From e2535154fa34ac0290ec3daaa44545be0b2b4606 Mon Sep 17 00:00:00 2001 From: Aleksandr Nikolaev Date: Tue, 20 Apr 2021 16:15:24 +0100 Subject: Fix for performance regression on G71 (mate9) for direct convolution This variant with GEMM solves regression problem only partially, but guarantees precision. Resolves: COMPMID-4387 Signed-off-by: Aleksandr Nikolaev Change-Id: I7fdb7d6ecf9d67926d994a9f666a80aeac3147a5 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/317792 Tested-by: bsgcomp Reviewed-by: Michele DiGiorgio Comments-Addressed: bsgcomp Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5470 Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins --- src/runtime/CL/functions/CLConvolutionLayer.cpp | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp index ac18b966af..aaabbe0cfc 100644 --- a/src/runtime/CL/functions/CLConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp @@ -226,6 +226,18 @@ ConvolutionMethod CLConvolutionLayer::get_convolution_method(const ITensorInfo * } if((weights->dimension(idx_h) > 7) && (input->dimension(idx_c) >= output->dimension(idx_c)) && (CLDirectConvolutionLayer::validate(input, weights, nullptr, output, conv_info, act_info))) { + if(gpu_target == GPUTarget::G71) + { + if(CLFFTConvolutionLayer::validate(input, weights, nullptr, output, conv_info, act_info, enable_fast_math)) + { + return ConvolutionMethod::FFT; + } + else + { + return ConvolutionMethod::GEMM; + } + } + return ConvolutionMethod::DIRECT; } if(input->dimension(idx_c) < 16) -- cgit v1.2.1