1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7 #include <math.h>
8 #include <stdbool.h>
9 #include <stddef.h>
10 #include <stdint.h>
11 #include <stdlib.h>
12 #include <string.h>
13
14 #include <xnnpack.h>
15 #include <xnnpack/allocator.h>
16 #include <xnnpack/operator.h>
17 #include <xnnpack/log.h>
18 #include <xnnpack/common.h>
19 #include <xnnpack/math.h>
20 #include <xnnpack/params.h>
21 #include <xnnpack/indirection.h>
22
23
compute_output_dimension(size_t input_dimension,size_t input_padding_dimension,size_t kernel_dimension)24 static inline size_t compute_output_dimension(
25 size_t input_dimension,
26 size_t input_padding_dimension,
27 size_t kernel_dimension)
28 {
29 return doz(kernel_dimension * input_dimension, input_padding_dimension);
30 }
31
xnn_create_unpooling2d_nhwc_x32(uint32_t input_padding_top,uint32_t input_padding_right,uint32_t input_padding_bottom,uint32_t input_padding_left,uint32_t pooling_height,uint32_t pooling_width,size_t channels,size_t input_pixel_stride,size_t output_pixel_stride,uint32_t flags,xnn_operator_t * unpooling_op_out)32 enum xnn_status xnn_create_unpooling2d_nhwc_x32(
33 uint32_t input_padding_top,
34 uint32_t input_padding_right,
35 uint32_t input_padding_bottom,
36 uint32_t input_padding_left,
37 uint32_t pooling_height,
38 uint32_t pooling_width,
39 size_t channels,
40 size_t input_pixel_stride,
41 size_t output_pixel_stride,
42 uint32_t flags,
43 xnn_operator_t* unpooling_op_out)
44 {
45 xnn_operator_t unpooling_op = NULL;
46 enum xnn_status status = xnn_status_uninitialized;
47
48 if (!xnn_params.initialized) {
49 xnn_log_error("failed to create Unpooling operator: XNNPACK is not initialized");
50 goto error;
51 }
52
53 status = xnn_status_invalid_parameter;
54
55 const uint32_t pooling_size = pooling_height * pooling_width;
56 if (pooling_size == 0) {
57 xnn_log_error(
58 "failed to create Unpooling operator with %" PRIu32 "x%" PRIu32 " pooling size: "
59 "pooling size dimensions must be non-zero",
60 pooling_width, pooling_height);
61 goto error;
62 }
63
64 if (pooling_size == 1) {
65 xnn_log_error(
66 "failed to create Unpooling operator with 1 pooling element: 1x1 unpooling is meaningless");
67 goto error;
68 }
69
70 if (channels == 0) {
71 xnn_log_error(
72 "failed to create Unpooling operator with %zu channels: number of channels must be non-zero",
73 channels);
74 goto error;
75 }
76
77 if (input_pixel_stride < channels) {
78 xnn_log_error(
79 "failed to create Unpooling operator with input pixel stride of %zu: "
80 "stride must be at least as large as the number of channels (%zu)",
81 input_pixel_stride, channels);
82 goto error;
83 }
84
85 if (output_pixel_stride < channels) {
86 xnn_log_error(
87 "failed to create Unpooling operator with output pixel stride of %zu: "
88 "stride must be at least as large as the number of channels (%zu)",
89 output_pixel_stride, channels);
90 goto error;
91 }
92
93 status = xnn_status_out_of_memory;
94
95 unpooling_op = xnn_allocate_zero_simd_memory(sizeof(struct xnn_operator));
96 if (unpooling_op == NULL) {
97 xnn_log_error("failed to allocate %zu bytes for Unpooling operator descriptor", sizeof(struct xnn_operator));
98 goto error;
99 }
100
101 unpooling_op->padding_top = input_padding_top;
102 unpooling_op->padding_right = input_padding_right;
103 unpooling_op->padding_bottom = input_padding_bottom;
104 unpooling_op->padding_left = input_padding_left;
105
106 unpooling_op->kernel_height = pooling_height;
107 unpooling_op->kernel_width = pooling_width;
108 unpooling_op->channels = channels;
109 unpooling_op->input_pixel_stride = input_pixel_stride;
110 unpooling_op->output_pixel_stride = output_pixel_stride;
111
112 unpooling_op->type = xnn_operator_type_unpooling_nhwc_x32;
113 unpooling_op->ukernel.type = xnn_ukernel_type_unpooling;
114
115 unpooling_op->state = xnn_run_state_invalid;
116
117 *unpooling_op_out = unpooling_op;
118 return xnn_status_success;
119
120 error:
121 xnn_delete_operator(unpooling_op);
122 return status;
123 }
124
xnn_setup_unpooling2d_nhwc_x32(xnn_operator_t unpooling_op,size_t batch_size,size_t input_height,size_t input_width,const void * input,const uint32_t * index,void * output,pthreadpool_t threadpool)125 enum xnn_status xnn_setup_unpooling2d_nhwc_x32(
