[GPU]fix kernel select validate in deconv kernel fsv16 for fs=8 input (#11909)

* fix kernel select validate in deconv kernel fsv16 for fs=8 input
* apply the WA for Windows only
* Narrow down condition of ref kernel selection
This commit is contained in:
Wilson Seok 2022-06-20 10:34:45 +09:00 committed by GitHub
parent ce5b2c6a45
commit 477bfa1c09
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -126,6 +126,14 @@ bool DeconvolutionKernel_b_fs_zyx_fsv16::Validate(const Params& p, const optiona
const auto& params = static_cast<const deconvolution_params&>(p);
const auto feature_block_size = 16;
#ifdef _WIN32
// NaN value in non-valid area from input makes an issue when Windows OS. This is WA to avoid the issue.
// The issue happens when input feature size = 8.
// TO-DO: Make the kernel handling this input case
if (params.inputs[0].Feature().v == 8)
return false;
#endif
// Check that padding features doesn't miss-align the blocks
if (params.inputs[0].Feature().pad.before % feature_block_size != 0 || params.outputs[0].Feature().pad.before % feature_block_size != 0)
return false;