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
465
intel
etiotto
@@ -129,56 +279,51 @@ class DotOpDPASConversionHelper { } } - /// Generate the GENX dialect dpas operation. Rules (for PVC): - /// - SD = 8 - /// - M = RC = 1,2,4,8 (we use 8) - /// - N = exec_size = SIMD_width = 16 - /// - Size of A, B element type = {32,16,8}, for {tf32,bf16/f16,u8/i8} - /// - K=...
Just return the RHS expression. No point storing it into a temp variable that is immediately returned.
intel-xpu-backend-for-triton
github_2023
cpp
465
intel
etiotto
@@ -129,56 +279,51 @@ class DotOpDPASConversionHelper { } } - /// Generate the GENX dialect dpas operation. Rules (for PVC): - /// - SD = 8 - /// - M = RC = 1,2,4,8 (we use 8) - /// - N = exec_size = SIMD_width = 16 - /// - Size of A, B element type = {32,16,8}, for {tf32,bf16/f16,u8/i8} - /// - K=...
auto -> unsigned (also the next line).
intel-xpu-backend-for-triton
github_2023
cpp
465
intel
etiotto
@@ -197,10 +342,10 @@ class DotOpDPASConversionHelper { if (isa<FloatType>(resElemType)) { if (width == 32) return GENX::PrecisionType::TF32;
collapse lines 342 and 343.
intel-xpu-backend-for-triton
github_2023
cpp
465
intel
etiotto
@@ -695,14 +695,17 @@ class ConvertTritonGPUOpToLLVMPatternBase { void emitDpasOffsetForCTA(const DpasEncodingAttr &dpasLayout, SmallVector<SmallVector<unsigned>> &offsets, unsigned ctaOffsetX, unsigned ctaOffsetY) const { - unsigned elemsPerThreadPerGroup...
auto -> unsigned (or uint32_t given that is used few lines below).
intel-xpu-backend-for-triton
github_2023
others
563
intel
whitneywhtsang
@@ -294,8 +310,23 @@ jobs: pip install vision/dist/*.whl python -c "import torchvision; print(torchvision.__version__)" - - name: Build torchtext wheels + - name: Save torchvision to a cache + if: ${{ (matrix.suite == 'timm_models' || matrix.suite == 'torchbench') && steps.torch...
```suggestion if: ${{ matrix.suite == 'torchbench' }} ```
intel-xpu-backend-for-triton
github_2023
others
563
intel
whitneywhtsang
@@ -294,8 +310,23 @@ jobs: pip install vision/dist/*.whl python -c "import torchvision; print(torchvision.__version__)" - - name: Build torchtext wheels + - name: Save torchvision to a cache + if: ${{ (matrix.suite == 'timm_models' || matrix.suite == 'torchbench') && steps.torch...
```suggestion if: ${{ matrix.suite == 'torchbench' && steps.torchtext-cache.outputs.status == 'miss'}} ```
intel-xpu-backend-for-triton
github_2023
others
563
intel
whitneywhtsang
@@ -308,8 +339,23 @@ jobs: pip install text/dist/*.whl python -c "import torchtext; print(torchtext.__version__)" - - name: Build torchaudio wheels + - name: Save torchtext to a cache + if: ${{ (matrix.suite == 'timm_models' || matrix.suite == 'torchbench') && steps.torchtext-ca...
```suggestion if: ${{ matrix.suite == 'torchbench' && steps.torchtext-cache.outputs.status == 'miss' }} ```
intel-xpu-backend-for-triton
github_2023
others
563
intel
whitneywhtsang
@@ -308,8 +339,23 @@ jobs: pip install text/dist/*.whl python -c "import torchtext; print(torchtext.__version__)" - - name: Build torchaudio wheels + - name: Save torchtext to a cache + if: ${{ (matrix.suite == 'timm_models' || matrix.suite == 'torchbench') && steps.torchtext-ca...
```suggestion if: ${{ matrix.suite == 'torchbench' }} ```
intel-xpu-backend-for-triton
github_2023
others
563
intel
whitneywhtsang
@@ -308,8 +339,23 @@ jobs: pip install text/dist/*.whl python -c "import torchtext; print(torchtext.__version__)" - - name: Build torchaudio wheels + - name: Save torchtext to a cache + if: ${{ (matrix.suite == 'timm_models' || matrix.suite == 'torchbench') && steps.torchtext-ca...
```suggestion if: ${{ matrix.suite == 'torchbench' && steps.torchaudio-cache.outputs.status == 'miss' }} ```
intel-xpu-backend-for-triton
github_2023
others
563
intel
whitneywhtsang
@@ -322,8 +368,15 @@ jobs: pip install audio/dist/*.whl python -c "import torchaudio; print(torchaudio.__version__)" - - name: Install PyTorch Benchmarks - if: ${{ matrix.suite == 'timm_models' || matrix.suite == 'torchbench' }} + - name: Save torchaudio to a cache + if: ...
```suggestion if: ${{ matrix.suite == 'torchbench' && steps.torchaudio-cache.outputs.status == 'miss' }} ```
intel-xpu-backend-for-triton
github_2023
others
558
intel
kurapov-peter
@@ -0,0 +1,329 @@ +//===-- GENOps.td - GEN IR dialect op definition file ------*- tablegen -*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===--...
