diff options
author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2017-07-13 11:03:35 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-09-17 14:16:42 +0100 |
commit | 7281834f8554670c9d8ae59069a4d80adec21fa7 (patch) | |
tree | 7ae65d26e86fd0f6f0067e630cef4353c54b3e3f | |
parent | e222941f931b1d44bb29e5827b6df748e60cefc4 (diff) | |
download | ComputeLibrary-7281834f8554670c9d8ae59069a4d80adec21fa7.tar.gz |
COMPMID-446: Add support for QS8/QS16 CL Arithmetic Add/Sub
Change-Id: I84fc457a9c28856a11322944822d2fabaf92e8e4
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/80528
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>
9 files changed, 638 insertions, 22 deletions
diff --git a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h index 7d736cdf44..e03f298a3c 100644 --- a/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h +++ b/arm_compute/core/CL/kernels/CLArithmeticAdditionKernel.h @@ -53,9 +53,9 @@ public: ~CLArithmeticAdditionKernel() = default; /** Initialise the kernel's inputs, output and convertion policy. * - * @param[in] input1 First tensor input. Data types supported: U8, S16, F16, F32. - * @param[in] input2 Second tensor input. Data types supported: U8, S16, F16, F32. - * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), S16, F16, F32. + * @param[in] input1 First tensor input. Data types supported: U8/QS8/QS16/S16/F16/F32. + * @param[in] input2 Second tensor input. Data types supported: U8/QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32. + * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32. * @param[in] policy Policy to use to handle overflow. */ void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy); diff --git a/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h b/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h index afecf6ed7d..0765d0f520 100644 --- a/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h +++ b/arm_compute/core/CL/kernels/CLArithmeticSubtractionKernel.h @@ -55,9 +55,9 @@ public: /** Initialise the kernel's inputs, output and convertion policy. * - * @param[in] input1 First tensor input. Data types supported: U8, S16, F16, F32. - * @param[in] input2 Second tensor input. Data types supported: U8, S16, F16, F32. - * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), S16, F16, F32. + * @param[in] input1 First tensor input. Data types supported: U8/QS8/QS16/S16/F16/F32. + * @param[in] input2 Second tensor input. Data types supported: U8/QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32. + * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32. * @param[in] policy Policy to use to handle overflow. */ void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy); diff --git a/arm_compute/runtime/CL/functions/CLArithmeticAddition.h b/arm_compute/runtime/CL/functions/CLArithmeticAddition.h index feadf39820..f888256b33 100644 --- a/arm_compute/runtime/CL/functions/CLArithmeticAddition.h +++ b/arm_compute/runtime/CL/functions/CLArithmeticAddition.h @@ -33,7 +33,7 @@ class ICLTensor; /** Basic function to run @ref CLArithmeticAdditionKernel * - * @note The tensor data type for the inputs must be U8, S16, F16, F32. + * @note The tensor data type for the inputs must be U8/QS8/QS16/S16/F16/F32. * @note The function performs an arithmetic addition between two tensors. */ class CLArithmeticAddition : public ICLSimpleFunction @@ -41,9 +41,9 @@ class CLArithmeticAddition : public ICLSimpleFunction public: /** Initialise the kernel's inputs, output and convertion policy. * - * @param[in] input1 First tensor input. Data types supported: U8, S16, F16, F32. - * @param[in] input2 Second tensor input. Data types supported: U8, S16, F16, F32. - * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), S16, F16, F32. + * @param[in] input1 First tensor input. Data types supported: U8/QS8/QS16/S16/F16/F32. + * @param[in] input2 Second tensor input. Data types supported: U8, QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32. + * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32. * @param[in] policy Policy to use to handle overflow. */ void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy); diff --git a/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h b/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h index d7bb21144e..0a3b00f980 100644 --- a/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h +++ b/arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h @@ -42,9 +42,9 @@ class CLArithmeticSubtraction : public ICLSimpleFunction public: /** Initialise the kernel's inputs, output and convertion policy. * - * @param[in] input1 First tensor input. Data types supported: U8, S16, F16, F32. - * @param[in] input2 Second tensor input. Data types supported: U8, S16, F16, F32. - * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), S16, F16, F32. + * @param[in] input1 First tensor input. Data types supported: U8/QS8/QS16/S16/F16/F32. + * @param[in] input2 Second tensor