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
734
intel
victor-eds
@@ -84,6 +84,37 @@ struct GPUIndexIntrinsicOpLowering : public ConvertOpToLLVMPattern<Op> { } }; +template <typename SourceOp, typename TargetOp> +class SingleDimLaunchConfigLowering : public ConvertOpToLLVMPattern<SourceOp> { +private: + unsigned indexBitwidth; + +public: + explicit SingleDimLaunchConfigLoweri...
Make `32` a constant like `resBitWidth`
intel-xpu-backend-for-triton
github_2023
cpp
734
intel
chengjunlu
@@ -84,6 +84,37 @@ struct GPUIndexIntrinsicOpLowering : public ConvertOpToLLVMPattern<Op> { } }; +template <typename SourceOp, typename TargetOp> +class SingleDimLaunchConfigLowering : public ConvertOpToLLVMPattern<SourceOp> { +private: + unsigned indexBitwidth; + +public: + explicit SingleDimLaunchConfigLoweri...
Is it better to add assert to protect 3-D dimension ops from GPU dialect?
intel-xpu-backend-for-triton
github_2023
cpp
734
intel
etiotto
@@ -84,6 +84,37 @@ struct GPUIndexIntrinsicOpLowering : public ConvertOpToLLVMPattern<Op> { } }; +template <typename SourceOp, typename TargetOp> +class SingleDimLaunchConfigLowering : public ConvertOpToLLVMPattern<SourceOp> { +private: + unsigned indexBitwidth; + +public: + explicit SingleDimLaunchConfigLoweri...
auto -> Location
intel-xpu-backend-for-triton
github_2023
others
734
intel
etiotto
@@ -127,6 +127,18 @@ def TritonGEN_GridDimZOp : TritonGEN_Op<"grid.dim.z", [Pure]> { }]; } +//===----------------------------------------------------------------------===// +// Subgroup index +//===----------------------------------------------------------------------===// + +def TritonGEN_SubgroupIdOp : TritonGE...
Add description of the operator and summary.
intel-xpu-backend-for-triton
github_2023
cpp
732
intel
whitneywhtsang
@@ -104,10 +104,80 @@ std::string translateLLVMIRToASM(llvm::Module &module, using ret = py::return_value_policy; -void findKernels(llvm::Module &M, std::set<llvm::Function *> &functions) { +static uint32_t findKernels(llvm::Module &M, + std::set<llvm::Function *> &functions) { + assert...
```suggestion ++numKernels; ```
intel-xpu-backend-for-triton
github_2023
others
722
intel
whitneywhtsang
@@ -1,8 +1,6 @@ // RUN: triton-opt %s --split-input-file --convert-triton-amdgpu-to-llvm | FileCheck %s -// FIXME: Currently unable to invoke amd passes. -// XFAIL: * -// CHECK-LABEL: wmma_dot +// CHECK-LABEL: wmma_dot
so reduce differences from openai repo: ```suggestion // CHECK-LABEL: wmma_dot ```
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +Triton architecture
Triton **A**rchitecture
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
please write **T**riton with the capital first letter
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Rewrite to: "Triton [c1]_ [c10]_ serves two main purposes: as a Triton language JIT compiler and as a low-level backend compiler for PyTorch 2.0 [c2]_. While we mainly discuss the latter, we also consider the former and support other frameworks."
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Suggested chnages are in bold: "From the PyTorch **perspective**, a model FX graph [c3]_ is captured by Dynamo [c4]_ (and AOT Autograd if necessary)**,** which is **then** passed to Inductor [c5]_. The Inductor converts the broad set of PyTorch’s ATen operators into a narrower set of stable primitive operators (PrimTor...
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Code **G**eneration **O**verview
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Triton MLIR **C**ompiler **D**etails
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
(**for example**, ptx/cubin in Nvidia case)
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
global: replace "e.g." with "for example"
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
In addition to the custom dialects**,** Triton reuses upstream ones like arith & math for math, scf/cf for control flow, index, func, nvvm for low-level GPU operations.
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
removing views
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
replace Future with Present tense
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
**Note** that Inductor can replace these native aten implementations with triton templated kernels.
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Triton **F**rontend
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
**P**ython-like
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
in this case, **it equals** 1024
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Triton **O**ptimizations
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
There **are** some common passes like inline,
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Triton Dialect
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
At this level, the optimizer runs:
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Shared Layout Class
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
an array, fastest changing axis first**.**
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Boolean value; when set to true, means when matrix is stored **in** shared memory, there will be an offset not only in the stride dimension, but also in the leading dimension.
