GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-07-08 22:21:49 +08:00
Code Issues Packages Projects Releases Wiki Activity

Extended the functionality of the TensorDeviceType classes

Browse Source
This commit is contained in:
Benoit Steiner 2014-07-08 16:30:48 -07:00
parent 7d53633e05
commit c285fda7f4
1 changed files with 56 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
unsupported/Eigen/CXX11/src/Tensor/TensorDeviceType.h
View File

@ -21,6 +21,12 @@ struct DefaultDevice {
EIGEN_STRONG_INLINE void deallocate(void* buffer) const { EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
internal::aligned_free(buffer); internal::aligned_free(buffer);
} }
EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
::memcpy(dst, src, n);
}
EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
::memset(buffer, c, n);
}
}; };
@ -37,6 +43,12 @@ struct ThreadPoolDevice {
EIGEN_STRONG_INLINE void deallocate(void* buffer) const { EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
internal::aligned_free(buffer); internal::aligned_free(buffer);
} }
EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
::memcpy(dst, src, n);
}
EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
::memset(buffer, c, n);
}
private: private:
// todo: NUMA, ... // todo: NUMA, ...
@ -47,20 +59,61 @@ struct ThreadPoolDevice {
// GPU offloading // GPU offloading
#ifdef EIGEN_USE_GPU #ifdef EIGEN_USE_GPU
static int m_numMultiProcessors = 0;
static int m_maxThreadsPerBlock = 0;
static int m_maxThreadsPerMultiProcessor = 0;
static inline int getNumCudaMultiProcessors() {
if (m_numMultiProcessors == 0) {
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, 0);
m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock;
m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor;
m_numMultiProcessors = deviceProp.multiProcessorCount;
}
return m_numMultiProcessors;
}
static inline int maxCudaThreadsPerBlock() {
if (m_maxThreadsPerBlock == 0) {
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, 0);
m_numMultiProcessors = deviceProp.multiProcessorCount;
m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor;
m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock;
}
return m_maxThreadsPerBlock;
}
static inline int maxCudaThreadsPerMultiProcessor() {
if (m_maxThreadsPerBlock == 0) {
cudaDeviceProp deviceProp;
cudaGetDeviceProperties(&deviceProp, 0);
m_numMultiProcessors = deviceProp.multiProcessorCount;
m_maxThreadsPerBlock = deviceProp.maxThreadsPerBlock;
m_maxThreadsPerMultiProcessor = deviceProp.maxThreadsPerMultiProcessor;
}
return m_maxThreadsPerMultiProcessor;
}
struct GpuDevice { struct GpuDevice {
// The cudastream is not owned: the caller is responsible for its initialization and eventual destruction. // The cudastream is not owned: the caller is responsible for its initialization and eventual destruction.
GpuDevice(const cudaStream_t* stream) : stream_(stream) { eigen_assert(stream); } GpuDevice(const cudaStream_t* stream) : stream_(stream) { eigen_assert(stream); }
EIGEN_STRONG_INLINE const cudaStream_t& stream() const { return *stream_; } EIGEN_STRONG_INLINE const cudaStream_t& stream() const { return *stream_; }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { /*EIGEN_DEVICE_FUNC*/ EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
void* result; void* result;
cudaMalloc(&result, num_bytes); cudaMalloc(&result, num_bytes);
return result; return result;
} }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const { /*EIGEN_DEVICE_FUNC */EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
cudaFree(buffer); cudaFree(buffer);
} }
EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
cudaMemcpyAsync(dst, src, n, cudaMemcpyDeviceToDevice, *stream_);
}
EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
cudaMemsetAsync(buffer, c, n, *stream_);
}
private: private:
// TODO: multigpu. // TODO: multigpu.