Issue #21
Welcome to LLVM GPU News, a bi-weekly newsletter on all the GPU things under the LLVM umbrella. This issue covers the period from September 25 to October 14 2021.
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 Conferences
- PoCL 1.8 was released and comes with Clang/LLVM 13.0 support and improved support for SPIR/SPIR-V on CUDA. PoCL is a portable open source implementation of the OpenCL standard for CPUs, NVIDIA and AMD GPUs, and TCE ASIPs.
LLVM and Clang
Discussions
- Nicholas Wilson asked about the signature of NVPTX surface intrinsics which have
i8in name but returni16results. There are no replies at the time of writing.
Commits
- AMDGPU meta instructions (e.g.,
KILLs) are ignored when forming clauses. D106042 - AMDGPU now removes frame indices which become dead after SGPR spills. D111150
- Clang’s CUDA implementation has experimental support for texture lookup. D110089
MLIR
Discussions
Commits
gpu.launchop now supports dynamic shared memory. D110800- SPIR-V to LLVM conversion now propagates location information from
spv.GlobalVariabletollvm.mlir.global. D110207 - SPIR-V to LLVM conversion now handles entry points with the same name but different execution modes. D111193
- Standard to SPIR-V conversion now supports min/max ops. D111143
- SPIR-V atomic ops now verifies memory semantics. D111510
OpenMP (Target Offloading)
Discussions
- Matthew Grismer is interested in OpenMP offloading for Apple Silicon. Johannes Doerfert was not aware of any ongoing efforts and mentioned codegen support for these series of GPUs as a prerequisite.
Commits
- Clang now has partial support for Static Device Libraries. The initially supported targets include AMDGPU and NVPTX. D105191
External Compilers
LLPC
- Immutable samplers are supported in relocatable shaders. LLPC#1416