GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-03 17:24:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Made the CUDA contract test more robust to numerical noise.

Browse Source
This commit is contained in:
Benoit Steiner 2016-01-30 10:28:43 -08:00
parent 4281eb1e2c
commit ba27c8a7de
1 changed files with 11 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
unsupported/test/cxx11_tensor_contract_cuda.cu
View File

@ -24,7 +24,7 @@ typedef Tensor<float, 1>::DimensionPair DimPair;
template<int DataLayout> template<int DataLayout>
static void test_cuda_contraction(int m_size, int k_size, int n_size) static void test_cuda_contraction(int m_size, int k_size, int n_size)
{ {
std::cout<<"Calling with ("<<m_size<<","<<k_size<<","<<n_size<<")"<<std::endl; std::cout << "Calling with (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
// with these dimensions, the output has 300 * 140 elements, which is // with these dimensions, the output has 300 * 140 elements, which is
// more than 30 * 1024, which is the number of threads in blocks on // more than 30 * 1024, which is the number of threads in blocks on
// a 15 SM GK110 GPU // a 15 SM GK110 GPU
@ -67,12 +67,16 @@ static void test_cuda_contraction(int m_size, int k_size, int n_size)
t_result = t_left.contract(t_right, dims); t_result = t_left.contract(t_right, dims);
cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost); cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) { for (size_t i = 0; i < t_result.size(); i++) {
if (fabs(t_result.data()[i] - t_result_gpu.data()[i]) >= 1e-4) { if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
std::cout << "mismatch detected at index " << i << ": " << t_result.data()[i] continue;
<< " vs " << t_result_gpu.data()[i] << std::endl;
assert(false);
} }
if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
continue;
}
std::cout << "mismatch detected at index " << i << ": " << t_result(i)
<< " vs " << t_result_gpu(i) << std::endl;
assert(false);
} }
cudaFree((void*)d_t_left); cudaFree((void*)d_t_left);
@ -83,7 +87,7 @@ static void test_cuda_contraction(int m_size, int k_size, int n_size)
void test_cxx11_tensor_cuda() void test_cxx11_tensor_cuda()
{ {
std::cout<<"Calling contraction tests"<<std::endl; std::cout << "Calling contraction tests" << std::endl;
CALL_SUBTEST(test_cuda_contraction<ColMajor>(128, 128, 128)); CALL_SUBTEST(test_cuda_contraction<ColMajor>(128, 128, 128));
CALL_SUBTEST(test_cuda_contraction<RowMajor>(128, 128, 128)); CALL_SUBTEST(test_cuda_contraction<RowMajor>(128, 128, 128));
for (int k = 32; k < 256; k++) { for (int k = 32; k < 256; k++) {