github/ROCm
1
1
Fork 0
mirror of https://github.com/ROCm/ROCm.git synced 2026-04-27 03:01:52 -04:00
Code Issues Packages Projects Releases Wiki Activity

upgrade llvm to b1115f8c (NFC) (#2403)

Browse Source 
Co-authored-by: Thomas Raoux <thomas.raoux@openai.com>
Co-authored-by: Keren Zhou <kerenzhou@openai.com>
Co-authored-by: Phil Tillet <phil@openai.com>
This commit is contained in:
Mehdi Amini
2023-10-16 16:38:49 -07:00
committed by GitHub
parent 87a223d76f
commit 721897fcc4
36 changed files with 128 additions and 109 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lib/Dialect/Triton/IR/Ops.cpp
View File

@@ -1,9 +1,9 @@
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinAttributes.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/FunctionImplementation.h"
#include "mlir/IR/FunctionInterfaces.h"
#include "mlir/IR/OperationSupport.h"
#include "mlir/Interfaces/FunctionImplementation.h"
#include "mlir/Interfaces/FunctionInterfaces.h"
#include "triton/Dialect/Triton/IR/Dialect.h"
#include "triton/Dialect/Triton/IR/Types.h"

2
lib/Dialect/TritonGPU/IR/CMakeLists.txt
View File

@@ -8,6 +8,6 @@ add_mlir_dialect_library(TritonGPUIR
TritonGPUAttrDefsIncGen
LINK_LIBS PUBLIC
MLIRGPUOps
MLIRGPUDialect
TritonIR
)

4
lib/Dialect/TritonGPU/IR/Dialect.cpp
View File

@@ -1360,7 +1360,7 @@ ParseResult parseInsertSliceOp(OpAsmParser &parser, OperationState &result) {
result.operands))
return failure();
// Deduce operand_segment_sizes from the number of the operands.
// Deduce operandSegmentSizes from the number of the operands.
auto operandSegmentSizesAttrName =
OpT::getOperandSegmentSizesAttrName(result.name);
result.addAttribute(
@@ -1373,7 +1373,7 @@ template <class OpT>
void printInsertSliceOp(OpAsmPrinter &printer, OpT insertSliceOp) {
printer << " ";
printer << insertSliceOp.getOperation()->getOperands();
// "operand_segment_sizes" can be deduced, so we don't print it.
// "operandSegmentSizes" can be deduced, so we don't print it.
printer.printOptionalAttrDict(
insertSliceOp->getAttrs(),
{insertSliceOp.getOperandSegmentSizesAttrName()});

12
lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp
View File

@@ -117,7 +117,10 @@ class BlockedToMMA : public mlir::RewritePattern {
int finalBitWidth = getElementTypeOrSelf(x).getIntOrFloatBitWidth();
int origBitWidth = finalBitWidth;
SetVector<Operation *> slice;
mlir::getBackwardSlice(x, &slice, bwdFilter);
mlir::BackwardSliceOptions opt;
opt.omitBlockArguments = true;
opt.filter = bwdFilter;
getBackwardSlice(x, &slice, opt);
Operation *firstOp = slice.empty() ? nullptr : *slice.begin();
if (firstOp)
if (Value arg = firstOp->getOperand(0))
@@ -213,8 +216,11 @@ public:
if (versionMajor == 1) {
SetVector<Operation *> aBwdSlices, bBwdSlices;
auto isCvt = [](Operation *op) { return isa<ConvertLayoutOp>(op); };
getBackwardSlice(a, &aBwdSlices, {isCvt});
getBackwardSlice(b, &bBwdSlices, {isCvt});
mlir::BackwardSliceOptions opt;
opt.omitBlockArguments = true;
opt.filter = isCvt;
getBackwardSlice(a, &aBwdSlices, opt);
getBackwardSlice(b, &bBwdSlices, opt);
// get the source of the first conversion found in slices
auto getCvtArgOrder = [](Operation *op) {
return cast<ConvertLayoutOp>(op)