input. Data types supported: U8/QS8 (only if @p input1 is QS8), QS16 (only if @p input1 is QS16), S16/F16/F32. + * @param[out] output Output tensor. Data types supported: U8 (Only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32. * @param[in] policy Policy to use to handle overflow. */ void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy); diff --git a/src/core/CL/cl_kernels/arithmetic_op.cl b/src/core/CL/cl_kernels/arithmetic_op.cl index 5102d34552..03414105e6 100644 --- a/src/core/CL/cl_kernels/arithmetic_op.cl +++ b/src/core/CL/cl_kernels/arithmetic_op.cl @@ -23,6 +23,10 @@ */ #include "helpers.h" +#if defined(FIXED_POINT_POSITION) +#include "fixed_point.h" +#endif /* FIXED_POINT_POSITION */ + #ifdef SATURATE #define ADD(x, y) add_sat((x), (y)) #define SUB(x, y) sub_sat((x), (y)) @@ -37,19 +41,19 @@ * e.g. -DDATA_TYPE_IN1=uchar -DDATA_TYPE_IN2=uchar -DDATA_TYPE_OUT=short * @attention To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used. * - * @param[in] in1_ptr Pointer to the source image. Supported data types: U8, S16 + * @param[in] in1_ptr Pointer to the source image. Supported data types: U8/QS8/QS16/S16/F16/F32 * @param[in] in1_stride_x Stride of the source image 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 image 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_offset_first_element_in_bytes The offset of the first element in the source image - * @param[in] in2_ptr Pointer to the source image. Supported data types: U8, S16 + * @param[in] in2_ptr Pointer to the source image. Supported data types: U8/QS8 (only if @p in1_ptr is QS8), QS16 (only if @p in1_ptr is QS16), S16/F16/F32 * @param[in] in2_stride_x Stride of the source image 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 image 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_offset_first_element_in_bytes The offset of the first element in the source image - * @param[out] out_ptr Pointer to the destination image. Supported data types: U8, S16 + * @param[out] out_ptr Pointer to the destination image. Supported data types: U8 (only if both inputs are U8), QS8 (only if both inputs are QS8), QS16 (only if both inputs are QS16), S16/F16/F32 * @param[in] out_stride_x Stride of the destination image 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 image in Y dimension (in bytes) diff --git a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp index 0cb0847784..65422c2bbf 100644 --- a/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp +++ b/src/core/CL/kernels/CLArithmeticAdditionKernel.cpp @@ -69,11 +69,16 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen } ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MSG(output->info()->data_type() == DataType::U8 && (input1->info()->data_type() != DataType::U8 || input2->info()->data_type() != DataType::U8), "Output can only be U8 if both inputs are U8"); + if(is_data_type_fixed_point(input1->info()->data_type()) || is_data_type_fixed_point(input2->info()->data_type()) || is_data_type_fixed_point(output->info()->data_type())) + { + // Check that all data types are the same and all fixed-point positions are the same + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2, output); + } _input1 = input1; _input2 = input2; @@ -87,6 +92,10 @@ void CLArithmeticAdditionKernel::configure(const ICLTensor *input1, const ICLTen 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())); + if(is_data_type_fixed_point(input1->info()->data_type())) + { + build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position())); + } // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_add", build_opts)); diff --git a/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp b/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp index 69f9ff17d3..c5183af7d7 100644 --- a/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp +++ b/src/core/CL/kernels/CLArithmeticSubtractionKernel.cpp @@ -62,11 +62,16 @@ void CLArithmeticSubtractionKernel::configure(const ICLTensor *input1, const ICL } ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::QS8, DataType::QS16, DataType::S16, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_MSG(output->info()->data_type() == DataType::U8 && (input1->info()->data_type() != DataType::U8 || input2->info()->data_type() != DataType::U8), "Output can only be U8 if both inputs are U8"); + if(is_data_type_fixed_point(input1->info()->data_type()) || is_data_type_fixed_point(input2->info()->data_type()) || is_data_type_fixed_point(output->info()->data_type())) + { + // Check that all data types are the same and all fixed-point positions are the same + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input1, input2, output); + } _input1 = input1; _input2 = input2; @@ -80,6 +85,10 @@ void CLArithmeticSubtractionKernel::configure(const ICLTensor *input1, const ICL 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())); + if(is_data_type_fixed_point(input1->info()->data_type())) + { + build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input1->info()->fixed_point_position())); + } // Create