Issue #40

Welcome to LLVM GPU News, a bi-weekly newsletter on all the GPU things under the LLVM umbrella. This issue covers the period from August 6 to August 19, 2022.

We welcome your feedback and suggestions. Let us know if we missed anything interesting, or want us to bring attention to your (sub)project, revisions under review, or proposals. Please see the bottom of the page for details on how to submit suggestions and contribute.

Industry News and Community Events

LLVM and Clang

Discussions

• Sameer Sahasrabuddhe gave an overview of his Uniformity Analysis RFC during the LLVM GPU Working Group Meeting on August 19. The presentation was recorded. Sameer is looking for reviewers and clarified that ‘the current intent of the RFC is to invite attention to how we are defining convergence and uniformity in irreducible CFGs’.

Commits

• Added support for capabilities and extensions to the SPIR-V backend. D131221
• A solver for non-trivial SchedGroup pipelines was added to the AMDGPU backend. The solver consists of an exponential-time exact algorithm and a greedy algorithm, with user-controlled knobs to select between the two. D130797

MLIR

Discussions

• Jakub Kuderski posted an RFC on adding integer add-with-carry op to the arith dialect. The goal is to support emulating 64-bit operations with 32-bit ones when lowering to both SPIR-V and LLVM dialects. An example of targets without 64-bit instructions are most mobile GPUs. The initial set of patches landed. D131893, D131908

Commits

• Added support for Intel joint matrix ops to the SPIR-V dialect. D131586
• Made SPIR-V conversion passes OperationPasses, so that downstream compilers can put them in a nested pass manager. D131591
• Refactored SPIR-V memory space mappings to better serve multiple client APIs (incl., Vulkan and OpenCL). D131409, D131410, D131411

OpenMP (Target Offloading)

Discussions

• Tom Deakin noticed a segfault at the end of taskgroup when waiting for all tasks, when targeting nvptx64. There are no replies at the time of writing.

Commits

• Fixed detecting CUDA compute capability between images (e.g., sm_60 image and sm_62 GPU). D131567
• libelf will be replaced with LLVM’s ELF handling. D131401
• dl will be replaced with LLVM’s dynamic library handling. D131401
• Fixed the driver crashing when trying to output multiple files in device-only mode. D132248
• Adding features to extract images from offloading images with clang-offload-packager. D129507

External Compilers

LLPC