[PATCH] D97318: [clang][CodeGen] Allow fp16 arg pass by register

tra added inline comments. Comment at: clang/lib/CodeGen/TargetInfo.cpp:2825 + // gcc where 16 bit integer is used in place of _Float16 or __fp16. + Lo = Integer; } Do we need to set `Hi`, too? We do set it for `int128`. CHANGES SINCE LAST ACTION

[PATCH] D97340: [HIP] Support Spack packages

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/AMDGPU.cpp:25-31 +// Look for sub-directory starts with Prefix under Path. If there is one and +// only one matching sub-directory found, append the sub-directory to Path. If +// there is no matching sub-directory

[PATCH] D97340: [HIP] Support Spack packages

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/AMDGPU.cpp:32 +static llvm::SmallString<0> findSPACKPackage(const Driver &D, + const llvm::SmallString<0> &Path, + StringRef

[PATCH] D103563: [HIP] Fix amdgcn builtin for long type

tra accepted this revision. tra added a comment. Still LGTM. CHANGES SINCE LAST ACTION https://reviews.llvm.org/D103563/new/ https://reviews.llvm.org/D103563 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mai

[PATCH] D103579: [LTO] Fix -fwhole-program-vtables handling after HIP ThinLTO patch

tra added inline comments. Comment at: clang/test/Driver/hip-options.hip:63 // RUN: %clang -### -target x86_64-unknown-linux-gnu -nogpuinc -nogpulib \ -// RUN: --cuda-gpu-arch=gfx906 -foffload-lto=thin %s 2>&1 \ -// RUN: | FileCheck -check-prefix=THINLTO %s +// RUN: --cuda

[PATCH] D103579: [LTO] Fix -fwhole-program-vtables handling after HIP ThinLTO patch

tra added inline comments. Comment at: clang/test/Driver/hip-options.hip:63 // RUN: %clang -### -target x86_64-unknown-linux-gnu -nogpuinc -nogpulib \ -// RUN: --cuda-gpu-arch=gfx906 -foffload-lto=thin %s 2>&1 \ -// RUN: | FileCheck -check-prefix=THINLTO %s +// RUN: --cuda

[PATCH] D103579: [LTO] Fix -fwhole-program-vtables handling after HIP ThinLTO patch

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. LGTM for CUDA. @yaxunl Sam, does the change make sense for HIP? Comment at: clang/test/Driver/cuda-options.cu:190-192 +// RUN: | FileCheck -check-prefix DEVICE -check-prefix DEVIC

[PATCH] D103658: CUDA/HIP: Change device-use-host-var.cu's NOT "external" check to include "addrspace"

tra added inline comments. Comment at: clang/test/CodeGenCUDA/device-use-host-var.cu:68 // NEG-NOT: @_ZL13var_host_only -// NEG-NOT: external +// NEG-NOT: external addrspace This may be too specific. What if we end up generating a variable in generic AS which

[PATCH] D101630: [HIP] Fix device-only compilation

tra added a comment. In D101630#2798975 , @yaxunl wrote: > For sure we will need -fgpu-bundle-device-output to control bundling of > intermediate files. Then adding -emit-gpu-object and -emit-gpu-bundle may be > redundant and can cause confusion. What i

[PATCH] D101630: [HIP] Fix device-only compilation

tra added a comment. In D101630#2799425 , @yaxunl wrote: > But how do we control emitting LLVM IR with or without bundle? `-emit-llvm > -emit-gpu-object` or `-emit-llvm -emit-gpu-bundle`? `-emit-*` is usually for > specifying a specific file type. Hmm.

[PATCH] D103108: [CUDA][HIP] Promote const variables to constant

tra added inline comments. Comment at: clang/test/CodeGenCUDA/device-use-host-var.cu:68 +// NEG-NOT: @_ZL13var_host_only +// NEG-NOT: external yaxunl wrote: > sbc100 wrote: > > This seems to break if the pathname where the llvm checkout lives contains > > the

[PATCH] D101630: [HIP] Add --gpu-bundle-output

tra accepted this revision. tra added inline comments. This revision is now accepted and ready to land. Comment at: clang/lib/Driver/Driver.cpp:2903 bool GPUSanitize; +Optional BundleOutput; We should document the behavior we expect from the `--gpu-bun

