aboutsummaryrefslogtreecommitdiffhomepage
path: root/unsupported/test/cxx11_tensor_of_float16_cuda.cu
diff options
context:
space:
mode:
authorGravatar Benoit Steiner <benoit.steiner.goog@gmail.com>2016-04-29 13:41:26 -0700
committerGravatar Benoit Steiner <benoit.steiner.goog@gmail.com>2016-04-29 13:41:26 -0700
commit07a247dcf4e86f9f741b68e1d8e0897de3eeca57 (patch)
treed103bd20faa1f103035bac2f21507ecc65f97f68 /unsupported/test/cxx11_tensor_of_float16_cuda.cu
parentfa5a8f055aebbf4f39fca26e857351103fab4d11 (diff)
parent0f3c4c8ff4a6635db77195a8919c743f34181cc2 (diff)
Pulled latest updates from upstream
Diffstat (limited to 'unsupported/test/cxx11_tensor_of_float16_cuda.cu')
-rw-r--r--unsupported/test/cxx11_tensor_of_float16_cuda.cu37
1 files changed, 37 insertions, 0 deletions
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index cb917bb37..154a72d5c 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -228,6 +228,42 @@ void test_cuda_reductions() {
gpu_device.deallocate(d_res_float);
}
+void test_cuda_forced_evals() {
+
+ Eigen::CudaStreamDevice stream;
+ Eigen::GpuDevice gpu_device(&stream);
+ int num_elem = 101;
+
+ float* d_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+ float* d_res_half = (float*)gpu_device.allocate(num_elem * sizeof(float));
+ float* d_res_float = (float*)gpu_device.allocate(num_elem * sizeof(float));
+
+ Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float(
+ d_float, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
+ d_res_half, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
+ d_res_float, num_elem);
+
+ gpu_float.device(gpu_device) = gpu_float.random() - gpu_float.constant(0.5f);
+ gpu_res_float.device(gpu_device) = gpu_float.abs();
+ gpu_res_half.device(gpu_device) = gpu_float.cast<Eigen::half>().abs().eval().cast<float>();
+
+ Tensor<float, 1> half_prec(num_elem);
+ Tensor<float, 1> full_prec(num_elem);
+ gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(float));
+ gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(float));
+ gpu_device.synchronize();
+
+ for (int i = 0; i < num_elem; ++i) {
+ std::cout << "Checking unary " << i << std::endl;
+ VERIFY_IS_APPROX(full_prec(i), half_prec(i));
+ }
+
+ gpu_device.deallocate(d_float);
+ gpu_device.deallocate(d_res_half);
+ gpu_device.deallocate(d_res_float);
+}
#endif
@@ -246,6 +282,7 @@ void test_cxx11_tensor_of_float16_cuda()
CALL_SUBTEST_1(test_cuda_elementwise());
CALL_SUBTEST_2(test_cuda_contractions());
CALL_SUBTEST_3(test_cuda_reductions());
+ CALL_SUBTEST_4(test_cuda_forced_evals());
}
else {
std::cout << "Half floats require compute capability of at least 5.3. This device only supports " << device.majorDeviceVersion() << "." << device.minorDeviceVersion() << ". Skipping the test" << std::endl;