Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Attach opencl.kernels metadata #2869

Open
victor-eds opened this issue Nov 28, 2024 · 1 comment · May be fixed by #2923
Open

Attach opencl.kernels metadata #2869

victor-eds opened this issue Nov 28, 2024 · 1 comment · May be fixed by #2923
Assignees
@victor-eds
Labels
enhancement New feature or request good first issue Good for newcomers performance
Milestone
0.3 [Triton] Lang...

Comments

@victor-eds
Copy link
Contributor

It's codegen's responsibility to add OCL's kernels metadata to the LLVM module. We are failing to do so and this prevents further optimizations via kernel metadata like setting a constant work-group size. Add MD in a similar way to https://github.com/intel/llvm/blob/a024380e33151b8d1e0d3abb937ff089ec1f24ba/llvm/test/CodeGen/AMDGPU/sampler-resource-id.ll#L44:

!opencl.kernels = !{!0, !1, !2}

!0 = !{ptr @test_0, !10, !20, !30, !40, !50}
!10 = !{!"kernel_arg_addr_space", i32 0, i32 1}
!20 = !{!"kernel_arg_access_qual", !"none", !"none"}
!30 = !{!"kernel_arg_type", !"sampler_t", !"int*"}
!40 = !{!"kernel_arg_base_type", !"sampler_t", !"int*"}
!50 = !{!"kernel_arg_type_qual", !"", !""}

!1 = !{ptr @test_1, !11, !21, !31, !41, !51}
!11 = !{!"kernel_arg_addr_space", i32 0, i32 0, i32 1}
!21 = !{!"kernel_arg_access_qual", !"none", !"none", !"none"}
!31 = !{!"kernel_arg_type", !"sampler_t", !"sampler_t", !"int*"}
!41 = !{!"kernel_arg_base_type", !"sampler_t", !"sampler_t", !"int*"}
!51 = !{!"kernel_arg_type_qual", !"", !"", !""}

!2 = !{ptr @test_2, !12, !22, !32, !42, !52}
!12 = !{!"kernel_arg_addr_space", i32 0, i32 0, i32 0, i32 1}
!22 = !{!"kernel_arg_access_qual", !"none", !"none", !"none", !"none"}
!32 = !{!"kernel_arg_type", !"sampler_t", !"sampler_t", !"sampler_t", !"int*"}
!42 = !{!"kernel_arg_base_type", !"sampler_t", !"sampler_t", !"sampler_t", !"int*"}
!52 = !{!"kernel_arg_type_qual", !"", !"", !"", !""}

What MD nodes to add is yet to be decided. At least the opencl.kernels and the nodes with the kernel references should be added.

We can do this by adding the required metadata to the function before lowering to LLVM dialect as a triton_gen attribute and handling the attribute in https://github.com/intel/intel-xpu-backend-for-triton/blob/main/third_party/intel/lib/Target/LLVMIR/Dialect/TritonGEN/TritonGENToLLVMIRTranslation.cpp.
@victor-eds victor-eds added enhancement New feature or request good first issue Good for newcomers performance labels Nov 28, 2024
@victor-eds victor-eds mentioned this issue Nov 28, 2024
Draft
@victor-eds victor-eds linked a pull request Dec 4, 2024 that will close this issue
Draft
@victor-eds victor-eds self-assigned this Dec 4, 2024
@victor-eds
Copy link
Contributor Author

Blocked by backend bug. May not be needed. Check after bug is fixed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@victor-eds victor-eds

Labels
enhancement New feature or request good first issue Good for newcomers performance
Projects
None yet
Milestone
0.3 [Triton] Language and Runtime
Development

Successfully merging a pull request may close this issue.

[XPU][LLVMIR] Generate opencl.kernels named metadata victor-eds/intel-xpu-backend-for-triton
2 participants
@victor-eds @vlad-penkin