Skip to content

Commit 1577f9f

Browse files
nvnprogramnvnprogram
committed
Fix io conflicts in vertex shader
1 parent 05645cb commit 1577f9f

2 files changed

Lines changed: 77 additions & 38 deletions

File tree

‎mesa-imported/codegen/nv50_ir_from_tgsi.cpp‎

Lines changed: 73 additions & 34 deletions
Original file line numberDiff line numberDiff line change
@@ -1829,6 +1829,18 @@ class Converter : public BuildUtil
18291829
Stack breakBBs; // end of / after loop
18301830

18311831
Value *viewport;
1832+
1833+
struct VertexOutput {
1834+
VertexOutput(const tgsi::Instruction::DstRegister& d, int channel, Value *value, Value *pointer)
1835+
: dst(d), c(channel), val(value), ptr(pointer) {}
1836+
1837+
tgsi::Instruction::DstRegister dst;
1838+
int c;
1839+
Value *val;
1840+
Value *ptr;
1841+
};
1842+
1843+
std::vector<VertexOutput> vertexOutputs;
18321844
};
18331845

18341846
Symbol *
@@ -2248,7 +2260,8 @@ Converter::storeDst(const tgsi::Instruction::DstRegister dst, int c,
22482260
assert(!ptr);
22492261
mkOp2(OP_WRSV, TYPE_U32, NULL, dstToSym(dst, c), val);
22502262
} else
2251-
if (f == TGSI_FILE_OUTPUT && prog->getType() != Program::TYPE_FRAGMENT) {
2263+
if (f == TGSI_FILE_OUTPUT && prog->getType() != Program::TYPE_FRAGMENT && prog->getType() != Program::TYPE_VERTEX) {
2264+
// This condition now excludes both fragment and vertex shaders
22522265

22532266
if (ptr || (info->out[idx].mask & (1 << c))) {
22542267
/* Save the viewport index into a scratch register so that it can be
@@ -2262,6 +2275,10 @@ Converter::storeDst(const tgsi::Instruction::DstRegister dst, int c,
22622275
info->out[idx].patch;
22632276
}
22642277
} else
2278+
if (f == TGSI_FILE_OUTPUT && prog->getType() == Program::TYPE_VERTEX) {
2279+
// For vertex shaders, collect outputs to emit later
2280+
vertexOutputs.push_back(VertexOutput(dst, c, val, ptr));
2281+
} else
22652282
if (f == TGSI_FILE_TEMPORARY ||
22662283
f == TGSI_FILE_ADDRESS ||
22672284
f == TGSI_FILE_OUTPUT) {
@@ -3835,7 +3852,7 @@ Converter::handleInstruction(const struct tgsi_full_instruction *insn)
38353852
BasicBlock *epilogue = BasicBlock::get(func->cfgExit);
38363853
bb->cfg.attach(&epilogue->cfg, Graph::Edge::TREE);
38373854
setPosition(epilogue, true);
3838-
if (prog->getType() == Program::TYPE_FRAGMENT)
3855+
if (prog->getType() == Program::TYPE_FRAGMENT || prog->getType() == Program::TYPE_VERTEX)
38393856
exportOutputs();
38403857
if ((prog->getType() == Program::TYPE_VERTEX ||
38413858
prog->getType() == Program::TYPE_TESSELLATION_EVAL
@@ -4313,40 +4330,62 @@ Converter::handleUserClipPlanes()
43134330
void
43144331
Converter::exportOutputs()
43154332
{
4316-
if (info->io.alphaRefBase) {
4317-
for (unsigned int i = 0; i < info->numOutputs; ++i) {
4318-
if (info->out[i].sn != TGSI_SEMANTIC_COLOR ||
4319-
info->out[i].si != 0)
4320-
continue;
4321-
const unsigned int c = 3;
4322-
if (!oData.exists(sub.cur->values, i, c))
4323-
continue;
4324-
Value *val = oData.load(sub.cur->values, i, c, NULL);
4325-
if (!val)
4326-
continue;
4327-
4328-
Symbol *ref = mkSymbol(FILE_MEMORY_CONST, info->io.auxCBSlot,
4329-
TYPE_U32, info->io.alphaRefBase);
4330-
Value *pred = new_LValue(func, FILE_PREDICATE);
4331-
mkCmp(OP_SET, CC_TR, TYPE_U32, pred, TYPE_F32, val,
4332-
mkLoadv(TYPE_U32, ref, NULL))
4333-
->subOp = 1;
4334-
mkOp(OP_DISCARD, TYPE_NONE, NULL)->setPredicate(CC_NOT_P, pred);
4335-
}
4336-
}
4333+
if (prog->getType() == Program::TYPE_FRAGMENT || prog->getType() == Program::TYPE_VERTEX) {
4334+
if (prog->getType() == Program::TYPE_FRAGMENT) {
4335+
if (info->io.alphaRefBase) {
4336+
for (unsigned int i = 0; i < info->numOutputs; ++i) {
4337+
if (info->out[i].sn != TGSI_SEMANTIC_COLOR ||
4338+
info->out[i].si != 0)
4339+
continue;
4340+
const unsigned int c = 3;
4341+
if (!oData.exists(sub.cur->values, i, c))
4342+
continue;
4343+
Value *val = oData.load(sub.cur->values, i, c, NULL);
4344+
if (!val)
4345+
continue;
4346+
4347+
Symbol *ref = mkSymbol(FILE_MEMORY_CONST, info->io.auxCBSlot,
4348+
TYPE_U32, info->io.alphaRefBase);
4349+
Value *pred = new_LValue(func, FILE_PREDICATE);
4350+
mkCmp(OP_SET, CC_TR, TYPE_U32, pred, TYPE_F32, val,
4351+
mkLoadv(TYPE_U32, ref, NULL))
4352+
->subOp = 1;
4353+
mkOp(OP_DISCARD, TYPE_NONE, NULL)->setPredicate(CC_NOT_P, pred);
4354+
}
4355+
}
43374356

