/**
* Called when a new variant is needed, we need to translate
* the ATI fragment shader to TGSI
*/
enum pipe_error
st_translate_atifs_program(
struct ureg_program *ureg,
struct ati_fragment_shader *atifs,
struct gl_program *program,
GLuint numInputs,
const GLuint inputMapping[],
const ubyte inputSemanticName[],
const ubyte inputSemanticIndex[],
const GLuint interpMode[],
GLuint numOutputs,
const GLuint outputMapping[],
const ubyte outputSemanticName[],
const ubyte outputSemanticIndex[])
{
enum pipe_error ret = PIPE_OK;
unsigned pass, i, r;
struct st_translate translate, *t;
t = &translate;
memset(t, 0, sizeof *t);
t->inputMapping = inputMapping;
t->outputMapping = outputMapping;
t->ureg = ureg;
t->atifs = atifs;
/*
* Declare input attributes.
*/
for (i = 0; i < numInputs; i++) {
t->inputs[i] = ureg_DECL_fs_input(ureg,
inputSemanticName[i],
inputSemanticIndex[i],
interpMode[i]);
}
/*
* Declare output attributes:
* we always have numOutputs=1 and it's FRAG_RESULT_COLOR
*/
t->outputs[0] = ureg_DECL_output(ureg,
TGSI_SEMANTIC_COLOR,
outputSemanticIndex[0]);
/* Emit constants and immediates. Mesa uses a single index space
* for these, so we put all the translated regs in t->constants.
*/
if (program->Parameters) {
t->constants = calloc(program->Parameters->NumParameters,
sizeof t->constants[0]);
if (t->constants == NULL) {
ret = PIPE_ERROR_OUT_OF_MEMORY;
goto out;
}
for (i = 0; i < program->Parameters->NumParameters; i++) {
switch (program->Parameters->Parameters[i].Type) {
case PROGRAM_STATE_VAR:
case PROGRAM_UNIFORM:
t->constants[i] = ureg_DECL_constant(ureg, i);
break;
case PROGRAM_CONSTANT:
t->constants[i] =
ureg_DECL_immediate(ureg,
(const float*)program->Parameters->ParameterValues[i],
4);
break;
default:
break;
}
}
}
/* texture samplers */
for (i = 0; i < MAX_NUM_FRAGMENT_REGISTERS_ATI; i++) {
if (program->SamplersUsed & (1 << i)) {
t->samplers[i] = ureg_DECL_sampler(ureg, i);
/* the texture target is still unknown, it will be fixed in the draw call */
ureg_DECL_sampler_view(ureg, i, TGSI_TEXTURE_2D,
TGSI_RETURN_TYPE_FLOAT,
TGSI_RETURN_TYPE_FLOAT,
TGSI_RETURN_TYPE_FLOAT,
TGSI_RETURN_TYPE_FLOAT);
}
}
/* emit instructions */
for (pass = 0; pass < atifs->NumPasses; pass++) {
t->current_pass = pass;
for (r = 0; r < MAX_NUM_FRAGMENT_REGISTERS_ATI; r++) {
struct atifs_setupinst *texinst = &atifs->SetupInst[pass][r];
compile_setupinst(t, r, texinst);
}
for (i = 0; i < atifs->numArithInstr[pass]; i++) {
struct atifs_instruction *inst = &atifs->Instructions[pass][i];
compile_instruction(t, inst);
}
}
finalize_shader(t, atifs->NumPasses);
out:
free(t->constants);
if (t->error) {
debug_printf("%s: translate error flag set\n", __func__);
}
return ret;
}
/**
* Translate Mesa program to TGSI format.
