1 //
2 // TensorUtils.cpp
3 // MNN
4 //
5 // Created by MNN on 2018/08/11.
6 // Copyright © 2018, Alibaba Group Holding Limited
7 //
8
9 #include "core/TensorUtils.hpp"
10 #include <float.h>
11 #include <math.h>
12 #include <stdio.h>
13 #include <cmath>
14 #include <cstring>
15 #include "core/Backend.hpp"
16 #include "core/Macro.h"
17
18 namespace MNN {
getDescribe(const Tensor * tensor)19 Tensor::InsideDescribe* TensorUtils::getDescribe(const Tensor* tensor) {
20 return tensor->mDescribe;
21 }
regionIsFull(Tensor * input)22 bool TensorUtils::regionIsFull(Tensor* input) {
23 auto des = TensorUtils::getDescribe(input);
24 if (des->memoryType != Tensor::InsideDescribe::MEMORY_VIRTUAL) {
25 return true;
26 }
27 int size = 1;
28 for (int i = 0; i < input->dimensions(); ++i) {
29 size *= input->length(i);
30 }
31 int regionSize = 0;
32 for (auto& region : des->regions) {
33 regionSize += region.size[1] * region.size[0] * region.size[2];
34 }
35 return regionSize == size;
36 }
37
makeFullSlice(Tensor * input)38 Tensor::InsideDescribe::Region TensorUtils::makeFullSlice(Tensor* input) {
39 Tensor::InsideDescribe::Region totalSlice;
40 totalSlice.src.offset = 0;
41 totalSlice.dst.offset = 0;
42 totalSlice.origin = input;
43 for (int i = 0; i < input->dimensions(); ++i) {
44 totalSlice.size[2] *= input->length(i);
45 }
46 totalSlice.dst.stride[1] = totalSlice.size[2];
47 totalSlice.dst.stride[0] = totalSlice.size[2];
48 totalSlice.src.stride[1] = totalSlice.size[2];
49 totalSlice.src.stride[0] = totalSlice.size[2];
50 return totalSlice;
51 }
reshapeSlice(Tensor::InsideDescribe::Region & slice,int outside,int inside,int axis)52 bool TensorUtils::reshapeSlice(Tensor::InsideDescribe::Region& slice, int outside, int inside, int axis) {
53 if (slice.size[1] == 1 && slice.size[0] == 1 && slice.size[2] == outside * inside * axis) {
54 slice.size[0] = outside;
55 slice.size[2] = inside;
56 slice.size[1] = axis;
57 slice.dst.stride[0] = inside * axis;
58 slice.dst.stride[1] = inside;
59
60 auto originStride = slice.src.stride[2];
61 slice.src.stride[0] = originStride * inside * axis;
62 slice.src.stride[1] = originStride * inside;
63 return true;
64 }
65 if (slice.size[0] == outside && slice.size[1] == axis && slice.size[2] == inside) {
66 return true;
67 }
68 return false;
69 }
70
setupTensorInfo(const Tensor * tensor,Tensor * wrapTensor,MNN_DATA_FORMAT mMidFormat)71 void TensorUtils::setupTensorInfo(const Tensor* tensor, Tensor* wrapTensor, MNN_DATA_FORMAT mMidFormat) {
72 TensorUtils::getDescribe(wrapTensor)->dimensionFormat = mMidFormat;
73 auto tensorFormat = TensorUtils::getDescribe(tensor)->dimensionFormat;
74 bool originCaffeFormat = (tensorFormat == MNN_DATA_FORMAT_NCHW || tensorFormat == MNN_DATA_FORMAT_NC4HW4);
75 bool wrapCaffeFormat = (mMidFormat == MNN_DATA_FORMAT_NCHW || mMidFormat == MNN_DATA_FORMAT_NC4HW4);
76 bool originTfFormat = (tensorFormat == MNN_DATA_FORMAT_NHWC || tensorFormat == MNN_DATA_FORMAT_NHWC4);
