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
1,572
intel
victor-eds
@@ -2174,10 +2174,9 @@ struct IndexCastOpLowering if (targetBits == sourceBits) return {operands[0][0]}; if (targetBits < sourceBits) - return {rewriter.replaceOpWithNewOp<LLVM::TruncOp>(op, elemTy, - operands[0][0])}; - return { - ...
We can use `mlir::getValueOrCreateCastToIndexLike` in `mlir/Dialect/Arith/Utils/Utils.h` for this
intel-xpu-backend-for-triton
github_2023
cpp
1,572
intel
victor-eds
@@ -811,19 +811,10 @@ struct FuncCallLowering { auto argType = rewriter.getIntegerType(32); auto arg = LLVM::createConstantI32(op->getLoc(), rewriter, dim); - intel::AttributeList attrs = createFunctionAttributes( - {{llvm::Attribute::NoUnwind, std::nullopt}, - {llvm::Attribute::WillReturn...
I am not sure about removing the attributes. I'd say we have to do quite the opposite: make the GPU to LLVM-SPV pass add the attributes expressing function semantics. We're running some LLVM optimization passes after this, so adding the attributes may enable optimization opportunities.
intel-xpu-backend-for-triton
github_2023
others
1,572
intel
whitneywhtsang
@@ -83,7 +83,9 @@ module attributes {"triton_gpu.num-warps" = 8 : i32, "triton_gpu.threads-per-war // CHECK: %[[VAL_80:.*]] = llvm.insertvalue %[[CST_1]], %[[VAL_79]][5] : !llvm.struct<(i32, i32, i64, i64, i64, i64, ptr<1>)> // CHECK: %[[BLOCK_PTR:.*]] = llvm.insertvalue %[[base]], %[[VAL_...
How about changing triton_gen.subgroup.id to return indextype?
intel-xpu-backend-for-triton
github_2023
cpp
1,572
intel
whitneywhtsang
@@ -606,14 +606,10 @@ struct FuncCallLowering { {llvm::Attribute::WillReturn, std::nullopt}, {llvm::Attribute::Memory, llvm::MemoryEffects::none().toIntValue()}}, ctx); + LLVM::CallOp callOp = createDeviceFunctionCall(rewriter, funcName, retType,
Should we change `retType` from `i64` to `index` type?
intel-xpu-backend-for-triton
github_2023
others
1,922
intel
pbchekin
@@ -216,24 +224,24 @@ build_ipex() { if [ ! -d "$IPEX_PROJ" ]; then echo "**** Cloning $IPEX_PROJ ****" cd $BASE - if [ "$FAKE_IPEX" = true ]; then + if [ "$NO_OP_IPEX" = true ]; then mkdir intel-extension-for-pytorch cd intel-extension-for-pytorch cat > setup.py <<EOF from setu...
Per https://peps.python.org/pep-0440/#local-version-identifiers Python local version identifier can contain only: * ASCII letters ([a-zA-Z]) * ASCII digits ([0-9]) * periods (.) I would recommend replacing ```suggestion version = "2.4.0+noop" ```
intel-xpu-backend-for-triton
github_2023
others
1,922
intel
vlad-penkin
@@ -49,7 +49,7 @@ for arg in "$@"; do shift ;; --help) - echo "Example usage: ./compile-pytorch-ipex.sh [--pytorch | --upstream-pytorch | --ipex | --fake-ipex | --pinned | --source | --clean | --venv]" + echo "Example usage: ./compile-pytorch-ipex.sh [--pytorch | --upstream-pytorch | --ipex...
@anmyachev to continue your line of thought - do we need two separate options `--pytorch` and `--ipex`?
intel-xpu-backend-for-triton
github_2023
others
1,875
intel
etiotto
@@ -799,6 +799,21 @@ def MmaEncodingTrait : AttrInterface<"MmaEncodingTrait"> { ]; } +//===----------------------------------------------------------------------===//
Adding this change downstream will cause us to deviate from upstream code more. Please post a PR upstream to introduce this change.
intel-xpu-backend-for-triton
github_2023
others
1,875
intel
victor-eds
@@ -799,6 +799,21 @@ def MmaEncodingTrait : AttrInterface<"MmaEncodingTrait"> { ]; } +//===----------------------------------------------------------------------===// +// LinearLayout Conversion Interface +//===----------------------------------------------------------------------===// +def LinearLayoutTrait : At...
```suggestion (ins "ArrayRef<int64_t>":$shape, "std::optional<int32_t>":$elemBitWidth)> ``` NIT
intel-xpu-backend-for-triton
github_2023
others
1,875
intel
victor-eds
@@ -11,46 +11,32 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : // CHECK-DAG: %[[CST_4:.*]] = llvm.mlir.constant(4 : i32) : i32 // CHECK-DAG: %[[CST_8:.*]] = llvm.mlir.constant(8 : i32) : i32 // CHECK-DAG: %[[CST_16:.*]] = llvm.mlir.const...
