1 | /******************************************************************************* |
2 | * Copyright 2016-2022 Intel Corporation |
3 | * |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
5 | * you may not use this file except in compliance with the License. |
6 | * You may obtain a copy of the License at |
7 | * |
8 | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | * |
10 | * Unless required by applicable law or agreed to in writing, software |
11 | * distributed under the License is distributed on an "AS IS" BASIS, |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | * See the License for the specific language governing permissions and |
14 | * limitations under the License. |
15 | *******************************************************************************/ |
16 | |
17 | #ifndef COMMON_CONVOLUTION_PD_HPP |
18 | #define COMMON_CONVOLUTION_PD_HPP |
19 | |
20 | #include "oneapi/dnnl/dnnl.h" |
21 | |
22 | #include "c_types_map.hpp" |
23 | #include "primitive_desc.hpp" |
24 | #include "utils.hpp" |
25 | |
26 | namespace dnnl { |
27 | namespace impl { |
28 | |
29 | status_t conv_desc_init(convolution_desc_t *conv_desc, prop_kind_t prop_kind, |
30 | alg_kind_t alg_kind, const memory_desc_t *src_desc, |
31 | const memory_desc_t *weights_desc, const memory_desc_t *bias_desc, |
32 | const memory_desc_t *dst_desc, const dims_t strides, |
33 | const dims_t dilates, const dims_t padding_l, const dims_t padding_r); |
34 | |
35 | memory_desc_t *conv_prop_invariant_src_d(convolution_desc_t *desc); |
36 | memory_desc_t *conv_prop_invariant_wei_d(convolution_desc_t *desc); |
37 | memory_desc_t *conv_prop_invariant_bia_d(convolution_desc_t *desc); |
38 | memory_desc_t *conv_prop_invariant_dst_d(convolution_desc_t *desc); |
39 | const memory_desc_t *conv_prop_invariant_src_d(const convolution_desc_t *desc); |
40 | const memory_desc_t *conv_prop_invariant_wei_d(const convolution_desc_t *desc); |
41 | const memory_desc_t *conv_prop_invariant_bia_d(const convolution_desc_t *desc); |
42 | const memory_desc_t *conv_prop_invariant_dst_d(const convolution_desc_t *desc); |
43 | |
44 | struct convolution_fwd_pd_t; |
45 | |
46 | struct convolution_pd_t : public primitive_desc_t { |
47 | static constexpr auto base_pkind = primitive_kind::convolution; |
48 | |
49 | const convolution_desc_t *desc() const { return &desc_; } |
50 | const op_desc_t *op_desc() const override { |
51 | return reinterpret_cast<const op_desc_t *>(this->desc()); |
52 | } |
53 | |
54 | status_t query(query_t what, int idx, void *result) const override { |
55 | switch (what) { |
56 | case query::prop_kind: |
57 | *(prop_kind_t *)result = desc()->prop_kind; |
58 | break; |
59 | case query::alg_kind: |
60 | *(alg_kind_t *)result = desc()->alg_kind; |
61 | break; |
62 | case query::strides: |
63 | *(const dims_t **)result = &desc()->strides; |
64 | break; |
65 | case query::dilations: |
66 | *(const dims_t **)result = &desc()->dilates; |
67 | break; |
68 | case query::padding_l: |
69 | *(const dims_t **)result = &desc()->padding[0]; |
70 | break; |
71 | case query::padding_r: |
72 | *(const dims_t **)result = &desc()->padding[1]; |
73 | break; |
74 | default: return primitive_desc_t::query(what, idx, result); |
75 | } |
76 | return status::success; |
77 | } |
78 | |
79 | /* common conv aux functions */ |
80 | |
81 | dim_t MB() const { return invariant_src_md()->dims[0]; } |
82 | |
83 | dim_t IC() const { return invariant_src_md()->dims[1]; } |
84 | dim_t OC() const { return invariant_dst_md()->dims[1]; } |
85 | dim_t G() const { return with_groups() ? invariant_wei_md()->dims[0] : 1; } |
86 | |
87 | dim_t ID() const { |