77 bool wrapTfFormat = (mMidFormat == MNN_DATA_FORMAT_NHWC || mMidFormat == MNN_DATA_FORMAT_NHWC4);
78 if ((originCaffeFormat && wrapCaffeFormat) || (originTfFormat && wrapTfFormat)) {
79 TensorUtils::copyShape(tensor, wrapTensor);
80 } else if (originCaffeFormat && wrapTfFormat) {
81 for (int i = 1; i < wrapTensor->dimensions() - 1; ++i) {
82 wrapTensor->setLength(i, tensor->length(i + 1));
83 }
84 wrapTensor->setLength(0, tensor->length(0));
85 wrapTensor->setLength(wrapTensor->dimensions() - 1, tensor->length(1));
86 } else if (originTfFormat && wrapCaffeFormat) {
87 for (int i = 2; i < wrapTensor->dimensions(); ++i) {
88 wrapTensor->setLength(i, tensor->length(i - 1));
89 }
90 wrapTensor->setLength(0, tensor->length(0));
91 wrapTensor->setLength(1, tensor->length(tensor->dimensions() - 1));
92 } else {
93 // will not reach here
94 MNN_ASSERT(false);
95 }
96 TensorUtils::setLinearLayout(wrapTensor);
97 wrapTensor->buffer().type = tensor->getType();
98 }
99
copyShape(const Tensor * source,Tensor * dest,bool copyFormat)100 void TensorUtils::copyShape(const Tensor* source, Tensor* dest, bool copyFormat) {
101 auto& ob = dest->buffer();
102 auto& ib = source->buffer();
103 ob.dimensions = ib.dimensions;
104 ::memcpy(ob.dim, ib.dim, ib.dimensions * sizeof(halide_dimension_t));
105 if (copyFormat) {
106 getDescribe(dest)->dimensionFormat = getDescribe(source)->dimensionFormat;
107 }
108 }
109
setShape(Tensor * dest,const std::vector<int> & alldims)110 void TensorUtils::setShape(Tensor* dest, const std::vector<int>& alldims) {
111 auto& ob = dest->buffer();
112 ob.dimensions = alldims.size();
113 int stride = 1;
114 for (int i = alldims.size() - 1; i >= 0; --i) {
115 ob.dim[i].stride = stride;
116 ob.dim[i].extent = alldims[i];
117 stride *= alldims[i];
118 }
119 return;
120 }
121
setLinearLayout(Tensor * tensor)122 void TensorUtils::setLinearLayout(Tensor* tensor) {
123 auto& buffer = tensor->buffer();
124 int size = 1;
125 for (int i = 0; i < buffer.dimensions; ++i) {
126 auto index = buffer.dimensions - i - 1;
127 auto extent = buffer.dim[index].extent;
128 if (1 == index && tensor->mDescribe->dimensionFormat == MNN_DATA_FORMAT_NC4HW4) {
129 extent = ROUND_UP(extent, 4);
130 }
131 buffer.dim[index].stride = size;
132 size *= extent;
133 }
134 }
135
clearHandleData(Tensor * tensor)136 void TensorUtils::clearHandleData(Tensor* tensor) {
137 if (tensor->buffer().type.code != halide_type_handle) {
138 return;
139 }
140 auto handle = tensor->host<void*>();
141 if (nullptr == handle) {
142 return;
143 }
144
145 MNN_ASSERT(tensor->mDescribe->extra.handleFreeFunction != nullptr);
146 for (int i = 0; i < tensor->elementSize(); ++i) {
147 if (nullptr != handle[i]) {
148 tensor->mDescribe->extra.handleFreeFunction(handle[i]);
149 handle[i] = nullptr;
150 }
151 }
152 }
153
createHostPlanar(const Tensor * source)154 static const Tensor* createHostPlanar(const Tensor* source) {
155 // check
156 auto bnType = MNN_FORWARD_CPU;
157 auto tensorBackend = TensorUtils::getDescribe(source)->backend;
