repo_name
stringlengths
1
62
dataset
stringclasses
1 value
lang
stringclasses
11 values
pr_id
int64
1
20.1k
owner
stringlengths
2
34
reviewer
stringlengths
2
39
diff_hunk
stringlengths
15
262k
code_review_comment
stringlengths
1
99.6k
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -328,6 +352,16 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { rewriter.replaceOpWithNewOp<LLVM::ShuffleVectorOp>(op, dstType, shfl01, shfl23, attr); } break; + case 8: { + unsigned num = 8; + Value undef ...
Can we come up with a lowering that works for all number of inputs? I don't think this is scalable at all.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
As it is in theory possible, I'd return failure here instead of asserting out.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
This case should not be upstream as tensors cannot be directly converted to LLVM dialect.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
Isn't `dstAttrType` the same type as `vecType`?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
```suggestion auto poison = rewriter.create<LLVM::PoisonOp>(loc, vecTy); auto splat = rewriter.create<LLVM::InsertElementOp>( loc, vecTy, poison, adaptor.getSrc(), rewriter.create<LLVM::ConstantOp>(loc, rewriter.getI32Type(), 0)); auto convertedTy = typeConverter->convertType(resultType...
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
Will we always get `adaptor.getSrc().getType() == op.getType()`?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
Can we just add the rest of the cases here now?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
Is this reduction pattern correct? Wouldn't we end up having each subgroup reduce a slice, but not the whole tensor?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +373,166 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
I'm not sure this and `tt.expand_dims` patterns are correct.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -380,6 +547,29 @@ class AddPtrOpConversion : public ConvertTritonGPUOpToLLVMPattern<AddPtrOp> { } }; +class ArithDivFOpLowering + : public ConvertTritonGPUOpToLLVMPattern<mlir::arith::DivFOp> { + using ConvertTritonGPUOpToLLVMPattern< + mlir::arith::DivFOp>::ConvertTritonGPUOpToLLVMPattern; + Logical...
Why do we need this?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -28,8 +28,13 @@ TritonIntelGPUToLLVMTypeConverter::TritonIntelGPUToLLVMTypeConverter( // tensor type is flattened and divided by 16 (subgroupSize). addConversion([&](mlir::RankedTensorType type) -> mlir::Type { - return mlir::VectorType::get(type.getNumElements() / 16, - ...
Dot operand encoding should go here too?
intel-xpu-backend-for-triton
github_2023
python
1,574
intel
etiotto
@@ -87,6 +87,45 @@ def make_ttgir(mod, metadata, opt): pm.run(mod) return mod + @staticmethod + def make_llir(mod, metadata, opt):
Can we use the default pass pipeline "make_llir" code ? We are producing TritonGPU IR and the existing "make_llir" should work fine. What is the reason it does not ?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +370,169 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
I'm surprised we need `VectorType` here as upstream already supports this.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -58,6 +60,44 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + + Lo...
Should we use threads per warp instead?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -36,6 +36,8 @@ #include "mlir/Transforms/DialectConversion.h" #include "mlir/Transforms/GreedyPatternRewriteDriver.h" #include "llvm/Support/FormatVariadic.h" +#include <cstdint> +#include <optional>
Are they really needed?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,75 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +namespace { +// FIXME: remove this when upstream ConstantOpLowering has such lowering +class ArithConstantOpLowering + : public ConvertTriton...
```suggestion // arith.constant should only have vector or tensor types. ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,75 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +namespace { +// FIXME: remove this when upstream ConstantOpLowering has such lowering +class ArithConstantOpLowering + : public ConvertTriton...
please remove commented code.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,75 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +namespace { +// FIXME: remove this when upstream ConstantOpLowering has such lowering
Should we attempt to upstream ourselves?
