diff options
Diffstat (limited to 'compute_kernel_writer/src/cl/helpers')
5 files changed, 884 insertions, 0 deletions
diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.cpp b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.cpp new file mode 100644 index 0000000000..7d16f35fbe --- /dev/null +++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.cpp @@ -0,0 +1,353 @@ +/* + * Copyright (c) 2023 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/cl/helpers/CLMemoryOpBufferHelper.h" + +#include "ckw/Error.h" +#include "ckw/TensorSampler.h" +#include "ckw/types/MemoryOperation.h" +#include "ckw/types/TensorStorageType.h" + +#include "src/cl/CLHelpers.h" +#include "src/cl/CLKernelWriter.h" +#include "src/cl/CLTensorArgument.h" +#include "src/cl/CLTile.h" +#include "src/ITensor.h" +#include "src/Tensor3dMapper.h" +#include "src/TileView.h" + +namespace ckw +{ +bool CLMemoryOpBufferHelper::validate(const CLKernelWriter *writer, + const ITensor *tensor, + const TensorSampler *sampler, + const Tensor3dMapper *mapper, + MemoryOperation op, + const TileView<CLTile> &dst) +{ + CKW_UNUSED(writer, tensor, mapper, op, dst); + + if (sampler->storage() != TensorStorageType::BufferUint8Ptr) + { + return false; + } + return true; +} + +/** Initialization and Finalizing Logic + * + * The meanings of if/elses in different dimensions and how they're constructed: + * - x: partial load/store + * - y: no load/store operation + * - z: no load/store operation + * if(x) + * { + * if(z) + * { + * if(y) + * { + * // full load/store width + * } + * else + * { + * // no load/store + * } + * } + * else + * { + * // no load/store + * } + * } + * else + * { + * if(z) + * { + * if(y) + * { + * // partial load/store width + * } + * else + * { + * // no load/store + * } + * } + * else + * { + * // no load/store + * } + * } + * + * In general, initialize() writes if conditions, and finalize() writes else conditions. + * The outermost block is x, then z and then y. This is why, if/else's covering for y are initialized + * at each row write. In some addressing modes, such as None, no if/else conditions are written. + */ +void CLMemoryOpBufferHelper::initialize(const CLTile *x, const CLTile *z, const CLTile *b) +{ + CKW_ASSERT(validate(_writer, _tensor, _sampler, _mapper.get(), _op, _dst)); + + _coord_x = x->scalar(0, 0).str; + _coord_z = z->scalar(0, 0).str; + _coord_b = b->scalar(0, 0).str; + _coord_orig_z = _coord_z; + + out_of_bound_initialize_x(_coord_x); + out_of_bound_initialize_z(_coord_z); +} + +void CLMemoryOpBufferHelper::write_row(int32_t row_id, const std::string &coord_y) +{ + // The only check required is on Y. + out_of_bound_initialize_y(coord_y); + + const std::string dst = _dst.vector(row_id).str; + const std::string address = to_buffer_address(_coord_x, coord_y, _coord_z, _coord_b); + const std::string ls_buf = to_statement(_op, _ls_width_full, dst, address); + + _writer->op_write_raw_code(ls_buf); + _writer->op_write_raw_code(";\n"); + + out_of_bound_finalize_y(dst); + + // The left over load/store will be written in the finalize stage + if (_ls_width_part.size() != 0) + { + int32_t col_start = 0; + const TileArea original_area = _dst.area(); + + for (int32_t partial_width : _ls_width_part) + { + // Set the active area + const TileArea area(original_area.row_start(), original_area.row_end(), col_start, + col_start + partial_width); + _dst.area(area); + + const std::string dst = _dst.vector(row_id).str; + const std::string coord_x = _coord_x + " + " + std::to_string(col_start); + const std::string address = to_buffer_address(coord_x, coord_y, _coord_z, _coord_b); + const std::string statement = to_statement(_op, partial_width, dst, address); + _leftovers_x.emplace_back(dst, coord_y, statement); + + col_start += partial_width; + } + // Restore the original area + _dst.area(original_area); + } +} + +void CLMemoryOpBufferHelper::finalize() +{ + out_of_bound_finalize_z(); + out_of_bound_finalize_x(); +} + +void CLMemoryOpBufferHelper::out_of_bound_initialize_x(const std::string &coord) +{ + if (_sampler->address_mode_x() == TensorSamplerAddressModeX::OverlappingMin) + { + TensorInfo tensor_info = _tensor->info(); + TensorShape shape = tensor_info.shape(); + + _ls_width_part = cl_decompose_vector_width(shape[0] % _ls_width_full); + if (_ls_width_part.size() != 0) + { + _writer->op_write_raw_code("if(" + coord + " > 0)\n{\n"); + } + } +} + +void CLMemoryOpBufferHelper::out_of_bound_finalize_x() +{ + if (_sampler->address_mode_x() == TensorSamplerAddressModeX::OverlappingMin) + { + if (_ls_width_part.size() != 0) + { + _writer->op_write_raw_code("}\nelse\n{\n"); + + out_of_bound_initialize_z(_coord_orig_z); + for (LeftoverDescriptor leftover_desc : _leftovers_x) + { + out_of_bound_initialize_y(leftover_desc.coord); + _writer->op_write_raw_code(leftover_desc.statement); + _writer->op_write_raw_code(";\n"); + out_of_bound_finalize_y(leftover_desc.dst); + } + out_of_bound_finalize_z(); + _writer->op_write_raw_code("}\n"); + } + } +} + +void CLMemoryOpBufferHelper::out_of_bound_initialize_y(const std::string &coord) +{ + std::string max = ""; + + const TensorSamplerAddressModeY address_mode_y = _sampler->address_mode_y(); + + switch (address_mode_y) + { + case TensorSamplerAddressModeY::ClampToBorderMaxOnly: + // Not to be moved outside the case because it marks the relevant tensor component as used even if we dont't use the variable + max = _mapper->dim_y().str; + _writer->op_write_raw_code("if(" + coord + " < " + max + ")\n{\n"); + break; + case TensorSamplerAddressModeY::SkipLessThanZero: + _writer->op_write_raw_code("if(" + coord + " >= 0)\n{\n"); + break; + case TensorSamplerAddressModeY::None: + break; + default: + CKW_THROW_MSG("Unsupported address mode for Y dimension"); + } +} + +void CLMemoryOpBufferHelper::out_of_bound_finalize_y(const std::string &dst) +{ + const TensorSamplerAddressModeY address_mode_y = _sampler->address_mode_y(); + + switch (address_mode_y) + { + case TensorSamplerAddressModeY::ClampToBorderMaxOnly: + _writer->op_write_raw_code("}\nelse\n{\n"); + _writer->op_write_raw_code(dst); + _writer->op_write_raw_code(" = 0.0f;\n}\n"); + break; + case TensorSamplerAddressModeY::SkipLessThanZero: + _writer->op_write_raw_code("}\n"); + break; + case TensorSamplerAddressModeY::None: + break; + default: + CKW_THROW_MSG("Unsupported address mode for Y dimension"); + } +} + +void CLMemoryOpBufferHelper::out_of_bound_initialize_z(const std::string &coord) +{ + CKW_UNUSED(coord); + + const TensorSamplerAddressModeZ address_mode_z = _sampler->address_mode_z(); + switch (address_mode_z) + { + case TensorSamplerAddressModeZ::None: + break; + default: + CKW_THROW_MSG("Unsupported address mode for Z dimension"); + } +} + +void CLMemoryOpBufferHelper::out_of_bound_finalize_z() +{ + const TensorSamplerAddressModeZ address_mode_z = _sampler->address_mode_z(); + + switch (address_mode_z) + { + case TensorSamplerAddressModeZ::None: + break; + default: + CKW_THROW_MSG("Unsupported address mode for Z dimension"); + } +} + +std::string CLMemoryOpBufferHelper::to_statement(MemoryOperation op, + int32_t vector_width, + const std::string &data, + const std::string &address) const +{ + switch (op) + { + case MemoryOperation::Load: + if (vector_width != 1) + { + return data + " = vload" + std::to_string(vector_width) + "(0, " + address + ")"; + } + else + { + return data + " = *(" + address + ")"; + } + break; + case MemoryOperation::Store: + if (vector_width != 1) + { + return "vstore" + std::to_string(vector_width) + "(" + data + ", 0, " + address + ")"; + } + else + { + return "*(" + address + ") = " + data; + } + break; + default: + CKW_THROW_MSG("Unsupported MemoryOperation"); + } + + return ""; +} + +std::string CLMemoryOpBufferHelper::to_buffer_address(const std::string &x, + const std::string &y, + const std::string &z, + const std::string &b) const +{ + TensorStorageType tensor_storage = _sampler->storage(); + CKW_ASSERT(tensor_storage == TensorStorageType::BufferUint8Ptr); + + const std::string ptr_buf = _tensor->storage(tensor_storage).val; + const std::string dst_type = cl_data_type_rounded_up_to_valid_vector_width(_dst.data_type(), 1); + + std::string address; + address += "(__global "; + address += dst_type; + address += "*)("; + address += ptr_buf; + if (x != "0" && (_mapper->dim_x().str != "1")) + { + address += " + ("; + address += x + ") * sizeof(" + dst_type + ")"; + } + if (y != "0") + { + const std::string stride_y = _mapper->stride_y().str; + address += " + ("; + address += y + ")"; + address += " * "; + address += stride_y; + } + if (z != "0" && (_mapper->dim_z().str != "1")) + { + const std::string stride_z = _mapper->stride_z().str; + address += " + ("; + address += z + ")"; + address += " * "; + address += stride_z; + } + if (b != "0" && (_mapper->dim_batch().str != "1")) + { + const std::string stride_b = _mapper->stride_batch().str; + address += " + ("; + address += b + ")"; + address += " * "; + address += stride_b; + } + address += ")"; + return address; +} +} // namespace ckw diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.h b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.h new file mode 100644 index 0000000000..a6b3272f32 --- /dev/null +++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpBufferHelper.h @@ -0,0 +1,108 @@ +/* + * Copyright (c) 2023 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 CKW_SRC_CL_HELPERS_CLMEMORYOPBUFFERHELPER_H +#define CKW_SRC_CL_HELPERS_CLMEMORYOPBUFFERHELPER_H + +#include "src/cl/helpers/ICLMemoryOpHelper.h" + +#include <cstdint> +#include <string> +#include <vector> + +namespace ckw +{ + +// Forward Declarations +class CLKernelWriter; +class CLTile; +template <class CLTile> +class TileView; +enum class MemoryOperation; + +/** Helper class to write memory operations (like load/store) in OpenCL + */ +class CLMemoryOpBufferHelper : public ICLMemoryOpHelper +{ +public: + /** Constructor similar to @ref ICLMemoryOpHelper() */ + CLMemoryOpBufferHelper(CLKernelWriter *writer, + ITensor *tensor, + TensorSampler *sampler, + MemoryOperation op, + const TileView<CLTile> &dst) + : ICLMemoryOpHelper(writer, tensor, sampler, op, dst) + { + } + + /** Copy constructor */ + CLMemoryOpBufferHelper(const CLMemoryOpBufferHelper &) = delete; + + /** Assignment operator overload */ + CLMemoryOpBufferHelper &operator=(const CLMemoryOpBufferHelper &) = delete; + + // Methods overridden + void initialize(const CLTile *x, const CLTile *z, const CLTile *b) override; + void write_row(int32_t row_id, const std::string &coord_y) override; + void finalize() override; + +private: + struct LeftoverDescriptor + { + LeftoverDescriptor(const std::string &dst, const std::string &coord, const std::string &statement) + : dst(dst), coord(coord), statement(statement) + { + } + + std::string dst{}; // Describes the destination tile or part of it + std::string coord{}; // Describes the coordinate to be used in boundary checks + std::string statement{}; // Describes the memory operation statement + }; + + std::vector<int32_t> _ls_width_part{}; + std::vector<LeftoverDescriptor> _leftovers_x{}; + std::string _coord_orig_z{}; + + static bool validate(const CLKernelWriter *writer, + const ITensor *tensor, + const TensorSampler *sampler, + const Tensor3dMapper *mapper, + MemoryOperation op, + const TileView<CLTile> &dst); + + void out_of_bound_initialize_x(const std::string &coord); + void out_of_bound_finalize_x(); + void out_of_bound_initialize_y(const std::string &coord); + void out_of_bound_finalize_y(const std::string &dst); + void out_of_bound_initialize_z(const std::string &coord); + void out_of_bound_finalize_z(); + + std::string + to_statement(MemoryOperation op, int32_t vector_width, const