158 if (tensorBackend) {
159 bnType = tensorBackend->type();
160 }
161 bool device = bnType != MNN_FORWARD_CPU;
162 bool chunky = TensorUtils::getDescribe(source)->dimensionFormat == MNN_DATA_FORMAT_NC4HW4;
163
164 // no convert needed
165 if (!device && !chunky) {
166 return source;
167 }
168
169 // convert
170 if (chunky) {
171 Tensor* result = source->createHostTensorFromDevice(source, false);
172 if (result->getDimensionType() == MNN::Tensor::TENSORFLOW) {
173 TensorUtils::getDescribe(result)->dimensionFormat = MNN_DATA_FORMAT_NHWC;
174 } else {
175 TensorUtils::getDescribe(result)->dimensionFormat = MNN_DATA_FORMAT_NCHW;
176 }
177 TensorUtils::setLinearLayout(result);
178
179 if (device) {
180 source->copyToHostTensor(result);
181 } else {
182 Backend::Info info;
183 info.type = MNN_FORWARD_CPU;
184 std::shared_ptr<Runtime> runtime(MNNGetExtraRuntimeCreator(MNN_FORWARD_CPU)->onCreate(info));
185 auto backend = runtime->onCreate();
186 backend->onCopyBuffer(source, result);
187 delete backend;
188 }
189 return result;
190 } else {
191 return source->createHostTensorFromDevice(source, true);
192 }
193 }
194
195 template <typename T>
copyTensorToFloat(const Tensor * source,double * dest)196 static void copyTensorToFloat(const Tensor* source, double* dest) {
197 auto srcData = source->host<T>();
198 auto size = source->elementSize();
199 for (int i = 0; i < size; ++i) {
200 dest[i] = srcData[i];
201 }
202 }
203
equals(const double * pa,const double * pb,size_t size,double tolerance,double epsilon,bool overall,bool prints)204 static bool equals(const double* pa, const double* pb, size_t size, double tolerance, double epsilon, bool overall,
205 bool prints) {
206 // get max if using overall torelance
207 double max = fabs(pb[0]);
208 if (overall) {
209 for (int i = 1; i < size; i++) {
210 max = std::max(max, fabs(pb[i]));
211 }
212 }
213
214 // compare
215 for (int i = 0; i < size; i++) {
216 float va = pa[i], vb = pb[i];
217 if (std::isinf(va) && std::isinf(vb)) {
218 continue;
219 }
220 if (fabs(va) < epsilon && fabs(vb) < epsilon) {
221 continue;
222 }
223 float div = overall ? max : fabsf(vb);
224 if (fabsf(va - vb) / div > tolerance) {
225 if (prints) {
226 MNN_PRINT("%d: %f != %f\n", i, va, vb);
227 }
228 return false;
229 }
230 }
231 return true;
232 }
233
compareTensors(const Tensor * compare,const Tensor * expect,float tolerance,bool overall,bool printsErrors,bool printsTensors)234 bool TensorUtils::compareTensors(const Tensor* compare, const Tensor* expect, float tolerance, bool overall,
235 bool printsErrors, bool printsTensors) {
236 // type
237 if (compare->getType().code != expect->getType().code || compare->getType().bits != expect->getType().bits) {
238 if (printsErrors) {
239 MNN_PRINT("NOT equal in type: %d/%d - %d/%d.\n", compare->getType().code, compare->getType().bits,
240 expect->getType().code, expect->getType().bits);
241 }
242 return false;
243 }
244
245 // dimensions
246 if (compare->dimensions() != expect->dimensions()) {
