IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Fixed pooling_imad kernel error and expanded corresponding tests (#8527)

Browse Source
This commit is contained in:
Ilya Znamenskiy 2021-11-12 22:19:58 +03:00 committed by GitHub
parent b305a01e21
commit 57eff5c287
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 31 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

39
inference-engine/thirdparty/clDNN/kernel_selector/core/cl_kernels/pooling_gpu_b_fs_zyx_fsv16_imad.cl vendored
View File

@ -126,26 +126,45 @@ KERNEL(pooling_gpu_b_fs_zyx_fsv16)(
ACTIVATION_VEC16 pool_result;
#if defined AVG_POOLING
#if INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE != 0
__attribute__((opencl_unroll_hint(INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE)))
for (uint i = 0; i < INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE; i++) {
#else
__attribute__((opencl_unroll_hint(FEATURE_SLICE_SIZE)))
for (uint i = 0; i < FEATURE_SLICE_SIZE; i++) {
#endif
if (last_in_f_group) {
__attribute__((opencl_unroll_hint(INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE)))
for (uint i = 0; i < INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE; i++) {
#if ENABLE_ROUND
#if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
pool_result[i] = convert_int(round(((float)result[i] / max(num_elements, (uint)1))));
pool_result[i] = convert_int(round(((float)result[i] / max(num_elements, (uint)1))));
#else
pool_result[i] = convert_int(round((float)result[i] / (int)(POOL_SIZE_X * INPUT0_SIZE_Z * INPUT0_SIZE_Y)));
pool_result[i] = convert_int(round((float)result[i] / (int)(POOL_SIZE_X * INPUT0_SIZE_Z * INPUT0_SIZE_Y)));
#endif
#else // ENABLE_ROUND
#if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
pool_result[i] = (float)result[i] / max(num_elements, (uint)1);
pool_result[i] = (float)result[i] / max(num_elements, (uint)1);
#else
pool_result[i] = (float)result[i] / (int)(POOL_SIZE_X * INPUT0_SIZE_Z * INPUT0_SIZE_Y);
pool_result[i] = (float)result[i] / (int)(POOL_SIZE_X * INPUT0_SIZE_Z * INPUT0_SIZE_Y);
#endif
#endif // ENABLE_ROUND
}
} else {
#endif
__attribute__((opencl_unroll_hint(FEATURE_SLICE_SIZE)))
for (uint i = 0; i < FEATURE_SLICE_SIZE; i++) {
// INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE != 0
#if ENABLE_ROUND
#if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
pool_result[i] = convert_int(round(((float)result[i] / max(num_elements, (uint)1))));
#else
pool_result[i] = convert_int(round((float)result[i] / (int)(POOL_SIZE_X * INPUT0_SIZE_Z * INPUT0_SIZE_Y)));
#endif
#else // ENABLE_ROUND
#if defined(DYNAMIC_KERNEL_DIVIDER) || defined(DYNAMIC_WITH_PADDING_KERNEL_DIVIDER)
pool_result[i] = (float)result[i] / max(num_elements, (uint)1);
#else
pool_result[i] = (float)result[i] / (int)(POOL_SIZE_X * INPUT0_SIZE_Z * INPUT0_SIZE_Y);
#endif
#endif // ENABLE_ROUND
}
#if INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE != 0
}
#endif // INPUT0_FEATURE_NUM % FEATURE_SLICE_SIZE != 0
#else // AVG_POOLING
pool_result = TO_ACTIVATION_VEC16(result);
#endif // AVG_POOLING

4
inference-engine/thirdparty/clDNN/tests/test_cases/pooling_gpu_test.cpp vendored
View File

@ -2719,7 +2719,7 @@ INSTANTIATE_TEST_SUITE_P(
smoke_low_precision_2d_spatial,
pooling_random_test,
testing::Combine(testing::Values(1, 2),
testing::Values(3, 8, 64),
testing::Values(3, 8, 19, 64),
testing::Values(std::tuple<size_t, size_t, size_t>(12, 12, 1), std::tuple<size_t, size_t, size_t>(24, 24, 1)),
testing::Values(std::tuple<size_t, size_t, size_t>(4, 4, 1), std::tuple<size_t, size_t, size_t>(2, 2, 1)),
testing::Values(std::tuple<int, int, int>(2, 2, 1)),
@ -2736,7 +2736,7 @@ INSTANTIATE_TEST_SUITE_P(
smoke_low_precision_3d_spatial,
pooling_random_test,
testing::Combine(testing::Values(1, 2),
testing::Values(3, 8, 64),
testing::Values(3, 8, 27, 64),
testing::Values(std::tuple<size_t, size_t, size_t>(12, 12, 12), std::tuple<size_t, size_t, size_t>(24, 24, 24)),
testing::Values(std::tuple<size_t, size_t, size_t>(4, 4, 4), std::tuple<size_t, size_t, size_t>(2, 2, 2)),
testing::Values(std::tuple<int, int, int>(2, 2, 2)),