1 | /* |
---|
2 | * CLRadeonExtender - Unofficial OpenCL Radeon Extensions Library |
---|
3 | * Copyright (C) 2014-2018 Mateusz Szpakowski |
---|
4 | * |
---|
5 | * This library is free software; you can redistribute it and/or |
---|
6 | * modify it under the terms of the GNU Lesser General Public |
---|
7 | * License as published by the Free Software Foundation; either |
---|
8 | * version 2.1 of the License, or (at your option) any later version. |
---|
9 | * |
---|
10 | * This library is distributed in the hope that it will be useful, |
---|
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
---|
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
---|
13 | * Lesser General Public License for more details. |
---|
14 | * |
---|
15 | * You should have received a copy of the GNU Lesser General Public |
---|
16 | * License along with this library; if not, write to the Free Software |
---|
17 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
---|
18 | */ |
---|
19 | |
---|
20 | #include <CLRX/Config.h> |
---|
21 | #include <cassert> |
---|
22 | #include <cstdio> |
---|
23 | #include <cstring> |
---|
24 | #include <cstdint> |
---|
25 | #include <string> |
---|
26 | #include <vector> |
---|
27 | #include <algorithm> |
---|
28 | #include <utility> |
---|
29 | #include <memory> |
---|
30 | #include <unordered_set> |
---|
31 | #include <CLRX/amdbin/ElfBinaries.h> |
---|
32 | #include <CLRX/utils/Utilities.h> |
---|
33 | #include <CLRX/utils/MemAccess.h> |
---|
34 | #include <CLRX/utils/InputOutput.h> |
---|
35 | #include <CLRX/utils/Containers.h> |
---|
36 | #include <CLRX/amdbin/ROCmBinaries.h> |
---|
37 | |
---|
38 | using namespace CLRX; |
---|
39 | |
---|
40 | /* |
---|
41 | * ROCm metadata YAML parser |
---|
42 | */ |
---|
43 | |
---|
44 | void ROCmKernelMetadata::initialize() |
---|
45 | { |
---|
46 | langVersion[0] = langVersion[1] = BINGEN_NOTSUPPLIED; |
---|
47 | reqdWorkGroupSize[0] = reqdWorkGroupSize[1] = reqdWorkGroupSize[2] = 0; |
---|
48 | workGroupSizeHint[0] = workGroupSizeHint[1] = workGroupSizeHint[2] = 0; |
---|
49 | kernargSegmentSize = BINGEN64_NOTSUPPLIED; |
---|
50 | groupSegmentFixedSize = BINGEN64_NOTSUPPLIED; |
---|
51 | privateSegmentFixedSize = BINGEN64_NOTSUPPLIED; |
---|
52 | kernargSegmentAlign = BINGEN64_NOTSUPPLIED; |
---|
53 | wavefrontSize = BINGEN_NOTSUPPLIED; |
---|
54 | sgprsNum = BINGEN_NOTSUPPLIED; |
---|
55 | vgprsNum = BINGEN_NOTSUPPLIED; |
---|
56 | maxFlatWorkGroupSize = BINGEN64_NOTSUPPLIED; |
---|
57 | fixedWorkGroupSize[0] = fixedWorkGroupSize[1] = fixedWorkGroupSize[2] = 0; |
---|
58 | spilledSgprs = BINGEN_NOTSUPPLIED; |
---|
59 | spilledVgprs = BINGEN_NOTSUPPLIED; |
---|
60 | } |
---|
61 | |
---|
62 | void ROCmMetadata::initialize() |
---|
63 | { |
---|
64 | version[0] = 1; |
---|
65 | version[1] = 0; |
---|
66 | } |
---|
67 | |
---|
68 | // return trailing spaces |
---|
69 | static size_t skipSpacesAndComments(const char*& ptr, const char* end, size_t& lineNo) |
---|
70 | { |
---|
71 | const char* lineStart = ptr; |
---|
72 | while (ptr != end) |
---|
73 | { |
---|
74 | lineStart = ptr; |
---|
75 | while (ptr != end && *ptr!='\n' && isSpace(*ptr)) ptr++; |
---|
76 | if (ptr == end) |
---|
77 | break; // end of stream |
---|
78 | if (*ptr=='#') |
---|
79 | { |
---|
80 | // skip comment |
---|
81 | while (ptr != end && *ptr!='\n') ptr++; |
---|
82 | if (ptr == end) |
---|
83 | return 0; // no trailing spaces and end |
---|
84 | } |
---|
85 | else if (*ptr!='\n') |
---|
86 | break; // no comment and no end of line |
---|
87 | else |
---|
88 | { |
---|
89 | ptr++; |
---|
90 | lineNo++; // next line |
---|
91 | } |
---|
92 | } |
---|
93 | return ptr - lineStart; |
---|
94 | } |
---|
95 | |
---|
96 | static inline void skipSpacesToLineEnd(const char*& ptr, const char* end) |
---|
97 | { |
---|
98 | while (ptr != end && *ptr!='\n' && isSpace(*ptr)) ptr++; |
---|
99 | } |
---|
100 | |
---|
101 | static void skipSpacesToNextLine(const char*& ptr, const char* end, size_t& lineNo) |
---|
102 | { |
---|
103 | skipSpacesToLineEnd(ptr, end); |
---|
104 | if (ptr != end && *ptr != '\n' && *ptr!='#') |
---|
105 | throw ParseException(lineNo, "Garbages at line"); |
---|
106 | if (ptr != end && *ptr == '#') |
---|
107 | // skip comment at end of line |
---|
108 | while (ptr!=end && *ptr!='\n') ptr++; |
---|
109 | if (ptr!=end) |
---|
110 | { // newline |
---|
111 | ptr++; |
---|
112 | lineNo++; |
---|
113 | } |
---|
114 | } |
---|
115 | |
---|
116 | enum class YAMLValType |
---|
117 | { |
---|
118 | NONE, |
---|
119 | NIL, |
---|
120 | BOOL, |
---|
121 | INT, |
---|
122 | FLOAT, |
---|
123 | STRING, |
---|
124 | SEQ |
---|
125 | }; |
---|
126 | |
---|
127 | static YAMLValType parseYAMLType(const char*& ptr, const char* end, size_t lineNo) |
---|
128 | { |
---|
129 | if (ptr+2>end || *ptr!='!' || ptr[1]!='!') |
---|
130 | return YAMLValType::NONE; // no type |
---|
131 | if (ptr+7 && ::strncmp(ptr+2, "null", 4)==0 && isSpace(ptr[6]) && ptr[6]!='\n') |
---|
132 | { |
---|
133 | ptr += 6; |
---|
134 | return YAMLValType::NIL; |
---|
135 | } |
---|
136 | else if (ptr+7 && ::strncmp(ptr+2, "bool", 4)==0 && isSpace(ptr[6]) && ptr[6]!='\n') |
---|
137 | { |
---|
138 | ptr += 6; |
---|
139 | return YAMLValType::BOOL; |
---|
140 | } |
---|
141 | else if (ptr+6 && ::strncmp(ptr+2, "int", 3)==0 && isSpace(ptr[5]) && ptr[5]!='\n') |
---|
142 | { |
---|
143 | ptr += 5; |
---|
144 | return YAMLValType::INT; |
---|
145 | } |
---|
146 | else if (ptr+8 && ::strncmp(ptr+2, "float", 5)==0 && isSpace(ptr[7]) && ptr[7]!='\n') |
---|
147 | { |
---|
148 | ptr += 7; |
---|
149 | return YAMLValType::FLOAT; |
---|
150 | } |
---|
151 | else if (ptr+6 && ::strncmp(ptr+2, "str", 3)==0 && isSpace(ptr[5]) && ptr[5]!='\n') |
---|
152 | { |
---|
153 | ptr += 5; |
---|
154 | return YAMLValType::STRING; |
---|
155 | } |
---|
156 | else if (ptr+6 && ::strncmp(ptr+2, "seq", 3)==0 && isSpace(ptr[5]) && ptr[5]!='\n') |
---|
157 | { |
---|
158 | ptr += 5; |
---|
159 | return YAMLValType::SEQ; |
---|
160 | } |
---|
161 | throw ParseException(lineNo, "Unknown YAML value type"); |
---|
162 | } |
---|
163 | |
---|
164 | // parse YAML key (keywords - recognized keys) |
---|
165 | static size_t parseYAMLKey(const char*& ptr, const char* end, size_t lineNo, |
---|
166 | size_t keywordsNum, const char** keywords) |
---|
167 | { |
---|
168 | const char* keyPtr = ptr; |
---|
169 | while (ptr != end && (isAlnum(*ptr) || *ptr=='_')) ptr++; |
---|
170 | if (keyPtr == end) |
---|
171 | throw ParseException(lineNo, "Expected key name"); |
---|
172 | const char* keyEnd = ptr; |
---|
173 | skipSpacesToLineEnd(ptr, end); |
---|
174 | if (ptr == end || *ptr!=':') |
---|
175 | throw ParseException(lineNo, "Expected colon"); |
---|
176 | ptr++; |
---|
177 | const char* afterColon = ptr; |
---|
178 | skipSpacesToLineEnd(ptr, end); |
---|
179 | if (afterColon == ptr && ptr != end && *ptr!='\n') |
---|
180 | // only if not immediate newline |
---|
181 | throw ParseException(lineNo, "After key and colon must be space"); |
---|
182 | CString keyword(keyPtr, keyEnd); |
---|
183 | const size_t index = binaryFind(keywords, keywords+keywordsNum, |
---|
184 | keyword.c_str(), CStringLess()) - keywords; |
---|
185 | return index; |
---|
186 | } |
---|
187 | |
---|
188 | // parse YAML integer value |
---|
189 | template<typename T> |
---|
190 | static T parseYAMLIntValue(const char*& ptr, const char* end, size_t& lineNo, |
---|
191 | bool singleValue = false) |
---|
192 | { |
---|
193 | skipSpacesToLineEnd(ptr, end); |
---|
194 | if (ptr == end || *ptr=='\n') |
---|
195 | throw ParseException(lineNo, "Expected integer value"); |
---|
196 | |
---|
197 | // skip !!int |
---|
198 | YAMLValType valType = parseYAMLType(ptr, end, lineNo); |
---|
199 | if (valType == YAMLValType::INT) |
---|
200 | { // if |
---|
201 | skipSpacesToLineEnd(ptr, end); |
---|
202 | if (ptr == end || *ptr=='\n') |
---|
203 | throw ParseException(lineNo, "Expected integer value"); |
---|
204 | } |
---|
205 | else if (valType != YAMLValType::NONE) |
---|
206 | throw ParseException(lineNo, "Expected value of integer type"); |
---|
207 | |
---|
208 | T value = 0; |
---|
209 | try |
---|
210 | { value = cstrtovCStyle<T>(ptr, end, ptr); } |
---|
211 | catch(const ParseException& ex) |
---|
212 | { throw ParseException(lineNo, ex.what()); } |
---|
213 | |
---|
214 | if (singleValue) |
---|
215 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
216 | return value; |
---|
217 | } |
---|
218 | |
---|
219 | // parse YAML boolean value |
---|
220 | static bool parseYAMLBoolValue(const char*& ptr, const char* end, size_t& lineNo, |
---|
221 | bool singleValue = false) |
---|
222 | { |
---|
223 | skipSpacesToLineEnd(ptr, end); |
---|
224 | if (ptr == end || *ptr=='\n') |
---|
225 | throw ParseException(lineNo, "Expected boolean value"); |
---|
226 | |
---|
227 | // skip !!bool |
---|
228 | YAMLValType valType = parseYAMLType(ptr, end, lineNo); |
---|
229 | if (valType == YAMLValType::BOOL) |
---|
230 | { // if |
---|
231 | skipSpacesToLineEnd(ptr, end); |
---|
232 | if (ptr == end || *ptr=='\n') |
---|
233 | throw ParseException(lineNo, "Expected boolean value"); |
---|
234 | } |
---|
235 | else if (valType != YAMLValType::NONE) |
---|
236 | throw ParseException(lineNo, "Expected value of boolean type"); |
---|
237 | |
---|
238 | const char* wordPtr = ptr; |
---|
239 | while(ptr != end && isAlnum(*ptr)) ptr++; |
---|
240 | CString word(wordPtr, ptr); |
---|
241 | |
---|
242 | bool value = false; |
---|
243 | bool isSet = false; |
---|
244 | for (const char* v: { "1", "true", "t", "on", "yes", "y"}) |
---|
245 | if (::strcasecmp(word.c_str(), v) == 0) |
---|
246 | { |
---|
247 | isSet = true; |
---|
248 | value = true; |
---|
249 | break; |
---|
250 | } |
---|
251 | if (!isSet) |
---|