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Distributed Layout Class
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Blocked **L**ayout
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
* **Order** – an array, fastest changing axis first**.**
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Below you can find a couple of examples from Triton’s inline doc
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
global: Use capital first letters for each word (except articles, prepositions, conjunctions) in headings
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
For Nvidia flow, these **are** emitted as inline assembly into LLVM
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Here is
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
device-specific lowering
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
From the execution perspective, the two modes are incompatible (in the driver)
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Thus, Intel GPU backed has two paths for Triton kernels compilation:
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
users install **Intel® oneAPI Base Toolkit** to have access to the runtime library.
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Triton is typically used with PyTorch, where Triton kernels consume PyTorch tensors to represent custom operations. As a result, the Triton runtime must interact with PyTorch components to ensure synchronization. These components include
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Using the lowest level possible has **the** following benefits:
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Having one component use a low-level runtime while the other uses a high-level one creates interoperability issues.
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Ultimately, this means that Triton is forced to have a dependency on SYCL runtime only to wrap the native queue handles for other components to respect the synchronization point.
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Rewrite to: Another option is to integrate SYCL kernel invocation support into the Unified runtime, making it agnostic to the input. This approach positions the Unified runtime as a versatile tool for all components. This enhancement enables Triton to operate independently of PyTorch, eliminating unnecessary depende...
intel-xpu-backend-for-triton
github_2023
others
528
intel
avladimi
@@ -0,0 +1,653 @@ +################### +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, howev...
Triton will be able to operate independently of PyTorch, eliminating the need for redundant dependencies on the SYCL runtime. This allows Triton to showcase all the advantageous qualities of relying on a low-level runtime.
intel-xpu-backend-for-triton
github_2023
others
677
intel
etiotto
@@ -0,0 +1,40 @@ +// RUN: triton-opt %s -split-input-file --convert-triton-to-tritongpu-warp="num-warps=32" | FileCheck %s +// CHECK: #triton_gpu.blocked<{sizePerThread = [32, 64], threadsPerWarp = [1, 1], warpsPerCTA = [8, 4], order = [1, 0]}> +module { + tt.func public @matmul_kernel_with_block_pointers_0d1d2d3d4d5...
[nit]: Let's us a simpler name for the function (no "0d1d2d3d4d5d")
intel-xpu-backend-for-triton
github_2023
others
677
intel
etiotto
@@ -0,0 +1,40 @@ +// RUN: triton-opt %s -split-input-file --convert-triton-to-tritongpu-warp="num-warps=32" | FileCheck %s +// CHECK: #triton_gpu.blocked<{sizePerThread = [32, 64], threadsPerWarp = [1, 1], warpsPerCTA = [8, 4], order = [1, 0]}>
Please add a CHECK in the function body to ensure this layout is used by the dot operation.
intel-xpu-backend-for-triton
github_2023
others
677
intel
etiotto
@@ -0,0 +1,41 @@ +// RUN: triton-opt %s -split-input-file | FileCheck %s +// CHECK: tt.func +#blocked = #triton_gpu.blocked<{sizePerThread = [32, 64], threadsPerWarp = [1, 1], warpsPerCTA = [8, 4], order = [1, 0]}> +module attributes {"triton_gpu.compute-capability" = 90 : i32, "triton_gpu.num-ctas" = 1 : i32, "triton_...
remove "_0d1d2d3d4d5d"
intel-xpu-backend-for-triton
github_2023
others
677
intel
etiotto
@@ -0,0 +1,41 @@ +// RUN: triton-opt %s -split-input-file | FileCheck %s
Can we explicitly add the the conversion pass to run here.
intel-xpu-backend-for-triton
github_2023
others
677
intel
etiotto
@@ -34,4 +34,26 @@ def ConvertTritonToTritonGPU: Pass<"convert-triton-to-tritongpu", "mlir::ModuleO ]; } +def ConvertTritonToTritonGPUWarp: Pass<"convert-triton-to-tritongpu-warp", "mlir::ModuleOp"> { + let summary = "Convert Triton to TritonGPU with warp distribute annotation"; + let description = [{ +
Missing description
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
etiotto
@@ -19,11 +19,15 @@ constexpr static char AttrNumThreadsPerWarp[] = "triton_gpu.threads-per-warp"; // Create the pass with numWarps passed from cl::opt. std::unique_ptr<OperationPass<ModuleOp>> createConvertTritonToTritonGPUPass(); +std::unique_ptr<OperationPass<ModuleOp>> +createConvertTritonToTritonGPUWarpPass();...