intel-xpu-backend-for-triton
github_2023
python
1,574
intel
whitneywhtsang
@@ -222,7 +222,8 @@ def make_ttgir(mod, metadata, opt, properties): return mod @staticmethod - def make_llir(src, metadata, options): + def make_llir(src, metadata, options, properties):
Why passing `properties` if it is not used?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -92,8 +92,13 @@ struct ConvertTritonGPUToLLVM int numCTAs = triton::gpu::TritonGPUDialect::getNumCTAs(mod); int threadsPerWarp = triton::gpu::TritonGPUDialect::getThreadsPerWarp(mod); - // Allocate shared memory and set barrier - if (!pipelineManager.skipSharedMemoryAllocation()) { + // FIXME: s...
should we change compiler.py to `metadata["shared"] = src.get_int_attr("triton_gpu.shared")` to check if `src.get_int_attr("triton_gpu.shared")` is none, if so `metadata["shared"] = 0` instead?
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
whitneywhtsang
@@ -1,5 +1,4 @@ // RUN: triton-opt %s -split-input-file --convert-triton-intel-gpu-to-llvm | FileCheck %s --check-prefixes=CHECK,CHECK-SCALAR -// RUN: env TRITON_INTEL_ADVANCED_PATH=1 triton-opt %s -split-input-file --convert-triton-intel-gpu-to-llvm | FileCheck %s --check-prefixes=CHECK,CHECK-VECTOR
Why is this removed?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -0,0 +1,72 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +namespace { +// FIXME: remove this when upstream ConstantOpLowering has such lowering +class ArithConstantOpLowering + : public ConvertTriton...
I'd say we don't need this as IGC should be handling it.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,77 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works for Advanced Path. +/// This is a temporary s...
```suggestion /// proper lowering strategy for these ops. ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,77 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works for Advanced Path. +/// This is a temporary s...
can we rely on upstream lowering for vector type?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,85 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works for the advanced path. +/// This is a tempora...
Can we have a test case to prevent regression of this lowering? There are no failures from removing this pattern.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -170,16 +174,19 @@ class LoadStorePrefetchOpConversion return truncatedShape; }; - Value surfaceW = calculateSurface(ptrOp.getShape()[1], true); - Value surfaceH = calculateSurface(ptrOp.getShape()[0], false); - Value surfaceP = calculateSurface(ptrOp.getStrides()[0], true); + Value surface...
[optional] ```suggestion Value surfaceW = calculateSurface( ptrOp.getShape()[!transpose], true); Value surfaceH = calculateSurface( ptrOp.getShape()[transpose], false); Value surfaceP = calculateSurface( ptrOp.getStrides()[transpose], true); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -189,6 +196,17 @@ class LoadStorePrefetchOpConversion getVectorType(cast<RankedTensorType>(op.getResult().getType()), isDword ? i32_ty : i16_ty); bool vnni = (idx == 1) && dataSize < 32; + + // FIXME: only support fp16/bf16 for now, add more support like tf32, fp8
Can we return failure for unsupported types?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -198,7 +216,12 @@ class LoadStorePrefetchOpConversion return failure(); } rewriter.replaceOp(op, bitcast(load, resType)); - } else if constexpr (std::is_same_v<OpType, PrefetchOp>) { + } else if constexpr (std::is_same_v<OpType, ttgi::PrefetchOp>) { + if (ptrOp.getOrder()[0] == 0) {...
remove?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -328,9 +352,17 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { rewriter.replaceOpWithNewOp<LLVM::ShuffleVectorOp>(op, dstType, shfl01, shfl23, attr); } break; - default: - llvm_unreachable("add more support for...
```suggestion for (auto i = 0; i < num; ++i) { ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -339,33 +370,142 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
`triton_intel_gpu.min_sg_size` means the minimum subgroup size of the platform, `triton_gpu.threads-per-warp` means the actual subgroup size used. We should use `triton_gpu.threads-per-warp` in this case.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -339,33 +370,142 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
```suggestion llvm_unreachable("Unhandled reduction kind"); ```
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
victor-eds
@@ -145,3 +145,104 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
```suggestion ``` Not needed
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
victor-eds
@@ -145,3 +145,104 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
```suggestion ``` Not needed
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
victor-eds
@@ -145,3 +145,104 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
```suggestion // COM: Checks the correct lowering of triton ops, including broadcast, splat, expand_dims, addptr. ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -0,0 +1,85 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works for the advanced path. +/// This is a tempora...
