[CUDA] Improve clang's ability to detect recent CUDA versions.

[tra]

Summary

CUDA-11.1 does not carry version.txt which causes clang to assume that it's
CUDA-7.0, which used to be the only CUDA version w/o version.txt.

In order to tell CUDA-7.0 apart from the new versions, clang now probes for the
presence of libdevice.10.bc which is not present in the old CUDA versions.

This should keep Clang working for CUDA-11.1.

Bug: https://bugs.llvm.org/show_bug.cgi?id=47332

[tra]

Added a test.

[yaxunl]

LGTM. Thanks.

[emankov]

clang/lib/Driver/ToolChains/Cuda.cpp

158

CUDA 11.0+, actually

161

Do we have any other mechanism besides version.txt for determining an exact CUDA version? Setting the latest version in case of absence of version.txt doesn't suit all the needs: sometimes the exact version is taken into account, for instance in hipify-clang.

clang/test/Driver/cuda-version-check.cu

13

CUDA-11.0 Update 1 doesn't carry version.txt as well.

72

missing space before {{

[tra]

addresses review comments.

clang/lib/Driver/ToolChains/Cuda.cpp

161

Not easily.

We could try running one of SDK binaries with --version. This would be fragile as the tool version does not necessarily match the SDK's and NVIDIA has already started versioning elements per-component. E.g. some shared libraries in 11.1 are already versioned as 10.2, 11.0, 11.1 and 11.2. There are also situations when we had to cherry-pick a tool from a different release in order to work around a critical bug. We don't want to change compiler's idea of CUDA version based on that.

We could parse CUDA headers and try finding CUDA_VERSION macro. That's feasible. It may be somewhat fragile if we just search for a text string '#define CUDA_VERSION XXXX' -- nvidia may change it. On the other hand it's not that much worse than relying on version.txt. It also does not carry complete version, only major.minor, so we will not be able to tell 11.0 apart from 11.0 update 1. Probably not a big deal. I don't think we've needed it so far.

We could do something like this:

• if version.txt is there -- use it
• otherwise if cuda.h is found, extract version from CUDA_VERSION macro
• if that failed, use libdevice bitcode format to detect CUDA-7.0
• fall back to "last supported version" otherwise

[tra]

clang/test/Driver/cuda-version-check.cu

13

CUDA-11.0 Update 1 doesn't carry version.txt as well.

This assertion appears to be incorrect. I've just installed 11.0 update1 using Ubuntu .run installer I've got from https://developer.nvidia.com/cuda-11.0-update1-download-archive and the installed version does have version.txt saying

CUDA Version 11.0.228

Did you use some other installer variant?

[emankov]

Thank you!
I've successfully applied your change in Cuda.cpp to the following clang versions: 10.0.0, 10.0.1, 11.0.0, and 12.0.0git; and have created the working patches (#206).

clang/test/Driver/cuda-version-check.cu

13

"c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0\bin\nvcc" --version                                                                                                             nvcc: NVIDIA (R) Cuda compiler driver                                                                                                                                                                               Copyright (c) 2005-2020 NVIDIA Corporation                                                                                                                                                                          Built on Wed_Jul_22_19:09:35_Pacific_Daylight_Time_2020                                                                                                                                                             Cuda compilation tools, release 11.0, V11.0.221                                                                                                                                                                     Build cuda_11.0_bu.relgpu_drvr445TC445_37.28845127_0

CUDA Version 11.0.221 (distributive cuda_11.0.3_451.82_win10.exe) doesn't have version.txt as well.

[tra]

In D89752#2344113, @emankov wrote:

Thank you!
I've successfully applied your change in Cuda.cpp to the following clang versions: 10.0.0, 10.0.1, 11.0.0, and 12.0.0git; and have created the working patches (#206).

Thank you. Can you also take a look at D89832 which extracts CUDA version from .h file, when version.txt is not available?

clang/test/Driver/cuda-version-check.cu

13

Oh, Windows... That explains the disparity.
Thank you for testing the patch there. Could you also give D89832 a try?

[emankov]

clang/lib/Driver/ToolChains/Cuda.cpp

161

I'd appreciate the appearance of the step with a version extraction from CUDA_VERSION macro.

[tra]

clang/lib/Driver/ToolChains/Cuda.cpp

161

D89832 does exactly that. PTAL.

