Fix performance regression in ClConv2D
- Call gemm-based convolution when the kernel is large and the stride is
unit
Resolves: COMPMID-5504
Change-Id: I8dd83bd012000ed76824b96a8e37c98c861c59e4
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8081
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Adnan AlSinan <adnan.alsinan@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/gpu/cl/operators/ClConv2d.cpp b/src/gpu/cl/operators/ClConv2d.cpp
index 8119fc8..cd64c8d 100644
--- a/src/gpu/cl/operators/ClConv2d.cpp
+++ b/src/gpu/cl/operators/ClConv2d.cpp
@@ -273,6 +273,11 @@
{
return ConvolutionMethod::WINOGRAD;
}
+
+ if(weights->dimension(idx_w) > 3 && weights->dimension(idx_h) > 3)
+ {
+ return ConvolutionMethod::WINOGRAD;
+ }
}
else
{
@@ -301,7 +306,11 @@
{
if( ((is_large_kernel_sz || is_m_one) && workload_gte_8192) || is_ofm_lte_8 )
{
- return ConvolutionMethod::DIRECT;
+ // Do not use direct convolution when the kernel is large and the stride is unit
+ if(!(is_large_kernel_sz && conv_info.stride().first == 1 && conv_info.stride().second == 1))
+ {
+ return ConvolutionMethod::DIRECT;
+ }
}
}
}