NIT: I think we can remove `-DAG` to simplify the tests as we know the operation generation order here.
intel-xpu-backend-for-triton
github_2023
cpp
1,875
intel
victor-eds
@@ -1,6 +1,7 @@ #ifndef TRITON_DIALECT_TRITON_INTEL_GPU_IR_ATTRIBUTES_H #define TRITON_DIALECT_TRITON_INTEL_GPU_IR_ATTRIBUTES_H +#include "triton/Dialect/TritonGPU/IR/Attributes.h" #define GET_ATTRDEF_CLASSES #include "intel/include/Dialect/TritonIntelGPU/IR/TritonIntelGPUAttrDefs.h.inc"
```suggestion #include "triton/Dialect/TritonGPU/IR/Attributes.h" #define GET_ATTRDEF_CLASSES #include "intel/include/Dialect/TritonIntelGPU/IR/TritonIntelGPUAttrDefs.h.inc" ``` NIT
intel-xpu-backend-for-triton
github_2023
cpp
1,875
intel
victor-eds
@@ -415,6 +416,12 @@ void DpasEncodingAttr::print(AsmPrinter &printer) const { << "}>"; } +LinearLayout +DpasEncodingAttr::toLinearLayout(ArrayRef<int64_t> shape, + std::optional<int32_t> elemBitWidth) const { + return DPAStoLinearLayout(shape, *this, 2 /*opIdx*/); +}
```suggestion LinearLayout DpasEncodingAttr::toLinearLayout(ArrayRef<int64_t> shape, std::optional<int32_t> elemBitWidth) const { return DPAStoLinearLayout(shape, *this); } ``` NIT: Can we exploit default value? If not, I think `/*opIdx=*/2` is preferred (at least clang-forma...
intel-xpu-backend-for-triton
github_2023
cpp
1,875
intel
victor-eds
@@ -547,36 +558,26 @@ DPAStoLinearLayout(ArrayRef<int64_t> shape, Attribute layout, unsigned opIdx) { LinearLayout::identity1D(repCluster[1], kRegister, outDimNames[1]); tileLayout *= LinearLayout::identity1D(repCluster[0], kRegister, outDimNames[0]); + + // The identical layout is repeated am...
```suggestion ArrayRef<int64_t> numReps = dpas.getDPASRepetitions(shape, opIdx); ``` NIT
intel-xpu-backend-for-triton
github_2023
others
1,881
intel
whitneywhtsang
@@ -92,9 +97,28 @@ jobs: IGC_DisableLoopUnroll=1 \ SYCL_PROGRAM_COMPILE_OPTIONS=" -vc-codegen -vc-disable-indvars-opt -doubleGRF -Xfinalizer ' -printregusage -enableBCR -DPASTokenReduction ' " \ python gemm_benchmark.py --reports $REPORTS + + TAG=${{ inputs.tag || 'ci' }} + ...
```suggestion - name: Run Triton GEMM kernel benchmark - default path ```
intel-xpu-backend-for-triton
github_2023
others
1,881
intel
whitneywhtsang
@@ -92,9 +97,28 @@ jobs: IGC_DisableLoopUnroll=1 \ SYCL_PROGRAM_COMPILE_OPTIONS=" -vc-codegen -vc-disable-indvars-opt -doubleGRF -Xfinalizer ' -printregusage -enableBCR -DPASTokenReduction ' " \ python gemm_benchmark.py --reports $REPORTS + + TAG=${{ inputs.tag || 'ci' }} + ...
```suggestion TAG=${{ inputs.tag || 'ci' }} ```
intel-xpu-backend-for-triton
github_2023
others
1,881
intel
whitneywhtsang
@@ -92,9 +97,28 @@ jobs: IGC_DisableLoopUnroll=1 \ SYCL_PROGRAM_COMPILE_OPTIONS=" -vc-codegen -vc-disable-indvars-opt -doubleGRF -Xfinalizer ' -printregusage -enableBCR -DPASTokenReduction ' " \ python gemm_benchmark.py --reports $REPORTS + + TAG=${{ inputs.tag || 'ci' }}
```suggestion TAG=${{ inputs.tag || 'ci' }}-advanced ```
intel-xpu-backend-for-triton
github_2023
others
1,881
intel
whitneywhtsang
@@ -92,9 +97,28 @@ jobs: IGC_DisableLoopUnroll=1 \ SYCL_PROGRAM_COMPILE_OPTIONS=" -vc-codegen -vc-disable-indvars-opt -doubleGRF -Xfinalizer ' -printregusage -enableBCR -DPASTokenReduction ' " \ python gemm_benchmark.py --reports $REPORTS + + TAG=${{ inputs.tag || 'ci' }} + ...
```suggestion python ../../scripts/build_report.py $REPORTS/matmul-performance.csv $REPORTS/gemm-triton-advanced-report.csv --benchmark gemm --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG ```
intel-xpu-backend-for-triton
github_2023
others
1,881
intel
whitneywhtsang
@@ -92,9 +97,28 @@ jobs: IGC_DisableLoopUnroll=1 \ SYCL_PROGRAM_COMPILE_OPTIONS=" -vc-codegen -vc-disable-indvars-opt -doubleGRF -Xfinalizer ' -printregusage -enableBCR -DPASTokenReduction ' " \ python gemm_benchmark.py --reports $REPORTS + + TAG=${{ inputs.tag || 'ci' }} + ...