```suggestion auto dstElementsAttr = dyn_cast<DenseElementsAttr>(op.getValue()); if (!dstElementsAttr || !dstElementsAttr.isSplat()) return failure(); ``` This will only work if the constant is a splat, e.g., it won't work for: ``` %0 = arith.constant dense[0, 1, 2, 3, ...] : tensor<128xi32> ```
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
victor-eds
@@ -145,3 +145,104 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
```suggestion ```
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
victor-eds
@@ -145,3 +145,104 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
```suggestion ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -122,8 +126,8 @@ class LoadStorePrefetchOpConversion ConversionPatternRewriter &rewriter) const override { auto ptrType = cast<PointerType>(op.getPtr().getType()); auto tensorType = cast<RankedTensorType>(ptrType.getPointeeType()); - assert(tensorType.getRank() <= 2 && - "on...
Should we return `failure()` instead of asserting here?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -189,6 +194,19 @@ class LoadStorePrefetchOpConversion getVectorType(cast<RankedTensorType>(op.getResult().getType()), isDword ? i32_ty : i16_ty); bool vnni = (idx == 1) && dataSize < 32; + + // FIXME: only support fp16/bf16 transpose for now, add more support like + ...
Can we return `failure()` here instead?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
victor-eds
@@ -339,33 +368,141 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { /// %extract = ttgi.extract %a[0] : tensor<8xf16> -> tensor<4xf16> /// is converted to /// %extract = llvm.shufflevector %a, %a : [0, 1, 2, 3] : vector<4xf16> -class ExtractOpConversion : public ConvertTritonGPUOpToLLVMP...
Can we return failure?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -170,16 +175,16 @@ class LoadStorePrefetchOpConversion return truncatedShape; }; - Value surfaceW = calculateSurface(ptrOp.getShape()[1], true); - Value surfaceH = calculateSurface(ptrOp.getShape()[0], false); - Value surfaceP = calculateSurface(ptrOp.getStrides()[0], true); + Value surface...
is it intentional to change from getStrides to getShape?
intel-xpu-backend-for-triton
github_2023
python
1,574
intel
FMarno
@@ -258,7 +258,7 @@ def make_llir(src, metadata, options): intel.post_process_llir(llvm_mod) # Get some metadata - metadata["shared"] = src.get_int_attr("triton_gpu.shared") + metadata["shared"] = src.get_int_attr("triton_gpu.shared") if src.get_int_attr("triton_gpu.shared") else 0
```suggestion metadata["shared"] = src.get_int_attr("triton_gpu.shared") or 0 ``` I'm not a python expert but I think this works. https://realpython.com/python-or-operator/#using-or-with-common-objects
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -0,0 +1,85 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works for the advanced path. +/// This is a tempora...
This check can be removed if the code at line 24 is modified to: ``` auto srcType = dyn_cast<RankedTensorType>(op.getType()); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -7,10 +7,12 @@ #include "triton/Analysis/Utility.h" #include "triton/Dialect/Triton/IR/Dialect.h" #include "triton/Dialect/Triton/IR/Types.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" +#include "llvm/ADT/ArrayRef.h" using namespace mlir; -using namespace mlir::triton; -using namespace mlir::tri...
why this change ? I think using the namespace as it was before was fine. We can revert this change and reduce the footprint of this PR.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -28,8 +28,14 @@ TritonIntelGPUToLLVMTypeConverter::TritonIntelGPUToLLVMTypeConverter( // tensor type is flattened and divided by 16 (subgroupSize). addConversion([&](mlir::RankedTensorType type) -> mlir::Type { - return mlir::VectorType::get(type.getNumElements() / 16, - ...
why is this change required ? In particular if `num == 1` can't we return a vector (of one element) to avoid diverging from other cases ?
