[IE CLDNN] Disable bs_fs_yx_bsv16_fsv16 format for int8 convolution

This commit is contained in:
Roman Lyamin 2021-10-20 17:47:47 +03:00 committed by myshevts
parent 4ddf63ab5f
commit 5125c66179

View File

@ -819,10 +819,7 @@ layout layout_optimizer::get_expected_layout(layout const& current_layout,
} else {
/* *************************** Native impls format selection part ************************** */
if (i8_u8_input) {
if ((_optimization_attributes.bs_fs_yx_bsv16_fsv16_network && expected_tensor.batch[0] % 16 == 0 &&
convolution_bs_fs_yx_bsv16_fsv16_opt(input_layout, output_layout, weights_layout, prim))) {
expected_format = cldnn::format::bs_fs_yx_bsv16_fsv16;
} else if ((_optimization_attributes.b_fs_yx_fsv16_network &&
if ((_optimization_attributes.b_fs_yx_fsv16_network &&
convolution_b_fs_yx_fsv16_opt(input_layout, output_layout, weights_layout, prim))) {
expected_format = cldnn::format::b_fs_yx_fsv16;
} else if ((_optimization_attributes.b_fs_zyx_fsv16_network &&