From 11d8415aa57b69fb6c83e86a37e3026c22d1d37d Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Wed, 28 Apr 2021 10:20:18 +0100 Subject: Port DepthConvert to new Api - Renames DepthConvert to Cast - Ports both NEDepthConverLayer and CLDepthConvert variants - Removes legacy shift capability from DepthConvert, allowing only shifts of 0 Signed-off-by: Georgios Pinitas Change-Id: I806a0f8eb23d23502b632c529fda7edde19c8176 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5565 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- src/core/CL/CLKernels.h | 1 - src/core/CL/cl_kernels/cast.cl | 134 +++++++++++++++++++ src/core/CL/cl_kernels/depth_convert.cl | 138 -------------------- src/core/CL/kernels/CLDepthConvertLayerKernel.cpp | 152 ---------------------- src/core/CL/kernels/CLDepthConvertLayerKernel.h | 91 ------------- 5 files changed, 134 insertions(+), 382 deletions(-) create mode 100644 src/core/CL/cl_kernels/cast.cl delete mode 100644 src/core/CL/cl_kernels/depth_convert.cl delete mode 100644 src/core/CL/kernels/CLDepthConvertLayerKernel.cpp delete mode 100644 src/core/CL/kernels/CLDepthConvertLayerKernel.h (limited to 'src/core/CL') diff --git a/src/core/CL/CLKernels.h b/src/core/CL/CLKernels.h index 1302d52180..c59eebacbb 100644 --- a/src/core/CL/CLKernels.h +++ b/src/core/CL/CLKernels.h @@ -35,7 +35,6 @@ #include "src/core/CL/kernels/CLComparisonKernel.h" #include "src/core/CL/kernels/CLDeconvolutionLayerUpsampleKernel.h" #include "src/core/CL/kernels/CLDeconvolutionReshapeOutputKernel.h" -#include "src/core/CL/kernels/CLDepthConvertLayerKernel.h" #include "src/core/CL/kernels/CLDepthToSpaceLayerKernel.h" #include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h" #include "src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h" diff --git a/src/core/CL/cl_kernels/cast.cl b/src/core/CL/cl_kernels/cast.cl new file mode 100644 index 0000000000..036a683ec7 --- /dev/null +++ b/src/core/CL/cl_kernels/cast.cl @@ -0,0 +1,134 @@ +/* + * Copyright (c) 2016-2021 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" + +#ifdef SATURATE +#define CONVERT_DOWN(x, type) CONVERT_SAT(x, type) +#else /* SATURATE */ +#define CONVERT_DOWN(x, type) CONVERT(x, type) +#endif /* SATURATE */ + +#define CONVERT_UP(x, type) CONVERT(x, type) + +/** This function performs a down-casting + * + * @attention For QSYMM8_PER_CHANNEL -> QASYMM8, it is user's responsibility to keep track of the quantization info. + * + * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT: + * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S8/QSYMM8_PER_CHANNEL/U16/S16/U32/S32/F16/F32 + * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] in_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] in_step_z in_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] in_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/S8/QASYMM8/U16/S16/U32/S32/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) + * @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 image + */ +__kernel void cast_down( + TENSOR3D_DECLARATION(in), + TENSOR3D_DECLARATION(out)) +{ + int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + + __global uchar *in_addr = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z; + __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z; + + // Load data + VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) + in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr); + +#if defined(IS_DATA_TYPE_QUANTIZED) + in_data ^= (VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE))0x80; +#endif // defined(IS_DATA_TYPE_QUANTIZED) + +#if defined(IS_DATA_TYPE_FLOAT) + VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) + res0 = CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)); + STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +#else /* defined(IS_DATA_TYPE_FLOAT) */ + VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) + res0 = CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)); + STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +#endif /* defined(IS_DATA_TYPE_FLOAT) */ +} + +/** This function performs a up-casting + * + * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT: + * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short + * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 + * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE + * + * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S8/U16/S16/U32/S32/F16/F32 + * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) + * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) + * @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] in_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] in_step_z in_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] in_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/U16/S16/U32/S32/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) + * @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 