diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2017-09-05 13:51:21 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:35:24 +0000 |
commit | bf17955e2bf36c635acbac7c3bb03fbbd7732671 (patch) | |
tree | 9032d87a13942d4b7a15b2db5e7570d79823c66f | |
parent | 583137cc60580023abfd9d05abf933e7e117e29f (diff) | |
download | ComputeLibrary-bf17955e2bf36c635acbac7c3bb03fbbd7732671.tar.gz |
COMPMID-522 - Added support for GlobalPooling in CLPoolingLayer and CLFlattening for 3D tensor
Change-Id: Ifc7db1e4d4af322a4dcbfeb3e132e5c326596872
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/86618
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
-rw-r--r-- | arm_compute/core/CL/kernels/CLPoolingLayerKernel.h | 7 | ||||
-rw-r--r-- | arm_compute/runtime/CL/CLFunctions.h | 1 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLFlattenLayer.h | 52 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLPoolingLayer.h | 2 | ||||
-rw-r--r-- | src/core/CL/CLKernelLibrary.cpp | 1 | ||||
-rw-r--r-- | src/core/CL/cl_kernels/pooling_layer.cl | 98 | ||||
-rw-r--r-- | src/core/CL/kernels/CLPoolingLayerKernel.cpp | 80 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLFlattenLayer.cpp | 37 | ||||
-rw-r--r-- | tests/datasets/ShapeDatasets.h | 25 | ||||
-rw-r--r-- | tests/validation/CL/Flatten.cpp | 117 | ||||
-rw-r--r-- | tests/validation/CL/GlobalPooling.cpp | 83 | ||||
-rw-r--r-- | tests/validation/CL/PoolingLayer.cpp | 2 | ||||
-rw-r--r-- | tests/validation/CPP/FlattenLayer.cpp | 58 | ||||
-rw-r--r-- | tests/validation/CPP/FlattenLayer.h | 44 | ||||
-rw-r--r-- | tests/validation/fixtures/FlattenLayerFixture.h | 128 | ||||
-rw-r--r-- | tests/validation/fixtures/PoolingLayerFixture.h | 11 |
16 files changed, 708 insertions, 38 deletions
diff --git a/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h b/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h index 971e1506af..9251a8ed98 100644 --- a/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h +++ b/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h @@ -49,10 +49,11 @@ public: /** Set the input and output tensors. * - * @param[in] input Source tensor. Data types supported: F16/F32. + * @note QS8 and QS16 are supported only for pool sizes 3, 5 and 7 + * + * @param[in] input Source tensor. Data types supported: QS8/QS16/F16/F32. * @param[out] output Destination tensor. Data types supported: Same as @p input. * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. - * Supported pooling sizes : 2, 3 and 7 */ void configure(const ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info); @@ -67,5 +68,5 @@ private: BorderSize _border_size; unsigned int _num_elems_processed_per_iteration; }; -} +} // namespace arm_compute #endif /*__ARM_COMPUTE_CLPOOLINGLAYERKERNEL_H__ */ diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h index 9c1177f43e..64b2deb3bf 100644 --- a/arm_compute/runtime/CL/CLFunctions.h +++ b/arm_compute/runtime/CL/CLFunctions.h @@ -53,6 +53,7 @@ #include "arm_compute/runtime/CL/functions/CLErode.h" #include "arm_compute/runtime/CL/functions/CLFastCorners.h" #include "arm_compute/runtime/CL/functions/CLFillBorder.h" +#include "arm_compute/runtime/CL/functions/CLFlattenLayer.h" #include "arm_compute/runtime/CL/functions/CLFloor.h" #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" #include "arm_compute/runtime/CL/functions/CLGEMM.h" diff --git a/arm_compute/runtime/CL/functions/CLFlattenLayer.h b/arm_compute/runtime/CL/functions/CLFlattenLayer.h new file mode 100644 index 0000000000..e19b0e465f --- /dev/null +++ b/arm_compute/runtime/CL/functions/CLFlattenLayer.h @@ -0,0 +1,52 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_CLFLATTENLAYER_H__ +#define __ARM_COMPUTE_CLFLATTENLAYER_H__ + +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/ICLSimpleFunction.h" + +namespace arm_compute +{ +class ICLTensor; + +/** Basic function to execute flatten. This function calls the following OpenCL kernel: +* +* -# @ref CLIm2ColKernel +* +*/ +class CLFlattenLayer : public ICLSimpleFunction +{ +public: + /** Initialise the kernel's input and output. + * + * @param[in] input First input tensor to flatten with at least 3 dimensions. The dimensions over the third will be interpreted as batches. Data types supported: QS8/QS16/F16/F32 + * @param[out] output Output tensor with shape [w*h*d, input_batches] where: + * w = width input tensor, h = height input tensor and d = depth input tensor. Data type supported: same as @p input + */ + void configure(const ICLTensor *input, ICLTensor *output); +}; +} // namespace arm_compute + +#endif /* __ARM_COMPUTE_CLFLATTENLAYER_H__ */