std::string &data, const std::string &address) const; + std::string + to_buffer_address(const std::string &x, const std::string &y, const std::string &z, const std::string &b) const; +}; +} // namespace ckw + +#endif // CKW_SRC_CL_HELPERS_CLMEMORYOPBUFFERHELPER_H diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.cpp b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.cpp new file mode 100644 index 0000000000..f392cd89cc --- /dev/null +++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.cpp @@ -0,0 +1,213 @@ +/* + * Copyright (c) 2023 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/cl/helpers/CLMemoryOpImage2dHelper.h" + +#include "ckw/Error.h" +#include "ckw/TensorSampler.h" +#include "ckw/types/MemoryOperation.h" +#include "ckw/types/TensorStorageType.h" + +#include "src/cl/CLKernelWriter.h" +#include "src/cl/CLTensorArgument.h" +#include "src/cl/CLTile.h" +#include "src/ITensor.h" +#include "src/Tensor3dMapper.h" +#include "src/TileView.h" + +namespace ckw +{ +void CLMemoryOpImage2dHelper::initialize(const CLTile *x, const CLTile *z, const CLTile *b) +{ + _coord_x = x->scalar(0, 0).str; + _coord_z = z->scalar(0, 0).str; + _coord_b = b->scalar(0, 0).str; +} + +void CLMemoryOpImage2dHelper::write_row(int32_t row_id, const std::string &coord_y) +{ + // The only check required is on Y. + out_of_bound_initialize_y(coord_y); + + const std::string dst = _dst.vector(row_id).str; + const std::string sampler = to_ls_image2d_sampler(); + const std::string coord = to_ls_image2d_address(_coord_x, coord_y, _coord_z, _coord_b); + const std::string ls_buf = to_ls_image2d(_op, _ls_width_full, dst, sampler, coord); + + _writer->op_write_raw_code(ls_buf + ";\n"); + + out_of_bound_finalize_y(); +} + +void CLMemoryOpImage2dHelper::finalize() +{ +} + +bool CLMemoryOpImage2dHelper::validate(const CLKernelWriter *writer, + const ITensor *tensor, + const TensorSampler *sampler, + const Tensor3dMapper *mapper, + MemoryOperation op, + const TileView<CLTile> &dst) +{ + CKW_UNUSED(writer, tensor, mapper); + + if (dst.width() != 4) + { + return false; + } + if (sampler->address_mode_x() != TensorSamplerAddressModeX::None) + { + return false; + } + if (sampler->address_mode_z() != TensorSamplerAddressModeZ::None) + { + return false; + } + if (sampler->storage() != TensorStorageType::Texture2dReadOnly && op == MemoryOperation::Load) + { + return false; + } + if (sampler->storage() != TensorStorageType::Texture2dWriteOnly && op == MemoryOperation::Store) + { + return false; + } + if ((dst.data_type() != DataType::Fp32) && (dst.data_type() != DataType::Fp16)) + { + return false; + } + return true; +} + +void CLMemoryOpImage2dHelper::out_of_bound_initialize_y(const std::string &coord) +{ + CKW_UNUSED(coord); + + const TensorSamplerAddressModeY address_mode_y = _sampler->address_mode_y(); + switch (address_mode_y) + { + case TensorSamplerAddressModeY::SkipLessThanZero: + _writer->op_write_raw_code("if(" + coord + " >= 0)\n{\n"); + break; + case TensorSamplerAddressModeY::ClampToBorderMaxOnly: + case TensorSamplerAddressModeY::None: + break; + default: + CKW_THROW_MSG("Unsupported address mode for Y dimension"); + } +} + +void CLMemoryOpImage2dHelper::out_of_bound_finalize_y() +{ + const TensorSamplerAddressModeY address_mode_y = _sampler->address_mode_y(); + switch (address_mode_y) + { + case TensorSamplerAddressModeY::SkipLessThanZero: + _writer->op_write_raw_code("}\n"); + break; + case TensorSamplerAddressModeY::ClampToBorderMaxOnly: + case TensorSamplerAddressModeY::None: + break; + default: + CKW_THROW_MSG("Unsupported address mode for Y dimension"); + } +} + +std::string CLMemoryOpImage2dHelper::to_ls_image2d(MemoryOperation op, + int32_t vector_width, + const std::string &data, + const std::string &sampler, + const std::string &address) const +{ + CKW_UNUSED(vector_width); + CKW_ASSERT_MSG(_dst.data_type() == DataType::Fp32 || _dst.data_type() == DataType::Fp16, + "Image2d only supports floating-point data type"); + + const TensorStorageType tensor_storage = _sampler->storage(); + const std::string image2d_obj = _tensor->storage(tensor_storage).val; + const std::string post_fix = _dst.data_type() == DataType::Fp32 ? "f" : "h"; + + switch (op) + { + case MemoryOperation::Load: + return data + " = read_image" + post_fix + "(" + image2d_obj + ", " + sampler + ", " + address + ")"; + break; + case MemoryOperation::Store: + return "write_image" + post_fix + "(" + image2d_obj + ", " + address + ", " + data + ")"; + default: + CKW_THROW_MSG("Unsupported MemoryOperation"); + } +} + +std::string CLMemoryOpImage2dHelper::to_ls_image2d_sampler() const +{ + const auto address_mode_y = _sampler->address_mode_y(); + + switch (address_mode_y) + { + case TensorSamplerAddressModeY::None: + return "CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST"; + case TensorSamplerAddressModeY::SkipLessThanZero: + case TensorSamplerAddressModeY::ClampToBorderMaxOnly: + return "CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST"; + default: + CKW_THROW_MSG("Unsupported address_mode_coord"); + } +} + +std::string CLMemoryOpImage2dHelper::to_ls_image2d_address(const std::string &x, + const std::string &y, + const std::string &z, + const std::string &b) const +{ + std::string coord_x = "(" + x + ") >> 2"; + std::string coord_y = "("; + + if (y != "0") + { + coord_y += y; + } + if (z != "0" && (_mapper->dim_z().str != "1")) + { + const std::string dim = _mapper->dim_y().str; + coord_y += " + ("; + coord_y += z + ")"; + coord_y += " * "; + coord_y += dim; + } + if (b != "0" && (_mapper->dim_batch().str != "1")) + { + const std::string dim0 = _mapper->dim_y().str; + const std::string dim1 = _mapper->dim_z().str; + coord_y += " + ("; + coord_y += b + ")"; + coord_y += " * "; + coord_y += dim0; + coord_y += " * "; + coord_y += dim1; + } + coord_y += ")"; + return "(int2)(" + coord_x + ", " + coord_y + ")"; +} + +} // namespace ckw diff --git a/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.h b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.h new file mode 100644 index 0000000000..6c42c132d9 --- /dev/null +++ b/compute_kernel_writer/src/cl/helpers/CLMemoryOpImage2dHelper.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2023 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 CKW_SRC_CL_HELPERS_CLMEMORYOPIMAGE2DHELPER_H +#define CKW_SRC_CL_HELPERS_CLMEMORYOPIMAGE2DHELPER_H + +#include "src/cl/helpers/ICLMemoryOpHelper.h" + +#include <string> + +namespace ckw +{ + +// Forward Declarations +class CLKernelWriter; +class CLTile; +template <class CLTile> +class TileView; +enum class MemoryOperation; + +/** Helper class to write memory operations (like load/store) in OpenCL for Image2d type */ +class CLMemoryOpImage2dHelper : public ICLMemoryOpHelper +{ +public: + /** Constructor similar to @ref ICLMemoryOpHelper() */ + CLMemoryOpImage2dHelper(CLKernelWriter *writer, + ITensor *tensor, + TensorSampler *sampler, + MemoryOperation op, + const TileView<CLTile> &dst) + : ICLMemoryOpHelper(writer, tensor, sampler, op, dst) + { + } + + /** Copy constructor */ + CLMemoryOpImage2dHelper(const CLMemoryOpImage2dHelper &) = delete; + + /** Assignment operator overload */ + CLMemoryOpImage2dHelper &operator=(const CLMemoryOpImage2dHelper &) = delete; + + // Methods overridden + void initialize(const CLTile *x, const CLTile *z, const CLTile *b) override; + void write_row(int32_t row_id, const std::string &coord_y) override; + void finalize() override; + +private: + static bool validate(const CLKernelWriter *writer, + const ITensor *tensor, + const TensorSampler *sampler, + const Tensor3dMapper *mapper, + MemoryOperation op, + const TileView<CLTile> &dst); + + void out_of_bound_initialize_y(const std::string &coord); + void out_of_bound_finalize_y(); + + std::string to_ls_image2d(MemoryOperation op, + int32_t