[PATCH] D104505: [HIP] Defer operator overloading errors

tra added a comment. I don't think we want to do this. struct S { S& operator <<(int x); }; __device__ void foo() { S s; s<<1; } :7:6: error: invalid operands to binary expression ('S' and 'int') s<<1; ~^ ~ :2:8: note: candidate function not viable:

[PATCH] D104505: [HIP] Defer operator overloading errors

tra added a comment. In D104505#2831644 , @yaxunl wrote: > However, this does cause source level incompatibilities, i.e. CUDA code that > passes nvcc does not pass clang. This patch somehow addresses that without > compromising clang's more sophisticate

[PATCH] D104505: [HIP] Defer operator overloading errors

tra added a comment. In D104505#2833271 , @yaxunl wrote: > Such host/device overloading resolution induced issue is not limited to > device functions calling host functions. It does not change the fact that the code in the test above is invalid, regard

[PATCH] D104505: [HIP] Defer operator overloading errors

tra added a comment. In D104505#2833943 , @yaxunl wrote: > We don't defer such diags by default. We only defer them under option > -fgpu-defer-diags, which users have to specify explicitly. Thank you for pointing this out. I've missed that all the test

[PATCH] D104505: [HIP] Defer operator overloading errors

tra accepted this revision. tra added inline comments. This revision is now accepted and ready to land. Comment at: clang/test/SemaCUDA/deferred-oeverload.cu:55 callee3(); // dev-error {{no matching function for call to 'callee3'}} callee4(); // com-error {{no matching functi

[PATCH] D102507: [HIP] Support in device code

tra added a comment. The key difference between C++ and CUDA/HIP, as implemented in clang, is that `__host__` and `__device__` attributes are considered during function overloading in CUDA and HIP, so `__host__ void foo()`, `__device__ void foo()` and `__host__ __device__ void foo()` are three

[PATCH] D104847: [Clang][NVPTX] Add NVPTX intrinsics and builtins for CUDA PTX 6.5 and 7.0 WMMA and MMA instructions

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. Nice. Thank you for adding support for these missing instructions! LGTM, modulo a few of cosmetic nits. Comment at: clang/include/clang/Basic/BuiltinsNVPTX.def:762 +// Builtins t

[PATCH] D106960: [OffloadArch] Library to query properties of current offload archicture

tra added a comment. In D106960#2925610 , @ye-luo wrote: > my second GPU is NVIDIA 3060Ti (sm_86) > I build my app daily with -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_80. > > About sm_80 binary able ot run on sm_86 > https://docs.nvidia.com/cuda/ampe

[PATCH] D107492: [clang] Replace asm with __asm__ in cuda header

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. LGTM in general. Comment at: clang/lib/Headers/__clang_cuda_device_functions.h:37 #if defined(__cplusplus) -__DEVICE__ void __brkpt() { asm volatile("brkpt;"); } +__DEVICE__ void

[PATCH] D107492: [clang] Replace asm with __asm__ in cuda header

tra added a comment. In D107492#2926871 , @JonChesterfield wrote: > Therefore I'd like to leave it as `__asm__ volatile`. Being the one who introduced inconsistent use of `__volatile__` and `volatile` in this header, I'm pretty sure that PTX's notion o

[PATCH] D106401: [CUDA, MemCpyOpt] Add a flag to force-enable memcpyopt and use it for CUDA.

tra updated this revision to Diff 364653. tra added a comment. Updated post D106769 Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D106401/new/ https://reviews.llvm.org/D106401 Files: clang/lib/Driver/ToolChains/

[PATCH] D106401: [CUDA, MemCpyOpt] Add a flag to force-enable memcpyopt and use it for CUDA.

tra added a comment. I've updated the patch and added a test to verify that the knob does work as expected. Please take a look. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D106401/new/ https://reviews.llvm.org/D106401 ___

[PATCH] D106401: [CUDA, MemCpyOpt] Add a flag to force-enable memcpyopt and use it for CUDA.

tra updated this revision to Diff 364847. tra added a comment. rebase to HEAD. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D106401/new/ https://reviews.llvm.org/D106401 Files: clang/lib/Driver/ToolChains/Cuda.cpp llvm/lib/Transforms/Scalar/Me

[PATCH] D106401: [CUDA, MemCpyOpt] Add a flag to force-enable memcpyopt and use it for CUDA.