```suggestion python ../../scripts/build_report.py $REPORTS/matmul-performance.csv $REPORTS/gemm-triton-report.csv --benchmark gemm --compiler triton --param_cols "B,M,K,N" --tflops_col Triton-TFlops --hbm_col "Triton-GB/s" --tag $TAG python ../../scripts/build_report.py $REPORTS/matmul-performanc...
intel-xpu-backend-for-triton
github_2023
others
1,881
intel
whitneywhtsang
@@ -188,12 +211,16 @@ jobs: - name: Run flash attention benchmarks run: | cd python/tutorials + # We need latest build report + # wget https://raw.githubusercontent.com/intel/intel-xpu-backend-for-triton/llvm-target/scripts/build_report.py + wget https://raw.githubu...
What is this change about?
intel-xpu-backend-for-triton
github_2023
c
1,884
intel
sommerlukas
@@ -104,8 +103,15 @@ static PyObject *getDeviceProperties(PyObject *self, PyObject *args) { mem_bus_width, "max_work_group_size", max_group_size, "sub_group_sizes", subgroup_sizes); } +void freeKernel(PyObject *p) { + delete reinterpret_cast<sycl::kernel *>(PyCapsule_Ge...
What is the reason to store `sycl::kernel` and `sycl::kernel_bundle` as a pointer? Both classes have common reference semantics, so storing them by value will work: https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#sec:reference-semantics
intel-xpu-backend-for-triton
github_2023
c
1,884
intel
sommerlukas
@@ -227,20 +225,19 @@ static PyObject *initContext(PyObject *self, PyObject *args) { if (!(queue = PyLong_AsVoidPtr(cap))) return NULL; sycl::queue *sycl_queue = static_cast<sycl::queue *>(queue);
`queue` also has common reference semantics.
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
jopperm
@@ -216,30 +230,140 @@ class ConvertTritonToTritonGPUWarp LDBG("\n"); return; case Workload::ElementWise: - case Workload::Reduction: - case Workload::Attention: + case Workload::Reduction: { + break; + } + case Workload::Attention: { + ...
Is it guaranteed that `chainOpsA/B/C` are disjoint?
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
jopperm
@@ -216,30 +230,140 @@ class ConvertTritonToTritonGPUWarp LDBG("\n"); return; case Workload::ElementWise: - case Workload::Reduction: - case Workload::Attention: + case Workload::Reduction: { + break; + } + case Workload::Attention: { + ...
Nit: Add curly braces here as well.
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
jopperm
@@ -216,30 +230,140 @@ class ConvertTritonToTritonGPUWarp LDBG("\n"); return; case Workload::ElementWise: - case Workload::Reduction: - case Workload::Attention: + case Workload::Reduction: { + break; + } + case Workload::Attention: { + ...
I think this style is hard to follow, would be preferable to return early for all cases you don't want to handle.
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
jopperm
@@ -216,30 +230,140 @@ class ConvertTritonToTritonGPUWarp LDBG("\n"); return; case Workload::ElementWise: - case Workload::Reduction: - case Workload::Attention: + case Workload::Reduction: { + break; + } + case Workload::Attention: { + ...
Can you generalise this somehow?
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
jopperm
@@ -249,6 +373,9 @@ class ConvertTritonToTritonGPUWarp transformScfForOp(loop); } else if (auto store = dyn_cast<tt::StoreOp>(op)) { transformStoreOp(store); + } else if (auto convert = dyn_cast<ttg::ConvertLayoutOp>(op)) { + ; + // arith, math, tt::ExpandDimsOp, ...
Same as above.
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
jopperm
@@ -462,6 +615,89 @@ class ConvertTritonToTritonGPUWarp else if (op->getDialect() == arithDialect || op->getDialect() == mathDialect) ops.push_back(op->getResults()[0]); + else if (isa<tt::ReduceOp, tt::ExpandDimsOp, tt::BroadcastOp>(op)) + ; + else if (auto dot1 = dyn_c...
If you iterate here over the uses (instead of users) ...
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
jopperm
@@ -462,6 +615,89 @@ class ConvertTritonToTritonGPUWarp else if (op->getDialect() == arithDialect || op->getDialect() == mathDialect) ops.push_back(op->getResults()[0]); + else if (isa<tt::ReduceOp, tt::ExpandDimsOp, tt::BroadcastOp>(op)) + ; + else if (auto dot1 = dyn_c...
... you can just query the result num `use.getOperandNumber()` here and drop the loop to compute `resNum` above.
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
etiotto
@@ -136,6 +137,19 @@ class ConvertTritonToTritonGPUWarp ConvertTritonToTritonGPUWarp() = default; ConvertTritonToTritonGPUWarp(unsigned numWarps) { this->numWarps = numWarps; } +private: + DenseMap<Value, Attribute> valueAttrMap; + Dialect *arithDialect; + Dialect *mathDialect; + +public: + LogicalResult in...
Is this called anywhere?
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
etiotto
@@ -216,39 +230,153 @@ class ConvertTritonToTritonGPUWarp LDBG("\n"); return; case Workload::ElementWise: - case Workload::Reduction: - case Workload::Attention: + case Workload::Reduction: { break; } + case Workload::Attention: { + ...