[emankov]

clang/lib/Driver/ToolChains/Cuda.cpp

161

Ok, I'll try it.

clang/lib/Driver/ToolChains/Cuda.cpp

for (const auto &Candidate : Candidates) {
     else if (FS.exists(InstallPath + "/lib"))
       LibPath = InstallPath + "/lib";
     else
       continue;
 
     llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> VersionFile =
         FS.getBufferForFile(InstallPath + "/version.txt");
     if (!VersionFile) {
-      // CUDA 7.0 doesn't have a version.txt, so guess that's our version if
-      // version.txt isn't present.
+      // CUDA 7.0 and CUDA 11.1+ do not have version.txt file.
+      // Use libdevice file to distinguish 7.0 from the new versions.
+      if (FS.exists(LibDevicePath + "/libdevice.10.bc")) {
+        Version = CudaVersion::LATEST;
+        DetectedVersionIsNotSupported = Version > CudaVersion::LATEST_SUPPORTED;
+      } else {
         Version = CudaVersion::CUDA_70;
+      }
     } else {
       ParseCudaVersionFile((*VersionFile)->getBuffer());
     }
 
     if (Version >= CudaVersion::CUDA_90) {
       // CUDA-9+ uses single libdevice file for all GPU variants.
       std::string FilePath = LibDevicePath + "/libdevice.10.bc";
       if (FS.exists(FilePath)) {

clang/test/Driver/cuda-version-check.cu

 // REQUIRES: clang-driver
 // REQUIRES: x86-registered-target
 // REQUIRES: nvptx-registered-target
 
 // RUN: %clang --target=x86_64-linux -v -### --cuda-gpu-arch=sm_20 --cuda-path=%S/Inputs/CUDA/usr/local/cuda 2>&1 %s | \
 // RUN:    FileCheck %s --check-prefix=OK
 // RUN: %clang --target=x86_64-linux -v -### --cuda-gpu-arch=sm_20 --cuda-path=%S/Inputs/CUDA_80/usr/local/cuda 2>&1 %s | \
 // RUN:    FileCheck %s --check-prefix=OK
 // RUN: %clang --target=x86_64-linux -v -### --cuda-gpu-arch=sm_60 --cuda-path=%S/Inputs/CUDA_80/usr/local/cuda 2>&1 %s | \
 // RUN:    FileCheck %s --check-prefix=OK
 // RUN: %clang --target=x86_64-linux -v -### --cuda-gpu-arch=sm_60 --cuda-path=%S/Inputs/CUDA-unknown/usr/local/cuda 2>&1 %s | \
 // RUN:    FileCheck %s --check-prefix=UNKNOWN_VERSION
+// CUDA-11.1 does not carry version.txt file. Make sure we still detect it as a
+// new version and handle it the same as we handle other new CUDA versions.
+// RUN: %clang --target=x86_64-linux -v -### --cuda-gpu-arch=sm_60 --cuda-path=%S/Inputs/CUDA_111/usr/local/cuda 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=UNKNOWN_VERSION
 // Make sure that we don't warn about CUDA version during C++ compilation.
 // RUN: %clang --target=x86_64-linux -v -### -x c++ --cuda-gpu-arch=sm_60 \
 // RUN:    --cuda-path=%S/Inputs/CUDA-unknown/usr/local/cuda 2>&1 %s | \
 // RUN:    FileCheck %s --check-prefix=UNKNOWN_VERSION_CXX
 
 // The installation at Inputs/CUDA is CUDA 7.0, which doesn't support sm_60.
 // RUN: %clang --target=x86_64-linux -v -### --cuda-gpu-arch=sm_60 --cuda-path=%S/Inputs/CUDA/usr/local/cuda 2>&1 %s | \
 // RUN:    FileCheck %s --check-prefix=ERR_SM60

 
 // We should only get one error per architecture.
 // ERR_SM60: error: GPU arch sm_60 {{.*}}
 // ERR_SM60-NOT: error: GPU arch sm_60
 
 // ERR_SM61: error: GPU arch sm_61 {{.*}}
 // ERR_SM61-NOT: error: GPU arch sm_61
 
-// UNKNOWN_VERSION: Unknown CUDA version 999.999. Assuming the latest supported version
+// UNKNOWN_VERSION: Unknown CUDA version{{.*}}. Assuming the latest supported version
 // UNKNOWN_VERSION_CXX-NOT: Unknown CUDA version