diff options
Diffstat (limited to 'src/runtime/CL')
-rw-r--r-- | src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp | 4 |
1 files changed, 4 insertions, 0 deletions
diff --git a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp index a70389ab6c..70bf3ae593 100644 --- a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp @@ -158,6 +158,10 @@ Status CLWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITen const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->data_layout()); + //FP16 implementation of winograd is slower than direct convolution. + //The following check needs to be removed when fp16 winograd is faster than direct convolution (COMPMID-1266) + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + // Check if the Winograd configuration requires fast math if(!enable_fast_math) { |