mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-01-30 17:40:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fixed the partial evaluation of non vectorizable tensor subexpressions

Browse Source
This commit is contained in:
Benoit Steiner 2016-04-25 10:43:03 -07:00
parent fd9401f260
commit 4a164d2c46
3 changed files with 39 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
View File

@ -95,7 +95,7 @@ struct TensorEvaluator<const TensorEvalToOp<ArgType>, Device>
enum {
IsAligned = true,
PacketAccess = true,
PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
Layout = TensorEvaluator<ArgType, Device>::Layout,
CoordAccess = false, // to be implemented
RawAccess = true

2
unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
View File

@ -90,7 +90,7 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
enum {
IsAligned = true,
PacketAccess = (internal::packet_traits<Scalar>::size > 1),
PacketAccess = (PacketSize > 1),
Layout = TensorEvaluator<ArgType, Device>::Layout,
RawAccess = true
};

37
unsupported/test/cxx11_tensor_of_float16_cuda.cu
View File

@ -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;