Sheri Zhang | b18252d | 2020-04-07 11:04:57 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2020 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 "arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h" |
| 25 | #include "arm_compute/core/CL/ICLTensor.h" |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 28 | #include "support/StringSupport.h" |
| 29 | |
| 30 | namespace arm_compute |
| 31 | { |
| 32 | namespace |
| 33 | { |
| 34 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) |
| 35 | { |
| 36 | ARM_COMPUTE_ERROR_ON_NULLPTR(input); |
| 37 | // Output auto inizialitation if not yet initialized |
| 38 | auto_init_if_empty(*output, *input); |
| 39 | |
| 40 | const uint32_t temp_num_elems_processed_per_iteration = max_cl_vector_width / input->element_size(); |
| 41 | /* If width is less then step, then make step same as width to avoid global size being step instead of actual width. */ |
| 42 | /* Or we should fix in arm_compute::enqueue() or arm_compute::calculate_max_window(). */ |
| 43 | const uint32_t num_elems_processed_per_iteration = (input->dimension(0) < temp_num_elems_processed_per_iteration) ? input->dimension(0) : temp_num_elems_processed_per_iteration; |
| 44 | |
| 45 | // This kernel doesn't need padding |
| 46 | Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); |
| 47 | output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); |
| 48 | |
| 49 | return std::make_pair(Status{}, win); |
| 50 | } |
| 51 | Status validate_arguments(const ITensorInfo *input, ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias) |
| 52 | { |
| 53 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weight, bias, output); |
| 54 | |
| 55 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions"); |
| 56 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(weight->num_dimensions() > 1, "Weight tensor cannot have more than 1 dimensions"); |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->num_dimensions() > 1, "Bias tensor cannot have more than 1 dimensions"); |
| 58 | |
| 59 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QSYMM16); |
| 60 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weight, 1, DataType::QSYMM16); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); |
| 62 | |
| 63 | ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().x() != weight->tensor_shape().x()); |
| 64 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(weight, bias); |
| 65 | |
| 66 | // Checks performed when output is configured |
| 67 | if(output->total_size() != 0) |
| 68 | { |
| 69 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 70 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 71 | } |
| 72 | return Status{}; |
| 73 | } |
| 74 | } // namespace |
| 75 | |
| 76 | CLQLSTMLayerNormalizationKernel::CLQLSTMLayerNormalizationKernel() |
| 77 | : _input(nullptr), _weight(nullptr), _bias(nullptr), _output(nullptr) |
| 78 | { |
| 79 | } |
| 80 | |
| 81 | void CLQLSTMLayerNormalizationKernel::configure(CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const ICLTensor *weight, const ICLTensor *bias) |
| 82 | { |
| 83 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weight, bias, output); |
| 84 | |
| 85 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), weight->info(), bias->info())); |
| 86 | |
| 87 | _input = input; |
| 88 | _weight = weight; |
| 89 | _bias = bias; |
| 90 | _output = output; |
| 91 | |
| 92 | const uint32_t num_elems_processed_per_iteration = max_cl_vector_width / input->info()->element_size(); |
| 93 | |
| 94 | int32_t output_multiplier{}; |
| 95 | int32_t output_shift{}; |
| 96 | const UniformQuantizationInfo quan_info = _weight->info()->quantization_info().uniform(); |
| 97 | const Status status = quantization::calculate_quantized_multiplier(quan_info.scale, &output_multiplier, &output_shift); |
| 98 | output_shift *= -1; |
| 99 | |
| 100 | // Set build options |
| 101 | CLBuildOptions build_opts; |
| 102 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); |
| 103 | build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); |
| 104 | build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0))); |
| 105 | build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier)); |
| 106 | build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift)); |
| 107 | build_opts.add_option("-DMIN_BOUND=" + support::cpp11::to_string(std::get<0>(quantization::get_min_max_values_from_quantized_data_type(input->info()->data_type())))); |
| 108 | build_opts.add_option("-DMAX_BOUND=" + support::cpp11::to_string(std::get<1>(quantization::get_min_max_values_from_quantized_data_type(input->info()->data_type())))); |
| 109 | |
| 110 | // Create kernel |
| 111 | _kernel = create_kernel(compile_context, "qlstm_layer_normalization", build_opts.options()); |
| 112 | |
| 113 | // Configure kernel window |
| 114 | auto win_config = validate_and_configure_window(input->info(), output->info()); |
| 115 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 116 | ICLKernel::configure_internal(win_config.second); |
| 117 | |
| 118 | // Set config_id for enabling LWS tuning |
| 119 | _config_id = "qlstm_layer_normalization_"; |
| 120 | _config_id += lower_string(string_from_data_type(input->info()->data_type())); |
| 121 | _config_id += "_"; |
| 122 | _config_id += support::cpp11::to_string(input->info()->dimension(0)); |
| 123 | _config_id += "_"; |
| 124 | _config_id += support::cpp11::to_string(input->info()->dimension(1)); |
| 125 | } |
| 126 | |
| 127 | void CLQLSTMLayerNormalizationKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *weight, const ICLTensor *bias) |
| 128 | { |
| 129 | configure(CLKernelLibrary::get().get_compile_context(), input, output, weight, bias); |
| 130 | } |
| 131 | |
| 132 | Status CLQLSTMLayerNormalizationKernel::validate(const ITensorInfo *input, ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias) |
| 133 | { |
| 134 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, weight, bias)); |
| 135 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); |
| 136 | return Status{}; |
| 137 | } |
| 138 | |
| 139 | void CLQLSTMLayerNormalizationKernel::run(const Window &window, cl::CommandQueue &queue) |
| 140 | { |
| 141 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 142 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 143 | |
| 144 | Window slice = window.first_slice_window_2D(); |
| 145 | // Set slice step equal to width to force gws[0] to 1, as each thread normalizes across all rows |
| 146 | slice.set_dimension_step(Window::DimX, _input->info()->dimension(0)); |
| 147 | |
| 148 | Window weight_window; |
| 149 | Window weight_slice; |
| 150 | |
| 151 | weight_window.use_tensor_dimensions(_weight->info()->tensor_shape()); |
| 152 | weight_slice = weight_window.first_slice_window_1D(); |
| 153 | |
| 154 | do |
| 155 | { |
| 156 | unsigned int idx = 0; |
| 157 | add_2D_tensor_argument(idx, _input, slice); |
| 158 | add_1D_tensor_argument(idx, _weight, weight_slice); |
| 159 | add_1D_tensor_argument(idx, _bias, weight_slice); |
| 160 | add_2D_tensor_argument(idx, _output, slice); |
| 161 | |
| 162 | enqueue(queue, *this, slice, lws_hint()); |
| 163 | } |
| 164 | while(window.slide_window_slice_2D(slice)); |
| 165 | } |
| 166 | } // namespace arm_compute |