247 if (printsErrors) {
248 MNN_PRINT("NOT equal in dimensions: %d - %d.\n", compare->dimensions(), expect->dimensions());
249 }
250 return false;
251 }
252 for (int i = 0; i < compare->dimensions(); i++) {
253 if (compare->length(i) == expect->length(i)) {
254 continue;
255 }
256 if (printsErrors) {
257 MNN_PRINT("NOT equal in dimensions[%d]: %d - %d.\n", i, compare->length(i), expect->length(i));
258 }
259 return false;
260 }
261
262 // convert to host if needed
263 auto a = createHostPlanar(compare), b = createHostPlanar(expect);
264
265 // get value as double
266 auto size = expect->elementSize();
267 std::vector<double> expectValue(expect->elementSize(), 0.0f);
268 std::vector<double> compareValue(compare->elementSize(), 0.0f);
269
270 auto result = false;
271 if (b->buffer().type.code == halide_type_uint) {
272 switch (b->buffer().type.bits) {
273 case 8:
274 copyTensorToFloat<uint8_t>(a, compareValue.data());
275 copyTensorToFloat<uint8_t>(b, expectValue.data());
276 break;
277 case 16:
278 copyTensorToFloat<uint16_t>(a, compareValue.data());
279 copyTensorToFloat<uint16_t>(b, expectValue.data());
280 break;
281 case 32:
282 copyTensorToFloat<uint32_t>(a, compareValue.data());
283 copyTensorToFloat<uint32_t>(b, expectValue.data());
284 break;
285 case 64:
286 copyTensorToFloat<uint64_t>(a, compareValue.data());
287 copyTensorToFloat<uint64_t>(b, expectValue.data());
288 break;
289 default:
290 break;
291 }
292 } else if (b->buffer().type.code == halide_type_int) {
293 switch (b->buffer().type.bits) {
294 case 8:
295 copyTensorToFloat<int8_t>(a, compareValue.data());
296 copyTensorToFloat<int8_t>(b, expectValue.data());
297 break;
298 case 16:
299 copyTensorToFloat<int16_t>(a, compareValue.data());
300 copyTensorToFloat<int16_t>(b, expectValue.data());
301 break;
302 case 32:
303 copyTensorToFloat<int32_t>(a, compareValue.data());
304 copyTensorToFloat<int32_t>(b, expectValue.data());
305 break;
306 case 64:
307 copyTensorToFloat<int64_t>(a, compareValue.data());
308 copyTensorToFloat<int64_t>(b, expectValue.data());
309 break;
310 default:
311 break;
312 }
313 } else if (b->buffer().type.code == halide_type_float) {
314 switch (b->buffer().type.bits) {
315 case 32:
316 copyTensorToFloat<float>(a, compareValue.data());
317 copyTensorToFloat<float>(b, expectValue.data());
318 break;
319 default:
320 break;
321 }
322 } else {
323 if (printsErrors) {
324 MNN_PRINT("unsupported data type.");
325 }
326 }
327 auto epsilon = FLT_EPSILON;
328 if ((NULL != compareValue.data()) && (NULL != expectValue.data())) {
329 result = equals(compareValue.data(), expectValue.data(), size, tolerance, epsilon, overall, printsErrors);
330 }
331
332 // clean up
333 if (a != compare) {
334 delete a;
335 }
336 if (b != expect) {
337 delete b;
338 }
339 return result;
340 }
341
342 // is copy only region
isCopyRegion(const Tensor::InsideDescribe::Region & region)343 bool TensorUtils::isCopyRegion(const Tensor::InsideDescribe::Region& region) {
