[CUDA] Check that our CUDA install supports the requested architectures.

[jlebar]

Summary

Raise an error if you're using a CUDA installation that's too old for
the requested architectures. In practice, this means that you need a
CUDA 8 install to compile for sm_6*.

[jlebar]

Friendly ping.

[tra]

include/clang/Basic/DiagnosticDriverKinds.td

32 ↗

(On Diff #62301)

Is it supposed to be two dashes or one?

include/clang/Driver/Options.td

1722–1724 ↗

(On Diff #62301)

IMO we should use double-dash for this option.

-nocudalib/-nocudainc were mimicking existing -nostdlib/-nostdinc and thus ended with single dash in front.
--nocuda-version-check is not constrained by this.

lib/Driver/ToolChains.cpp

1704 ↗

(On Diff #62301)

Might add an example of what's in that file, so it's easier to understand what the code below is doing.

1793 ↗

(On Diff #62301)

What if it's cuda-8, but we've failed to read the file due to permissions.
Perhaps we should differentiate no-such-file from other errors.

4711 ↗

(On Diff #62301)

-nocudainc should not preclude CUDA version check, IMO.
Primary use of version check is to make sure we can compile for a given GPU architecture. I think we still want to do that even if -nocudainc is passed.

[jlebar]

Address review comments.

include/clang/Driver/Options.td

1722–1724 ↗

(On Diff #62301)

I guess I have to do a double dash because --no-cuda-noopt-device-debug is thus spelled. Notice there's also a dash after "no".

...this is so screwed up.

lib/Driver/ToolChains.cpp

1793 ↗

(On Diff #62301)

Urgh.

I feel like simpler is probably better. Like, I would rather say:

• We consider /usr/local/cuda, /usr/local/cuda-8.0, ...
• We pick the first one that has a few key files
• We infer its version by looking for version.txt. If we can't read it, for whatever reason, we assume 7.0.

Than the more complicated

• We infer its version by looking for version.txt. If it's present but we can't read it, we try to look at the file path (/usr/local/cuda-8.0/...).

Especially because as we make things more complicated, our next step will probably be to change the algorithm to something like:

• For each directory D in /usr/local/cuda, /usr/local/cuda-8.0, ...
  • Infer D's version using the algorithm above.
  • If D is not compatible with all of the cuda_archs specified, break and try the next one...

I do agree that if someone hits a permission problem on version.txt, that would be very difficult for them to debug. I suppose we could raise a proper error if version.txt is present but not readable? I'm not sure it's worth doing that, though...

4711 ↗

(On Diff #62301)

We also do a check if you invoke ptxas. So the only circumstances under which we don't do a check is -nocudainc and no invocation of ptxas. In which case the only thing we're using from the CUDA install is fatbinary, which doesn't care about the sm version (afaik). We could add a version check on the fatbin invocation too if you want. My goal was just that if we don't use anything from the CUDA install, we should not check its version.

[tra]

LGTM.

lib/Driver/ToolChains.cpp

1798 ↗

(On Diff #63094)

Good point. It may be worth a separate patch. Falling back to 7.0 on any error is OK for now.

4727 ↗

(On Diff #63094)

OK.

cfe/trunk/include/clang/Basic/DiagnosticDriverKinds.td

def err_drv_unknown_stdin_type_clang_cl : Error<
   "use /Tc or /Tp to set input type for standard input">;
 def err_drv_unknown_language : Error<"language not recognized: '%0'">;
 def err_drv_invalid_arch_name : Error<
   "invalid arch name '%0'">;
 def err_drv_cuda_bad_gpu_arch : Error<"Unsupported CUDA gpu architecture: %0">;
 def err_drv_no_cuda_installation : Error<
   "cannot find CUDA installation.  Provide its path via --cuda-path, or pass "
   "-nocudainc to build without CUDA includes.">;
+def err_drv_cuda_version_too_low : Error<
+  "GPU arch %1 requires CUDA version at least %3, but installation at %0 is %2. "
+  "Use --cuda-path to specify a different CUDA install, or pass "
+  "--no-cuda-version-check.">;
 def err_drv_invalid_thread_model_for_target : Error<
   "invalid thread model '%0' in '%1' for this target">;
 def err_drv_invalid_linker_name : Error<
   "invalid linker name in argument '%0'">;
 def err_drv_invalid_pgo_instrumentor : Error<
   "invalid PGO instrumentor in argument '%0'">;
 def err_drv_invalid_rtlib_name : Error<
   "invalid runtime library name in argument '%0'">;

