diff options
Diffstat (limited to 'unsupported/test/cxx11_tensor_cuda.cu')
| -rw-r--r-- | unsupported/test/cxx11_tensor_cuda.cu | 186 |
1 files changed, 177 insertions, 9 deletions
diff --git a/unsupported/test/cxx11_tensor_cuda.cu b/unsupported/test/cxx11_tensor_cuda.cu index 4d8465756..134359611 100644 --- a/unsupported/test/cxx11_tensor_cuda.cu +++ b/unsupported/test/cxx11_tensor_cuda.cu @@ -122,6 +122,43 @@ void test_cuda_elementwise() cudaFree(d_out); } +void test_cuda_props() { + Tensor<float, 1> in1(200); + Tensor<bool, 1> out(200); + in1.setRandom(); + + std::size_t in1_bytes = in1.size() * sizeof(float); + std::size_t out_bytes = out.size() * sizeof(bool); + + float* d_in1; + bool* d_out; + cudaMalloc((void**)(&d_in1), in1_bytes); + cudaMalloc((void**)(&d_out), out_bytes); + + cudaMemcpy(d_in1, in1.data(), in1_bytes, cudaMemcpyHostToDevice); + + Eigen::CudaStreamDevice stream; + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_in1( + d_in1, 200); + Eigen::TensorMap<Eigen::Tensor<bool, 1>, Eigen::Aligned> gpu_out( + d_out, 200); + + gpu_out.device(gpu_device) = (gpu_in1.isnan)(); + + assert(cudaMemcpyAsync(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost, + gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 200; ++i) { + VERIFY_IS_EQUAL(out(i), (std::isnan)(in1(i))); + } + + cudaFree(d_in1); + cudaFree(d_out); +} + void test_cuda_reduction() { Tensor<float, 4> in1(72,53,97,113); @@ -627,6 +664,132 @@ void test_cuda_digamma() } template <typename Scalar> +void test_cuda_zeta() +{ + Tensor<Scalar, 1> in_x(6); + Tensor<Scalar, 1> in_q(6); + Tensor<Scalar, 1> out(6); + Tensor<Scalar, 1> expected_out(6); + out.setZero(); + + in_x(0) = Scalar(1); + in_x(1) = Scalar(1.5); + in_x(2) = Scalar(4); + in_x(3) = Scalar(-10.5); + in_x(4) = Scalar(10000.5); + in_x(5) = Scalar(3); + + in_q(0) = Scalar(1.2345); + in_q(1) = Scalar(2); + in_q(2) = Scalar(1.5); + in_q(3) = Scalar(3); + in_q(4) = Scalar(1.0001); + in_q(5) = Scalar(-2.5); + + expected_out(0) = std::numeric_limits<Scalar>::infinity(); + expected_out(1) = Scalar(1.61237534869); + expected_out(2) = Scalar(0.234848505667); + expected_out(3) = Scalar(1.03086757337e-5); + expected_out(4) = Scalar(0.367879440865); + expected_out(5) = Scalar(0.054102025820864097); + + std::size_t bytes = in_x.size() * sizeof(Scalar); + + Scalar* d_in_x; + Scalar* d_in_q; + Scalar* d_out; + cudaMalloc((void**)(&d_in_x), bytes); + cudaMalloc((void**)(&d_in_q), bytes); + cudaMalloc((void**)(&d_out), bytes); + + cudaMemcpy(d_in_x, in_x.data(), bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in_q, in_q.data(), bytes, cudaMemcpyHostToDevice); + + Eigen::CudaStreamDevice stream; + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 6); + Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_q(d_in_q, 6); + Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 6); + + gpu_out.device(gpu_device) = gpu_in_x.zeta(gpu_in_q); + + assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + VERIFY_IS_EQUAL(out(0), expected_out(0)); + VERIFY_IS_APPROX_OR_LESS_THAN(out(3), expected_out(3)); + + for (int i = 1; i < 6; ++i) { + if (i != 3) { + VERIFY_IS_APPROX(out(i), expected_out(i)); + } + } +} + +template <typename Scalar> +void test_cuda_polygamma() +{ + Tensor<Scalar, 1> in_x(7); + Tensor<Scalar, 1> in_n(7); + Tensor<Scalar, 1> out(7); + Tensor<Scalar, 