\ No newline at end of file diff --git a/arm_compute/runtime/CL/functions/CLPoolingLayer.h b/arm_compute/runtime/CL/functions/CLPoolingLayer.h index b9675b94e0..80233d400e 100644 --- a/arm_compute/runtime/CL/functions/CLPoolingLayer.h +++ b/arm_compute/runtime/CL/functions/CLPoolingLayer.h @@ -48,5 +48,5 @@ public: */ void configure(ICLTensor *input, ICLTensor *output, const PoolingLayerInfo &pool_info); }; -} +} // namespace arm_compute #endif /* __ARM_COMPUTE_CLPOOLINGLAYER_H__ */ diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 6602b4dc78..4cd0a78a92 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -239,6 +239,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map = { "pooling_layer_3", "pooling_layer.cl" }, { "pooling_layer_3_optimized", "pooling_layer.cl" }, { "pooling_layer_7", "pooling_layer.cl" }, + { "pooling_layer_N", "pooling_layer.cl" }, { "quantization_layer", "quantization_layer.cl" }, { "reduction_operation", "reduction_operation.cl" }, { "remap_nearest_neighbour", "remap.cl" }, diff --git a/src/core/CL/cl_kernels/pooling_layer.cl b/src/core/CL/cl_kernels/pooling_layer.cl index 18ad4a69a8..0497bf4b91 100644 --- a/src/core/CL/cl_kernels/pooling_layer.cl +++ b/src/core/CL/cl_kernels/pooling_layer.cl @@ -415,3 +415,101 @@ __kernel void pooling_layer_7( // Store result *(__global DATA_TYPE *)output.ptr = res; } + +#if defined(POOL_SIZE) + +// Set the initial value for the pooling operation accordingly with the data type +#if defined(POOL_AVG) +#define INITIAL_VALUE 0 +#else // POOL_AVG +#ifdef FIXED_POINT_POSITION +#define MIN_VAL_EXPAND(type) type##_MIN +#define MIN_VAL(type) MIN_VAL_EXPAND(type) +#define INITIAL_VALUE MIN_VAL(DATA_TYPE) +#define INITIAL_VALUE 0 +#else // FIXED_POINT_POSITION +#if FP16 +#define INITIAL_VALUE -HALF_MAX +#else // FP16 +#define INITIAL_VALUE -FLT_MAX +#endif // FP16 +#endif // FIXED_POINT_POSITION + +#endif // POOL_AVG + +/** Performs a pooling function of pool size equal to N + * + * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16/F32; + * @note -DFP16 must be passed at compile time if half float data type is used + * @note Pool size must be passed using -DPOOL_SIZE e.g. -DPOOL_SIZE=13; + * @note In case of average pooling the following information must be passed at compile time: + * -DPOOL_AVG must be provided otherwise max pooling will be performed. + * -DMAX_WIDTH and -DMAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad) + * -DSTRIDE_X and -DSTRIDE_Y which are the steps of the window along the x and y directions + * -DPAD_X and -DPAD_Y which are the pooling paddings in x and y dimension + * + * @param[in] input_ptr Pointer to the source image. Supported data types: F16/F32 + * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image + * @param[out] output_ptr Pointer to the destination image. Supported data types: same as @p input_ptr + * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) + * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) + * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image + */ +__kernel void pooling_layer_N( + TENSOR3D_DECLARATION(input), + TENSOR3D_DECLARATION(output)) +{ + // Get pixels pointer + Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); + Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); + + VEC_DATA_TYPE(DATA_TYPE, 8) + vdata = INITIAL_VALUE; + DATA_TYPE sdata = INITIAL_VALUE; + + // Load data + for(int y = 0; y < POOL_SIZE; y++) + { + int x = 0; + for(; x <= ((int)POOL_SIZE - 8); x += 8) + { + VEC_DATA_TYPE(DATA_TYPE, 8) + data0 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); + vdata = POOL_OP(vdata, data0); + } + + // Leftover + for(; x < (int)POOL_SIZE; ++x) + { + DATA_TYPE data0 = *((__global DATA_TYPE *)tensor3D_offset(&input, x, y, 0)); + sdata = POOL_OP(sdata, data0); + } + } + + // Reduce result + VEC_DATA_TYPE(DATA_TYPE, 4) + reduce4 = POOL_OP(vdata.s0123, vdata.s4567); + VEC_DATA_TYPE(DATA_TYPE, 2) + reduce2 = POOL_OP(reduce4.s01, reduce4.s23); + DATA_TYPE res = POOL_OP(reduce2.s0, reduce2.s1); + res = POOL_OP(res, sdata); + + // Divide by pool region in case of average pooling +#ifdef POOL_AVG + res = DIV_OP(res, calculate_avg_scale(POOL_SIZE, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y)); +#endif /* POOL_AVG */ + + // Store result + *(__global DATA_TYPE *)output.ptr = res; +} +#endif // defined(POOL_SIZE)
\ No newline at end of file diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp index 22d29c66aa..22c7730963 100644 --- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp +++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp @@ -64,13 +64,10 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output, std::tie(pool_pad_x, pool_pad_y) = pad_stride_info.pad(); std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride(); - static const std::set<int> supported_pool_sizes = { 2, 3, 7 }; - ARM_COMPUTE_UNUSED(supported_pool_sizes); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); ARM_COMPUTE_ERROR_ON_NULLPTR(output); - ARM_COMPUTE_ERROR_ON(supported_pool_sizes.find(pool_size) == supported_pool_sizes.end()); ARM_COMPUTE_ERROR_ON(pool_pad_x >= pool_size || pool_pad_y >= pool_size); + ARM_COMPUTE_ERROR_ON(pool_size > 7 && is_data_type_fixed_point(input->info()->data_type())); // Check output dimensions std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(0), @@ -92,29 +89,14 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pooled_w) || (output->info()->dimension(1) != pooled_h)); - // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenCL kernel where - // each thread computes 4 output elements - const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3) && !is_data_type_fixed_point(input->info()->data_type()); - - int num_elements_read_per_iteration = (pool_size == 7) ? 8 : pool_size; - if(is_pool3x3_stride_le3) - { - // Change the number of elements processed and number of elements read per iteration for pooling 3x3 with stride less equal than 3 - _num_elems_processed_per_iteration = 4; - num_elements_read_per_iteration = pool_size * (pool_stride_x + 1); - } - const int input_width = input->info()->dimension(0); - const int input_height = input->info()->dimension(1); - const int upper_bound_w = ((pooled_w - 1) * pool_stride_x - pool_pad_x + num_elements_read_per_iteration) - input_width; - const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height; + const int input_width = input->info()->dimension(0); + const int input_height = input->info()->dimension(1); // Set instance variables - _input = input; - _output = output; - _pool_info = pool_info; - _border_size = BorderSize(pool_pad_y, pool_pad_x); - _border_size.right = std::max(upper_bound_w, pool_pad_x); - _border_size.bottom = std::max(upper_bound_h, pool_pad_y); + _input = input; + _output = output; + _pool_info = pool_info; + _border_size = BorderSize(pool_pad_y, pool_pad_x); // Set build options std::set<std::string> build_opts; @@ -136,14 +118,52 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output, } // Create kernel - std::string kernel_name = "pooling_layer_" + support::cpp11::to_string(pool_size); - if(is_pool3x3_stride_le3) + if(pool_size <= 7) { - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name + "_optimized", build_opts)); + // Check if we have pool3x3 with stride_x less equal than 3. In these cases, run an optimized OpenCL kernel where + // each thread computes 4 output elements + const bool is_pool3x3_stride_le3 = (pool_size == 3) && (pool_stride_x <= 3) && !is_data_type_fixed_point(input->info()->data_type()); + + int num_elements_read_per_iteration = (pool_size == 7) ? 