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -385,14 +521,20 @@ class AddPtrOpConversion : public ConvertTritonGPUOpToLLVMPattern<AddPtrOp> { void mlir::triton::intel::populateTritonOpsToLLVMPatterns( TritonIntelGPUToLLVMTypeConverter &typeConverter, RewritePatternSet &patterns, PatternBenefit benefit) { - patterns.add<MakeTensorPtrOpConversion>(ty...
kudos for using alphabetical order here !
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -0,0 +1,85 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works for the advanced path. +/// This is a tempora...
The cast is unsafe, should modify the check at line 33 to check for this.,
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -0,0 +1,85 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works for the advanced path. +/// This is a tempora...
Can use `replaceOpWithNewOp` here.
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
etiotto
@@ -145,3 +145,100 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
[nit]: The indentation of the CHECKS is not consistent with the style used by the test cases already in this file.
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
etiotto
@@ -145,3 +145,100 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
These should be CHECK-DAG
intel-xpu-backend-for-triton
github_2023
others
1,574
intel
etiotto
@@ -145,3 +145,100 @@ module attributes {"triton_intel_gpu.support_sg_2d_block", "triton_intel_gpu.sup tt.return } } + +// ----- + +// COM: Checks the correct lowering of sub-group reductions. + +#warp = #triton_intel_gpu.warp<{sizePerThread = [16, 64], threadsPerWarp = [1, 1], order = [1, 0]}> +module attribu...
Existing tests use a different style to match, example: ``` // CHECK: [[VAL_1:%.*]] = llvm.mlir.constant(0 : i32) : i32 ``` Let's be consistent.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,79 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/Dialect/LLVMIR/LLVMDialect.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works...
```suggestion if (!dyn_cast_or_null<VectorType>(dstType)) ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
whitneywhtsang
@@ -0,0 +1,79 @@ +#include "PatternTritonGPUOpToLLVM.h" + +#include "mlir/Conversion/LLVMCommon/TypeConverter.h" +#include "mlir/Dialect/LLVMIR/LLVMDialect.h" +#include "mlir/IR/BuiltinTypes.h" + +using namespace mlir; + +/// Custom lowering for converting arith ops to LLVMIR dialect. +/// These ops lowering ONLY works...
```suggestion if (!dyn_cast_or_null<VectorType>(dstType)) ```
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -328,9 +346,17 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { rewriter.replaceOpWithNewOp<LLVM::ShuffleVectorOp>(op, dstType, shfl01, shfl23, attr); } break; - default: - llvm_unreachable("add more support for...
Drop the semicolon here.
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -346,16 +372,120 @@ class ExtractOpConversion : public ConvertTritonGPUOpToLLVMPattern<ExtractOp> { LogicalResult matchAndRewrite(ExtractOp op, OpAdaptor adaptor, ConversionPatternRewriter &rewriter) const override { + Location loc = op.getLoc(); Value base = adaptor.getBase(); - ...
Use the `extract_element` macro
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -189,6 +192,19 @@ class LoadStorePrefetchOpConversion getVectorType(cast<RankedTensorType>(op.getResult().getType()), isDword ? i32_ty : i16_ty); bool vnni = (idx == 1) && dataSize < 32; + + // FIXME: only support fp16/bf16 transpose for now, add more support like + ...
use the `lshr` macro
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -328,9 +346,17 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { rewriter.replaceOpWithNewOp<LLVM::ShuffleVectorOp>(op, dstType, shfl01, shfl23, attr); } break; - default: - llvm_unreachable("add more support for...
use the `insert_element` macro
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -328,9 +346,17 @@ class GlueOpConversion : public ConvertTritonGPUOpToLLVMPattern<GlueOp> { rewriter.replaceOpWithNewOp<LLVM::ShuffleVectorOp>(op, dstType, shfl01, shfl23, attr); } break; - default: - llvm_unreachable("add more support for...
auto -> int
intel-xpu-backend-for-triton
github_2023
cpp
1,574
intel
etiotto
@@ -346,16 +372,120 @@ class ExtractOpConversion : public ConvertTritonGPUOpToLLVMPattern<ExtractOp> { LogicalResult matchAndRewrite(ExtractOp op, OpAdaptor adaptor, ConversionPatternRewriter &rewriter) const override { + Location loc = op.getLoc(); Value base = adaptor.getBase(); - ...
