Skip to content

[TritonToLinalg](fix) fix zero-stride load boundary #1509

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
zhanwei33 wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

[TritonToLinalg](fix) fix zero-stride load boundary #1509

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
76 changes: 68 additions & 8 deletions third_party/ascend/lib/TritonToLinalg/LoadStoreConverter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,56 @@ using namespace triton;
const std::string MayImplicitTransposeWithLastAxisTAG =
"MayImplicitTransposeWithLastAxis";

namespace {

Value getRemappedOrOriginal(Value value, ConversionPatternRewriter &rewriter) {
if (Value remapped = rewriter.getRemappedValue(value))
return remapped;
return value;
}

bool hasStaticZeroStride(triton::MakeTensorPtrOp makeTensorPtrOp) {
return llvm::any_of(makeTensorPtrOp.getStrides(), [](Value stride) {
auto constantStride = getConstantIntValue(stride);
return constantStride.has_value() && constantStride.value() == 0;
});
}

SmallVector<OpFoldResult> getBoundarySizesFromMakeTensorPtr(
triton::MakeTensorPtrOp makeTensorPtrOp,
llvm::ArrayRef<int32_t> boundaryCheck, llvm::ArrayRef<int64_t> tileShape,
const Location &loc, ConversionPatternRewriter &rewriter) {
assert(makeTensorPtrOp.getShape().size() == tileShape.size());
assert(makeTensorPtrOp.getOffsets().size() == tileShape.size());

SmallVector<OpFoldResult> boundarySizes =
getAsIndexOpFoldResult(rewriter.getContext(), tileShape);
const OpFoldResult zero = rewriter.getIndexAttr(0);

for (size_t i = 0; i < tileShape.size(); ++i) {
if (llvm::find(boundaryCheck, i) == boundaryCheck.end())
continue;

OpFoldResult shape = getOpFoldResultOfLayoutInfo(
getRemappedOrOriginal(makeTensorPtrOp.getShape()[i], rewriter),
rewriter);
OpFoldResult offset = getOpFoldResultOfLayoutInfo(
getRemappedOrOriginal(makeTensorPtrOp.getOffsets()[i], rewriter),
rewriter);
OpFoldResult nonNegativeOffset =
maxOpFoldResult(offset, zero, loc, rewriter);
OpFoldResult remaining = maxOpFoldResult(
subOpFoldResult(shape, nonNegativeOffset, loc, rewriter), zero, loc,
rewriter);
boundarySizes[i] =
minOpFoldResult(boundarySizes[i], remaining, loc, rewriter);
}

return boundarySizes;
}

} // namespace

LogicalResult
AddPtrConverter::matchAndRewrite(triton::AddPtrOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Expand Down Expand Up @@ -325,13 +375,19 @@ LogicalResult LoadConverter::replaceMaskedLoadWithTensorOther(
LogicalResult
LoadConverter::matchAndRewrite(triton::LoadOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto makeTensorPtrOp = op.getPtr().getDefiningOp<triton::MakeTensorPtrOp>();
const bool hasZeroStrideMTP =
makeTensorPtrOp && hasStaticZeroStride(makeTensorPtrOp);

// Check if tt.load is modified by AddPtrConverter to a specified state.
if (checkModifiedByAddPtrConverter(op).succeeded()) {
if (!hasZeroStrideMTP && checkModifiedByAddPtrConverter(op).succeeded()) {
return continueModifyFromAddPtrConverter(op, adaptor, rewriter);
}

auto ptr = adaptor.getPtr();
Value ptr = hasZeroStrideMTP ? rewriter.getRemappedValue(op.getPtr())
: adaptor.getPtr();
if (!ptr)
return rewriter.notifyMatchFailure(op, "missing remapped tensor pointer");
auto mask = op.getMask();
auto other = op.getOther();
auto loc = op.getLoc();
Expand Down Expand Up @@ -367,7 +423,7 @@ LoadConverter::matchAndRewrite(triton::LoadOp op, OpAdaptor adaptor,
}