This might need further clarification
intel-xpu-backend-for-triton
github_2023
others
541
intel
whitneywhtsang
@@ -0,0 +1,41 @@ +add_triton_library(TritonIntelGPUToLLVM + ConvertLayoutOpToLLVM/SharedToDotOperandDPAS.cpp + ConvertLayoutOpToLLVM/SharedToDotOperandFMA.cpp + ConvertLayoutOpToLLVM/SharedToDotOperandMMAv1.cpp + ConvertLayoutOpToLLVM/SharedToDotOperandMMAv2.cpp + ConvertLayoutOpToLLVM.cpp + DotOpToLL...
why `TritonGPUToLLVM` doesn't dpends on `NVGPUConversionPassIncGen`, but `TritonIntelGPUToLLVM` dpends on `NVGPUIntelConversionPassIncGen` here?
intel-xpu-backend-for-triton
github_2023
cpp
515
intel
whitneywhtsang
@@ -498,8 +498,8 @@ static Value getModuleWarpSize(ConversionPatternRewriter &rewriter, static Value getClusterCTAId(ConversionPatternRewriter &rewriter, Location loc) { - return rewriter.create<triton::nvgpu::ClusterCTAIdOp>(loc, - ...
should we do different codegen for different target? For example, ``` Value getClusterCTAId(ConversionPatternRewriter &rewriter, Location loc) const { switch (target) { case triton::Target::NVVM: return rewriter.create<triton::nvgpu::ClusterCTAIdOp>( loc, rewrit...
intel-xpu-backend-for-triton
github_2023
others
517
intel
chengjunlu
@@ -46,7 +46,7 @@ def TritonGPU_Dialect : Dialect { static int getThreadsPerWarp(ModuleOp mod) { Attribute threadsPerWarp = mod->getDiscardableAttr("triton_gpu.threads-per-warp"); if(!threadsPerWarp) { - return 32; + return triton::gpu::TritonGPUDialect::getThreadsPerWarp(mod);
It goes endless nested recursive calling `getThreadsPerWarp` -> `getThreadsPerWarp`. The default number of 32 is for the case that the old `ttgir` without the `threads-per-warp` attribute.
intel-xpu-backend-for-triton
github_2023
cpp
520
intel
etiotto
@@ -1663,9 +1663,12 @@ struct FpToFpOpConversion cvt(res, operand); return builder.launch(rewriter, loc, f32_ty, false); } break; - case mlir::triton::Target::GENX: - return rewriter.create<GENX::FpToFpOp>(loc, f32_ty, v, - GENX::RoundingMode::RTE);...
Add a TODO: the rounding mode should not be passed when we do an extension.
intel-xpu-backend-for-triton
github_2023
python
516
intel
whitneywhtsang
@@ -2577,28 +2577,40 @@ def kernel(In, Out, # for col_b in [True, False]] + [(64, 64, 64, 4, False, False, 'chain-dot', False, 'bfloat16', 'float32')]) @pytest.mark.parametrize("num_ctas", num_ctas_list) def test_dot(M, N, K, num_warps, col_a, col_b, epilogue, allow_tf32, in_dtype, out_dtype, num_ctas, device)...
Do we expect to fix this failure? If yes, `pytest.skip("FIXME: shared memory out of resource")`, else `pytest.xfail("shared memory out of resource")`
intel-xpu-backend-for-triton
github_2023
others
510
intel
leshikus
@@ -62,13 +55,11 @@ jobs: python3 -m pre_commit run --show-diff-on-failure --color=always --all-files --verbose - name: Save pip cache - if: ${{ hashFiles('.pip-cache') == '' }} - run: | - TMPDIR=/cache/${{ github.run_id }}-$RANDOM - mkdir $TMPDIR - cp -r $HO...
is the `path` actually the action output value? it seems more like the action parameter
intel-xpu-backend-for-triton
github_2023
python
512
intel
whitneywhtsang
@@ -1135,12 +1135,17 @@ def test_atomic_rmw(op, dtype_x_str, mode, sem, device): 'test_atomic_rmw for HIP currently broken in https://github.com/openai/triton. Use https://github.com/ROCmSoftwarePlatform/triton' ) - check_cuda_only(device)
Note that `check_cuda_only` is in the upstream version of test_core.py.
intel-xpu-backend-for-triton
github_2023
others
503
intel
whitneywhtsang
@@ -212,52 +231,57 @@ jobs: path: ${{ steps.triton-cache.outputs.path }} dest: ${{ steps.triton-cache.outputs.dest }} - # TODO: move to the base image - - name: Install torchvision dependencies - run: | - sudo apt update -y - sudo apt install -y libpng-dev libj...
pandas is needed for huggingface too
intel-xpu-backend-for-triton
github_2023
others
503
intel
whitneywhtsang
@@ -212,52 +233,56 @@ jobs: path: ${{ steps.triton-cache.outputs.path }} dest: ${{ steps.triton-cache.outputs.dest }} - # TODO: move to the base image - - name: Install torchvision dependencies + - name: Install python test dependencies run: | - sudo apt update -...
remove?
intel-xpu-backend-for-triton
github_2023
others
462
intel
pbchekin
@@ -0,0 +1,61 @@ +# Building Triton for Intel GPUs + +## Dependencies + +### Drivers + +To build triton from source, we need to start with installing GPU drivers. If you're using Data Center (PVC) GPUs, please follow the instructions from [here](https://dgpu-docs.intel.com/driver/installation.html#install-steps) of you...
```suggestion Next, we install python3 and venv in our system. For Ubuntu 22.04, follow these instructions ```
intel-xpu-backend-for-triton
github_2023
others
462
intel
pbchekin
@@ -0,0 +1,61 @@ +# Building Triton for Intel GPUs + +## Dependencies + +### Drivers + +To build triton from source, we need to start with installing GPU drivers. If you're using Data Center (PVC) GPUs, please follow the instructions from [here](https://dgpu-docs.intel.com/driver/installation.html#install-steps) of you...
This requires oneAPI BaseKit mentioned here: https://github.com/intel/intel-xpu-backend-for-triton/blob/llvm-target/README.md?plain=1#L26.
intel-xpu-backend-for-triton
github_2023
others
462
intel
pbchekin
@@ -0,0 +1,61 @@ +# Building Triton for Intel GPUs + +## Dependencies + +### Drivers + +To build triton from source, we need to start with installing GPU drivers. If you're using Data Center (PVC) GPUs, please follow the instructions from [here](https://dgpu-docs.intel.com/driver/installation.html#install-steps) of you...
It would be nice to mention that Python needs to be 3.9 or 3.10 for other distributions.
intel-xpu-backend-for-triton
github_2023
others
462
intel
pbchekin
@@ -0,0 +1,63 @@ +# Building Triton for Intel GPUs + +## Dependencies + +### Drivers and OneAPI Toolkit + +To build triton from source, we need to start with installing GPU drivers. If you're using Data Center (PVC) GPUs, please follow the instructions from [here](https://dgpu-docs.intel.com/driver/installation.html#in...