Use: ``` int subgroupSize = triton::gpu::TritonGPUDialect::getThreadsPerWarp(mod);
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -0,0 +1,18 @@ +#ifndef TRITON_INTEL_ANALYSIS_UTILITY_H +#define TRITON_INTEL_ANALYSIS_UTILITY_H + +#include "triton/Analysis/Utility.h" + +namespace mlir { +namespace triton { +namespace gpu { +namespace intel {
[Suggestion]: use `namespace mlir::triton::gpu::intel`
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -0,0 +1,32 @@ +#include "triton/Analysis/Utility.h" +#include "intel/include/Analysis/Utility.h" +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" + +namespace mlir { +namespace triton { +namespace gpu { +namespace intel {
use `namespace mlir::triton::gpu::intel`
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -0,0 +1,18 @@ +#ifndef TRITON_INTEL_ANALYSIS_UTILITY_H +#define TRITON_INTEL_ANALYSIS_UTILITY_H + +#include "triton/Analysis/Utility.h" + +namespace mlir { +namespace triton { +namespace gpu { +namespace intel { + +bool isDpasToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy);
Document what srcTy and dstTy are in this context (srcTy is the output type of a tt.dot operation and dstTy is the input type of one of the operand of another tt.dot operation)
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -0,0 +1,32 @@ +#include "triton/Analysis/Utility.h" +#include "intel/include/Analysis/Utility.h" +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" + +namespace mlir { +namespace triton { +namespace gpu { +namespace intel { + +bool isDpasToDotShortcut(RankedTensorType srcTy, RankedTensorType dstTy) { + a...
Can we avoid hard coding the magic numbers (16,8,2) ?
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -455,6 +459,122 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
[nit]: use the same order you used in the then branch
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -455,6 +459,122 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
auto -> Location
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -455,6 +459,122 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
auto -> Type
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -455,6 +459,122 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
Early exit: `if (!intel:: ....) return failure();`
intel-xpu-backend-for-triton
github_2023
others
1,729
intel
etiotto
@@ -0,0 +1,1223 @@ +// RUN: triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s + +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, ...
mma -> dpas
intel-xpu-backend-for-triton
github_2023
others
1,729
intel
etiotto
@@ -0,0 +1,1223 @@ +// RUN: triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s + +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, ...
remove attributes noinline
intel-xpu-backend-for-triton
github_2023
others
1,729
intel
etiotto
@@ -0,0 +1,1223 @@ +// RUN: triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s + +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, ...
hmmm. this is very verbose. Can we use CHECK-COUNT here ?
intel-xpu-backend-for-triton
github_2023
others
1,729
intel
etiotto
@@ -0,0 +1,1223 @@ +// RUN: triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s + +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, ...
use CHECK-COUNT ?
intel-xpu-backend-for-triton
github_2023
others
1,729
intel
etiotto
@@ -0,0 +1,1223 @@ +// RUN: triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s + +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, ...
mma -> dpas
intel-xpu-backend-for-triton
github_2023
others
1,729
intel
etiotto
@@ -0,0 +1,1223 @@ +// RUN: triton-opt %s -split-input-file --intel-allocate-shared-memory --convert-triton-intel-gpu-to-llvm | FileCheck %s + +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [32, 1], repCluster = [1, 2], A = [8, ...
use CHECK-COUNT ?
