github/ROCm
1
1
Fork 0
mirror of https://github.com/ROCm/ROCm.git synced 2026-04-05 03:01:17 -04:00
Code Issues Packages Projects Releases Wiki Activity

[BACKEND] Init values before load to avoid ptxas issues (#1396)

Browse Source
This commit is contained in:
Keren Zhou
2023-03-23 20:24:03 -04:00
committed by GitHub
parent 3239c93a93
commit c9f47d9094
2 changed files with 25 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
lib/Conversion/TritonGPUToLLVM/LoadStoreOpToLLVM.cpp
View File

@@ -122,6 +122,7 @@ struct LoadOpConversion
const size_t width = std::min(totalWidth, maxWordWidth);
const size_t nWords = std::max<size_t>(1, totalWidth / width);
const size_t wordNElems = width / valueElemNbits;
const size_t movWidth = width < 16 ? 16 : width;
assert(wordNElems * nWords * numVecs == numElems);
// TODO(Superjomn) Add cache policy fields to StoreOp.
@@ -137,11 +138,18 @@ struct LoadOpConversion
const std::string writeConstraint =
(width == 64) ? "=l" : ((width == 32) ? "=r" : "=c");
PTXInstr &init =
ptxBuilder.create<>("mov")->o("u" + std::to_string(movWidth));
PTXInstr::Operand *zero = ptxBuilder.newConstantOperand(0);
// prepare asm operands
auto *dstsOpr = ptxBuilder.newListOperand();
for (size_t wordIdx = 0; wordIdx < nWords; ++wordIdx) {
auto *opr = ptxBuilder.newOperand(writeConstraint); // =r operations
dstsOpr->listAppend(opr);
// Initialize the destination register, otherwise the register will
// be undefined if the predicate is false.
init(opr, zero);
}
auto *addrOpr =
@@ -175,7 +183,6 @@ struct LoadOpConversion
if (other) {
for (size_t ii = 0; ii < nWords; ++ii) {
// PTX doesn't support mov.u8, so we need to use mov.u16
auto movWidth = width < 16 ? 16 : width;
PTXInstr &mov =
ptxBuilder.create<>("mov")->o("u" + std::to_string(movWidth));