kernel _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("arithmetic_sub", build_opts)); diff --git a/tests/validation/CL/ArithmeticAddition.cpp b/tests/validation/CL/ArithmeticAddition.cpp new file mode 100644 index 0000000000..3390ed6ada --- /dev/null +++ b/tests/validation/CL/ArithmeticAddition.cpp @@ -0,0 +1,305 @@ +/* + * Copyright (c) 2017 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 "CL/CLAccessor.h" +#include "PaddingCalculator.h" +#include "TensorLibrary.h" +#include "TypePrinter.h" +#include "tests/Globals.h" +#include "tests/Utils.h" +#include "validation/Datasets.h" +#include "validation/Reference.h" +#include "validation/Validation.h" + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/functions/CLArithmeticAddition.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" + +#include "boost_wrapper.h" + +#include <random> +#include <string> + +using namespace arm_compute; +using namespace arm_compute::test; +using namespace arm_compute::test::cl; +using namespace arm_compute::test::validation; + +namespace +{ +/** Compute Neon arithmetic addition function. + * + * @param[in] shape Shape of the input and output tensors. + * @param[in] dt_in0 Data type of first input tensor. + * @param[in] dt_in1 Data type of second input tensor. + * @param[in] dt_out Data type of the output tensor. + * @param[in] policy Overflow policy of the operation. + * @param[in] fixed_point_position (Optional) Fixed point position that expresses the number of bits for the fractional part of the number when the tensor's data type is QS8 or QS16 (default = 0). + * + * @return Computed output tensor. + */ +CLTensor compute_arithmetic_addition(const TensorShape &shape, DataType dt_in0, DataType dt_in1, DataType dt_out, ConvertPolicy policy, int fixed_point_position = 0) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, dt_in0, 1, fixed_point_position); + CLTensor src2 = create_tensor<CLTensor>(shape, dt_in1, 1, fixed_point_position); + CLTensor dst = create_tensor<CLTensor>(shape, dt_out, 1, fixed_point_position); + + // Create and configure function + CLArithmeticAddition add; + add.configure(&src1, &src2, &dst, policy); + + // Allocate tensors + src1.allocator()->allocate(); + src2.allocator()->allocate(); + dst.allocator()->allocate(); + + BOOST_TEST(!src1.info()->is_resizable()); + BOOST_TEST(!src2.info()->is_resizable()); + BOOST_TEST(!dst.info()->is_resizable()); + + // Fill tensors + library->fill_tensor_uniform(CLAccessor(src1), 0); + library->fill_tensor_uniform(CLAccessor(src2), 1); + + // Compute function + add.run(); + + return dst; +} + +void validate_configuration(const CLTensor &src1, const CLTensor &src2, CLTensor &dst, TensorShape shape, ConvertPolicy policy) +{ + BOOST_TEST(src1.info()->is_resizable()); + BOOST_TEST(src2.info()->is_resizable()); + BOOST_TEST(dst.info()->is_resizable()); + + // Create and configure function + CLArithmeticAddition add; + add.configure(&src1, &src2, &dst, policy); + + // Validate valid region + const ValidRegion valid_region = shape_to_valid_region(shape); + validate(src1.info()->valid_region(), valid_region); + validate(src2.info()->valid_region(), valid_region); + validate(dst.info()->valid_region(), valid_region); + + // Validate padding + const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding(); + validate(src1.info()->padding(), padding); + validate(src2.info()->padding(), padding); + validate(dst.info()->padding(), padding); +} +} // namespace + +#ifndef DOXYGEN_SKIP_THIS +BOOST_AUTO_TEST_SUITE(CL) +BOOST_AUTO_TEST_SUITE(ArithmeticAddition) + +BOOST_AUTO_TEST_SUITE(U8) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(Configuration, (SmallShapes() + LargeShapes()) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor src2 = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor dst = create_tensor<CLTensor>(shape, DataType::U8); + + validate_configuration(src1, src2, dst, shape, policy); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::U8, DataType::U8, DataType::U8, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::U8, DataType::U8, DataType::U8, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE(S16) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(Configuration, (SmallShapes() + LargeShapes()) * boost::unit_test::data::make({ DataType::U8, DataType::S16 }) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, dt, policy) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, dt); + CLTensor src2 = create_tensor<CLTensor>(shape, DataType::S16); + CLTensor dst = create_tensor<CLTensor>(shape, DataType::S16); + + validate_configuration(src1, src2, dst, shape, policy); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * boost::unit_test::data::make({ DataType::U8, DataType::S16 }) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, dt, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, dt, DataType::S16, DataType::S16, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, dt, DataType::S16, DataType::S16, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * boost::unit_test::data::make({ DataType::U8, DataType::S16 }) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, dt, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, dt, DataType::S16, DataType::S16, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, dt, DataType::S16, DataType::S16, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE(Quantized) +BOOST_AUTO_TEST_SUITE(QS8) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 7), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 7), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE(QS16) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 15), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 15), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() +BOOST_AUTO_TEST_SUITE_END() + +#ifdef ARM_COMPUTE_ENABLE_FP16 +BOOST_AUTO_TEST_SUITE(F16) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes(), shape) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::F16, DataType::F16, DataType::F16, ConvertPolicy::WRAP); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::F16, DataType::F16, DataType::F16, ConvertPolicy::WRAP); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() +#endif /* ARM_COMPUTE_ENABLE_FP16 */ + +BOOST_AUTO_TEST_SUITE(F32) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(Configuration, (SmallShapes() + LargeShapes()) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, DataType::F32); + CLTensor src2 = create_tensor<CLTensor>(shape, DataType::F32); + CLTensor dst = create_tensor<CLTensor>(shape, DataType::F32); + + validate_configuration(src1, src2, dst, shape, policy); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes(), shape) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::F32, DataType::F32, DataType::F32, ConvertPolicy::WRAP); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::F32, DataType::F32, DataType::F32, ConvertPolicy::WRAP); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_addition(shape, DataType::F32, DataType::F32, DataType::F32, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_addition(shape, DataType::F32, DataType::F32, DataType::F32, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE_END() +BOOST_AUTO_TEST_SUITE_END() +#endif /* DOXYGEN_SKIP_THIS */ diff --git a/tests/validation/CL/ArithmeticSubtraction.cpp b/tests/validation/CL/ArithmeticSubtraction.cpp new file mode 100644 index 0000000000..397e363590 --- /dev/null +++ b/tests/validation/CL/ArithmeticSubtraction.cpp @@ -0,0 +1,289 @@ +/* + * Copyright (c) 2017 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 "CL/CLAccessor.h" +#include "PaddingCalculator.h" +#include "TensorLibrary.h" +#include "TypePrinter.h" +#include "tests/Globals.h" +#include "tests/Utils.h" +#include "validation/Datasets.h" +#include "validation/Reference.h" +#include "validation/Validation.h" + +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/functions/CLArithmeticSubtraction.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" + +#include "boost_wrapper.h" + +#include <random> +#include <string> + +using namespace arm_compute; +using namespace arm_compute::test; +using namespace arm_compute::test::cl; +using namespace arm_compute::test::validation; + +namespace +{ +/** Compute Neon arithmetic subtraction function. + * + * @param[in] shape Shape of the input and output tensors. + * @param[in] dt_in0 Data type of first input tensor. + * @param[in] dt_in1 Data type of second input tensor. + * @param[in] dt_out Data type of the output tensor. + * @param[in] policy Overflow policy of the operation. + * @param[in] fixed_point_position (Optional) Fixed point position that expresses the number of bits for the fractional part of the number when the tensor's data type is QS8 or QS16 (default = 0). + * + * @return Computed output tensor. + */ +CLTensor compute_arithmetic_subtraction(const TensorShape &shape, DataType dt_in0, DataType dt_in1, DataType dt_out, ConvertPolicy policy, int fixed_point_position = 0) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, dt_in0, 1, fixed_point_position); + CLTensor src2 = create_tensor<CLTensor>(shape, dt_in1, 1, fixed_point_position); + CLTensor dst = create_tensor<CLTensor>(shape, dt_out, 1, fixed_point_position); + + // Create and configure function + CLArithmeticSubtraction sub; + sub.configure(&src1, &src2, &dst, policy); + + // Allocate tensors + src1.allocator()->allocate(); + src2.allocator()->allocate(); + dst.allocator()->allocate(); + + BOOST_TEST(!src1.info()->is_resizable()); + BOOST_TEST(!src2.info()->is_resizable()); + BOOST_TEST(!dst.info()->is_resizable()); + + // Fill tensors + library->fill_tensor_uniform(CLAccessor(src1), 0); + library->fill_tensor_uniform(CLAccessor(src2), 1); + + // Compute function + sub.run(); + + return dst; +} + +void validate_configuration(const