Fix shuffle bug in CodeGen C. by mcourteaux · Pull Request #8567 · halide/Halide · GitHub
Skip to content

Fix shuffle bug in CodeGen C.#8567

Merged
mcourteaux merged 7 commits into
halide:mainfrom
mcourteaux:fix-shuffle-bug
Feb 18, 2025
Merged

Fix shuffle bug in CodeGen C.#8567
mcourteaux merged 7 commits into
halide:mainfrom
mcourteaux:fix-shuffle-bug

Conversation

@mcourteaux

@mcourteaux mcourteaux commented Feb 9, 2025

Copy link
Copy Markdown
Contributor

Reopened this PR from a different branch, as I wasn't expecting this would take a month to merge, so I moved it to a new branch (instead of my main branch).

Shuffle emitting in OpenCL was broken when the input to the Shuffle node were actual vectors instead of scalar. For some reason, in most of the scenarios the codegen makes it's way to the Shuffle nodes, the Shuffles are containing all vectors with 1 lane, causing no real issue without this PR. Today I ran into codegen having to shuffle multiple actual vectors in the OpenCL codegen.

@derek-gerstmann I had to disable Vulkan testing, because there is an issue on my machine with that. Could you check that out? Last excerpt from DEBUG_CODEGEN=1 output:

Skipping Hexagon offload...
Offloading GPU loops...
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32x8 vectors=2 is_interleave=false is_extract_element=false
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32x4 vectors=4 is_interleave=false is_extract_element=false
    vector shuffle x4 : 0 1 2 3 
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32x4 vectors=4 is_interleave=false is_extract_element=false
    vector shuffle x4 : 0 1 2 3 
    vector shuffle x2 : 3 1 6 7 2 4 0 5 
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32x4 vectors=1 is_interleave=false is_extract_element=false
    vector shuffle x1 : 0 1 2 3 
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32x4 vectors=1 is_interleave=false is_extract_element=false
    vector shuffle x1 : 4 5 6 7 
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32x2 vectors=1 is_interleave=false is_extract_element=false
    vector shuffle x1 : 0 1 
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32x2 vectors=1 is_interleave=false is_extract_element=false
    vector shuffle x1 : 2 3 
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32 vectors=1 is_interleave=false is_extract_element=true
 CodeGen_Vulkan_Dev::SPIRV_Emitter::visit(Shuffle): type=int32 vectors=1 is_interleave=false is_extract_element=true
Vulkan: Using static workgroup local size [8, 1, 1]...
  kernel_count = 1
  spirv_module_size[0] = 2432 bytes
Lowering Parallel Tasks...
Embedding image vulkan_buf
Embedding image vulkan_gpu_source_kernels
Target triple of initial module: x86_64--linux-gnu
Generating llvm bitcode...
Generating llvm bitcode prolog for function g...
Generating llvm bitcode for function g...
JIT compiling g for x86-64-linux-tune_znver1-avx-avx2-f16c-fma-jit-sse41-user_context-vk_v13-vulkan
[New Thread 0x7fffed4006c0 (LWP 434073)]
[New Thread 0x7fffeca006c0 (LWP 434074)]
[New Thread 0x7fffe5c006c0 (LWP 434075)]
[New Thread 0x7fffe40006c0 (LWP 434077)]
[New Thread 0x7fffe36006c0 (LWP 434079)]
NVVM compilation failed: 1
Vulkan [WARNING]: (user_context=0x7fffffffc510, id=2, name:NVIDIA) CreatePipeline: failed to compile internal representation
Vulkan [WARNING]: (user_context=0x7fffffffc510, id=2, name:NVIDIA) CreatePipeline: unexpected compilation failure
Vulkan [WARNING]: (user_context=0x7fffffffc510, id=2, name:NVIDIA) CreateComputePipeline: unexpected failure compiling SPIR-V shader: 0x9c4841a2cbb3db9d
User error triggered at /home/martijn/zec/3rd/halide/src/JITModule.cpp:1232
Error: Vulkan: Failed to create compute pipeline! vkCreateComputePipelines returned <Unknown Vulkan Result Code>
Vulkan: Failed to create compute pipeline!
Vulkan: Failed to setup compute pipeline!

@mcourteaux mcourteaux mentioned this pull request Feb 9, 2025
Closed
@abadams

abadams commented Feb 14, 2025

Copy link
Copy Markdown
Member

@mcourteaux mcourteaux requested review from halidebuildbots and removed request for halidebuildbots February 14, 2025 12:53
@derek-gerstmann derek-gerstmann self-requested a review February 14, 2025 13:45
@derek-gerstmann

derek-gerstmann commented Feb 14, 2025

Copy link
Copy Markdown
Contributor

@abadams Sure, I’ll have a look. I’ll need some time to diagnose the Vulkan codegen issues, but I’ll review what’s been submitted so far.

derek-gerstmann
derek-gerstmann requested changes Feb 14, 2025
View reviewed changes
Comment thread src/CodeGen_GPU_Dev.cpp
Comment thread src/CodeGen_GPU_Dev.cpp
Comment thread src/CodeGen_GPU_Dev.cpp
Comment thread test/correctness/shuffle.cpp
@derek-gerstmann

derek-gerstmann commented Feb 14, 2025

Copy link
Copy Markdown
Contributor

What’s the status on all the other backends? Cuda, D3D12, etc? Does this only affect GPU C-Source based code paths?

@mcourteaux

mcourteaux commented Feb 14, 2025 •
edited
Loading

Copy link
Copy Markdown
Contributor Author

What’s the status on all the other backends? Cuda, D3D12, etc? Does this only affect GPU C-Source based code paths?

Yes, this only affected the C-based backends (OpenCL, D3D12, WebGPU). The PTX codegen is based on the LLVM codegen-backend, which doesn't have this issue. The Vulkan backend works via SPIRV, which I opened the issue for in #8580.

@mcourteaux

mcourteaux commented Feb 18, 2025

Copy link
Copy Markdown
Contributor Author

@derek-gerstmann Can you update the review? I'd like to start clearing my PRs.

@derek-gerstmann

derek-gerstmann commented Feb 18, 2025

Copy link
Copy Markdown
Contributor

@mcourteaux mcourteaux mentioned this pull request Feb 18, 2025
Closed
@mcourteaux mcourteaux merged commit 2e36da4 into halide:main Feb 18, 2025
@BrewTestBot BrewTestBot mentioned this pull request Sep 16, 2025
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@derek-gerstmann derek-gerstmann derek-gerstmann approved these changes
@shoaibkamil shoaibkamil Awaiting requested review from shoaibkamil
@halidebuildbots halidebuildbots Awaiting requested review from halidebuildbots

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@mcourteaux @abadams @derek-gerstmann