int -> unsigned
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
etiotto
@@ -0,0 +1,537 @@ +#include "triton/Conversion/TritonToTritonGPU/TritonToTritonGPUPass.h" + +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" +#include "mlir/Dialect/GPU/IR/GPUDialect.h" +#include "mlir/Dialect/Index/IR/IndexDialect.h" +#include "mlir/Pass/Pass.h" +#incl...
Please document what this pass does
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
etiotto
@@ -0,0 +1,537 @@ +#include "triton/Conversion/TritonToTritonGPU/TritonToTritonGPUPass.h" + +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" +#include "mlir/Dialect/GPU/IR/GPUDialect.h" +#include "mlir/Dialect/Index/IR/IndexDialect.h" +#include "mlir/Pass/Pass.h" +#incl...
int -> unsigned
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
etiotto
@@ -0,0 +1,537 @@ +#include "triton/Conversion/TritonToTritonGPU/TritonToTritonGPUPass.h"
Add copyright
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
etiotto
@@ -0,0 +1,537 @@ +#include "triton/Conversion/TritonToTritonGPU/TritonToTritonGPUPass.h" + +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" +#include "mlir/Dialect/GPU/IR/GPUDialect.h" +#include "mlir/Dialect/Index/IR/IndexDialect.h" +#include "mlir/Pass/Pass.h" +#incl...
These traces should not be emitted by default. I suggest to use LLVM_DEBUG facilities to print out the traces.
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
etiotto
@@ -0,0 +1,537 @@ +#include "triton/Conversion/TritonToTritonGPU/TritonToTritonGPUPass.h" + +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" +#include "mlir/Dialect/GPU/IR/GPUDialect.h" +#include "mlir/Dialect/Index/IR/IndexDialect.h" +#include "mlir/Pass/Pass.h" +#incl...
These `if`, `else if` chain should be changed into a switch.
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
etiotto
@@ -0,0 +1,537 @@ +#include "triton/Conversion/TritonToTritonGPU/TritonToTritonGPUPass.h" + +#include "mlir/Dialect/Arith/IR/Arith.h" +#include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" +#include "mlir/Dialect/GPU/IR/GPUDialect.h" +#include "mlir/Dialect/Index/IR/IndexDialect.h" +#include "mlir/Pass/Pass.h" +#incl...
assert statement should have a msg. e.g. assert(cond && "msg");
intel-xpu-backend-for-triton
github_2023
others
677
intel
whitneywhtsang
@@ -34,4 +34,28 @@ def ConvertTritonToTritonGPU: Pass<"convert-triton-to-tritongpu", "mlir::ModuleO ]; } +def ConvertTritonToTritonGPUWarp: Pass<"convert-triton-to-tritongpu-warp", "mlir::ModuleOp"> { + let summary = "Convert Triton to TritonGPU with warp distribute annotation"; + let description = [{ + ...
```suggestion This pass first analyze the kernel's workload pattern (elementwise/reduction/gemm/attention), and then figure out the best layout for key/anchor operation (dot in gemm case). Afterwards, we can get all other operation’s layout through def/use chain. Finally, each tensor...
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
```suggestion /// This file implements a pass to convert triton to tritongpu with warp /// distribute annotation. This pass first analyze the kernel's workload /// pattern (elementwise/reduction/gemm/attention), /// and then figure out the best layout for key/anchor operation (dot in /// gemm case). Afterwards, we...
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
Why do we need to define `DBGS`? It is only used in `LDBG`.
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
This comment is confusing to me.
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
```suggestion // TODO: add more ```
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
Can we guard all `dump`s in `LLVM_DEBUG`?
