devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add missing CUDA kernel to tensor scan op

Browse Source 
The TensorScanOp implementation was missing a CUDA kernel launch.
This adds a simple placeholder implementation.
This commit is contained in:
Igor Babuschkin
2016-06-29 11:54:35 +01:00
parent 328c5d876a
commit 85699850d9
3 changed files with 159 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
unsupported/test/CMakeLists.txt
View File

@@ -220,7 +220,7 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
ei_add_test(cxx11_tensor_reduction_cuda)
ei_add_test(cxx11_tensor_argmax_cuda)
ei_add_test(cxx11_tensor_cast_float16_cuda)
# ei_add_test(cxx11_tensor_scan_cuda)
ei_add_test(cxx11_tensor_scan_cuda)
# The random number generation code requires arch 3.5 or greater.
if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 34)

118
unsupported/test/cxx11_tensor_scan.cpp
View File

@@ -14,87 +14,73 @@
using Eigen::Tensor;
template <int DataLayout, typename Type=float>
template <int DataLayout, typename Type=float, bool Exclusive = false>
static void test_1d_scan()
{
int size = 50;
Tensor<Type, 1, DataLayout> tensor(size);
tensor.setRandom();
Tensor<Type, 1, DataLayout> result = tensor.cumsum(0);
int size = 50;
Tensor<Type, 1, DataLayout> tensor(size);
tensor.setRandom();
Tensor<Type, 1, DataLayout> result = tensor.cumsum(0, Exclusive);
VERIFY_IS_EQUAL(tensor.dimension(0), result.dimension(0));
VERIFY_IS_EQUAL(tensor.dimension(0), result.dimension(0));
float accum = 0;
for (int i = 0; i < size; i++) {
float accum = 0;
for (int i = 0; i < size; i++) {
if (Exclusive) {
VERIFY_IS_EQUAL(result(i), accum);
accum += tensor(i);
} else {
accum += tensor(i);
VERIFY_IS_EQUAL(result(i), accum);
}
}
accum = 1;
result = tensor.cumprod(0);
for (int i = 0; i < size; i++) {
accum = 1;
result = tensor.cumprod(0, Exclusive);
for (int i = 0; i < size; i++) {
if (Exclusive) {
VERIFY_IS_EQUAL(result(i), accum);
accum *= tensor(i);
} else {
accum *= tensor(i);
VERIFY_IS_EQUAL(result(i), accum);
}
}
template <int DataLayout, typename Type=float>
static void test_1d_inclusive_scan()
{
int size = 50;
Tensor<Type, 1, DataLayout> tensor(size);
tensor.setRandom();
Tensor<Type, 1, DataLayout> result = tensor.cumsum(0, true);
VERIFY_IS_EQUAL(tensor.dimension(0), result.dimension(0));
float accum = 0;
for (int i = 0; i < size; i++) {
VERIFY_IS_EQUAL(result(i), accum);
accum += tensor(i);
}
accum = 1;
result = tensor.cumprod(0, true);
for (int i = 0; i < size; i++) {
VERIFY_IS_EQUAL(result(i), accum);
accum *= tensor(i);
}
}
}
template <int DataLayout, typename Type=float>
static void test_4d_scan()
{
int size = 5;
Tensor<Type, 4, DataLayout> tensor(size, size, size, size);
tensor.setRandom();
int size = 5;
Tensor<Type, 4, DataLayout> tensor(size, size, size, size);
tensor.setRandom();
Tensor<Type, 4, DataLayout> result(size, size, size, size);
Tensor<Type, 4, DataLayout> result(size, size, size, size);
result = tensor.cumsum(0);
float accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(i, 0, 0, 0);
VERIFY_IS_EQUAL(result(i, 0, 0, 0), accum);
}
result = tensor.cumsum(1);
accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(0, i, 0, 0);
VERIFY_IS_EQUAL(result(0, i, 0, 0), accum);
}
result = tensor.cumsum(2);
accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(0, 0, i, 0);
VERIFY_IS_EQUAL(result(0, 0, i, 0), accum);
}
result = tensor.cumsum(3);
accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(0, 0, 0, i);
VERIFY_IS_EQUAL(result(0, 0, 0, i), accum);
}
result = tensor.cumsum(0);
float accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(i, 1, 2, 3);
VERIFY_IS_EQUAL(result(i, 1, 2, 3), accum);
}
result = tensor.cumsum(1);
accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(1, i, 2, 3);
VERIFY_IS_EQUAL(result(1, i, 2, 3), accum);
}
result = tensor.cumsum(2);
accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(1, 2, i, 3);
VERIFY_IS_EQUAL(result(1, 2, i, 3), accum);
}
result = tensor.cumsum(3);
accum = 0;
for (int i = 0; i < size; i++) {
accum += tensor(1, 2, 3, i);
VERIFY_IS_EQUAL(result(1, 2, 3, i), accum);
}
}
template <int DataLayout>
@@ -113,8 +99,10 @@ static void test_tensor_maps() {
}
void test_cxx11_tensor_scan() {
CALL_SUBTEST(test_1d_scan<ColMajor>());
CALL_SUBTEST(test_1d_scan<RowMajor>());
CALL_SUBTEST((test_1d_scan<ColMajor, float, true>()));
CALL_SUBTEST((test_1d_scan<ColMajor, float, false>()));
CALL_SUBTEST((test_1d_scan<RowMajor, float, true>()));
CALL_SUBTEST((test_1d_scan<RowMajor, float, false>()));
CALL_SUBTEST(test_4d_scan<ColMajor>());
CALL_SUBTEST(test_4d_scan<RowMajor>());
CALL_SUBTEST(test_tensor_maps<ColMajor>());