252 | for (const char* v: { "0", "false", "f", "off", "no", "n"}) |
---|
253 | if (::strcasecmp(word.c_str(), v) == 0) |
---|
254 | { |
---|
255 | isSet = true; |
---|
256 | value = false; |
---|
257 | break; |
---|
258 | } |
---|
259 | if (!isSet) |
---|
260 | throw ParseException(lineNo, "This is not boolean value"); |
---|
261 | |
---|
262 | if (singleValue) |
---|
263 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
264 | return value; |
---|
265 | } |
---|
266 | |
---|
267 | // trim spaces (remove spaces from start and end) |
---|
268 | static std::string trimStrSpaces(const std::string& str) |
---|
269 | { |
---|
270 | size_t i = 0; |
---|
271 | const size_t sz = str.size(); |
---|
272 | while (i!=sz && isSpace(str[i])) i++; |
---|
273 | if (i == sz) return ""; |
---|
274 | size_t j = sz-1; |
---|
275 | while (j>i && isSpace(str[j])) j--; |
---|
276 | return str.substr(i, j-i+1); |
---|
277 | } |
---|
278 | |
---|
279 | static std::string parseYAMLString(const char*& linePtr, const char* end, |
---|
280 | size_t& lineNo) |
---|
281 | { |
---|
282 | std::string strarray; |
---|
283 | if (linePtr == end || (*linePtr != '"' && *linePtr != '\'')) |
---|
284 | { |
---|
285 | while (linePtr != end && !isSpace(*linePtr) && *linePtr != ',') linePtr++; |
---|
286 | throw ParseException(lineNo, "Expected string"); |
---|
287 | } |
---|
288 | const char termChar = *linePtr; |
---|
289 | linePtr++; |
---|
290 | |
---|
291 | // main loop, where is character parsing |
---|
292 | while (linePtr != end && *linePtr != termChar) |
---|
293 | { |
---|
294 | if (*linePtr == '\\') |
---|
295 | { |
---|
296 | // escape |
---|
297 | linePtr++; |
---|
298 | uint16_t value; |
---|
299 | if (linePtr == end) |
---|
300 | throw ParseException(lineNo, "Unterminated character of string"); |
---|
301 | if (*linePtr == 'x') |
---|
302 | { |
---|
303 | // hex literal |
---|
304 | linePtr++; |
---|
305 | if (linePtr == end) |
---|
306 | throw ParseException(lineNo, "Unterminated character of string"); |
---|
307 | value = 0; |
---|
308 | if (isXDigit(*linePtr)) |
---|
309 | for (; linePtr != end; linePtr++) |
---|
310 | { |
---|
311 | cxuint digit; |
---|
312 | if (*linePtr >= '0' && *linePtr <= '9') |
---|
313 | digit = *linePtr-'0'; |
---|
314 | else if (*linePtr >= 'a' && *linePtr <= 'f') |
---|
315 | digit = *linePtr-'a'+10; |
---|
316 | else if (*linePtr >= 'A' && *linePtr <= 'F') |
---|
317 | digit = *linePtr-'A'+10; |
---|
318 | else |
---|
319 | break; |
---|
320 | value = (value<<4) + digit; |
---|
321 | } |
---|
322 | else |
---|
323 | throw ParseException(lineNo, "Expected hexadecimal character code"); |
---|
324 | value &= 0xff; |
---|
325 | } |
---|
326 | else if (isODigit(*linePtr)) |
---|
327 | { |
---|
328 | // octal literal |
---|
329 | value = 0; |
---|
330 | for (cxuint i = 0; linePtr != end && i < 3; i++, linePtr++) |
---|
331 | { |
---|
332 | if (!isODigit(*linePtr)) |
---|
333 | break; |
---|
334 | value = (value<<3) + uint64_t(*linePtr-'0'); |
---|
335 | // checking range |
---|
336 | if (value > 255) |
---|
337 | throw ParseException(lineNo, "Octal code out of range"); |
---|
338 | } |
---|
339 | } |
---|
340 | else |
---|
341 | { |
---|
342 | // normal escapes |
---|
343 | const char c = *linePtr++; |
---|
344 | switch (c) |
---|
345 | { |
---|
346 | case 'a': |
---|
347 | value = '\a'; |
---|
348 | break; |
---|
349 | case 'b': |
---|
350 | value = '\b'; |
---|
351 | break; |
---|
352 | case 'r': |
---|
353 | value = '\r'; |
---|
354 | break; |
---|
355 | case 'n': |
---|
356 | value = '\n'; |
---|
357 | break; |
---|
358 | case 'f': |
---|
359 | value = '\f'; |
---|
360 | break; |
---|
361 | case 'v': |
---|
362 | value = '\v'; |
---|
363 | break; |
---|
364 | case 't': |
---|
365 | value = '\t'; |
---|
366 | break; |
---|
367 | case '\\': |
---|
368 | value = '\\'; |
---|
369 | break; |
---|
370 | case '\'': |
---|
371 | value = '\''; |
---|
372 | break; |
---|
373 | case '\"': |
---|
374 | value = '\"'; |
---|
375 | break; |
---|
376 | default: |
---|
377 | value = c; |
---|
378 | } |
---|
379 | } |
---|
380 | strarray.push_back(value); |
---|
381 | } |
---|
382 | else // regular character |
---|
383 | { |
---|
384 | if (*linePtr=='\n') |
---|
385 | lineNo++; |
---|
386 | strarray.push_back(*linePtr++); |
---|
387 | } |
---|
388 | } |
---|
389 | if (linePtr == end) |
---|
390 | throw ParseException(lineNo, "Unterminated string"); |
---|
391 | linePtr++; |
---|
392 | return strarray; |
---|
393 | } |
---|
394 | |
---|
395 | static std::string parseYAMLStringValue(const char*& ptr, const char* end, size_t& lineNo, |
---|
396 | cxuint prevIndent, bool singleValue = false, bool blockAccept = true) |
---|
397 | { |
---|
398 | skipSpacesToLineEnd(ptr, end); |
---|
399 | if (ptr == end) |
---|
400 | return ""; |
---|
401 | |
---|
402 | // skip !!str |
---|
403 | YAMLValType valType = parseYAMLType(ptr, end, lineNo); |
---|
404 | if (valType == YAMLValType::STRING) |
---|
405 | { // if |
---|
406 | skipSpacesToLineEnd(ptr, end); |
---|
407 | if (ptr == end) |
---|
408 | return ""; |
---|
409 | } |
---|
410 | else if (valType != YAMLValType::NONE) |
---|
411 | throw ParseException(lineNo, "Expected value of string type"); |
---|
412 | |
---|
413 | std::string buf; |
---|
414 | if (*ptr=='"' || *ptr== '\'') |
---|
415 | buf = parseYAMLString(ptr, end, lineNo); |
---|
416 | // otherwise parse stream |
---|
417 | else if (*ptr == '|' || *ptr == '>') |
---|
418 | { |
---|
419 | if (!blockAccept) |
---|
420 | throw ParseException(lineNo, "Illegal block string start"); |
---|
421 | // multiline |
---|
422 | bool newLineFold = *ptr=='>'; |
---|
423 | ptr++; |
---|
424 | skipSpacesToLineEnd(ptr, end); |
---|
425 | if (ptr!=end && *ptr!='\n') |
---|
426 | throw ParseException(lineNo, "Garbages at string block"); |
---|
427 | if (ptr == end) |
---|
428 | return ""; // end |
---|
429 | lineNo++; |
---|
430 | ptr++; // skip newline |
---|
431 | const char* lineStart = ptr; |
---|
432 | skipSpacesToLineEnd(ptr, end); |
---|
433 | size_t indent = ptr - lineStart; |
---|
434 | if (indent <= prevIndent) |
---|
435 | throw ParseException(lineNo, "Unindented string block"); |
---|
436 | |
---|
437 | std::string buf; |
---|
438 | while(ptr != end) |
---|
439 | { |
---|
440 | const char* strStart = ptr; |
---|
441 | while (ptr != end && *ptr!='\n') ptr++; |
---|
442 | buf.append(strStart, ptr); |
---|
443 | |
---|
444 | if (ptr != end) // if new line |
---|
445 | { |
---|
446 | lineNo++; |
---|
447 | ptr++; |
---|
448 | } |
---|
449 | else // end of stream |
---|
450 | break; |
---|
451 | |
---|
452 | const char* lineStart = ptr; |
---|
453 | skipSpacesToLineEnd(ptr, end); |
---|
454 | bool emptyLines = false; |
---|
455 | while (size_t(ptr - lineStart) <= indent) |
---|
456 | { |
---|
457 | if (ptr != end && *ptr=='\n') |
---|
458 | { |
---|
459 | // empty line |
---|
460 | buf.append("\n"); |
---|
461 | ptr++; |
---|
462 | lineNo++; |
---|
463 | lineStart = ptr; |
---|
464 | skipSpacesToLineEnd(ptr, end); |
---|
465 | emptyLines = true; |
---|
466 | continue; |
---|
467 | } |
---|
468 | // if smaller indent |
---|
469 | if (size_t(ptr - lineStart) < indent) |
---|
470 | { |
---|
471 | buf.append("\n"); // always add newline at last line |
---|
472 | if (ptr != end) |
---|
473 | ptr = lineStart; |
---|
474 | return buf; |
---|
475 | } |
---|
476 | else // if this same and not end of line |
---|
477 | break; |
---|
478 | } |
---|
479 | |
---|
480 | if (!emptyLines || !newLineFold) |
---|
481 | // add missing newline after line with text |
---|
482 | // only if no emptyLines or no newLineFold |
---|
483 | buf.append(newLineFold ? " " : "\n"); |
---|
484 | // to indent |
---|
485 | ptr = lineStart + indent; |
---|
486 | } |
---|
487 | return buf; |
---|
488 | } |
---|
489 | else |
---|
490 | { |
---|
491 | // single line string (unquoted) |
---|
492 | const char* strStart = ptr; |
---|
493 | // automatically trim spaces at ends |
---|
494 | const char* strEnd = ptr; |
---|
495 | while (ptr != end && *ptr!='\n' && *ptr!='#') |
---|
496 | { |
---|
497 | if (!isSpace(*ptr)) |
---|
498 | strEnd = ptr; // to trim at end |
---|
499 | ptr++; |
---|
500 | } |
---|
501 | if (strEnd != end && !isSpace(*strEnd)) |
---|
502 | strEnd++; |
---|
503 | |
---|
504 | buf.assign(strStart, strEnd); |
---|
505 | } |
---|
506 | |
---|
507 | if (singleValue) |
---|
508 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
509 | return buf; |
---|
510 | } |
---|
511 | |
---|
512 | /// element consumer class |
---|
513 | class CLRX_INTERNAL YAMLElemConsumer |
---|
514 | { |
---|
515 | public: |
---|
516 | virtual void consume(const char*& ptr, const char* end, size_t& lineNo, |
---|
517 | cxuint prevIndent, bool singleValue, bool blockAccept) = 0; |
---|
518 | }; |
---|
519 | |
---|
520 | static void parseYAMLValArray(const char*& ptr, const char* end, size_t& lineNo, |
---|
521 | size_t prevIndent, YAMLElemConsumer* elemConsumer, bool singleValue = false) |
---|
522 | { |
---|
523 | skipSpacesToLineEnd(ptr, end); |
---|
524 | if (ptr == end) |
---|
525 | return; |
---|
526 | |
---|
527 | // skip !!int |
---|
528 | YAMLValType valType = parseYAMLType(ptr, end, lineNo); |
---|
529 | if (valType == YAMLValType::SEQ) |
---|
530 | { // if |
---|
531 | skipSpacesToLineEnd(ptr, end); |
---|
532 | if (ptr == end) |
---|
533 | return; |
---|
534 | } |
---|
535 | else if (valType != YAMLValType::NONE) |
---|
536 | throw ParseException(lineNo, "Expected value of sequence type"); |
---|
537 | |
---|
538 | if (*ptr == '[') |
---|
539 | { |
---|
540 | // parse array [] |
---|
541 | ptr++; |
---|
542 | skipSpacesAndComments(ptr, end, lineNo); |
---|
543 | while (ptr != end) |
---|
544 | { |
---|
545 | // parse in line |
---|
546 | elemConsumer->consume(ptr, end, lineNo, 0, false, false); |
