diff options
| author | Frank Lei <frank.lei@arm.com> | 2017-12-05 10:43:33 +0800 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:42:17 +0000 |
| commit | b9d38ee6378f3035f8dbad442223d3d9e2f3dc4f (patch) | |
| tree | 89a4b81430100a4a91902d5987ae42edc438012c /src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp | |
| parent | 397d58aa40b02a26923c34d8cd4ba274eac45963 (diff) | |
| download | ComputeLibrary-b9d38ee6378f3035f8dbad442223d3d9e2f3dc4f.tar.gz | |
APPBROWSER-312 Fully connected performance optimization
Change-Id: Ie93fd630ebbad7b6ca8812cb5044b3f1908b45fd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111830
Reviewed-by: Stephen Li <stephen.li@arm.com>
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Diffstat (limited to 'src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp')
| -rw-r--r-- | src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp | 30 |
1 files changed, 29 insertions, 1 deletions
diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp index a75ab6b609..8179525470 100644 --- a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp @@ -118,9 +118,23 @@ void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTen switch(input0->info()->data_type()) { case DataType::F16: + build_opts.emplace("#define DATA_TYPE_FP16"); + +#define MM_PROCESS_4X_OPTIMIZED + +#if defined(MM_PROCESS_4X) + num_elems_processed_per_iteration_x = 4; + num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4); + build_opts.emplace("#define MM_PROCESS_4X"); +#elif defined(MM_PROCESS_4X_OPTIMIZED) /* MM_PROCESS_4X */ num_elems_processed_per_iteration_x = 4; + num_elems_processed_per_iteration_y = std::min(static_cast<int>(output->info()->dimension(1)), 4); + build_opts.emplace("#define MM_PROCESS_4X_OPTIMIZED"); +#elif defined(MM_PROCESS_8X) /* MM_PROCESS_4X */ + num_elems_processed_per_iteration_x = 8; num_elems_processed_per_iteration_y = 1; - build_opts.emplace("#define DATA_TYPE_FP16"); + build_opts.emplace("#define MM_PROCESS_8X"); +#endif /* MM_PROCESS_4X */ break; case DataType::F32: @@ -143,8 +157,12 @@ void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTen win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); +#if defined(MM_PROCESS_4X_OPTIMIZED) + AccessWindowStatic input0_access(input0->info(), 0, 0, ceil_to_multiple(input0->info()->dimension(0), 8), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y)); +#else /* MM_PROCESS_4X_OPTIMIZED */ AccessWindowStatic input0_access(input0->info(), 0, 0, ceil_to_multiple(input0->info()->dimension(0), num_elems_processed_per_iteration_x), ceil_to_multiple(input0->info()->dimension(1), num_elems_processed_per_iteration_y)); +#endif /* MM_PROCESS_4X_OPTIMIZED */ AccessWindowStatic input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1)); AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); @@ -185,9 +203,19 @@ void GCGEMMMatrixMultiplyKernel::run(const Window &window) switch(_input0->info()->data_type()) { case DataType::F16: +#if defined(MM_PROCESS_4X) add_2D_tensor_argument(idx, _input0, BufferParam(1, 2), slice); add_2D_tensor_argument(idx, _input1, BufferParam(2, 3), slice_b); add_2D_tensor_argument(idx, _output, BufferParam(3, 3), slice); +#elif defined(MM_PROCESS_4X_OPTIMIZED) /* MM_PROCESS_4X */ + add_2D_tensor_argument(idx, _input0, BufferParam(1, 4), slice); + add_2D_tensor_argument(idx, _input1, BufferParam(2, 3), slice_b); + add_2D_tensor_argument(idx, _output, BufferParam(3, 3), slice); +#elif defined(MM_PROCESS_8X) /* MM_PROCESS_4X */ + add_2D_tensor_argument(idx, _input0, BufferParam(1, 4), slice); + add_2D_tensor_argument(idx, _input1, BufferParam(2, 4), slice_b); + add_2D_tensor_argument(idx, _output, BufferParam(3, 4), slice); +#endif /* MM_PROCESS_4X */ break; case DataType::F32: |