This revision was landed with ongoing or failed builds. This revision was automatically updated to reflect the committed changes. Closed by commit rG6a9cf21f5a2d: [CUDA, MemCpyOpt] Add a flag to force-enable memcpyopt and use it for CUDA. (authored by tra). Repository: rG LLVM Github Monorepo

[PATCH] D107718: [cuda] Mark builtin texture/surface reference variable as 'externally_initialized'.

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. CUDA lowers texture/surface types into a special handle. I do not think `externally_initialized` matters for it. AFAICT this change is a no-op for CUDA. Repository: rG LLVM Github Monorepo CHANG

[PATCH] D77670: [CUDA] Add partial support for recent CUDA versions.

tra added a comment. In D77670#2943753 , @Hahnfeld wrote: > @tra The split between `LATEST` and `LATEST_SUPPORTED` leads to very weird > warning and error messages: Agreed, it's far from ideal. There's also more than one issue involved. > clang-14: warn

[PATCH] D77670: [CUDA] Add partial support for recent CUDA versions.

tra added a comment. In D77670#292 , @Hahnfeld wrote: >> It's also time to bump the default GPU target to something that's supported >> by the CUDA versions we reasonably expect to see. That should probably be >> sm_35 as that's probably the oldest G

[PATCH] D108235: [CUDA] Bump default GPU architecture to sm_35.

tra created this revision. tra added a reviewer: Hahnfeld. Herald added subscribers: ormris, steven_wu, bixia, hiraditya, yaxunl, emaste. tra requested review of this revision. Herald added a project: clang. It's the oldest GPU architecture currently supported by all CUDA versions clang can use.

[PATCH] D108235: [CUDA] Bump default GPU architecture to sm_35.

tra added a comment. cuda-options-freebsd.cu has been removed as it does not test anything useful -- it's a somewhat out of date copy of cuda-options.cu with replaced host triple, which does not matter for processing GPU options. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION

[PATCH] D108239: [CUDA] Add support for CUDA-11.4.

tra created this revision. tra added a reviewer: yaxunl. Herald added subscribers: dexonsmith, bixia, hiraditya, jholewinski. tra requested review of this revision. Herald added projects: clang, LLVM. Herald added a subscriber: llvm-commits. Repository: rG LLVM Github Monorepo https://reviews.l

[PATCH] D108247: [CUDA] Improve CUDA version detection and diagnostics.

tra created this revision. tra added reviewers: Hahnfeld, yaxunl. Herald added subscribers: dexonsmith, bixia. tra requested review of this revision. Herald added a project: clang. Always use cuda.h to detect CUDA version. It's a more universal approach compared to version.txt which is no longer p

[PATCH] D108248: [CUDA] Bump the latest supported CUDA version to 11.4.

tra created this revision. tra added reviewers: Hahnfeld, yaxunl. Herald added subscribers: dexonsmith, bixia. tra requested review of this revision. Herald added a project: clang. This should reduce the amount of noise issued by clang for the recent-ish CUDA versions. Clang still does not supp

[PATCH] D77670: [CUDA] Add partial support for recent CUDA versions.

tra added a comment. See the patch stack at https://reviews.llvm.org/D108248 Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D77670/new/ https://reviews.llvm.org/D77670 ___ cfe-commits mailing list cfe-comm

[PATCH] D108247: [CUDA] Improve CUDA version detection and diagnostics.

tra updated this revision to Diff 367246. tra edited the summary of this revision. tra added a comment. Do not report the version if we don't know it. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D108247/new/ https://reviews.llvm.org/D108247 Files

[PATCH] D108247: [CUDA] Improve CUDA version detection and diagnostics.

tra added inline comments. Comment at: clang/test/Driver/cuda-version-check.cu:75 -// UNKNOWN_VERSION_V: unknown CUDA version: version.txt:{{.*}}; assuming the latest supported version -// UNKNOWN_VERSION_H: unknown CUDA version: cuda.h: CUDA_VERSION={{.*}}; assuming the late

[PATCH] D108248: [CUDA] Bump the latest supported CUDA version to 11.4.

tra updated this revision to Diff 367259. tra edited the summary of this revision. tra added a comment. Prepend space to version string. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D108248/new/ https://reviews.llvm.org/D108248 Files: clang/incl

[PATCH] D108248: [CUDA] Bump the latest supported CUDA version to 11.4.

tra updated this revision to Diff 367260. tra added a comment. Undo unintentional change. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D108248/new/ https://reviews.llvm.org/D108248 Files: clang/include/clang/Basic/Cuda.h Index: clang/include/c

[PATCH] D108247: [CUDA] Improve CUDA version detection and diagnostics.

tra updated this revision to Diff 367261. tra added a comment. Prepend space to the version string. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D108247/new/ https://reviews.llvm.org/D108247 Files: clang/include/clang/Basic/Cuda.h clang/includ

[PATCH] D108247: [CUDA] Improve CUDA version detection and diagnostics.

