aboutsummaryrefslogtreecommitdiffhomepage
path: root/unsupported/test/cxx11_tensor_of_float16_cuda.cu
diff options
context:
space:
mode:
Diffstat (limited to 'unsupported/test/cxx11_tensor_of_float16_cuda.cu')
-rw-r--r--unsupported/test/cxx11_tensor_of_float16_cuda.cu226
1 files changed, 188 insertions, 38 deletions
diff --git a/unsupported/test/cxx11_tensor_of_float16_cuda.cu b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
index cb917bb37..dceac793e 100644
--- a/unsupported/test/cxx11_tensor_of_float16_cuda.cu
+++ b/unsupported/test/cxx11_tensor_of_float16_cuda.cu
@@ -134,6 +134,68 @@ void test_cuda_elementwise() {
gpu_device.deallocate(d_res_float);
}
+void test_cuda_trancendental() {
+ Eigen::CudaStreamDevice stream;
+ Eigen::GpuDevice gpu_device(&stream);
+ int num_elem = 101;
+
+ float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+ float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+ Eigen::half* d_res1_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+ Eigen::half* d_res1_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+ Eigen::half* d_res2_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+ Eigen::half* d_res2_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+
+ Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float1(
+ d_float1, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_float2(
+ d_float2, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_half(
+ d_res1_half, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res1_float(
+ d_res1_float, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_half(
+ d_res2_half, num_elem);
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res2_float(
+ d_res2_float, num_elem);
+
+ gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
+ gpu_float2.device(gpu_device) = gpu_float2.random() + gpu_float1.constant(0.5f);
+ gpu_res1_float.device(gpu_device) = gpu_float1.exp().cast<Eigen::half>();
+ gpu_res2_float.device(gpu_device) = gpu_float2.log().cast<Eigen::half>();
+ gpu_res1_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().exp();
+ gpu_res2_half.device(gpu_device) = gpu_float2.cast<Eigen::half>().log();
+
+ Tensor<float, 1> input1(num_elem);
+ Tensor<Eigen::half, 1> half_prec1(num_elem);
+ Tensor<Eigen::half, 1> full_prec1(num_elem);
+ Tensor<float, 1> input2(num_elem);
+ Tensor<Eigen::half, 1> half_prec2(num_elem);
+ Tensor<Eigen::half, 1> full_prec2(num_elem);
+ gpu_device.memcpyDeviceToHost(input1.data(), d_float1, num_elem*sizeof(float));
+ gpu_device.memcpyDeviceToHost(input2.data(), d_float2, num_elem*sizeof(float));
+ gpu_device.memcpyDeviceToHost(half_prec1.data(), d_res1_half, num_elem*sizeof(Eigen::half));
+ gpu_device.memcpyDeviceToHost(full_prec1.data(), d_res1_float, num_elem*sizeof(Eigen::half));
+ gpu_device.memcpyDeviceToHost(half_prec2.data(), d_res2_half, num_elem*sizeof(Eigen::half));
+ gpu_device.memcpyDeviceToHost(full_prec2.data(), d_res2_float, num_elem*sizeof(Eigen::half));
+ gpu_device.synchronize();
+
+ for (int i = 0; i < num_elem; ++i) {
+ std::cout << "Checking elemwise exp " << i << " input = " << input1(i) << " full = " << full_prec1(i) << " half = " << half_prec1(i) << std::endl;
+ VERIFY_IS_APPROX(full_prec1(i), half_prec1(i));
+ }
+ for (int i = 0; i < num_elem; ++i) {
+ std::cout << "Checking elemwise log " << i << " input = " << input2(i) << " full = " << full_prec2(i) << " half = " << half_prec2(i) << std::endl;
+ VERIFY_IS_APPROX(full_prec2(i), half_prec2(i));
+ }
+ gpu_device.deallocate(d_float1);
+ gpu_device.deallocate(d_float2);
+ gpu_device.deallocate(d_res1_half);
+ gpu_device.deallocate(d_res1_float);
+ gpu_device.deallocate(d_res2_half);
+ gpu_device.deallocate(d_res2_float);
+}
+
void test_cuda_contractions() {
Eigen::CudaStreamDevice stream;
@@ -144,36 +206,38 @@ void test_cuda_contractions() {
float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
