Gian Marco Iodice | 5d01681 | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2022 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 27 | #include "arm_compute/core/CL/ICLTensor.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
| 29 | #include "arm_compute/core/KernelDescriptors.h" |
| 30 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 31 | #include "src/core/CL/CLValidate.h" |
| 32 | #include "src/core/helpers/AutoConfiguration.h" |
| 33 | #include "src/core/helpers/WindowHelpers.h" |
| 34 | #include "support/Cast.h" |
| 35 | #include "support/StringSupport.h" |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | namespace opencl |
| 40 | { |
| 41 | namespace kernels |
| 42 | { |
| 43 | namespace |
| 44 | { |
| 45 | Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, |
| 46 | const PadStrideInfo &conv_info, const DirectConvComputeKernelInfo &desc) |
| 47 | { |
| 48 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src); |
| 49 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); |
| 50 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights); |
| 51 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC); |
| 52 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != src->dimension(0), "Weights feature map dimension should match the respective src's one"); |
| 53 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional"); |
| 54 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(desc.m0 <= 0 || desc.m0 > 8, "M0 can only be greater than 0 and less than or equal to 8"); |
| 55 | |
| 56 | // Checks performed when dst is configured |
| 57 | if(dst->total_size() != 0) |
| 58 | { |
| 59 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(dst->tensor_shape(), |
| 60 | misc::shape_calculator::compute_indirect_buffer_shape(src->tensor_shape(), |
| 61 | src->data_layout(), |
| 62 | weights->tensor_shape(), |
| 63 | conv_info, |
| 64 | desc)); |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst, 1, DataType::S32); |
| 66 | } |
| 67 | |
| 68 | return Status{}; |
| 69 | } |
| 70 | } // namespace |
| 71 | |
| 72 | ClIndirectConv2dAddressPrecalculationKernel::ClIndirectConv2dAddressPrecalculationKernel() |
| 73 | { |
| 74 | _type = CLKernelType::ELEMENTWISE; |
| 75 | } |
| 76 | |
| 77 | void ClIndirectConv2dAddressPrecalculationKernel::configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *dst, |
| 78 | const PadStrideInfo &conv_info, const DirectConvComputeKernelInfo &desc) |
| 79 | { |
| 80 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); |
| 81 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, dst, conv_info, desc)); |
| 82 | |
Gian Marco Iodice | 76335eb | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 83 | constexpr unsigned int width_idx = 1; |
| 84 | constexpr unsigned int height_idx = 2; |
Gian Marco Iodice | 5d01681 | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 85 | |
| 86 | // Get dst shape |
| 87 | TensorShape output_shape = misc::shape_calculator::compute_indirect_buffer_shape(src->tensor_shape(), |
| 88 | src->data_layout(), |
| 89 | weights->tensor_shape(), |
| 90 | conv_info, |
| 91 | desc); |
| 92 | |
| 93 | TensorShape output_conv_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *weights, conv_info); |
| 94 | |
| 95 | // Output auto inizialitation if not yet initialized |
| 96 | auto_init_if_empty(*dst, output_shape, 1, DataType::S32); |
| 97 | |
| 98 | // Configure kernel window |
| 99 | Window win; |
| 100 | |
| 101 | // Create window and update padding |
| 102 | win = calculate_max_window(output_shape, Steps(1)); |
| 103 | |
| 104 | ICLKernel::configure_internal(win); |
| 105 | |
| 106 | std::stringstream kernel_name; |
| 107 | CLBuildOptions build_options; |
| 108 | |
| 109 | kernel_name << "indirect_convolution_address_precalculation"; |
| 110 | |
| 111 | const unsigned int pad_left = conv_info.pad_left(); |
| 112 | const unsigned int pad_top = conv_info.pad_top(); |
| 113 | const unsigned int conv_stride_x = std::get<0>(conv_info.stride()); |
| 114 | const unsigned int conv_stride_y = std::get<1>(conv_info.stride()); |
| 115 | const auto dst_data_type = dst->data_type(); |
| 116 | |
| 117 | build_options.add_option("-DSRC_CONV_WIDTH=" + support::cpp11::to_string(src->dimension(width_idx))); |
| 118 | build_options.add_option("-DSRC_CONV_HEIGHT=" + support::cpp11::to_string(src->dimension(height_idx))); |
| 119 | build_options.add_option("-DDST_CONV_WIDTH=" + support::cpp11::to_string(output_conv_shape[width_idx])); |
| 120 | build_options.add_option("-DDST_CONV_HEIGHT=" + support::cpp11::to_string(output_conv_shape[height_idx])); |
| 121 | build_options.add_option("-DDST_TENSOR_TYPE=BUFFER"); |
| 122 | build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst_data_type)); |
| 123 | build_options.add_option("-DWEI_CONV_WIDTH=" + support::cpp11::to_string(weights->dimension(width_idx))); |
| 124 | build_options.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_stride_x)); |
| 125 | build_options.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_stride_y)); |
| 126 | build_options.add_option("-DPAD_LEFT=" + support::cpp11::to_string(pad_left)); |
| 127 | build_options.add_option("-DPAD_TOP=" + support::cpp11::to_string(pad_top)); |
| 128 | build_options.add_option("-DM0=" + support::cpp11::to_string(desc.m0)); |
| 129 | |
Gian Marco Iodice | 76335eb | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 130 | // A macro guard to compile ONLY the kernel of interest |
| 131 | build_options.add_option("-D" + upper_string(kernel_name.str())); |
| 132 | |
Gian Marco Iodice | 5d01681 | 2022-11-17 11:03:39 +0000 | [diff] [blame] | 133 | _kernel = create_kernel(compile_context, kernel_name.str(), build_options.options()); |
| 134 | |
| 135 | // Since this kernel should be called only once, we do not need to set the config_id for tuning |
| 136 | } |
| 137 | |
| 138 | Status ClIndirectConv2dAddressPrecalculationKernel::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, |
| 139 | const PadStrideInfo &conv_info, const DirectConvComputeKernelInfo &desc) |
| 140 | { |
| 141 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, dst, conv_info, desc)); |
| 142 | return Status{}; |
| 143 | } |
| 144 | |
| 145 | void ClIndirectConv2dAddressPrecalculationKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) |
| 146 | { |
| 147 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 148 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 149 | |
| 150 | // Get initial windows |
| 151 | const Window slice = window.first_slice_window_3D(); |
| 152 | |
| 153 | auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); |
| 154 | |
| 155 | unsigned int idx = 0; |
| 156 | add_4d_tensor_nhwc_argument(idx, dst); |
| 157 | enqueue(queue, *this, slice); |
| 158 | } |
| 159 | } // namespace kernels |
| 160 | } // namespace opencl |
| 161 | } // namespace arm_compute |