* \param program the program to translate
* \param numInputs number of input registers used
* \param inputMapping maps Mesa fragment program inputs to TGSI generic
* input indexes
* \param inputSemanticName the TGSI_SEMANTIC flag for each input
* \param inputSemanticIndex the semantic index (ex: which texcoord) for
* each input
* \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
* \param numOutputs number of output registers used
* \param outputMapping maps Mesa fragment program outputs to TGSI
* generic outputs
* \param outputSemanticName the TGSI_SEMANTIC flag for each output
* \param outputSemanticIndex the semantic index (ex: which texcoord) for
* each output
*
* \return PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
*/
enum pipe_error
st_translate_mesa_program(
struct gl_context *ctx,
uint procType,
struct ureg_program *ureg,
const struct gl_program *program,
GLuint numInputs,
const GLuint inputMapping[],
const ubyte inputSemanticName[],
const ubyte inputSemanticIndex[],
const GLuint interpMode[],
GLuint numOutputs,
const GLuint outputMapping[],
const ubyte outputSemanticName[],
const ubyte outputSemanticIndex[],
boolean passthrough_edgeflags,
boolean clamp_color)
{
struct st_translate translate, *t;
unsigned i;
enum pipe_error ret = PIPE_OK;
assert(numInputs <= ARRAY_SIZE(t->inputs));
assert(numOutputs <= ARRAY_SIZE(t->outputs));
t = &translate;
memset(t, 0, sizeof *t);
t->procType = procType;
t->inputMapping = inputMapping;
t->outputMapping = outputMapping;
t->ureg = ureg;
/*_mesa_print_program(program);*/
/*
* Declare input attributes.
*/
if (procType == TGSI_PROCESSOR_FRAGMENT) {
for (i = 0; i < numInputs; i++) {
t->inputs[i] = ureg_DECL_fs_input(ureg,
inputSemanticName[i],
inputSemanticIndex[i],
interpMode[i]);
}
if (program->InputsRead & VARYING_BIT_POS) {
/* Must do this after setting up t->inputs, and before
* emitting constant references, below:
*/
emit_wpos(st_context(ctx), t, program, ureg);
}
if (program->InputsRead & VARYING_BIT_FACE) {
emit_face_var( t, program );
}
/*
* Declare output attributes.
*/
for (i = 0; i < numOutputs; i++) {
switch (outputSemanticName[i]) {
case TGSI_SEMANTIC_POSITION:
t->outputs[i] = ureg_DECL_output( ureg,
TGSI_SEMANTIC_POSITION, /* Z / Depth */
outputSemanticIndex[i] );
t->outputs[i] = ureg_writemask( t->outputs[i],
TGSI_WRITEMASK_Z );
break;
case TGSI_SEMANTIC_STENCIL:
t->outputs[i] = ureg_DECL_output( ureg,
TGSI_SEMANTIC_STENCIL, /* Stencil */
outputSemanticIndex[i] );
t->outputs[i] = ureg_writemask( t->outputs[i],
TGSI_WRITEMASK_Y );
break;
case TGSI_SEMANTIC_COLOR:
t->outputs[i] = ureg_DECL_output( ureg,
TGSI_SEMANTIC_COLOR,
outputSemanticIndex[i] );
break;
default:
debug_assert(0);
return 0;
}
}
}
else if (procType == TGSI_PROCESSOR_GEOMETRY) {
for (i = 0; i < numInputs; i++) {
t->inputs[i] = ureg_DECL_input(ureg,
inputSemanticName[i],
inputSemanticIndex[i], 0, 1);
}
for (i = 0; i < numOutputs; i++) {
t->outputs[i] = ureg_DECL_output( ureg,
outputSemanticName[i],
outputSemanticIndex[i] );
}
}
else {
assert(procType == TGSI_PROCESSOR_VERTEX);
for (i = 0; i < numInputs; i++) {
t->inputs[i] = ureg_DECL_vs_input(ureg, i);
}
for (i = 0; i < numOutputs; i++) {
t->outputs[i] = ureg_DECL_output( ureg,
outputSemanticName[i],
outputSemanticIndex[i] );
if (outputSemanticName[i] == TGSI_SEMANTIC_FOG) {
/* force register to contain a fog coordinate in the form (F, 0, 0, 1). */
ureg_MOV(ureg,
ureg_writemask(t->outputs[i], TGSI_WRITEMASK_YZW),
ureg_imm4f(ureg, 0.0f, 0.0f, 0.0f, 1.0f));
t->outputs[i] = ureg_writemask(t->outputs[i], TGSI_WRITEMASK_X);
}
}
if (passthrough_edgeflags)
emit_edgeflags( t, program );
}
/* Declare address register.
*/
if (program->NumAddressRegs > 0) {
debug_assert( program->NumAddressRegs == 1 );
t->address[0] = ureg_DECL_address( ureg );
}
/* Declare misc input registers
*/
{
GLbitfield sysInputs = program->SystemValuesRead;
unsigned numSys = 0;
for (i = 0; sysInputs; i++) {
if (sysInputs & (1 << i)) {
unsigned semName = _mesa_sysval_to_semantic[i];
t->systemValues[i] = ureg_DECL_system_value(ureg, numSys, semName, 0);
if (semName == TGSI_SEMANTIC_INSTANCEID ||
semName == TGSI_SEMANTIC_VERTEXID) {
/* From Gallium perspective, these system values are always
* integer, and require native integer support. However, if
* native integer is supported on the vertex stage but not the
* pixel stage (e.g, i915g + draw), Mesa will generate IR that
* assumes these system values are floats. To resolve the
* inconsistency, we insert a U2F.