344 bool eq = true;
345 for (int i = 0; i < 3; i++) {
346 eq &= ((region.src.stride[i] == region.dst.stride[i]) || (region.size[i] <= 1));
347 }
348 return eq;
349 }
350
351 // compute offset through region
offsetCompute(Tensor::InsideDescribe::Region reg,int offset,bool backward)352 static inline int offsetCompute(Tensor::InsideDescribe::Region reg, int offset, bool backward) {
353 if (backward) {
354 auto tmp = reg.src;
355 reg.src = reg.dst;
356 reg.dst = tmp;
357 }
358 int res = 0;
359 for (int i = 0; i < 3; i++) {
360 if (reg.size[i] > 1) {
361 res += offset / reg.src.stride[i] * reg.dst.stride[i];
362 offset %= reg.src.stride[i];
363 }
364 }
365 return res;
366 }
367
368 // expand src stride with expand value
expandSrc(std::vector<int> & src,std::vector<int> & dst,std::vector<int> & size,int expandValue)369 static inline bool expandSrc(std::vector<int>& src, std::vector<int>& dst, std::vector<int>& size, int expandValue) {
370 if (expandValue <= 0) {
371 return false;
372 }
373 for (int i = size.size()-1; i >= 0; i--) {
374 int splitSize = expandValue / src[i];
375 if (!(expandValue % src[i] || size[i] % splitSize)) {
376 src.insert(src.begin()+i, expandValue);
377 dst.insert(dst.begin()+i, splitSize * dst[i]);
378 size[i] /= splitSize;
379 size.insert(size.begin()+i+1, splitSize);
380 return true;
381 }
382 }
383 return false;
384 }
385 // expand stride and size with expand value
expandStrideSize(int * src,int * dst,int * size,int & num,int expandValue)386 static inline bool expandStrideSize(int* src, int* dst, int* size, int& num, int expandValue) {
387 #define MNN_3_INT_INSERT(x, i, y) if (i == 2) { x[2] = y; } else if (i == 1) { x[2] = x[1]; x[1] = y; } else if (i == 0) { x[2] = x[1]; x[1] = x[0]; x[0] = y; } else { return false; }
388 for (int i = num-1; i >= 0; i--) {
389 int splitSize = expandValue / src[i];
390 if (!(expandValue % src[i] || size[i] % splitSize)) {
391 MNN_3_INT_INSERT(src, i, expandValue)
392 MNN_3_INT_INSERT(dst, i, (splitSize * dst[i]))
393 size[i] /= splitSize;
394 MNN_3_INT_INSERT(size, (i+1), splitSize)
395 if (++num > 3) return false;
396 return true;
397 }
398 }
399 return false;
400 #undef MNN_3_INT_INSERT
401 }
402
403 // fuse srcRegion and dstRegion to dstRegion if return true
fuseRegion(Tensor::InsideDescribe::Region & srcReg,Tensor::InsideDescribe::Region & dstReg)404 bool TensorUtils::fuseRegion(Tensor::InsideDescribe::Region& srcReg, Tensor::InsideDescribe::Region& dstReg) {
405 // src data isnot full data of dst
406 if (srcReg.dst.offset > dstReg.src.offset ||
407 srcReg.dst.stride[1] > srcReg.size[2] ||
408 srcReg.dst.stride[2] > srcReg.size[1] * srcReg.size[2]) {
409 return false;
410 }
411 int dstTotalSize = 1, srcTotalSize = 1;
412 for (int i = 0; i < 3; i++) {
413 if (dstReg.size[i] > 1) {
414 dstTotalSize *= dstReg.size[i];
415 }
416 if (srcReg.size[i] > 1) {
417 srcTotalSize *= srcReg.size[i];
418 }
419 }
420 // src data is not full data of dst
421 if (dstTotalSize > srcTotalSize) {
