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/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.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/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp | 61 |
1 files changed, 18 insertions, 43 deletions
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp index d403d67173..38e0474dde 100644 --- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp @@ -42,7 +42,7 @@ namespace arm_compute namespace { Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, - int min, int max, unsigned int output_3d_depth) + int min, int max) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32); ARM_COMPUTE_RETURN_ERROR_ON(max > 255); @@ -58,10 +58,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con if(output->total_size() != 0) { - const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_output_stage_shape(*input, output_3d_depth, true); - const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(output_shape); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); } return Status{}; @@ -69,7 +67,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output) { - constexpr unsigned int num_elems_processed_per_iteration = 16; + constexpr unsigned int num_elems_processed_per_iteration = 4; // Configure kernel window Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); @@ -103,15 +101,15 @@ class Coordinates; } // namespace arm_compute CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel() - : _input(nullptr), _bias(nullptr), _output(nullptr), _reinterpret_as_3d(false) + : _input(nullptr), _bias(nullptr), _output(nullptr) { } Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, - int min, int max, unsigned int output_3d_depth) + int min, int max) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max, output_3d_depth)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max)); ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (bias != nullptr) ? bias->clone().get() : nullptr, output->clone().get()) @@ -122,22 +120,20 @@ Status CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, - int min, int max, unsigned int output_3d_depth) + int min, int max) { // Perform validate step ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Output auto inizialitation if not yet initialized - const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_output_stage_shape(*input->info(), output_3d_depth, true); - auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(DataType::QASYMM8).set_tensor_shape(output_shape)); + auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(DataType::QASYMM8)); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), - min, max, output_3d_depth)); + min, max)); - _input = input; - _bias = bias; - _output = output; - _reinterpret_as_3d = output_3d_depth > 1; + _input = input; + _bias = bias; + _output = output; // Set the arguments to pass at compile time CLBuildOptions build_opts; @@ -147,7 +143,6 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min)); build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max)); build_opts.add_option_if(bias != nullptr, "-DADD_BIAS"); - build_opts.add_option_if(_reinterpret_as_3d, "-DDST_HEIGHT=" + support::cpp11::to_string(input->info()->tensor_shape().y() / output_3d_depth)); // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down_fixedpoint", build_opts.options())); @@ -177,32 +172,12 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window add_1D_tensor_argument(idx1, _bias, biases_slice); } - if(_reinterpret_as_3d) + do { - // Create output window - Window window_out; - window_out.use_tensor_dimensions(_output->info()->tensor_shape()); - Window collapsed_out = window_out.collapse_if_possible(window_out, 3); - Window slice_out = collapsed.first_slice_window_4D(); - - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, _input, slice); - add_4D_tensor_argument(idx1, _output, slice_out); - enqueue(queue, *this, slice); - } - while(collapsed.slide_window_slice_3D(slice) && collapsed_out.slide_window_slice_4D(slice_out)); - } - else - { - do - { - unsigned int idx = 0; - add_3D_tensor_argument(idx, _input, slice); - add_3D_tensor_argument(idx1, _output, slice); - enqueue(queue, *this, slice); - } - while(collapsed.slide_window_slice_3D(slice)); + unsigned int idx = 0; + add_3D_tensor_argument(idx, _input, slice); + add_3D_tensor_argument(idx1, _output, slice); + enqueue(queue, *this, slice); } + while(collapsed.slide_window_slice_3D(slice)); } |