cfe/trunk/include/clang/Driver/Options.td

HelpText<"Compile CUDA code for host only.  Has no effect on non-CUDA "
            "compilations.">;
 def cuda_compile_host_device : Flag<["--"], "cuda-compile-host-device">,
   HelpText<"Compile CUDA code for both host and device (default).  Has no "
            "effect on non-CUDA compilations.">;
 def cuda_gpu_arch_EQ : Joined<["--"], "cuda-gpu-arch=">, Flags<[DriverOption]>,
   HelpText<"CUDA GPU architecture (e.g. sm_35).  May be specified more than once.">;
 def cuda_noopt_device_debug : Flag<["--"], "cuda-noopt-device-debug">,
   HelpText<"Enable device-side debug info generation. Disables ptxas optimizations.">;
+def no_cuda_version_check : Flag<["--"], "no-cuda-version-check">,
+  HelpText<"Don't error out if the detected version of the CUDA install is "
+           "too low for the requested CUDA gpu architecture.">;
 def no_cuda_noopt_device_debug : Flag<["--"], "no-cuda-noopt-device-debug">;
 def cuda_path_EQ : Joined<["--"], "cuda-path=">, Group<i_Group>,
   HelpText<"CUDA installation path">;
 def fcuda_flush_denormals_to_zero : Flag<["-"], "fcuda-flush-denormals-to-zero">,
   Flags<[CC1Option]>, HelpText<"Flush denormal floating point values to zero in CUDA device mode.">;
 def fno_cuda_flush_denormals_to_zero : Flag<["-"], "fno-cuda-flush-denormals-to-zero">;
 def fcuda_approx_transcendentals : Flag<["-"], "fcuda-approx-transcendentals">,
   Flags<[CC1Option]>, HelpText<"Use approximate transcendental functions">;

cfe/trunk/lib/Driver/ToolChains.h

 //===--- ToolChains.h - ToolChain Implementations ---------------*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_H
 #define LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_H
 
 #include "Tools.h"
+#include "clang/Basic/Cuda.h"
 #include "clang/Basic/VersionTuple.h"
 #include "clang/Driver/Action.h"
 #include "clang/Driver/Multilib.h"
 #include "clang/Driver/ToolChain.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/Optional.h"
+#include "llvm/ADT/SmallSet.h"
 #include "llvm/Support/Compiler.h"
 #include <set>
 #include <vector>
 
 namespace clang {
 namespace driver {
 namespace toolchains {
 

 protected:
   GCCInstallationDetector GCCInstallation;
 
   // \brief A class to find a viable CUDA installation
   class CudaInstallationDetector {
   private:
     const Driver &D;
     bool IsValid = false;
+    CudaVersion Version = CudaVersion::UNKNOWN;
     std::string InstallPath;
     std::string BinPath;
     std::string LibPath;
     std::string LibDevicePath;
     std::string IncludePath;
     llvm::StringMap<std::string> LibDeviceMap;
 
+    // CUDA architectures for which we have raised an error in
+    // CheckCudaVersionSupportsArch.
+    mutable llvm::SmallSet<CudaArch, 4> ArchsWithVersionTooLowErrors;
+
   public:
     CudaInstallationDetector(const Driver &D) : D(D) {}
     void init(const llvm::Triple &TargetTriple, const llvm::opt::ArgList &Args);
 
