| /external/tensorflow/tensorflow/compiler/xla/service/gpu/ |
| gpu_executable.h | 50 // cubin (i.e. the compiled ptx) may be empty, in which case we leave
52 GpuExecutable(const string& ptx, const std::vector<uint8>& cubin,
71 // Returns the cubin (compiled PTX) stored in this GpuExecutable. May be
73 const std::vector<uint8>& cubin() const { return cubin_; } function in class:xla::gpu::GpuExecutable
|
| kernel_thunk.cc | 52 if (!executable.cubin().empty()) {
54 reinterpret_cast<const char*>(executable.cubin().data()), kernel_name_);
|
| gpu_compiler.cc | 399 // code (i.e. a cubin) as a byte array.
426 return InternalError("couldn't get temp CUBIN file name");
429 // CUBIN file may never be created, so the failure to delete it should not
456 string cubin; local
458 cubin_path, &cubin));
459 std::vector<uint8> cubin_vector(cubin.begin(), cubin.end());
647 const std::vector<uint8> cubin = local
668 ptx, cubin, {cc_major, cc_minor}, std::move(thunk_schedule),
682 Tracing::TraceMe annotation("PTX->CUBIN", /*is_expensive=*/true)
[all...] |
| gpu_executable.cc | 114 const string& ptx, const std::vector<uint8>& cubin,
124 cubin_(cubin),
|
| /external/tensorflow/tensorflow/stream_executor/cuda/ |
| cuda_gpu_executor.cc | 73 // Hook that can be used to CUBIN-ate PTX before it is loaded into the driver.
76 // therefore, it's useful to have this hook to hack in uniform CUBIN-ation of
220 const char *cubin = spec.cuda_cubin_in_memory().bytes(); local
223 std::tie(module, module_refcount) = gpu_binary_to_module_[cubin];
226 auto load_status = CUDADriver::LoadCubin(context_, cubin, &module);
228 LOG(ERROR) << "failed to load CUBIN: " << load_status;
232 VLOG(3) << "Loaded CUBIN " << static_cast<const void *>(cubin)
236 VLOG(3) << "CUBIN " << static_cast<const void *>(cubin)
[all...] |