*/
struct st_context *st = st_context(ctx);
struct pipe_screen *pscreen = st->pipe->screen;
assert(procType == TGSI_PROCESSOR_VERTEX);
assert(pscreen->get_shader_param(pscreen, PIPE_SHADER_VERTEX, PIPE_SHADER_CAP_INTEGERS));
(void) pscreen; /* silence non-debug build warnings */
if (!ctx->Const.NativeIntegers) {
struct ureg_dst temp = ureg_DECL_local_temporary(t->ureg);
ureg_U2F( t->ureg, ureg_writemask(temp, TGSI_WRITEMASK_X), t->systemValues[i]);
t->systemValues[i] = ureg_scalar(ureg_src(temp), 0);
}
}
numSys++;
sysInputs &= ~(1 << i);
}
}
}
if (program->IndirectRegisterFiles & (1 << PROGRAM_TEMPORARY)) {
/* If temps are accessed with indirect addressing, declare temporaries
* in sequential order. Else, we declare them on demand elsewhere.
*/
for (i = 0; i < program->NumTemporaries; i++) {
/* XXX use TGSI_FILE_TEMPORARY_ARRAY when it's supported by ureg */
t->temps[i] = ureg_DECL_temporary( t->ureg );
}
}
/* Emit constants and immediates. Mesa uses a single index space
* for these, so we put all the translated regs in t->constants.
*/
if (program->Parameters) {
t->constants = calloc( program->Parameters->NumParameters,
sizeof t->constants[0] );
if (t->constants == NULL) {
ret = PIPE_ERROR_OUT_OF_MEMORY;
goto out;
}
for (i = 0; i < program->Parameters->NumParameters; i++) {
switch (program->Parameters->Parameters[i].Type) {
case PROGRAM_STATE_VAR:
case PROGRAM_UNIFORM:
t->constants[i] = ureg_DECL_constant( ureg, i );
break;
/* Emit immediates only when there's no indirect addressing of
* the const buffer.
* FIXME: Be smarter and recognize param arrays:
* indirect addressing is only valid within the referenced
* array.
*/
case PROGRAM_CONSTANT:
if (program->IndirectRegisterFiles & PROGRAM_ANY_CONST)
t->constants[i] = ureg_DECL_constant( ureg, i );
else
t->constants[i] =
ureg_DECL_immediate( ureg,
(const float*) program->Parameters->ParameterValues[i],
4 );
break;
default:
break;
}
}
}
/* texture samplers */
for (i = 0; i < ctx->Const.Program[MESA_SHADER_FRAGMENT].MaxTextureImageUnits; i++) {
if (program->SamplersUsed & (1 << i)) {
t->samplers[i] = ureg_DECL_sampler( ureg, i );
}
}
/* Emit each instruction in turn:
*/
for (i = 0; i < program->NumInstructions; i++) {
set_insn_start( t, ureg_get_instruction_number( ureg ));
compile_instruction( ctx, t, &program->Instructions[i], clamp_color );
}
/* Fix up all emitted labels:
*/
for (i = 0; i < t->labels_count; i++) {
ureg_fixup_label( ureg,
t->labels[i].token,
t->insn[t->labels[i].branch_target] );
}
out:
free(t->insn);
free(t->labels);
free(t->constants);
if (t->error) {
debug_printf("%s: translate error flag set\n", __func__);
}
return ret;
}
/**
* Translate Mesa program to TGSI format.