+    /// \brief Emit an error if Version does not support the given Arch.
+    ///
+    /// If either Version or Arch is unknown, does not emit an error.  Emits at
+    /// most one error per Arch.
+    void CheckCudaVersionSupportsArch(CudaArch Arch) const;
+
     /// \brief Check whether we detected a valid Cuda install.
     bool isValid() const { return IsValid; }
     /// \brief Print information about the detected CUDA installation.
     void print(raw_ostream &OS) const;
 
+    /// \brief Get the deteced Cuda install's version.
+    CudaVersion version() const { return Version; }
     /// \brief Get the detected Cuda installation path.
     StringRef getInstallPath() const { return InstallPath; }
     /// \brief Get the detected path to Cuda's bin directory.
     StringRef getBinPath() const { return BinPath; }
     /// \brief Get the detected Cuda Include path.
     StringRef getIncludePath() const { return IncludePath; }
     /// \brief Get the detected Cuda library path.
     StringRef getLibPath() const { return LibPath; }
TranslateArgs(const llvm::opt::DerivedArgList &Args,
                 const char *BoundArch) const override;
   void addClangTargetOptions(const llvm::opt::ArgList &DriverArgs,
                              llvm::opt::ArgStringList &CC1Args) const override;
 
   // Never try to use the integrated assembler with CUDA; always fork out to
   // ptxas.
   bool useIntegratedAs() const override { return false; }
 
+  void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs,
+                          llvm::opt::ArgStringList &CC1Args) const override;
+
+  const Generic_GCC::CudaInstallationDetector &cudaInstallation() const {
+    return CudaInstallation;
+  }
+  Generic_GCC::CudaInstallationDetector &cudaInstallation() {
+    return CudaInstallation;
+  }
+
 protected:
   Tool *buildAssembler() const override;  // ptxas
   Tool *buildLinker() const override;     // fatbinary (ok, not really a linker)
 };
 
 class LLVM_LIBRARY_VISIBILITY MipsLLVMToolChain : public Linux {
 protected:
   Tool *buildLinker() const override;

cfe/trunk/lib/Driver/ToolChains.cpp

 //===--- ToolChains.cpp - ToolChain Implementations -------------*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 
 #include "ToolChains.h"
+#include "clang/Basic/Cuda.h"
 #include "clang/Basic/ObjCRuntime.h"
 #include "clang/Basic/Version.h"
 #include "clang/Basic/VirtualFileSystem.h"
 #include "clang/Config/config.h" // for GCC_INSTALL_PREFIX
 #include "clang/Driver/Compilation.h"
 #include "clang/Driver/Driver.h"
 #include "clang/Driver/DriverDiagnostic.h"
 #include "clang/Driver/Options.h"
/*static*/ void Generic_GCC::GCCInstallationDetector::CollectLibDirsAndTriples(
   // match any of our aliases.
   TripleAliases.push_back(TargetTriple.str());
 
   // Also include the multiarch variant if it's different.
   if (TargetTriple.str() != BiarchTriple.str())
     BiarchTripleAliases.push_back(BiarchTriple.str());
 }
 
+// Parses the contents of version.txt in an CUDA installation.  It should
+// contain one line of the from e.g. "CUDA Version 7.5.2".
+static CudaVersion ParseCudaVersionFile(llvm::StringRef V) {
+  if (!V.startswith("CUDA Version "))
+    return CudaVersion::UNKNOWN;
+  V = V.substr(strlen("CUDA Version "));
+  int Major = -1, Minor = -1;
+  auto First = V.split('.');
+  auto Second = First.second.split('.');
+  if (!First.first.getAsInteger(10, Major) ||
+      !Second.first.getAsInteger(10, Minor))
+    return CudaVersion::UNKNOWN;
+
+  if (Major == 7 && Minor == 0) {
+    // This doesn't appear to ever happen -- version.txt doesn't exist in the
+    // CUDA 7 installs I've seen.  But no harm in checking.
+    return CudaVersion::CUDA_70;
+  }
+  if (Major == 7 && Minor == 5)
+    return CudaVersion::CUDA_75;
+  if (Major == 8 && Minor == 0)
+    return CudaVersion::CUDA_80;
+  return CudaVersion::UNKNOWN;
+}
+
 // \brief -- try common CUDA installation paths looking for files we need for
 // CUDA compilation.
-
 void Generic_GCC::CudaInstallationDetector::init(
     const llvm::Triple &TargetTriple, const llvm::opt::ArgList &Args) {
   SmallVector<std::string, 4> CudaPathCandidates;
 