Use static type instead of auto (unless RHS of assignment makes the type obvious).
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
etiotto
@@ -216,39 +230,153 @@ class ConvertTritonToTritonGPUWarp LDBG("\n"); return; case Workload::ElementWise: - case Workload::Reduction: - case Workload::Attention: + case Workload::Reduction: { break; } + case Workload::Attention: { + ...
asserts should have a message: `assert(cond && "...")`
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
etiotto
@@ -462,6 +616,80 @@ class ConvertTritonToTritonGPUWarp else if (op->getDialect() == arithDialect || op->getDialect() == mathDialect) ops.push_back(op->getResults()[0]); + else if (isa<tt::ReduceOp, tt::ExpandDimsOp, tt::BroadcastOp>(op)) + ; + else if (auto dot1 = dyn_c...
Would this work? ```suggestion if (op->getDialect() == arithDialect || op->getDialect() == mathDialect) { Value result = op->getResults()[0]; if (chainedVals.count(result) == 0) { chainedVals.insert(result); expandUseChain(result, chainedVals); } for (...
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
etiotto
@@ -462,6 +616,80 @@ class ConvertTritonToTritonGPUWarp else if (op->getDialect() == arithDialect || op->getDialect() == mathDialect) ops.push_back(op->getResults()[0]); + else if (isa<tt::ReduceOp, tt::ExpandDimsOp, tt::BroadcastOp>(op)) + ; + else if (auto dot1 = dyn_c...
```suggestion auto loop = cast<scf::ForOp>(arg.getOwner()->getParentOp()); ```
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
etiotto
@@ -136,6 +137,19 @@ class ConvertTritonToTritonGPUWarp ConvertTritonToTritonGPUWarp() = default; ConvertTritonToTritonGPUWarp(unsigned numWarps) { this->numWarps = numWarps; } +private: + DenseMap<Value, Attribute> valueAttrMap; + Dialect *arithDialect; + Dialect *mathDialect;
```suggestion Dialect *arithDialect = nullptr; Dialect *mathDialect = nullptr; ```
intel-xpu-backend-for-triton
github_2023
cpp
1,866
intel
etiotto
@@ -136,6 +137,19 @@ class ConvertTritonToTritonGPUWarp ConvertTritonToTritonGPUWarp() = default; ConvertTritonToTritonGPUWarp(unsigned numWarps) { this->numWarps = numWarps; } +private: + DenseMap<Value, Attribute> valueAttrMap; + Dialect *arithDialect; + Dialect *mathDialect; + +public: + LogicalResult in...
I think this function could just return void.
intel-xpu-backend-for-triton
github_2023
others
1,889
intel
etiotto
@@ -0,0 +1,47 @@ +// RUN: triton-opt %s -split-input-file --tritonintelgpu-materialize-block-pointer | FileCheck %s + +#blocked = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 16], warpsPerCTA = [2, 4], order = [1, 0]}>
not used
intel-xpu-backend-for-triton
github_2023
others
1,889
intel
etiotto
@@ -0,0 +1,47 @@ +// RUN: triton-opt %s -split-input-file --tritonintelgpu-materialize-block-pointer | FileCheck %s + +#blocked = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 16], warpsPerCTA = [2, 4], order = [1, 0]}> +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSiz...
Some of the module annotations aren't required for this test case to run (correctly).
intel-xpu-backend-for-triton
github_2023
others
1,889
intel
etiotto
@@ -270,4 +270,16 @@ def TritonIntelGPUScheduleLoad : Pass<"tritonintelgpu-schedule-load", "mlir::Mod "mlir::triton::gpu::TritonGPUDialect"]; } +def TritonIntelGPUMaterializeBlockPointer : Pass<"tritonintelgpu-materialize-block-pointer", "mlir::ModuleOp"> { + let summary = "Intel GPU mat...
Need a description of the pass here
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,133 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Utility.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/Tensor/IR/Tensor.h" +#include "trit...
`v.getDefiningOp()` will fail when `v` is a function argument (no defining operation in that case).
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,133 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Utility.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/Tensor/IR/Tensor.h" +#include "trit...
[nit]: Triton::gpu -> ttg (given that there is a namespace alias at the top of this file).
intel-xpu-backend-for-triton
github_2023
others
1,889
intel
etiotto
@@ -0,0 +1,47 @@ +// RUN: triton-opt %s -split-input-file --tritonintelgpu-materialize-block-pointer | FileCheck %s + +#blocked = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 16], warpsPerCTA = [2, 4], order = [1, 0]}> +#mma = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSiz...
Instead of extending the `i32` arguments to `i64` we can just pass in an i64 argument.
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,133 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Utility.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/Tensor/IR/Tensor.h" +#include "trit...
Several of the include statements aren't necessary and can be removed.
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,133 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Utility.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/Tensor/IR/Tensor.h" +#include "trit...
After returning from a block the next else or else if isn't necessary.
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,133 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Utility.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/Tensor/IR/Tensor.h" +#include "trit...
Already defined at line 76
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,133 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Utility.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/Tensor/IR/Tensor.h" +#include "trit...
auto -> Value
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,133 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Utility.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/Tensor/IR/Tensor.h" +#include "trit...