float* d_float2 = (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::half* d_res_half = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
+ Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(num_elem * sizeof(Eigen::half));
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
d_float1, rows, cols);
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
d_float2, rows, cols);
- Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_res_half(
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_half(
d_res_half, rows, cols);
- Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_res_float(
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 2>, Eigen::Aligned> gpu_res_float(
d_res_float, rows, cols);
gpu_float1.device(gpu_device) = gpu_float1.random() - gpu_float1.constant(0.5f);
- gpu_float2.device(gpu_device) = gpu_float2.random() - gpu_float1.constant(0.5f);
+ gpu_float2.device(gpu_device) = gpu_float2.random() - gpu_float2.constant(0.5f);
typedef Tensor<float, 2>::DimensionPair DimPair;
Eigen::array<DimPair, 1> dims(DimPair(1, 0));
- gpu_res_float.device(gpu_device) = gpu_float1.contract(gpu_float2, dims);
- gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().contract(gpu_float2.cast<Eigen::half>(), dims).cast<float>();
+ gpu_res_float.device(gpu_device) = gpu_float1.contract(gpu_float2, dims).cast<Eigen::half>();
+ gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().contract(gpu_float2.cast<Eigen::half>(), dims);
- Tensor<float, 2> half_prec(rows, cols);
- Tensor<float, 2> full_prec(rows, cols);
- 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));
+ Tensor<Eigen::half, 2> half_prec(rows, cols);
+ Tensor<Eigen::half, 2> full_prec(rows, cols);
+ gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, num_elem*sizeof(Eigen::half));
+ gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, num_elem*sizeof(Eigen::half));
gpu_device.synchronize();
for (int i = 0; i < rows; ++i) {
for (int j = 0; j < cols; ++j) {
- std::cout << "Checking contract " << i << " " << j << std::endl;
- VERIFY_IS_APPROX(full_prec(i, j), half_prec(i, j));
+ std::cout << "Checking contract " << i << " " << j << full_prec(i, j) << " " << half_prec(i, j) << std::endl;
+ if (numext::abs(full_prec(i, j) - half_prec(i, j)) > Eigen::half(1e-2f)) {
+ VERIFY_IS_APPROX(full_prec(i, j), half_prec(i, j));
+ }
}
}
@@ -192,29 +256,29 @@ void test_cuda_reductions() {
float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
- float* d_res_half = (float*)gpu_device.allocate(size * sizeof(float));
- float* d_res_float = (float*)gpu_device.allocate(size * sizeof(float));
+ Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(size * sizeof(Eigen::half));
+ Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(size * sizeof(Eigen::half));
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
d_float1, size, size);
Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
d_float2, size, size);
- Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_half(
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_half(
d_res_half, size);
- Eigen::TensorMap<Eigen::Tensor<float, 1>, Eigen::Aligned> gpu_res_float(
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 1>, Eigen::Aligned> gpu_res_float(
d_res_float, size);
gpu_float1.device(gpu_device) = gpu_float1.random();
gpu_float2.device(gpu_device) = gpu_float2.random();
Eigen::array<int, 1> redux_dim = {{0}};
- gpu_res_float.device(gpu_device) = gpu_float1.sum(redux_dim);
- gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum(redux_dim).cast<float>();
+ gpu_res_float.device(gpu_device) = gpu_float1.sum(redux_dim).cast<Eigen::half>();
+ gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum(redux_dim);
- Tensor<float, 1> half_prec(size);
- Tensor<float, 1> full_prec(size);
- gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, size*sizeof(float));
- gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, size*sizeof(float));
