diff options
author | Sangwon Ha <sangwon.ha@arm.com> | 2023-10-04 00:35:22 +0100 |
---|---|---|
committer | Sang Won Ha <sangwon.ha@arm.com> | 2023-10-09 18:29:04 +0000 |
commit | 1434155d8d937c5a07af67439f8241064f841ad8 (patch) | |
tree | 0c67d04cf62b7bb0d52201c129ed2905d1feb019 /src/runtime/CL | |
parent | 68b6dcebf90f0a9d22ba84682d7591fc8923213f (diff) | |
download | ComputeLibrary-1434155d8d937c5a07af67439f8241064f841ad8.tar.gz |
Change heuristics for FP16 Deconv
- For FP16, disable direct Deconv method when oiutput channel count is
greater than 32.
Resolves COMPMID-6311
Change-Id: I14d9dbf1a1b95736ccd09488d633df4775a01dcb
Signed-off-by: Sangwon Ha <sangwon.ha@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10446
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime/CL')
-rw-r--r-- | src/runtime/CL/functions/CLDeconvolutionLayer.cpp | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp index e988ab0ac4..4e0d1501ba 100644 --- a/src/runtime/CL/functions/CLDeconvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLDeconvolutionLayer.cpp @@ -178,9 +178,9 @@ DeconvolutionMethod CLDeconvolutionLayer::get_deconvolution_method(const ITensor if (weights->dimension(idx_w) != deconv_info.stride().first || weights->dimension(idx_h) != deconv_info.stride().second) { - // We observe better performance for FP32 types only when ofm <= 16. - // A better heuristic is required for selecting the method for FP16 data types. - if (input->data_layout() == DataLayout::NHWC && !((input->data_type() == DataType::F32) && (ofm > 16))) + // We observe better performance for FP32 types only when ofm <= 16, and for FP16 only when ofm <= 32. + if (input->data_layout() == DataLayout::NHWC && !((input->data_type() == DataType::F32) && (ofm > 16)) && + !((input->data_type() == DataType::F16) && (ofm > 32))) { return DeconvolutionMethod::DIRECT; } |