Would be cleaner to early exit here.
intel-xpu-backend-for-triton
github_2023
others
1,889
intel
etiotto
@@ -0,0 +1,45 @@ +// RUN: triton-opt %s --tritonintelgpu-materialize-block-pointer | FileCheck %s + +#dpas = #triton_intel_gpu.dpas<{repeatCount = 8, systolicDepth = 8, executionSize = 16, opsPerChan = 2, threadsPerWarp = 16, warpsPerCTA = [4, 2], repCluster = [1, 1], A = [8, 16], B = [16, 16], C = [8, 16]}> +#dot_a = ...
remove
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,130 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
v.getDefiningOp()->getOperand(0) --> op
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,130 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
loadOp->getResult(0) ==> loadOp.getResult().
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,130 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
For Intel XPU a dot operand will always have parent DPAS layout. This can just become an assert.
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
etiotto
@@ -0,0 +1,130 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
Actually why do we care whether the loaded value feeds a dot operation ? This pass just annotates the load operation and it doesn't need to know that the value loaded is used by a tt.dot operation.
intel-xpu-backend-for-triton
github_2023
others
1,889
intel
whitneywhtsang
@@ -270,4 +270,21 @@ def TritonIntelGPUScheduleLoad : Pass<"tritonintelgpu-schedule-load", "mlir::Mod "mlir::triton::gpu::TritonGPUDialect"]; } +def TritonIntelGPUMaterializeBlockPointer : Pass<"tritonintelgpu-materialize-block-pointer", "mlir::ModuleOp"> { + let summary = "annotate the ...
```suggestion by the Intel block IO while it is much easier deduced at TTGIR optimization phase. ```
intel-xpu-backend-for-triton
github_2023
others
1,889
intel
whitneywhtsang
@@ -4,6 +4,7 @@ add_triton_library(TritonIntelGPUTransforms MatchTargetSize.cpp Pipeliner/MatmulLoopPipeline.cpp Pipeliner/SoftwarePipeliner.cpp + MaterializeBlockPointer.cpp
alphabetical
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
whitneywhtsang
@@ -85,6 +85,8 @@ void init_triton_intel_passes_ttgpuir(py::module &&m) { bool, unsigned); ADD_PASS_WRAPPER_0("add_reduce_data_duplication", gpu::intel::createTritonIntelGPUReduceDataDuplication); + ADD_PASS_WRAPPER_0("add_materialize_block_pointer",
alphabetical
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
whitneywhtsang
@@ -0,0 +1,121 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
IIRC, we have this function in RewriteTensorPointer, can we common them?
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
whitneywhtsang
@@ -0,0 +1,121 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
does return mean interrupt or advance the walk?
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
whitneywhtsang
@@ -0,0 +1,121 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
since we already early exit for not isTensorPointerType, we can change this to a cast and remove the if block.
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
whitneywhtsang
@@ -0,0 +1,121 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
[optional] since the only user of isConstantExp is here is check if it is one, for simplicity I would name the function isOne and remove the second arg.
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
whitneywhtsang
@@ -0,0 +1,121 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
is isConstantVal a better name?
intel-xpu-backend-for-triton
github_2023
cpp
1,889
intel
whitneywhtsang
@@ -0,0 +1,121 @@ +#include "intel/include/Dialect/TritonIntelGPU/IR/Dialect.h" +#include "intel/include/Dialect/TritonIntelGPU/Transforms/Passes.h" +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/IR/Visitors.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/Transforms/Utility.h" + ...
Since https://github.com/intel/intel-xpu-backend-for-triton/pull/1834, can we change this to OWord(128 bits)?
intel-xpu-backend-for-triton
github_2023
python
1,861
intel
pbchekin
@@ -5,15 +5,35 @@ from contextlib import contextmanager from typing import Any, Dict, List from . import language as tl +import torch.xpu +import logging -if "intel_extension_for_pytorch" in sys.modules: + +def _support_elapsed_time():
I would recommend adding cache (for example, `@functools.cache`).
intel-xpu-backend-for-triton
github_2023
python
1,861
intel
alexbaden
@@ -5,15 +5,36 @@ from contextlib import contextmanager from typing import Any, Dict, List from . import language as tl +import torch.xpu +import logging -if "intel_extension_for_pytorch" in sys.modules: + +@functools.cache +def _support_elapsed_time(): + e1 = torch.xpu.Event(enable_timing=True) + e1.record(...
Why not just return true here?
intel-xpu-backend-for-triton
github_2023
cpp
1,888
intel
victor-eds
@@ -74,22 +76,31 @@ namespace mlir::triton::intel { Value convertBf16ToFp32(Location loc, ConversionPatternRewriter &rewriter, Value v) { auto moduleOp = v.getDefiningOp()->getParentWithTrait<OpTrait::SymbolTable>(); - constexpr StringLiteral baseName = "__spirv_ConvertBF16ToFINTEL"; - Ty...