vector_width, + const std::string &data, + const std::string &sampler, + const std::string &address) const; + std::string to_ls_image2d_sampler() const; + std::string + to_ls_image2d_address(const std::string &x, const std::string &y, const std::string &z, const std::string &b) const; +}; +} // namespace ckw + +#endif // CKW_SRC_CL_HELPERS_CLMEMORYOPIMAGE2DHELPER_H diff --git a/compute_kernel_writer/src/cl/helpers/ICLMemoryOpHelper.h b/compute_kernel_writer/src/cl/helpers/ICLMemoryOpHelper.h new file mode 100644 index 0000000000..a5b679ac03 --- /dev/null +++ b/compute_kernel_writer/src/cl/helpers/ICLMemoryOpHelper.h @@ -0,0 +1,121 @@ +/* + * Copyright (c) 2023 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 CKW_SRC_CL_HELPERS_ICLMEMORYOPHELPER_H +#define CKW_SRC_CL_HELPERS_ICLMEMORYOPHELPER_H + +#include "ckw/TensorSampler.h" + +#include "src/Tensor3dMapper.h" +#include "src/TileView.h" + +#include <cstdint> +#include <memory> +#include <string> + +namespace ckw +{ + +// Forward Declarations +class CLTile; +class CLKernelWriter; +class ITensor; +class TensorSampler; +enum class MemoryOperation; + +/** Base class OpenCL memory operation helper classes + * that helps writing code for memory operations like load/store. + */ +class ICLMemoryOpHelper +{ +public: + /** Constructor + * + * @param[in] writer @ref ckw::CLKernelWriter object to write the code + * @param[in] tensor @ref ckw::ITensor object to perform the memory operation on + * @param[in] sampler @ref ckw::TensorSampler object that tells how to sample a tensor + * @param[in] op The memory operation to be done (e.g. Load/Store) + * @param[in] dst The tile to perform the memory operation on + */ + ICLMemoryOpHelper(CLKernelWriter *writer, + ITensor *tensor, + TensorSampler *sampler, + MemoryOperation op, + const TileView<CLTile> &dst) + : _writer(writer), _tensor(tensor), _sampler(sampler), _op(op), _dst(dst) + { + _mapper = std::make_unique<Tensor3dMapper>(tensor, sampler->format()); + _ls_width_full = _dst.width(); + } + + /** Copy constructor */ + ICLMemoryOpHelper(const ICLMemoryOpHelper &) = delete; + + /** Assignment operator overload */ + ICLMemoryOpHelper &operator=(const ICLMemoryOpHelper &) = delete; + + /** Destructor */ + virtual ~ICLMemoryOpHelper() = default; + + /** Initialization method that takes a 3D tensor's x, z dimensions and + * the batch offset as a tile object, and initializes the code inside + * the writer object. + * + * @param[in] x tile object that describes the x-coordinate of the tensor involved + * @param[in] z tile object that describes the z-coordinate of the tensor involved + * @param[in] b tile object that describes the batch offset of the tensor involved + */ + virtual void initialize(const CLTile *x, const CLTile *z, const CLTile *b) = 0; + + /** Method that writes the actual code to the writer that performs the mentioned memory + * operation on the tile initialized. It writes the code for a specific row given in the + * arguments. + * + * @param[in] row_id row id + * @param[in] coord_y y-coordinate as string + */ + virtual void write_row(int32_t row_id, const std::string &coord_y) = 0; + + /** Method that finalizes the code in the writer object. This part is usually for taking + * care of finalizing anything that's been initialized inside @ref IMemoryHelper::initialize() + * such as matching compound statements, checking certain boundary conditions etc. No inputs + * and/or outputs, only the writer object is affected. + */ + virtual void finalize() = 0; + +protected: + CLKernelWriter *_writer{nullptr}; + ITensor *_tensor{nullptr}; + TensorSampler *_sampler{nullptr}; + MemoryOperation _op; + std::unique_ptr<Tensor3dMapper> _mapper{nullptr}; + TileView<CLTile> _dst{}; + int32_t _ls_width_full{0}; + std::string _coord_x{}; + std::string _coord_z{}; + std::string _coord_b{}; +}; +} // namespace ckw + +#endif // CKW_SRC_CL_HELPERS_ICLMEMORYOPHELPER_H |