   if (Args.hasArg(options::OPT_cuda_path_EQ))
     CudaPathCandidates.push_back(
         Args.getLastArgValue(options::OPT_cuda_path_EQ));
   else {
for (llvm::sys::fs::directory_iterator LI(LibDevicePath, EC), LE;
         LibDeviceMap["sm_52"] = FilePath;
         LibDeviceMap["sm_53"] = FilePath;
         LibDeviceMap["sm_60"] = FilePath;
         LibDeviceMap["sm_61"] = FilePath;
         LibDeviceMap["sm_62"] = FilePath;
       }
     }
 
+    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.
+      Version = CudaVersion::CUDA_70;
+    } else {
+      Version = ParseCudaVersionFile((*VersionFile)->getBuffer());
+    }
+
     IsValid = true;
     break;
   }
 }
 
+void Generic_GCC::CudaInstallationDetector::CheckCudaVersionSupportsArch(
+    CudaArch Arch) const {
+  if (Arch == CudaArch::UNKNOWN || Version == CudaVersion::UNKNOWN ||
+      ArchsWithVersionTooLowErrors.count(Arch) > 0)
+    return;
+
+  auto RequiredVersion = MinVersionForCudaArch(Arch);
+  if (Version < RequiredVersion) {
+    ArchsWithVersionTooLowErrors.insert(Arch);
+    D.Diag(diag::err_drv_cuda_version_too_low)
+        << InstallPath << CudaArchToString(Arch) << CudaVersionToString(Version)
+        << CudaVersionToString(RequiredVersion);
+  }
+}
+
 void Generic_GCC::CudaInstallationDetector::print(raw_ostream &OS) const {
   if (isValid())
-    OS << "Found CUDA installation: " << InstallPath << "\n";
+    OS << "Found CUDA installation: " << InstallPath << ", version "
+       << CudaVersionToString(Version) << "\n";
 }
 
 namespace {
 // Filter to remove Multilibs that don't exist as a suffix to Path
 class FilterNonExistent {
   StringRef Base, File;
   vfs::FileSystem &VFS;
 
if (!LibDeviceFile.empty()) {
     // Libdevice in CUDA-7.0 requires PTX version that's more recent
     // than LLVM defaults to. Use PTX4.2 which is the PTX version that
     // came with CUDA-7.0.
     CC1Args.push_back("-target-feature");
     CC1Args.push_back("+ptx42");
   }
 }
 
+void CudaToolChain::AddCudaIncludeArgs(const ArgList &DriverArgs,
+                                       ArgStringList &CC1Args) const {
+  // Check our CUDA version if we're going to include the CUDA headers.
+  if (!DriverArgs.hasArg(options::OPT_nocudainc) &&
+      !DriverArgs.hasArg(options::OPT_nocuda_version_check)) {
+    StringRef Arch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
+    assert(!Arch.empty() && "Must have an explicit GPU arch.");
+    CudaInstallation.CheckCudaVersionSupportsArch(StringToCudaArch(Arch));
+  }
+  Linux::AddCudaIncludeArgs(DriverArgs, CC1Args);
+}
+
 llvm::opt::DerivedArgList *
 CudaToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
                              const char *BoundArch) const {
   DerivedArgList *DAL = new DerivedArgList(Args.getBaseArgs());
   const OptTable &Opts = getDriver().getOpts();
 
   for (Arg *A : Args) {
     if (A->getOption().matches(options::OPT_Xarch__)) {

cfe/trunk/lib/Driver/Tools.cpp

const auto &TC =
       static_cast<const toolchains::CudaToolChain &>(getToolChain());
   assert(TC.getTriple().isNVPTX() && "Wrong platform");
 
   std::vector<std::string> gpu_archs =
       Args.getAllArgValues(options::OPT_march_EQ);
   assert(gpu_archs.size() == 1 && "Exactly one GPU Arch required for ptxas.");
   const std::string& gpu_arch = gpu_archs[0];
 