422 return false;
423 }
424 // dont deal size > 1 && stride <= 0
425 for (int i = 0; i < 3; i++) {
426 if (srcReg.size[i] > 1 && (srcReg.src.stride[i] <= 0 || srcReg.dst.stride[i] <= 0)) {
427 return false;
428 }
429 if (dstReg.size[i] > 1 && (dstReg.src.stride[i] <= 0 || dstReg.dst.stride[i] <= 0)) {
430 return false;
431 }
432 }
433 // src copy fuse
434 if (isCopyRegion(srcReg)) {
435 dstReg.origin = srcReg.origin;
436 dstReg.src.offset += srcReg.src.offset - srcReg.dst.offset;
437 return true;
438 }
439 // dst copy fuse
440 if (isCopyRegion(dstReg) && dstTotalSize == srcTotalSize) {
441 int srcOff = dstReg.src.offset - srcReg.dst.offset;
442 int dstOff = dstReg.dst.offset;
443 srcOff = offsetCompute(srcReg, srcOff, true) + srcReg.src.offset;
444 if (srcReg.src.stride[2] > 0 && srcOff % srcReg.src.stride[2] != 0) {
445 // when transpose + slice, offset is not align can't fuse
446 return false;
447 }
448 dstReg.origin = srcReg.origin;
449 dstReg.dst = srcReg.dst;
450 dstReg.src = srcReg.src;
451 dstReg.src.offset = srcOff;
452 dstReg.dst.offset = dstOff;
453 dstReg.size[0] = srcReg.size[0];
454 dstReg.size[1] = srcReg.size[1];
455 dstReg.size[2] = srcReg.size[2];
456 return true;
457 }
458 #define MNN_FAST_FUSE_WITHOUT_STL
459 #ifdef MNN_FAST_FUSE_WITHOUT_STL
460 // general fuse
461 int srcDst[3], srcSrc[3], dstSrc[3], dstDst[3], srcSize[3], dstSize[3], newSrc[3], dstStride[3], srcStride[3];
462 #define MNN_3_INT_INIT(x, y) { x[0] = y; x[1] = y; x[2] = y; }
463 MNN_3_INT_INIT(dstStride, -1)
464 MNN_3_INT_INIT(srcStride, -1)
465 #undef MNN_3_INT_INIT
466 int srcNum = 0, dstNum = 0, sizeNum = 0;
467 for (int i = 0; i < 3; i++) {
468 if (srcReg.size[i] > 1) {
469 srcStride[srcNum] = srcReg.dst.stride[i];
470 srcDst[srcNum] = srcReg.dst.stride[i];
471 srcSrc[srcNum] = srcReg.src.stride[i];
472 srcSize[srcNum] = srcReg.size[i];
473 srcNum++;
474 }
475 if (dstReg.size[i] > 1) {
476 dstStride[dstNum] = dstReg.src.stride[i];
477 dstDst[dstNum] = dstReg.dst.stride[i];
478 dstSrc[dstNum] = dstReg.src.stride[i];
479 dstSize[dstNum] = dstReg.size[i];
480 dstNum++;
481 }
482 }
483 sizeNum = dstNum;
484 #define MNN_3_INT_DIFF(r, x, y, i) if ((x[i] != y[0]) && (x[i] != y[1]) && (x[i] != y[2])) { if (r > 0) { return false; } else { r = x[i]; } }
485 int srcExtra = -1, dstExtra = -1;
486 MNN_3_INT_DIFF(srcExtra, srcStride, dstStride, 0)
487 MNN_3_INT_DIFF(srcExtra, srcStride, dstStride, 1)
488 MNN_3_INT_DIFF(srcExtra, srcStride, dstStride, 2)
489 MNN_3_INT_DIFF(dstExtra, dstStride, srcStride, 0)
490 MNN_3_INT_DIFF(dstExtra, dstStride, srcStride, 1)
491 MNN_3_INT_DIFF(dstExtra, dstStride, srcStride, 2)
492 #undef MNN_3_INT_DIFF
493 if (dstExtra > 0) {
494 if (!expandStrideSize(srcDst, srcSrc, srcSize, srcNum, dstExtra)) {
495 return false;
496 }
497 }
498 if (srcExtra > 0) {
499 if (!expandStrideSize(dstSrc, dstDst, dstSize, dstNum, srcExtra)) {
500 return false;
501 }
502 }
503 // reorder srcSrc to newSrc by align srcDst and dstSrc
504 for (int i = 0; i < dstNum; i++) {