```suggestion We also need to install OneAPI Basekit with version >= 2024.0.1. Please follow the instructions from [here](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html) to install it. ```
intel-xpu-backend-for-triton
github_2023
others
462
intel
pbchekin
@@ -0,0 +1,63 @@ +# Building Triton for Intel GPUs + +## Dependencies + +### Drivers and OneAPI Toolkit + +To build triton from source, we need to start with installing GPU drivers. If you're using Data Center (PVC) GPUs, please follow the instructions from [here](https://dgpu-docs.intel.com/driver/installation.html#in...
Please add trailing new line to pass pre-commit check: ```suggestion This should not throw any error! ```
intel-xpu-backend-for-triton
github_2023
python
492
intel
whitneywhtsang
@@ -150,6 +150,8 @@ def test_attention_fwd_bwd(block, dtype, device, input_scale=1.0, scale=1 / 8.0, if capability[0] < 7: pytest.skip("Only test tl.dot() on devices with sm >= 70") + # Triton result is correct compared to CUDA result + # Torch is currently producing nan values pytest...
```suggestion # Note: Triton result on XPU is correct compared to the result on CUDA. pytest.skip("FIXME: Torch produces incorrect nan values") ```
intel-xpu-backend-for-triton
github_2023
others
476
intel
whitneywhtsang
@@ -0,0 +1,275 @@ +name: E2E + +on: + workflow_dispatch: + schedule: + - cron: "5 1 * * *" + +permissions: read-all + +env: + BASE: /home/runner + LLVM_SYSPATH: /home/runner/packages/llvm + BACKEND: XPU + TRITON_DISABLE_LINE_INFO: 1 + USE_AOT_DEVLIST: pvc + PYTHON_VERSION: "3.10" + PYTORCH_REPO: https://git...
Note: my runs used the wheel provided by IPEX team. Let's see if that may affect the results.
intel-xpu-backend-for-triton
github_2023
others
476
intel
whitneywhtsang
@@ -0,0 +1,275 @@ +name: E2E + +on: + workflow_dispatch: + schedule: + - cron: "5 1 * * *" + +permissions: read-all + +env: + BASE: /home/runner + LLVM_SYSPATH: /home/runner/packages/llvm + BACKEND: XPU + TRITON_DISABLE_LINE_INFO: 1 + USE_AOT_DEVLIST: pvc + PYTHON_VERSION: "3.10" + PYTORCH_REPO: https://git...
Note: `skip-triton` option will be unable after https://github.com/intel/intel-xpu-backend-for-triton/pull/480/files. ```suggestion ./scripts/compile-triton.sh --llvm ```
intel-xpu-backend-for-triton
github_2023
others
476
intel
whitneywhtsang
@@ -0,0 +1,275 @@ +name: E2E + +on: + workflow_dispatch: + schedule: + - cron: "5 1 * * *" + +permissions: read-all + +env: + BASE: /home/runner + LLVM_SYSPATH: /home/runner/packages/llvm + BACKEND: XPU + TRITON_DISABLE_LINE_INFO: 1 + USE_AOT_DEVLIST: pvc + PYTHON_VERSION: "3.10" + PYTORCH_REPO: https://git...
Can we get the transformers version from https://github.com/Stonepia/pytorch/blob/dev/triton-test-3.0/.ci/docker/ci_commit_pins/huggingface.txt?
intel-xpu-backend-for-triton
github_2023
others
476
intel
whitneywhtsang
@@ -0,0 +1,275 @@ +name: E2E + +on: + workflow_dispatch: + schedule: + - cron: "5 1 * * *" + +permissions: read-all + +env: + BASE: /home/runner + LLVM_SYSPATH: /home/runner/packages/llvm + BACKEND: XPU + TRITON_DISABLE_LINE_INFO: 1 + USE_AOT_DEVLIST: pvc + PYTHON_VERSION: "3.10" + PYTORCH_REPO: https://git...
According to https://github.com/intel/intel-xpu-backend-for-triton/wiki/End-to-End-Test#install-torch-vision, we also need to install `ffmpeg`, but the latest `ffmpeg` doesn't work with the pinned_commit, so we can do something like `conda install -c conda-forge 'ffmpeg<4.4'`.
intel-xpu-backend-for-triton
github_2023
others
476
intel
whitneywhtsang
@@ -0,0 +1,54 @@ +name: load +description: Load a directory from a cache based on a shared directory +inputs: + root: + description: + default: /cache + path: + description: A directory to load from a cache + required: true + key: + description: A cache key + required: true + symlink: + descrip...
```suggestion description: Create a symlink instead of copying from cache ```
intel-xpu-backend-for-triton
github_2023
others
476
intel
whitneywhtsang
@@ -0,0 +1,286 @@ +name: E2E + +on: + workflow_dispatch: + schedule: + - cron: "5 1 * * *"
Just curious, why you pick 5 for minute?
intel-xpu-backend-for-triton
github_2023
others
418
intel
aregm
@@ -0,0 +1,514 @@ +# Triton architecture + +The goal of this document is to describe the high-level architecture and roadmap of the solution outlining components and interactions. The intended audience is triton & triton backend developers. No a priori knowledge is assumed, however, the introduction is concise (use lin...
/s/pytorch/PyTorch
intel-xpu-backend-for-triton
github_2023
others
418
intel
aregm
@@ -0,0 +1,514 @@ +# Triton architecture + +The goal of this document is to describe the high-level architecture and roadmap of the solution outlining components and interactions. The intended audience is triton & triton backend developers. No a priori knowledge is assumed, however, the introduction is concise (use lin...
...call to the operator's native implementation in Aten.
intel-xpu-backend-for-triton
github_2023
others
418
intel
aregm
@@ -0,0 +1,514 @@ +# Triton architecture + +The goal of this document is to describe the high-level architecture and roadmap of the solution outlining components and interactions. The intended audience is triton & triton backend developers. No a priori knowledge is assumed, however, the introduction is concise (use lin...
