diff options
| author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-10-18 10:21:02 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:55:45 +0000 |
| commit | 4b90865ab985d571f70c60583cdfb8c7a65f1670 (patch) | |
| tree | f116a4ffef5f5e823689dd00c1e5c9d987f3d295 /src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp | |
| parent | c55beee7ef70fa08a5d217619083b288a74fcb27 (diff) | |
| download | ComputeLibrary-4b90865ab985d571f70c60583cdfb8c7a65f1670.tar.gz | |
COMPMID-1413 - Improve the performance of GEMMLowp with 8 bit dot product on OpenCL
COMPMID-1424 - Add dot product support for CLDepthwise QASYMM8 3x3 NHWC non-unit stride
With this patch we are able to improve the performance of MobileNet v1-qasymm8 by 37 %
Tried to use the dot product instruction in CLDepthwise QASYMM8 3x3 NHWC non-unit stride
but I have not seen any benefit (maybe because we have few arithemtic operation and we
do not have more load instructions). However Depthwise convolution has been improved by
30%
Change-Id: Id768a99c2e53a04276707e427af5d0ec93419ada
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/155082
Tested-by: bsgcomp <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp | 9 |
1 files changed, 7 insertions, 2 deletions
diff --git a/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp index 9cf5d1fb6a..225c358b20 100644 --- a/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp @@ -24,6 +24,7 @@ #include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h" #include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/CL/CLHelpers.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" @@ -59,7 +60,7 @@ std::pair<Status, Window> validate_and_configure_window_matrix_a_reduction(ITens Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); - AccessWindowStatic input_access(input, 0, 0, ceil_to_multiple(input->dimension(0), 16), input->dimension(1)); + AccessWindowStatic input_access(input, 0, 0, input->dimension(0), input->dimension(1)); AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); bool window_changed = update_window_and_padding(win, input_access, output_access); @@ -115,8 +116,12 @@ void CLGEMMLowpMatrixAReductionKernel::configure(const ICLTensor *mtx_a, ICLTens CLBuildOptions build_opts; build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(mtx_a->info()->dimension(0))); + const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device()); + + std::string kernel_name = "gemmlowp_matrix_a_reduction" + std::string(is_dot8_supported ? "_dot8" : ""); + // Create kernel - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_matrix_a_reduction", build_opts.options())); + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options())); // Configure kernel window auto win_config = validate_and_configure_window_matrix_a_reduction(_input->info(), _output->info()); |