diff options
author | Gian Marco <gianmarco.iodice@arm.com> | 2018-01-10 15:56:30 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:43:42 +0000 |
commit | 7b4d547800d3ea49e7e6d9f497ec2766411cb948 (patch) | |
tree | efff27f22ed692314690a666dc4c0e23838b78c5 /src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp | |
parent | dde3ad94dc11fae29dba862a1ad657f551f36763 (diff) | |
download | ComputeLibrary-7b4d547800d3ea49e7e6d9f497ec2766411cb948.tar.gz |
COMPMID-816 - Optimizing CLGEMMLowpMatrixMultiplyCore - Part1
The performance improvements have been reported at the following
confluence page:
https://confluence.arm.com/display/MLENG/GEMMLowp+performance%3A+ACL+18.02
Config3 of McVail looks improved by 29x
Change-Id: I8b203c0b75fc368f85cea863b7eed398fab3e79a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/115783
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp | 12 |
1 files changed, 8 insertions, 4 deletions
diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp index 423592b79c..2f96724210 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h" #include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/CLKernelLibrary.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/CL/OpenCL.h" @@ -94,8 +95,8 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITe else { // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor. num_elems_processed_per_iteration_x - num_elems_processed_per_iteration_x = 16; - num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 4); + num_elems_processed_per_iteration_x = 4; + num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->dimension(1)), 5); // Configure window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); @@ -132,6 +133,9 @@ void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const IC ElementsProcessed num_elements_processed{}; + // Get target architecture + GPUTarget arch_target = get_arch_from_target(get_target()); + // Configure kernel window auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info(), is_interleaved_transposed, num_elements_processed); ARM_COMPUTE_ERROR_THROW_ON(win_config.first); @@ -150,7 +154,7 @@ void CLGEMMLowpMatrixMultiplyKernel::configure(const ICLTensor *input0, const IC build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(input0->info()->dimension(0))); build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_X=" + support::cpp11::to_string(num_elements_processed.x())); build_opts.add_option("-DNUM_ELEMS_PROCESSED_PER_THREAD_Y=" + support::cpp11::to_string(num_elements_processed.y())); - kernel_name = "gemmlowp_mm"; + kernel_name = "gemmlowp_mm_" + string_from_target(arch_target); } // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); |