[llvm-dev] LLVM GPU News #23, November 12 2021

Sun Nov 14 19:33:18 PST 2021

Hi folks,

The 23rd issue of LLVM GPU News, a bi-weekly newsletter on all the GPU
things under the LLVM umbrella, is out: <
https://llvm-gpu-news.github.io/2021/11/12/issue-23.html>.
I also paste the content below, in case you prefer to read in your email
client.

-Jakub

======================================================================

# LLVM GPU News #23, November 12 2021
Authors: Jakub Kuderski, Alexey Bader

Welcome to LLVM GPU News, a bi-weekly newsletter on all the GPU things
under the LLVM umbrella.
This issue covers the period from October 29 to November 11 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

*  [2021 LLVM Developers' Meeting](https://llvm.org/devmtg/2021-11/) is on
November 17-19. [The agenda](https://llvm.swoogo.com/2021devmtg/agenda)
features a number of GPU-related sessions:
   -  [LTO and JIT Support in LLVM OpenMP Target Offloading](
https://llvm.swoogo.com/2021DevMtg/session/705664/lightning-talks)
(Thursday)
   -  [OpenCL support in clang status: OpenCL C 3.0, improvements and
future directions](
https://llvm.swoogo.com/2021DevMtg/session/705664/lightning-talks)
(Thursday)
   -  [Late Divergence Analysis](
https://llvm.swoogo.com/2021DevMtg/session/705672/quick-talks) (Thursday)
   -  [SPIR-V support in LLVM and Clang](
https://llvm.swoogo.com/2021DevMtg/session/705687/spir-v-support-in-llvm-and-clang)
(Friday)
   -  [Optimizing OpenMP GPU Execution in LLVM](
https://llvm.swoogo.com/2021DevMtg/session/705685/optimizing-openmp-gpu-execution-in-llvm)
(Friday)

##  LLVM and Clang

### Discussions

*  Johannes Doerfert is gathering feedback on [starting a 'GPU working
group'](https://lists.llvm.org/pipermail/llvm-dev/2021-November/153654.html)
which would focus on 'development of GPU-specific optimizations,
organization of driver and backend code, dealing with issues like
convergent functions,' etc. The thread received a number of positive
replies with some ideas for additional topics of interest.

### Commits

*  Two new architectures, `spirv32` and `spirv64`, were ported from the
[LLVM SPIR-V backend Khronos repository](
https://github.com/KhronosGroup/LLVM-SPIRV-Backend). This initially reuses
the target information and code generation from the existing SPIR support.
[D109144](https://reviews.llvm.org/D109144)
*  The latest supported CUDA version was bumped to 11.5. [D113249](
https://reviews.llvm.org/D113249)
*  Debug locations will not be added to code inside AMDGPU function
[prologue](https://llvm.org/docs/AMDGPUUsage.html#kernel-prolog) as it is
entirely compiler-generated. [D113100](https://reviews.llvm.org/D113100)
*  The `AMDGPU_Gfx` calling convention now defines `s[4:29]` [SGPR
registers](https://llvm.org/docs/AMDGPUUsage.html#register-identifier) as
callee-saved. [D111637](https://reviews.llvm.org/D111637)
*  CUDA landed support for assumptions predicates for pointer address
spaces, e.g., `__builtin_assume(__isGlobal(p))`. These predicates are
portable across Clang and NVCC. [D112041](https://reviews.llvm.org/D112041)
*  Special NVPTX registers were marked as reserved to avoid machine
verifier errors. [D113008](https://reviews.llvm.org/D113008)

## MLIR

### Discussions

### Commits

*  SPIR-V to LLVM conversion now supports `shufflevector` with static
indices. [D112161](https://reviews.llvm.org/D112161)
*  `arith.bitcast`, `std.br`, and `std.cond_br` to SPIR-V conversions were
added. [D112819](https://reviews.llvm.org/D112819)

## OpenMP (Target Offloading)

### Discussions

### Commits

## External Compilers

### LLPC

*  Middle-end now supports the New Pass Manager but still default to the
Legacy Pass Manager. [LLPC#1459](
https://github.com/GPUOpen-Drivers/llpc/pull/1459)

### oneAPI DPC++

#### CUDA/HIP support

*  Added initial support for the [Tensorcore matrix extension](
https://github.com/intel/llvm/pull/4695).
*  CUDA and HIP plugins were renamed to align with the SYCL 2020
specification.
*  Fixed preprocessor-only compilation mode for the NVPTX target.

#### SYCL 2020 support

*  Implemented property traits.
*  Fixed interoperability kernel API returning invalid kernel bundle.
*  Improved performance of `sycl::queue::submit()` for commands without
dependencies.

#### Non-standard extensions

*  Fixed a bug in SVM gather/scatter emulation on host for ESIMD extension.
*  Implemented `no_offset` compile-time accessor property.
*  Enhanced matrix extension specification with JIT support for AMX and
DPAS ISA.
*  The `SPV_INTEL_hw_thread_queries` extension support was enabled by
default.

#### Misc

*  Release notes for September [were published](
https://github.com/intel/llvm/blob/sycl/sycl/ReleaseNotes.md#september21-release-notes
).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20211114/44ac9bd6/attachment.html>