diff options
author | Aleksandr Nikolaev <alenik01@dsgdps-gold-1.lab.cambridge.arm.com> | 2021-04-20 16:15:24 +0100 |
---|---|---|
committer | Aleksandr Nikolaev <aleksandr.nikolaev@arm.com> | 2021-04-28 15:40:34 +0000 |
commit | e2535154fa34ac0290ec3daaa44545be0b2b4606 (patch) | |
tree | 523cb4509e69e15fcd12e25d317aadc9bbdd95d4 /src/runtime/CL/functions/CLConvolutionLayer.cpp | |
parent | 660eea4dcb74c37d407a6aa148db1b30123e7d21 (diff) | |
download | ComputeLibrary-e2535154fa34ac0290ec3daaa44545be0b2b4606.tar.gz |
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 <aleksandr.nikolaev@arm.com>
Change-Id: I7fdb7d6ecf9d67926d994a9f666a80aeac3147a5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/VisualCompute/ComputeLibrary/+/317792
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Comments-Addressed: bsgcomp <bsgcomp@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5470
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime/CL/functions/CLConvolutionLayer.cpp')
-rw-r--r-- | src/runtime/CL/functions/CLConvolutionLayer.cpp | 12 |
1 files changed, 12 insertions, 0 deletions
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) |