---|
547 | skipSpacesAndComments(ptr, end, lineNo); |
---|
548 | if (ptr!=end && *ptr==']') |
---|
549 | // just end |
---|
550 | break; |
---|
551 | else if (ptr==end || *ptr!=',') |
---|
552 | throw ParseException(lineNo, "Expected ','"); |
---|
553 | ptr++; |
---|
554 | skipSpacesAndComments(ptr, end, lineNo); |
---|
555 | } |
---|
556 | if (ptr == end) |
---|
557 | throw ParseException(lineNo, "Unterminated array"); |
---|
558 | ptr++; |
---|
559 | |
---|
560 | if (singleValue) |
---|
561 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
562 | return; |
---|
563 | } |
---|
564 | // parse sequence |
---|
565 | size_t oldLineNo = lineNo; |
---|
566 | size_t indent0 = skipSpacesAndComments(ptr, end, lineNo); |
---|
567 | if (ptr == end || lineNo == oldLineNo) |
---|
568 | throw ParseException(lineNo, "Expected sequence of values"); |
---|
569 | |
---|
570 | if (indent0 < prevIndent) |
---|
571 | throw ParseException(lineNo, "Unindented sequence of objects"); |
---|
572 | |
---|
573 | // main loop to parse sequence |
---|
574 | while (ptr != end) |
---|
575 | { |
---|
576 | if (*ptr != '-') |
---|
577 | throw ParseException(lineNo, "No '-' before element value"); |
---|
578 | ptr++; |
---|
579 | const char* afterMinus = ptr; |
---|
580 | skipSpacesToLineEnd(ptr, end); |
---|
581 | if (afterMinus == ptr) |
---|
582 | throw ParseException(lineNo, "No spaces after '-'"); |
---|
583 | elemConsumer->consume(ptr, end, lineNo, indent0, true, true); |
---|
584 | |
---|
585 | size_t indent = skipSpacesAndComments(ptr, end, lineNo); |
---|
586 | if (indent < indent0) |
---|
587 | { |
---|
588 | // if parent level |
---|
589 | ptr -= indent; |
---|
590 | break; |
---|
591 | } |
---|
592 | if (indent != indent0) |
---|
593 | throw ParseException(lineNo, "Wrong indentation of element"); |
---|
594 | } |
---|
595 | } |
---|
596 | |
---|
597 | // integer element consumer |
---|
598 | template<typename T> |
---|
599 | class CLRX_INTERNAL YAMLIntArrayConsumer: public YAMLElemConsumer |
---|
600 | { |
---|
601 | private: |
---|
602 | size_t elemsNum; |
---|
603 | size_t requiredElemsNum; |
---|
604 | public: |
---|
605 | T* array; |
---|
606 | |
---|
607 | YAMLIntArrayConsumer(size_t reqElemsNum, T* _array) |
---|
608 | : elemsNum(0), requiredElemsNum(reqElemsNum), array(_array) |
---|
609 | { } |
---|
610 | |
---|
611 | virtual void consume(const char*& ptr, const char* end, size_t& lineNo, |
---|
612 | cxuint prevIndent, bool singleValue, bool blockAccept) |
---|
613 | { |
---|
614 | if (elemsNum == requiredElemsNum) |
---|
615 | throw ParseException(lineNo, "Too many elements"); |
---|
616 | try |
---|
617 | { array[elemsNum] = cstrtovCStyle<T>(ptr, end, ptr); } |
---|
618 | catch(const ParseException& ex) |
---|
619 | { throw ParseException(lineNo, ex.what()); } |
---|
620 | elemsNum++; |
---|
621 | if (singleValue) |
---|
622 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
623 | } |
---|
624 | }; |
---|
625 | |
---|
626 | // printf info string consumer |
---|
627 | class CLRX_INTERNAL YAMLPrintfVectorConsumer: public YAMLElemConsumer |
---|
628 | { |
---|
629 | private: |
---|
630 | std::unordered_set<cxuint> printfIds; |
---|
631 | public: |
---|
632 | std::vector<ROCmPrintfInfo>& printfInfos; |
---|
633 | |
---|
634 | YAMLPrintfVectorConsumer(std::vector<ROCmPrintfInfo>& _printInfos) |
---|
635 | : printfInfos(_printInfos) |
---|
636 | { } |
---|
637 | |
---|
638 | virtual void consume(const char*& ptr, const char* end, size_t& lineNo, |
---|
639 | cxuint prevIndent, bool singleValue, bool blockAccept) |
---|
640 | { |
---|
641 | const size_t oldLineNo = lineNo; |
---|
642 | std::string str = parseYAMLStringValue(ptr, end, lineNo, prevIndent, |
---|
643 | singleValue, blockAccept); |
---|
644 | // parse printf string |
---|
645 | ROCmPrintfInfo printfInfo{}; |
---|
646 | |
---|
647 | const char* ptr2 = str.c_str(); |
---|
648 | const char* end2 = str.c_str() + str.size(); |
---|
649 | skipSpacesToLineEnd(ptr2, end2); |
---|
650 | try |
---|
651 | { printfInfo.id = cstrtovCStyle<uint32_t>(ptr2, end2, ptr2); } |
---|
652 | catch(const ParseException& ex) |
---|
653 | { throw ParseException(oldLineNo, ex.what()); } |
---|
654 | |
---|
655 | // check printf id uniqueness |
---|
656 | if (!printfIds.insert(printfInfo.id).second) |
---|
657 | throw ParseException(oldLineNo, "Duplicate of printf id"); |
---|
658 | |
---|
659 | skipSpacesToLineEnd(ptr2, end2); |
---|
660 | if (ptr2==end || *ptr2!=':') |
---|
661 | throw ParseException(oldLineNo, "No colon after printf callId"); |
---|
662 | ptr2++; |
---|
663 | skipSpacesToLineEnd(ptr2, end2); |
---|
664 | uint32_t argsNum = cstrtovCStyle<uint32_t>(ptr2, end2, ptr2); |
---|
665 | skipSpacesToLineEnd(ptr2, end2); |
---|
666 | if (ptr2==end || *ptr2!=':') |
---|
667 | throw ParseException(oldLineNo, "No colon after printf argsNum"); |
---|
668 | ptr2++; |
---|
669 | |
---|
670 | printfInfo.argSizes.resize(argsNum); |
---|
671 | |
---|
672 | // parse arg sizes |
---|
673 | for (size_t i = 0; i < argsNum; i++) |
---|
674 | { |
---|
675 | skipSpacesToLineEnd(ptr2, end2); |
---|
676 | printfInfo.argSizes[i] = cstrtovCStyle<uint32_t>(ptr2, end2, ptr2); |
---|
677 | skipSpacesToLineEnd(ptr2, end2); |
---|
678 | if (ptr2==end || *ptr2!=':') |
---|
679 | throw ParseException(lineNo, "No colon after printf argsNum"); |
---|
680 | ptr2++; |
---|
681 | } |
---|
682 | // format |
---|
683 | printfInfo.format.assign(ptr2, end2); |
---|
684 | |
---|
685 | printfInfos.push_back(printfInfo); |
---|
686 | } |
---|
687 | }; |
---|
688 | |
---|
689 | // skip YAML value after key |
---|
690 | static void skipYAMLValue(const char*& ptr, const char* end, size_t& lineNo, |
---|
691 | cxuint prevIndent, bool singleValue = true) |
---|
692 | { |
---|
693 | skipSpacesToLineEnd(ptr, end); |
---|
694 | if (ptr+2 >= end && ptr[0]=='!' && ptr[1]=='!') |
---|
695 | { // skip !!xxxxx |
---|
696 | ptr+=2; |
---|
697 | while (ptr!=end && isAlpha(*ptr)) ptr++; |
---|
698 | skipSpacesToLineEnd(ptr, end); |
---|
699 | } |
---|
700 | |
---|
701 | if (ptr==end || (*ptr!='\'' && *ptr!='"' && *ptr!='|' && *ptr!='>' && *ptr !='[' && |
---|
702 | *ptr!='#' && *ptr!='\n')) |
---|
703 | { |
---|
704 | while (ptr!=end && *ptr!='\n') ptr++; |
---|
705 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
706 | return; |
---|
707 | } |
---|
708 | // string |
---|
709 | if (*ptr=='\'' || *ptr=='"') |
---|
710 | { |
---|
711 | const char delim = *ptr++; |
---|
712 | bool escape = false; |
---|
713 | while(ptr!=end && (escape || *ptr!=delim)) |
---|
714 | { |
---|
715 | if (!escape && *ptr=='\\') |
---|
716 | escape = true; |
---|
717 | else if (escape) |
---|
718 | escape = false; |
---|
719 | if (*ptr=='\n') lineNo++; |
---|
720 | ptr++; |
---|
721 | } |
---|
722 | if (ptr==end) |
---|
723 | throw ParseException(lineNo, "Unterminated string"); |
---|
724 | ptr++; |
---|
725 | if (singleValue) |
---|
726 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
727 | } |
---|
728 | else if (*ptr=='[') |
---|
729 | { // otherwise [array] |
---|
730 | ptr++; |
---|
731 | skipSpacesAndComments(ptr, end, lineNo); |
---|
732 | while (ptr != end) |
---|
733 | { |
---|
734 | // parse in line |
---|
735 | if (ptr!=end && (*ptr=='\'' || *ptr=='"')) |
---|
736 | // skip YAML string |
---|
737 | skipYAMLValue(ptr, end, lineNo, 0, false); |
---|
738 | else |
---|
739 | while (ptr!=end && *ptr!='\n' && |
---|
740 | *ptr!='#' && *ptr!=',' && *ptr!=']') ptr++; |
---|
741 | skipSpacesAndComments(ptr, end, lineNo); |
---|
742 | |
---|
743 | if (ptr!=end && *ptr==']') |
---|
744 | // just end |
---|
745 | break; |
---|
746 | else if (ptr!=end && *ptr!=',') |
---|
747 | throw ParseException(lineNo, "Expected ','"); |
---|
748 | ptr++; |
---|
749 | skipSpacesAndComments(ptr, end, lineNo); |
---|
750 | } |
---|
751 | if (ptr == end) |
---|
752 | throw ParseException(lineNo, "Unterminated array"); |
---|
753 | ptr++; |
---|
754 | skipSpacesToNextLine(ptr, end, lineNo); |
---|
755 | } |
---|
756 | else |
---|
757 | { // block value |
---|
758 | bool blockValue = false; |
---|
759 | if (ptr!=end && (*ptr=='|' || *ptr=='>')) |
---|
760 | { |
---|
761 | ptr++; // skip '|' or '>' |
---|
762 | blockValue = true; |
---|
763 | } |
---|
764 | if (ptr!=end && *ptr=='#') |
---|
765 | while (ptr!=end && *ptr!='\n') ptr++; |
---|
766 | else |
---|
767 | skipSpacesToLineEnd(ptr, end); |
---|
768 | if (ptr!=end && *ptr!='\n') |
---|
769 | throw ParseException(lineNo, "Garbages before block or children"); |
---|
770 | ptr++; |
---|
771 | lineNo++; |
---|
772 | // skip all lines indented beyound previous level |
---|
773 | while (ptr != end) |
---|
774 | { |
---|
775 | const char* lineStart = ptr; |
---|
776 | skipSpacesToLineEnd(ptr, end); |
---|
777 | if (ptr == end) |
---|
778 | { |
---|
779 | ptr++; |
---|
780 | lineNo++; |
---|
781 | continue; |
---|
782 | } |
---|
783 | if (size_t(ptr-lineStart) <= prevIndent && *ptr!='\n' && |
---|
784 | (blockValue || *ptr!='#')) |
---|
785 | // if indent is short and not empty line (same spaces) or |
---|
786 | // or with only comment and not blockValue |
---|
787 | { |
---|
788 | ptr = lineStart; |
---|
789 | break; |
---|
790 | } |
---|
791 | |
---|
792 | while (ptr!=end && *ptr!='\n') ptr++; |
---|
793 | if (ptr!=end) |
---|
794 | { |
---|
795 | lineNo++; |
---|
796 | ptr++; |
---|
797 | } |
---|
798 | } |
---|
799 | } |
---|
800 | } |
---|
801 | |
---|
802 | enum { |
---|
803 | ROCMMT_MAIN_KERNELS = 0, ROCMMT_MAIN_PRINTF, ROCMMT_MAIN_VERSION |
---|
804 | }; |
---|
805 | |
---|
806 | static const char* mainMetadataKeywords[] = |
---|
807 | { |
---|
808 | "Kernels", "Printf", "Version" |
---|
809 | }; |
---|
810 | |
---|
811 | static const size_t mainMetadataKeywordsNum = |
---|