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -0,0 +1,27 @@ +#include "triton/Analysis/Utility.h" +#include "intel/include/Analysis/Utility.h" +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" + +namespace mlir::triton::gpu::intel { + +bool isDpasToDotShortcut(RankedTensorType dpasTy, RankedTensorType dotTy) { + auto dpasLayout = dyn_cast<DpasEncod...
no auto pls
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -457,6 +461,118 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
size_t -> unsigned. auto is ok here because the RHS type is clear already
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -457,6 +461,118 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
no auto
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -457,6 +461,118 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
unsigned ?
intel-xpu-backend-for-triton
github_2023
cpp
1,729
intel
etiotto
@@ -457,6 +461,118 @@ struct ConvertLayoutOpConversion return success(); } + using ValueTable = std::map<std::pair<unsigned, unsigned>, Value>; + + ValueTable getValuesFromDpasLayoutStruct(Location loc, + ConversionPatternRewriter &rewriter, + ...
Can we use SmallVector ?
intel-xpu-backend-for-triton
github_2023
cpp
1,737
intel
Dewei-Wang-sh
@@ -412,11 +413,17 @@ class ScfPattern : public OpRewritePattern<scf::ForOp> { if (llvm::any_of(arg.getUsers(), [](Operation *user) { return !isa<ttgi::ExtractOp>(user); - })) { + })) return false; - } } + // Bail out if any user of the loop result is not ...
not the way we want. we can have loop results that result0's user is extract, result1's user is not extract, and still this should work.
intel-xpu-backend-for-triton
github_2023
cpp
1,684
intel
whitneywhtsang
@@ -0,0 +1,25 @@ +// Conversions from TritonIntelGPU DpasEncodingAttr to LinearLayout. + +#ifndef TRITON_DIALECT_TRITONINTELGPU_IR_LINEARLAYOUTCONVERSIONS_H +#define TRITON_DIALECT_TRITONINTELGPU_IR_LINEARLAYOUTCONVERSIONS_H + +#include <optional> + +#include "intel/include/Dialect/TritonIntelGPU/IR/Attributes.h" +#inc...
In the comment, it is written that default to -1, but here it is default to 2.
intel-xpu-backend-for-triton
github_2023
others
1,684
intel
whitneywhtsang
@@ -1,11 +1,12 @@ add_triton_library(TritonIntelGPUIR Dialect.cpp Ops.cpp - + LinearLayoutConversions.cpp DEPENDS TritonIntelGPUTableGen TritonIntelGPUAttrDefsIncGen LINK_LIBS PUBLIC MLIRSideEffectInterfaces +
Remove.
intel-xpu-backend-for-triton
github_2023
cpp
1,684
intel
whitneywhtsang
@@ -747,19 +754,32 @@ loadSharedToDistributed(Value dst, Value src, SharedMemoryObject &shrMemObj, auto srcTy = cast<MemDescType>(src.getType()); SmallVector<Value> ret; - if (emitTransferBetweenRegistersAndShared( - dstTy, srcTy, elemTy, /*maxVecElems=*/std::nullopt, - shrMemObj.getBase(), s...
```suggestion } bool success = emitTransferBetweenRegistersAndShared( dstTy, srcTy, elemTy, /*maxVecElems=*/std::nullopt, shrMemObj.getBase(), shrMemObj.getStrides(), loc, rewriter, target, [&](VectorType vecTy, Value vecAddr) { auto vecVal...
intel-xpu-backend-for-triton
github_2023
cpp
1,684
intel
chengjunlu
@@ -140,3 +142,142 @@ Value llPrintf(ConversionPatternRewriter &rewriter, StringRef msg, } } // namespace mlir::LLVM::intel + +namespace mlir::triton::intel { +bool emitTransferBetweenDPASAndShared( + RankedTensorType registerTy, MemDescType sharedTy, Type elemLlvmTy, + std::optional<int32_t> maxVecElems, Val...