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:103-104 +std::string VersionString = CudaVersionToString(Version); +if (!VersionString.empty()) + VersionString += " "; +D.Diag(diag::warn_drv_new_cuda_version) Hahnfeld

[PATCH] D108239: [CUDA] Add support for CUDA-11.4.

tra updated this revision to Diff 367267. tra added a comment. Fixed typos Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D108239/new/ https://reviews.llvm.org/D108239 Files: clang/include/clang/Basic/Cuda.h clang/lib/Basic/Cuda.cpp clang/lib/

[PATCH] D108239: [CUDA] Add support for CUDA-11.4.

tra added inline comments. Comment at: clang/lib/Basic/Targets/NVPTX.cpp:48-49 PTXVersion = llvm::StringSwitch(Feature) + .Case("+ptx72", 74) + .Case("+ptx71", 73) .Case("+ptx72", 72) Hahnfeld wro

[PATCH] D108248: [CUDA] Bump the latest supported CUDA version to 11.4.

tra updated this revision to Diff 367282. tra added a comment. Updated release notes. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D108248/new/ https://reviews.llvm.org/D108248 Files: clang/docs/ReleaseNotes.rst clang/include/clang/Basic/Cuda.

[PATCH] D108235: [CUDA] Bump default GPU architecture to sm_35.

tra added a comment. In D108235#2952714 , @Hahnfeld wrote: > (might be good to have an entry in the release notes?) I've updated release notes in D108248 . Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTI

[PATCH] D108247: [CUDA] Improve CUDA version detection and diagnostics.

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:209-211 + Version = FS.exists(LibDevicePath + "/libdevice.10.bc") +? Version = CudaVersion::NEW +: Version = CudaVersion::CUDA_70; Hahnfeld wr

[PATCH] D108247: [CUDA] Improve CUDA version detection and diagnostics.

tra updated this revision to Diff 367532. tra added a comment. Fixed an error spotted by reviewer. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D108247/new/ https://reviews.llvm.org/D108247 Files: clang/include/clang/Basic/Cuda.h clang/include

[PATCH] D104847: [Clang][NVPTX] Add NVPTX intrinsics and builtins for CUDA PTX 6.5 and 7.0 WMMA and MMA instructions

tra accepted this revision. tra added a comment. LGTM. Would you like me to land the patch on your behalf? Comment at: clang/lib/CodeGen/CGBuiltin.cpp:16397 unsigned NumEltsD; std::array Variants; A comment here describing expected arrangement of the va

[PATCH] D104847: [Clang][NVPTX] Add NVPTX intrinsics and builtins for CUDA PTX 6.5 and 7.0 WMMA and MMA instructions

This revision was automatically updated to reflect the committed changes. Closed by commit rG3644726a78e3: [Clang][NVPTX] Add NVPTX intrinsics and builtins for CUDA PTX 6.5 and 7.0 WMMA… (authored by steffenlarsen, committed by tra). Changed prior to commit: https://reviews.llvm.org/D104847?vs

[PATCH] D105226: [Clang] allow overriding -fbasic-block-sections

tra created this revision. tra added reviewers: tmsriram, jlebar. Herald added subscribers: pengfei, bixia. tra requested review of this revision. Herald added a project: clang. We should not error out on non-x86 targets if `-fbasic-block-sections=none` is in effect. Also, filter it out for GPU-

[PATCH] D105226: [Clang] allow overriding -fbasic-block-sections

tra updated this revision to Diff 355681. tra edited the summary of this revision. tra added a comment. comment typo fix. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D105226/new/ https://reviews.llvm.org/D105226 Files: clang/lib/Driver/ToolChai

[PATCH] D105226: [Clang] allow overriding -fbasic-block-sections

This revision was landed with ongoing or failed builds. This revision was automatically updated to reflect the committed changes. Closed by commit rGcab5f89cfd9e: [Clang] allow overriding -fbasic-block-sections (authored by tra). Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION

