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/ClTransposedConvolutionKernel.h"
25
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
28 #include "src/core/CL/CLValidate.h"
29 #include "src/core/helpers/AutoConfiguration.h"
30 #include "src/core/helpers/WindowHelpers.h"
31 #include "support/Cast.h"
32
33 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
34
35 namespace arm_compute
36 {
37 namespace opencl
38 {
39 namespace kernels
40 {
41 namespace
42 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * output,const PadStrideInfo & deconv_info)43 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
44 const PadStrideInfo &deconv_info)
45 {
46 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
47 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QASYMM8_SIGNED, DataType::QASYMM8);
48 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
49 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
50 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(weights, DataLayout::NHWC);
51
52 constexpr unsigned int channel_idx = 0;
53 constexpr unsigned int width_idx = 1;
54 constexpr unsigned int height_idx = 2;
55 constexpr unsigned int batch_idx = 3;
56
57 ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(channel_idx) != input->dimension(channel_idx), "Weights feature map dimension should match the respective src's one");
58 ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->num_dimensions() > 4, "Weights can be at most 4 dimensional");
59
60 if(biases != nullptr)
61 {
62 if(is_data_type_quantized_asymmetric(input->data_type()))
63 {
64 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
65 }
66 else
67 {
68 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
69 }
70
71 ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->dimension(channel_idx) != weights->dimension(batch_idx),
72 "Biases size and number of dst feature maps should match");
73 ARM_COMPUTE_RETURN_ERROR_ON_MSG(biases->num_dimensions() > 1, "Biases should be one dimensional");
74 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
75 }
76
77 // Checks performed when output is configured
78 if(output->total_size() != 0)
79 {
80 const size_t input_width = input->dimension(width_idx);
81 const size_t input_height = input->dimension(height_idx);
82 const size_t weights_width = weights->dimension(width_idx);
83 const size_t weights_height = weights->dimension(height_idx);
84
85 auto out_dims = deconvolution_output_dimensions(input_width, input_height, weights_width, weights_height, deconv_info);
86 TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights);
87
88 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
89 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
90 ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(output, DataLayout::NHWC);
91 }
92
93 return Status{};
94 }
95 } // namespace
96
configure(const CLCompileContext & compile_context,const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,ITensorInfo * output,const PadStrideInfo & deconv_info)97 void ClTransposedConvolutionKernel::configure(const CLCompileContext &compile_context, const ITensorInfo *input, const ITensorInfo *weights,
98 const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &deconv_info)
99 {
100 ARM_COMPUTE_UNUSED(biases, deconv_info);
101 ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
102
103 // Perform validation
104 ARM_COMPUTE_ERROR_THROW_ON(validate(input, weights, biases, output, deconv_info));
105
106 constexpr unsigned int channel_idx = 0;
107 constexpr unsigned int width_idx = 1;
108 constexpr unsigned int height_idx = 2;
109
110 const size_t input_channels = input->dimension(channel_idx); // same as weight channels
111 const size_t input_width = input->dimension(width_idx);
112 const size_t input_height = input->dimension(height_idx);
113 const size_t weights_width = weights->dimension(width_idx);
114 const size_t weights_height = weights->dimension(height_idx);
115 const size_t output_width = output->dimension(width_idx);
116 const size_t output_height = output->dimension(height_idx);
117 const size_t output_channels = output->dimension(channel_idx);
118
119 // Calculate output shape
120 auto out_dims = deconvolution_output_dimensions(input_width, input_height, weights_width, weights_height, deconv_info);
121 TensorShape output_shape = misc::shape_calculator::compute_deconvolution_output_shape(out_dims, *input, *weights);
122 auto_init_if_empty(*output, output_shape, 1, input->data_type(), input->quantization_info());
123
124 // Calculate updated paddings
125 // p' = k - p - 1 (k: kernel dimensions)
126 const uint32_t pad_left = weights_width - deconv_info.pad_left() - 1;
127 const uint32_t pad_top = weights_height - deconv_info.pad_top() - 1;
128
129 // Configure kernel window
130 Window win;
131 output_shape.collapse(2U, 1U); // Collapse width and height into single dimension
132
133 const unsigned int n0 = adjust_vec_size(16 / output->element_size(), output_channels);
134 const unsigned int m0 = 1;
135 const unsigned int k0 = adjust_vec_size(16 / input->element_size(), input_channels);
136 const unsigned int partial_store_n0 = output_channels % n0;
137
138 // Create window and update padding
139 win = calculate_max_window(output_shape, Steps(n0, m0));
140 ICLKernel::configure_internal(win);
141
142 const std::string kernel_name = "transposed_convolution_nhwc";
143 CLBuildOptions build_options;
144
145 const DataType input_data_type = input->data_type();
146 const PaddingInfo strides = deconv_info.stride();
147
148 if(biases != nullptr)
149 {
150 build_options.add_option(std::string("-DHAS_BIAS"));
151 build_options.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->data_type())));
