diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2018-06-11 16:30:23 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:53:09 +0000 |
commit | 0a887922c73bbe7c5d42b1eb3ae55730f0d9a139 (patch) | |
tree | 3b4908c9ea3490569a9adaca44697a1c9e498c7c /src/core | |
parent | 32af1f8ed8466647abb4f0532c70f72530a1a9ca (diff) | |
download | ComputeLibrary-0a887922c73bbe7c5d42b1eb3ae55730f0d9a139.tar.gz |
COMPMID-1222 Implementing CLArithmeticDivision - FP32 / FP16
Change-Id: I2e3f725ef5ed1454755086b9640ab84a81f4d40e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/135170
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/CL/CLKernelLibrary.cpp | 1 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/arithmetic_op.cl | 56 | ||||
-rw-r--r-- | src/core/CL/kernels/CLArithmeticDivisionKernel.cpp | 185 |
3 files changed, 240 insertions, 2 deletions
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 207efa6aa1..0b2f414c71 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -151,6 +151,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map = { "activation_layer_qa8", "activation_layer_qa8.cl" }, { "arithmetic_add", "arithmetic_op.cl" }, { "arithmetic_sub", "arithmetic_op.cl" }, + { "arithmetic_div", "arithmetic_op.cl" }, { "batchnormalization_layer_nchw", "batchnormalization_layer.cl" }, { "batchnormalization_layer_nhwc", "batchnormalization_layer.cl" }, { "bitwise_or", "bitwise_op.cl" }, diff --git a/src/core/CL/cl_kernels/arithmetic_op.cl b/src/core/CL/cl_kernels/arithmetic_op.cl index 12963473c5..8bd28230b7 100644 --- a/src/core/CL/cl_kernels/arithmetic_op.cl +++ b/src/core/CL/cl_kernels/arithmetic_op.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2016, 2017 ARM Limited. + * Copyright (c) 2016-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -35,6 +35,8 @@ #define SUB(x, y) (x) - (y) #endif /* SATURATE */ +#define DIV(x, y) (x) / (y) + /** This function adds two tensors. * * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT: @@ -86,7 +88,7 @@ __kernel void arithmetic_add( vstore16(ADD(in_a, in_b), 0, (__global DATA_TYPE_OUT *)out.ptr); } -/** This function subtracts one tensors from another. +/** This function subtracts one tensor from another. * * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT: * e.g. -DDATA_TYPE_IN1=uchar -DDATA_TYPE_IN2=uchar -DDATA_TYPE_OUT=short @@ -136,3 +138,53 @@ __kernel void arithmetic_sub( // Calculate and store result vstore16(SUB(in_a, in_b), 0, (__global DATA_TYPE_OUT *)out.ptr); } + +/** This function divides one tensor from another. + * + * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN1, -DDATA_TYPE_IN2 and -DDATA_TYPE_OUT: + * e.g. -DDATA_TYPE_IN1=float -DDATA_TYPE_IN2=float -DDATA_TYPE_OUT=float + * + * @param[in] in1_ptr Pointer to the source tensor. Supported data types: F16/F32 + * @param[in] in1_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] in1_step_x in1_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] in1_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] in1_step_y in1_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] in1_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] in1_step_z in1_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] in1_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[in] in2_ptr Pointer to the source tensor. Supported data types: Same as @p in1_ptr + * @param[in] in2_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] in2_step_x in2_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] in2_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] in2_step_y in2_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] in2_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] in2_step_z in2_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] in2_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] out_ptr Pointer to the destination tensor. Supported data types: Same as @p in1_ptr + * @param[in] out_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] out_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] out_step_y out_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] out_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] out_step_z out_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination tensor + */ +__kernel void arithmetic_div( + TENSOR3D_DECLARATION(in1), + TENSOR3D_DECLARATION(in2), + TENSOR3D_DECLARATION(out)) +{ + // Get pixels pointer + Tensor3D in1 = CONVERT_TO_TENSOR3D_STRUCT(in1); + Tensor3D in2 = CONVERT_TO_TENSOR3D_STRUCT(in2); + Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(out); + + // Load values + VEC_DATA_TYPE(DATA_TYPE_OUT, 16) + in_a = CONVERT(vload16(0, (__global DATA_TYPE_IN1 *)in1.ptr), VEC_DATA_TYPE(DATA_TYPE_OUT, 16)); + VEC_DATA_TYPE(DATA_TYPE_OUT, 16) + in_b = CONVERT(vload16(0, (__global DATA_TYPE_IN2 *)in2.ptr), VEC_DATA_TYPE(DATA_TYPE_OUT, 16)); + + // Calculate and store result + vstore16(DIV(in_a, in_b), 0, (__global DATA_TYPE_OUT *)out.ptr); +} diff --git a/src/core/CL/kernels/CLArithmeticDivisionKernel.cpp b/src/core/CL/kernels/CLArithmeticDivisionKernel.cpp new file mode 100644 index 0000000000..9bd0da15a3 --- /dev/null +++ b/src/core/CL/kernels/CLArithmeticDivisionKernel.cpp @@ -0,0 +1,185 