[PATCH] D105135: [Internalize] Preserve variables externally initialized.

tra added inline comments. Comment at: clang/test/CodeGenCUDA/host-used-device-var.cu:21-24 -__device__ int v1; - -// DEV-NEG-NOT: @v2 -__constant__ int v2; These should be changed to positive checks to verify that they are emitted. Ditto for other tests. =

[PATCH] D105295: [CUDA] Only allow NVIDIA offload-arch during CUDA compilation.

tra created this revision. tra added a reviewer: yaxunl. Herald added a subscriber: bixia. tra requested review of this revision. Herald added a project: clang. Otherwise, if someone specifies a valid AMD arch, we may end up triggering an assertion on unexpected arch later on. Current tests didn'

[PATCH] D104847: [Clang][NVPTX] Add NVPTX intrinsics and builtins for CUDA PTX 6.5 and 7.0 WMMA and MMA instructions

tra added inline comments. Comment at: clang/include/clang/Basic/BuiltinsNVPTX.def:727 TARGET_BUILTIN(__bmma_m8n8k128_ld_c, "vi*iC*UiIi", "", AND(SM_75,PTX63)) TARGET_BUILTIN(__bmma_m8n8k128_mma_xor_popc_b1, "vi*iC*iC*iC*Ii", "", AND(SM_75,PTX63)) TARGET_BUILTIN(__bmma_m8n8k1

[PATCH] D104847: [Clang][NVPTX] Add NVPTX intrinsics and builtins for CUDA PTX 6.5 and 7.0 WMMA and MMA instructions

tra added inline comments. Comment at: clang/test/CodeGen/builtins-nvptx-mma.cu:781-786 + // CHECK_PTX70_SM80: call {{.*}} @llvm.nvvm.wmma.m16n16k8.load.c.col.stride.f32 + // expected-error-re@+1 {{'__mma_tf32_m16n16k8_ld_c' needs target feature (sm_80{{.*}},(ptx70{{.* +

[PATCH] D104847: [Clang][NVPTX] Add NVPTX intrinsics and builtins for CUDA PTX 6.5 and 7.0 WMMA and MMA instructions

tra added inline comments. Comment at: clang/test/CodeGen/builtins-nvptx-mma.cu:781-786 + // CHECK_PTX70_SM80: call {{.*}} @llvm.nvvm.wmma.m16n16k8.load.c.col.stride.f32 + // expected-error-re@+1 {{'__mma_tf32_m16n16k8_ld_c' needs target feature (sm_80{{.*}},(ptx70{{.* +

[PATCH] D105384: [NVPTX, CUDA] Add .and.popc variant of the b1 MMA instruction.

tra created this revision. tra added a reviewer: steffenlarsen. Herald added subscribers: bixia, hiraditya, yaxunl, jholewinski. tra requested review of this revision. Herald added a subscriber: jdoerfert. Herald added projects: clang, LLVM. Extends the changes in D104847

[PATCH] D105384: [NVPTX, CUDA] Add .and.popc variant of the b1 MMA instruction.

tra updated this revision to Diff 358041. tra marked an inline comment as done. tra edited the summary of this revision. tra added a comment. Addressed review comments. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D105384/new/ https://reviews.llvm.

[PATCH] D105384: [NVPTX, CUDA] Add .and.popc variant of the b1 MMA instruction.

tra added inline comments. Comment at: clang/test/CodeGen/builtins-nvptx-mma.py:35-38 +def make_mma_ops(geoms, types_a, types_b, types_c, types_d, b1ops=None): ops = [] + if b1ops is None: +b1ops = [""] steffenlarsen wrote: > Default initializers that us

[PATCH] D105295: [CUDA] Only allow NVIDIA offload-arch during CUDA compilation.