What is the original code to see before/after?
intel-xpu-backend-for-triton
github_2023
others
418
intel
aregm
@@ -0,0 +1,514 @@ +# Triton architecture + +The goal of this document is to describe the high-level architecture and roadmap of the solution outlining components and interactions. The intended audience is triton & triton backend developers. No a priori knowledge is assumed, however, the introduction is concise (use lin...
This is good.
intel-xpu-backend-for-triton
github_2023
others
418
intel
aregm
@@ -0,0 +1,514 @@ +# Triton architecture + +The goal of this document is to describe the high-level architecture and roadmap of the solution outlining components and interactions. The intended audience is triton & triton backend developers. No a priori knowledge is assumed, however, the introduction is concise (use lin...
import torch import triton import triton.language as tl
intel-xpu-backend-for-triton
github_2023
others
418
intel
aregm
@@ -0,0 +1,514 @@ +# Triton architecture + +The goal of this document is to describe the high-level architecture and roadmap of the solution outlining components and interactions. The intended audience is triton & triton backend developers. No a priori knowledge is assumed, however, the introduction is concise (use lin...
What is SW?
intel-xpu-backend-for-triton
github_2023
others
480
intel
whitneywhtsang
@@ -33,18 +23,57 @@ for arg in "$@"; do VENV=true shift ;; - --skip-triton)
CI uses this option, please update CI accordingly.
intel-xpu-backend-for-triton
github_2023
others
480
intel
whitneywhtsang
@@ -33,18 +23,57 @@ for arg in "$@"; do VENV=true shift ;; - --skip-triton) - SKIP_TRITON=true + --help) + echo "Example usage: ./compile-triton.sh [--llvm | --triton | --clean| --venv]"
[nit] ```suggestion echo "Example usage: ./compile-triton.sh [--llvm | --triton | --clean | --venv]" ```
intel-xpu-backend-for-triton
github_2023
others
480
intel
pbchekin
@@ -33,18 +23,57 @@ for arg in "$@"; do VENV=true shift ;; - --skip-triton) - SKIP_TRITON=true + --help) + echo "Example usage: ./compile-triton.sh [--llvm | --triton | --clean | --venv]" + exit 1 + ;; + *) + ARGS+="${arg} " shift ;; esac done +...
I know it was in the previous version, but `deactivate` generates an error if we are not in venv. The error is ignored, because the script does not have `set -e` but anyways it is an error. ```suggestion elif [ -n "$VIRTUAL_ENV" ]; then echo "**** Cleaning up Python virtualenv ****" deactivate fi ```
intel-xpu-backend-for-triton
github_2023
others
480
intel
whitneywhtsang
@@ -119,7 +148,6 @@ function build_llvm { ninja check-mlir check_rc } -build_llvm ############################################################################ # Install libGenISAIntrinsics.a
`Install libGenISAIntrinsics.a` needs to happen after `build_llvm`, or else `$PACKAGES_DIR/llvm/lib` is overwritten. Suggest to move the `Install libGenISAIntrinsics.a` step to the end of `build_llvm`.
intel-xpu-backend-for-triton
github_2023
python
484
intel
whitneywhtsang
@@ -3187,8 +3187,8 @@ def kernel(x): @pytest.mark.parametrize("device", ['xpu', 'cpu', 'cpu_pinned']) def test_pointer_arguments(device): - if is_xpu() and device in ['cpu', 'cpu_pinned']: - pytest.skip("FIXME: Incorrect result on XPU") + if is_xpu() and device in ['cpu_pinned']: + pytest.skip("...
```suggestion pytest.xfail("RuntimeError: Pinned memory requires CUDA.") ```
intel-xpu-backend-for-triton
github_2023
python
475
intel
etiotto
@@ -313,11 +313,11 @@ def test_layer_norm(M, N, dtype, eps=1e-5, device='xpu'): y_ref.backward(dy, retain_graph=True) dx_ref, dw_ref, db_ref = [_.grad.clone() for _ in [x, weight, bias]] # compare - assert torch.allclose(y_tri, y_ref, atol=1e-2, rtol=0) - assert torch.allclose(dx_tri, dx_ref, atol=...
If we call this function with float32 (line 380) can we just put back the original code form the OpenAI version (which uses 1e-2 for all 4 comparisons) has ? Rationale: It would be best to avoid potential merge conflicts.
intel-xpu-backend-for-triton
github_2023
others
442
intel
whitneywhtsang
@@ -223,3 +227,11 @@ jobs: cp -r $HOME/.cache/pip/* $TMPDIR/ # ignore error if other job created a cache with the same key already mv $TMPDIR $PIP_CACHE || true + + - name: Pass rate + run: | + python3 scripts/pass_rate.py --reports ~/reports + mkdir -p /ca...
Does this work? ```suggestion python3 scripts/pass_rate.py --reports ~/reports --json > $TMPFILE.json ```
intel-xpu-backend-for-triton
github_2023
python
442
intel
whitneywhtsang
@@ -0,0 +1,122 @@ +"""Calculates and reports pass rate for unit tests.""" + +import argparse +import dataclasses +import datetime +import json +import os +import pathlib +import platform +import xml.etree.ElementTree as ET +from typing import List + + +@dataclasses.dataclass +class ReportStats: + """Report stats."""...
Given that we xfailed all NVidia specific test cases, we may want to just print the second pass rate for simplicity.
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
whitneywhtsang
@@ -45,6 +46,7 @@ inline void registerTritonDialects(mlir::DialectRegistry &registry) { // TODO: register Triton & TritonGPU passes registry.insert<mlir::triton::TritonDialect, mlir::cf::ControlFlowDialect, mlir::triton::nvidia_gpu::TritonNvidiaGPUDialect, + mlir::triton::gpu::...
