src/Tensor/TensorDeviceDefault.h

0001 // This file is part of Eigen, a lightweight C++ template library
0002 // for linear algebra.
0003 //
0004 // Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
0005 //
0006 // This Source Code Form is subject to the terms of the Mozilla
0007 // Public License v. 2.0. If a copy of the MPL was not distributed
0008 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
0009
0010 #ifndef EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H
0011 #define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H
0012
0013
0014 namespace Eigen {
0015
0016 // Default device for the machine (typically a single cpu core)
0017 struct DefaultDevice {
0018   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
0019     return internal::aligned_malloc(num_bytes);
0020   }
0021   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate(void* buffer) const {
0022     internal::aligned_free(buffer);
0023   }
0024     EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void* allocate_temp(size_t num_bytes) const {
0025     return allocate(num_bytes);
0026   }
0027   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void deallocate_temp(void* buffer) const {
0028     deallocate(buffer);
0029   }
0030   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpy(void* dst, const void* src, size_t n) const {
0031     ::memcpy(dst, src, n);
0032   }
0033   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyHostToDevice(void* dst, const void* src, size_t n) const {
0034     memcpy(dst, src, n);
0035   }
0036   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memcpyDeviceToHost(void* dst, const void* src, size_t n) const {
0037     memcpy(dst, src, n);
0038   }
0039   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void memset(void* buffer, int c, size_t n) const {
0040     ::memset(buffer, c, n);
0041   }
0042   template<typename Type>
0043   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Type get(Type data) const {
0044     return data;
0045   }
0046
0047   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t numThreads() const {
0048 #if !defined(EIGEN_GPU_COMPILE_PHASE)
0049     // Running on the host CPU
0050     return 1;
0051 #elif defined(EIGEN_HIP_DEVICE_COMPILE)
0052     // Running on a HIP device
0053     return 64;
0054 #else
0055     // Running on a CUDA device
0056     return 32;
0057 #endif
0058   }
0059
0060   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t firstLevelCacheSize() const {
0061 #if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
0062     // Running on the host CPU
0063     return l1CacheSize();
0064 #elif defined(EIGEN_HIP_DEVICE_COMPILE)
0065     // Running on a HIP device
0066     return 48*1024; // FIXME : update this number for HIP
0067 #else
0068     // Running on a CUDA device, return the amount of shared memory available.
0069     return 48*1024;
0070 #endif
0071   }
0072
0073   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE size_t lastLevelCacheSize() const {
0074 #if !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY)
0075     // Running single threaded on the host CPU
0076     return l3CacheSize();
0077 #elif defined(EIGEN_HIP_DEVICE_COMPILE)
0078     // Running on a HIP device
0079     return firstLevelCacheSize(); // FIXME : update this number for HIP
0080 #else
0081     // Running on a CUDA device
0082     return firstLevelCacheSize();
0083 #endif
0084   }
0085
0086   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int majorDeviceVersion() const {
0087 #if !defined(EIGEN_GPU_COMPILE_PHASE)
0088     // Running single threaded on the host CPU
0089     // Should return an enum that encodes the ISA supported by the CPU
0090     return 1;
0091 #elif defined(EIGEN_HIP_DEVICE_COMPILE)
0092     // Running on a HIP device
0093     // return 1 as major for HIP
0094     return 1;
0095 #else
0096     // Running on a CUDA device
0097     return EIGEN_CUDA_ARCH / 100;
0098 #endif
0099   }
0100 };
0101
0102 }  // namespace Eigen
0103
0104 #endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_DEFAULT_H