This revision was landed with ongoing or failed builds. This revision was automatically updated to reflect the committed changes. Closed by commit rG01d3a3dcabaf: [CUDA] Only allow NVIDIA offload-arch during CUDA compilation. (authored by tra). Repository: rG LLVM Github Monorepo CHANGES SINCE

[PATCH] D105384: [NVPTX, CUDA] Add .and.popc variant of the b1 MMA instruction.

tra updated this revision to Diff 358365. tra added a comment. Updated LD/ST generation. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D105384/new/ https://reviews.llvm.org/D105384 Files: clang/include/clang/Basic/BuiltinsNVPTX.def clang/lib/Co

[PATCH] D105295: [CUDA] Only allow NVIDIA offload-arch during CUDA compilation.

tra added a comment. Ugh. I broke the cuda-bad-arch.cu test. Comment at: clang/test/Driver/cuda-bad-arch.cu:30 // RUN: | FileCheck -check-prefix OK %s // RUN: %clang -### -target x86_64-linux-gnu --cuda-gpu-arch=gfx90a -c %s 2>&1 \ // RUN: | FileCheck -check-prefix OK %s --

[PATCH] D105295: [CUDA] Only allow NVIDIA offload-arch during CUDA compilation.

tra added a comment. In D105295#2874935 , @tra wrote: > Ugh. I broke the cuda-bad-arch.cu test. Should be fixed in 25629bb45f0a4b8c8e99dbde4f4a7e3d980b9fd7 Repository: rG LLVM Git

[PATCH] D105384: [NVPTX, CUDA] Add .and.popc variant of the b1 MMA instruction.

tra marked an inline comment as done. tra added inline comments. Comment at: clang/test/CodeGen/builtins-nvptx-mma.py:84 + # It uses __mma_tf32_m16n16k8_ld_c but __mma_m16n16k8_st_c_f32. + make_ldst_ops(["m16n16k8"], ["a", "b", "c", "d"], ["tf32", "f32"])) ---

[PATCH] D112284: [Clang][NFC] Clang CUDA codegen clean-up

tra added a comment. In D112284#3086499 , @bondhugula wrote: > @tra While on this, I also wanted to ask as to why clang cuda codegen is > using an argument on the global ctor and the dtor it's generating. It's a good question, and I don't have a good a

[PATCH] D112492: [HIP] Do not use kernel handle for MSVC target

tra added a comment. As phrased the summary would likely be rather confusing for anyone other than you and me. > Currently Visual Studio 2019 has a linker issue which causes linking error > when a template kernel is instantiated in different compilation units. It's not clear what exactly is the

[PATCH] D112041: [InferAddressSpaces] Support assumed addrspaces from addrspace predicates.

tra added a comment. LGTM in general, modulo remaining nits. Comment at: llvm/lib/Transforms/Scalar/InferAddressSpaces.cpp:196 void inferAddressSpaces(ArrayRef Postorder, - ValueToAddrSpaceMapTy *InferredAddrSpace) const; + V

[PATCH] D112492: [HIP] Do not use kernel handle for MSVC target

tra added a comment. > With these changes, we should have consistent name mangling for kernel stubs > and kernel launching mechanism on Linux and Windows. Nice! Thank you for figuring out the root causes. CHANGES SINCE LAST ACTION https://reviews.llvm.org/D112492/new/ https://reviews.llvm.o

[PATCH] D113249: [CUDA] Bump CUDA version to 11.5

tra requested changes to this revision. tra added a comment. This revision now requires changes to proceed. I think we're missing few more changes here: - The driver needs to enable ptx75 when it constructs cc1 command line in clang/lib/Driver/ToolChains/Cuda.cpp - We also need to handle PTX75 i

[PATCH] D113249: [CUDA] Bump CUDA version to 11.5

tra added a comment. In D113249#3112279 , @Hahnfeld wrote: > Experimental support for `__int128` is new in CUDA 11.5, not sure if Clang > enables this for CUDA. I think we've added support for i128 a while back: https://godbolt.org/z/18bEbhMYb > The r

[PATCH] D111443: [Driver] Fix ToolChain::getSanitizerArgs

tra added a comment. I'll defer to @eugenis. Overall it looks OK to be. CHANGES SINCE LAST ACTION https://reviews.llvm.org/D111443/new/ https://reviews.llvm.org/D111443 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/

[PATCH] D113249: [CUDA] Bump CUDA version to 11.5

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. In D113249#3113666 , @carlosgalvezp wrote: >> - The driver needs to enable ptx75 when it constructs cc1 command line in >> clang/lib/Driver/ToolChains/Cuda

[PATCH] D113491: [HIP] Fix device stub name for Windows

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. LGTM in general. Comment at: clang/lib/AST/MicrosoftMangle.cpp:975-976 + llvm::SmallString<128> Buf; + mangleSourceName((llvm::Twine("__device_stub__") + II->getN