812 | sizeof(mainMetadataKeywords) / sizeof(const char*); |
---|
813 | |
---|
814 | enum { |
---|
815 | ROCMMT_KERNEL_ARGS = 0, ROCMMT_KERNEL_ATTRS, ROCMMT_KERNEL_CODEPROPS, |
---|
816 | ROCMMT_KERNEL_LANGUAGE, ROCMMT_KERNEL_LANGUAGE_VERSION, |
---|
817 | ROCMMT_KERNEL_NAME, ROCMMT_KERNEL_SYMBOLNAME |
---|
818 | }; |
---|
819 | |
---|
820 | static const char* kernelMetadataKeywords[] = |
---|
821 | { |
---|
822 | "Args", "Attrs", "CodeProps", "Language", "LanguageVersion", "Name", "SymbolName" |
---|
823 | }; |
---|
824 | |
---|
825 | static const size_t kernelMetadataKeywordsNum = |
---|
826 | sizeof(kernelMetadataKeywords) / sizeof(const char*); |
---|
827 | |
---|
828 | enum { |
---|
829 | ROCMMT_ATTRS_REQD_WORK_GROUP_SIZE = 0, ROCMMT_ATTRS_RUNTIME_HANDLE, |
---|
830 | ROCMMT_ATTRS_VECTYPEHINT, ROCMMT_ATTRS_WORK_GROUP_SIZE_HINT |
---|
831 | }; |
---|
832 | |
---|
833 | static const char* kernelAttrMetadataKeywords[] = |
---|
834 | { |
---|
835 | "ReqdWorkGroupSize", "RuntimeHandle", "VecTypeHint", "WorkGroupSizeHint" |
---|
836 | }; |
---|
837 | |
---|
838 | static const size_t kernelAttrMetadataKeywordsNum = |
---|
839 | sizeof(kernelAttrMetadataKeywords) / sizeof(const char*); |
---|
840 | |
---|
841 | enum { |
---|
842 | ROCMMT_CODEPROPS_FIXED_WORK_GROUP_SIZE = 0, ROCMMT_CODEPROPS_GROUP_SEGMENT_FIXED_SIZE, |
---|
843 | ROCMMT_CODEPROPS_KERNARG_SEGMENT_ALIGN, ROCMMT_CODEPROPS_KERNARG_SEGMENT_SIZE, |
---|
844 | ROCMMT_CODEPROPS_MAX_FLAT_WORK_GROUP_SIZE, ROCMMT_CODEPROPS_NUM_SGPRS, |
---|
845 | ROCMMT_CODEPROPS_NUM_SPILLED_SGPRS, ROCMMT_CODEPROPS_NUM_SPILLED_VGPRS, |
---|
846 | ROCMMT_CODEPROPS_NUM_VGPRS, ROCMMT_CODEPROPS_PRIVATE_SEGMENT_FIXED_SIZE, |
---|
847 | ROCMMT_CODEPROPS_WAVEFRONT_SIZE |
---|
848 | }; |
---|
849 | |
---|
850 | static const char* kernelCodePropsKeywords[] = |
---|
851 | { |
---|
852 | "FixedWorkGroupSize", "GroupSegmentFixedSize", "KernargSegmentAlign", |
---|
853 | "KernargSegmentSize", "MaxFlatWorkGroupSize", "NumSGPRs", |
---|
854 | "NumSpilledSGPRs", "NumSpilledVGPRs", "NumVGPRs", "PrivateSegmentFixedSize", |
---|
855 | "WavefrontSize" |
---|
856 | }; |
---|
857 | |
---|
858 | static const size_t kernelCodePropsKeywordsNum = |
---|
859 | sizeof(kernelCodePropsKeywords) / sizeof(const char*); |
---|
860 | |
---|
861 | enum { |
---|
862 | ROCMMT_ARGS_ACCQUAL = 0, ROCMMT_ARGS_ACTUALACCQUAL, ROCMMT_ARGS_ADDRSPACEQUAL, |
---|
863 | ROCMMT_ARGS_ALIGN, ROCMMT_ARGS_ISCONST, ROCMMT_ARGS_ISPIPE, ROCMMT_ARGS_ISRESTRICT, |
---|
864 | ROCMMT_ARGS_ISVOLATILE, ROCMMT_ARGS_NAME, ROCMMT_ARGS_POINTEE_ALIGN, |
---|
865 | ROCMMT_ARGS_SIZE, ROCMMT_ARGS_TYPENAME, ROCMMT_ARGS_VALUEKIND, |
---|
866 | ROCMMT_ARGS_VALUETYPE |
---|
867 | }; |
---|
868 | |
---|
869 | static const char* kernelArgInfosKeywords[] = |
---|
870 | { |
---|
871 | "AccQual", "ActualAccQual", "AddrSpaceQual", "Align", "IsConst", "IsPipe", |
---|
872 | "IsRestrict", "IsVolatile", "Name", "PointeeAlign", "Size", "TypeName", |
---|
873 | "ValueKind", "ValueType" |
---|
874 | }; |
---|
875 | |
---|
876 | static const size_t kernelArgInfosKeywordsNum = |
---|
877 | sizeof(kernelArgInfosKeywords) / sizeof(const char*); |
---|
878 | |
---|
879 | static const std::pair<const char*, ROCmValueKind> rocmValueKindNamesMap[] = |
---|
880 | { |
---|
881 | { "ByValue", ROCmValueKind::BY_VALUE }, |
---|
882 | { "DynamicSharedPointer", ROCmValueKind::DYN_SHARED_PTR }, |
---|
883 | { "GlobalBuffer", ROCmValueKind::GLOBAL_BUFFER }, |
---|
884 | { "HiddenCompletionAction", ROCmValueKind::HIDDEN_COMPLETION_ACTION }, |
---|
885 | { "HiddenDefaultQueue", ROCmValueKind::HIDDEN_DEFAULT_QUEUE }, |
---|
886 | { "HiddenGlobalOffsetX", ROCmValueKind::HIDDEN_GLOBAL_OFFSET_X }, |
---|
887 | { "HiddenGlobalOffsetY", ROCmValueKind::HIDDEN_GLOBAL_OFFSET_Y }, |
---|
888 | { "HiddenGlobalOffsetZ", ROCmValueKind::HIDDEN_GLOBAL_OFFSET_Z }, |
---|
889 | { "HiddenMultiGridSyncArg", ROCmValueKind::HIDDEN_MULTIGRID_SYNC_ARG }, |
---|
890 | { "HiddenNone", ROCmValueKind::HIDDEN_NONE }, |
---|
891 | { "HiddenPrintfBuffer", ROCmValueKind::HIDDEN_PRINTF_BUFFER }, |
---|
892 | { "Image", ROCmValueKind::IMAGE }, |
---|
893 | { "Pipe", ROCmValueKind::PIPE }, |
---|
894 | { "Queue", ROCmValueKind::QUEUE }, |
---|
895 | { "Sampler", ROCmValueKind::SAMPLER } |
---|
896 | }; |
---|
897 | |
---|
898 | static const size_t rocmValueKindNamesNum = |
---|
899 | sizeof(rocmValueKindNamesMap) / sizeof(std::pair<const char*, ROCmValueKind>); |
---|
900 | |
---|
901 | static const std::pair<const char*, ROCmValueType> rocmValueTypeNamesMap[] = |
---|
902 | { |
---|
903 | { "F16", ROCmValueType::FLOAT16 }, |
---|
904 | { "F32", ROCmValueType::FLOAT32 }, |
---|
905 | { "F64", ROCmValueType::FLOAT64 }, |
---|
906 | { "I16", ROCmValueType::INT16 }, |
---|
907 | { "I32", ROCmValueType::INT32 }, |
---|
908 | { "I64", ROCmValueType::INT64 }, |
---|
909 | { "I8", ROCmValueType::INT8 }, |
---|
910 | { "Struct", ROCmValueType::STRUCTURE }, |
---|
911 | { "U16", ROCmValueType::UINT16 }, |
---|
912 | { "U32", ROCmValueType::UINT32 }, |
---|
913 | { "U64", ROCmValueType::UINT64 }, |
---|
914 | { "U8", ROCmValueType::UINT8 } |
---|
915 | }; |
---|
916 | |
---|
917 | static const size_t rocmValueTypeNamesNum = |
---|
918 | sizeof(rocmValueTypeNamesMap) / sizeof(std::pair<const char*, ROCmValueType>); |
---|
919 | |
---|
920 | static const char* rocmAddrSpaceTypesTbl[] = |
---|
921 | { "Private", "Global", "Constant", "Local", "Generic", "Region" }; |
---|
922 | |
---|
923 | static const char* rocmAccessQualifierTbl[] = |
---|
924 | { "Default", "ReadOnly", "WriteOnly", "ReadWrite" }; |
---|
925 | |
---|
926 | static void parseROCmMetadata(size_t metadataSize, const char* metadata, |
---|
927 | ROCmMetadata& metadataInfo) |
---|
928 | { |
---|
929 | const char* ptr = metadata; |
---|
930 | const char* end = metadata + metadataSize; |
---|
931 | size_t lineNo = 1; |
---|
932 | // init metadata info object |
---|
933 | metadataInfo.kernels.clear(); |
---|
934 | metadataInfo.printfInfos.clear(); |
---|
935 | metadataInfo.version[0] = metadataInfo.version[1] = 0; |
---|
936 | |
---|
937 | std::vector<ROCmKernelMetadata>& kernels = metadataInfo.kernels; |
---|
938 | |
---|
939 | cxuint levels[6] = { UINT_MAX, UINT_MAX, UINT_MAX, UINT_MAX, UINT_MAX, UINT_MAX }; |
---|
940 | cxuint curLevel = 0; |
---|
941 | bool inKernels = false; |
---|
942 | bool inKernel = false; |
---|
943 | bool inKernelArgs = false; |
---|
944 | bool inKernelArg = false; |
---|
945 | bool inKernelCodeProps = false; |
---|
946 | bool inKernelAttrs = false; |
---|
947 | bool canToNextLevel = false; |
---|
948 | |
---|
949 | size_t oldLineNo = 0; |
---|
950 | while (ptr != end) |
---|
951 | { |
---|
952 | cxuint level = skipSpacesAndComments(ptr, end, lineNo); |
---|
953 | if (ptr == end || lineNo == oldLineNo) |
---|
954 | throw ParseException(lineNo, "Expected new line"); |
---|
955 | |
---|
956 | if (levels[curLevel] == UINT_MAX) |
---|
957 | levels[curLevel] = level; |
---|
958 | else if (levels[curLevel] < level) |
---|
959 | { |
---|
960 | if (canToNextLevel) |
---|
961 | // go to next nesting level |
---|
962 | levels[++curLevel] = level; |
---|
963 | else |
---|
964 | throw ParseException(lineNo, "Unexpected nesting level"); |
---|
965 | canToNextLevel = false; |
---|
966 | } |
---|
967 | else if (levels[curLevel] > level) |
---|
968 | { |
---|
969 | while (curLevel != UINT_MAX && levels[curLevel] > level) |
---|
970 | curLevel--; |
---|
971 | if (curLevel == UINT_MAX) |
---|
972 | throw ParseException(lineNo, "Indentation smaller than in main level"); |
---|
973 | |
---|
974 | // pop from previous level |
---|
975 | if (curLevel < 3) |
---|
976 | { |
---|
977 | if (inKernelArgs) |
---|
978 | { |
---|
979 | // leave from kernel args |
---|
980 | inKernelArgs = false; |
---|
981 | inKernelArg = false; |
---|
982 | } |
---|
983 | |
---|
984 | inKernelCodeProps = false; |
---|
985 | inKernelAttrs = false; |
---|
986 | } |
---|
987 | if (curLevel < 1 && inKernels) |
---|
988 | { |
---|
989 | // leave from kernels |
---|
990 | inKernels = false; |
---|
991 | inKernel = false; |
---|
992 | } |
---|
993 | |
---|
994 | if (levels[curLevel] != level) |
---|
995 | throw ParseException(lineNo, "Unexpected nesting level"); |
---|
996 | } |
---|
997 | |
---|
998 | oldLineNo = lineNo; |
---|
999 | if (curLevel == 0) |
---|
1000 | { |
---|
1001 | if (lineNo==1 && ptr+3 <= end && *ptr=='-' && ptr[1]=='-' && ptr[2]=='-' && |
---|
1002 | (ptr+3==end || (ptr+3 < end && ptr[3]=='\n'))) |
---|
1003 | { |
---|
1004 | ptr += 3; |
---|
1005 | if (ptr!=end) |
---|
1006 | { |
---|
1007 | lineNo++; |
---|
1008 | ptr++; // to newline |
---|
1009 | } |
---|
1010 | continue; // skip document start |
---|
1011 | } |
---|
1012 | |
---|
1013 | if (ptr+3 <= end && *ptr=='.' && ptr[1]=='.' && ptr[2]=='.' && |
---|
1014 | (ptr+3==end || (ptr+3 < end && ptr[3]=='\n'))) |
---|
1015 | break; // end of the document |
---|
1016 | |
---|
1017 | const size_t keyIndex = parseYAMLKey(ptr, end, lineNo, |
---|
1018 | mainMetadataKeywordsNum, mainMetadataKeywords); |
---|
1019 | |
---|
1020 | switch(keyIndex) |
---|
1021 | { |
---|
1022 | case ROCMMT_MAIN_KERNELS: |
---|
1023 | inKernels = true; |
---|
1024 | canToNextLevel = true; |
---|
1025 | break; |
---|
1026 | case ROCMMT_MAIN_PRINTF: |
---|
1027 | { |
---|
1028 | YAMLPrintfVectorConsumer consumer(metadataInfo.printfInfos); |
---|
1029 | parseYAMLValArray(ptr, end, lineNo, levels[curLevel], &consumer, true); |
---|
1030 | break; |
---|
1031 | } |
---|
1032 | case ROCMMT_MAIN_VERSION: |
---|
1033 | { |
---|
1034 | YAMLIntArrayConsumer<uint32_t> consumer(2, metadataInfo.version); |
---|
1035 | parseYAMLValArray(ptr, end, lineNo, levels[curLevel], &consumer, true); |
---|
1036 | break; |
---|
1037 | } |
---|
1038 | default: |
---|