152 }
153
154 const auto output_data_type = output->data_type();
155
156 build_options.add_option("-cl-fast-relaxed-math");
157 build_options.add_option("-DSRC_TENSOR_TYPE=BUFFER");
158 build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(input_data_type));
159 build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(input_channels));
160 build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width));
161 build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input_height));
162 build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(output_channels));
163 build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(output_width));
164 build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(output_height));
165 build_options.add_option("-DDST_TENSOR_TYPE=BUFFER");
166 build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(output_data_type));
167 build_options.add_option("-DWEI_TENSOR_TYPE=BUFFER");
168 build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights_width));
169 build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights_height));
170 build_options.add_option("-DWEI_DATA_TYPE=" + get_cl_type_from_data_type(weights->data_type()));
171 build_options.add_option("-DSTRIDE_X=" + support::cpp11::to_string(strides.first));
172 build_options.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(strides.second));
173 build_options.add_option("-DPAD_LEFT=" + support::cpp11::to_string(pad_left));
174 build_options.add_option("-DPAD_TOP=" + support::cpp11::to_string(pad_top));
175 build_options.add_option("-DN0=" + support::cpp11::to_string(n0));
176 build_options.add_option("-DM0=" + support::cpp11::to_string(m0));
177 build_options.add_option("-DK0=" + support::cpp11::to_string(k0));
178 build_options.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0));
179 build_options.add_option_if((input_channels % k0) != 0, "-DLEFTOVER_LOOP");
180
181 if(is_data_type_quantized(output_data_type))
182 {
183 const UniformQuantizationInfo iqinfo = input->quantization_info().uniform();
184 const UniformQuantizationInfo wqinfo = weights->quantization_info().uniform();
185 const UniformQuantizationInfo oqinfo = output->quantization_info().uniform();
186
187 PixelValue zero_value = PixelValue(0, input->data_type(), input->quantization_info());
188 int zero_value_s32;
189 zero_value.get(zero_value_s32);
190
191 float multiplier = iqinfo.scale * wqinfo.scale / oqinfo.scale;
192 int output_multiplier = 0;
193 int output_shift = 0;
194
195 quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
196 build_options.add_option("-DIS_QUANTIZED");
197 build_options.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
198 build_options.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift));
199 build_options.add_option("-DSRC_OFFSET=" + support::cpp11::to_string(-iqinfo.offset));
200 build_options.add_option("-DWEI_OFFSET=" + support::cpp11::to_string(-wqinfo.offset));
201 build_options.add_option("-DDST_OFFSET=" + support::cpp11::to_string(oqinfo.offset));
202 build_options.add_option("-DZERO_VALUE=" + support::cpp11::to_string(zero_value_s32));
203 build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(DataType::S32));
204 }
205 else
206 {
207 build_options.add_option("-DACC_DATA_TYPE=" + get_cl_type_from_data_type(input_data_type));
208 build_options.add_option("-DZERO_VALUE=" + support::cpp11::to_string(0));
209 }
210
211 if(compile_context.get_ddk_version() >= 30)
212 {
213 build_options.add_option("-fregister-allocation=64");
214 }
215
216 _kernel = create_kernel(compile_context, kernel_name, build_options.options());
217
218 // Set config_id for enabling LWS tuning
219 _config_id = kernel_name;
220 _config_id += "_";
221 _config_id += lower_string(string_from_data_type(input_data_type));
222 _config_id += "_";
223 _config_id += support::cpp11::to_string(weights_width);
224 _config_id += "_";
225 _config_id += support::cpp11::to_string(strides.first);
226 _config_id += "_";
227 _config_id += support::cpp11::to_string(strides.second);
228 _config_id += "_";
229 _config_id += support::cpp11::to_string(output_width);
230 _config_id += "_";
231 _config_id += support::cpp11::to_string(m0);
232 _config_id += "_";
233 _config_id += support::cpp11::to_string(n0);
234 }
235
validate(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * dst,const PadStrideInfo & deconv_info)236 Status ClTransposedConvolutionKernel::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases,
237 const ITensorInfo *dst, const PadStrideInfo &deconv_info)
238 {
239 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, deconv_info));
240 return Status{};
241 }
242
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)243 void ClTransposedConvolutionKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
244 {
245 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
246 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
247
248 // Get initial windows
249 Window slice = window.first_slice_window_3D();
250
251 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
252 const auto weights = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
253 const auto biases = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2));
254 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
255
256 unsigned int idx = 0;
257 add_4d_tensor_nhwc_argument(idx, src);
258 add_4d_tensor_nhwc_argument(idx, dst);
259
260 add_4d_tensor_nhwc_argument(idx, weights);
261 if(biases != nullptr)
262 {
263 add_1D_tensor_argument(idx, biases, slice);
264 }
265
266 enqueue(queue, *this, slice, lws_hint());
267 }
268 } // namespace kernels
269 } // namespace opencl
270 } // namespace arm_compute
271