88 | return ndims() >= 5 ? invariant_src_md()->dims[ndims() - 3] : 1; |
89 | } |
90 | dim_t IH() const { |
91 | return ndims() >= 4 ? invariant_src_md()->dims[ndims() - 2] : 1; |
92 | } |
93 | dim_t IW() const { return invariant_src_md()->dims[ndims() - 1]; } |
94 | |
95 | dim_t OD() const { |
96 | return ndims() >= 5 ? invariant_dst_md()->dims[ndims() - 3] : 1; |
97 | } |
98 | dim_t OH() const { |
99 | return ndims() >= 4 ? invariant_dst_md()->dims[ndims() - 2] : 1; |
100 | } |
101 | dim_t OW() const { return invariant_dst_md()->dims[ndims() - 1]; } |
102 | |
103 | dim_t KD() const { |
104 | return ndims() >= 5 |
105 | ? invariant_wei_md()->dims[ndims() + with_groups() - 3] |
106 | : 1; |
107 | } |
108 | dim_t KH() const { |
109 | return ndims() >= 4 |
110 | ? invariant_wei_md()->dims[ndims() + with_groups() - 2] |
111 | : 1; |
112 | } |
113 | dim_t KW() const { |
114 | return invariant_wei_md()->dims[ndims() + with_groups() - 1]; |
115 | } |
116 | |
117 | dim_t KSD() const { return ndims() >= 5 ? desc_.strides[ndims() - 5] : 1; } |
118 | dim_t KSH() const { return ndims() >= 4 ? desc_.strides[ndims() - 4] : 1; } |
119 | dim_t KSW() const { return desc_.strides[ndims() - 3]; } |
120 | |
121 | dim_t KDD() const { return ndims() >= 5 ? desc_.dilates[ndims() - 5] : 0; } |
122 | dim_t KDH() const { return ndims() >= 4 ? desc_.dilates[ndims() - 4] : 1; } |
123 | dim_t KDW() const { return desc_.dilates[ndims() - 3]; } |
124 | |
125 | dim_t padFront() const { |
126 | return ndims() >= 5 ? desc_.padding[0][ndims() - 5] : 0; |
127 | } |
128 | dim_t padBack() const { |
129 | return ndims() >= 5 ? desc_.padding[1][ndims() - 5] : 0; |
130 | } |
131 | dim_t padT() const { |
132 | return ndims() >= 4 ? desc_.padding[0][ndims() - 4] : 0; |
133 | } |
134 | dim_t padB() const { |
135 | return ndims() >= 4 ? desc_.padding[1][ndims() - 4] : 0; |
136 | } |
137 | dim_t padL() const { return desc_.padding[0][ndims() - 3]; } |
138 | dim_t padR() const { return desc_.padding[1][ndims() - 3]; } |
139 | |
140 | int ndims() const { return invariant_src_md()->ndims; } |
141 | |
142 | bool with_bias() const { |
143 | auto *bia_d = desc()->prop_kind == prop_kind::backward_weights |
144 | ? &desc()->diff_bias_desc |
145 | : &desc()->bias_desc; |
146 | return !memory_desc_wrapper(bia_d).is_zero(); |
147 | } |
148 | bool with_groups() const { |
149 | return invariant_wei_md()->ndims == ndims() + 1; |
150 | } |
151 | |
152 | bool is_fwd() const { |
153 | return utils::one_of(desc_.prop_kind, prop_kind::forward_training, |
154 | prop_kind::forward_inference); |
155 | } |
156 | |
157 | bool is_bwd_d() const { |
158 | return desc_.prop_kind == prop_kind::backward_data; |
159 | } |
160 | |
161 | bool is_bwd_w() const { |
162 | return desc_.prop_kind == prop_kind::backward_weights; |
163 | } |
164 | |
165 | bool has_zero_dim_memory() const { |
166 | const auto s_d = memory_desc_wrapper(*invariant_src_md()); |
167 | const auto d_d = memory_desc_wrapper(*invariant_dst_md()); |
168 | return s_d.has_zero_dim() || d_d.has_zero_dim(); |
169 | } |
170 | |
171 | const memory_desc_t *invariant_src_md() const { |
172 | return desc()->prop_kind == prop_kind::backward_data ? diff_src_md() |
173 | : src_md(); |
174 | } |
175 | const memory_desc_t *invariant_wei_md(int index = 0) const { |
176 | return desc()->prop_kind == prop_kind::backward_weights |
177 | ? diff_weights_md(index) |
178 | : weights_md(index); |
179 | } |
180 | const memory_desc_t *invariant_bia_md() const { |
181 | return invariant_wei_md(1); |
182 | } |
183 | const memory_desc_t *invariant_dst_md() const { |
184 | return is_fwd() ? dst_md() : diff_dst_md(); |
185 | } |
186 | memory_desc_t *invariant_src_md() { |