8 : pool_size; + if(is_pool3x3_stride_le3) + { + // Change the number of elements processed and number of elements read per iteration for pooling 3x3 with stride less equal than 3 + _num_elems_processed_per_iteration = 4; + num_elements_read_per_iteration = pool_size * (pool_stride_x + 1); + } + + const int upper_bound_w = ((pooled_w - 1) * pool_stride_x - pool_pad_x + num_elements_read_per_iteration) - input_width; + const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height; + + _border_size.right = std::max(upper_bound_w, pool_pad_x); + _border_size.bottom = std::max(upper_bound_h, pool_pad_y); + + std::string kernel_name = "pooling_layer_" + support::cpp11::to_string(pool_size); + if(is_pool3x3_stride_le3) + { + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name + "_optimized", build_opts)); + } + else + { + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); + } } - else + else // Run general case { - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); + _num_elems_processed_per_iteration = 1; + + const int upper_bound_w = ((pooled_w - 1) * pool_stride_x - pool_pad_x + pool_size) - input_width; + const int upper_bound_h = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height; + + _border_size.right = std::max(upper_bound_w, pool_pad_x); + _border_size.bottom = std::max(upper_bound_h, pool_pad_y); + + build_opts.emplace(("-DPOOL_SIZE=" + support::cpp11::to_string(pool_size))); + if(input->info()->data_type() == DataType::F16) + { + build_opts.emplace("-DFP16"); + } + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("pooling_layer_N", build_opts)); } // Configure kernel window diff --git a/src/runtime/CL/functions/CLFlattenLayer.cpp b/src/runtime/CL/functions/CLFlattenLayer.cpp new file mode 100644 index 0000000000..9f571b2036 --- /dev/null +++ b/src/runtime/CL/functions/CLFlattenLayer.cpp @@ -0,0 +1,37 @@ +/* + * 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 "arm_compute/runtime/CL/functions/CLFlattenLayer.h" + +#include "arm_compute/core/CL/kernels/CLIm2ColKernel.h" +#include "arm_compute/core/Size2D.h" +#include "support/ToolchainSupport.h" + +using namespace arm_compute; + +void CLFlattenLayer::configure(const ICLTensor *input, ICLTensor *output) +{ + auto k = arm_compute::support::cpp14::make_unique<CLIm2ColKernel>(); + k->configure(input, output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false); + _kernel = std::move(k); +} diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h index ddcf6a7a95..aa653ace60 100644 --- a/tests/datasets/ShapeDatasets.h +++ b/tests/datasets/ShapeDatasets.h @@ -101,14 +101,14 @@ public: : ShapeDataset("Shape", { // Batch size 1 - TensorShape{ 7U, 7U }, + TensorShape{ 9U, 9U }, TensorShape{ 27U, 13U, 2U }, TensorShape{ 128U, 64U, 1U, 3U }, // Batch size 4 - TensorShape{ 7U, 7U, 3U, 4U }, + TensorShape{ 9U, 9U, 3U, 4U }, TensorShape{ 27U, 13U, 2U, 4U }, // Arbitrary batch size - TensorShape{ 7U, 7U, 3U, 5U } + TensorShape{ 9U, 9U, 3U, 5U } }) { } @@ -218,6 +218,25 @@ public: } }; +/** Data set containing global pooling tensor shapes. */ +class GlobalPoolingShapes final : public ShapeDataset +{ +public: + GlobalPoolingShapes() + : ShapeDataset("Shape", + { + // Batch size 1 + TensorShape{ 9U, 9U }, + TensorShape{ 13U, 13U, 2U }, + TensorShape{ 27U, 27U, 1U, 3U }, + // Batch size 4 + TensorShape{ 31U, 31U, 3U, 4U }, + TensorShape{ 34U, 34U, 2U, 4U } + }) + { + } +}; + } // namespace datasets } // namespace test } // namespace arm_compute diff --git a/tests/validation/CL/Flatten.cpp b/tests/validation/CL/Flatten.cpp new file mode 100644 index 0000000000..3350a72910 --- /dev/null +++ b/tests/validation/CL/Flatten.cpp @@ -0,0 +1,117 @@ +/* + * 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 "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "arm_compute/runtime/CL/functions/CLFlattenLayer.h" +#include "tests/CL/CLAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/FlattenLayerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(CL) +TEST_SUITE(FlattenLayer) + +template <typename T> +using CLFlattenLayerFixture = FlattenLayerValidationFixture<CLTensor, CLAccessor, CLFlattenLayer, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<float>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLFlattenLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<half>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLFlattenLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() +TEST_SUITE_END() + +TEST_SUITE(Quantized) +TEST_SUITE(QS8) +FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<int8_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), + framework::dataset::make("DataType", DataType::QS8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLFlattenLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), + framework::dataset::make("DataType", DataType::QS8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(QS16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<int16_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), + framework::dataset::make("DataType", DataType::QS16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLFlattenLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), + framework::dataset::make("DataType", DataType::QS16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() +TEST_SUITE_END() + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/GlobalPooling.cpp b/tests/validation/CL/GlobalPooling.cpp new file mode 100644 index 0000000000..c5c9d00b27 --- /dev/null +++ b/tests/validation/CL/GlobalPooling.cpp @@ -0,0 +1,83 @@ +/* + * 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 "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "arm_compute/runtime/CL/functions/CLPoolingLayer.h" +#include "tests/CL/CLAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/PoolingTypesDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/PoolingLayerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +/** Input data set for float data types */ +const auto GlobalPoolingLayerDataset = combine(datasets::GlobalPoolingShapes(), datasets::PoolingTypes()); + +/** Input data set for quantized data types */ +constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for float types */ +constexpr AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for float types */ +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(GlobalPoolingLayer) + +template <typename T> +using CLGlobalPoolingLayerFixture = GlobalPoolingLayerValidationFixture<CLTensor, CLAccessor, CLPoolingLayer, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunGlobalPooling, CLGlobalPoolingLayerFixture<float>, framework::DatasetMode::ALL, combine(GlobalPoolingLayerDataset, framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunGlobalPooling, CLGlobalPoolingLayerFixture<half>, framework::DatasetMode::ALL, combine(GlobalPoolingLayerDataset, framework::dataset::make("DataType", + DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() +TEST_SUITE_END() + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/PoolingLayer.cpp b/tests/validation/CL/PoolingLayer.cpp index 4dca93a9ac..e82df07a91 100644 --- a/tests/validation/CL/PoolingLayer.cpp +++ b/tests/validation/CL/PoolingLayer.cpp @@ -44,7 +44,7 @@ namespace validation namespace { /** Input data set for float data types */ -const auto PoolingLayerDatasetFP = combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { 2, 3, 7 })), +const auto PoolingLayerDatasetFP = combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { 2, 3, 7, 9 })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })); /** Input data set for quantized data types */ diff --git a/tests/validation/CPP/FlattenLayer.cpp b/tests/validation/CPP/FlattenLayer.cpp new file mode 100644 index 0000000000..611701d8cf --- /dev/null +++ b/tests/validation/CPP/FlattenLayer.cpp @@ -0,0 +1,58 @@ +/* + * 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 "FlattenLayer.h" + +#include "tests/validation/FixedPoint.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<T> flatten_layer(const SimpleTensor<T> &src) +{ + TensorShape shape_flatten(src.shape()); + shape_flatten.set(0, src.shape()[0] * src.shape()[1] * src.shape()[2]); + shape_flatten.remove_dimension(1); + shape_flatten.remove_dimension(1); + SimpleTensor<T> dst(shape_flatten, src.data_type(), 1, src.fixed_point_position()); + + // Note: Since the reference implementation does not use padding bytes, we can copy directly the content of the source tensor + std::copy(src.data(), src.data() + src.num_elements(), dst.data()); + + return dst; +} + +template SimpleTensor<float> flatten_layer(const SimpleTensor<float> &src); +template