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/CL/kernels/CLArithmeticDivisionKernel.h" + +#include "arm_compute/core/CL/CLHelpers.h" +#include "arm_compute/core/CL/CLValidate.h" +#include "arm_compute/core/CL/ICLTensor.h" + +using namespace arm_compute; + +namespace +{ +constexpr unsigned int num_elems_processed_per_iteration = 16; + +Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2); + + const TensorShape out_shape = TensorShape::broadcast_shape(input1->tensor_shape(), input2->tensor_shape()); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(out_shape.total_size() == 0, "Inputs are not broadcast compatible"); + + // Validate in case of configured output + if(output->total_size() > 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, output); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output->tensor_shape(), 0), + "Wrong shape for output"); + } + + return Status{}; +} + +std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output) +{ + const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*input1, *input2); + const TensorShape &out_shape = broadcast_pair.first; + const ValidRegion &valid_region = broadcast_pair.second; + + // Auto initialize output if not initialized + { + set_shape_if_empty(*output, out_shape); + + if(input1->data_type() == DataType::F16 && input2->data_type() == DataType::F16) + { + set_format_if_unknown(*output, Format::F16); + } + else if(input1->data_type() == DataType::F32 || input2->data_type() == DataType::F32) + { + set_format_if_unknown(*output, Format::F32); + } + } + + Window win = calculate_max_window(valid_region, Steps(num_elems_processed_per_iteration)); + Window win_input1 = win.broadcast_if_dimension_le_one(*input1); + Window win_input2 = win.broadcast_if_dimension_le_one(*input2); + + AccessWindowHorizontal input1_access(input1, 0, num_elems_processed_per_iteration); + AccessWindowHorizontal input2_access(input2, 0, num_elems_processed_per_iteration); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); + + bool window_changed = update_window_and_padding(win_input1, input1_access) + || update_window_and_padding(win_input2, input2_access) + || update_window_and_padding(win, output_access); + + output_access.set_valid_region(win, valid_region); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} +} // namespace + +CLArithmeticDivisionKernel::CLArithmeticDivisionKernel() + : _input1(nullptr), _input2(nullptr), _output(nullptr) +{ +} + +void CLArithmeticDivisionKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info())); + + // Configure kernel window + auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info()); + ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + + _input1 = input1; + _input2 = input2; + _output = output; + + // Set kernel build options + std::set<std::string> build_opts; + build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); + build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); + build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); + + // Create kernel + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_div", build_opts)); + + ICLKernel::configure(win_config.second); +} + +Status CLArithmeticDivisionKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get()).first); + + return Status{}; +} + +void CLArithmeticDivisionKernel::run(const Window &window, cl::CommandQueue &queue) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + + const TensorShape &in_shape1 = _input1->info()->tensor_shape(); + const TensorShape &in_shape2 = _input2->info()->tensor_shape(); + const TensorShape &out_shape = _output->info()->tensor_shape(); + + bool can_collapse = true; + if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1) + { + can_collapse = (std::min(in_shape1.num_dimensions(), in_shape2.num_dimensions()) > Window::DimZ); + for(size_t d = Window::DimZ; can_collapse && (d < out_shape.num_dimensions()); d++) + { + can_collapse = (in_shape1[d] == in_shape2[d]); + } + } + + bool has_collapsed = false; + Window collapsed = can_collapse ? window.collapse_if_possible(ICLKernel::window(), Window::DimZ, &has_collapsed) : window; + + const TensorShape &in_shape1_collapsed = has_collapsed ? in_shape1.collapsed_from(Window::DimZ) : in_shape1; + const TensorShape &in_shape2_collapsed = has_collapsed ? in_shape2.collapsed_from(Window::DimZ) : in_shape2; + + Window slice = collapsed.first_slice_window_3D(); + Window slice_input1 = slice.broadcast_if_dimension_le_one(in_shape1_collapsed); + Window slice_input2 = slice.broadcast_if_dimension_le_one(in_shape2_collapsed); + + do + { + unsigned int idx = 0; + + add_3D_tensor_argument(idx, _input1, slice_input1); + add_3D_tensor_argument(idx, _input2, slice_input2); + add_3D_tensor_argument(idx, _output, slice); + + enqueue(queue, *this, slice); + + collapsed.slide_window_slice_3D(slice_input1); + collapsed.slide_window_slice_3D(slice_input2); + } + while(collapsed.slide_window_slice_3D(slice)); +} + +BorderSize CLArithmeticDivisionKernel::border_size() const +{ + const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0)); + const unsigned int border = std::min<unsigned int>(num_elems_processed_per_iteration - 1U, replicateSize); + return BorderSize(0, border, 0, 0); +} |