1039 | skipYAMLValue(ptr, end, lineNo, level); |
---|
1040 | break; |
---|
1041 | } |
---|
1042 | } |
---|
1043 | |
---|
1044 | if (curLevel==1 && inKernels) |
---|
1045 | { |
---|
1046 | // enter to kernel level |
---|
1047 | if (ptr == end || *ptr != '-') |
---|
1048 | throw ParseException(lineNo, "No '-' before kernel object"); |
---|
1049 | ptr++; |
---|
1050 | const char* afterMinus = ptr; |
---|
1051 | skipSpacesToLineEnd(ptr, end); |
---|
1052 | levels[++curLevel] = level + 1 + ptr-afterMinus; |
---|
1053 | level = levels[curLevel]; |
---|
1054 | inKernel = true; |
---|
1055 | |
---|
1056 | kernels.push_back(ROCmKernelMetadata()); |
---|
1057 | kernels.back().initialize(); |
---|
1058 | } |
---|
1059 | |
---|
1060 | if (curLevel==2 && inKernel) |
---|
1061 | { |
---|
1062 | // in kernel |
---|
1063 | const size_t keyIndex = parseYAMLKey(ptr, end, lineNo, |
---|
1064 | kernelMetadataKeywordsNum, kernelMetadataKeywords); |
---|
1065 | |
---|
1066 | ROCmKernelMetadata& kernel = kernels.back(); |
---|
1067 | switch(keyIndex) |
---|
1068 | { |
---|
1069 | case ROCMMT_KERNEL_ARGS: |
---|
1070 | inKernelArgs = true; |
---|
1071 | canToNextLevel = true; |
---|
1072 | kernel.argInfos.clear(); |
---|
1073 | break; |
---|
1074 | case ROCMMT_KERNEL_ATTRS: |
---|
1075 | inKernelAttrs = true; |
---|
1076 | canToNextLevel = true; |
---|
1077 | // initialize kernel attributes values |
---|
1078 | kernel.reqdWorkGroupSize[0] = 0; |
---|
1079 | kernel.reqdWorkGroupSize[1] = 0; |
---|
1080 | kernel.reqdWorkGroupSize[2] = 0; |
---|
1081 | kernel.workGroupSizeHint[0] = 0; |
---|
1082 | kernel.workGroupSizeHint[1] = 0; |
---|
1083 | kernel.workGroupSizeHint[2] = 0; |
---|
1084 | kernel.runtimeHandle.clear(); |
---|
1085 | kernel.vecTypeHint.clear(); |
---|
1086 | break; |
---|
1087 | case ROCMMT_KERNEL_CODEPROPS: |
---|
1088 | // initialize CodeProps values |
---|
1089 | kernel.kernargSegmentSize = BINGEN64_DEFAULT; |
---|
1090 | kernel.groupSegmentFixedSize = BINGEN64_DEFAULT; |
---|
1091 | kernel.privateSegmentFixedSize = BINGEN64_DEFAULT; |
---|
1092 | kernel.kernargSegmentAlign = BINGEN64_DEFAULT; |
---|
1093 | kernel.wavefrontSize = BINGEN_DEFAULT; |
---|
1094 | kernel.sgprsNum = BINGEN_DEFAULT; |
---|
1095 | kernel.vgprsNum = BINGEN_DEFAULT; |
---|
1096 | kernel.spilledSgprs = BINGEN_NOTSUPPLIED; |
---|
1097 | kernel.spilledVgprs = BINGEN_NOTSUPPLIED; |
---|
1098 | kernel.maxFlatWorkGroupSize = BINGEN64_DEFAULT; |
---|
1099 | kernel.fixedWorkGroupSize[0] = 0; |
---|
1100 | kernel.fixedWorkGroupSize[1] = 0; |
---|
1101 | kernel.fixedWorkGroupSize[2] = 0; |
---|
1102 | inKernelCodeProps = true; |
---|
1103 | canToNextLevel = true; |
---|
1104 | break; |
---|
1105 | case ROCMMT_KERNEL_LANGUAGE: |
---|
1106 | kernel.language = parseYAMLStringValue(ptr, end, lineNo, level, true); |
---|
1107 | break; |
---|
1108 | case ROCMMT_KERNEL_LANGUAGE_VERSION: |
---|
1109 | { |
---|
1110 | YAMLIntArrayConsumer<uint32_t> consumer(2, kernel.langVersion); |
---|
1111 | parseYAMLValArray(ptr, end, lineNo, levels[curLevel], &consumer); |
---|
1112 | break; |
---|
1113 | } |
---|
1114 | case ROCMMT_KERNEL_NAME: |
---|
1115 | kernel.name = parseYAMLStringValue(ptr, end, lineNo, level, true); |
---|
1116 | break; |
---|
1117 | case ROCMMT_KERNEL_SYMBOLNAME: |
---|
1118 | kernel.symbolName = parseYAMLStringValue(ptr, end, lineNo, level, true); |
---|
1119 | break; |
---|
1120 | default: |
---|
1121 | skipYAMLValue(ptr, end, lineNo, level); |
---|
1122 | break; |
---|
1123 | } |
---|
1124 | } |
---|
1125 | |
---|
1126 | if (curLevel==3 && inKernelAttrs) |
---|
1127 | { |
---|
1128 | // in kernel attributes |
---|
1129 | const size_t keyIndex = parseYAMLKey(ptr, end, lineNo, |
---|
1130 | kernelAttrMetadataKeywordsNum, kernelAttrMetadataKeywords); |
---|
1131 | |
---|
1132 | ROCmKernelMetadata& kernel = kernels.back(); |
---|
1133 | switch(keyIndex) |
---|
1134 | { |
---|
1135 | case ROCMMT_ATTRS_REQD_WORK_GROUP_SIZE: |
---|
1136 | { |
---|
1137 | YAMLIntArrayConsumer<cxuint> consumer(3, kernel.reqdWorkGroupSize); |
---|
1138 | parseYAMLValArray(ptr, end, lineNo, level, &consumer); |
---|
1139 | break; |
---|
1140 | } |
---|
1141 | case ROCMMT_ATTRS_RUNTIME_HANDLE: |
---|
1142 | kernel.runtimeHandle = parseYAMLStringValue( |
---|
1143 | ptr, end, lineNo, level, true); |
---|
1144 | break; |
---|
1145 | case ROCMMT_ATTRS_VECTYPEHINT: |
---|
1146 | kernel.vecTypeHint = parseYAMLStringValue( |
---|
1147 | ptr, end, lineNo, level, true); |
---|
1148 | break; |
---|
1149 | case ROCMMT_ATTRS_WORK_GROUP_SIZE_HINT: |
---|
1150 | { |
---|
1151 | YAMLIntArrayConsumer<cxuint> consumer(3, kernel.workGroupSizeHint); |
---|
1152 | parseYAMLValArray(ptr, end, lineNo, level, &consumer, true); |
---|
1153 | break; |
---|
1154 | } |
---|
1155 | default: |
---|
1156 | skipYAMLValue(ptr, end, lineNo, level); |
---|
1157 | break; |
---|
1158 | } |
---|
1159 | } |
---|
1160 | |
---|
1161 | if (curLevel==3 && inKernelCodeProps) |
---|
1162 | { |
---|
1163 | // in kernel codeProps |
---|
1164 | const size_t keyIndex = parseYAMLKey(ptr, end, lineNo, |
---|
1165 | kernelCodePropsKeywordsNum, kernelCodePropsKeywords); |
---|
1166 | |
---|
1167 | ROCmKernelMetadata& kernel = kernels.back(); |
---|
1168 | switch(keyIndex) |
---|
1169 | { |
---|
1170 | case ROCMMT_CODEPROPS_FIXED_WORK_GROUP_SIZE: |
---|
1171 | { |
---|
1172 | YAMLIntArrayConsumer<cxuint> consumer(3, kernel.fixedWorkGroupSize); |
---|
1173 | parseYAMLValArray(ptr, end, lineNo, level, &consumer); |
---|
1174 | break; |
---|
1175 | } |
---|
1176 | case ROCMMT_CODEPROPS_GROUP_SEGMENT_FIXED_SIZE: |
---|
1177 | kernel.groupSegmentFixedSize = |
---|
1178 | parseYAMLIntValue<cxuint>(ptr, end, lineNo, true); |
---|
1179 | break; |
---|
1180 | case ROCMMT_CODEPROPS_KERNARG_SEGMENT_ALIGN: |
---|
1181 | kernel.kernargSegmentAlign = |
---|
1182 | parseYAMLIntValue<uint64_t>(ptr, end, lineNo, true); |
---|
1183 | break; |
---|
1184 | case ROCMMT_CODEPROPS_KERNARG_SEGMENT_SIZE: |
---|
1185 | kernel.kernargSegmentSize = |
---|
1186 | parseYAMLIntValue<uint64_t>(ptr, end, lineNo, true); |
---|
1187 | break; |
---|
1188 | case ROCMMT_CODEPROPS_MAX_FLAT_WORK_GROUP_SIZE: |
---|
1189 | kernel.maxFlatWorkGroupSize = |
---|
1190 | parseYAMLIntValue<uint64_t>(ptr, end, lineNo, true); |
---|
1191 | break; |
---|
1192 | case ROCMMT_CODEPROPS_NUM_SGPRS: |
---|
1193 | kernel.sgprsNum = parseYAMLIntValue<cxuint>(ptr, end, lineNo, true); |
---|
1194 | break; |
---|
1195 | case ROCMMT_CODEPROPS_NUM_SPILLED_SGPRS: |
---|
1196 | kernel.spilledSgprs = |
---|
1197 | parseYAMLIntValue<cxuint>(ptr, end, lineNo, true); |
---|
1198 | break; |
---|
1199 | case ROCMMT_CODEPROPS_NUM_SPILLED_VGPRS: |
---|
1200 | kernel.spilledVgprs = |
---|
1201 | parseYAMLIntValue<cxuint>(ptr, end, lineNo, true); |
---|
1202 | break; |
---|
1203 | case ROCMMT_CODEPROPS_NUM_VGPRS: |
---|
1204 | kernel.vgprsNum = parseYAMLIntValue<cxuint>(ptr, end, lineNo, true); |
---|
1205 | break; |
---|
1206 | case ROCMMT_CODEPROPS_PRIVATE_SEGMENT_FIXED_SIZE: |
---|
1207 | kernel.privateSegmentFixedSize = |
---|
1208 | parseYAMLIntValue<uint64_t>(ptr, end, lineNo, true); |
---|
1209 | break; |
---|
1210 | case ROCMMT_CODEPROPS_WAVEFRONT_SIZE: |
---|
1211 | kernel.wavefrontSize = |
---|
1212 | parseYAMLIntValue<cxuint>(ptr, end, lineNo, true); |
---|
1213 | break; |
---|
1214 | default: |
---|
1215 | skipYAMLValue(ptr, end, lineNo, level); |
---|
1216 | break; |
---|
1217 | } |
---|
1218 | } |
---|
1219 | |
---|
1220 | if (curLevel==3 && inKernelArgs) |
---|
1221 | { |
---|
1222 | // enter to kernel argument level |
---|
1223 | if (ptr == end || *ptr != '-') |
---|
1224 | throw ParseException(lineNo, "No '-' before argument object"); |
---|
1225 | ptr++; |
---|
1226 | const char* afterMinus = ptr; |
---|
1227 | skipSpacesToLineEnd(ptr, end); |
---|
1228 | levels[++curLevel] = level + 1 + ptr-afterMinus; |
---|
1229 | level = levels[curLevel]; |
---|
1230 | inKernelArg = true; |
---|
1231 | |
---|
1232 | kernels.back().argInfos.push_back(ROCmKernelArgInfo{}); |
---|
1233 | } |
---|
1234 | |
---|
1235 | if (curLevel==4 && inKernelArg) |
---|
1236 | { |
---|
1237 | // in kernel argument |
---|
1238 | const size_t keyIndex = parseYAMLKey(ptr, end, lineNo, |
---|
1239 | kernelArgInfosKeywordsNum, kernelArgInfosKeywords); |
---|
1240 | |
---|
1241 | ROCmKernelArgInfo& kernelArg = kernels.back().argInfos.back(); |
---|
1242 | |
---|
1243 | size_t valLineNo = lineNo; |
---|
1244 | switch(keyIndex) |
---|
1245 | { |
---|
1246 | case ROCMMT_ARGS_ACCQUAL: |
---|
1247 | case ROCMMT_ARGS_ACTUALACCQUAL: |
---|
1248 | { |
---|
1249 | const std::string acc = trimStrSpaces(parseYAMLStringValue( |
---|
1250 | ptr, end, lineNo, level, true)); |
---|
1251 | size_t accIndex = 0; |
---|
1252 | for (; accIndex < 6; accIndex++) |
---|
1253 | if (::strcmp(rocmAccessQualifierTbl[accIndex], acc.c_str())==0) |
---|
1254 | break; |
---|
1255 | if (accIndex == 4) |
---|
1256 | throw ParseException(lineNo, "Wrong access qualifier"); |
---|
1257 | if (keyIndex == ROCMMT_ARGS_ACCQUAL) |
---|
1258 | kernelArg.accessQual = ROCmAccessQual(accIndex); |
---|
1259 | else |
---|
1260 | kernelArg.actualAccessQual = ROCmAccessQual(accIndex); |
---|
1261 | break; |
---|
1262 | } |
---|
1263 | case ROCMMT_ARGS_ADDRSPACEQUAL: |
---|
1264 | { |
---|
1265 | const std::string aspace = trimStrSpaces(parseYAMLStringValue( |
---|
1266 | ptr, end, lineNo, level, true)); |
---|
1267 | size_t aspaceIndex = 0; |
---|
1268 | for (; aspaceIndex < 6; aspaceIndex++) |
---|
1269 | if (::strcmp(rocmAddrSpaceTypesTbl[aspaceIndex], |
---|
1270 | aspace.c_str())==0) |
---|
1271 | break; |
---|
1272 | if (aspaceIndex == 6) |
---|
1273 | throw ParseException(valLineNo, "Wrong address space"); |
---|
1274 | kernelArg.addressSpace = ROCmAddressSpace(aspaceIndex+1); |