CLTensor &src1, const CLTensor &src2, CLTensor &dst, TensorShape shape, ConvertPolicy policy) +{ + BOOST_TEST(src1.info()->is_resizable()); + BOOST_TEST(src2.info()->is_resizable()); + BOOST_TEST(dst.info()->is_resizable()); + + // Create and configure function + CLArithmeticSubtraction sub; + sub.configure(&src1, &src2, &dst, policy); + + // Validate valid region + const ValidRegion valid_region = shape_to_valid_region(shape); + validate(src1.info()->valid_region(), valid_region); + validate(src2.info()->valid_region(), valid_region); + validate(dst.info()->valid_region(), valid_region); + + // Validate padding + const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding(); + validate(src1.info()->padding(), padding); + validate(src2.info()->padding(), padding); + validate(dst.info()->padding(), padding); +} +} // namespace + +#ifndef DOXYGEN_SKIP_THIS +BOOST_AUTO_TEST_SUITE(CL) +BOOST_AUTO_TEST_SUITE(ArithmeticSubtraction) + +BOOST_AUTO_TEST_SUITE(U8) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(Configuration, (SmallShapes() + LargeShapes()) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor src2 = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor dst = create_tensor<CLTensor>(shape, DataType::U8); + + validate_configuration(src1, src2, dst, shape, policy); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, DataType::U8, DataType::U8, DataType::U8, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, DataType::U8, DataType::U8, DataType::U8, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE(S16) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(Configuration, (SmallShapes() + LargeShapes()) * boost::unit_test::data::make({ DataType::U8, DataType::S16 }) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, dt, policy) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, dt); + CLTensor src2 = create_tensor<CLTensor>(shape, DataType::S16); + CLTensor dst = create_tensor<CLTensor>(shape, DataType::S16); + + validate_configuration(src1, src2, dst, shape, policy); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * boost::unit_test::data::make({ DataType::U8, DataType::S16 }) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, dt, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, dt, DataType::S16, DataType::S16, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, dt, DataType::S16, DataType::S16, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * boost::unit_test::data::make({ DataType::U8, DataType::S16 }) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, dt, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, dt, DataType::S16, DataType::S16, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, dt, DataType::S16, DataType::S16, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE(Quantized) +BOOST_AUTO_TEST_SUITE(QS8) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 7), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 7), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, DataType::QS8, DataType::QS8, DataType::QS8, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE(QS16) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 15), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * ConvertPolicies() * boost::unit_test::data::xrange(1, 15), + shape, policy, fixed_point_position) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, DataType::QS16, DataType::QS16, DataType::QS16, policy, fixed_point_position); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE(Float) +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(Configuration, (SmallShapes() + LargeShapes()) * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Create tensors + CLTensor src1 = create_tensor<CLTensor>(shape, DataType::F32); + CLTensor src2 = create_tensor<CLTensor>(shape, DataType::F32); + CLTensor dst = create_tensor<CLTensor>(shape, DataType::F32); + + validate_configuration(src1, src2, dst, shape, policy); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit")) +BOOST_DATA_TEST_CASE(RunSmall, SmallShapes(), shape) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, DataType::F32, DataType::F32, DataType::F32, ConvertPolicy::WRAP); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, DataType::F32, DataType::F32, DataType::F32, ConvertPolicy::WRAP); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly")) +BOOST_DATA_TEST_CASE(RunLarge, LargeShapes() * boost::unit_test::data::make({ ConvertPolicy::SATURATE, ConvertPolicy::WRAP }), + shape, policy) +{ + // Compute function + CLTensor dst = compute_arithmetic_subtraction(shape, DataType::F32, DataType::F32, DataType::F32, policy); + + // Compute reference + RawTensor ref_dst = Reference::compute_reference_arithmetic_subtraction(shape, DataType::F32, DataType::F32, DataType::F32, policy); + + // Validate output + validate(CLAccessor(dst), ref_dst); +} +BOOST_AUTO_TEST_SUITE_END() + +BOOST_AUTO_TEST_SUITE_END() +BOOST_AUTO_TEST_SUITE_END() +#endif /* DOXYGEN_SKIP_THIS */ |