Skip to content

[MLIR] [AMX] Fix strides used by AMX lowering for tile loads and stores. #113476

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

Merged
merged 2 commits into from
Oct 30, 2024
Merged

[MLIR] [AMX] Fix strides used by AMX lowering for tile loads and stores. #113476

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0c0f415
Fix strides used by AMX lowering.
ienkovich Oct 23, 2024
41f66cd
Fix review comments.
ienkovich Oct 28, 2024
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
16 changes: 10 additions & 6 deletions mlir/lib/Dialect/AMX/Transforms/LegalizeForLLVMExport.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,21 +55,25 @@ Value getStride(ConversionPatternRewriter &rewriter,
const LLVMTypeConverter &typeConverter, MemRefType mType,
Value base, Location loc) {
assert(mType.getRank() >= 2);
int64_t last = mType.getRank() - 1;
int64_t preLast = mType.getRank() - 2;
Type llvmInt64Type = IntegerType::get(&typeConverter.getContext(), 64);
unsigned width = mType.getElementType().getIntOrFloatBitWidth();
assert(llvm::isPowerOf2_64(width) && width >= 8);
unsigned bytes = width >> 3;
if (mType.isDynamicDim(last)) {
// Dynamic size needs code to compute the stride at runtime.
int64_t offset;
SmallVector<int64_t, 4> strides;
getStridesAndOffset(mType, strides, offset);
if (strides[preLast] == ShapedType::kDynamic) {
// Dynamic stride needs code to compute the stride at runtime.
MemRefDescriptor memrefDescriptor(base);
auto attr = rewriter.getI64IntegerAttr(bytes);
Value scale = rewriter.create<LLVM::ConstantOp>(loc, llvmInt64Type, attr);
return rewriter.create<LLVM::MulOp>(
loc, llvmInt64Type, scale, memrefDescriptor.size(rewriter, loc, last));
loc, llvmInt64Type, scale,
memrefDescriptor.stride(rewriter, loc, preLast));
}
// Use direct constant for static size.
auto attr = rewriter.getI64IntegerAttr(mType.getDimSize(last) * bytes);
// Use direct constant for static stride.
auto attr = rewriter.getI64IntegerAttr(strides[preLast] * bytes);
return rewriter.create<LLVM::ConstantOp>(loc, llvmInt64Type, attr);
}

Expand Down
28 changes: 28 additions & 0 deletions mlir/test/Dialect/AMX/legalize-for-llvm.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,3 +43,31 @@ func.func @mulf(%arg0: memref<?x?xbf16>, %arg1: memref<?x?xf32>) {
amx.tile_store %arg1[%0, %0], %4 : memref<?x?xf32>, vector<16x16xf32>
return
}

// CHECK-LABEL: strides(
// CHECK: %[[CST_64_1:.+]] = llvm.mlir.constant(64 : i64) : i64
// CHECK: "amx.tileloadd64"(%{{.+}}, %{{.+}}, %{{.+}}, %[[CST_64_1]]
// CHECK: %[[CST_128_1:.+]] = llvm.mlir.constant(128 : i64) : i64
// CHECK: "amx.tileloadd64"(%{{.+}}, %{{.+}}, %{{.+}}, %[[CST_128_1]]
// CHECK: llvm.mlir.constant(2 : i64) : i64
// CHECK: llvm.extractvalue %{{.+}}[4, 0]
// CHECK: %[[STRIDE_1:.+]] = llvm.mul
// CHECK: "amx.tileloadd64"(%{{.+}}, %{{.+}}, %{{.+}}, %[[STRIDE_1]]
// CHECK: %[[CST_64_2:.+]] = llvm.mlir.constant(64 : i64) : i64
// CHECK: "amx.tilestored64"(%{{.+}}, %{{.+}}, %{{.+}}, %[[CST_64_2]]
// CHECK: %[[CST_128_2:.+]] = llvm.mlir.constant(128 : i64) : i64
// CHECK: "amx.tilestored64"(%{{.+}}, %{{.+}}, %{{.+}}, %[[CST_128_2]]
// CHECK: llvm.mlir.constant(2 : i64) : i64
// CHECK: llvm.extractvalue %{{.+}}[4, 0]
// CHECK: %[[STRIDE_2:.+]] = llvm.mul
// CHECK: "amx.tilestored64"(%{{.+}}, %{{.+}}, %{{.+}}, %[[STRIDE_2]]
func.func @strides(%arg0: memref<16x32xbf16>, %arg1: memref<16x32xbf16, strided<[64, 1]>>, %arg2: memref<16x32xbf16, strided<[?, 1]>>) {
%0 = arith.constant 0 : index
%1 = amx.tile_load %arg0[%0, %0] : memref<16x32xbf16> into vector<16x32xbf16>
%2 = amx.tile_load %arg1[%0, %0] : memref<16x32xbf16, strided<[64, 1]>> into vector<16x32xbf16>
%3 = amx.tile_load %arg2[%0, %0] : memref<16x32xbf16, strided<[?, 1]>> into vector<16x32xbf16>
amx.tile_store %arg0[%0, %0], %3 : memref<16x32xbf16>, vector<16x32xbf16>
amx.tile_store %arg1[%0, %0], %1 : memref<16x32xbf16, strided<[64, 1]>>, vector<16x32xbf16>
amx.tile_store %arg2[%0, %0], %2 : memref<16x32xbf16, strided<[?, 1]>>, vector<16x32xbf16>
return
}
Loading