126 xnn_operator_t unpooling_op,
127 size_t batch_size,
128 size_t input_height,
129 size_t input_width,
130 const void* input,
131 const uint32_t* index,
132 void* output,
133 pthreadpool_t threadpool)
134 {
135 if (unpooling_op->type != xnn_operator_type_unpooling_nhwc_x32) {
136 xnn_log_error("failed to setup Unpooling (X32) operator: operator type mismatch");
137 return xnn_status_invalid_parameter;
138 }
139 unpooling_op->state = xnn_run_state_invalid;
140
141 if (!xnn_params.initialized) {
142 xnn_log_error("failed to setup Unpooling operator: XNNPACK is not initialized");
143 return xnn_status_uninitialized;
144 }
145
146 if (input_width == 0 || input_height == 0) {
147 xnn_log_error(
148 "failed to setup Unpooling operator with %zux%zu input: input dimensions must be non-zero",
149 input_width, input_height);
150 return xnn_status_invalid_parameter;
151 }
152
153 if (batch_size == 0) {
154 unpooling_op->state = xnn_run_state_skip;
155 return xnn_status_success;
156 }
157
158 unpooling_op->batch_size = batch_size;
159 unpooling_op->input_height = input_height;
160 unpooling_op->input_width = input_width;
161 unpooling_op->input = input;
162
163 unpooling_op->output_height = compute_output_dimension(
164 input_height, unpooling_op->padding_top + unpooling_op->padding_bottom,
165 unpooling_op->kernel_height);
166 unpooling_op->output_width = compute_output_dimension(
167 input_width, unpooling_op->padding_left + unpooling_op->padding_right,
168 unpooling_op->kernel_width);
169 unpooling_op->output = output;
170
171 size_t valid_batch_size = 0;
172 if (output == unpooling_op->last_output &&
173 input_height == unpooling_op->last_input_height &&
174 input_width == unpooling_op->last_input_width)
175 {
176 valid_batch_size = unpooling_op->valid_batch_size;
177 if (batch_size <= valid_batch_size) {
178 unpooling_op->compute.range[0] = batch_size * input_height;
179 unpooling_op->state = xnn_run_state_ready;
180 return xnn_status_success;
181 }
182 }
183
184 const size_t pooling_height = unpooling_op->kernel_height;
185 const size_t pooling_width = unpooling_op->kernel_width;
186 const size_t pooling_size = pooling_height * pooling_width;
187
188 const size_t indirection_buffer_size = sizeof(void*) * (batch_size * input_height * input_width * pooling_size);
189
190 void** indirection_buffer = (void**) xnn_reallocate_memory(unpooling_op->indirection_buffer, indirection_buffer_size);
191 if (indirection_buffer == NULL) {
192 xnn_log_error("failed to allocate %zu bytes for indirection buffer", indirection_buffer_size);
193 return xnn_status_out_of_memory;
194 }
195 unpooling_op->indirection_buffer = (const void**) indirection_buffer;
196
197 xnn_indirection_init_unpool2d(unpooling_op, valid_batch_size, 2 /* log2(sizeof(type32)) */);
198
199 const size_t channels = unpooling_op->channels;
200 const size_t input_pixel_stride_in_bytes = unpooling_op->input_pixel_stride * sizeof(float);
201 unpooling_op->context.unpooling = (struct unpooling_context) {
202 .input = input,
203 .input_height_stride = input_width * input_pixel_stride_in_bytes,
204 .input_width_stride = input_pixel_stride_in_bytes,
205 .index = index,
206 .index_height_stride = input_width * channels * sizeof(uint32_t),
207 .index_width_stride = channels * sizeof(uint32_t),
208 .indirect_output = indirection_buffer,
209 .indirect_output_height_stride = input_width * pooling_size * sizeof(void*),
210 .indirect_output_width_stride = pooling_size * sizeof(void*),
211 .pooling_size = pooling_size,
212 .channels = channels,
213 .fill_value = 0,
214 .ukernel = xnn_params.x32.unpool,
215 };
216 unpooling_op->compute.type = xnn_parallelization_type_2d;
217 unpooling_op->compute.task_2d = (pthreadpool_task_2d_t) xnn_compute_unpooling;
218 unpooling_op->compute.range[0] = batch_size * input_height;
219 unpooling_op->compute.range[1] = input_width;
220 unpooling_op->state = xnn_run_state_ready;
221
222 unpooling_op->last_output = output;
223 unpooling_op->last_input_height = input_height;
224 unpooling_op->last_input_width = input_width;
225 unpooling_op->valid_batch_size = max(valid_batch_size, batch_size);
226
227 return xnn_status_success;
228 }
229