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
```suggestion assert(dots.size() <= 2 && "only support 1 or 2 dot in a loop"); ```
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
remove?
intel-xpu-backend-for-triton
github_2023
cpp
677
intel
whitneywhtsang
@@ -0,0 +1,465 @@ +//===- TritonToTritonGPUWarpPass.cpp - ------------------------*- C++ -*-===// +// +// Copyright 2024 Intel Corporation +// Part of the intel-xpu-backend-for-trito Project, under the Apache License +// v2.0 with LLVM Exceptions. See https://llvm.org/LICENSE.txt for license +// information. SPDX-Lice...
is it possible to have both generic and block pointers in one loop? Should we give up in that case?
intel-xpu-backend-for-triton
github_2023
cpp
713
intel
whitneywhtsang
@@ -1706,22 +1706,42 @@ struct FpToFpOpConversion static Value convertFp32ToFp16(Location loc, ConversionPatternRewriter &rewriter, const Value &v, const RoundingMode rounding) { - auto ctx = rewriter.getContext(); + MLIRContext *ctx = rewrite...
Please put a FIXME to indicate that we don't want to use `llvm.genx.GenISA.ftof`.
intel-xpu-backend-for-triton
github_2023
others
711
intel
pbchekin
@@ -60,10 +60,47 @@ export SCRIPTS_DIR=$(dirname "$0") python3 -m pip install lit python3 -m pip install pytest pytest-xdist pytest-rerunfailures -$SCRIPTS_DIR/compile-pytorch-ipex.sh $ARGS -if [ $? -ne 0 ]; then - echo "FAILED: return code $?" - exit $? +# Determine if the installed PyTorch version is the same a...
What about using more universal way, like this: ```suggestion PYTHON_VERSION=$( python -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')" ) cd wheels-py${PYTHON_VERSION}* ```
intel-xpu-backend-for-triton
github_2023
others
711
intel
etiotto
@@ -83,14 +138,13 @@ build_pytorch() { echo "**** Cloning $PYTORCH_PROJ ****" cd $BASE git clone --single-branch -b dev/triton-test-3.0 --recurse-submodules --jobs 8 https://github.com/Stonepia/pytorch.git + PYTORCH_COMMIT_ID="$(<$BASE/intel-xpu-backend-for-triton/.github/pins/pytorch.txt)"
If `--pinned` is in effect then you can build the pinned commit, otherwise I think build whatever commit is in the source code tree (for both PyTorch and IPEX).
intel-xpu-backend-for-triton
github_2023
others
711
intel
etiotto
@@ -17,6 +19,14 @@ for arg in "$@"; do BUILD_IPEX=true shift ;; + --pinned) + BUILD_PINNED=true
After reading the changes I realized --pinned means to build a pinned commit (for PyTorch and/or IPEX). The `--source` flag is orthogonal so `--source --pinned` means to build a pinned commit from source, and `--pinned` means to download a pinned wheel. Of course if the `gc` command is missing... then downloading t...
intel-xpu-backend-for-triton
github_2023
cpp
704
intel
ienkovich
@@ -58,7 +57,20 @@ computeWarpLevelHistogram(Location loc, RankedTensorType srcType, numThreadPerWarp, rewriter); ballotBits.push_back(bit); } - Value fullMask = i32_val(0xFFFFFFFF); + + uint64_t fullMaskValue = 0x0;
You can use shorter `((uint64_t)1 << numThreadPerWarp) - 1` for all numThreadPerWarp < 64.
intel-xpu-backend-for-triton
github_2023
python
704
intel
whitneywhtsang
@@ -37,7 +37,7 @@ class XPUOptions: num_ctas: int = 1 num_stages: int = 2 cluster_dims: tuple = (1, 1, 1) - threads_per_warp: int = 32
We need to careful that this change doesn't cause performance regression in TorchDynamo benchmarks.
intel-xpu-backend-for-triton
github_2023
cpp
704
intel
ienkovich
@@ -58,7 +57,9 @@ computeWarpLevelHistogram(Location loc, RankedTensorType srcType, numThreadPerWarp, rewriter); ballotBits.push_back(bit); } - Value fullMask = i32_val(0xFFFFFFFF); + + uint64_t fullMaskValue = (1u << numThreadPerWarp) - 1u;
`(1u << 32)` is 0, so you should use 64bit integer here.
intel-xpu-backend-for-triton
github_2023
cpp
710
intel
whitneywhtsang
@@ -230,23 +231,24 @@ Type getSharedMemTy(Type argType) { template <unsigned opIdx> std::function<void(int, int)> -getLoadMatrixFn(Value tensor, const SharedMemoryObject &smemObj, +getLoadMatrixFn(MemDescType descTy, const SharedMemoryObject &smemObj, DpasEncodingAttr dpasLayout, unsigned warpsPerT...
