1 // Copyright 2020 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 <stddef.h>
9 #include <stdint.h>
10
11 #include <xnnpack.h>
12 #include <xnnpack/log.h>
13 #include <xnnpack/operator.h>
14 #include <xnnpack/params.h>
15 #include <xnnpack/requantization.h>
16 #include <xnnpack/subgraph.h>
17 #include <xnnpack/subgraph-validation.h>
18
19
create_max_pooling_operator(const struct xnn_node * node,const struct xnn_value * values,size_t num_values,struct xnn_operator_data * opdata,const struct xnn_caches * caches)20 static enum xnn_status create_max_pooling_operator(
21 const struct xnn_node* node,
22 const struct xnn_value* values,
23 size_t num_values,
24 struct xnn_operator_data* opdata,
25 const struct xnn_caches* caches)
26 {
27 assert(node->num_inputs == 1);
28 const uint32_t input_id = node->inputs[0];
29 assert(input_id != XNN_INVALID_VALUE_ID);
30 assert(input_id < num_values);
31
32 assert(node->num_outputs == 1);
33 const uint32_t output_id = node->outputs[0];
34 assert(output_id != XNN_INVALID_VALUE_ID);
35 assert(output_id < num_values);
36
37 const size_t channel_dim = values[input_id].shape.dim[3];
38 assert(channel_dim == values[output_id].shape.dim[3]);
39
40 enum xnn_status status;
41 switch (node->compute_type) {
42 #ifndef XNN_NO_F16_OPERATORS
43 case xnn_compute_type_fp16:
44 status = xnn_create_max_pooling2d_nhwc_f16(
45 node->params.pooling_2d.padding_top,
46 node->params.pooling_2d.padding_right,
47 node->params.pooling_2d.padding_bottom,
48 node->params.pooling_2d.padding_left,
49 node->params.pooling_2d.pooling_height,
50 node->params.pooling_2d.pooling_width,
51 node->params.pooling_2d.stride_height,
52 node->params.pooling_2d.stride_width,
53 node->params.pooling_2d.dilation_height,
54 node->params.pooling_2d.dilation_width,
55 channel_dim /* channels */, channel_dim /* input stride */, channel_dim /* output stride */,
56 node->activation.output_min,
57 node->activation.output_max,
58 node->flags,
59 &opdata->operator_objects[0]);
60 break;
61 #endif // !defined(XNN_NO_F16_OPERATORS)
62 case xnn_compute_type_fp32:
63 status = xnn_create_max_pooling2d_nhwc_f32(
64 node->params.pooling_2d.padding_top,
65 node->params.pooling_2d.padding_right,
66 node->params.pooling_2d.padding_bottom,
67 node->params.pooling_2d.padding_left,
68 node->params.pooling_2d.pooling_height,
69 node->params.pooling_2d.pooling_width,
70 node->params.pooling_2d.stride_height,
71 node->params.pooling_2d.stride_width,
72 node->params.pooling_2d.dilation_height,
73 node->params.pooling_2d.dilation_width,
74 channel_dim /* channels */, channel_dim /* input stride */, channel_dim /* output stride */,
75 node->activation.output_min,
76 node->activation.output_max,
77 node->flags,
78 &opdata->operator_objects[0]);
79 break;
80 #ifndef XNN_NO_S8_OPERATORS
81 case xnn_compute_type_qs8:
82 {
83 const float output_scale = values[output_id].quantization.scale;
84 const int32_t output_zero_point = values[output_id].quantization.zero_point;
85 const int8_t output_min = xnn_qs8_quantize(node->activation.output_min, output_scale, output_zero_point);
86 const int8_t output_max = xnn_qs8_quantize(node->activation.output_max, output_scale, output_zero_point);
87 status = xnn_create_max_pooling2d_nhwc_s8(
88 node->params.pooling_2d.padding_top,
89 node->params.pooling_2d.padding_right,
90 node->params.pooling_2d.padding_bottom,
91 node->params.pooling_2d.padding_left,
92 node->params.pooling_2d.pooling_height,
93 node->params.pooling_2d.pooling_width,
94 node->params.pooling_2d.stride_height,
95 node->params.pooling_2d.stride_width,
96 node->params.pooling_2d.dilation_height,
97 node->params.pooling_2d.dilation_width,
98 channel_dim /* channels */, channel_dim /* input stride */, channel_dim /* output stride */,
99 output_min,
100 output_max,
101 node->flags,
102 &opdata->operator_objects[0]);
103 break;
104 }
105 #endif // !defined(XNN_NO_S8_OPERATORS)
106 #ifndef XNN_NO_U8_OPERATORS
107 case xnn_compute_type_qu8:
108 {
109 const float output_scale = values[output_id].quantization.scale;