187 | auto *const_this = (const convolution_pd_t *)this; |
188 | return const_cast<memory_desc_t *>(const_this->invariant_src_md()); |
189 | } |
190 | memory_desc_t *invariant_wei_md(int index = 0) { |
191 | auto *const_this = (const convolution_pd_t *)this; |
192 | return const_cast<memory_desc_t *>(const_this->invariant_wei_md(index)); |
193 | } |
194 | memory_desc_t *invariant_bia_md() { |
195 | auto *const_this = (const convolution_pd_t *)this; |
196 | return const_cast<memory_desc_t *>(const_this->invariant_bia_md()); |
197 | } |
198 | memory_desc_t *invariant_dst_md() { |
199 | auto *const_this = (const convolution_pd_t *)this; |
200 | return const_cast<memory_desc_t *>(const_this->invariant_dst_md()); |
201 | } |
202 | |
203 | protected: |
204 | convolution_desc_t desc_; |
205 | const convolution_fwd_pd_t *hint_fwd_pd_; |
206 | |
207 | convolution_pd_t(const convolution_desc_t *adesc, |
208 | const primitive_attr_t *attr, |
209 | const convolution_fwd_pd_t *hint_fwd_pd) |
210 | : primitive_desc_t(attr, base_pkind) |
211 | , desc_(*adesc) |
212 | , hint_fwd_pd_(hint_fwd_pd) {} |
213 | |
214 | bool set_default_formats_common_template(memory_desc_t &src_md, |
215 | format_tag_t src_tag, memory_desc_t &wei_md, format_tag_t wei_tag, |
216 | memory_desc_t &dst_md, format_tag_t dst_tag, |
217 | memory_desc_t &bia_md) { |
218 | using namespace format_tag; |
219 | |
220 | #define IS_OK(f) \ |
221 | do { \ |
222 | if ((f) != status::success) return false; \ |
223 | } while (0) |
224 | if (src_md.format_kind == format_kind::any |
225 | && !utils::one_of(src_tag, any, undef)) |
226 | IS_OK(memory_desc_init_by_tag(src_md, src_tag)); |
227 | if (dst_md.format_kind == format_kind::any |
228 | && !utils::one_of(dst_tag, any, undef)) |
229 | IS_OK(memory_desc_init_by_tag(dst_md, dst_tag)); |
230 | if (wei_md.format_kind == format_kind::any |
231 | && !utils::one_of(wei_tag, any, undef)) |
232 | IS_OK(memory_desc_init_by_tag(wei_md, wei_tag)); |
233 | if (with_bias() && bia_md.format_kind == format_kind::any) |
234 | IS_OK(memory_desc_init_by_tag(bia_md, x)); |
235 | #undef IS_OK |
236 | |
237 | return true; |
238 | } |
239 | |
240 | bool set_default_alg_kind(alg_kind_t alg_kind) { |
241 | assert(utils::one_of(alg_kind, alg_kind::convolution_direct, |
242 | alg_kind::convolution_winograd)); |
243 | if (desc_.alg_kind == alg_kind::convolution_auto) |
244 | desc_.alg_kind = alg_kind; |
245 | return desc_.alg_kind == alg_kind; |
246 | } |
247 | |
248 | bool expect_data_types(data_type_t src_dt, data_type_t wei_dt, |
249 | data_type_t bia_dt, data_type_t dst_dt, data_type_t acc_dt) const { |
250 | bool ok = true |
251 | && (src_dt == data_type::undef |
252 | || invariant_src_md()->data_type == src_dt) |
253 | && (wei_dt == data_type::undef |
254 | || invariant_wei_md()->data_type == wei_dt) |
255 | && (dst_dt == data_type::undef |
256 | || invariant_dst_md()->data_type == dst_dt) |
257 | && (acc_dt == data_type::undef |
258 | || desc_.accum_data_type == acc_dt); |
259 | if (with_bias() && bia_dt != data_type::undef) |
260 | ok = ok && invariant_bia_md()->data_type == bia_dt; |
261 | return ok; |
262 | } |
263 | }; |
264 | |
265 | struct convolution_fwd_pd_t : public convolution_pd_t { |
266 | typedef convolution_fwd_pd_t base_class; |
267 | typedef convolution_fwd_pd_t hint_class; |
268 | |
269 | arg_usage_t arg_usage(int arg) const override { |
270 | if (utils::one_of(arg, DNNL_ARG_SRC, DNNL_ARG_WEIGHTS)) |
271 | return arg_usage_t::input; |
272 | |
273 | if (arg == DNNL_ARG_BIAS && with_bias()) return arg_usage_t::input; |
274 | |
275 | if (arg == DNNL_ARG_DST) return arg_usage_t::output; |
