experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[XLA:GPU] Simplify libdevice-search code.

Browse Source 
We can hardcode a path to libdevice for CUDA 9+, and that's all we support nowadays.

PiperOrigin-RevId: 254093960
This commit is contained in:
Justin Lebar 2019-06-19 16:23:59 -07:00 committed by TensorFlower Gardener
parent c8f1b3826f
commit 835bcfba8f
1 changed files with 15 additions and 48 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
tensorflow/compiler/xla/service/gpu/llvm_gpu_backend/nvptx_backend_lib.cc
View File

@ -68,47 +68,6 @@ namespace {
// Default inline threshold value to use in llvm. // Default inline threshold value to use in llvm.
const int kDefaultInlineThreshold = 1100; const int kDefaultInlineThreshold = 1100;
// Gets the libdevice filename for a particular compute capability. When
// presented with a GPU we don't recognize, we just return the libdevice from
// compute_20.
static string GetLibdeviceFilename(const string& libdevice_dir_path,
std::pair<int, int> compute_capability) {
// Since CUDA 9.0, all GPU versions are included in a single file
const char* unified_libdevice_filename = "libdevice.10.bc";
std::vector<string> unified_libdevice_files;
const Status status = tensorflow::Env::Default()->GetMatchingPaths(
tensorflow::io::JoinPath(libdevice_dir_path, unified_libdevice_filename),
&unified_libdevice_files);
if (status.ok() && unified_libdevice_files.size() == 1) {
return unified_libdevice_filename;
}
// There are only four libdevice files: compute_{20,30,35,50}. Each GPU
// version gets mapped to one of these. Note in particular that sm_60 and
// sm_61 map to libdevice.compute_30.
static auto* m = new std::map<std::pair<int, int>, int>({{{2, 0}, 20},
{{2, 1}, 20},
{{3, 0}, 30},
{{3, 2}, 30},
{{3, 5}, 35},
{{3, 7}, 35},
{{5, 0}, 50},
{{5, 2}, 50},
{{5, 3}, 50},
{{6, 0}, 30},
{{6, 1}, 30},
{{6, 2}, 30}});
int libdevice_version = 20;
auto it = m->find(compute_capability);
if (it != m->end()) {
libdevice_version = it->second;
} else {
LOG(WARNING) << "Unknown compute capability (" << compute_capability.first
<< ", " << compute_capability.second << ") ."
<< "Defaulting to libdevice for compute_" << libdevice_version;
}
return absl::StrCat("libdevice.compute_", libdevice_version, ".10.bc");
}
// Gets the GPU name as it's known to LLVM for a given compute capability. If // Gets the GPU name as it's known to LLVM for a given compute capability. If
// we see an unrecognized compute capability, we return "sm_35". // we see an unrecognized compute capability, we return "sm_35".
static string GetSmName(std::pair<int, int> compute_capability) { static string GetSmName(std::pair<int, int> compute_capability) {
@ -303,14 +262,22 @@ Status LinkLibdeviceIfNecessary(llvm::Module* module,
return Status::OK(); return Status::OK();
} }
llvm::Linker linker(*module); // CUDA 9+ uses a single libdevice file for all devices, and we don't support
string libdevice_path = tensorflow::io::JoinPath( // older CUDAs.
libdevice_dir_path, string libdevice_path =
GetLibdeviceFilename(libdevice_dir_path, compute_capability)); tensorflow::io::JoinPath(libdevice_dir_path, "libdevice.10.bc");
TF_RETURN_IF_ERROR(tensorflow::Env::Default()->FileExists(libdevice_path)); if (!tensorflow::Env::Default()->FileExists(libdevice_path).ok()) {
LOG(WARNING)
<< "libdevice is required by this HLO module but was not found at "
<< libdevice_path;
return xla::InternalError("libdevice not found at %s", libdevice_path);
}
VLOG(1) << "Linking with libdevice from: " << libdevice_path; VLOG(1) << "Linking with libdevice from: " << libdevice_path;
std::unique_ptr<llvm::Module> libdevice_module = std::unique_ptr<llvm::Module> libdevice_module =
LoadIRModule(libdevice_path, &module->getContext()); LoadIRModule(libdevice_path, &module->getContext());
llvm::Linker linker(*module);
if (linker.linkInModule( if (linker.linkInModule(
std::move(libdevice_module), llvm::Linker::Flags::LinkOnlyNeeded, std::move(libdevice_module), llvm::Linker::Flags::LinkOnlyNeeded,
[](Module& M, const StringSet<>& GVS) { [](Module& M, const StringSet<>& GVS) {
@ -318,8 +285,8 @@ Status LinkLibdeviceIfNecessary(llvm::Module* module,
return !GV.hasName() || (GVS.count(GV.getName()) == 0); return !GV.hasName() || (GVS.count(GV.getName()) == 0);
}); });
})) { })) {
return tensorflow::errors::Internal( return xla::InternalError("Error linking libdevice from %s",
absl::StrCat("Error linking libdevice from ", libdevice_path)); libdevice_path);
} }
return Status::OK(); return Status::OK();
} }