4338-
for (unsigned int i = 0; i < info->numOutputs; ++i) {
4339-
for (unsigned int c = 0; c < 4; ++c) {
4340-
if (!oData.exists(sub.cur->values, i, c))
4341-
continue;
4342-
Symbol *sym = mkSymbol(FILE_SHADER_OUTPUT, 0, TYPE_F32,
4343-
info->out[i].slot[c] * 4);
4344-
Value *val = oData.load(sub.cur->values, i, c, NULL);
4345-
if (val) {
4346-
if (info->out[i].sn == TGSI_SEMANTIC_POSITION)
4347-
mkOp1(OP_SAT, TYPE_F32, val, val);
4348-
mkStore(OP_EXPORT, TYPE_F32, sym, NULL, val);
4357+
for (unsigned int i = 0; i < info->numOutputs; ++i) {
4358+
for (unsigned int c = 0; c < 4; ++c) {
4359+
if (!oData.exists(sub.cur->values, i, c))
4360+
continue;
4361+
Symbol *sym = mkSymbol(FILE_SHADER_OUTPUT, 0, TYPE_F32,
4362+
info->out[i].slot[c] * 4);
4363+
Value *val = oData.load(sub.cur->values, i, c, NULL);
4364+
if (val) {
4365+
if (info->out[i].sn == TGSI_SEMANTIC_POSITION)
4366+
mkOp1(OP_SAT, TYPE_F32, val, val);
4367+
mkStore(OP_EXPORT, TYPE_F32, sym, NULL, val);
4368+
}
4369+
}
4370+
}
4371+
} else if (prog->getType() == Program::TYPE_VERTEX) {
4372+
// Emit all collected vertex outputs in the original format
4373+
for (size_t i = 0; i < vertexOutputs.size(); ++i) {
4374+
const VertexOutput &output = vertexOutputs[i];
4375+
const unsigned idx = output.dst.getIndex(0);
4376+
4377+
if (output.ptr || (info->out[idx].mask & (1 << output.c))) {
4378+
if (info->out[idx].sn == TGSI_SEMANTIC_VIEWPORT_INDEX &&
4379+
viewport != NULL) {
4380+
mkOp1(OP_MOV, TYPE_U32, viewport, output.val);
4381+
} else {
4382+
mkStore(OP_EXPORT, TYPE_U32, dstToSym(output.dst, output.c),
4383+
output.ptr, output.val)->perPatch = info->out[idx].patch;
4384+
}
4385+
}
43494386
}
4387+
// Clear the vector after emitting all outputs
4388+
vertexOutputs.clear();
43504389
}
43514390
}
43524391
}

‎mesa-imported/codegen/nv50_ir_peephole.cpp‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -3166,8 +3166,7 @@ MemoryOpt::runOpt(BasicBlock *bb)
31663166
continue;
31673167
if (ldst->perPatch) // TODO: create separate per-patch lists
31683168
continue;
3169-
3170-
if (isLoad) {
3169+
else if (isLoad) {
31713170
DataFile file = ldst->src(0).getFile();
31723171

31733172
// if ld l[]/g[] look for previous store to eliminate the reload
@@ -3183,9 +3182,10 @@ MemoryOpt::runOpt(BasicBlock *bb)
31833182
if (rec) {
31843183
if (!isAdjacent)
31853184
keep = !replaceLdFromLd(ldst, rec);
3186-
else
3185+
else if (ldst->op != OP_VFETCH) // edit: nx doesnt combine vfetch
31873186
// or combine a previous load with this one
31883187
keep = !combineLd(rec, ldst);
3188+
31893189
}
31903190
if (keep)
31913191
lockStores(ldst);
@@ -3194,7 +3194,7 @@ MemoryOpt::runOpt(BasicBlock *bb)
31943194
if (rec) {
31953195
if (!isAdjacent)
31963196
keep = !replaceStFromSt(ldst, rec);
3197-
else
3197+
else if (ldst->op != OP_EXPORT) // edit: nx doesnt combine export
31983198
keep = !combineSt(rec, ldst);
31993199
}
32003200
if (keep)

0 commit comments

Comments
 (0)