Page MenuHomePhabricator
Differential D84743 Diff 281183 clang/lib/Headers/offload_macros.h

Changeset View

Standalone View

View Options

clang/lib/Headers/offload_macros.h

  • This file was added.
//===--- offload_macros.h - Universal _DEVICE Offloading Macros Header ---===//
//
jdoerfertUnsubmitted
Not Done
ReplyInline Actions

After @MaskRay noticed this, I think this should be __offload_macros.h to make it clear this is an internal header.

jdoerfert: After @MaskRay noticed this, I think this should be `__offload_macros.h` to make it clear this…
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===-----------------------------------------------------------------------===
//
// This header creates macros _DEVICE_ARCH and _DEVICE_GPU with values. This
// header exists because compiler macros are inconsistent in specifying if a
// compiliation is a device pass or a host pass. There is also inconsistency in
// how the device architecture and type are specified during a device pass. The
// inconsistencies are between OpenMP, CUDA, HIP, and OpenCL. The macro logic
// in this header is aware of these inconsistencies and sets useful values for
// _DEVICE_ARCH and _DEVICE_GPU during a device compilation. The macros will
// not be defined during a host compilation pass. So "#ifndef _DEVICE_ARCH" can
// be used by users to imply a host compilation. This header must remain a
// preprocessing header only because it is intended to be used by different
// languages.
//
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
jdoerfertUnsubmitted
Not Done
ReplyInline Actions

Nit duplicate

jdoerfert: Nit duplicate
#ifndef __OFFLOAD_MACROS_H__
Lint: Pre-merge checks
Inline Actions

clang-tidy: warning: header guard does not follow preferred style [llvm-header-guard]
not useful

Lint: Pre-merge checks: clang-tidy: warning: header guard does not follow preferred style [llvm-header-guard] [[https…
#define __OFFLOAD_MACROS_H__
#undef _DEVICE_GPU
#undef _DEVICE_ARCH
traUnsubmitted
Not Done
ReplyInline Actions

The naming seems to conflict with the current notation that GPU arch is the specific GPU variant. E.g. gfx900 or sm_60.
Perhaps we should use a higher level term. kind, vendor?

tra: The naming seems to conflict with the current notation that GPU `arch` is the specific GPU…
#if defined(_OPENMP)
// OpenMP does not set architecture macros on host pass.
// So if either set, this is an OpenMP device pass.
#if defined(__AMDGCN__) || defined(__NVPTX__)
#if defined(__AMDGCN__)
traUnsubmitted
Not Done
ReplyInline Actions

I'd just split it into separate if sections for AMDGCN and NVPTX. One less nesting level for preprocessor conditionals would be easier to follow.

tra: I'd just split it into separate `if` sections for AMDGCN and NVPTX. One less nesting level for…
#define _DEVICE_ARCH amdgcn
// _DEVICE_GPU set below
traUnsubmitted
Not Done
ReplyInline Actions

What exactly is amdgcn and how can it be used in practice? I.e. I can't use it in preprocessor conditionals.

I think you need to have numberic constants defined for the different ARCH variants.

tra: What exactly is `amdgcn` and how can it be used in practice? I.e. I can't use it in…
#else
#define _DEVICE_ARCH nvptx64
traUnsubmitted
Not Done
ReplyInline Actions

Please add a comment tracking which conditional this else is for. E.g. // __AMDGCN__

tra: Please add a comment tracking which conditional this `else` is for. E.g. `// __AMDGCN__`
#define _DEVICE_GPU __CUDA_ARCH__
traUnsubmitted
Not Done
ReplyInline Actions

Nit -- there's techically 32-bit nvptx, even though it's getting obsolete.

