From 5e9a1e7a7a7eccbb20a2c4eb44141727b0943f11 Mon Sep 17 00:00:00 2001 From: Mehdi Goli Date: Wed, 8 Mar 2017 14:17:48 +0000 Subject: Adding sycl Benchmarks. --- bench/tensors/tensor_benchmarks.h | 106 +++++++++++++++++++++++++++++++++----- 1 file changed, 92 insertions(+), 14 deletions(-) (limited to 'bench/tensors/tensor_benchmarks.h') diff --git a/bench/tensors/tensor_benchmarks.h b/bench/tensors/tensor_benchmarks.h index c2fb3dede..325026113 100644 --- a/bench/tensors/tensor_benchmarks.h +++ b/bench/tensors/tensor_benchmarks.h @@ -35,6 +35,11 @@ template class BenchmarkSuite { void memcpy(int num_iters) { eigen_assert(m_ == k_ && k_ == n_); +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + device_.memcpy(c_, a_, m_ * m_ * sizeof(T)); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { device_.memcpy(c_, a_, m_ * m_ * sizeof(T)); @@ -55,7 +60,11 @@ template class BenchmarkSuite { } const TensorMap, Eigen::Aligned> A((int*)a_, sizes); TensorMap, Eigen::Aligned> B(b_, sizes); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + B.device(device_) = A.template cast(); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { B.device(device_) = A.template cast(); @@ -70,7 +79,6 @@ template class BenchmarkSuite { sizes[0] = m_; sizes[1] = m_; TensorMap, Eigen::Aligned> C(c_, sizes); - StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = C.random(); @@ -93,7 +101,18 @@ template class BenchmarkSuite { const Eigen::DSizes second_quadrant(0, m_/2); const Eigen::DSizes third_quadrant(m_/2, 0); const Eigen::DSizes fourth_quadrant(m_/2, m_/2); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.slice(first_quadrant, quarter_sizes).device(device_) = + A.slice(first_quadrant, quarter_sizes); + C.slice(second_quadrant, quarter_sizes).device(device_) = + B.slice(second_quadrant, quarter_sizes); + C.slice(third_quadrant, quarter_sizes).device(device_) = + A.slice(third_quadrant, quarter_sizes); + C.slice(fourth_quadrant, quarter_sizes).device(device_) = + B.slice(fourth_quadrant, quarter_sizes); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.slice(first_quadrant, quarter_sizes).device(device_) = @@ -118,7 +137,11 @@ template class BenchmarkSuite { Eigen::array output_size; output_size[0] = n_; TensorMap, Eigen::Aligned> C(c_, output_size); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = B.chip(iter % k_, 0); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = B.chip(iter % k_, 0); @@ -135,7 +158,11 @@ template class BenchmarkSuite { Eigen::array output_size; output_size[0] = n_; TensorMap, Eigen::Aligned> C(c_, output_size); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = B.chip(iter % n_, 1); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = B.chip(iter % n_, 1); @@ -158,7 +185,11 @@ template class BenchmarkSuite { Eigen::array shuffle; shuffle[0] = 1; shuffle[1] = 0; - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + B.device(device_) = A.shuffle(shuffle); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { B.device(device_) = A.shuffle(shuffle); @@ -186,7 +217,11 @@ template class BenchmarkSuite { paddings[0] = Eigen::IndexPair(0, 0); paddings[1] = Eigen::IndexPair(2, 1); #endif - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + B.device(device_) = A.pad(paddings); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { B.device(device_) = A.pad(paddings); @@ -216,6 +251,11 @@ template class BenchmarkSuite { Eigen::IndexList, Eigen::type2index<2> > strides; #endif +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + B.device(device_) = A.stride(strides); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { B.device(device_) = A.stride(strides); @@ -245,6 +285,11 @@ template class BenchmarkSuite { broadcast.set(1, n_); #endif +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = A.broadcast(broadcast); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = A.broadcast(broadcast); @@ -261,7 +306,11 @@ template class BenchmarkSuite { const TensorMap, Eigen::Aligned> A(a_, sizes); const TensorMap, Eigen::Aligned> B(b_, sizes); TensorMap, Eigen::Aligned> C(c_, sizes); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = A * A.constant(static_cast(3.14)) + B * B.constant(static_cast(2.7)); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = A * A.constant(static_cast(3.14)) + B * B.constant(static_cast(2.7)); @@ -280,6 +329,11 @@ template class BenchmarkSuite { const TensorMap, Eigen::Aligned> B(b_, sizes); TensorMap, Eigen::Aligned> C(c_, sizes); +#ifdef EIGEN_USE_SYCL // warmup for sycl +for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = A.rsqrt() + B.sqrt() * B.square(); +} +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = A.rsqrt() + B.sqrt() * B.square(); @@ -297,7 +351,11 @@ template class BenchmarkSuite { const TensorMap, Eigen::Aligned> A(a_, sizes); const TensorMap, Eigen::Aligned> B(b_, sizes); TensorMap, Eigen::Aligned> C(c_, sizes); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = A.exp() + B.log(); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = A.exp() + B.log(); @@ -325,7 +383,11 @@ template class BenchmarkSuite { // optimize the code. Eigen::IndexList> sum_along_dim; #endif - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = B.sum(sum_along_dim); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = B.sum(sum_along_dim); @@ -355,7 +417,11 @@ template class BenchmarkSuite { // optimize the code. Eigen::IndexList> sum_along_dim; #endif - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = B.sum(sum_along_dim); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = B.sum(sum_along_dim); @@ -375,7 +441,11 @@ template class BenchmarkSuite { Eigen::array output_size; TensorMap, Eigen::Aligned> C( c_, output_size); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = B.sum(); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = B.sum(); @@ -404,7 +474,11 @@ template class BenchmarkSuite { typedef typename Tensor::DimensionPair DimPair; Eigen::array dims; dims[0] = DimPair(1, 0); - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = A.contract(B, dims); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = A.contract(B, dims); @@ -430,7 +504,11 @@ template class BenchmarkSuite { Eigen::array dims; dims[0] = 0; dims[1] = 1; - +#ifdef EIGEN_USE_SYCL // warmup for sycl + for (int iter = 0; iter < 10; ++iter) { + C.device(device_) = A.convolve(B, dims); + } +#endif StartBenchmarkTiming(); for (int iter = 0; iter < num_iters; ++iter) { C.device(device_) = A.convolve(B, dims); -- cgit v1.2.3