int64_t lastStride = -1;
if (isa<BlockArgument>(ptr)) {
if (!hasZeroStrideMTP && isa<BlockArgument>(ptr)) {
auto u = ptr;
while (auto blkArg = dyn_cast<BlockArgument>(u)) {
if (auto forOp = dyn_cast<scf::ForOp>(blkArg.getOwner()->getParentOp())) {
Expand All @@ -392,14 +448,15 @@ LoadConverter::matchAndRewrite(triton::LoadOp op, OpAdaptor adaptor,
return rewriter.notifyMatchFailure(
op, "LoadOp expects a memref, not a memref of pointers");
}
if (!op->hasAttr(ConverterUtils::GeneratedByMakeTensorPtrTAG)) {
if (!hasZeroStrideMTP &&
!op->hasAttr(ConverterUtils::GeneratedByMakeTensorPtrTAG)) {
auto memrefOp = dyn_cast<memref::ReinterpretCastOp>(ptr.getDefiningOp());
auto ret = mlir::ConverterUtils::getLastStrideOfReinterpretCastOp(memrefOp);
if (ret.has_value())
lastStride = *ret;
}
bool mayImplicitTransposeWithLastAxis =
(existDotFlag) &&
(!hasZeroStrideMTP) && (existDotFlag) &&
(!op->hasAttr(ConverterUtils::GeneratedByMakeTensorPtrTAG)) &&
(lastStride != 1 &&
mlir::ConverterUtils::isaPermutedMemRefType(memRefType));
Expand Down Expand Up @@ -464,9 +521,12 @@ LoadConverter::matchAndRewrite(triton::LoadOp op, OpAdaptor adaptor,
// boundary check
auto boundaryCheck = op.getBoundaryCheck();
if (!boundaryCheck.empty()) {
auto makeTensorPtrOp = op.getPtr().getDefiningOp<triton::MakeTensorPtrOp>();
auto boundarySizes = mlir::ConverterUtils::getBoundarySizes(
boundaryCheck, /*remapped*/ ptr, loc, rewriter);
auto boundarySizes =
hasZeroStrideMTP
? getBoundarySizesFromMakeTensorPtr(makeTensorPtrOp, boundaryCheck,
memRefShape, loc, rewriter)
: mlir::ConverterUtils::getBoundarySizes(
boundaryCheck, /*remapped*/ ptr, loc, rewriter);
// handle the padding
auto padding = op.getPadding();
SmallVector<OpFoldResult> srcOffsets(boundarySizes.size(),
Expand Down
128 changes: 128 additions & 0 deletions ...d_party/ascend/unittest/Conversion/General/TritonToLinalg/zero_stride_block_ptr_load.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,128 @@
// RUN: triton-opt "--triton-to-linalg=global-kernel=false named-ops=True" --split-input-file %s | FileCheck %s

// A direct MTP load with zero strides must derive its boundary from the MTP
// logical shape/offset. The converted physical offset is always zero and
// cannot be divided by a zero physical stride to recover that information.
// CHECK-LABEL: func.func @zero_stride_dynamic
// CHECK: memref.reinterpret_cast {{.*}}strides: [0, 0]
// CHECK: arith.subi
// CHECK: arith.maxsi
// CHECK: arith.minsi
// CHECK: memref.copy
module attributes {hacc.target = #hacc.target<"Ascend910B2">} {
tt.func public @zero_stride_dynamic(
%src: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%dst: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%shape_m: i32, %shape_n: i32, %offset_m: i32, %offset_n: i32) {
%c0_i32 = arith.constant 0 : i32
%c0_i64 = arith.constant 0 : i64
%c1_i64 = arith.constant 1 : i64
%c4_i64 = arith.constant 4 : i64
%shape_m_i64 = arith.extsi %shape_m : i32 to i64
%shape_n_i64 = arith.extsi %shape_n : i32 to i64
%src_block = tt.make_tensor_ptr %src, [%shape_m_i64, %shape_n_i64],
[%c0_i64, %c0_i64], [%offset_m, %offset_n]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
%value = tt.load %src_block {boundaryCheck = array<i32: 0, 1>, padding = 1 : i32}
: !tt.ptr<tensor<4x4xf32>>
%dst_block = tt.make_tensor_ptr %dst, [%c4_i64, %c4_i64],
[%c4_i64, %c1_i64], [%c0_i32, %c0_i32]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
tt.store %dst_block, %value : !tt.ptr<tensor<4x4xf32>>
tt.return
}
}