* \param program the program to translate
* \param numInputs number of input registers used
* \param inputMapping maps Mesa fragment program inputs to TGSI generic
* input indexes
* \param inputSemanticName the TGSI_SEMANTIC flag for each input
* \param inputSemanticIndex the semantic index (ex: which texcoord) for each input
* \param interpMode the TGSI_INTERPOLATE_LINEAR/PERSP mode for each input
* \param numOutputs number of output registers used
* \param outputMapping maps Mesa fragment program outputs to TGSI
* generic outputs
* \param outputSemanticName the TGSI_SEMANTIC flag for each output
* \param outputSemanticIndex the semantic index (ex: which texcoord) for each output
* \param tokens array to store translated tokens in
* \param maxTokens size of the tokens array
*
* \return number of tokens placed in 'tokens' buffer, or zero if error
*/
GLuint
st_translate_mesa_program(
GLcontext *ctx,
uint procType,
const struct gl_program *program,
GLuint numInputs,
const GLuint inputMapping[],
const ubyte inputSemanticName[],
const ubyte inputSemanticIndex[],
const GLuint interpMode[],
const GLbitfield inputFlags[],
GLuint numOutputs,
const GLuint outputMapping[],
const ubyte outputSemanticName[],
const ubyte outputSemanticIndex[],
const GLbitfield outputFlags[],
struct tgsi_token *tokens,
GLuint maxTokens )
{
GLuint i;
GLuint ti; /* token index */
struct tgsi_header *header;
struct tgsi_processor *processor;
GLuint preamble_size = 0;
GLuint immediates[1000];
GLuint numImmediates = 0;
GLboolean insideSubroutine = GL_FALSE;
GLboolean indirectAccess = GL_FALSE;
GLboolean tempsUsed[MAX_PROGRAM_TEMPS + 1];
GLint wposTemp = -1, winHeightConst = -1;
assert(procType == TGSI_PROCESSOR_FRAGMENT ||
procType == TGSI_PROCESSOR_VERTEX);
find_temporaries(program, tempsUsed);
if (procType == TGSI_PROCESSOR_FRAGMENT) {
if (program->InputsRead & FRAG_BIT_WPOS) {
/* Fragment program uses fragment position input.
* Need to replace instances of INPUT[WPOS] with temp T
* where T = INPUT[WPOS] by y is inverted.
*/
static const gl_state_index winSizeState[STATE_LENGTH]
= { STATE_INTERNAL, STATE_FB_SIZE, 0, 0, 0 };
winHeightConst = _mesa_add_state_reference(program->Parameters,
winSizeState);
wposTemp = find_free_temporary(tempsUsed);
}
}
*(struct tgsi_version *) &tokens[0] = tgsi_build_version();
header = (struct tgsi_header *) &tokens[1];
*header = tgsi_build_header();
processor = (struct tgsi_processor *) &tokens[2];
*processor = tgsi_build_processor( procType, header );
ti = 3;
/*
* Declare input attributes.
*/
if (procType == TGSI_PROCESSOR_FRAGMENT) {
for (i = 0; i < numInputs; i++) {
struct tgsi_full_declaration fulldecl;
fulldecl = make_input_decl(i,
GL_TRUE, interpMode[i],
TGSI_WRITEMASK_XYZW,
GL_TRUE, inputSemanticName[i],
inputSemanticIndex[i],
inputFlags[i]);
ti += tgsi_build_full_declaration(&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
}
}
else {
/* vertex prog */
/* XXX: this could probaby be merged with the clause above.
* the only difference is the semantic tags.
*/
for (i = 0; i < numInputs; i++) {
struct tgsi_full_declaration fulldecl;
fulldecl = make_input_decl(i,
GL_FALSE, 0,
TGSI_WRITEMASK_XYZW,
GL_FALSE, 0, 0,
inputFlags[i]);
ti += tgsi_build_full_declaration(&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
}
}
/*
* Declare output attributes.
*/
if (procType == TGSI_PROCESSOR_FRAGMENT) {
for (i = 0; i < numOutputs; i++) {
struct tgsi_full_declaration fulldecl;
switch (outputSemanticName[i]) {
case TGSI_SEMANTIC_POSITION:
fulldecl = make_output_decl(i,
TGSI_SEMANTIC_POSITION, /* Z / Depth */
outputSemanticIndex[i],
TGSI_WRITEMASK_Z,
outputFlags[i]);
break;
case TGSI_SEMANTIC_COLOR:
fulldecl = make_output_decl(i,
TGSI_SEMANTIC_COLOR,
outputSemanticIndex[i],
TGSI_WRITEMASK_XYZW,
outputFlags[i]);
break;
default:
assert(0);
return 0;
}
ti += tgsi_build_full_declaration(&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
}
}
else {
/* vertex prog */
for (i = 0; i < numOutputs; i++) {
struct tgsi_full_declaration fulldecl;
fulldecl = make_output_decl(i,
outputSemanticName[i],
outputSemanticIndex[i],
TGSI_WRITEMASK_XYZW,
outputFlags[i]);
ti += tgsi_build_full_declaration(&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
}
}
/* temporary decls */
{
GLboolean inside_range = GL_FALSE;
GLuint start_range = 0;
tempsUsed[MAX_PROGRAM_TEMPS] = GL_FALSE;
for (i = 0; i < MAX_PROGRAM_TEMPS + 1; i++) {
if (tempsUsed[i] && !inside_range) {
inside_range = GL_TRUE;
start_range = i;
}
else if (!tempsUsed[i] && inside_range) {
struct tgsi_full_declaration fulldecl;
inside_range = GL_FALSE;
fulldecl = make_temp_decl( start_range, i - 1 );
ti += tgsi_build_full_declaration(
&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
}
}
}
/* Declare address register.