[nit] ```suggestion mlir::triton::gpu::intel::TritonIntelGPUDialect, mlir::triton::nvidia_gpu::TritonNvidiaGPUDialect, ```
intel-xpu-backend-for-triton
github_2023
others
461
intel
whitneywhtsang
@@ -2,3 +2,4 @@ add_subdirectory(Triton) add_subdirectory(TritonGPU) add_subdirectory(TritonNvidiaGPU) add_subdirectory(NVGPU) +add_subdirectory(TritonIntelGPU)
[nit] ```suggestion add_subdirectory(TritonIntelGPU) add_subdirectory(TritonNvidiaGPU) add_subdirectory(NVGPU) ```
intel-xpu-backend-for-triton
github_2023
others
461
intel
whitneywhtsang
@@ -0,0 +1,96 @@ +#ifndef TRITON_INTEL_GPU_ATTRDEFS +#define TRITON_INTEL_GPU_ATTRDEFS + +include "mlir/IR/AttrTypeBase.td" +include "triton/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td" + +//===-----------------------------------------------------------...
```suggestion "unsigned":$opsPerChannel, ```
intel-xpu-backend-for-triton
github_2023
others
461
intel
whitneywhtsang
@@ -0,0 +1,96 @@ +#ifndef TRITON_INTEL_GPU_ATTRDEFS +#define TRITON_INTEL_GPU_ATTRDEFS + +include "mlir/IR/AttrTypeBase.td" +include "triton/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td" + +//===-----------------------------------------------------------...
```suggestion SmallVector<int64_t> getDPASRep(ArrayRef<int64_t> shape, int opIdx) const; ```
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
whitneywhtsang
@@ -13,14 +14,15 @@ using namespace mlir; namespace tt = mlir::triton; namespace ttg = mlir::triton::gpu; +namespace ttig = mlir::triton::gpu::intel; namespace { using tt::DotOp; using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; -using ttg::DpasEncodingAttr; using t...
```suggestion using triton::gpu::intel::DpasEncodingAttr; ```
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
whitneywhtsang
@@ -13,14 +14,15 @@ using namespace mlir; namespace tt = mlir::triton; namespace ttg = mlir::triton::gpu; +namespace ttig = mlir::triton::gpu::intel;
```suggestion ```
intel-xpu-backend-for-triton
github_2023
python
461
intel
whitneywhtsang
@@ -163,15 +163,17 @@ def __str__(self): class DpasLayout: - def __init__(self, repeatCount, warps_per_cta, ctas_per_cga, cta_split_num, cta_order): + def __init__(self, repeatCount, systolic_depth, execution_size, ops_per_chan, threads_per_warp, warps_per_cta): self.repeatCount = repeatCount + ...
remove?
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
whitneywhtsang
@@ -108,7 +109,8 @@ void init_triton_intel(py::module &&m){ m.def("load_dialects", [](mlir::MLIRContext &context) { mlir::DialectRegistry registry; registry.insert<mlir::triton::nvidia_gpu::TritonNvidiaGPUDialect, - mlir::triton::nvgpu::NVGPUDialect>(); + mlir::triton:...
```suggestion mlir::triton::gpu::intel::TritonIntelGPUDialect, mlir::triton::nvgpu::NVGPUDialect>(); ```
intel-xpu-backend-for-triton
github_2023
others
461
intel
whitneywhtsang
@@ -0,0 +1,96 @@ +#ifndef TRITON_INTEL_GPU_ATTRDEFS +#define TRITON_INTEL_GPU_ATTRDEFS + +include "mlir/IR/AttrTypeBase.td" +include "triton/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td" + +//===-----------------------------------------------------------...
```suggestion The shape of the of XMX tensor core operation is defined by systolic depth, repeat count, execution size and operands per channel. ```
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
whitneywhtsang
@@ -0,0 +1,431 @@ +#include "triton/Dialect/Triton/IR/Dialect.h" + +#include <numeric> + +#include "mlir/IR/DialectImplementation.h" +#include "mlir/IR/OpImplementation.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/IR/Dialect.h" +#include "triton/Dialect/TritonIntelGPU/IR/Dialect.h" + +#i...
remove
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
etiotto
@@ -0,0 +1,26 @@ +#ifndef TRITON_DIALECT_TRITON_INTEL_GPU_IR_DIALECT_H_
Drop trailing underscore: TRITON_DIALECT_TRITON_INTEL_GPU_IR_DIALECT_H_ -> TRITON_DIALECT_TRITON_INTEL_GPU_IR_DIALECT_H
intel-xpu-backend-for-triton
github_2023
others
461
intel
etiotto
@@ -0,0 +1,96 @@ +#ifndef TRITON_INTEL_GPU_ATTRDEFS +#define TRITON_INTEL_GPU_ATTRDEFS + +include "mlir/IR/AttrTypeBase.td" +include "triton/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td" + +//===-----------------------------------------------------------...
MMA -> DPAS
intel-xpu-backend-for-triton
github_2023
others
461
intel
etiotto
@@ -0,0 +1,96 @@ +#ifndef TRITON_INTEL_GPU_ATTRDEFS +#define TRITON_INTEL_GPU_ATTRDEFS + +include "mlir/IR/AttrTypeBase.td" +include "triton/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td" + +//===-----------------------------------------------------------...
Remove 2 empty lines pls.
intel-xpu-backend-for-triton
github_2023
others
461
intel
etiotto
@@ -0,0 +1,96 @@ +#ifndef TRITON_INTEL_GPU_ATTRDEFS +#define TRITON_INTEL_GPU_ATTRDEFS + +include "mlir/IR/AttrTypeBase.td" +include "triton/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td" + +//===-----------------------------------------------------------...
type,. -> type.
intel-xpu-backend-for-triton
github_2023
others
461
intel
etiotto
@@ -0,0 +1,31 @@ +#ifndef TRITON_INTEL_GPU_DIALECT +#define TRITON_INTEL_GPU_DIALECT + +include "mlir/IR/OpBase.td" + +def TritonIntelGPU_Dialect : Dialect { + let name = "triton_intel_gpu"; + + let cppNamespace = "::mlir::triton::gpu::intel"; + + let hasOperationAttrVerify = 1; + + let description = [{ + Triton...
why does this dialect depends on the TensorDialect ?
intel-xpu-backend-for-triton
github_2023
others
461
intel
etiotto
@@ -0,0 +1,19 @@ +#ifndef TRITON_INTEL_GPU_OPSDEFS +#define TRITON_INTEL_GPU_OPSDEFS + +include "triton/Dialect/Triton/IR/TritonTypes.td" +include "triton/Dialect/Triton/IR/TritonAttrDefs.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUTypes.td" +include "tr...