Do we wanna support vectors here too as the builtin path does?
intel-xpu-backend-for-triton
github_2023
cpp
1,888
intel
victor-eds
@@ -112,6 +123,13 @@ Value convertFp32ToBf16(Location loc, ConversionPatternRewriter &rewriter, auto exponent_not_all1s = icmp_ne(check_exponent, i32_val(0)); auto exponent_all1s = icmp_eq(check_exponent, i32_val(0)); Value rounded = as_uint32; + if (rounding == RoundingMode::RTNE) { + rounded = + a...
Complex stuff
intel-xpu-backend-for-triton
github_2023
python
1,888
intel
vlad-penkin
@@ -137,6 +137,7 @@ def parse_target(self, tgt_prop) -> dict: dev_prop['has_subgroup_matrix_multiply_accumulate_tensor_float32'] = tgt_prop.get( 'has_subgroup_matrix_multiply_accumulate_tensor_float32', False) dev_prop['has_subgroup_2d_block_io'] = tgt_prop.get('has_subgroup_2d_block_io',...
`has_bf16_conversion` property value is the derivative of the device architecture. I suggest to set it explicitly to False for MTL only.
intel-xpu-backend-for-triton
github_2023
python
1,888
intel
vlad-penkin
@@ -185,7 +186,7 @@ def make_ttgir(mod, metadata, opt, properties): intel.passes.ttgpuir.add_triton_annotate_module(pm, min(properties["sub_group_sizes"]), properties["has_subgroup_2d_block_io"], p...
For consistency please move `opt.threads_per_warp)` to the next line.
intel-xpu-backend-for-triton
github_2023
others
1,846
intel
FMarno
@@ -0,0 +1,151 @@ +# Comparing CI Runs + +This script can be used to compare the results of two runs of the +"Performance E2E" CI workflow. + +## Prerequisites + +To evaluate the data from CI runs, the Python packages `pandas` and `numpy` and +their corresponding dependencies are required. + +If you also plan to plot t...
```suggestion * `-p`/`--path`: The output path, `DIR`. Per default, the script would store all intermediate data and ``` This description takes a while to get to the point.
intel-xpu-backend-for-triton
github_2023
others
1,846
intel
FMarno
@@ -0,0 +1,151 @@ +# Comparing CI Runs + +This script can be used to compare the results of two runs of the +"Performance E2E" CI workflow. + +## Prerequisites + +To evaluate the data from CI runs, the Python packages `pandas` and `numpy` and +their corresponding dependencies are required. + +If you also plan to plot t...
```suggestion * `--no-plot`: Disable plotting. Per default, the script will also visualize the output in a plot. ```
intel-xpu-backend-for-triton
github_2023
others
1,846
intel
FMarno
@@ -0,0 +1,151 @@ +# Comparing CI Runs + +This script can be used to compare the results of two runs of the +"Performance E2E" CI workflow. + +## Prerequisites + +To evaluate the data from CI runs, the Python packages `pandas` and `numpy` and +their corresponding dependencies are required. + +If you also plan to plot t...
```suggestion The evaluation mode can be used by adding the `-e`/`--eval-only` option. ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
FMarno
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
what about eval?
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
Nit: is it more reasonable to use `subprocess.check_call` (https://docs.python.org/3/library/subprocess.html#subprocess.check_call) in this context?. It throws an exception if process exit code is not 0 and prints process output to stdout/stderr.
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
```suggestion ["gh", "run", "download", "-R", "intel/intel-xpu-backend-for-triton", "-D", name, run], ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
Nit: ```suggestion path = Path(file).absolute ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
```suggestion for file in Path(directory).glob("**/*performance.csv"): df = parse_data(config, df, file) ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
```suggestion f" relative difference > 0.0 means {numerator} performs better," f" relative difference < 0.0 means {denominator} performs better") ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
vlad-penkin
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
I suggest to print the rows which are skipped.
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
vlad-penkin
@@ -0,0 +1,235 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +import re +from pathlib import Path + + +def get_config(ident): + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def download(ident): + if not shutil.whic...
Do we want to show / print rows which are present in current results only?
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,241 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +from pathlib import Path + + +def get_config(ident: str) -> str: + """Retrieve configuration name"""
Nit: ```suggestion """Retrieve configuration name.""" ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,241 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +from pathlib import Path + + +def get_config(ident: str) -> str: + """Retrieve configuration name""" + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def ...
Nit: ```suggestion """Download artifacts for given configuration and CI run ID from GitHub.""" ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,241 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +from pathlib import Path + + +def get_config(ident: str) -> str: + """Retrieve configuration name""" + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def ...
```suggestion def get_raw_data(args: argparse.Namespace) -> tuple[Optional[Path], Optional[Path]]: ``` This will also require ``` from typing import Optional ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,241 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +from pathlib import Path + + +def get_config(ident: str) -> str: + """Retrieve configuration name""" + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def ...
```suggestion def parse_directory(config: str, previous: pd.DataFrame, directory: Path) -> pd.DataFrame: ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,241 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +from pathlib import Path + + +def get_config(ident: str) -> str: + """Retrieve configuration name""" + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def ...
Is it expected to be in the previous loop? To merge all data frames for all csv files.
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,241 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +from pathlib import Path + + +def get_config(ident: str) -> str: + """Retrieve configuration name""" + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def ...