```suggestion ```
intel-xpu-backend-for-triton
github_2023
cpp
710
intel
etiotto
@@ -44,11 +44,10 @@ Value lowerSharedToDotOperandDPAS( const DpasEncodingAttr &dpasLayout, const DotOperandEncodingAttr &dotOperandLayout, bool isOuter) { auto loc = op.getLoc(); - Value src = op.getSrc(); + auto src = op.getSrc();
Triton `uses` auto pervasively. The LLVM coding guidelines (https://llvm.org/docs/CodingStandards.html#use-auto-type-deduction-to-make-code-more-readable) state that `auto` should be used more sparingly: ``` Use auto if and only if it makes the code more readable or easier to maintain. Don’t “almost always” use ...
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
whitneywhtsang
@@ -21,12 +21,25 @@ namespace py = pybind11; void init_triton_intel_passes_ttgpuir(py::module &&m) { using namespace mlir::triton::gpu; + + // Device arch + py::enum_<mlir::triton::gpu::intel::DeviceArch>(m, "DEVICE_ARCH", + py::module_local()) + .value("ATS...
Why do we need unknown, from device properties, we supposed to always know the device architecture, and if it is not ats or pvc, then we should add to the list?
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
whitneywhtsang
@@ -12,8 +12,22 @@ #include "mlir/Pass/Pass.h" namespace mlir { +namespace triton { +namespace gpu { +namespace intel { -std::unique_ptr<Pass> createTritonIntelGPUAccelerateMatmulPass(); +enum class DeviceArch { + ATS = 0, + PVC = 1, + UNKNOWN,
remove?
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
whitneywhtsang
@@ -24,24 +25,152 @@ using namespace mlir; namespace tt = mlir::triton; namespace ttg = mlir::triton::gpu; -namespace ttgi = mlir::triton::gpu::intel; +namespace ttig = mlir::triton::gpu::intel;
Got the review comment before to change this from ttig to ttgi
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
whitneywhtsang
@@ -24,24 +25,152 @@ using namespace mlir; namespace tt = mlir::triton; namespace ttg = mlir::triton::gpu; -namespace ttgi = mlir::triton::gpu::intel; +namespace ttig = mlir::triton::gpu::intel; namespace { using tt::DotOp; using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodin...
If DeviceArch::UNKNOWN is not allowed, then we don't need this assert.
intel-xpu-backend-for-triton
github_2023
others
690
intel
whitneywhtsang
@@ -0,0 +1,71 @@ +// RUN: triton-opt %s -split-input-file --tritonintelgpu-accelerate-matmul=device-architecture=PVC | FileCheck %s --check-prefix=CHECK-PVC +// RUN: triton-opt %s -split-input-file --tritonintelgpu-accelerate-matmul=device-architecture=ATS | FileCheck %s --check-prefix=CHECK-ATS + +// CHECK-PVC: #[[MMA...
add CHECK-ATS to ensure dpas layout is not generated for ATS.
intel-xpu-backend-for-triton
github_2023
others
690
intel
whitneywhtsang
@@ -0,0 +1,71 @@ +// RUN: triton-opt %s -split-input-file --tritonintelgpu-accelerate-matmul=device-architecture=PVC | FileCheck %s --check-prefix=CHECK-PVC +// RUN: triton-opt %s -split-input-file --tritonintelgpu-accelerate-matmul=device-architecture=ATS | FileCheck %s --check-prefix=CHECK-ATS + +// CHECK-PVC: #[[MMA...
add CHECK-PVC to ensure dpas layout is not generated for PVC.
intel-xpu-backend-for-triton
github_2023
python
690
intel
whitneywhtsang
@@ -111,7 +111,7 @@ def make_ttgir(mod, metadata, opt, capability): intel.passes.ttnvgpuir.add_plan_cta(pm, cluster_info) passes.ttgpuir.add_remove_layout_conversions(pm) passes.ttgpuir.add_optimize_thread_locality(pm) - passes.ttgpuir.add_accelerate_matmul(pm, capability) + int...
should we pass `DeviceArch::PVC` here?