505 int index = 0;
506 for (int j = 0; j < srcNum; j++) {
507 if (dstSrc[j] == srcDst[i]) {
508 index = j;
509 }
510 }
511 newSrc[index] = srcSrc[i];
512 }
513 // set final size and set expandIdx if expand val is 1
514 int expandIdx = -1;
515 if (dstNum > sizeNum) {
516 for (int i = 2; i >= 0; i--) {
517 if (i < dstNum) {
518 if (dstSize[i] == 1) {
519 expandIdx = i;
520 }
521 dstReg.size[i] = dstSize[i];
522 } else {
523 dstReg.size[i] = 1;
524 }
525 }
526 }
527 #else
528 // general fuse
529 std::set<int> dstStride, srcStride, dstDiff, srcDiff;
530 std::vector<int> dstDst, dstSrc, srcDst, srcSrc, newSrc, dstSize, srcSize;
531 for (int i = 0; i < 3; i++) {
532 if (srcReg.size[i] > 1) {
533 srcStride.insert(srcReg.dst.stride[i]);
534 srcDst.push_back(srcReg.dst.stride[i]);
535 srcSrc.push_back(srcReg.src.stride[i]);
536 srcSize.push_back(srcReg.size[i]);
537 }
538 if (dstReg.size[i] > 1) {
539 dstStride.insert(dstReg.src.stride[i]);
540 dstDst.push_back(dstReg.dst.stride[i]);
541 dstSrc.push_back(dstReg.src.stride[i]);
542 dstSize.push_back(dstReg.size[i]);
543 }
544 }
545 int sizeNum = dstSize.size();
546 std::set_difference(dstStride.begin(), dstStride.end(), srcStride.begin(), srcStride.end(), std::inserter(dstDiff, dstDiff.begin()));
547 std::set_difference(srcStride.begin(), srcStride.end(), dstStride.begin(), dstStride.end(), std::inserter(srcDiff, srcDiff.begin()));
548 if (dstDiff.size() > 1 || srcDiff.size() > 1) {
549 // many diff stride, now dont deal
550 return false;
551 }
552 // expand stride when middle tensor's stride diff
553 if (!dstDiff.empty()) {
554 if (!expandSrc(srcDst, srcSrc, srcSize, *dstDiff.begin())) {
555 return false;
556 }
557 }
558 if (!srcDiff.empty()) {
559 if (!expandSrc(dstSrc, dstDst, dstSize, *srcDiff.begin())) {
560 return false;
561 }
562 }
563 if (dstSize.size() > 3) {
564 // need splite region, dont deal
565 return false;
566 }
567 // reorder srcSrc to newSrc by align srcDst and dstSrc
568 newSrc.resize(srcSrc.size());
569 for (int i = 0; i < dstSrc.size(); i++) {
570 int index = std::distance(dstSrc.begin(), std::find(dstSrc.begin(), dstSrc.end(), srcDst[i]));
571 newSrc[index] = srcSrc[i];
572 }
573 // set final size and set expandIdx if expand val is 1
574 int expandIdx = -1;
575 if (dstSize.size() > sizeNum) {
576 for (int i = 2; i >= 0; i--) {
577 if (i < dstSize.size()) {
578 if (dstSize[i] == 1) {
579 expandIdx = i;
580 }
581 dstReg.size[i] = dstSize[i];
582 } else {
583 dstReg.size[i] = 1;
584 }
585 }
586 }
587 #endif
588 int idx = 0;
589 for (int i = 0; i < 3; i++) {
590 if (dstReg.size[i] > 1 || i == expandIdx) {
591 dstReg.src.stride[i] = newSrc[idx];
592 dstReg.dst.stride[i] = dstDst[idx++];
593 }
594 }
595 dstReg.origin = srcReg.origin;
596 dstReg.src.offset = offsetCompute(srcReg, dstReg.src.offset - srcReg.dst.offset, true) + srcReg.src.offset;
597 return true;
598 }
adjustTensorForCompability(Tensor * newTensor)599 void TensorUtils::adjustTensorForCompability(Tensor* newTensor) {
