COMPMID-1559 : (Nightly) CLWinograd FP16 SIGABORT and mismatches
Current implementation of winograd fp16 is not accurate enough for large runs.
disabling its use and reopening ticket(COMPMID-1266) to fix it.
The sigbus error that was originally reported against COMPMID-1559 is being tracked as COMPMID-1606
Change-Id: I45129aa366d5710402bc54b623c5fbfb865b3cd5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/151543
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
diff --git a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
index a70389a..70bf3ae 100644
--- a/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLWinogradConvolutionLayer.cpp
@@ -158,6 +158,10 @@
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)
{