This branch is taken only when the layout is the DPAS layout. It seems this condition is always false.
intel-xpu-backend-for-triton
github_2023
cpp
1,684
intel
chengjunlu
@@ -140,3 +142,142 @@ Value llPrintf(ConversionPatternRewriter &rewriter, StringRef msg, } } // namespace mlir::LLVM::intel + +namespace mlir::triton::intel { +bool emitTransferBetweenDPASAndShared( + RankedTensorType registerTy, MemDescType sharedTy, Type elemLlvmTy, + std::optional<int32_t> maxVecElems, Val...
The tensor in SLM should always have the SharedLayout. It seems this condition is always false.
intel-xpu-backend-for-triton
github_2023
cpp
1,684
intel
chengjunlu
@@ -140,3 +142,142 @@ Value llPrintf(ConversionPatternRewriter &rewriter, StringRef msg, } } // namespace mlir::LLVM::intel + +namespace mlir::triton::intel { +bool emitTransferBetweenDPASAndShared(
The LinearLayout allows us to use the common lowering code in TritonGPUToLLVM. Is there any limitation in the function for us to re-use. https://github.com/triton-lang/triton/blob/9c7f9f8e3f6c769d0c32d0c51265c04ba8b164ca/lib/Conversion/TritonGPUToLLVM/Utility.cpp#L267C6-L267C43
intel-xpu-backend-for-triton
github_2023
cpp
1,720
intel
etiotto
@@ -49,7 +49,10 @@ struct PrintOpConversion LLVM::intel::llPrintf(rewriter, formatStr, {pid[0], pid[1], pid[2], prefixStr}); } else { + assert(op.getNumOperands() == op.getIsSigned().size());
add an assert message pls.
intel-xpu-backend-for-triton
github_2023
others
1,694
intel
whitneywhtsang
@@ -4,6 +4,9 @@ # Example usage: # ./triton-opt.sh --triton-opt=<path to triton-opt> 01-vector-add-xpu.ttir --target=llir -mlir-print-ir-after-all +echo "Warning: This script is deprecated, if you think it could be still helpful and needs to be updated, please create an Issue." >&2
```suggestion echo "Warning: This script is deprecated, if you think it could be still helpful and needs to be updated, please create an GitHub issue." >&2 ```
intel-xpu-backend-for-triton
github_2023
cpp
1,709
intel
whitneywhtsang
@@ -10,34 +10,28 @@ using namespace mlir::triton::gpu; // blocked -> shared. // Swizzling in shared memory to avoid bank conflict. Normally used for // A/B operands of dots. -void lowerDistributedToShared(LocalAllocOp op, LocalAllocOpAdaptor adaptor, +void lowerDistributedToShared(Location loc, Value src, Value dst,...
Can you please also change `mlir::triton::intel::storeDistributedToShared` signature to match the common `storeDistributedToShared`?
intel-xpu-backend-for-triton
github_2023
cpp
1,709
intel
whitneywhtsang
@@ -10,34 +10,28 @@ using namespace mlir::triton::gpu; // blocked -> shared. // Swizzling in shared memory to avoid bank conflict. Normally used for // A/B operands of dots. -void lowerDistributedToShared(LocalAllocOp op, LocalAllocOpAdaptor adaptor, +void lowerDistributedToShared(Location loc, Value src, Value dst,...
Can we change back to below in this PR? ``` Value smemBase = LLVM::intel::getSharedMemoryBase(loc, rewriter, op.getOperation()); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,713
intel
FMarno
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + + Lo...
I think you should add a check for the uniform attribute, even if just to report that it isn't supported.
intel-xpu-backend-for-triton
github_2023
others
1,713
intel
FMarno
@@ -0,0 +1,18 @@ +// RUN: triton-opt -convert-gpu-to-tritongen %s | FileCheck %s + +module attributes { + "triton_gpu.threads-per-warp" = 16 : i32, + spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Kernel, Addresses, GroupNonUniformShuffle, Int64], []>, #spirv.resource_limits<subgroup_size = 16>>
are both of these needed
intel-xpu-backend-for-triton
github_2023
cpp
1,713
intel
FMarno
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + + Lo...