600 if (newTensor->dimensions() < 4) {
601 for (int n = newTensor->dimensions(); n < 4; ++n) {
602 newTensor->setLength(n, 1);
603 }
604 }
605 }
606
getDimType(const Tensor * t)607 Tensor::DimensionType TensorUtils::getDimType(const Tensor* t) {
608 auto format = TensorUtils::getDescribe(t)->dimensionFormat;
609 switch (format) {
610 case MNN_DATA_FORMAT_NCHW:
611 return Tensor::CAFFE;
612 case MNN_DATA_FORMAT_NC4HW4:
613 return Tensor::CAFFE_C4;
614 case MNN_DATA_FORMAT_NHWC:
615 return Tensor::TENSORFLOW;
616 default:
617 break;
618 }
619 return Tensor::TENSORFLOW;
620 }
621
DataTypeToHalideType(DataType t)622 halide_type_t TensorUtils::DataTypeToHalideType(DataType t) {
623 switch (t) {
624 case DataType_DT_DOUBLE:
625 case DataType_DT_FLOAT:
626 return halide_type_of<float>();
627 case DataType_DT_BFLOAT16:
628 return halide_type_t(halide_type_float, 16);
629 case DataType_DT_QINT32:
630 case DataType_DT_INT32:
631 case DataType_DT_BOOL:
632 case DataType_DT_INT64:
633 return halide_type_of<int32_t>();
634 case DataType_DT_QINT8:
635 case DataType_DT_INT8:
636 return halide_type_of<int8_t>();
637 case DataType_DT_QUINT8:
638 case DataType_DT_UINT8:
639 return halide_type_of<uint8_t>();
640 case DataType_DT_QUINT16:
641 case DataType_DT_UINT16:
642 return halide_type_of<uint16_t>();
643 case DataType_DT_QINT16:
644 case DataType_DT_INT16:
645 return halide_type_of<int16_t>();
646 case DataType_DT_STRING:
647 default:
648 MNN_PRINT("Unsupported data type!");
649 MNN_ASSERT(false);
650 return halide_type_of<float>();
651 }
652 }
653
HaildeTypeToDataType(halide_type_t t)654 DataType TensorUtils::HaildeTypeToDataType(halide_type_t t) {
655 if (t == halide_type_of<int8_t>()) {
656 return DataType_DT_INT8;
657 }
658 if (t == halide_type_of<int16_t>()) {
659 return DataType_DT_INT16;
660 }
661 if (t == halide_type_of<int32_t>()) {
662 return DataType_DT_INT32;
663 }
664 if (t == halide_type_of<int64_t>()) {
665 return DataType_DT_INT64;
666 }
667 if (t == halide_type_of<uint8_t>()) {
668 return DataType_DT_UINT8;
669 }
670 if (t == halide_type_of<uint16_t>()) {
671 return DataType_DT_UINT16;
672 }
673 if (t == halide_type_t(halide_type_float, 16)) {
674 return DataType_DT_BFLOAT16;
675 }
676 if (t == halide_type_of<float>()) {
677 return DataType_DT_FLOAT;
678 }
679 if (t == halide_type_of<double>()) {
680 return DataType_DT_DOUBLE;
681 }
682 MNN_PRINT("Unsupported data type!");
683 MNN_ASSERT(false);
684 return DataType_DT_INVALID;
685 }
getQuantInfo(const Tensor * t)686 std::vector<float> TensorUtils::getQuantInfo(const Tensor* t) {
687 float scale = getDescribe(t)->quantAttr ? getDescribe(t)->quantAttr->scale : 0.0f;
688 float zero = getDescribe(t)->quantAttr ? getDescribe(t)->quantAttr->zero : 0.0f;
689 float min = getDescribe(t)->quantAttr ? getDescribe(t)->quantAttr->min : -127.0f;
690 float max = getDescribe(t)->quantAttr ? getDescribe(t)->quantAttr->max : 127.0f;
691 return {scale, zero, min, max};
692 }
693 } // namespace MNN
694