// -----

// CHECK-LABEL: func.func @mixed_zero_stride_dynamic
// CHECK: memref.reinterpret_cast {{.*}}strides: [0, 1]
// CHECK: arith.subi
// CHECK: arith.maxsi
// CHECK: arith.minsi
// CHECK: memref.copy
module attributes {hacc.target = #hacc.target<"Ascend910B2">} {
tt.func public @mixed_zero_stride_dynamic(
%src: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%dst: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%shape_m: i32, %shape_n: i32, %offset_m: i32, %offset_n: i32) {
%c0_i32 = arith.constant 0 : i32
%c0_i64 = arith.constant 0 : i64
%c1_i64 = arith.constant 1 : i64
%c4_i64 = arith.constant 4 : i64
%shape_m_i64 = arith.extsi %shape_m : i32 to i64
%shape_n_i64 = arith.extsi %shape_n : i32 to i64
%src_block = tt.make_tensor_ptr %src, [%shape_m_i64, %shape_n_i64],
[%c0_i64, %c1_i64], [%offset_m, %offset_n]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
%value = tt.load %src_block {boundaryCheck = array<i32: 0, 1>, padding = 1 : i32}
: !tt.ptr<tensor<4x4xf32>>
%dst_block = tt.make_tensor_ptr %dst, [%c4_i64, %c4_i64],
[%c4_i64, %c1_i64], [%c0_i32, %c0_i32]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
tt.store %dst_block, %value : !tt.ptr<tensor<4x4xf32>>
tt.return
}
}

// -----

// CHECK-LABEL: func.func @zero_stride_negative_offset
// CHECK: memref.reinterpret_cast {{.*}}strides: [0, 0]
// CHECK: memref.subview {{.*}}[2, 0]
// CHECK: memref.copy
module attributes {hacc.target = #hacc.target<"Ascend910B2">} {
tt.func public @zero_stride_negative_offset(
%src: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%dst: !tt.ptr<f32> {tt.divisibility = 16 : i32}) {
%cneg2_i32 = arith.constant -2 : i32
%c0_i32 = arith.constant 0 : i32
%c0_i64 = arith.constant 0 : i64
%c1_i64 = arith.constant 1 : i64
%c4_i64 = arith.constant 4 : i64
%c5_i64 = arith.constant 5 : i64
%c6_i64 = arith.constant 6 : i64
%src_block = tt.make_tensor_ptr %src, [%c6_i64, %c5_i64],
[%c0_i64, %c0_i64], [%cneg2_i32, %c0_i32]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
%value = tt.load %src_block {boundaryCheck = array<i32: 0, 1>, padding = 1 : i32}
: !tt.ptr<tensor<4x4xf32>>
%dst_block = tt.make_tensor_ptr %dst, [%c4_i64, %c4_i64],
[%c4_i64, %c1_i64], [%c0_i32, %c0_i32]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
tt.store %dst_block, %value : !tt.ptr<tensor<4x4xf32>>
tt.return
}
}

// -----

// A direct MTP with statically nonzero strides must keep the legacy physical
// offset reconstruction path.
// CHECK-LABEL: func.func @nonzero_stride_dynamic
// CHECK: arith.divsi
// CHECK: arith.remsi
// CHECK: memref.copy
module attributes {hacc.target = #hacc.target<"Ascend910B2">} {
tt.func public @nonzero_stride_dynamic(
%src: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%dst: !tt.ptr<f32> {tt.divisibility = 16 : i32},
%shape_m: i32, %shape_n: i32, %offset_m: i32, %offset_n: i32) {
%c0_i32 = arith.constant 0 : i32
%c1_i64 = arith.constant 1 : i64
%c4_i64 = arith.constant 4 : i64
%c8_i64 = arith.constant 8 : i64
%shape_m_i64 = arith.extsi %shape_m : i32 to i64
%shape_n_i64 = arith.extsi %shape_n : i32 to i64
%src_block = tt.make_tensor_ptr %src, [%shape_m_i64, %shape_n_i64],
[%c8_i64, %c1_i64], [%offset_m, %offset_n]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
%value = tt.load %src_block {boundaryCheck = array<i32: 0, 1>, padding = 1 : i32}
: !tt.ptr<tensor<4x4xf32>>
%dst_block = tt.make_tensor_ptr %dst, [%c4_i64, %c4_i64],
[%c4_i64, %c1_i64], [%c0_i32, %c0_i32]
{order = array<i32: 1, 0>} : <tensor<4x4xf32>>
tt.store %dst_block, %value : !tt.ptr<tensor<4x4xf32>>
tt.return
}
}
Loading