Is there not a helper function to get the threads per warp/subgroup size?
intel-xpu-backend-for-triton
github_2023
cpp
1,713
intel
FMarno
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + + Lo...
you could use `rewriter.replaceOpWithNewOp` here
intel-xpu-backend-for-triton
github_2023
cpp
1,713
intel
victor-eds
@@ -58,6 +58,39 @@ namespace { /// Import the GPU Ops to TritonGEN Patterns. #include "GPUToTritonGEN.cpp.inc" +struct GPUSubgroupReduceOpLowering + : public ConvertOpToLLVMPattern<mlir::gpu::SubgroupReduceOp> { + using ConvertOpToLLVMPattern< + mlir::gpu::SubgroupReduceOp>::ConvertOpToLLVMPattern; + + Lo...
Can we this as a function and support remaining types? Shouldn't be much, right?
intel-xpu-backend-for-triton
github_2023
cpp
1,687
intel
whitneywhtsang
@@ -413,8 +413,12 @@ void init_triton_llvm(py::module &&m) { // FIXME: Temporary workaround to avoid __devicelib_assert_fail // optimization with InternalLinkage, which causes // test_subprocess.py::test_assert to fail. - if (fn.getName().str() == "__devicelib_assert_fail") - ...
Why is it not enough to add optnone? ```suggestion if (fn.getName().str() == "__devicelib_assert_fail") fn.addFnAttr(llvm::Attribute::OptimizeNone); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,706
intel
whitneywhtsang
@@ -105,12 +106,18 @@ static Value getSharedMemoryBase(Location loc, auto ptrTy = LLVM::LLVMPointerType::get( rewriter.getContext(), TritonGEN::TritonGENMemorySpace::kWorkgroup); FunctionOpInterface func = op->getParentOfType<FunctionOpInterface>(); - size_t offset = 0; - if (op->hasAttr("allocation.offs...
use llvm_unreachable instead.
intel-xpu-backend-for-triton
github_2023
python
1,695
intel
whitneywhtsang
@@ -401,9 +401,8 @@ def benchmark(B, M, N, K, provider): func = getattr(xetla_kernel, name) xetla_fn = lambda: func(a, b, c, acc, cnt) torch_fn = lambda: torch.matmul(a, b).to(torch.float32) - rtol = 1e-2 if a.dtype == torch.bfloat16 else 1e-3 # FIXME: XeTLA batch GEMM impleme...
```suggestion # benchmark_suit.assert_close(xetla_fn(), torch_fn(), atol=1e-4, rtol=1.0, err_msg="xetla to torch") ```
intel-xpu-backend-for-triton
github_2023
others
1,681
intel
whitneywhtsang
@@ -951,6 +1008,7 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : tt.func @matmul_fmadot(%ptr:!tt.ptr<f32> {tt.divisibility = 16 : i32},
Please help update memdesc with `#triton_gpu.shared_memory`.
intel-xpu-backend-for-triton
github_2023
others
1,681
intel
whitneywhtsang
@@ -1655,6 +1655,7 @@ module attributes {"triton_gpu.target" = "xpu:DEVICE_ARCH.PVC", "triton_gpu.num- #blocked0 = #triton_gpu.blocked<{sizePerThread = [2], threadsPerWarp = [32], warpsPerCTA = [4], order = [0], CTAsPerCGA = [1], CTASplitNum = [1], CTAOrder = [0]}> module attributes {"triton_gpu.num-ctas" = 1 : i32, ...
Can you open an issue to investigate?
intel-xpu-backend-for-triton
github_2023
others
1,693
intel
pbchekin
@@ -20,19 +20,28 @@ This is the development repository of Intel® XPU Backend for Triton\*, a new [T # Install from source ``` -git clone https://github.com/intel/intel-xpu-backend-for-triton.git -b llvm-target -cd intel-xpu-backend-for-triton -scripts/compile-triton.sh +git clone https://github.com/triton-lang/trit...
The instructions are confusing: this builds an upstream Triton instead of Triton XPU from this repository. Not sure why we need this.