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

[IE CLDNN] Enable int8 activation for fsv16 format (#516)

Browse Source 
This change enables int8/uint8 standalone activation to use optimized
block format (b_fs_yx_fsv16). This should eliminate cases where such
activation had reorders before and after.

Support for this is already provided by activation_kernel_ref implementation.

Related JIRA: CVS-28494
This commit is contained in:
Konrad Dobros 2020-05-27 04:37:38 +02:00 committed by GitHub
parent 6788153ba9
commit d3ea03bbfc
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
inference-engine/thirdparty/clDNN/src/gpu/activation_gpu.cpp vendored
View File

@ -103,6 +103,8 @@ attach_activation_gpu::attach_activation_gpu() {
// block f16 format
{std::make_tuple(engine_types::ocl, data_types::f16, format::b_fs_yx_fsv16), val_fw},
{std::make_tuple(engine_types::ocl, data_types::f32, format::b_fs_yx_fsv16), val_fw},
{std::make_tuple(engine_types::ocl, data_types::i8, format::b_fs_yx_fsv16), val_fw},
{std::make_tuple(engine_types::ocl, data_types::u8, format::b_fs_yx_fsv16), val_fw},
// 3D
{std::make_tuple(engine_types::ocl, data_types::f32, format::bfzyx), val_fw},
{std::make_tuple(engine_types::ocl, data_types::f16, format::bfzyx), val_fw},