276 | |
277 | return primitive_desc_t::arg_usage(arg); |
278 | } |
279 | |
280 | const memory_desc_t *arg_md(int arg) const override { |
281 | switch (arg) { |
282 | case DNNL_ARG_SRC: return src_md(0); |
283 | case DNNL_ARG_WEIGHTS: return weights_md(0); |
284 | case DNNL_ARG_BIAS: return weights_md(1); |
285 | case DNNL_ARG_DST: return dst_md(0); |
286 | default: return convolution_pd_t::arg_md(arg); |
287 | } |
288 | } |
289 | |
290 | const memory_desc_t *src_md(int index = 0) const override { |
291 | return index == 0 ? &src_md_ : &glob_zero_md; |
292 | } |
293 | const memory_desc_t *dst_md(int index = 0) const override { |
294 | return index == 0 ? &dst_md_ : &glob_zero_md; |
295 | } |
296 | const memory_desc_t *weights_md(int index = 0) const override { |
297 | if (index == 0) return &weights_md_; |
298 | if (index == 1 && with_bias()) return &bias_md_; |
299 | return &glob_zero_md; |
300 | } |
301 | |
302 | int n_inputs() const override { |
303 | return 2 + with_bias() + attr_post_op_dw_inputs() + n_binary_po_inputs() |
304 | + n_prelu_po_inputs(); |
305 | } |
306 | |
307 | int n_outputs() const override { return 1; } |
308 | |
309 | protected: |
310 | memory_desc_t src_md_; |
311 | memory_desc_t weights_md_; |
312 | memory_desc_t bias_md_; |
313 | memory_desc_t dst_md_; |
314 | |
315 | convolution_fwd_pd_t(const convolution_desc_t *adesc, |
316 | const primitive_attr_t *attr, |
317 | const convolution_fwd_pd_t *hint_fwd_pd) |
318 | : convolution_pd_t(adesc, attr, hint_fwd_pd) |
319 | , src_md_(desc_.src_desc) |
320 | , weights_md_(desc_.weights_desc) |
321 | , bias_md_(desc_.bias_desc) |
322 | , dst_md_(desc_.dst_desc) {} |
323 | |
324 | bool set_default_formats_common( |
325 | format_tag_t src_tag, format_tag_t wei_tag, format_tag_t dst_tag) { |
326 | return set_default_formats_common_template(src_md_, src_tag, |
327 | weights_md_, wei_tag, dst_md_, dst_tag, bias_md_); |
328 | } |
329 | |
330 | int attr_post_op_dw_inputs() const { |
331 | const auto &po = attr_.post_ops_; |
332 | int conv = po.find(primitive_kind::convolution); |
333 | if (conv == -1) return 0; |
334 | return po.entry_[conv].depthwise_conv.bias_dt == data_type::undef ? 1 |
335 | : 2; |
336 | } |
337 | }; |
338 | |
339 | struct convolution_bwd_data_pd_t : public convolution_pd_t { |
340 | typedef convolution_bwd_data_pd_t base_class; |
341 | typedef convolution_fwd_pd_t hint_class; |
342 | |
343 | arg_usage_t arg_usage(int arg) const override { |
344 | if (utils::one_of(arg, DNNL_ARG_WEIGHTS, DNNL_ARG_DIFF_DST)) |
345 | return arg_usage_t::input; |
346 | |
347 | if (arg == DNNL_ARG_DIFF_SRC) return arg_usage_t::output; |
348 | |
349 | return primitive_desc_t::arg_usage(arg); |
350 | } |
351 | |
352 | const memory_desc_t *arg_md(int arg) const override { |
353 | switch (arg) { |
354 | case DNNL_ARG_DIFF_SRC: return diff_src_md(0); |
355 | case DNNL_ARG_WEIGHTS: return weights_md(0); |
356 | case DNNL_ARG_BIAS: return weights_md(1); |
357 | case DNNL_ARG_DIFF_DST: return diff_dst_md(0); |
358 | default: return convolution_pd_t::arg_md(arg); |
359 | } |
360 | } |
361 | |
362 | const memory_desc_t *diff_src_md(int index = 0) const override { |
363 | return index == 0 ? &diff_src_md_ : &glob_zero_md; |
364 | } |
365 | const memory_desc_t *diff_dst_md(int index = 0) const override { |
366 | return index == 0 ? &diff_dst_md_ : &glob_zero_md; |
367 | } |
368 | const memory_desc_t *weights_md(int index = 0) const override { |
369 | if (index == 0) return &weights_md_; |
370 | if (index == 1 && with_bias()) return &bias_md_; |
371 | return &glob_zero_md; |
372 | } |
373 | |
374 | int n_inputs() const override { return 2 + with_bias(); } |
