Issue #10
Welcome to LLVM GPU News, a bi-weekly newsletter on all the GPU things under the LLVM umbrella. This issue covers the period from April 2 to April 15 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 Conference Talks
- Nvidia’s GPU Technology Conference (GTC) was live this week.
- Registration for the SIGGRAPH 2021 conference (virtual, August 9-13) is open.
- Vulkan 1.2.175 has been published. New notable extensions include video encode and decode. You can share your feedback under the Vulkan Video GitHub issue.
- Heterogeneous Parallel Virtual Machine (HPVM) v1.0 has been released. HPVM is a retargetable compiler infrastructure that targets CPUs, GPUs, FPGAs, and accelerators. This release is based on LLVM 9.0.
LLVM and Clang
Discussions
Commits
- The work on supporting AMDGPU GFX90* continues:
- GFX90a memory model got updated. This includes documentation changes.
- GFX90a code path will reject assembly using the
scc
bit. - GFX90c code object v2 is added for backwards compatibility.
- NVPTX learned to avoid some copying in function argument lowering.
- CUDA/HIP option
-fcuda-flush-denormals-to-zero
has been renamed to-fgpu-flush-denormals-to-zero
.
MLIR
Discussions
Commits
- SPIR-V conversion now allows explicitly controlling bitwidth emulation for bitwidth unsupported in the target environment.
- A few fixes landed in SPIR-V conversion to handle dynamic ranked
memref
better. - A few utility functions are added in SPIR-V conversion for creating push constant blocks.
- Boolean
memref
s are now properly handled when converting to SPIR-V.
OpenMP (Target Offloading)
Discussions
- Kai Plociennik is asking about creating a static library containing OpenMP offloading code. There are no replies as of writing.
Commits
External Compilers
LLPC
Mesa
- Support for LLVM 8, 9, and 10 has been dropped in the
amd
compiler. This allows for a code cleanup, these versions do not support RDNA2.