---|
1275 | break; |
---|
1276 | } |
---|
1277 | case ROCMMT_ARGS_ALIGN: |
---|
1278 | kernelArg.align = parseYAMLIntValue<uint64_t>(ptr, end, lineNo, true); |
---|
1279 | break; |
---|
1280 | case ROCMMT_ARGS_ISCONST: |
---|
1281 | kernelArg.isConst = parseYAMLBoolValue(ptr, end, lineNo, true); |
---|
1282 | break; |
---|
1283 | case ROCMMT_ARGS_ISPIPE: |
---|
1284 | kernelArg.isPipe = parseYAMLBoolValue(ptr, end, lineNo, true); |
---|
1285 | break; |
---|
1286 | case ROCMMT_ARGS_ISRESTRICT: |
---|
1287 | kernelArg.isRestrict = parseYAMLBoolValue(ptr, end, lineNo, true); |
---|
1288 | break; |
---|
1289 | case ROCMMT_ARGS_ISVOLATILE: |
---|
1290 | kernelArg.isVolatile = parseYAMLBoolValue(ptr, end, lineNo, true); |
---|
1291 | break; |
---|
1292 | case ROCMMT_ARGS_NAME: |
---|
1293 | kernelArg.name = parseYAMLStringValue(ptr, end, lineNo, level, true); |
---|
1294 | break; |
---|
1295 | case ROCMMT_ARGS_POINTEE_ALIGN: |
---|
1296 | kernelArg.pointeeAlign = |
---|
1297 | parseYAMLIntValue<uint64_t>(ptr, end, lineNo, true); |
---|
1298 | break; |
---|
1299 | case ROCMMT_ARGS_SIZE: |
---|
1300 | kernelArg.size = parseYAMLIntValue<uint64_t>(ptr, end, lineNo); |
---|
1301 | break; |
---|
1302 | case ROCMMT_ARGS_TYPENAME: |
---|
1303 | kernelArg.typeName = |
---|
1304 | parseYAMLStringValue(ptr, end, lineNo, level, true); |
---|
1305 | break; |
---|
1306 | case ROCMMT_ARGS_VALUEKIND: |
---|
1307 | { |
---|
1308 | const std::string vkind = trimStrSpaces(parseYAMLStringValue( |
---|
1309 | ptr, end, lineNo, level, true)); |
---|
1310 | const size_t vkindIndex = binaryMapFind(rocmValueKindNamesMap, |
---|
1311 | rocmValueKindNamesMap + rocmValueKindNamesNum, vkind.c_str(), |
---|
1312 | CStringLess()) - rocmValueKindNamesMap; |
---|
1313 | // if unknown kind |
---|
1314 | if (vkindIndex == rocmValueKindNamesNum) |
---|
1315 | throw ParseException(valLineNo, "Wrong argument value kind"); |
---|
1316 | kernelArg.valueKind = rocmValueKindNamesMap[vkindIndex].second; |
---|
1317 | break; |
---|
1318 | } |
---|
1319 | case ROCMMT_ARGS_VALUETYPE: |
---|
1320 | { |
---|
1321 | const std::string vtype = trimStrSpaces(parseYAMLStringValue( |
---|
1322 | ptr, end, lineNo, level, true)); |
---|
1323 | const size_t vtypeIndex = binaryMapFind(rocmValueTypeNamesMap, |
---|
1324 | rocmValueTypeNamesMap + rocmValueTypeNamesNum, vtype.c_str(), |
---|
1325 | CStringLess()) - rocmValueTypeNamesMap; |
---|
1326 | // if unknown type |
---|
1327 | if (vtypeIndex == rocmValueTypeNamesNum) |
---|
1328 | throw ParseException(valLineNo, "Wrong argument value type"); |
---|
1329 | kernelArg.valueType = rocmValueTypeNamesMap[vtypeIndex].second; |
---|
1330 | break; |
---|
1331 | } |
---|
1332 | default: |
---|
1333 | skipYAMLValue(ptr, end, lineNo, level); |
---|
1334 | break; |
---|
1335 | } |
---|
1336 | } |
---|
1337 | } |
---|
1338 | } |
---|
1339 | |
---|
1340 | void ROCmMetadata::parse(size_t metadataSize, const char* metadata) |
---|
1341 | { |
---|
1342 | parseROCmMetadata(metadataSize, metadata, *this); |
---|
1343 | } |
---|
1344 | |
---|
1345 | /* |
---|
1346 | * ROCm metadata MsgPack parser |
---|
1347 | */ |
---|
1348 | |
---|
1349 | static void parseMsgPackNil(const cxbyte*& dataPtr, const cxbyte* dataEnd) |
---|
1350 | { |
---|
1351 | if (dataPtr>=dataEnd || *dataPtr != 0xc0) |
---|
1352 | throw ParseException("MsgPack: Can't parse nil value"); |
---|
1353 | dataPtr++; |
---|
1354 | } |
---|
1355 | |
---|
1356 | static bool parseMsgPackBool(const cxbyte*& dataPtr, const cxbyte* dataEnd) |
---|
1357 | { |
---|
1358 | if (dataPtr>=dataEnd || ((*dataPtr)&0xfe) != 0xc2) |
---|
1359 | throw ParseException("MsgPack: Can't parse bool value"); |
---|
1360 | const bool v = (*dataPtr==0xc3); |
---|
1361 | dataPtr++; |
---|
1362 | return v; |
---|
1363 | } |
---|
1364 | |
---|
1365 | enum: cxbyte { |
---|
1366 | MSGPACK_WS_UNSIGNED = 0, // only unsigned |
---|
1367 | MSGPACK_WS_SIGNED = 1, // only signed |
---|
1368 | MSGPACK_WS_BOTH = 2 // both signed and unsigned range checking |
---|
1369 | }; |
---|
1370 | |
---|
1371 | |
---|
1372 | static uint64_t parseMsgPackInteger(const cxbyte*& dataPtr, const cxbyte* dataEnd, |
---|
1373 | cxbyte signess = MSGPACK_WS_BOTH) |
---|
1374 | { |
---|
1375 | if (dataPtr>=dataEnd) |
---|
1376 | throw ParseException("MsgPack: Can't parse integer value"); |
---|
1377 | uint64_t v = 0; |
---|
1378 | if (*dataPtr < 0x80) |
---|
1379 | v = *dataPtr++; |
---|
1380 | else if (*dataPtr >= 0xe0) |
---|
1381 | v = uint64_t(-32) + ((*dataPtr++) & 0x1f); |
---|
1382 | else |
---|
1383 | { |
---|
1384 | const cxbyte code = *dataPtr++; |
---|
1385 | switch(code) |
---|
1386 | { |
---|
1387 | case 0xcc: |
---|
1388 | case 0xd0: |
---|
1389 | if (dataPtr>=dataEnd) |
---|
1390 | throw ParseException("MsgPack: Can't parse integer value"); |
---|
1391 | v = *dataPtr++; |
---|
1392 | break; |
---|
1393 | case 0xcd: |
---|
1394 | case 0xd1: |
---|
1395 | if (dataPtr+1>=dataEnd) |
---|
1396 | throw ParseException("MsgPack: Can't parse integer value"); |
---|
1397 | v = *dataPtr++; |
---|
1398 | v |= uint32_t(*dataPtr++)<<8; |
---|
1399 | break; |
---|
1400 | case 0xce: |
---|
1401 | case 0xd2: |
---|
1402 | if (dataPtr+3>=dataEnd) |
---|
1403 | throw ParseException("MsgPack: Can't parse integer value"); |
---|
1404 | for (cxuint i = 0; i < 32; i+=8) |
---|
1405 | v |= uint32_t(*dataPtr++)<<i; |
---|
1406 | break; |
---|
1407 | case 0xcf: |
---|
1408 | case 0xd3: |
---|
1409 | if (dataPtr+7>=dataEnd) |
---|
1410 | throw ParseException("MsgPack: Can't parse integer value"); |
---|
1411 | for (cxuint i = 0; i < 64; i+=8) |
---|
1412 | v |= uint64_t(*dataPtr++)<<i; |
---|
1413 | break; |
---|
1414 | default: |
---|
1415 | throw ParseException("MsgPack: Can't parse integer value"); |
---|
1416 | } |
---|
1417 | |
---|
1418 | if (signess == MSGPACK_WS_UNSIGNED && code >= 0xd0 && v >= (1ULL<<63)) |
---|
1419 | throw ParseException("MsgPack: Negative value for unsigned integer"); |
---|
1420 | if (signess == MSGPACK_WS_SIGNED && code < 0xd0 && v >= (1ULL<<63)) |
---|
1421 | throw ParseException("MsgPack: Positive value out of range for signed integer"); |
---|
1422 | } |
---|
1423 | return v; |
---|
1424 | } |
---|
1425 | |
---|
1426 | static double parseMsgPackFloat(const cxbyte*& dataPtr, const cxbyte* dataEnd) |
---|
1427 | { |
---|
1428 | if (dataPtr>=dataEnd) |
---|
1429 | throw ParseException("MsgPack: Can't parse float value"); |
---|
1430 | const cxbyte code = *dataPtr++; |
---|
1431 | if (code == 0xca) |
---|
1432 | { |
---|
1433 | union { |
---|
1434 | uint32_t v; |
---|
1435 | float vf; |
---|
1436 | } v; |
---|
1437 | v.v = 0; |
---|
1438 | if (dataPtr+3>=dataEnd) |
---|
1439 | throw ParseException("MsgPack: Can't parse float value"); |
---|
1440 | for (cxuint i = 0; i < 32; i+=8) |
---|
1441 | v.v |= uint32_t(*dataPtr++)<<i; |
---|
1442 | return v.vf; |
---|
1443 | } |
---|
1444 | else if (code == 0xcb) |
---|
1445 | { |
---|
1446 | union { |
---|
1447 | uint64_t v; |
---|
1448 | double vf; |
---|
1449 | } v; |
---|
1450 | v.v = 0; |
---|
1451 | if (dataPtr+7>=dataEnd) |
---|
1452 | throw ParseException("MsgPack: Can't parse float value"); |
---|
1453 | for (cxuint i = 0; i < 64; i+=8) |
---|
1454 | v.v |= uint64_t(*dataPtr++)<<i; |
---|
1455 | return v.vf; |
---|
1456 | } |
---|
1457 | else |
---|
1458 | throw ParseException("MsgPack: Can't parse float value"); |
---|
1459 | } |
---|
1460 | |
---|
1461 | static CString parseMsgPackString(const cxbyte*& dataPtr, const cxbyte* dataEnd) |
---|
1462 | { |
---|
1463 | if (dataPtr>=dataEnd) |
---|
1464 | throw ParseException("MsgPack: Can't parse string"); |
---|
1465 | size_t size = 0; |
---|
1466 | |
---|
1467 | if ((*dataPtr&0xe0) == 0xa0) |
---|
1468 | size = (*dataPtr++) & 0x1f; |
---|
1469 | else |
---|
1470 | { |
---|
1471 | const cxbyte code = *dataPtr++; |
---|
1472 | switch (code) |
---|
1473 | { |
---|
1474 | case 0xd9: |
---|
1475 | if (dataPtr>=dataEnd) |
---|
1476 | throw ParseException("MsgPack: Can't parse string size"); |
---|
1477 | size = *dataPtr++; |
---|
1478 | break; |
---|
1479 | case 0xda: |
---|
1480 | if (dataPtr+1>=dataEnd) |
---|
1481 | throw ParseException("MsgPack: Can't parse string size"); |
---|
1482 | size = *dataPtr++; |
---|
1483 | size |= uint32_t(*dataPtr++)<<8; |
---|
1484 | break; |
---|
1485 | case 0xdb: |
---|
1486 | if (dataPtr+3>=dataEnd) |
---|
1487 | throw ParseException("MsgPack: Can't parse string size"); |
---|
1488 | for (cxuint i = 0; i < 32; i+=8) |
---|
1489 | size |= uint32_t(*dataPtr++)<<i; |
---|
1490 | break; |
---|
1491 | default: |
---|
1492 | throw ParseException("MsgPack: Can't parse string"); |
---|
1493 | } |
---|
1494 | } |
---|
1495 | |
---|
1496 | if (dataPtr+size > dataEnd) |
---|
1497 | throw ParseException("MsgPack: Can't parse string"); |
---|
1498 | const char* strData = reinterpret_cast<const char*>(dataPtr); |
---|
1499 | CString out(strData, strData + size); |
---|
1500 | dataPtr += size; |
---|
1501 | return out; |
---|
1502 | } |
---|
1503 | |
---|
1504 | static Array<cxbyte> parseMsgPackData(const cxbyte*& dataPtr, const cxbyte* dataEnd) |
---|
1505 | { |
---|
1506 | if (dataPtr>=dataEnd) |
---|
1507 | throw ParseException("MsgPack: Can't parse byte-array"); |
---|
1508 | const cxbyte code = *dataPtr++; |
---|
1509 | size_t size = 0; |
---|
1510 | switch (code) |
---|
1511 | { |
---|
1512 | case 0xc4: |
---|
1513 | if (dataPtr>=dataEnd) |
---|
1514 | throw ParseException("MsgPack: Can't parse byte-array size"); |
---|
1515 | size = *dataPtr++; |
---|
1516 | break; |
---|
1517 | case 0xc5: |
---|
1518 | if (dataPtr+1>=dataEnd) |
---|
1519 | throw ParseException("MsgPack: Can't parse byte-array size"); |
---|
1520 | size = *dataPtr++; |
---|
1521 | size |= uint32_t(*dataPtr++)<<8; |
---|