Is this needed or can it be removed ?
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
etiotto
@@ -285,10 +285,15 @@ Value loadOperand(ConversionPatternRewriter &rewriter, Location loc, auto sharedLayout = tensorTy.getEncoding().cast<SharedEncodingAttr>(); const ArrayRef<unsigned> order = sharedLayout.getOrder(); - Type elemTy = tensorTy.getElementType(); - SmallVector<int64_t> elemsPerInstr = - e...
getXMXRep -> getDPASRepetitions
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
etiotto
@@ -212,9 +212,6 @@ bool ReduceOpHelper::isSupportedLayout() { if (auto mmaLayout = srcLayout.dyn_cast<triton::gpu::MmaEncodingTrait>()) { return mmaLayout.supportReduction(); } - if (auto dpasLayout = srcLayout.dyn_cast<triton::gpu::DpasEncodingAttr>()) {
Why can this be removed ?
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
etiotto
@@ -0,0 +1,431 @@ +#include "triton/Dialect/Triton/IR/Dialect.h" + +#include <numeric> + +#include "mlir/IR/DialectImplementation.h" +#include "mlir/IR/OpImplementation.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/IR/Dialect.h" +#include "triton/Dialect/TritonIntelGPU/IR/Dialect.h" + +#i...
Please do fill this in this PR.
intel-xpu-backend-for-triton
github_2023
others
461
intel
etiotto
@@ -0,0 +1,28 @@ +#ifndef TRITON_INTEL_GPU_DIALECT +#define TRITON_INTEL_GPU_DIALECT + +include "mlir/IR/OpBase.td" + +def TritonIntelGPU_Dialect : Dialect { + let name = "triton_intel_gpu"; + + let cppNamespace = "::mlir::triton::gpu::intel"; + + let description = [{ + Triton Intel GPU Dialect. + }]; + + let d...
Remove, is empty.
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
etiotto
@@ -0,0 +1,423 @@ +#include "triton/Dialect/Triton/IR/Dialect.h" + +#include <numeric> + +#include "mlir/IR/DialectImplementation.h" +#include "mlir/IR/OpImplementation.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/IR/Dialect.h" +#include "triton/Dialect/TritonIntelGPU/IR/Dialect.h" + +#i...
MMA -> DPAS
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
etiotto
@@ -0,0 +1,423 @@ +#include "triton/Dialect/Triton/IR/Dialect.h" + +#include <numeric> + +#include "mlir/IR/DialectImplementation.h" +#include "mlir/IR/OpImplementation.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/IR/Dialect.h" +#include "triton/Dialect/TritonIntelGPU/IR/Dialect.h" + +#i...
mma -> DPAS
intel-xpu-backend-for-triton
github_2023
cpp
461
intel
etiotto
@@ -0,0 +1,423 @@ +#include "triton/Dialect/Triton/IR/Dialect.h" + +#include <numeric> + +#include "mlir/IR/DialectImplementation.h" +#include "mlir/IR/OpImplementation.h" +#include "triton/Analysis/Utility.h" +#include "triton/Dialect/TritonGPU/IR/Dialect.h" +#include "triton/Dialect/TritonIntelGPU/IR/Dialect.h" + +#i...
mmaRetEncoding -> dpasRetEncoding
intel-xpu-backend-for-triton
github_2023
others
461
intel
Jianhui-Li
@@ -537,123 +537,123 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : // ----- -#dpas = #triton_gpu.dpas<{repeatCount=8, warpsPerCTA=[1,1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [0, 1]}> -#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#dpas}> -#dot_operand_b...
if it is i8, then ops per chain should be 4, not 2.
intel-xpu-backend-for-triton
github_2023
others
461
intel
Jianhui-Li
@@ -537,123 +537,123 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : // ----- -#dpas = #triton_gpu.dpas<{repeatCount=8, warpsPerCTA=[1,1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [0, 1]}> -#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#dpas}> -#dot_operand_b...
what is purpose of "allowTF32 = true" when the data type is i8 and fp16?
intel-xpu-backend-for-triton
github_2023
others
461
intel
Jianhui-Li
@@ -537,123 +537,123 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : // ----- -#dpas = #triton_gpu.dpas<{repeatCount=8, warpsPerCTA=[1,1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [0, 1]}> -#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#dpas}> -#dot_operand_b...
warpsPerCTA = [1, 1] but "triton_gpu.num-warps" = 4 : i32 is the number of warps 1, or 4?
intel-xpu-backend-for-triton
github_2023
others
461
intel
Jianhui-Li
@@ -0,0 +1,94 @@ +#ifndef TRITON_INTEL_GPU_ATTRDEFS +#define TRITON_INTEL_GPU_ATTRDEFS + +include "mlir/IR/AttrTypeBase.td" +include "triton/Dialect/TritonIntelGPU/IR/TritonIntelGPUDialect.td" +include "triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td" + +//===-----------------------------------------------------------...
It is better to use Intel terminology consistently here. Say subgroupsPerWorkgroup
intel-xpu-backend-for-triton
github_2023
others
461
intel
Jianhui-Li
@@ -537,123 +537,123 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 1 : // ----- -#dpas = #triton_gpu.dpas<{repeatCount=8, warpsPerCTA=[1,1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [0, 1]}> -#dot_operand_a = #triton_gpu.dot_op<{opIdx=0, parent=#dpas}> -#dot_operand_b...
Need to assert (#blocked.threadsPerWarp[0] * #blocked.threadsPerWarp[0] == #mma.threadPerWarp)
intel-xpu-backend-for-triton
github_2023
cpp
430
intel
etiotto
@@ -30,14 +54,14 @@ computeWarpLevelHistogram(Location loc, RankedTensorType srcType, // The histogram is distributed across threads, each thread owns `numBins / // numThreadPerWarp` bins. SmallVector<Value> warpLevelHistogram(numBins / numThreadPerWarp, zero); + Value laneId = and_(threadId, i32_val(numThrea...
