From 7900a9efa421f0ac8faf9bb939f5c536b42c62b0 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Fri, 23 Nov 2018 11:44:58 +0000 Subject: COMPMID-1716: CL Comparison operations Adds support for Equal,NotEqual,Less,LessEqual,Greater,GreaterEqual Change-Id: If0cdf4aae7f95c94709b195eee485f6663f45909 --- src/core/CL/CLKernelLibrary.cpp | 16 +++ src/core/CL/cl_kernels/comparisons.cl | 149 +++++++++++++++++++ src/core/CL/kernels/CLComparisonKernel.cpp | 224 +++++++++++++++++++++++++++++ src/runtime/CL/functions/CLComparison.cpp | 78 ++++++++++ 4 files changed, 467 insertions(+) create mode 100644 src/core/CL/cl_kernels/comparisons.cl create mode 100644 src/core/CL/kernels/CLComparisonKernel.cpp create mode 100644 src/runtime/CL/functions/CLComparison.cpp (limited to 'src') diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 6e5e97e3e1..a9c4074310 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -180,6 +180,18 @@ const std::map CLKernelLibrary::_kernel_program_map = { "channel_extract_YUYV422", "channel_extract.cl" }, { "combine_gradients_L1", "canny.cl" }, { "combine_gradients_L2", "canny.cl" }, + { "compare_equal", "comparisons.cl" }, + { "compare_equal_quantized", "comparisons.cl" }, + { "compare_notequal", "comparisons.cl" }, + { "compare_notequal_quantized", "comparisons.cl" }, + { "compare_greater", "comparisons.cl" }, + { "compare_greater_quantized", "comparisons.cl" }, + { "compare_greaterequal", "comparisons.cl" }, + { "compare_greaterequal_quantized", "comparisons.cl" }, + { "compare_less", "comparisons.cl" }, + { "compare_less_quantized", "comparisons.cl" }, + { "compare_lessequal", "comparisons.cl" }, + { "compare_lessequal_quantized", "comparisons.cl" }, { "concatenate_depth", "concatenate.cl" }, { "concatenate_width", "concatenate.cl" }, { "concatenate_width_x2", "concatenate.cl" }, @@ -536,6 +548,10 @@ const std::map CLKernelLibrary::_program_source_map = { "col2im.cl", #include "./cl_kernels/col2im.clembed" + }, + { + "comparisons.cl", +#include "./cl_kernels/comparisons.clembed" }, { "concatenate.cl", diff --git a/src/core/CL/cl_kernels/comparisons.cl b/src/core/CL/cl_kernels/comparisons.cl new file mode 100644 index 0000000000..8824b136b2 --- /dev/null +++ b/src/core/CL/cl_kernels/comparisons.cl @@ -0,0 +1,149 @@ +/* + * 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 "helpers.h" + +#define EQUAL(x, y) ((x) == (y)) +#define NOTEQUAL(x, y) ((x) != (y)) +#define GREATER(x, y) ((x) > (y)) +#define GREATEREQUAL(x, y) ((x) >= (y)) +#define LESS(x, y) ((x) < (y)) +#define LESSEQUAL(x, y) ((x) <= (y)) + +#define DEFINE_KERNEL_STR(name) compare_##name +#define DEFINE_KERNEL(name) DEFINE_KERNEL_STR(name) + +#define DEFINE_KERNEL_QUANTIZED_STR(name) compare_##name##_quantized +#define DEFINE_KERNEL_QUANTIZED(name) DEFINE_KERNEL_QUANTIZED_STR(name) + +#if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OP) && defined(OP_NAME) +/** This function compares two tensors. + * + * @attention The inputs' data type need to be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float + * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @attention The comparison operation should be given as a preprocessor argument using -DOP=operation. e.g. -DOP=LESS + * + * @param[in] in1_ptr Pointer to the source tensor. Supported data types: U8/S16/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: U8/S16/F16/F32 + * @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: U8 (only if both inputs are U8), S16/F16/F32 + * @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 DEFINE_KERNEL(OP_NAME)( + 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, VEC_SIZE) + in_a = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in1.ptr); + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + in_b = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in2.ptr); + + // Calculate and store result + VSTORE(VEC_SIZE) + (CONVERT(OP(in_a, in_b), VEC_DATA_TYPE(uchar, VEC_SIZE)), 0, (__global uchar *)out.ptr); +} +#endif /* defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OP) && defined(OP_NAME) */ + +#if defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(SCALE_IN1) && defined(SCALE_IN2) +/** This function compares two quantized tensors. + * + * @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10 + * @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10 + * @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10 + * @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10 + * + * @param[in] in1_ptr Pointer to the source tensor. Supported data types: QASYMM8 + * @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 DEFINE_KERNEL_QUANTIZED(OP_NAME)( + 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); + + int16 in_a = CONVERT(vload16(0, (__global uchar *)in1.ptr), int16); + int16 in_b = CONVERT(vload16(0, (__global uchar *)in2.ptr), int16); + + in_a = in_a - (int16)((int)OFFSET_IN1); + in_b = in_b - (int16)((int)OFFSET_IN2); + + const float16 in1f32 = convert_float16(in_a) * (float16)((float)SCALE_IN1); + const float16 in2f32 = convert_float16(in_b) * (float16)((float)SCALE_IN2); + const int16 res = OP(in1f32, in2f32); + + // Store result + vstore16(convert_uchar16(res), 0, (__global uchar *)out.ptr); +} +#endif /* defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(SCALE_IN1) && defined(SCALE_IN2) */ \ No newline at end of file diff --git a/src/core/CL/kernels/CLComparisonKernel.cpp b/src/core/CL/kernels/CLComparisonKernel.cpp new file mode 100644 index 0000000000..f5f5a0fbd6 --- /dev/null +++ b/src/core/CL/kernels/CLComparisonKernel.cpp @@ -0,0 +1,224 @@ +/* + * 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/CLComparisonKernel.h" + +#include "arm_compute/core/CL/CLHelpers.h" +#include "arm_compute/core/CL/CLValidate.h" +#include "arm_compute/core/CL/ICLTensor.h" + +#include + +namespace arm_compute +{ +namespace +{ +// Create supported comparisons map +const std::map supported_comparison_ops = +{ + { ComparisonOperation::Equal, "EQUAL" }, + { ComparisonOperation::NotEqual, "NOTEQUAL" }, + { ComparisonOperation::Greater, "GREATER" }, + { ComparisonOperation::GreaterEqual, "GREATEREQUAL" }, + { ComparisonOperation::Less, "LESS" }, + { ComparisonOperation::LessEqual, "LESSEQUAL" }, +}; + +int calculate_num_elems_processed_per_iteration(const ITensorInfo &input) +{ + return 16 / input.element_size(); +} + +Status validate_arguments(const ITensorInfo &input1, const ITensorInfo &input2, const ITensorInfo &output, ComparisonOperation operation) +{ + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(&input1); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&input1, + 1, + DataType::U8, DataType::S8, DataType::QASYMM8, + DataType::U16, DataType::S16, + DataType::U32, DataType::S32, + DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input1, &input2); + ARM_COMPUTE_RETURN_ERROR_ON(supported_comparison_ops.count(operation) == 0); + + 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_DATA_TYPE_CHANNEL_NOT_IN(&output, 1, DataType::U8); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(detail::have_different_dimensions(out_shape, output.tensor_shape(), 0), + "Wrong shape for output"); + } + + return Status{}; +} + +std::pair validate_and_configure_window(ITensorInfo &input1, ITensorInfo &input2, ITensorInfo &output) +{ + const std::pair broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(input1, input2); + const TensorShape &out_shape = broadcast_pair.first; + const ValidRegion &valid_region = broadcast_pair.second; + + const unsigned int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration(input1); + + // Auto initialize output if not initialized + auto_init_if_empty(output, out_shape, 1, DataType::U8, QuantizationInfo()); + + 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 + +CLComparisonKernel::CLComparisonKernel() + : _input1(nullptr), _input2(nullptr), _output(nullptr) +{ +} + +void CLComparisonKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ComparisonOperation operation) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(*input1->info(), *input2->info(), *output->info(), operation)); + + // 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; + + const std::string &operation_name = supported_comparison_ops.at(operation); + std::string kernel_name = "compare_" + lower_string(operation_name); + + // Set kernel build options + std::set