image + */ +__kernel void cast_up( + TENSOR3D_DECLARATION(in), + TENSOR3D_DECLARATION(out)) +{ + int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); + + __global uchar *in_addr = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z; + __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z; + + // Load data + VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) + in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr); + +#if defined(IS_DATA_TYPE_FLOAT) + VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) + res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)); + STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +#else /* defined(IS_DATA_TYPE_FLOAT) */ + VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) + res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)); + STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) +#endif /* defined(IS_DATA_TYPE_FLOAT) */ +} diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl deleted file mode 100644 index a888d7b9bc..0000000000 --- a/src/core/CL/cl_kernels/depth_convert.cl +++ /dev/null @@ -1,138 +0,0 @@ -/* - * Copyright (c) 2016-2021 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" - -#ifdef SATURATE -#define CONVERT_DOWN(x, type) CONVERT_SAT(x, type) -#else /* SATURATE */ -#define CONVERT_DOWN(x, type) CONVERT(x, type) -#endif /* SATURATE */ - -#define CONVERT_UP(x, type) CONVERT(x, type) - -/** This function performs a down-scaling depth conversion. - * - * @attention For QSYMM8_PER_CHANNEL -> QASYMM8, it is user's responsibility to keep track of the quantization info. - * - * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT: - * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short - * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 - * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S8/QSYMM8_PER_CHANNEL/U16/S16/U32/S32/F16/F32 - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] in_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] in_step_z in_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] in_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/S8/QASYMM8/U16/S16/U32/S32/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) - * @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 image - * @param[in] shift The integer shift amount value. Supported data types: S32 - */ -__kernel void convert_depth_down( - TENSOR3D_DECLARATION(in), - TENSOR3D_DECLARATION(out), - const int shift) -{ - int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); - - __global uchar *in_addr = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z; - __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z; - - // Load data - VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) - in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr); - -#if defined(IS_DATA_TYPE_QUANTIZED) - in_data ^= (VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE))0x80; -#endif // defined(IS_DATA_TYPE_QUANTIZED) - -#if defined(IS_DATA_TYPE_FLOAT) - VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) - res0 = CONVERT_DOWN(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)); - STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) -#else /* defined(IS_DATA_TYPE_FLOAT) */ - VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) - res0 = CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)); - STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) -#endif /* defined(IS_DATA_TYPE_FLOAT) */ -} - -/** This function performs a up-scaling depth conversion. - * - * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT: - * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short - * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 - * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE - * - * @param[in] in_ptr Pointer to the source image. Supported data types: U8/S8/U16/S16/U32/S32/F16/F32 - * @param[in] in_stride_x Stride of the source image in X dimension (in bytes) - * @param[in] in_step_x in_stride_x * number of elements along X processed per workitem(in bytes) - * @param[in] in_stride_y Stride of the source image in Y dimension (in bytes) - * @param[in] in_step_y in_stride_y * number of elements along Y processed per workitem(in bytes) - * @param[in] in_stride_z Stride of the source tensor in Z dimension (in bytes) - * @param[in] in_step_z in_stride_z * number of elements along Z processed per workitem(in bytes) - * @param[in] in_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/U16/S16/U32/S32/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) - * @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 image - * @param[in] shift The integer shift amount value. Supported data types: S32 - */ -__kernel void convert_depth_up( - TENSOR3D_DECLARATION(in), - TENSOR3D_DECLARATION(out), - const