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

removing the workaround for hipMalloc / hipHostMalloc

Browse Source
This commit is contained in:
Deven Desai 2019-03-07 14:22:22 +00:00
parent fc9e5a9435
commit 699ff43685
3 changed files with 8 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/stream_executor/gpu/gpu_types.h
View File

@ -20,6 +20,8 @@ limitations under the License.
#if TENSORFLOW_USE_ROCM #if TENSORFLOW_USE_ROCM
#define __HIP_DISABLE_CPP_FUNCTIONS__
#include "rocm/include/hip/hip_complex.h" #include "rocm/include/hip/hip_complex.h"
#include "rocm/include/hip/hip_runtime.h" #include "rocm/include/hip/hip_runtime.h"
#include "rocm/include/hiprand/hiprand.h" #include "rocm/include/hiprand/hiprand.h"

4
tensorflow/stream_executor/rocm/rocm_driver.cc
View File

@ -671,7 +671,7 @@ GpuDriver::ContextGetSharedMemConfig(GpuContext* context) {
uint64 bytes) { uint64 bytes) {
ScopedActivateContext activated{context}; ScopedActivateContext activated{context};
hipDeviceptr_t result = 0; hipDeviceptr_t result = 0;
hipError_t res = tensorflow::wrap::hipMallocVanilla(&result, bytes); hipError_t res = tensorflow::wrap::hipMalloc(&result, bytes);
if (res != hipSuccess) { if (res != hipSuccess) {
LOG(ERROR) << "failed to allocate " LOG(ERROR) << "failed to allocate "
<< port::HumanReadableNumBytes::ToString(bytes) << " (" << bytes << port::HumanReadableNumBytes::ToString(bytes) << " (" << bytes
@ -717,7 +717,7 @@ GpuDriver::ContextGetSharedMemConfig(GpuContext* context) {
ScopedActivateContext activation{context}; ScopedActivateContext activation{context};
void* host_mem = nullptr; void* host_mem = nullptr;
// "Portable" memory is visible to all ROCM contexts. Safe for our use model. // "Portable" memory is visible to all ROCM contexts. Safe for our use model.
hipError_t res = tensorflow::wrap::hipHostMallocVanilla( hipError_t res = tensorflow::wrap::hipHostMalloc(
&host_mem, bytes, hipHostMallocPortable); &host_mem, bytes, hipHostMallocPortable);
if (res != hipSuccess) { if (res != hipSuccess) {
LOG(ERROR) << "failed to alloc " << bytes LOG(ERROR) << "failed to alloc " << bytes

19
tensorflow/stream_executor/rocm/rocm_driver_wrapper.h
View File

@ -20,6 +20,8 @@ limitations under the License.
#ifndef TENSORFLOW_STREAM_EXECUTOR_ROCM_ROCM_DRIVER_WRAPPER_H_ #ifndef TENSORFLOW_STREAM_EXECUTOR_ROCM_ROCM_DRIVER_WRAPPER_H_
#define TENSORFLOW_STREAM_EXECUTOR_ROCM_ROCM_DRIVER_WRAPPER_H_ #define TENSORFLOW_STREAM_EXECUTOR_ROCM_ROCM_DRIVER_WRAPPER_H_
#define __HIP_DISABLE_CPP_FUNCTIONS__
#include "rocm/include/hip/hip_runtime.h" #include "rocm/include/hip/hip_runtime.h"
#include "tensorflow/stream_executor/lib/env.h" #include "tensorflow/stream_executor/lib/env.h"
#include "tensorflow/stream_executor/platform/dso_loader.h" #include "tensorflow/stream_executor/platform/dso_loader.h"
@ -48,21 +50,6 @@ namespace wrap {
#define TO_STR_(x) #x #define TO_STR_(x) #x
#define TO_STR(x) TO_STR_(x) #define TO_STR(x) TO_STR_(x)
// hipMalloc and hipHostMalloc are defined as funtion templates in the
// HIP header files, and hence their names get mangled and the attempt
// to resolve their name when trying to dynamically load them will fail
// Updating the HIP header files to make them C functions is underway.
// Until that change flows through, we will workaround the issue by
// creating dummy wrappers for them here
hipError_t hipMallocVanilla(void** ptr, size_t size) {
return hipErrorNotInitialized;
}
hipError_t hipHostMallocVanilla(void** ptr, size_t size, unsigned int flags) {
return hipErrorNotInitialized;
}
#define STREAM_EXECUTOR_HIP_WRAP(hipSymbolName) \ #define STREAM_EXECUTOR_HIP_WRAP(hipSymbolName) \
template <typename... Args> \ template <typename... Args> \
auto hipSymbolName(Args... args)->decltype(::hipSymbolName(args...)) { \ auto hipSymbolName(Args... args)->decltype(::hipSymbolName(args...)) { \
@ -107,9 +94,11 @@ hipError_t hipHostMallocVanilla(void** ptr, size_t size, unsigned int flags) {
__macro(hipGetDeviceCount) \ __macro(hipGetDeviceCount) \
__macro(hipGetDeviceProperties) \ __macro(hipGetDeviceProperties) \
__macro(hipHostFree) \ __macro(hipHostFree) \
__macro(hipHostMalloc) \
__macro(hipHostRegister) \ __macro(hipHostRegister) \
__macro(hipHostUnregister) \ __macro(hipHostUnregister) \
__macro(hipInit) \ __macro(hipInit) \
__macro(hipMalloc) \
__macro(hipMemGetAddressRange) \ __macro(hipMemGetAddressRange) \
__macro(hipMemGetInfo) \ __macro(hipMemGetInfo) \
__macro(hipMemcpyDtoD) \ __macro(hipMemcpyDtoD) \