110 const int32_t output_zero_point = values[output_id].quantization.zero_point;
111 const uint8_t output_min = xnn_qu8_quantize(node->activation.output_min, output_scale, output_zero_point);
112 const uint8_t output_max = xnn_qu8_quantize(node->activation.output_max, output_scale, output_zero_point);
113 status = xnn_create_max_pooling2d_nhwc_u8(
114 node->params.pooling_2d.padding_top,
115 node->params.pooling_2d.padding_right,
116 node->params.pooling_2d.padding_bottom,
117 node->params.pooling_2d.padding_left,
118 node->params.pooling_2d.pooling_height,
119 node->params.pooling_2d.pooling_width,
120 node->params.pooling_2d.stride_height,
121 node->params.pooling_2d.stride_width,
122 node->params.pooling_2d.dilation_height,
123 node->params.pooling_2d.dilation_width,
124 channel_dim /* channels */, channel_dim /* input stride */, channel_dim /* output stride */,
125 output_min,
126 output_max,
127 node->flags,
128 &opdata->operator_objects[0]);
129 break;
130 }
131 #endif // !defined(XNN_NO_U8_OPERATORS)
132 default:
133 XNN_UNREACHABLE;
134 }
135 if (status == xnn_status_success) {
136 opdata->batch_size = values[input_id].shape.dim[0];
137 opdata->input_height = values[input_id].shape.dim[1];
138 opdata->input_width = values[input_id].shape.dim[2];
139 opdata->inputs[0] = input_id;
140 opdata->outputs[0] = output_id;
141 }
142 return status;
143 }
144
setup_max_pooling_operator(const struct xnn_operator_data * opdata,const struct xnn_blob * blobs,size_t num_blobs,pthreadpool_t threadpool)145 static enum xnn_status setup_max_pooling_operator(
146 const struct xnn_operator_data* opdata,
147 const struct xnn_blob* blobs,
148 size_t num_blobs,
149 pthreadpool_t threadpool)
150 {
151 const uint32_t input_id = opdata->inputs[0];
152 assert(input_id != XNN_INVALID_VALUE_ID);
153 assert(input_id < num_blobs);
154
155 const uint32_t output_id = opdata->outputs[0];
156 assert(output_id != XNN_INVALID_VALUE_ID);
157 assert(output_id < num_blobs);
158
159 const struct xnn_blob* input_blob = blobs + input_id;
160 const void* input_data = input_blob->data;
161 assert(input_data != NULL);
162
163 const struct xnn_blob* output_blob = blobs + output_id;
164 void* output_data = output_blob->data;
165 assert(output_data != NULL);
166
167 switch (opdata->operator_objects[0]->type) {
168 #ifndef XNN_NO_F16_OPERATORS
169 case xnn_operator_type_max_pooling_nhwc_f16:
170 return xnn_setup_max_pooling2d_nhwc_f16(
171 opdata->operator_objects[0],
172 opdata->batch_size,
173 opdata->input_height,
174 opdata->input_width,
175 input_data,
176 output_data,
177 threadpool);
178 #endif // !defined(XNN_NO_F16_OPERATORS)
179 case xnn_operator_type_max_pooling_nhwc_f32:
180 return xnn_setup_max_pooling2d_nhwc_f32(
181 opdata->operator_objects[0],
182 opdata->batch_size,
183 opdata->input_height,
184 opdata->input_width,
185 input_data,
186 output_data,
187 threadpool);
188 #ifndef XNN_NO_S8_OPERATORS
189 case xnn_operator_type_max_pooling_nhwc_s8:
190 return xnn_setup_max_pooling2d_nhwc_s8(
191 opdata->operator_objects[0],
192 opdata->batch_size,
193 opdata->input_height,
194 opdata->input_width,
195 input_data,
196 output_data,
197 threadpool);
198 #endif // !defined(XNN_NO_S8_OPERATORS)
199 #ifndef XNN_NO_U8_OPERATORS
200 case xnn_operator_type_max_pooling_nhwc_u8:
201 return xnn_setup_max_pooling2d_nhwc_u8(
202 opdata->operator_objects[0],
203 opdata->batch_size,
204 opdata->input_height,
205 opdata->input_width,
206 input_data,
207 output_data,
208 threadpool);
209 #endif // !defined(XNN_NO_U8_OPERATORS)
210 default:
211 XNN_UNREACHABLE;
212 }
213 }
214
xnn_define_max_pooling_2d(xnn_subgraph_t subgraph,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,uint32_t stride_height,uint32_t stride_width,uint32_t dilation_height,uint32_t dilation_width,float output_min,float output_max,uint32_t input_id,uint32_t output_id,uint32_t flags)215 enum xnn_status xnn_define_max_pooling_2d(
216 xnn_subgraph_t subgraph,
217 uint32_t input_padding_top,
218 uint32_t input_padding_right,
219 uint32_t input_padding_bottom,