*/
if (program->NumAddressRegs > 0) {
struct tgsi_full_declaration fulldecl;
assert( program->NumAddressRegs == 1 );
fulldecl = make_addr_decl( 0, 0 );
ti += tgsi_build_full_declaration(
&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
indirectAccess = GL_TRUE;
}
/* immediates/literals */
memset(immediates, ~0, sizeof(immediates));
/* Emit immediates only when there is no address register in use.
* FIXME: Be smarter and recognize param arrays -- indirect addressing is
* only valid within the referenced array.
*/
if (program->Parameters && !indirectAccess) {
for (i = 0; i < program->Parameters->NumParameters; i++) {
if (program->Parameters->Parameters[i].Type == PROGRAM_CONSTANT) {
struct tgsi_full_immediate fullimm;
fullimm = make_immediate( program->Parameters->ParameterValues[i], 4 );
ti += tgsi_build_full_immediate(
&fullimm,
&tokens[ti],
header,
maxTokens - ti );
immediates[i] = numImmediates;
numImmediates++;
}
}
}
/* constant buffer refs */
if (program->Parameters) {
GLint start = -1, end = -1;
for (i = 0; i < program->Parameters->NumParameters; i++) {
GLboolean emit = (i == program->Parameters->NumParameters - 1);
GLboolean matches;
switch (program->Parameters->Parameters[i].Type) {
case PROGRAM_ENV_PARAM:
case PROGRAM_STATE_VAR:
case PROGRAM_NAMED_PARAM:
case PROGRAM_UNIFORM:
matches = GL_TRUE;
break;
case PROGRAM_CONSTANT:
matches = indirectAccess;
break;
default:
matches = GL_FALSE;
}
if (matches) {
if (start == -1) {
/* begin a sequence */
start = i;
end = i;
}
else {
/* continue sequence */
end = i;
}
}
else {
if (start != -1) {
/* end of sequence */
emit = GL_TRUE;
}
}
if (emit && start >= 0) {
struct tgsi_full_declaration fulldecl;
fulldecl = make_constant_decl( start, end );
ti += tgsi_build_full_declaration(
&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
start = end = -1;
}
}
}
/* texture samplers */
for (i = 0; i < ctx->Const.MaxTextureImageUnits; i++) {
if (program->SamplersUsed & (1 << i)) {
struct tgsi_full_declaration fulldecl;
fulldecl = make_sampler_decl( i );
ti += tgsi_build_full_declaration(
&fulldecl,
&tokens[ti],
header,
maxTokens - ti );
}
}
/* invert WPOS fragment input */
if (wposTemp >= 0) {
ti += emit_inverted_wpos(&tokens[ti], wposTemp, winHeightConst,
inputMapping[FRAG_ATTRIB_WPOS],
header, maxTokens - ti);
preamble_size = 2; /* two instructions added */
}
for (i = 0; i < program->NumInstructions; i++) {
struct tgsi_full_instruction fullinst;
compile_instruction(
&program->Instructions[i],
&fullinst,
inputMapping,
outputMapping,
immediates,
indirectAccess,
preamble_size,
procType,
&insideSubroutine,
wposTemp);
ti += tgsi_build_full_instruction(
&fullinst,
&tokens[ti],
header,
maxTokens - ti );
}
#if DEBUG
if(!tgsi_sanity_check(tokens)) {
debug_printf("Due to sanity check failure(s) above the following shader program is invalid:\n");
debug_printf("\nOriginal program:\n%s", program->String);
debug_printf("\nMesa program:\n");
_mesa_print_program(program);
debug_printf("\nTGSI program:\n");
tgsi_dump(tokens, 0);
assert(0);
}
#endif
return ti;
}