int shift) -{ - int x_offs = max((int)(get_global_id(0) * VEC_SIZE - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE), 0); - - __global uchar *in_addr = in_ptr + in_offset_first_element_in_bytes + sizeof(DATA_TYPE_IN) * x_offs + get_global_id(1) * in_stride_y + get_global_id(2) * in_stride_z; - __global uchar *out_addr = out_ptr + out_offset_first_element_in_bytes + sizeof(DATA_TYPE_OUT) * x_offs + get_global_id(1) * out_stride_y + get_global_id(2) * out_stride_z; - - // Load data - VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) - in_data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)in_addr); - -#if defined(IS_DATA_TYPE_FLOAT) - VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) - res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)); - STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) -#else /* defined(IS_DATA_TYPE_FLOAT) */ - VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE) - res0 = CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)) << shift; - STORE_VECTOR_SELECT(res, DATA_TYPE_OUT, out_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) -#endif /* defined(IS_DATA_TYPE_FLOAT) */ -} diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp deleted file mode 100644 index 0d5c7a4881..0000000000 --- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp +++ /dev/null @@ -1,152 +0,0 @@ -/* - * Copyright (c) 2016-2021 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 "src/core/CL/kernels/CLDepthConvertLayerKernel.h" - -#include "arm_compute/core/CL/CLHelpers.h" -#include "arm_compute/core/CL/CLKernelLibrary.h" -#include "arm_compute/core/CL/ICLTensor.h" -#include "arm_compute/core/CL/OpenCL.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Utils.h" -#include "arm_compute/core/Validate.h" -#include "src/core/CL/CLValidate.h" -#include "src/core/helpers/AutoConfiguration.h" -#include "src/core/helpers/WindowHelpers.h" -#include "support/StringSupport.h" - -#include -#include -#include - -namespace arm_compute -{ -namespace -{ -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift) -{ - ARM_COMPUTE_UNUSED(policy); - ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_RETURN_ERROR_ON(input == output); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, - 1, - DataType::U8, DataType::S8, DataType::QSYMM8_PER_CHANNEL, DataType::S16, - DataType::U16, DataType::U32, DataType::S32, DataType::F16, - DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, - 1, - DataType::U8, DataType::S8, DataType::QASYMM8, DataType::S16, - DataType::U16, DataType::U32, DataType::S32, DataType::F16, - DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == output->data_type(), "Input and output data types must be different"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_float(input->data_type()) && shift != 0, "Shift is used only with integer non-quantized inputs"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_quantized(input->data_type()) && shift != 0, "Shift is used only with integer non-quantized inputs"); - ARM_COMPUTE_RETURN_ERROR_ON(shift >= 8); - - // Validate in case of configured output - if(output->total_size() > 0) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); - } - - return Status{}; -} -} // namespace - -void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift) -{ - configure(CLKernelLibrary::get().get_compile_context(), input, output, policy, shift); -} - -void CLDepthConvertLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift) -{ - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - - _input = input; - _output = output; - - // Auto initialize output shape if not initialized (We can only auto-configure the shape, datatype must be given) - set_shape_if_empty(*output->info(), input->info()->tensor_shape()); - - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy, shift)); - - auto padding_info = get_padding_info({ input, output }); - - // Get data sizes - const size_t input_size = data_size_from_type(input->info()->data_type()); - const size_t output_size = data_size_from_type(output->info()->data_type()); - - // Get number of elements to process per iterations - const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->info()->element_size(), input->info()->dimension(0)); - - // Set build options - CLBuildOptions build_opts; - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->info()->dimension(0) % num_elems_processed_per_iteration)); - build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type())); - build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); - // Conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined - build_opts.add_option_if(is_data_type_float(input->info()->data_type()) || policy == ConvertPolicy::SATURATE, "-DSATURATE"); - build_opts.add_option_if(is_data_type_float(input->info()->data_type()) || is_data_type_float(output->info()->data_type()), "-DIS_DATA_TYPE_FLOAT"); - build_opts.add_option_if(is_data_type_quantized(input->info()->data_type()), "-DIS_DATA_TYPE_QUANTIZED"); - - // Create kernel - const std::string kernel_name = (input_size >= output_size) ? "convert_depth_down" : "convert_depth_up"; - _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - - // Set shift arg - unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters - _kernel.setArg(idx++, shift); - - // Configure kernel - Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); - ICLKernel::configure_internal(win); - - // Collapse window - const Window &full_window = window(); - Window collapsed_window = full_window.collapse_if_possible(full_window, Window::DimZ); - ICLKernel::configure_internal(collapsed_window); - - ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); - - // Set config_id for enabling LWS tuning - _config_id = kernel_name; - _config_id += "_"; - _config_id += lower_string(string_from_data_type(input->info()->data_type())); - _config_id += "_"; - _config_id += support::cpp11::to_string(input->info()->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(input->info()->dimension(1)); - _config_id += "_"; - _config_id += support::cpp11::to_string(output->info()->dimension(0)); - _config_id += "_"; - _config_id += support::cpp11::to_string(output->info()->dimension(1)); -} - -Status CLDepthConvertLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift) -{ - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, policy, shift)); - - return Status{}; -} -} // namespace arm_compute diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.h b/src/core/CL/kernels/CLDepthConvertLayerKernel.h deleted file mode 100644 index 8b511c6707..0000000000 --- a/src/core/CL/kernels/CLDepthConvertLayerKernel.h +++ /dev/null @@ -1,91 +0,0 @@ -/* - * Copyright (c) 2016-2020 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_CLDEPTHCONVERTKERNEL_H -#define ARM_COMPUTE_CLDEPTHCONVERTKERNEL_H - -#include "arm_compute/core/Types.h" -#include "src/core/CL/ICLSimple3DKernel.h" - -#include - -namespace arm_compute -{ -class ICLTensor; - -/** Interface for the depth conversion kernel. */ -class CLDepthConvertLayerKernel : public ICLSimple3DKernel -{ -public: - /** Set the input and output of the kernel. - * - * Valid conversions Input -> Output : - * - * - QSYMM8_PER_CHANNEL -> QASYMM8 (ATTENTION: it is the user's responsibility to keep track of the quantization info in the TensorInfo meta-data) - * - U8 -> S8, U16, S16, U32, S32, F16, F32 - * - U16 -> U8, S8, S16, U32, S32, F16, F32 - * - S16 -> U8, S8, U16, U32, S32, F16, F32 - * - U32 -> U8, S8, U16, S16, S32, F16, F32 - * - S32 -> U8, S8, U16, S16, U32, F16, F32 - * - F16 -> U8, S8, U16, S16, U32, F32 - * - F32 -> U8, S8, U16, S16, U32, F16 - * - * @param[in] input The input tensor to convert. Data types supported: U8/S8/QSYMM8_PER_CHANNEL/U16/S16/U32/S32/F16/F32. - * @param[out] output The output tensor. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32. - * @param[in] policy Conversion policy - * @param[in] shift Value for down/up conversions. Must be 0 <= shift < 8. - */ - void configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift); - /** Set the input and output of the kernel. - * - * Valid conversions Input -> Output : - * - * - QSYMM8_PER_CHANNEL -> QASYMM8 (ATTENTION: it is the user's responsibility to keep track of the quantization info in the TensorInfo meta-data) - * - U8 -> S8, U16, S16, U32, S32, F16, F32 - * - U16 -> U8, S8, S16, U32, S32, F16, F32 - * - S16 -> U8, S8, U16, U32, S32, F16, F32 - * - U32 -> U8, S8, U16, S16, S32, F16, F32 - * - S32 -> U8, S8, U16, S16, U32, F16, F32 - * - F16 -> U8, S8, U16, S16, U32, F32 - * - F32 -> U8, S8, U16, S16, U32, F16 - * - * @param[in] compile_context The compile context to be used. - * @param[in] input The input tensor to convert. Data types supported: U8/S8/QSYMM8_PER_CHANNEL/U16/S16/U32/S32/F16/F32. - * @param[out] output The output tensor. Data types supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32. - * @param[in] policy Conversion policy - * @param[in] shift Value for down/up conversions. Must be 0 <= shift < 8. - */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift); - /** Static function to check if given info will lead to a valid configuration of @ref CLDepthConvertLayerKernel - * - * @param[in] input Source tensor info. Data types supported: U8/S8/QSYMM8_PER_CHANNEL/U16/S16/U32/S32/F16/F32. - * @param[in] output Destination tensor info. Data type supported: U8/S8/QASYMM8/U16/S16/U32/S32/F16/F32. - * @param[in] policy Conversion policy - * @param[in] shift Value for down/up conversions. Must be 0 <= shift < 8. - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift); -}; -} // namespace arm_compute -#endif /*ARM_COMPUTE_CLDEPTHCONVERTKERNEL_H */ -- cgit v1.2.1