tra: Nit -- there's techically 32-bit nvptx, even though it's getting obsolete.
#endif
#endif
#elif defined(__CUDA_ARCH__)
// CUDA sets macros __NVPTX__ on host pass. So use __CUDA_ARCH__
// to determine if this is device pass.
#define _DEVICE_ARCH nvptx64
#define _DEVICE_GPU __CUDA_ARCH__
#elif defined(__HIP_DEVICE_COMPILE__)
// HIP sets macros __AMDGCN__ on host pass. So use __HIP_DEVICE_COMPILE__
// to determine if this is device pass.
#define _DEVICE_ARCH amdgcn
// _DEVICE_GPU set below
#elif defined(__SYCL_DEVICE_ONLY__)
#if defined(__AMDGCN__)
#define _DEVICE_ARCH amdgcn
// _DEVICE_GPU set below
#else
#define _DEVICE_ARCH nvptx64
#define _DEVICE_GPU __CUDA_ARCH__
#endif
#elif defined(__OPENCL_C_VERSION__) || defined(__OPENCL_CPP_VERSION__)
#if defined(__AMDGCN__)
#define _DEVICE_ARCH amdgcn
// _DEVICE_GPU set below
#endif
#if defined(__NVPTX__)
#define _DEVICE_ARCH nvptx64
#define _DEVICE_GPU __CUDA_ARCH__
#endif
#endif
jdoerfertUnsubmitted
Not Done
ReplyInline Actions

I guess the pattern

#if defined(__AMDGCN__)
#define _DEVICE_ARCH amdgcn
// _DEVICE_GPU set below
#endif
#if defined(__NVPTX__)
#define _DEVICE_ARCH nvptx64
#define _DEVICE_GPU __CUDA_ARCH__
#endif

is repeating here but it might make sense to lists all the cases one by one instead of a single conditional, e.g., ifdef OPENMP || SYCL || OPENCL || ...

jdoerfert: I guess the pattern ``` #if defined(__AMDGCN__) #define _DEVICE_ARCH amdgcn // _DEVICE_GPU set…
#if defined(_DEVICE_ARCH) && (_DEVICE_ARCH == amdgcn)
// AMD uses binary macros only, so create a value for _DEVICE_GPU
traUnsubmitted
Not Done
ReplyInline Actions

This does not work, does it? https://godbolt.org/z/Kn3r4x

tra: This does not work, does it? https://godbolt.org/z/Kn3r4x
#if defined(__gfx906__)
#define _DEVICE_GPU 9060
#elif defined(__gfx900__)
#define _DEVICE_GPU 9000
#elif defined(__gfx601__)
#define _DEVICE_GPU 6010
#elif defined(__gfx700__)
#define _DEVICE_GPU 7000
#elif defined(__gfx701__)
#define _DEVICE_GPU 7010
#elif defined(__gfx702__)
#define _DEVICE_GPU 7020
#elif defined(__gfx703__)
#define _DEVICE_GPU 7030
#elif defined(__gfx801__)
#define _DEVICE_GPU 8010
#elif defined(__gfx802__)
#define _DEVICE_GPU 8020
#elif defined(__gfx803__)
#define _DEVICE_GPU 8030
#elif defined(__gfx810__)
#define _DEVICE_GPU 8100
#elif defined(__gfx900__)
#define _DEVICE_GPU 9000
#elif defined(__gfx902__)
#define _DEVICE_GPU 9020
#elif defined(__gfx904__)
#define _DEVICE_GPU 9040
#elif defined(__gfx906__)
#define _DEVICE_GPU 9060
#elif defined(__gfx909__)
#define _DEVICE_GPU 9090
#elif defined(__gfx1010__)
#define _DEVICE_GPU 10100
#elif defined(__gfx1011__)
#define _DEVICE_GPU 10110
#elif defined(__gfx1012__)
#define _DEVICE_GPU 10120
#elif defined(__gfx1030__)
#define _DEVICE_GPU 10300
#else
#define _DEVICE_GPU UNKNOWN
#endif
#endif
#endif // __OFFLOAD_MACROS_H__
No newline at end of file