diff options
author | Gian Marco <gianmarco.iodice@arm.com> | 2018-01-30 13:35:54 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:47:18 +0000 |
commit | 19835e591cb0b66a0f5000ae1505bf299e50337d (patch) | |
tree | 525ee8b233a2cefe3b2734d76fdb91093b8c2d50 /src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp | |
parent | 6fa009e05ae32e64f397f54087885c3eb68f0b4b (diff) | |
download | ComputeLibrary-19835e591cb0b66a0f5000ae1505bf299e50337d.tar.gz |
COMPMID-882 - Optimizing GEMMLowp on OpenCL reshaping matrices
This new optimization allows to achieve 36.3 % of MAC utilisation on Mate 9 @ 1GHz.
The performance have been reported here
https://confluence.arm.com/display/MLENG/GEMMLowp+performance%3A+ACL+18.02
Change-Id: I71b6a217068763dfdc11bbf3574ee0eb94f93679
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118531
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp | 12 |
1 files changed, 10 insertions, 2 deletions
diff --git a/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp b/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp index d05939fcf5..221a1566b9 100644 --- a/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -91,7 +91,7 @@ Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vecto std::pair<Status, Window> validate_and_configure_window(ITensorInfo *mm_result, ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, int32_t a_offset, int32_t b_offset) { - constexpr unsigned int num_elems_processed_per_iteration = 16; + constexpr unsigned int num_elems_processed_per_iteration = 4; bool window_changed = false; // Configure kernel window @@ -160,6 +160,14 @@ void CLGEMMLowpOffsetContributionKernel::configure(ICLTensor *mm_result, const I a_offset, b_offset); // NOLINT ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICLKernel::configure(win_config.second); + + // Set config_id for enabling LWS tuning + _config_id = "gemmlowp_offset_contribution_"; + _config_id += support::cpp11::to_string(mm_result->info()->dimension(0)); + _config_id += "_"; + _config_id += support::cpp11::to_string(mm_result->info()->dimension(1)); + _config_id += "_"; + _config_id += support::cpp11::to_string(mm_result->info()->dimension(2)); } Status CLGEMMLowpOffsetContributionKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, |