+ Tensor<Eigen::half, 1> half_prec(size);
+ Tensor<Eigen::half, 1> full_prec(size);
+ gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, size*sizeof(Eigen::half));
+ gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, size*sizeof(Eigen::half));
gpu_device.synchronize();
for (int i = 0; i < size; ++i) {
@@ -222,6 +286,61 @@ void test_cuda_reductions() {
VERIFY_IS_APPROX(full_prec(i), half_prec(i));
}
+ redux_dim = {{1}};
+ gpu_res_float.device(gpu_device) = gpu_float1.sum(redux_dim).cast<Eigen::half>();
+ gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum(redux_dim);
+
+ gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, size*sizeof(Eigen::half));
+ gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, size*sizeof(Eigen::half));
+ gpu_device.synchronize();
+
+ for (int i = 0; i < size; ++i) {
+ std::cout << "Checking redux " << i << std::endl;
+ VERIFY_IS_APPROX(full_prec(i), half_prec(i));
+ }
+
+ gpu_device.deallocate(d_float1);
+ gpu_device.deallocate(d_float2);
+ gpu_device.deallocate(d_res_half);
+ gpu_device.deallocate(d_res_float);
+}
+
+
+
+void test_cuda_full_reductions() {
+ Eigen::CudaStreamDevice stream;
+ Eigen::GpuDevice gpu_device(&stream);
+ int size = 13;
+ int num_elem = size*size;
+
+ float* d_float1 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+ float* d_float2 = (float*)gpu_device.allocate(num_elem * sizeof(float));
+ Eigen::half* d_res_half = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
+ Eigen::half* d_res_float = (Eigen::half*)gpu_device.allocate(1 * sizeof(Eigen::half));
+
+ Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float1(
+ d_float1, size, size);
+ Eigen::TensorMap<Eigen::Tensor<float, 2>, Eigen::Aligned> gpu_float2(
+ d_float2, size, size);
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_half(
+ d_res_half);
+ Eigen::TensorMap<Eigen::Tensor<Eigen::half, 0>, Eigen::Aligned> gpu_res_float(
+ d_res_float);
+
+ gpu_float1.device(gpu_device) = gpu_float1.random();
+ gpu_float2.device(gpu_device) = gpu_float2.random();
+
+ gpu_res_float.device(gpu_device) = gpu_float1.sum().cast<Eigen::half>();
+ gpu_res_half.device(gpu_device) = gpu_float1.cast<Eigen::half>().sum();
+
+ Tensor<Eigen::half, 0> half_prec;
+ Tensor<Eigen::half, 0> full_prec;
+ gpu_device.memcpyDeviceToHost(half_prec.data(), d_res_half, sizeof(Eigen::half));
+ gpu_device.memcpyDeviceToHost(full_prec.data(), d_res_float, sizeof(Eigen::half));
+ gpu_device.synchronize();
+
+ VERIFY_IS_APPROX(full_prec(), half_prec());
+
gpu_device.deallocate(d_float1);
gpu_device.deallocate(d_float2);
gpu_device.deallocate(d_res_half);
@@ -229,27 +348,58 @@ void test_cuda_reductions() {
}
+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
void test_cxx11_tensor_of_float16_cuda()
{
#ifdef EIGEN_HAS_CUDA_FP16
- Eigen::CudaStreamDevice stream;
- Eigen::GpuDevice device(&stream);
- if (device.majorDeviceVersion() > 5 ||
- (device.majorDeviceVersion() == 5 && device.minorDeviceVersion() >= 3)) {
- std::cout << "Running test on device with capability " << device.majorDeviceVersion() << "." << device.minorDeviceVersion() << std::endl;
-
- CALL_SUBTEST_1(test_cuda_conversion());
- CALL_SUBTEST_1(test_cuda_unary());
- CALL_SUBTEST_1(test_cuda_elementwise());
- CALL_SUBTEST_2(test_cuda_contractions());
- CALL_SUBTEST_3(test_cuda_reductions());
- }
- 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;
- }
+ CALL_SUBTEST_1(test_cuda_conversion());
+ CALL_SUBTEST_1(test_cuda_unary());
+ CALL_SUBTEST_1(test_cuda_elementwise());
+ CALL_SUBTEST_1(test_cuda_trancendental());
+ CALL_SUBTEST_2(test_cuda_contractions());
+ CALL_SUBTEST_3(test_cuda_reductions());
+ CALL_SUBTEST_3(test_cuda_full_reductions());
+ CALL_SUBTEST_4(test_cuda_forced_evals());
+
#else
std::cout << "Half floats are not supported by this version of cuda: skipping the test" << std::endl;
#endif
generated by cgit v1.2.3 (git 2.39.1) at 2024-07-03 05:03:55 +0000