opcode), inst->RGB.Saturate ? "_SAT" : "");
	if (inst->RGB.WriteMask)
		_mesa_printf(" TEMP[%i].%s%s%s", inst->RGB.DestIndex,
			(inst->RGB.WriteMask & 1) ? "x" : "",
			(inst->RGB.WriteMask & 2) ? "y" : "",
			(inst->RGB.WriteMask & 4) ? "z" : "");
	if (inst->RGB.OutputWriteMask)
		_mesa_printf(" COLOR.%s%s%s",
			(inst->RGB.OutputWriteMask & 1) ? "x" : "",
			(inst->RGB.OutputWriteMask & 2) ? "y" : "",
			(inst->RGB.OutputWriteMask & 4) ? "z" : "");
	nargs = num_pairinst_args(inst->RGB.Opcode);
	for(i = 0; i < nargs; ++i) {
		const char* abs = inst->RGB.Arg[i].Abs ? "|" : "";
		const char* neg = inst->RGB.Arg[i].Negate ? "-" : "";
		_mesa_printf(", %s%sSrc%i.%c%c%c%s", neg, abs, inst->RGB.Arg[i].Source,
			swizzle_char(GET_SWZ(inst->RGB.Arg[i].Swizzle, 0)),
			swizzle_char(GET_SWZ(inst->RGB.Arg[i].Swizzle, 1)),
			swizzle_char(GET_SWZ(inst->RGB.Arg[i].Swizzle, 2)),
			abs);
	}
	_mesa_printf("\n");

	_mesa_printf("  %s%s", opcode_string(inst->Alpha.Opcode), inst->Alpha.Saturate ? "_SAT" : "");
	if (inst->Alpha.WriteMask)
		_mesa_printf(" TEMP[%i].w", inst->Alpha.DestIndex);
	if (inst->Alpha.OutputWriteMask)
		_mesa_printf(" COLOR.w");
	if (inst->Alpha.DepthWriteMask)
		_mesa_printf(" DEPTH.w");
	nargs = num_pairinst_args(inst->Alpha.Opcode);
	for(i = 0; i < nargs; ++i) {
		const char* abs = inst->Alpha.Arg[i].Abs ? "|" : "";
		const char* neg = inst->Alpha.Arg[i].Negate ? "-" : "";
		_mesa_printf(", %s%sSrc%i.%c%s", neg, abs, inst->Alpha.Arg[i].Source,
			swizzle_char(inst->Alpha.Arg[i].Swizzle), abs);
	}
	_mesa_printf("\n");
}