To simplify merge conflict resolution it would work better to have: ``` if (target == Target::NVVM) { // NV code here } else { // call generateVoteBallot } ``` This way, if upstream code changes these lines we do not have to push the changes inside `generateVoteBallot`
intel-xpu-backend-for-triton
github_2023
cpp
430
intel
etiotto
@@ -30,14 +54,14 @@ computeWarpLevelHistogram(Location loc, RankedTensorType srcType, // The histogram is distributed across threads, each thread owns `numBins / // numThreadPerWarp` bins. SmallVector<Value> warpLevelHistogram(numBins / numThreadPerWarp, zero); + Value laneId = and_(threadId, i32_val(numThrea...
Can we just pass `threadMask` rather than passing -1 ?
intel-xpu-backend-for-triton
github_2023
cpp
430
intel
etiotto
@@ -30,14 +54,14 @@ computeWarpLevelHistogram(Location loc, RankedTensorType srcType, // The histogram is distributed across threads, each thread owns `numBins / // numThreadPerWarp` bins. SmallVector<Value> warpLevelHistogram(numBins / numThreadPerWarp, zero); + Value laneId = and_(threadId, i32_val(numThrea...
Can you compute this inside the `generateVoteBallot` (useful to reduce potential merge conflicts).
intel-xpu-backend-for-triton
github_2023
cpp
414
intel
whitneywhtsang
@@ -503,7 +503,9 @@ class ConvertTritonGPUOpToLLVMPatternBase { auto warpsPerCTA = triton::gpu::getWarpsPerCTA(layout); auto order = triton::gpu::getOrder(layout); auto shapePerCTATile = triton::gpu::getShapePerCTATile(layout, shape); - Value warpSize = i32_val(32); + auto mod = rewriter....
AMD changed this line to `Value warpSize = i32_val(triton::gpu::getWarpSize(layout));`, can we do something similar? https://github.com/intel/intel-xpu-backend-for-triton/blob/llvm-target/third_party/amd/lib/TritonAMDGPUToLLVM/TritonGPUToLLVMBase.h#L577
intel-xpu-backend-for-triton
github_2023
python
414
intel
etiotto
@@ -369,11 +370,24 @@ def supports_target(target: tuple): def __init__(self, target: tuple) -> None: super().__init__(target) - self.capability = target[1] + self.properties = target[1] + self.capability = self.properties["device_arch"] assert isinstance(self.capability, in...
Instead of using 16 hard coded here, can you use the class member at line 340 you added ?
intel-xpu-backend-for-triton
github_2023
python
414
intel
etiotto
@@ -407,7 +407,12 @@ def format_of(ty): sycl::queue stream = *(static_cast<sycl::queue*>(pStream)); sycl::kernel kernel = *(static_cast<sycl::kernel*>(pKrnl)); - auto threads_per_warp = 32; + int threads_per_warp = 32;
why is this 32 and not 16?
intel-xpu-backend-for-triton
github_2023
python
414
intel
whitneywhtsang
@@ -495,7 +496,7 @@ def make_llir(src, metadata, options, capability): if options.extern_libs: for name, path in options.extern_libs: llvm.link_extern_lib(llvm_mod, path) - llvm.optimize_module(llvm_mod, llvm.OPTIMIZE_O3) + llvm.optimize_module(llvm_mod, llvm.OPTIMIZ...
```suggestion llvm.optimize_module(llvm_mod, llvm.OPTIMIZE_O3) ```
intel-xpu-backend-for-triton
github_2023
cpp
414
intel
whitneywhtsang
@@ -349,6 +346,10 @@ struct ReduceOpConversion auto threadsPerWarp = triton::gpu::getThreadsPerWarpWithUniqueData(srcLayout, srcShape); auto order = getOrder(srcLayout); + Value warpSize = getModuleWarpSize(rewriter, loc);
Can we move them back to the original position?
intel-xpu-backend-for-triton
github_2023
python
414
intel
etiotto
@@ -2048,7 +2048,7 @@ def test_scan_layouts(M, N, src_layout, axis, device): capability = 0 if torch.cuda.is_available(): capability = torch.cuda.get_device_capability() - kernel = triton.compile(f.name, target=(device, capability)) + kernel = triton.compile(f.name, target=(...
Do we have to explicitly set the threads per warp ? The default should work.
intel-xpu-backend-for-triton
github_2023
python
414
intel
etiotto
@@ -49,15 +49,15 @@ def test_assert(func: str): x = torch.arange(0, shape[0], dtype=torch.int32, device='xpu') y = torch.zeros(shape, dtype=x.dtype, device="xpu") if func == "device_assert": - kernel_device_assert[(1, )](x, y, BLOCK=shape[0]) + kernel_device_assert[(1, )](x, y, BLOCK=shape[...
why do we need to explicitly state what the threads per warp size is ?
intel-xpu-backend-for-triton
github_2023
python
414
intel
etiotto
@@ -85,9 +85,9 @@ def test_print(func: str, data_type: str): elif func == "static_print": kernel_static_print[(1, )](x, y, BLOCK=shape[0], PLACEHOLDER=uuid.uuid4()) elif func == "no_arg_print": - kernel_no_arg_print[(1, )](num_warps=4) + kernel_no_arg_print[(1, )](num_warps=4, threads_p...
This should work with any warp size and we shouldn't specify what is the size of the warp here.
intel-xpu-backend-for-triton
github_2023
python
448
intel
etiotto
@@ -4356,7 +4361,7 @@ def test_fp8_dot_acc(in_type_str, low_precision_acc, device): pytest.skip('test_fp8_dot_acc for HIP currently broken in upstream.') if is_xpu(): - pytest.xfail('test_fp8_dot_acc not supported on XPU.') + pytest.skip('FIXME: test_fp8_dot_acc not supported on XPU.')
Let's skip and investigate whether this is something we can support.
intel-xpu-backend-for-triton
github_2023
cpp
434
intel
whitneywhtsang
@@ -212,39 +212,44 @@ struct ReduceOpConversion // region and the accumulator values as source. void warpReduce(ConversionPatternRewriter &rewriter, Location loc, SmallVector<Value> &acc, triton::ReduceOp op, - unsigned numLaneToReduce, unsigned interleave) const { - if (aut...