220 uint32_t input_padding_left,
221 uint32_t pooling_height,
222 uint32_t pooling_width,
223 uint32_t stride_height,
224 uint32_t stride_width,
225 uint32_t dilation_height,
226 uint32_t dilation_width,
227 float output_min,
228 float output_max,
229 uint32_t input_id,
230 uint32_t output_id,
231 uint32_t flags)
232 {
233 enum xnn_status status;
234 if ((status = xnn_subgraph_check_xnnpack_initialized(xnn_node_type_max_pooling_2d)) != xnn_status_success) {
235 return status;
236 }
237
238 const uint32_t pooling_size = pooling_height * pooling_width;
239 if (pooling_size == 0) {
240 xnn_log_error(
241 "failed to define %s operator with %" PRIu32 "x%" PRIu32 " pooling size: "
242 "pooling size dimensions must be non-zero",
243 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), pooling_width, pooling_height);
244 return xnn_status_invalid_parameter;
245 }
246
247 if (pooling_size == 1) {
248 xnn_log_error(
249 "failed to define %s operator with 1 pooling element: 1x1 pooling is meaningless",
250 xnn_node_type_to_string(xnn_node_type_max_pooling_2d));
251 return xnn_status_invalid_parameter;
252 }
253
254 if (stride_height == 0 || stride_width == 0) {
255 xnn_log_error(
256 "failed to define %s operator with %" PRIu32 "x%" PRIu32 " stride: stride dimensions must be non-zero",
257 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), stride_width, stride_height);
258 return xnn_status_invalid_parameter;
259 }
260
261 if (dilation_height == 0 || dilation_width == 0) {
262 xnn_log_error(
263 "failed to define %s operator with %" PRIu32 "x%" PRIu32 " dilation: dilation dimensions must be non-zero",
264 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), dilation_width, dilation_height);
265 return xnn_status_invalid_parameter;
266 }
267
268 if (stride_height > pooling_height) {
269 xnn_log_error(
270 "failed to define %s operator with %" PRIu32 " stride height: must be less than pooling height %" PRIu32,
271 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), stride_height, pooling_height);
272 return xnn_status_invalid_parameter;
273 }
274
275 if (stride_width > pooling_width) {
276 xnn_log_error(
277 "failed to define %s operator with %" PRIu32 " stride width: must be less than pooling width %" PRIu32,
278 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), stride_width, pooling_width);
279 return xnn_status_invalid_parameter;
280 }
281
282 status = xnn_subgraph_check_output_min_max(xnn_node_type_max_pooling_2d, output_min, output_max);
283 if (status != xnn_status_success) {
284 return status;
285 }
286
287 const bool any_padding = (input_padding_left | input_padding_top | input_padding_right | input_padding_bottom) != 0;
288 if ((flags & XNN_FLAG_TENSORFLOW_SAME_PADDING) != 0) {
289 if (any_padding) {
290 xnn_log_error(
291 "failed to define %s operator with %" PRIu32 "+%" PRIu32 "x%" PRIu32 "+%" PRIu32" padding: "
292 "TensorFlow SAME padding can't be combined with explicit padding specification",
293 xnn_node_type_to_string(xnn_node_type_max_pooling_2d),
294 input_padding_top, input_padding_left, input_padding_bottom, input_padding_right);
295 return xnn_status_invalid_parameter;
296 }
297 }
298
299 if ((status = xnn_subgraph_check_input_node_id(xnn_node_type_max_pooling_2d, input_id, subgraph->num_values)) !=
300 xnn_status_success) {
301 return status;
302 }
303
304 const struct xnn_value* input_value = &subgraph->values[input_id];
305 status = xnn_subgraph_check_input_type_dense(xnn_node_type_max_pooling_2d, input_id, input_value);
306 if (status != xnn_status_success) {
307 return status;
308 }
309
310 switch (input_value->datatype) {
311 case xnn_datatype_fp32:
312 #ifndef XNN_NO_S8_OPERATORS
313 case xnn_datatype_qint8:
314 #endif // !defined(XNN_NO_S8_OPERATORS)
315 #ifndef XNN_NO_U8_OPERATORS
316 case xnn_datatype_quint8:
317 #endif // !defined(XNN_NO_U8_OPERATORS)
318 break;
319 default:
320 xnn_log_error(
321 "failed to define %s operator with input ID #%" PRIu32 ": unsupported Value datatype %s (%d)",
322 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), input_id,