+  // Check that our installation's ptxas supports gpu_arch.
+  if (!Args.hasArg(options::OPT_nocuda_version_check)) {
+    TC.cudaInstallation().CheckCudaVersionSupportsArch(
+        StringToCudaArch(gpu_arch));
+  }
+
   ArgStringList CmdArgs;
   CmdArgs.push_back(TC.getTriple().isArch64Bit() ? "-m64" : "-m32");
   if (Args.hasFlag(options::OPT_cuda_noopt_device_debug,
                    options::OPT_no_cuda_noopt_device_debug, false)) {
     // ptxas does not accept -g option if optimization is enabled, so
     // we ignore the compiler's -O* options if we want debug info.
     CmdArgs.push_back("-g");
     CmdArgs.push_back("--dont-merge-basicblocks");

cfe/trunk/test/Driver/Inputs/CUDA_80/usr/local/cuda/version.txt

+CUDA Version 8.0.42

cfe/trunk/test/Driver/cuda-version-check.cu

+// REQUIRES: clang-driver
+// REQUIRES: x86-registered-target
+// REQUIRES: nvptx-registered-target
+
+// RUN: %clang -v -### --cuda-gpu-arch=sm_20 --sysroot=%S/Inputs/CUDA 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=OK
+// RUN: %clang -v -### --cuda-gpu-arch=sm_20 --sysroot=%S/Inputs/CUDA_80 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=OK
+// RUN: %clang -v -### --cuda-gpu-arch=sm_60 --sysroot=%S/Inputs/CUDA_80 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=OK
+
+// The installation at Inputs/CUDA is CUDA 7.0, which doesn't support sm_60.
+// RUN: %clang -v -### --cuda-gpu-arch=sm_60 --sysroot=%S/Inputs/CUDA 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=ERR_SM60
+
+// This should only complain about sm_60, not sm_35.
+// RUN: %clang -v -### --cuda-gpu-arch=sm_60 --cuda-gpu-arch=sm_35 \
+// RUN:    --sysroot=%S/Inputs/CUDA 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=ERR_SM60 --check-prefix=OK_SM35
+
+// We should get two errors here, one for sm_60 and one for sm_61.
+// RUN: %clang -v -### --cuda-gpu-arch=sm_60 --cuda-gpu-arch=sm_61 \
+// RUN:    --sysroot=%S/Inputs/CUDA 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=ERR_SM60 --check-prefix=ERR_SM61
+
+// We should still get an error if we pass -nocudainc, because this compilation
+// would invoke ptxas, and we do a version check on that, too.
+// RUN: %clang -v -### --cuda-gpu-arch=sm_60 -nocudainc --sysroot=%S/Inputs/CUDA 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=ERR_SM60
+
+// If with -nocudainc and -E, we don't touch the CUDA install, so we
+// shouldn't get an error.
+// RUN: %clang -v -### -E --cuda-device-only --cuda-gpu-arch=sm_60 -nocudainc \
+// RUN:    --sysroot=%S/Inputs/CUDA 2>&1 %s | \
+// RUN:    FileCheck %s --check-prefix=OK
+
+// -nocuda-version-check should suppress all of these errors.
+// RUN: %clang -v -### --cuda-gpu-arch=sm_60 --sysroot=%S/Inputs/CUDA 2>&1 \
+// RUN:    -nocuda-version-check %s | \
+// RUN:    FileCheck %s --check-prefix=OK
+
+// OK-NOT: error: GPU arch
+
+// OK_SM35-NOT: error: GPU arch sm_35
+
+// 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