```suggestion if (target == Target::GENX) return; ```
intel-xpu-backend-for-triton
github_2023
cpp
379
intel
whitneywhtsang
@@ -1540,6 +1545,15 @@ struct InsertSliceAsyncOpConversion LogicalResult matchAndRewrite(triton::gpu::InsertSliceAsyncOp op, OpAdaptor adaptor, ConversionPatternRewriter &rewriter) const override { + + // This function should not be called on the genx target since all + // InsertSliceOps ...
```suggestion // InsertSliceAsyncOp would be decomposed into InsertSliceOps by the ```
intel-xpu-backend-for-triton
github_2023
others
379
intel
whitneywhtsang
@@ -530,45 +530,6 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 8 : // ----- -#block0 = #triton_gpu.blocked<{sizePerThread = [1], threadsPerWarp = [32], warpsPerCTA = [4], order = [0], CTAsPerCGA = [1], CTASplitNum = [1], CTAOrder = [0]}>
Note to other reviewers, this test is removed in https://github.com/intel/intel-xpu-backend-for-triton/commit/b39ae5550dc69e5ac8d981633cb7631642a9d224.
intel-xpu-backend-for-triton
github_2023
others
432
intel
pbchekin
@@ -81,8 +81,12 @@ fi ############################################################################ ## Configure and build the llvm project. -C_COMPILER=/usr/bin/gcc -CXX_COMPILER=/usr/bin/g++ +if [ ! -d "$C_COMPILER" ] || [ ! -d "$CXX_COMPILER" ]; then
Wrong test `-d` for the variables. Also the logic is not clear: why if one is not set, then set both? ``` if [ -z "$C_COMPILER" ]; then C_COMPILER="gcc" fi if [ -z "$CXX_COMPILER" ]; then CXX_COMPILER="g++" fi ```
intel-xpu-backend-for-triton
github_2023
others
432
intel
pbchekin
@@ -1,7 +1,7 @@ #!/usr/bin/env bash set +o xtrace -if [ ! -d "$BASE" ]; then +if [ ! -z "$BASE" ]; then
Bash `-z` is a test for empty string. ```suggestion if [ -z "$BASE" ]; then ```
intel-xpu-backend-for-triton
github_2023
others
432
intel
pbchekin
@@ -81,8 +81,14 @@ fi ############################################################################ ## Configure and build the llvm project. -C_COMPILER=/usr/bin/gcc -CXX_COMPILER=/usr/bin/g++ +if [ ! -z "$C_COMPILER" ]; then
```suggestion if [ -z "$C_COMPILER" ]; then ```
intel-xpu-backend-for-triton
github_2023
others
432
intel
pbchekin
@@ -81,8 +81,14 @@ fi ############################################################################ ## Configure and build the llvm project. -C_COMPILER=/usr/bin/gcc -CXX_COMPILER=/usr/bin/g++ +if [ ! -z "$C_COMPILER" ]; then + C_COMPILER=`which gcc` + echo "**** C_COMPILER is set to $C_COMPILER ****" +fi +if [ ! ...
```suggestion if [ -z "$CXX_COMPILER" ]; then ```
intel-xpu-backend-for-triton
github_2023
python
424
intel
whitneywhtsang
@@ -40,10 +40,6 @@ def is_spirv(): else: THREADS_PER_WARP = 32 -if is_spirv():
can we remove `is_spirv` function as well, this is the only user.
intel-xpu-backend-for-triton
github_2023
c
429
intel
whitneywhtsang
@@ -226,8 +229,14 @@ static PyObject *loadSyclBinary(PyObject *self, PyObject *args) { auto l0_context = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(ctx); auto l0_module = create_module(l0_context, l0_device, binary_ptr, binary_size); - +
```suggestion ```
intel-xpu-backend-for-triton
github_2023
others
420
intel
pbchekin
@@ -4,7 +4,7 @@ set +o xtrace if [ ! -d "$BASE" ]; then echo "**** BASE is not given *****" echo "**** Default BASE is set to /iusers/$USER ****"
This comment needs to be updated as well.
intel-xpu-backend-for-triton
github_2023
others
420
intel
pbchekin
@@ -4,7 +4,7 @@ set +o xtrace if [ ! -d "$BASE" ]; then echo "**** BASE is not given *****" echo "**** Default BASE is set to /iusers/$USER ****" - BASE=/iusers/$USER + BASE=$(dirname "$0")/../..
What about setting it to $HOME instead? This script requires `$HOME/packages` anyway to compile Triton.
intel-xpu-backend-for-triton
github_2023
others
420
intel
pbchekin
@@ -3,8 +3,8 @@ set +o xtrace if [ ! -d "$BASE" ]; then echo "**** BASE is not given *****" - echo "**** Default BASE is set to /iusers/$USER ****" - BASE=/iusers/$USER + BASE=$(dirname "$0")/../..
I am looking to the script and looks like $BASE needs to be absolute and currently it can be relative. For example, if i am running `./scripts/compile-triton.sh`, then BASE is `./scripts/../..` and we do `cd` in the following order: ``` cd $PACKAGES_DIR # cd ./scripts/../../packages, cd $BASE # cd ./scripts/../....
intel-xpu-backend-for-triton
github_2023
others
420
intel
pbchekin
@@ -43,8 +43,8 @@ fi set +o xtrace if [ ! -d "$BASE" ]; then echo "**** BASE is not given *****" - echo "**** Default BASE is set to /iusers/$USER ****" - BASE=/iusers/$USER + BASE=$(dirname "$0")/../..
Nit: use absolute BASE for consistency with `compile-triton.sh`.
intel-xpu-backend-for-triton
github_2023
cpp
397
intel
etiotto
@@ -1331,8 +1331,7 @@ struct FpToFpOpConversion return builder.launch(rewriter, loc, f32_ty, false); } break; case mlir::triton::Target::GENX: - return rewriter.create<GENX::FpToFpOp>(loc, f32_ty, v, - GENX::RoundingMode::RTE); + return rewriter.cre...
Remove