TensorDeviceDefault.h (3427B)
1 // This file is part of Eigen, a lightweight C++ template library 2 // for linear algebra. 3 // 4 // Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com> 5 // 6 // This Source Code Form is subject to the terms of the Mozilla 7 // Public License v. 2.0. If a copy of the MPL was not distributed 8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 9 10 #ifndef EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H 11 #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H 12 13 14 namespace Eigen { 15 16 // Default device for the machine (typically a single cpu core) 17 struct DefaultDevice { 18 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const { 19 return internal::aligned_malloc(num_bytes); 20 } 21 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const { 22 internal::aligned_free(buffer); 23 } 24 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const { 25 return allocate(num_bytes); 26 } 27 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const { 28 deallocate(buffer); 29 } 30 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const { 31 ::memcpy(dst, src, n); 32 } 33 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const { 34 memcpy(dst, src, n); 35 } 36 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const { 37 memcpy(dst, src, n); 38 } 39 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const { 40 ::memset(buffer, c, n); 41 } 42 template<typename Type> 43 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Type get(Type data) const { 44 return data; 45 } 46 47 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t numThreads() const { 48 #if !defined(EIGEN_GPU_COMPILE_PHASE) 49 // Running on the host CPU 50 return 1; 51 #elif defined(EIGEN_HIP_DEVICE_COMPILE) 52 // Running on a HIP device 53 return 64; 54 #else 55 // Running on a CUDA device 56 return 32; 57 #endif 58 } 59 60 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const { 61 #if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY) 62 // Running on the host CPU 63 return l1CacheSize(); 64 #elif defined(EIGEN_HIP_DEVICE_COMPILE) 65 // Running on a HIP device 66 return 48*1024; // FIXME : update this number for HIP 67 #else 68 // Running on a CUDA device, return the amount of shared memory available. 69 return 48*1024; 70 #endif 71 } 72 73 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const { 74 #if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY) 75 // Running single threaded on the host CPU 76 return l3CacheSize(); 77 #elif defined(EIGEN_HIP_DEVICE_COMPILE) 78 // Running on a HIP device 79 return firstLevelCacheSize(); // FIXME : update this number for HIP 80 #else 81 // Running on a CUDA device 82 return firstLevelCacheSize(); 83 #endif 84 } 85 86 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const { 87 #if !defined(EIGEN_GPU_COMPILE_PHASE) 88 // Running single threaded on the host CPU 89 // Should return an enum that encodes the ISA supported by the CPU 90 return 1; 91 #elif defined(EIGEN_HIP_DEVICE_COMPILE) 92 // Running on a HIP device 93 // return 1 as major for HIP 94 return 1; 95 #else 96 // Running on a CUDA device 97 return EIGEN_CUDA_ARCH / 100; 98 #endif 99 } 100 }; 101 102 } // namespace Eigen 103 104 #endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H