1522 | break; |
---|
1523 | case 0xc6: |
---|
1524 | if (dataPtr+3>=dataEnd) |
---|
1525 | throw ParseException("MsgPack: Can't parse byte-array size"); |
---|
1526 | for (cxuint i = 0; i < 32; i+=8) |
---|
1527 | size |= uint32_t(*dataPtr++)<<i; |
---|
1528 | break; |
---|
1529 | default: |
---|
1530 | throw ParseException("MsgPack: Can't parse byte-array"); |
---|
1531 | } |
---|
1532 | |
---|
1533 | if (dataPtr+size > dataEnd) |
---|
1534 | throw ParseException("MsgPack: Can't parse byte-array"); |
---|
1535 | Array<cxbyte> out(dataPtr, dataPtr + size); |
---|
1536 | dataPtr += size; |
---|
1537 | return out; |
---|
1538 | } |
---|
1539 | |
---|
1540 | static void skipMsgPackObject(const cxbyte*& dataPtr, const cxbyte* dataEnd) |
---|
1541 | { |
---|
1542 | if (dataPtr>=dataEnd) |
---|
1543 | throw ParseException("MsgPack: Can't skip object"); |
---|
1544 | if (*dataPtr==0xc0 || *dataPtr==0xc2 || *dataPtr==0xc3 || |
---|
1545 | *dataPtr < 0x80 || *dataPtr >= 0xe0) |
---|
1546 | dataPtr++; |
---|
1547 | else if (*dataPtr==0xcc || *dataPtr==0xd0) |
---|
1548 | { |
---|
1549 | if (dataPtr+1>=dataEnd) |
---|
1550 | throw ParseException("MsgPack: Can't skip object"); |
---|
1551 | dataPtr += 2; |
---|
1552 | } |
---|
1553 | else if (*dataPtr==0xcd || *dataPtr==0xd1) |
---|
1554 | { |
---|
1555 | if (dataPtr+2>=dataEnd) |
---|
1556 | throw ParseException("MsgPack: Can't skip object"); |
---|
1557 | dataPtr += 3; |
---|
1558 | } |
---|
1559 | else if (*dataPtr==0xce || *dataPtr==0xd2 || *dataPtr==0xca) |
---|
1560 | { |
---|
1561 | if (dataPtr+4>=dataEnd) |
---|
1562 | throw ParseException("MsgPack: Can't skip object"); |
---|
1563 | dataPtr += 5; |
---|
1564 | } |
---|
1565 | else if (*dataPtr==0xcf || *dataPtr==0xd3 || *dataPtr==0xcb) |
---|
1566 | { |
---|
1567 | if (dataPtr+8>=dataEnd) |
---|
1568 | throw ParseException("MsgPack: Can't skip object"); |
---|
1569 | dataPtr += 9; |
---|
1570 | } |
---|
1571 | else if(((*dataPtr)&0xe0)==0xa0) |
---|
1572 | { |
---|
1573 | const size_t size = *dataPtr&0x1f; |
---|
1574 | if (dataPtr+size>=dataEnd) |
---|
1575 | throw ParseException("MsgPack: Can't skip object"); |
---|
1576 | dataPtr += size+1; |
---|
1577 | } |
---|
1578 | else if (*dataPtr == 0xc4 || *dataPtr == 0xd9) |
---|
1579 | { |
---|
1580 | dataPtr++; |
---|
1581 | if (dataPtr>=dataEnd) |
---|
1582 | throw ParseException("MsgPack: Can't skip object"); |
---|
1583 | const size_t size = *dataPtr++; |
---|
1584 | if (dataPtr+size>=dataEnd) |
---|
1585 | throw ParseException("MsgPack: Can't skip object"); |
---|
1586 | dataPtr += size; |
---|
1587 | } |
---|
1588 | else if (*dataPtr == 0xc5 || *dataPtr == 0xda) |
---|
1589 | { |
---|
1590 | dataPtr++; |
---|
1591 | if (dataPtr+1>=dataEnd) |
---|
1592 | throw ParseException("MsgPack: Can't skip object"); |
---|
1593 | size_t size = *dataPtr++; |
---|
1594 | size |= (*dataPtr++)<<8; |
---|
1595 | if (dataPtr+size>=dataEnd) |
---|
1596 | throw ParseException("MsgPack: Can't skip object"); |
---|
1597 | dataPtr += size; |
---|
1598 | } |
---|
1599 | else if (*dataPtr == 0xc6 || *dataPtr == 0xdb) |
---|
1600 | { |
---|
1601 | dataPtr++; |
---|
1602 | if (dataPtr+1>=dataEnd) |
---|
1603 | throw ParseException("MsgPack: Can't skip object"); |
---|
1604 | size_t size = 0; |
---|
1605 | for (cxuint i = 0; i < 32; i+=8) |
---|
1606 | size |= (*dataPtr++)<<i; |
---|
1607 | if (dataPtr+size>=dataEnd) |
---|
1608 | throw ParseException("MsgPack: Can't skip object"); |
---|
1609 | dataPtr += size; |
---|
1610 | } |
---|
1611 | else if ((*dataPtr&0xf0) == 0x90 || (*dataPtr&0xf0) == 0x80) |
---|
1612 | { |
---|
1613 | const bool isMap = (*dataPtr<0x90); |
---|
1614 | size_t size = (*dataPtr++)&15; |
---|
1615 | if (isMap) |
---|
1616 | size <<= 1; |
---|
1617 | for (size_t i = 0; i < size; i++) |
---|
1618 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1619 | } |
---|
1620 | else if (*dataPtr == 0xdc || *dataPtr==0xde) |
---|
1621 | { |
---|
1622 | const bool isMap = (*dataPtr==0xde); |
---|
1623 | dataPtr++; |
---|
1624 | if (dataPtr>=dataEnd) |
---|
1625 | throw ParseException("MsgPack: Can't skip object"); |
---|
1626 | size_t size = *dataPtr++; |
---|
1627 | size |= (*dataPtr++)<<8; |
---|
1628 | if (dataPtr+size>=dataEnd) |
---|
1629 | throw ParseException("MsgPack: Can't skip object"); |
---|
1630 | if (isMap) |
---|
1631 | size<<=1; |
---|
1632 | for (size_t i = 0; i < size; i++) |
---|
1633 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1634 | } |
---|
1635 | else if (*dataPtr == 0xdd || *dataPtr==0xdf) |
---|
1636 | { |
---|
1637 | const bool isMap = (*dataPtr==0xdf); |
---|
1638 | dataPtr++; |
---|
1639 | if (dataPtr>=dataEnd) |
---|
1640 | throw ParseException("MsgPack: Can't skip object"); |
---|
1641 | size_t size = 0; |
---|
1642 | for (cxuint i = 0; i < 32; i+=8) |
---|
1643 | size |= (*dataPtr++)<<i; |
---|
1644 | if (dataPtr+size>=dataEnd) |
---|
1645 | throw ParseException("MsgPack: Can't skip object"); |
---|
1646 | if (isMap) |
---|
1647 | size<<=1; |
---|
1648 | for (size_t i = 0; i < size; i++) |
---|
1649 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1650 | } |
---|
1651 | } |
---|
1652 | |
---|
1653 | class CLRX_INTERNAL MsgPackMapParser; |
---|
1654 | |
---|
1655 | class CLRX_INTERNAL MsgPackArrayParser |
---|
1656 | { |
---|
1657 | private: |
---|
1658 | const cxbyte*& dataPtr; |
---|
1659 | const cxbyte* dataEnd; |
---|
1660 | size_t count; |
---|
1661 | void handleErrors(); |
---|
1662 | public: |
---|
1663 | MsgPackArrayParser(const cxbyte*& _dataPtr, const cxbyte* _dataEnd); |
---|
1664 | |
---|
1665 | void parseNil(); |
---|
1666 | bool parseBool(); |
---|
1667 | uint64_t parseInteger(cxbyte signess); |
---|
1668 | double parseFloat(); |
---|
1669 | CString parseString(); |
---|
1670 | Array<cxbyte> parseData(); |
---|
1671 | MsgPackArrayParser parseArray(); |
---|
1672 | MsgPackMapParser parseMap(); |
---|
1673 | size_t end(); // return left elements |
---|
1674 | |
---|
1675 | bool haveElements() const |
---|
1676 | { return count!=0; } |
---|
1677 | }; |
---|
1678 | |
---|
1679 | class CLRX_INTERNAL MsgPackMapParser |
---|
1680 | { |
---|
1681 | private: |
---|
1682 | const cxbyte*& dataPtr; |
---|
1683 | const cxbyte* dataEnd; |
---|
1684 | size_t count; |
---|
1685 | bool keyLeft; |
---|
1686 | void handleErrors(bool key); |
---|
1687 | public: |
---|
1688 | MsgPackMapParser(const cxbyte*& _dataPtr, const cxbyte* _dataEnd); |
---|
1689 | |
---|
1690 | void parseKeyNil(); |
---|
1691 | bool parseKeyBool(); |
---|
1692 | uint64_t parseKeyInteger(cxbyte signess); |
---|
1693 | double parseKeyFloat(); |
---|
1694 | CString parseKeyString(); |
---|
1695 | Array<cxbyte> parseKeyData(); |
---|
1696 | MsgPackArrayParser parseKeyArray(); |
---|
1697 | MsgPackMapParser parseKeyMap(); |
---|
1698 | void parseValueNil(); |
---|
1699 | bool parseValueBool(); |
---|
1700 | uint64_t parseValueInteger(cxbyte signess); |
---|
1701 | double parseValueFloat(); |
---|
1702 | CString parseValueString(); |
---|
1703 | Array<cxbyte> parseValueData(); |
---|
1704 | MsgPackArrayParser parseValueArray(); |
---|
1705 | MsgPackMapParser parseValueMap(); |
---|
1706 | void skipValue(); |
---|
1707 | size_t end(); // return left elements |
---|
1708 | |
---|
1709 | bool haveElements() const |
---|
1710 | { return count!=0; } |
---|
1711 | }; |
---|
1712 | |
---|
1713 | ////////////////// |
---|
1714 | MsgPackArrayParser::MsgPackArrayParser(const cxbyte*& _dataPtr, const cxbyte* _dataEnd) |
---|
1715 | : dataPtr(_dataPtr), dataEnd(_dataEnd), count(0) |
---|
1716 | { |
---|
1717 | if (dataPtr==dataEnd) |
---|
1718 | throw ParseException("MsgPack: Can't parse array of elements"); |
---|
1719 | |
---|
1720 | if (((*dataPtr) & 0xf0) == 0x90) |
---|
1721 | count = (*dataPtr++) & 15; |
---|
1722 | else |
---|
1723 | { |
---|
1724 | const cxbyte code = *dataPtr++; |
---|
1725 | if (code == 0xdc) |
---|
1726 | { |
---|
1727 | if (dataPtr+1 >= dataEnd) |
---|
1728 | throw ParseException("MsgPack: Can't parse array size"); |
---|
1729 | count = *dataPtr++; |
---|
1730 | count |= (*dataPtr++)<<8; |
---|
1731 | } |
---|
1732 | else if (code == 0xdd) |
---|
1733 | { |
---|
1734 | if (dataPtr+3 >= dataEnd) |
---|
1735 | throw ParseException("MsgPack: Can't parse array size"); |
---|
1736 | for (cxuint i = 0; i < 32; i+=8) |
---|
1737 | count |= uint32_t(*dataPtr++)<<i; |
---|
1738 | } |
---|
1739 | else |
---|
1740 | throw ParseException("MsgPack: Can't parse array of elements"); |
---|
1741 | } |
---|
1742 | } |
---|
1743 | |
---|
1744 | void MsgPackArrayParser::handleErrors() |
---|
1745 | { |
---|
1746 | if (count == 0) |
---|
1747 | throw ParseException("MsgPack: No left element to parse"); |
---|
1748 | } |
---|
1749 | |
---|
1750 | void MsgPackArrayParser::parseNil() |
---|
1751 | { |
---|
1752 | handleErrors(); |
---|
1753 | parseMsgPackNil(dataPtr, dataEnd); |
---|
1754 | count--; |
---|
1755 | } |
---|
1756 | |
---|
1757 | bool MsgPackArrayParser::parseBool() |
---|
1758 | { |
---|
1759 | handleErrors(); |
---|
1760 | auto v = parseMsgPackBool(dataPtr, dataEnd); |
---|
1761 | count--; |
---|
1762 | return v; |
---|
1763 | } |
---|
1764 | |
---|
1765 | uint64_t MsgPackArrayParser::parseInteger(cxbyte signess) |
---|
1766 | { |
---|
1767 | handleErrors(); |
---|
1768 | auto v = parseMsgPackInteger(dataPtr, dataEnd, signess); |
---|
1769 | count--; |
---|
1770 | return v; |
---|
1771 | } |
---|
1772 | |
---|
1773 | double MsgPackArrayParser::parseFloat() |
---|
1774 | { |
---|
1775 | handleErrors(); |
---|
1776 | auto v = parseMsgPackFloat(dataPtr, dataEnd); |
---|
1777 | count--; |
---|
1778 | return v; |
---|
1779 | } |
---|
1780 | |
---|
1781 | CString MsgPackArrayParser::parseString() |
---|
1782 | { |
---|
1783 | handleErrors(); |
---|
