1 #ifdef EIGEN_USE_SYCL 2 3 #include <CL/sycl.hpp> 4 #include <iostream> 5 6 #include "tensor_benchmarks.h" 7 8 cl::sycl::gpu_selector selector; 9 Eigen::QueueInterface queue(selector); 10 #define BM_FuncWithInput2DimsGPU(FUNC, D1, D2) \ 11 static void BM_##FUNC##_##D1##x##D2(int iters, int N) { \ 12 StopBenchmarkTiming(); \ 13 Eigen::SyclDevice device(&queue); \ 14 BenchmarkSuite<Eigen::SyclDevice, float> suite(device, D1, D2); \ 15 suite.FUNC(iters); \ 16 } \ 17 BENCHMARK_RANGE(BM_##FUNC##_##D1##x##D2, 10, 10); 18 19 BM_FuncWithInput2DimsGPU(rowReduction, 256, 100352); 20 BM_FuncWithInput2DimsGPU(rowReduction, 64, 100352); 21 BM_FuncWithInput2DimsGPU(rowReduction, 512, 25088); 22 BM_FuncWithInput2DimsGPU(rowReduction, 128, 25088); 23 BM_FuncWithInput2DimsGPU(rowReduction, 102, 6272); 24 BM_FuncWithInput2DimsGPU(rowReduction, 256, 6272); 25 BM_FuncWithInput2DimsGPU(rowReduction, 204, 1568); 26 BM_FuncWithInput2DimsGPU(rowReduction, 512, 1568); 27 BM_FuncWithInput2DimsGPU(rowReduction, 1024, 1568); 28 BM_FuncWithInput2DimsGPU(rowReduction, 2048, 1568); 29 30 BM_FuncWithInput2DimsGPU(colReduction, 100352, 256); 31 BM_FuncWithInput2DimsGPU(colReduction, 100352, 64); 32 BM_FuncWithInput2DimsGPU(colReduction, 25088, 512); 33 BM_FuncWithInput2DimsGPU(colReduction, 6272, 102); 34 BM_FuncWithInput2DimsGPU(colReduction, 25088, 128); 35 BM_FuncWithInput2DimsGPU(colReduction, 6272, 256); 36 BM_FuncWithInput2DimsGPU(colReduction, 1568, 204); 37 BM_FuncWithInput2DimsGPU(colReduction, 1568, 512); 38 BM_FuncWithInput2DimsGPU(colReduction, 1568, 1024); 39 BM_FuncWithInput2DimsGPU(colReduction, 1568, 2048); 40 BM_FuncWithInput2DimsGPU(fullReduction, 1001, 1); 41 BM_FuncWithInput2DimsGPU(fullReduction, 2050048, 1); 42 BM_FuncWithInput2DimsGPU(fullReduction, 2097152, 1); 43 BM_FuncWithInput2DimsGPU(fullReduction, 2048, 1); 44 BM_FuncWithInput2DimsGPU(fullReduction, 262144, 1); 45 BM_FuncWithInput2DimsGPU(fullReduction, 256, 1); 46 BM_FuncWithInput2DimsGPU(fullReduction, 589824, 1); 47 BM_FuncWithInput2DimsGPU(fullReduction, 1024, 1); 48 BM_FuncWithInput2DimsGPU(fullReduction, 524288, 1); 49 BM_FuncWithInput2DimsGPU(fullReduction, 512, 1); 50 BM_FuncWithInput2DimsGPU(fullReduction, 2359296, 1); 51 BM_FuncWithInput2DimsGPU(fullReduction, 1048576, 1); 52 BM_FuncWithInput2DimsGPU(fullReduction, 131072, 1); 53 BM_FuncWithInput2DimsGPU(fullReduction, 16384, 1); 54 BM_FuncWithInput2DimsGPU(fullReduction, 9408, 1); 55 BM_FuncWithInput2DimsGPU(fullReduction, 64, 1); 56 BM_FuncWithInput2DimsGPU(fullReduction, 4096, 1); 57 BM_FuncWithInput2DimsGPU(fullReduction, 36864, 1); 58 BM_FuncWithInput2DimsGPU(fullReduction, 32768, 1); 59 BM_FuncWithInput2DimsGPU(fullReduction, 128, 1); 60 BM_FuncWithInput2DimsGPU(fullReduction, 147456, 1); 61 BM_FuncWithInput2DimsGPU(fullReduction, 65536, 1); 62 #define BM_FuncGPU(FUNC) \ 63 static void BM_##FUNC(int iters, int N) { \ 64 StopBenchmarkTiming(); \ 