build_opts; + build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input1->info()->data_type())); + build_opts.emplace("-DVEC_SIZE=" + support::cpp11::to_string(calculate_num_elems_processed_per_iteration(*input1->info()))); + build_opts.emplace("-DOP=" + operation_name); + build_opts.emplace("-DOP_NAME=" + lower_string(operation_name)); + if(is_data_type_quantized_asymmetric(input1->info()->data_type())) + { + build_opts.emplace("-DOFFSET_IN1=" + support::cpp11::to_string(input1->info()->quantization_info().offset)); + build_opts.emplace("-DOFFSET_IN2=" + support::cpp11::to_string(input2->info()->quantization_info().offset)); + build_opts.emplace("-DSCALE_IN1=" + float_to_string_with_full_precision(input1->info()->quantization_info().scale)); + build_opts.emplace("-DSCALE_IN2=" + float_to_string_with_full_precision(input2->info()->quantization_info().scale)); + kernel_name += "_quantized"; + } + + // Create kernel + _kernel = static_cast(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); + + ICLKernel::configure_internal(win_config.second); + + // Set config_id for enabling LWS tuning + _config_id = kernel_name; + _config_id += "_"; + _config_id += lower_string(string_from_data_type(input1->info()->data_type())); + _config_id += "_"; + _config_id += support::cpp11::to_string(output->info()->dimension(0)); + _config_id += "_"; + _config_id += support::cpp11::to_string(output->info()->dimension(1)); + _config_id += lower_string(string_from_data_layout(input1->info()->data_layout())); +} + +Status CLComparisonKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ComparisonOperation operation) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); + + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(*input1, *input2, *output, operation)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(*input1->clone(), *input2->clone(), *output->clone()).first); + + return Status{}; +} + +void CLComparisonKernel::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; + const bool is_vector = in_shape1.num_dimensions() == 1 || in_shape2.num_dimensions() == 1; + if(std::min(in_shape1.total_size(), in_shape2.total_size()) > 1 && !is_vector) + { + 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, lws_hint()); + + collapsed.slide_window_slice_3D(slice_input1); + collapsed.slide_window_slice_3D(slice_input2); + } + while(collapsed.slide_window_slice_3D(slice)); +} + +BorderSize CLComparisonKernel::border_size() const +{ + const int num_elems_processed_per_iteration = calculate_num_elems_processed_per_iteration(*_input1->info()); + + const unsigned int replicateSize = _output->info()->dimension(0) - std::min(_input1->info()->dimension(0), _input2->info()->dimension(0)); + const unsigned int border = std::min(num_elems_processed_per_iteration - 1U, replicateSize); + return BorderSize(0, border, 0, 0); +} +} // namespace arm_compute diff --git a/src/runtime/CL/functions/CLComparison.cpp b/src/runtime/CL/functions/CLComparison.cpp new file mode 100644 index 0000000000..0c02962491 --- /dev/null +++ b/src/runtime/CL/functions/CLComparison.cpp @@ -0,0 +1,78 @@ +/* + * 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/runtime/CL/functions/CLComparison.h" + +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/CL/kernels/CLComparisonKernel.h" +#include "arm_compute/core/Types.h" +#include "support/ToolchainSupport.h" + +namespace arm_compute +{ +void CLComparison::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ComparisonOperation operation) +{ + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(input1, input2, output, operation); + _kernel = std::move(k); + + if(output->info()->dimension(0) > 1) + { + ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2; + + if(broadcasted_info->info()->dimension(0) == 1) + { + _border_handler.configure(broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE); + } + } +} + +Status CLComparison::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ComparisonOperation operation) +{ + return CLComparisonKernel::validate(input1, input2, output, operation); +} + +template +void CLComparisonStatic::configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output) +{ + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(input1, input2, output, COP); + _kernel = std::move(k); + + if(output->info()->dimension(0) > 1) + { + ICLTensor *broadcasted_info = (input1->info()->dimension(0) == 1) ? input1 : input2; + + if(broadcasted_info->info()->dimension(0) == 1) + { + _border_handler.configure(broadcasted_info, _kernel->border_size(), BorderMode::REPLICATE); + } + } +} + +template +Status CLComparisonStatic::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) +{ + return CLComparisonKernel::validate(input1, input2, output, COP); +} +} // namespace arm_compute -- cgit v1.2.1