I think `Path()` creates a path relative to the current directory, `os.getcwd()` is not required. ```suggestion numDir = Path(get_config(args.numerator)) denomDir = Path(get_config(args.denominator)). ```
intel-xpu-backend-for-triton
github_2023
python
1,846
intel
pbchekin
@@ -0,0 +1,241 @@ +#! /usr/bin/env python3 + +import argparse +import shutil +import subprocess +import os +import pandas as pd +from pathlib import Path + + +def get_config(ident: str) -> str: + """Retrieve configuration name""" + if ":" in ident: + return ident.split(":")[0] + + return ident + + +def ...
Not used
intel-xpu-backend-for-triton
github_2023
others
1,846
intel
whitneywhtsang
@@ -0,0 +1,153 @@ +# Comparing CI Runs + +This script can be used to compare the results of two runs of the +"Performance E2E" CI workflow. + +## Prerequisites + +To evaluate the data from CI runs, the Python packages `pandas` and its +corresponding dependencies are required. + +If you also plan to plot the performance...
```suggestion the script are independent from the mode of execution and will be explained ```
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
etiotto
@@ -206,24 +209,29 @@ LogicalResult TritonGEN::Matrix2DBlockLoadOp::verify() { } assert(!getVnniTransform() && "Expecting vnni_transform should be false"); - switch (getElemSizeInBits()) { - case 8: - if (tileWidth != 32) - return emitOpError("tile_width for 8 bit elements when vnni_transform is " - ...
Wait a sec. When we use GenISA there will be some limitations still on the element size right ? We should have code to diagnose invalid element sizes also when GenISA APIs are used.
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
etiotto
@@ -399,14 +399,36 @@ struct LoadOpConversion Type unpackedDPASOperandType = LLVM::getFixedVectorType( typeConverter->convertType(eltTy), elemsPerLanePerDPASInst); - // pack scalars for operand A and B. - Type elemType = (isOperandA && eltTy != f32_ty) ? i16_ty : i32_ty; + // If the element typ...
The "DPAS type" to me means the result of the DPAS instruction. Please add a better description here. Also "opaque packed type" is confusing because there is nothing opaque in a packed type (to me packed type means packing elements into a larger data type which doesn't make the larger data type opaque).
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
etiotto
@@ -69,11 +70,21 @@ bool shouldRemove(tt::MakeTensorPtrOp &op, bool isUsedByStoreOp) { !(isUsedByStoreOp && ttgi::hasDpasEncoding(tensorType))) return true; - // FIXME: Temporary workaround to avoid - // compile error on fp8 2d block read - Type eltType = tensorType.getElementType(); - if (eltType.isF...
Can you add a comment to explain why this code is needed ?
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
whitneywhtsang
@@ -69,11 +70,21 @@ bool shouldRemove(tt::MakeTensorPtrOp &op, bool isUsedByStoreOp) { !(isUsedByStoreOp && ttgi::hasDpasEncoding(tensorType))) return true; - // FIXME: Temporary workaround to avoid - // compile error on fp8 2d block read - Type eltType = tensorType.getElementType(); - if (eltType.isF...
Why do we need to check dotLayout? Didn't we already check that earlier?
intel-xpu-backend-for-triton
github_2023
others
1,818
intel
etiotto
@@ -333,3 +333,84 @@ llvm.func @triton_gen.2Dblockload(%ptr : !llvm.ptr<1>, %base_width : i32, %base_ %0 = triton_gen.2Dblockload %ptr, %base_width, %base_height, %base_pitch, %x, %y {elem_size_in_bits=8, tile_width=32, tile_height=8, v_blocks=2, transpose=false, vnni_transform=false, cache_control=Default} : (!llvm...
[nit]: For consistency with other tests in this file can we "sink" the `%` inside the `[[...]]` ? So: `%[[ELEM_SIZE_IN_BITS:.*]]` --> `[[ELEM_SIZE_IN_BITS:%.*]]`
intel-xpu-backend-for-triton
github_2023
others
1,818
intel
etiotto
@@ -284,37 +284,13 @@ llvm.func @matrix_2Dblockload(%ptr : !llvm.ptr, %base_width : i32, %base_height // ----- llvm.func @matrix_2Dblockload(%ptr : !llvm.ptr, %base_width : i32, %base_height : i32, %base_pitch : i32, %x : i32, %y : i32) { - // expected-error @+1 {{'triton_gen.2Dblockload' op tile_width when vnni_t...
Hmmm. This test doesn't contain any tt.dot operation (lowered to DPAS instructions). So referring to DPAS execution size in the message is out of context.
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
etiotto
@@ -331,40 +331,17 @@ LogicalResult TritonGEN::Matrix2DBlockLoadOp::verify() { if (getElemSizeInBits() == 32 || getVnniTransform()) { if (resElemTySize != 32) return emitOpError() << "expecting result element type to be 32 bits"; - } else if (resElemTySize != 16) { - return emitOpError() << "expectin...
Referring to DPAS execution size in the message is not general enough because we will want to lower `Matrix2DBlockLoadOp` to use 2B block reads even when the load doesn't "feed" a tt.dot operation.