SimpleTensor<half> flatten_layer(const SimpleTensor<half> &src); +template SimpleTensor<qint8_t> flatten_layer(const SimpleTensor<qint8_t> &src); +template SimpleTensor<qint16_t> flatten_layer(const SimpleTensor<qint16_t> &src); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CPP/FlattenLayer.h b/tests/validation/CPP/FlattenLayer.h new file mode 100644 index 0000000000..b1286fe2bd --- /dev/null +++ b/tests/validation/CPP/FlattenLayer.h @@ -0,0 +1,44 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_TEST_FLATTEN_LAYER_H__ +#define __ARM_COMPUTE_TEST_FLATTEN_LAYER_H__ + +#include "tests/SimpleTensor.h" +#include "tests/validation/Helpers.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<T> flatten_layer(const SimpleTensor<T> &src); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* __ARM_COMPUTE_TEST_FLATTEN_LAYER_H__ */ diff --git a/tests/validation/fixtures/FlattenLayerFixture.h b/tests/validation/fixtures/FlattenLayerFixture.h new file mode 100644 index 0000000000..faf53c70a3 --- /dev/null +++ b/tests/validation/fixtures/FlattenLayerFixture.h @@ -0,0 +1,128 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_TEST_FLATTEN_LAYER_FIXTURE +#define ARM_COMPUTE_TEST_FLATTEN_LAYER_FIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/runtime/Tensor.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/CPP/FlattenLayer.h" + +#include <random> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class FlattenLayerValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(TensorShape shape, DataType data_type) + { + _fractional_bits = is_data_type_fixed_point(data_type) ? 4 : 0; + _target = compute_target(shape, data_type); + _reference = compute_reference(shape, data_type); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(_target.info()->tensor_shape(), _reference.shape()); + } + +protected: + template <typename U> + void fill(U &&tensor) + { + if(_fractional_bits == 0) + { + std::uniform_real_distribution<> distribution(-1.f, 1.f); + library->fill(tensor, distribution, 0); + } + else + { + const int one_fixed = 1 << _fractional_bits; + std::uniform_int_distribution<> distribution(-one_fixed, one_fixed); + library->fill(tensor, distribution, 0); + } + } + + TensorType compute_target(const TensorShape &shape, DataType data_type) + { + TensorShape shape_flatten(shape); + shape_flatten.set(0, shape[0] * shape[1] * shape[2]); + shape_flatten.remove_dimension(1); + shape_flatten.remove_dimension(1); + + // Create tensors + TensorType src = create_tensor<TensorType>(shape, data_type, 1, _fractional_bits); + TensorType dst = create_tensor<TensorType>(shape_flatten, data_type, 1, _fractional_bits); + + // Create and configure function + FunctionType flatten_layer; + flatten_layer.configure(&src, &dst); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + src.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(src)); + + // Compute function + flatten_layer.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type) + { + // Create reference + SimpleTensor<T> src{ shape, data_type, 1, _fractional_bits }; + + // Fill reference + fill(src); + + return reference::flatten_layer<T>(src); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; + int _fractional_bits{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_FLATTEN_LAYER_FIXTURE */ diff --git a/tests/validation/fixtures/PoolingLayerFixture.h b/tests/validation/fixtures/PoolingLayerFixture.h index 5ce4aa6755..775c4125fc 100644 --- a/tests/validation/fixtures/PoolingLayerFixture.h +++ b/tests/validation/fixtures/PoolingLayerFixture.h @@ -128,6 +128,17 @@ public: PoolingLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, pool_type, pool_size, pad_stride_info, data_type, 0); } }; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class GlobalPoolingLayerValidationFixture : public PoolingLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + void setup(TensorShape shape, PoolingType pool_type, DataType data_type) + { + PoolingLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, pool_type, shape.x(), PadStrideInfo(1, 1, 0, 0), data_type, 0); + } +}; } // namespace validation } // namespace test } // namespace arm_compute |