323 xnn_datatype_to_string(input_value->datatype), input_value->datatype);
324 return xnn_status_invalid_parameter;
325 }
326
327 status = xnn_subgraph_check_output_node_id(xnn_node_type_max_pooling_2d, output_id, subgraph->num_values);
328 if (status != xnn_status_success) {
329 return status;
330 }
331
332 const struct xnn_value* output_value = &subgraph->values[output_id];
333 status = xnn_subgraph_check_output_type_dense(xnn_node_type_max_pooling_2d, output_id, output_value);
334 if (status != xnn_status_success) {
335 return status;
336 }
337
338 enum xnn_compute_type compute_type = xnn_compute_type_invalid;
339 switch (output_value->datatype) {
340 case xnn_datatype_fp32:
341 compute_type = xnn_compute_type_fp32;
342 break;
343 #ifndef XNN_NO_S8_OPERATORS
344 case xnn_datatype_qint8:
345 compute_type = xnn_compute_type_qs8;
346 break;
347 #endif // !defined(XNN_NO_S8_OPERATORS)
348 #ifndef XNN_NO_U8_OPERATORS
349 case xnn_datatype_quint8:
350 compute_type = xnn_compute_type_qu8;
351 break;
352 #endif // !defined(XNN_NO_U8_OPERATORS)
353 default:
354 xnn_log_error(
355 "failed to define %s operator with output ID #%" PRIu32 ": unsupported Value datatype %s (%d)",
356 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), output_id,
357 xnn_datatype_to_string(output_value->datatype), output_value->datatype);
358 return xnn_status_invalid_parameter;
359 }
360
361 status = xnn_subgraph_check_datatype_matches(
362 xnn_node_type_max_pooling_2d, input_id, input_value, output_id, output_value);
363 if (status != xnn_status_success) {
364 return status;
365 }
366
367 #if !defined(XNN_NO_S8_OPERATORS) || !defined(XNN_NO_U8_OPERATORS)
368 if (output_value->datatype == xnn_datatype_qint8 || output_value->datatype == xnn_datatype_quint8) {
369 if (input_value->quantization.zero_point != output_value->quantization.zero_point) {
370 xnn_log_error(
371 "failed to define %s operator with input ID #%" PRIu32 " and output ID #%" PRIu32
372 ": mismatching zero point quantization parameter across input (%"PRId32") and output (%"PRId32")",
373 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), input_id, output_id,
374 input_value->quantization.zero_point, output_value->quantization.zero_point);
375 return xnn_status_invalid_parameter;
376 }
377 if (input_value->quantization.scale != output_value->quantization.scale) {
378 xnn_log_error(
379 "failed to define %s operator with input ID #%" PRIu32 " and output ID #%" PRIu32
380 ": mismatching zero point quantization parameter across input (%.7g) and output (%.7g)",
381 xnn_node_type_to_string(xnn_node_type_max_pooling_2d), input_id, output_id,
382 input_value->quantization.scale, output_value->quantization.scale);
383 return xnn_status_invalid_parameter;
384 }
385 }
386 #endif // !defined(XNN_NO_S8_OPERATORS) || !defined(XNN_NO_U8_OPERATORS)
387
388 struct xnn_node* node = xnn_subgraph_new_node(subgraph);
389 if (node == NULL) {
390 return xnn_status_out_of_memory;
391 }
392
393 node->type = xnn_node_type_max_pooling_2d;
394 node->compute_type = compute_type;
395 node->params.pooling_2d.padding_top = input_padding_top;
396 node->params.pooling_2d.padding_right = input_padding_right;
397 node->params.pooling_2d.padding_bottom = input_padding_bottom;
398 node->params.pooling_2d.padding_left = input_padding_left;
399 node->params.pooling_2d.pooling_height = pooling_height;
400 node->params.pooling_2d.pooling_width = pooling_width;
401 node->params.pooling_2d.stride_height = stride_height;
402 node->params.pooling_2d.stride_width = stride_width;
403 node->params.pooling_2d.dilation_height = dilation_height;
404 node->params.pooling_2d.dilation_width = dilation_width;
405 node->activation.output_min = output_min;
406 node->activation.output_max = output_max;
407 node->num_inputs = 1;
408 node->inputs[0] = input_id;
409 node->num_outputs = 1;
410 node->outputs[0] = output_id;
411 node->flags = flags;
412
413 node->create = create_max_pooling_operator;
414 node->setup = setup_max_pooling_operator;
415
416 return xnn_status_success;
417 }
418