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
etiotto
@@ -399,14 +399,45 @@ struct LoadOpConversion Type unpackedDPASOperandType = LLVM::getFixedVectorType( typeConverter->convertType(eltTy), elemsPerLanePerDPASInst); - // pack scalars for operand A and B. - Type elemType = (isOperandA && eltTy != f32_ty) ? i16_ty : i32_ty; + Type loadResultElemTy...
[nit]: remove empty line
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
etiotto
@@ -200,6 +200,38 @@ loadCacheControlToCacheControls(Builder &builder, return builder.getAttr<TritonGEN::DecorationCacheControlAttr>(decorations); } +static bool isOCLBuiltinAvailable(TritonGEN::Matrix2DBlockLoadOp op) { + VectorType resTy = op.getRes().getType(); + unsigned resElemTySize = resTy.getElementType...
```suggestion bool needsResElemSizeEqualTo32 = op.getElemSizeInBits() == 32 || op.getVnniTransform(); assert((!needsResElemSizeEqualTo32 || resElemTySize == 32) && "Expecting 32-bit element type"); if (!needsResElemSizeEqualTo32 && resElemTySize != 16) return false; ```
intel-xpu-backend-for-triton
github_2023
cpp
1,818
intel
etiotto
@@ -200,6 +200,38 @@ loadCacheControlToCacheControls(Builder &builder, return builder.getAttr<TritonGEN::DecorationCacheControlAttr>(decorations); } +static bool isOCLBuiltinAvailable(TritonGEN::Matrix2DBlockLoadOp op) { + VectorType resTy = op.getRes().getType(); + unsigned resElemTySize = resTy.getElementType...
```suggestion if (op.getVnniTransform()) return true; uint32_t tileWidth = op.getTileWidth(); switch (op.getElemSizeInBits()) { case 8: return (tileWidth == 32); case 16: return (tileWidth == 16); case 32: return (tileWidth == 8 || tileWidth == 16); default: llvm_unreacha...
intel-xpu-backend-for-triton
github_2023
others
1,870
intel
etiotto
@@ -1072,18 +1070,6 @@ module attributes {"triton_gpu.target" = "xpu", "triton_gpu.num-ctas" = 1 : i32, // CHECK-NEXT: llvm.br ^bb2([[CMPXCHG_RES]] : i32) // CHECK-NEXT: ^bb2([[RES:%.*]]: i32): // CHECK-NEXT: [[RES_CAST:%.*]] = llvm.bitcast [[RES]] : i32 to f32 - // CHECK: llvm.call spir_fu...
Why removing the checks for the store and the barrier after the store ? Too many unnecessary changes
intel-xpu-backend-for-triton
github_2023
others
1,870
intel
etiotto
@@ -1142,21 +1129,14 @@ module attributes {"triton_gpu.target" = "xpu", "triton_gpu.num-ctas" = 1 : i32, // CHECK: [[IE1:%.*]] = llvm.insertelement %arg2, [[UNDEF1]][{{.*}} : i32] : vector<1xf32> // CHECK: [[PRED:%.*]] = llvm.and [[AND1]], %arg1 : i1 // CHECK-NEXT: [[ZERO:%.*]] = llvm.mlir.con...
Please add back this section of checks
intel-xpu-backend-for-triton
github_2023
cpp
1,870
intel
chengjunlu
@@ -1200,6 +1200,8 @@ struct AtomicRMWOpConversion emulateFp16AtomicRmw(rewriter, loc, atomicRmwAttr, valueElemTy, rmwPtr, rmwVal, rmwMask, {zero}); } else { + if (!atomicNeedsSharedMemory(op.getResult()))
Can we move the code closer to the AtomicRMWOp and add more comments about why it is required?
intel-xpu-backend-for-triton
github_2023
cpp
1,859
intel
whitneywhtsang
@@ -387,9 +387,97 @@ createBlock2DReadWithAddressPayloadUpdate(TritonGEN::Matrix2DBlockLoadOp op, attrs); }; + auto createBlock2DReadGenIsa = [&](Value ptr, + TritonGEN::Matrix2DBlockLoadOp op) { + assert(isa<LLVM::LLVMPointerType>(ptr.ge...
can we use `getGenISATypeMangling`?
intel-xpu-backend-for-triton
github_2023
cpp
1,859
intel
whitneywhtsang
@@ -387,9 +387,97 @@ createBlock2DReadWithAddressPayloadUpdate(TritonGEN::Matrix2DBlockLoadOp op, attrs); }; + auto createBlock2DReadGenIsa = [&](Value ptr, + TritonGEN::Matrix2DBlockLoadOp op) { + assert(isa<LLVM::LLVMPointerType>(ptr.ge...
can we use utilities in utility.h? like `i32_val(op.getElemSizeInBits())`
intel-xpu-backend-for-triton
github_2023
cpp
1,859
intel
whitneywhtsang
@@ -387,9 +387,97 @@ createBlock2DReadWithAddressPayloadUpdate(TritonGEN::Matrix2DBlockLoadOp op, attrs); }; + auto createBlock2DReadGenIsa = [&](Value ptr,
[nit] ```suggestion auto createBlock2DReadGenISA = [&](Value ptr, ```