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/GCTransposeKernel.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/GCTransposeKernel.cpp')
| -rw-r--r-- | src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp | 32 |
1 files changed, 27 insertions, 5 deletions
diff --git a/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp index 5bd34c2c85..acb998840b 100644 --- a/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp +++ b/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp @@ -64,12 +64,25 @@ void GCTransposeKernel::configure(const IGCTensor *input, IGCTensor *output) build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)); build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)); + // Configure kernel window + unsigned int num_elems_processed_per_iteration = 4; + + if(input->info()->data_type() == DataType::F16) + { +#define TRANSPOSE_8X8 + +#if defined(TRANSPOSE_4X4) + build_opts.emplace(("#define TRANSPOSE_4X4")); + num_elems_processed_per_iteration = 4; +#elif defined(TRANSPOSE_8X8) /* TRANSPOSE_4X4 */ + build_opts.emplace(("#define TRANSPOSE_8X8")); + num_elems_processed_per_iteration = 8; +#endif /* TRANSPOSE_4X4 */ + } + // Create kernel _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("transpose", build_opts)); - // Configure kernel window - const unsigned int num_elems_processed_per_iteration = 4; - Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration, num_elems_processed_per_iteration)); AccessWindowRectangle input_access(input->info(), 0, 0, num_elems_processed_per_iteration, num_elems_processed_per_iteration); @@ -100,8 +113,17 @@ void GCTransposeKernel::run(const Window &window) } else if(_input->info()->data_type() == DataType::F16) { - add_2D_tensor_argument(idx, _input, BufferParam(1, 3), slice); - add_2D_tensor_argument(idx, _output, BufferParam(2, 3), slice); +#if defined(TRANSPOSE_4X4) + BufferParam param = { 1, 3 }; + add_2D_tensor_argument(idx, _input, param, slice); + param.binding_point = 2; + add_2D_tensor_argument(idx, _output, param, slice); +#elif defined(TRANSPOSE_8X8) /* TRANSPOSE_4X4 */ + BufferParam param = { 1, 4 }; + add_2D_tensor_argument(idx, _input, param, slice); + param.binding_point = 2; + add_2D_tensor_argument(idx, _output, param, slice); +#endif /* TRANSPOSE_4X4 */ } _kernel.update_shader_params(); |