[PATCH] D113490: [NFC] Let Microsoft mangler accept GlobalDecl

tra added a reviewer: rnk. tra added a subscriber: rnk. tra added a comment. + @rnk as it's a windows-specific change. LGTM in general. CHANGES SINCE LAST ACTION https://reviews.llvm.org/D113490/new/ https://reviews.llvm.org/D113490 ___ cfe-commit

[PATCH] D112492: [CUDA][HIP] Allow comdat for kernels

tra added subscribers: kpyzhov, rnk. tra added inline comments. Comment at: clang/lib/CodeGen/CodeGenModule.cpp:4290-4293 - // Do not set COMDAT attribute for CUDA/HIP stub functions to prevent - // them being "merged" by the COMDAT Folding linker optimization. - if (D.hasAttr

[PATCH] D112492: [CUDA][HIP] Allow comdat for kernels

tra added a comment. Yes, we do need to merge identical functions with **identical names** for templates. The comdat-folding issue is different. IIUIC, it allows merging two functions with identical code and **different names**, into one function with two names. That will break CUDA as we do n

[PATCH] D86376: [HIP] Emit kernel symbol

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. So, to summarize how the patch changes the under-the-hood kernel launch machinery: - device-side is unchanged. Kernel function is generated with the real kernel name - host-side stub is still gener

[PATCH] D97708: [CUDA] Remove `noreturn` attribute from __assertfail().

tra created this revision. tra added a reviewer: jlebar. Herald added subscribers: bixia, yaxunl. tra requested review of this revision. Herald added a project: clang. `noreturn` complicates control flow and tends to trigger a known bug in ptxas if the assert is used within loops in sufficiently c

[PATCH] D97708: [CUDA] Remove `noreturn` attribute from __assertfail().

tra updated this revision to Diff 327197. tra edited the summary of this revision. tra added a comment. Added a comment. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D97708/new/ https://reviews.llvm.org/D97708 Files: clang/lib/Headers/__clang_cu

[PATCH] D97340: [HIP] Support Spack packages

tra added inline comments. Comment at: clang/include/clang/Driver/Options.td:3535-3536 HelpText<"Print the registered targets">; +def print_rocm_search_dirs : Flag<["-", "--"], "print-rocm-search-dirs">, + HelpText<"Print the paths used for finding ROCm installation">; def p

[PATCH] D97708: [CUDA] Remove `noreturn` attribute from __assertfail().

This revision was landed with ongoing or failed builds. This revision was automatically updated to reflect the committed changes. Closed by commit rG32e064527623: [CUDA] Remove `noreturn` attribute from __assertfail(). (authored by tra). Repository: rG LLVM Github Monorepo CHANGES SINCE LAST A

[PATCH] D97340: [HIP] Support Spack packages

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. LGTM modulo couple of nits. Comment at: clang/include/clang/Driver/Options.td:3535-3536 HelpText<"Print the registered targets">; +def print_rocm_search_dirs : Flag<["-", "--"],

[PATCH] D98068: Remove asserts for LocalInstantiationScope

tra added a comment. Godbolt appears to be OK with the code for both gcc and clang: https://godbolt.org/z/enec44 Debug build does assert here: clang++: /work/llvm/repo/clang/lib/Sema/SemaTemplateInstantiate.cpp:3630: void clang::LocalInstantiationScope::InstantiatedLocal(const clang::Decl *,

[PATCH] D98193: [CUDA][HIP] Allow non-ODR use of host var in device

tra added inline comments. Comment at: clang/test/SemaCUDA/device-use-host-var.cu:41 *out = global_const_var; + *out = global_const_struct_var.x; I do not think it should be allowed. We end up instantiating the variable on device, even though the variable

[PATCH] D98193: [CUDA][HIP] Allow non-ODR use of host var in device

tra added a reviewer: rsmith. tra added a subscriber: rsmith. tra added a comment. LGTM I've added @rsmith to double check that we're handling it correctly. CHANGES SINCE LAST ACTION https://reviews.llvm.org/D98193/new/ https://reviews.llvm.org/D98193 ___

[PATCH] D107054: [Clang][CUDA] Add descriptors, mappings, and features for missing CUDA and PTX versions