1> expected_out(7); + out.setZero(); + + in_n(0) = Scalar(1); + in_n(1) = Scalar(1); + in_n(2) = Scalar(1); + in_n(3) = Scalar(17); + in_n(4) = Scalar(31); + in_n(5) = Scalar(28); + in_n(6) = Scalar(8); + + in_x(0) = Scalar(2); + in_x(1) = Scalar(3); + in_x(2) = Scalar(25.5); + in_x(3) = Scalar(4.7); + in_x(4) = Scalar(11.8); + in_x(5) = Scalar(17.7); + in_x(6) = Scalar(30.2); + + expected_out(0) = Scalar(0.644934066848); + expected_out(1) = Scalar(0.394934066848); + expected_out(2) = Scalar(0.0399946696496); + expected_out(3) = Scalar(293.334565435); + expected_out(4) = Scalar(0.445487887616); + expected_out(5) = Scalar(-2.47810300902e-07); + expected_out(6) = Scalar(-8.29668781082e-09); + + std::size_t bytes = in_x.size() * sizeof(Scalar); + + Scalar* d_in_x; + Scalar* d_in_n; + Scalar* d_out; + cudaMalloc((void**)(&d_in_x), bytes); + cudaMalloc((void**)(&d_in_n), bytes); + cudaMalloc((void**)(&d_out), bytes); + + cudaMemcpy(d_in_x, in_x.data(), bytes, cudaMemcpyHostToDevice); + cudaMemcpy(d_in_n, in_n.data(), bytes, cudaMemcpyHostToDevice); + + Eigen::CudaStreamDevice stream; + Eigen::GpuDevice gpu_device(&stream); + + Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_x(d_in_x, 7); + Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_in_n(d_in_n, 7); + Eigen::TensorMap<Eigen::Tensor<Scalar, 1> > gpu_out(d_out, 7); + + gpu_out.device(gpu_device) = gpu_in_n.polygamma(gpu_in_x); + + assert(cudaMemcpyAsync(out.data(), d_out, bytes, cudaMemcpyDeviceToHost, gpu_device.stream()) == cudaSuccess); + assert(cudaStreamSynchronize(gpu_device.stream()) == cudaSuccess); + + for (int i = 0; i < 7; ++i) { + VERIFY_IS_APPROX(out(i), expected_out(i)); + } +} + +template <typename Scalar> void test_cuda_igamma() { Tensor<Scalar, 2> a(6, 6); @@ -841,6 +1004,7 @@ void test_cxx11_tensor_cuda() { CALL_SUBTEST_1(test_cuda_elementwise_small()); CALL_SUBTEST_1(test_cuda_elementwise()); + CALL_SUBTEST_1(test_cuda_props()); CALL_SUBTEST_1(test_cuda_reduction()); CALL_SUBTEST_2(test_cuda_contraction<ColMajor>()); CALL_SUBTEST_2(test_cuda_contraction<RowMajor>()); @@ -862,8 +1026,10 @@ void test_cxx11_tensor_cuda() CALL_SUBTEST_4(test_cuda_lgamma<float>(0.01f)); CALL_SUBTEST_4(test_cuda_lgamma<float>(0.001f)); - CALL_SUBTEST_4(test_cuda_digamma<float>()); - + CALL_SUBTEST_4(test_cuda_lgamma<double>(1.0)); + CALL_SUBTEST_4(test_cuda_lgamma<double>(100.0)); + CALL_SUBTEST_4(test_cuda_lgamma<double>(0.01)); + CALL_SUBTEST_4(test_cuda_lgamma<double>(0.001)); CALL_SUBTEST_4(test_cuda_erf<float>(1.0f)); CALL_SUBTEST_4(test_cuda_erf<float>(100.0f)); @@ -876,13 +1042,6 @@ void test_cxx11_tensor_cuda() CALL_SUBTEST_4(test_cuda_erfc<float>(0.01f)); CALL_SUBTEST_4(test_cuda_erfc<float>(0.001f)); - CALL_SUBTEST_4(test_cuda_lgamma<double>(1.0)); - CALL_SUBTEST_4(test_cuda_lgamma<double>(100.0)); - CALL_SUBTEST_4(test_cuda_lgamma<double>(0.01)); - CALL_SUBTEST_4(test_cuda_lgamma<double>(0.001)); - - CALL_SUBTEST_4(test_cuda_digamma<double>()); - CALL_SUBTEST_4(test_cuda_erf<double>(1.0)); CALL_SUBTEST_4(test_cuda_erf<double>(100.0)); CALL_SUBTEST_4(test_cuda_erf<double>(0.01)); @@ -894,6 +1053,15 @@ void test_cxx11_tensor_cuda() CALL_SUBTEST_4(test_cuda_erfc<double>(0.01)); CALL_SUBTEST_4(test_cuda_erfc<double>(0.001)); + CALL_SUBTEST_5(test_cuda_digamma<float>()); + CALL_SUBTEST_5(test_cuda_digamma<double>()); + + CALL_SUBTEST_5(test_cuda_polygamma<float>()); + CALL_SUBTEST_5(test_cuda_polygamma<double>()); + + CALL_SUBTEST_5(test_cuda_zeta<float>()); + CALL_SUBTEST_5(test_cuda_zeta<double>()); + CALL_SUBTEST_5(test_cuda_igamma<float>()); CALL_SUBTEST_5(test_cuda_igammac<float>()); |