mirror of
https://gitlab.com/libeigen/eigen.git
synced 2025-03-07 18:27:40 +08:00
Made the CUDA contract test more robust to numerical noise.
This commit is contained in:
parent
4281eb1e2c
commit
ba27c8a7de
@ -24,7 +24,7 @@ typedef Tensor<float, 1>::DimensionPair DimPair;
|
||||
template<int DataLayout>
|
||||
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
|
||||
// more than 30 * 1024, which is the number of threads in blocks on
|
||||
// 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);
|
||||
|
||||
cudaMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, cudaMemcpyDeviceToHost);
|
||||
for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) {
|
||||
if (fabs(t_result.data()[i] - t_result_gpu.data()[i]) >= 1e-4) {
|
||||
std::cout << "mismatch detected at index " << i << ": " << t_result.data()[i]
|
||||
<< " vs " << t_result_gpu.data()[i] << std::endl;
|
||||
assert(false);
|
||||
for (size_t i = 0; i < t_result.size(); i++) {
|
||||
if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
|
||||
continue;
|
||||
}
|
||||
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);
|
||||
@ -83,7 +87,7 @@ static void test_cuda_contraction(int m_size, int k_size, int n_size)
|
||||
|
||||
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<RowMajor>(128, 128, 128));
|
||||
for (int k = 32; k < 256; k++) {
|
||||
|
Loading…
Reference in New Issue
Block a user