375 | int n_outputs() const override { return 1; } |
376 | |
377 | virtual bool support_bias() const { return false; } |
378 | |
379 | protected: |
380 | memory_desc_t diff_src_md_; |
381 | memory_desc_t weights_md_; |
382 | memory_desc_t bias_md_; |
383 | memory_desc_t diff_dst_md_; |
384 | |
385 | convolution_bwd_data_pd_t(const convolution_desc_t *adesc, |
386 | const primitive_attr_t *attr, |
387 | const convolution_fwd_pd_t *hint_fwd_pd) |
388 | : convolution_pd_t(adesc, attr, hint_fwd_pd) |
389 | , diff_src_md_(desc_.diff_src_desc) |
390 | , weights_md_(desc_.weights_desc) |
391 | , bias_md_(desc_.bias_desc) |
392 | , diff_dst_md_(desc_.diff_dst_desc) {} |
393 | |
394 | bool set_default_formats_common(format_tag_t diff_src_tag, |
395 | format_tag_t wei_tag, format_tag_t diff_dst_tag) { |
396 | return set_default_formats_common_template(diff_src_md_, diff_src_tag, |
397 | weights_md_, wei_tag, diff_dst_md_, diff_dst_tag, bias_md_); |
398 | } |
399 | }; |
400 | |
401 | struct convolution_bwd_weights_pd_t : public convolution_pd_t { |
402 | typedef convolution_bwd_weights_pd_t base_class; |
403 | typedef convolution_fwd_pd_t hint_class; |
404 | |
405 | convolution_bwd_weights_pd_t(const convolution_desc_t *adesc, |
406 | const primitive_attr_t *attr, |
407 | const convolution_fwd_pd_t *hint_fwd_pd) |
408 | : convolution_pd_t(adesc, attr, hint_fwd_pd) |
409 | , src_md_(desc_.src_desc) |
410 | , diff_weights_md_(desc_.diff_weights_desc) |
411 | , diff_bias_md_(desc_.diff_bias_desc) |
412 | , diff_dst_md_(desc_.diff_dst_desc) {} |
413 | |
414 | arg_usage_t arg_usage(int arg) const override { |
415 | if (utils::one_of(arg, DNNL_ARG_SRC, DNNL_ARG_DIFF_DST)) |
416 | return arg_usage_t::input; |
417 | |
418 | if (arg == DNNL_ARG_DIFF_WEIGHTS) return arg_usage_t::output; |
419 | |
420 | if (arg == DNNL_ARG_DIFF_BIAS && with_bias()) |
421 | return arg_usage_t::output; |
422 | |
423 | return primitive_desc_t::arg_usage(arg); |
424 | } |
425 | |
426 | const memory_desc_t *arg_md(int arg) const override { |
427 | switch (arg) { |
428 | case DNNL_ARG_SRC: return src_md(0); |
429 | case DNNL_ARG_DIFF_WEIGHTS: return diff_weights_md(0); |
430 | case DNNL_ARG_DIFF_BIAS: return diff_weights_md(1); |
431 | case DNNL_ARG_DIFF_DST: return diff_dst_md(0); |
432 | default: return convolution_pd_t::arg_md(arg); |
433 | } |
434 | } |
435 | |
436 | const memory_desc_t *src_md(int index = 0) const override { |
437 | return index == 0 ? &src_md_ : &glob_zero_md; |
438 | } |
439 | const memory_desc_t *diff_dst_md(int index = 0) const override { |
440 | return index == 0 ? &diff_dst_md_ : &glob_zero_md; |
441 | } |
442 | const memory_desc_t *diff_weights_md(int index = 0) const override { |
443 | if (index == 0) return &diff_weights_md_; |
444 | if (index == 1 && with_bias()) return &diff_bias_md_; |
445 | return &glob_zero_md; |
446 | } |
447 | |
448 | int n_inputs() const override { return 2; } |
449 | int n_outputs() const override { return 1 + with_bias(); } |
450 | |
451 | protected: |
452 | memory_desc_t src_md_; |
453 | memory_desc_t diff_weights_md_; |
454 | memory_desc_t diff_bias_md_; |
455 | memory_desc_t diff_dst_md_; |
456 | |
457 | bool set_default_formats_common(format_tag_t src_tag, |
458 | format_tag_t diff_wei_tag, format_tag_t diff_dst_tag) { |
459 | return set_default_formats_common_template(src_md_, src_tag, |
460 | diff_weights_md_, diff_wei_tag, diff_dst_md_, diff_dst_tag, |
461 | diff_bias_md_); |
462 | } |
463 | }; |
464 | |
465 | } // namespace impl |
466 | } // namespace dnnl |
467 | |
468 | #endif |
469 | |
470 | // vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s |
471 | |