1 /*
2 * Copyright (c) 2020-2021 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/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h"
25 #include "arm_compute/core/CL/ICLTensor.h"
26 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
27 #include "src/core/helpers/AutoConfiguration.h"
28 #include "src/core/helpers/WindowHelpers.h"
29 #include "support/StringSupport.h"
30
31 namespace arm_compute
32 {
33 namespace
34 {
compute_output_qinfo()35 QuantizationInfo compute_output_qinfo()
36 {
37 return QuantizationInfo(1.f / 4096);
38 }
39
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)40 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
41 {
42 ARM_COMPUTE_ERROR_ON_NULLPTR(input);
43 // Output auto inizialitation if not yet initialized
44 auto_init_if_empty(*output, *input);
45 output->set_quantization_info(compute_output_qinfo());
46
47 const uint32_t temp_num_elems_processed_per_iteration = max_cl_vector_width / input->element_size();
48 /* If width is less then step, then make step same as width to avoid global size being step instead of actual width. */
49 /* Or we should fix in arm_compute::enqueue() or arm_compute::calculate_max_window(). */
50 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;
51
52 // This kernel doesn't need padding
53 Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
54
55 return std::make_pair(Status{}, win);
56 }
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * weight,const ITensorInfo * bias)57 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
58 {
59 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weight, bias, output);
60
61 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions");
62 ARM_COMPUTE_RETURN_ERROR_ON_MSG(weight->num_dimensions() > 1, "Weight tensor cannot have more than 1 dimensions");
63 ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->num_dimensions() > 1, "Bias tensor cannot have more than 1 dimensions");
64
65 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QSYMM16);
66 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weight, 1, DataType::QSYMM16);
67 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
68
69 ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().x() != weight->tensor_shape().x());
70 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(weight, bias);
71
72 // Checks performed when output is configured
73 if(output->total_size() != 0)
74 {
75 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
76 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
77 }
78 return Status{};
79 }
80 } // namespace
81
CLQLSTMLayerNormalizationKernel()82 CLQLSTMLayerNormalizationKernel::CLQLSTMLayerNormalizationKernel()
83 : _input(nullptr), _weight(nullptr), _bias(nullptr), _output(nullptr)
84 {
85 _type = CLKernelType::ELEMENTWISE;
86 }
87
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const ICLTensor * weight,const ICLTensor * bias)88 void CLQLSTMLayerNormalizationKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const ICLTensor *weight, const ICLTensor *bias)
89 {
90 ARM_COMPUTE_ERROR_ON_NULLPTR(input, weight, bias, output);
91 auto padding_info = get_padding_info({ input, weight, bias, output });
92
93 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), weight->info(), bias->info()));
94
95 _input = input;
96 _weight = weight;
97 _bias = bias;
98 _output = output;
99
100 const uint32_t num_elems_processed_per_iteration = max_cl_vector_width / input->info()->element_size();
101
102 int32_t output_multiplier{};
103 int32_t output_shift{};
104 const UniformQuantizationInfo quan_info = _weight->info()->quantization_info().uniform();
105 const Status status = quantization::calculate_quantized_multiplier(quan_info.scale, &output_multiplier, &output_shift);
106 output_shift *= -1;
107
108 // Set build options
109 CLBuildOptions build_opts;
110 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
111 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
112 build_opts.add_option("-DWIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
113 build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
114 build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
115 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()))));
116 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()))));
117
118 // Create kernel
119 _kernel = create_kernel(compile_context, "qlstm_layer_normalization", build_opts.options());
120
121 // Configure kernel window
122 auto win_config = validate_and_configure_window(input->info(), output->info());
123 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
124 ICLKernel::configure_internal(win_config.second);
125
126 // Set config_id for enabling LWS tuning
127 _config_id = "qlstm_layer_normalization_";
128 _config_id += lower_string(string_from_data_type(input->info()->data_type()));
129 _config_id += "_";
130 _config_id += support::cpp11::to_string(input->info()->dimension(0));
131 _config_id += "_";
132 _config_id += support::cpp11::to_string(input->info()->dimension(1));
133 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
134 }
135
configure(const ICLTensor * input,ICLTensor * output,const ICLTensor * weight,const ICLTensor * bias)136 void CLQLSTMLayerNormalizationKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *weight, const ICLTensor *bias)
137 {
138 configure(CLKernelLibrary::get().get_compile_context(), input, output, weight, bias);
139 }
140
validate(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * weight,const ITensorInfo * bias)141 Status CLQLSTMLayerNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias)
142 {
143 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, weight, bias));
144 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
145 return Status{};
146 }
147
run(const Window & window,cl::CommandQueue & queue)148 void CLQLSTMLayerNormalizationKernel::run(const Window &window, cl::CommandQueue &queue)
149 {
150 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
151 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
152
153 Window slice = window.first_slice_window_2D();
154 // Set slice step equal to width to force gws[0] to 1, as each thread normalizes across all rows
155 slice.set_dimension_step(Window::DimX, _input->info()->dimension(0));
156
157 Window weight_window;
158 Window weight_slice;
159
160 weight_window.use_tensor_dimensions(_weight->info()->tensor_shape());
161 weight_slice = weight_window.first_slice_window_1D();
162
163 do
164 {
165 unsigned int idx = 0;
166 add_2D_tensor_argument(idx, _input, slice);
167 add_1D_tensor_argument(idx, _weight, weight_slice);
168 add_1D_tensor_argument(idx, _bias, weight_slice);
169 add_2D_tensor_argument(idx, _output, slice);
170
171 enqueue(queue, *this, slice, lws_hint());
172 }
173 while(window.slide_window_slice_2D(slice));
174 }
175 } // namespace arm_compute
176