Issue #30

Welcome to LLVM GPU News, a bi-weekly newsletter on all the GPU things under the LLVM umbrella. This issue covers the period from February 25 to March 18, 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

• Nicolai Hähnle blogged about convergence control intrinsics.
• The Khronos group released the Vulkan SC 1.0 standard for safety-critical graphics and compute. The standard is aimed at industries like automotive and avionics. One notable difference compared to regular Vulkan is that Vulkan SC does not support online (runtime) pipeline compilation and all shader/pipeline compilations happen offline (ahead-of-time).

LLVM and Clang

Discussions

• Chris B posted an RFC on ‘Adding HLSL and DirectX support to Clang & LLVM’. The plan assumes landing the implementation piece-by-piece, instead of directly merging with the DXC codebase. HLSL can target two intermediate languages: DXIL (based on LLVM IR 3.7) and SPIR-V. The proposal was also discussed in the 3rd LLVM GPU Working Group meeting on March 18. An initial stack of revisions is already available on Phabricator.

Commits

• New AMDGPU target definitions were added for gfx940 and gfx1036. D120688, D120846

MLIR

Discussions

Commits

• gpu.global_id is added to the GPU dialect. D121548
• spv.VectorTimesScalar and spv.AssumeTrueKHROp ops are defined in the SPIR-V dialect. D121247 D121601
• A pass to unify aliased resource variables was added to SPIR-V dialect. D119872
• A canonicalization pass specifically for GLSL is exposed in SPIR-V dialect. D121222
• SPIR-V entry point ABI local size is made optional to support OpenCL. D120399
• spv.GLSL.{U|S}Clamp op type checking is fixed. D121238
• gpu.barrier is lowered to spv.ControlBarrier now. D120722

OpenMP (Target Offloading)

Discussions

Commits

• Worked around a bug in libomptarget where necessary globals were removed if the code contained no kernels. D121007

External Compilers

LLPC

• Initial work towards adding task shader support. LLPC#1716, LLPC#1720, LLPC#1721, LLPC#1733

oneAPI DPC++

CUDA/HIP support

• Added bf16 builtins operating on storage types DPCPP#5748 and optimized half builtins for fma, fmin, fmax and, exp2 DPCPP#5724.
• Optimized atomic operations DPCPP#5710, math functions DPCPP#5747, and async_work_group_copy DPCPP#5611 for the CUDA backend.

SYCL 2020 support

• Added default argument support for the work_group_size_hint attribute. DPCPP#5565
• Fixed link and compile options. DPCPP#5741, DPCPP#5476

Non-standard extensions

• Added compile-time properties extension support. DPCPP#4976
• Added Clang support for device_global. DPCPP#5597, DPCPP#5576

Explicit SIMD

• Moved a number of APIs from the experimental namespace. DPCPP#5729, DPCPP#5785, DPCPP#5773
• Improved single_task support by the ESIMD emulator. DPCPP#5671
• Enabled SVM gather/scatter for 1, 2, and 4 elements. DPCPP#5780
• Add support for lsc mem access APIs. DPCPP#5512

FPGA

• Exposed value_type and min_capacity from SYCL pipes extension class. DPCPP#5471
• Fixed max_work_group_size and reqd_work_group_size attribute arguments check. DPCPP#5592
• Added support for the USM Buffer Location Properties extension. DPCPP#5634

Misc

• Prepared SYCL compiler for opaque pointers. DPCPP#5830
• Optimized Level Zero plug-in:
  • Enabled round-robin submissions to multiple compute CCS. DPCPP#5657
  • Set minimum pooled USM allocation size on device to 512. DPCPP#5635
• Improved -lname static linking with shared objects DPCPP#5790
• Fixed a memory leak in reduction resources. DPCPP#5653
• Added XPTI-based tooling for SYCL applications for tracing profiling and sanitization. DPCPP#5389
• Fixed a bug with constexpr recursion. DPCPP#4257