65 Eigen::SyclDevice device(&queue); \ 66 BenchmarkSuite<Eigen::SyclDevice, float> suite(device, N); \ 67 suite.FUNC(iters); \ 68 } \ 69 BENCHMARK_RANGE(BM_##FUNC, 10, 5000); 70 71 BM_FuncGPU(rowReduction); 72 BM_FuncGPU(colReduction); 73 BM_FuncGPU(fullReduction); 74 75 BM_FuncGPU(memcpy); 76 BM_FuncGPU(typeCasting); 77 BM_FuncGPU(random); 78 BM_FuncGPU(slicing); 79 BM_FuncGPU(rowChip); 80 BM_FuncGPU(colChip); 81 BM_FuncGPU(shuffling); 82 BM_FuncGPU(padding); 83 BM_FuncGPU(striding); 84 BM_FuncGPU(broadcasting); 85 BM_FuncGPU(coeffWiseOp); 86 BM_FuncGPU(algebraicFunc); 87 BM_FuncGPU(transcendentalFunc); 88 // Contractions 89 #define BM_FuncWithInputDimsGPU(FUNC, D1, D2, D3) \ 90 static void BM_##FUNC##_##D1##x##D2##x##D3(int iters, int N) { \ 91 StopBenchmarkTiming(); \ 92 Eigen::SyclDevice device(&queue); \ 93 BenchmarkSuite<Eigen::SyclDevice, float> suite(device, D1, D2, D3); \ 94 suite.FUNC(iters); \ 95 } \ 96 BENCHMARK_RANGE(BM_##FUNC##_##D1##x##D2##x##D3, 10, 5000); 97 98 BM_FuncWithInputDimsGPU(contraction, N, N, N); 99 BM_FuncWithInputDimsGPU(contraction, 64, N, N); 100 BM_FuncWithInputDimsGPU(contraction, N, 64, N); 101 BM_FuncWithInputDimsGPU(contraction, N, N, 64); 102 103 BM_FuncWithInputDimsGPU(contractionRowMajor, N, N, N); 104 BM_FuncWithInputDimsGPU(contractionRowMajor, 64, N, N); 105 BM_FuncWithInputDimsGPU(contractionRowMajor, N, 64, N); 106 BM_FuncWithInputDimsGPU(contractionRowMajor, N, N, 64); 107 108 BM_FuncWithInputDimsGPU(contractionRowMajorAT, N, N, N); 109 BM_FuncWithInputDimsGPU(contractionRowMajorAT, 64, N, N); 110 BM_FuncWithInputDimsGPU(contractionRowMajorAT, N, 64, N); 111 BM_FuncWithInputDimsGPU(contractionRowMajorAT, N, N, 64); 112 113 BM_FuncWithInputDimsGPU(contractionRowMajorBT, N, N, N); 114 BM_FuncWithInputDimsGPU(contractionRowMajorBT, 64, N, N); 115 BM_FuncWithInputDimsGPU(contractionRowMajorBT, N, 64, N); 116 BM_FuncWithInputDimsGPU(contractionRowMajorBT, N, N, 64); 117 118 119 BM_FuncWithInputDimsGPU(contractionRowMajorABT, N, N, N); 120 BM_FuncWithInputDimsGPU(contractionRowMajorABT, 64, N, N); 121 BM_FuncWithInputDimsGPU(contractionRowMajorABT, N, 64, N); 122 BM_FuncWithInputDimsGPU(contractionRowMajorABT, N, N, 64); 123 124 // Convolutions 125 #define BM_FuncWithKernelDimsGPU(FUNC, DIM1, DIM2) \ 126 static void BM_##FUNC##_##DIM1##x##DIM2(int iters, int N) { \ 127 StopBenchmarkTiming(); \ 128 Eigen::SyclDevice device(&queue); \ 129 BenchmarkSuite<Eigen::SyclDevice, float> suite(device, N); \ 130 suite.FUNC(iters, DIM1, DIM2); \ 131 } \ 132 BENCHMARK_RANGE(BM_##FUNC##_##DIM1##x##DIM2, 128, 5000); 133 134 BM_FuncWithKernelDimsGPU(convolution, 7, 1); 135 BM_FuncWithKernelDimsGPU(convolution, 1, 7); 136 BM_FuncWithKernelDimsGPU(convolution, 7, 4); 137 BM_FuncWithKernelDimsGPU(convolution, 4, 7); 138 BM_FuncWithKernelDimsGPU(convolution, 7, 64); 139 BM_FuncWithKernelDimsGPU(convolution, 64, 7); 140 #endif 141