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. I think this patch has been obsoleted by https://reviews.llvm.org/D113249 which has already landed. My apologies for letting the patch slip through the cracks. CHANGES SINCE LAST ACTION https://

[PATCH] D114326: Update the list of CUDA versions up to 11.5

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:131 + std::initializer_list Versions = { + "11.5", "11.4", "11.3", "11.2", "11.1", "11.0", "10.2", "10.1", + "10.0", "9.2", "9.1", "9.0", "8.0", "7.5", "7.0"}; mojca wrote

[PATCH] D110618: [HIPSPV][2/4] Add HIPSPV tool chain

tra accepted this revision. tra added a comment. This revision is now accepted and ready to land. LGTM in general, modulo push_back/append nits. Comment at: clang/include/clang/Driver/Options.td:3701 " do not include the default CUDA/HIP wrapper headers">; +def nohipwrapperi

[PATCH] D110549: [HIPSPV][1/4] Refactor HIP tool chain

tra added a subscriber: echristo. tra added inline comments. Comment at: clang/lib/Driver/ToolChains/HIPUtility.cpp:119-133 + // Add MC directives to embed target binaries. We ensure that each + // section and image is 16-byte aligned. This is not mandatory, but + // increases

[PATCH] D114326: Update the list of CUDA versions up to 11.5

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:131 + std::initializer_list Versions = { + "11.5", "11.4", "11.3", "11.2", "11.1", "11.0", "10.2", "10.1", + "10.0", "9.2", "9.1", "9.0", "8.0", "7.5", "7.0"}; tra wrote:

[PATCH] D114326: Update the list of CUDA versions up to 11.5

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:131 + std::initializer_list Versions = { + "11.5", "11.4", "11.3", "11.2", "11.1", "11.0", "10.2", "10.1", + "10.0", "9.2", "9.1", "9.0", "8.0", "7.5", "7.0"}; carlosgalve

[PATCH] D114326: Update the list of CUDA versions up to 11.5

tra added a comment. In D114326#3154228 , @mojca wrote: > Somewhat off-topic from a discussion earlier in the thread. > What's the purpose of the following code then if users are supposed to > explicitly specify the `-L` flag anyway? Good point, it is i

[PATCH] D114601: Read path to CUDA from env. variable CUDA_PATH on Windows

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:137 } else if (HostTriple.isOSWindows()) { -for (const char *Ver : Versions) - Candidates.emplace_back( Do we want to keep this as the fall-back for cases when `CUDA_PATH` is

[PATCH] D114326: Update the list of CUDA versions up to 11.5

tra added a comment. In D114326#3159122 , @mojca wrote: > @tra: this is not yet 100% ready since the unit tests are now failing > (expecting to find CUDA 8.0). > I can fix the unit test, but I suppose that someone needs to install > additional SDK somew

[PATCH] D114601: Read path to CUDA from env. variable CUDA_PATH on Windows

tra added inline comments. Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:137 } else if (HostTriple.isOSWindows()) { -for (const char *Ver : Versions) - Candidates.emplace_back( mojca wrote: > tra wrote: > > Do we want to keep this as the fall-back

[PATCH] D114326: Update the list of CUDA versions up to 11.5

tra requested changes to this revision. tra added a comment. This revision now requires changes to proceed. With D114601 , this patch would no longer be needed. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D114326/

[PATCH] D114812: [HIP] Add pre-defined macro `__HIPCC_RDC__`

tra added a comment. In D114812#3162282 , @yaxunl wrote: > I am not sure whether we want to define a similar macro for cuda-clang. > > Maybe `__CLANG_RDC__` is better? I think it would make sense. For CUDA compatibility we can the define __CUDACC_RDC__

[PATCH] D136311: [CUDA] Propagate __bf16 type info from the host compilation.

tra created this revision. Herald added subscribers: mattd, gchakrabarti, asavonic, bixia, yaxunl. Herald added a project: All. tra updated this revision to Diff 469453. tra added a comment. Herald added a subscriber: hiraditya. tra updated this revision to Diff 469460. tra updated this revision to

[PATCH] D136311: [CUDA] Propagate __bf16 type info from the host compilation.

tra updated this revision to Diff 469663. tra added a comment. Make __bf16 available regradless of its availability on the host. Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D136311/new/ https://reviews.llvm.org/D136311 Files: clang/lib/Basic/Ta

<    6   7   8   9   10   11   12   13   14   15   >