4
lib/Dialect/TritonGPU/Transforms/OptimizeDotOperands.cpp
View File

@@ -98,7 +98,9 @@ public:
// and all operations between the load and the conversion
// should be layout preserving
SetVector<Operation *> slice;
getBackwardSlice(op, &slice);
mlir::BackwardSliceOptions opt;
opt.omitBlockArguments = true;
getBackwardSlice(op, &slice, opt);
int loadIdx = -1;
bool checkOp = false;
for (int i = 0; i < slice.size(); i++) {

2
lib/Dialect/TritonNvidiaGPU/Transforms/FenceInsertion.cpp
View File

@@ -91,7 +91,7 @@ private:
// suport ForOp only
if (auto forOp = dyn_cast<scf::ForOp>(argOwner)) {
// prologue
auto iterOperands = forOp.getIterOperands();
auto iterOperands = forOp.getInitArgs();
if (argNum == 0)
return false;
if (dependOnSharedEncOperand(iterOperands[argNum - 1]))

13
lib/Dialect/TritonNvidiaGPU/Transforms/PlanCTA.cpp
View File

@@ -628,12 +628,13 @@ bool CTAPlanner::isElementwiseOp(Operation *op) const {
arith::CeilDivUIOp, arith::DivFOp, arith::DivSIOp,
arith::DivUIOp, arith::ExtFOp, arith::ExtSIOp, arith::ExtUIOp,
arith::FloorDivSIOp, arith::FPToSIOp, arith::FPToUIOp,
arith::MaxFOp, arith::MaxSIOp, arith::MaxUIOp, arith::MinFOp,
arith::MinSIOp, arith::MinUIOp, arith::MulFOp, arith::MulIOp,
arith::NegFOp, arith::OrIOp, arith::RemFOp, arith::RemSIOp,
arith::RemUIOp, arith::ShLIOp, arith::ShRSIOp, arith::ShRUIOp,
arith::SIToFPOp, arith::SubFOp, arith::SubIOp, arith::TruncFOp,
arith::TruncIOp, arith::UIToFPOp, arith::XOrIOp>(op))
arith::MaximumFOp, arith::MaxSIOp, arith::MaxUIOp,
arith::MinimumFOp, arith::MinSIOp, arith::MinUIOp,
arith::MulFOp, arith::MulIOp, arith::NegFOp, arith::OrIOp,
arith::RemFOp, arith::RemSIOp, arith::RemUIOp, arith::ShLIOp,
arith::ShRSIOp, arith::ShRUIOp, arith::SIToFPOp, arith::SubFOp,
arith::SubIOp, arith::TruncFOp, arith::TruncIOp,
arith::UIToFPOp, arith::XOrIOp>(op))
return true;
if (llvm::isa<math::AbsFOp, math::AbsIOp, math::AtanOp, math::Atan2Op,
math::CeilOp, math::CopySignOp, math::CosOp, math::SinOp,

4
lib/Dialect/TritonNvidiaGPU/Transforms/WSDecomposing.cpp
View File

@@ -220,7 +220,9 @@ public:
SetVector<Operation *> backwardSlice;
mod.walk([&](triton::MakeTensorPtrOp op) -> void {
assert(isa<triton::FuncOp>(op->getParentOp()));
getBackwardSlice(op.getOperation(), &backwardSlice);
mlir::BackwardSliceOptions opt;
opt.omitBlockArguments = true;
getBackwardSlice(op.getOperation(), &backwardSlice, opt);
op->removeAttr("async_agent");
});
for (auto op : backwardSlice) {

2
lib/Dialect/TritonNvidiaGPU/Transforms/WSPipeline.cpp
View File

@@ -259,7 +259,7 @@ scf::ForOp createNewMathLoop(scf::ForOp forOp, int numStages,
// 3. create newLoopArgs
SmallVector<Value> newLoopArgs;
for (auto operand : forOp.getIterOperands())
for (auto operand : forOp.getInitArgs())
newLoopArgs.push_back(operand);
builder.setInsertionPoint(forOp);