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
whitneywhtsang
@@ -12,8 +12,22 @@ #include "mlir/Pass/Pass.h" namespace mlir { +namespace triton { +namespace gpu { +namespace intel { -std::unique_ptr<Pass> createTritonIntelGPUAccelerateMatmulPass(); +enum class DeviceArch { + ATS = 0, + PVC = 1, + NUMBER_INVALID,
Let's remove `NUMBER_INVALID`.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -12,8 +12,22 @@ #include "mlir/Pass/Pass.h" namespace mlir { +namespace triton { +namespace gpu { +namespace intel { -std::unique_ptr<Pass> createTritonIntelGPUAccelerateMatmulPass(); +enum class DeviceArch { + ATS = 0, + PVC = 1, + NUMBER_INVALID, +}; + +} // namespace intel +} // namespace gpu +} // names...
The default should be UNKNOWN
intel-xpu-backend-for-triton
github_2023
others
690
intel
etiotto
@@ -27,6 +27,15 @@ def TritonIntelGPUAccelerateMatmul "mlir::triton::gpu::intel::TritonIntelGPUDialect", "mlir::triton::TritonDialect" ]; + + let options = [ + Option<"deviceArch", "device-architecture", + "mlir::triton::gpu::intel::DeviceArch", /*default*/" mlir::triton::gpu::intel::DeviceA...
add unknown as well
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -382,17 +379,6 @@ static void decomposeMixedModeDotOp(ModuleOp mod) { if (!isFP8 || (isNativeHopperFP8 && mmaLayout.isHopper())) return; promoteType = builder.getF16Type(); - } else if (auto dpasLayout = D.getType()
Note: if this is common code it *should* handle all devices (from all vendors). IF it is not common code, then it should be in the third_party directory. OpenAI will prob. move this code into the third_party directory in the future.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
Remove the 2 and make the return small vector capable of handling any size.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
```suggestion SmallVector<unsigned> ret{1, 1}; ```
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
The c-like cast is ugly. Remove it and change the enum class underlying data type to be uint32_t
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
Instead of assigning can we just initialize (saves a copy).
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
Please add comments to explain what this code snipped is doing.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
Add comments to explain the code gen strategy this class implements, which cases it can handle and which it cannot.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
Do not use auto when the return type is not immediately clear from the context please.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
int -> unsigned? numWarps can never be negative.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
auto -> IntelDPASCapability
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -31,17 +32,144 @@ using ttg::BlockedEncodingAttr; using ttg::ConvertLayoutOp; using ttg::DotOperandEncodingAttr; using ttg::SliceEncodingAttr; +using ttgi::DeviceArch; using ttgi::DpasEncodingAttr; +struct IntelDPASCapability { + uint32_t systolicDepth; + uint32_t repeatCount; + uint32_t executionSize; + u...
auto -> unsigned... let's use the proper C++ data type throughout.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -108,21 +235,25 @@ class TritonIntelGPUAccelerateMatmulPass TritonIntelGPUAccelerateMatmulPass> { public: TritonIntelGPUAccelerateMatmulPass() = default; + TritonIntelGPUAccelerateMatmulPass(ttgi::DeviceArch arch) { + this->deviceArch = arch; + }
```suggestion TritonIntelGPUAccelerateMatmulPass(ttgi::DeviceArch arch) : deviceArch(arch) { } ```
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -108,21 +235,25 @@ class TritonIntelGPUAccelerateMatmulPass TritonIntelGPUAccelerateMatmulPass> { public: TritonIntelGPUAccelerateMatmulPass() = default;
Can we remove this constructor, we should always pass a target arch.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -14,8 +14,30 @@ namespace triton { namespace gpu { namespace intel { -bool supportDPAS(DotOp op) { - return getDPASType(op) != DPASEngineType::NOT_APPLICABLE; +bool supportDPAS(DotOp op, DeviceArch arch) { + assert(arch < DeviceArch::NUMBER_INVALID && "invalid arch");
If we had an Unknow arch this function should return false if the architecture is not known. So we can remove the assert.
intel-xpu-backend-for-triton
github_2023
cpp
690
intel
etiotto
@@ -21,12 +21,24 @@ namespace py = pybind11; void init_triton_intel_passes_ttgpuir(py::module &&m) { using namespace mlir::triton::gpu; + + // Device arch + py::enum_<mlir::triton::gpu::intel::DeviceArch>(m, "DEVICE_ARCH", + py::module_local()) + .value("ATS...
It should not be PVC by default. It should be unknown. When pytorch passes to Triton the device architecture we can use that information to affect the matmul codegen but by default we will have to generate FMAs.