Home | History | Annotate | Download | only in tensors 
      1 #define EIGEN_USE_GPU
      2 
      3 #include <cuda.h>
      4 #include <cuda_runtime.h>
      5 #include <iostream>
      6 
      7 #include "tensor_benchmarks.h"
      8 
      9 // Simple functions
     10 #define BM_FuncGPU(FUNC)                                                       \
     11   static void BM_##FUNC(int iters, int N) {                                    \
     12     StopBenchmarkTiming();                                                     \
     13     Eigen::CudaStreamDevice stream;                                            \
     14     Eigen::GpuDevice device(&stream);                                          \
     15     BenchmarkSuite<Eigen::GpuDevice, Eigen::half> suite(device, N);            \
     16     cudaDeviceSynchronize();                                                   \
     17     suite.FUNC(iters);                                                         \
     18   }                                                                            \
     19   BENCHMARK_RANGE(BM_##FUNC, 10, 5000);
     20 
     21 BM_FuncGPU(memcpy);
     22 BM_FuncGPU(typeCasting);
     23 //BM_FuncGPU(random);
     24 BM_FuncGPU(slicing);
     25 BM_FuncGPU(rowChip);
     26 BM_FuncGPU(colChip);
     27 BM_FuncGPU(shuffling);
     28 BM_FuncGPU(padding);
     29 BM_FuncGPU(striding);
     30 BM_FuncGPU(broadcasting);
     31 BM_FuncGPU(coeffWiseOp);
     32 BM_FuncGPU(algebraicFunc);
     33 BM_FuncGPU(transcendentalFunc);
     34 BM_FuncGPU(rowReduction);
     35 BM_FuncGPU(colReduction);
     36 BM_FuncGPU(fullReduction);
     37 
     38 
     39 // Contractions
     40 #define BM_FuncWithInputDimsGPU(FUNC, D1, D2, D3)                              \
     41   static void BM_##FUNC##_##D1##x##D2##x##D3(int iters, int N) {               \
     42     StopBenchmarkTiming();                                                     \
     43     Eigen::CudaStreamDevice stream;                                            \
     44     Eigen::GpuDevice device(&stream);                                          \
     45     BenchmarkSuite<Eigen::GpuDevice, Eigen::half> suite(device, D1, D2, D3);   \
     46     cudaDeviceSynchronize();                                                   \
     47     suite.FUNC(iters);                                                         \
     48   }                                                                            \
     49   BENCHMARK_RANGE(BM_##FUNC##_##D1##x##D2##x##D3, 10, 5000);
     50 
     51 
     52 BM_FuncWithInputDimsGPU(contraction, N, N, N);
     53 BM_FuncWithInputDimsGPU(contraction, 64, N, N);
     54 BM_FuncWithInputDimsGPU(contraction, N, 64, N);
     55 BM_FuncWithInputDimsGPU(contraction, N, N, 64);
     56 
     57 
     58 // Convolutions
     59 #define BM_FuncWithKernelDimsGPU(FUNC, DIM1, DIM2)                             \
     60   static void BM_##FUNC##_##DIM1##x##DIM2(int iters, int N) {                  \
     61     StopBenchmarkTiming();                                                     \
     62     Eigen::CudaStreamDevice stream;                                            \
     63     Eigen::GpuDevice device(&stream);                                          \
     64     BenchmarkSuite<Eigen::GpuDevice, Eigen::half> suite(device, N);            \
     65     cudaDeviceSynchronize();                                                   \
     66     suite.FUNC(iters, DIM1, DIM2);                                             \
     67   }                                                                            \
     68   BENCHMARK_RANGE(BM_##FUNC##_##DIM1##x##DIM2, 128, 5000);
     69 
     70 /*
     71 BM_FuncWithKernelDimsGPU(convolution, 7, 1);
     72 BM_FuncWithKernelDimsGPU(convolution, 1, 7);
     73 BM_FuncWithKernelDimsGPU(convolution, 7, 4);
     74 BM_FuncWithKernelDimsGPU(convolution, 4, 7);
     75 BM_FuncWithKernelDimsGPU(convolution, 7, 64);
     76 BM_FuncWithKernelDimsGPU(convolution, 64, 7);
     77 */
     78