1784 | auto v = parseMsgPackString(dataPtr, dataEnd); |
---|
1785 | count--; |
---|
1786 | return v; |
---|
1787 | } |
---|
1788 | |
---|
1789 | Array<cxbyte> MsgPackArrayParser::parseData() |
---|
1790 | { |
---|
1791 | handleErrors(); |
---|
1792 | auto v = parseMsgPackData(dataPtr, dataEnd); |
---|
1793 | count--; |
---|
1794 | return v; |
---|
1795 | } |
---|
1796 | |
---|
1797 | MsgPackArrayParser MsgPackArrayParser::parseArray() |
---|
1798 | { |
---|
1799 | handleErrors(); |
---|
1800 | auto v = MsgPackArrayParser(dataPtr, dataEnd); |
---|
1801 | count--; |
---|
1802 | return v; |
---|
1803 | } |
---|
1804 | |
---|
1805 | MsgPackMapParser MsgPackArrayParser::parseMap() |
---|
1806 | { |
---|
1807 | handleErrors(); |
---|
1808 | auto v = MsgPackMapParser(dataPtr, dataEnd); |
---|
1809 | count--; |
---|
1810 | return v; |
---|
1811 | } |
---|
1812 | |
---|
1813 | size_t MsgPackArrayParser::end() |
---|
1814 | { |
---|
1815 | for (size_t i = 0; i < count; i++) |
---|
1816 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1817 | return count; |
---|
1818 | } |
---|
1819 | |
---|
1820 | ////////////////// |
---|
1821 | MsgPackMapParser::MsgPackMapParser(const cxbyte*& _dataPtr, const cxbyte* _dataEnd) |
---|
1822 | : dataPtr(_dataPtr), dataEnd(_dataEnd), count(0), keyLeft(true) |
---|
1823 | { |
---|
1824 | if (dataPtr==dataEnd) |
---|
1825 | throw ParseException("MsgPack: Can't parse map"); |
---|
1826 | |
---|
1827 | if (((*dataPtr) & 0xf0) == 0x80) |
---|
1828 | count = (*dataPtr++) & 15; |
---|
1829 | else |
---|
1830 | { |
---|
1831 | const cxbyte code = *dataPtr++; |
---|
1832 | if (code == 0xde) |
---|
1833 | { |
---|
1834 | if (dataPtr+1 >= dataEnd) |
---|
1835 | throw ParseException("MsgPack: Can't parse map size"); |
---|
1836 | count = *dataPtr++; |
---|
1837 | count |= (*dataPtr++)<<8; |
---|
1838 | } |
---|
1839 | else if (code == 0xdf) |
---|
1840 | { |
---|
1841 | if (dataPtr+3 >= dataEnd) |
---|
1842 | throw ParseException("MsgPack: Can't parse map size"); |
---|
1843 | for (cxuint i = 0; i < 32; i+=8) |
---|
1844 | count |= uint32_t(*dataPtr++)<<i; |
---|
1845 | } |
---|
1846 | else |
---|
1847 | throw ParseException("MsgPack: Can't parse map"); |
---|
1848 | } |
---|
1849 | } |
---|
1850 | |
---|
1851 | void MsgPackMapParser::handleErrors(bool key) |
---|
1852 | { |
---|
1853 | if (count == 0) |
---|
1854 | throw ParseException("MsgPack: No left element to parse"); |
---|
1855 | if (key && !keyLeft) |
---|
1856 | throw ParseException("MsgPack: Key already parsed"); |
---|
1857 | if (!key && keyLeft) |
---|
1858 | throw ParseException("MsgPack: Value already parsed"); |
---|
1859 | } |
---|
1860 | |
---|
1861 | void MsgPackMapParser::parseKeyNil() |
---|
1862 | { |
---|
1863 | handleErrors(true); |
---|
1864 | parseMsgPackNil(dataPtr, dataEnd); |
---|
1865 | keyLeft = false; |
---|
1866 | } |
---|
1867 | |
---|
1868 | bool MsgPackMapParser::parseKeyBool() |
---|
1869 | { |
---|
1870 | handleErrors(true); |
---|
1871 | auto v = parseMsgPackBool(dataPtr, dataEnd); |
---|
1872 | keyLeft = false; |
---|
1873 | return v; |
---|
1874 | } |
---|
1875 | |
---|
1876 | uint64_t MsgPackMapParser::parseKeyInteger(cxbyte signess) |
---|
1877 | { |
---|
1878 | handleErrors(true); |
---|
1879 | auto v = parseMsgPackInteger(dataPtr, dataEnd, signess); |
---|
1880 | keyLeft = false; |
---|
1881 | return v; |
---|
1882 | } |
---|
1883 | |
---|
1884 | CString MsgPackMapParser::parseKeyString() |
---|
1885 | { |
---|
1886 | handleErrors(true); |
---|
1887 | auto v = parseMsgPackString(dataPtr, dataEnd); |
---|
1888 | keyLeft = false; |
---|
1889 | return v; |
---|
1890 | } |
---|
1891 | |
---|
1892 | Array<cxbyte> MsgPackMapParser::parseKeyData() |
---|
1893 | { |
---|
1894 | handleErrors(true); |
---|
1895 | auto v = parseMsgPackData(dataPtr, dataEnd); |
---|
1896 | keyLeft = false; |
---|
1897 | return v; |
---|
1898 | } |
---|
1899 | |
---|
1900 | MsgPackArrayParser MsgPackMapParser::parseKeyArray() |
---|
1901 | { |
---|
1902 | handleErrors(true); |
---|
1903 | auto v = MsgPackArrayParser(dataPtr, dataEnd); |
---|
1904 | keyLeft = false; |
---|
1905 | return v; |
---|
1906 | } |
---|
1907 | |
---|
1908 | MsgPackMapParser MsgPackMapParser::parseKeyMap() |
---|
1909 | { |
---|
1910 | handleErrors(true); |
---|
1911 | auto v = MsgPackMapParser(dataPtr, dataEnd); |
---|
1912 | keyLeft = false; |
---|
1913 | return v; |
---|
1914 | } |
---|
1915 | |
---|
1916 | void MsgPackMapParser::parseValueNil() |
---|
1917 | { |
---|
1918 | handleErrors(false); |
---|
1919 | parseMsgPackNil(dataPtr, dataEnd); |
---|
1920 | keyLeft = true; |
---|
1921 | count--; |
---|
1922 | } |
---|
1923 | |
---|
1924 | bool MsgPackMapParser::parseValueBool() |
---|
1925 | { |
---|
1926 | handleErrors(false); |
---|
1927 | auto v = parseMsgPackBool(dataPtr, dataEnd); |
---|
1928 | keyLeft = true; |
---|
1929 | count--; |
---|
1930 | return v; |
---|
1931 | } |
---|
1932 | |
---|
1933 | uint64_t MsgPackMapParser::parseValueInteger(cxbyte signess) |
---|
1934 | { |
---|
1935 | handleErrors(false); |
---|
1936 | auto v = parseMsgPackInteger(dataPtr, dataEnd, signess); |
---|
1937 | keyLeft = true; |
---|
1938 | count--; |
---|
1939 | return v; |
---|
1940 | } |
---|
1941 | |
---|
1942 | CString MsgPackMapParser::parseValueString() |
---|
1943 | { |
---|
1944 | handleErrors(false); |
---|
1945 | auto v = parseMsgPackString(dataPtr, dataEnd); |
---|
1946 | keyLeft = true; |
---|
1947 | count--; |
---|
1948 | return v; |
---|
1949 | } |
---|
1950 | |
---|
1951 | Array<cxbyte> MsgPackMapParser::parseValueData() |
---|
1952 | { |
---|
1953 | handleErrors(false); |
---|
1954 | auto v = parseMsgPackData(dataPtr, dataEnd); |
---|
1955 | keyLeft = true; |
---|
1956 | count--; |
---|
1957 | return v; |
---|
1958 | } |
---|
1959 | |
---|
1960 | MsgPackArrayParser MsgPackMapParser::parseValueArray() |
---|
1961 | { |
---|
1962 | handleErrors(false); |
---|
1963 | auto v = MsgPackArrayParser(dataPtr, dataEnd); |
---|
1964 | keyLeft = true; |
---|
1965 | count--; |
---|
1966 | return v; |
---|
1967 | } |
---|
1968 | |
---|
1969 | MsgPackMapParser MsgPackMapParser::parseValueMap() |
---|
1970 | { |
---|
1971 | handleErrors(false); |
---|
1972 | auto v = MsgPackMapParser(dataPtr, dataEnd); |
---|
1973 | keyLeft = true; |
---|
1974 | count--; |
---|
1975 | return v; |
---|
1976 | } |
---|
1977 | |
---|
1978 | void MsgPackMapParser::skipValue() |
---|
1979 | { |
---|
1980 | handleErrors(false); |
---|
1981 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1982 | keyLeft = true; |
---|
1983 | count--; |
---|
1984 | } |
---|
1985 | |
---|
1986 | size_t MsgPackMapParser::end() |
---|
1987 | { |
---|
1988 | if (!keyLeft) |
---|
1989 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1990 | for (size_t i = 0; i < count; i++) |
---|
1991 | { |
---|
1992 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1993 | skipMsgPackObject(dataPtr, dataEnd); |
---|
1994 | } |
---|
1995 | return count; |
---|
1996 | } |
---|
1997 | |
---|
1998 | template<typename T> |
---|
1999 | static void parseMsgPackValueTypedArrayForMap(MsgPackMapParser& map, T* out, |
---|
2000 | size_t elemsNum, cxbyte signess) |
---|
2001 | { |
---|
2002 | MsgPackArrayParser arrParser = map.parseValueArray(); |
---|
2003 | for (size_t i = 0; i < elemsNum; i++) |
---|
2004 | out[i] = arrParser.parseInteger(signess); |
---|
2005 | if (arrParser.haveElements()) |
---|
2006 | throw ParseException("Typed Array has too many elements"); |
---|
2007 | } |
---|
2008 | |
---|
2009 | enum { |
---|
2010 | ROCMMP_KERNEL_ARGS = 0, ROCMMP_KERNEL_DEVICE_ENQUEUE_SYMBOL, |
---|
2011 | ROCMMP_KERNEL_GROUP_SEGMENT_FIXED_SIZE, ROCMMP_KERNEL_KERNARG_SEGMENT_ALIGN, |
---|
2012 | ROCMMP_KERNEL_KERNARG_SEGMENT_SIZE, ROCMMP_KERNEL_LANGUAGE, |
---|
2013 | ROCMMP_KERNEL_LANGUAGE_VERSION, ROCMMP_KERNEL_MAX_FLAT_WORKGROUP_SIZE, |
---|
2014 | ROCMMP_KERNEL_NAME, ROCMMP_KERNEL_PRIVATE_SEGMENT_FIXED_SIZE, |
---|
2015 | ROCMMP_KERNEL_REQD_WORKGROUP_SIZE, ROCMMP_KERNEL_SGPR_COUNT, |
---|
2016 | ROCMMP_KERNEL_SGPR_SPILL_COUNT, ROCMMP_KERNEL_SYMBOL, |
---|
2017 | ROCMMP_KERNEL_VEC_TYPE_HINT, ROCMMP_KERNEL_VGPR_COUNT, |
---|
2018 | ROCMMP_KERNEL_VGPR_SPILL_COUNT, ROCMMP_KERNEL_WAVEFRONT_SIZE, |
---|
2019 | ROCMMP_KERNEL_WORKGROUP_SIZE_HINT |
---|
2020 | }; |
---|
2021 | |
---|
2022 | static const char* rocmMetadataMPKernelNames[] = |
---|
2023 | { |
---|
2024 | "args", "device_enqueue_symbol", "group_segment_fixed_size", "kernarg_segment_align", |
---|
2025 | "kernarg_segment_size", "language", "language_version", "max_flat_workgroup_size", |
---|
2026 | "name", "private_segment_fixed_size", "reqd_workgroup_size", "sgpr_count", |
---|
2027 | "sgpr_spill_count", "symbol", "vec_type_hint", "vgpr_count", "vgpr_spill_count", |
---|
2028 | "wavefront_size", "workgroup_size_hint" |
---|
2029 | }; |
---|
2030 | |
---|
2031 | static const size_t rocmMetadataMPKernelNamesSize = sizeof(rocmMetadataMPKernelNames) / |
---|
2032 | sizeof(const char*); |
---|
2033 | |
---|
2034 | static void parseROCmMetadataKernelMsgPack(MsgPackArrayParser& kernelsParser, |
---|
2035 | ROCmKernelMetadata& kernel) |
---|
2036 | { |
---|
2037 | MsgPackMapParser kParser = kernelsParser.parseMap(); |
---|
2038 | while (kParser.haveElements()) |
---|
2039 | { |
---|
2040 | const CString name = kParser.parseKeyString(); |
---|
2041 | const size_t index = binaryFind(rocmMetadataMPKernelNames, |
---|
2042 | rocmMetadataMPKernelNames + rocmMetadataMPKernelNamesSize, |
---|
2043 | name.c_str()) - rocmMetadataMPKernelNames; |
---|
2044 | |
---|
2045 | switch(index) |
---|
2046 | { |
---|
2047 | case ROCMMP_KERNEL_ARGS: |
---|
2048 | break; |
---|
2049 | case ROCMMP_KERNEL_DEVICE_ENQUEUE_SYMBOL: |
---|
2050 | kernel.deviceEnqueueSymbol = kParser.parseValueString(); |
---|
2051 | break; |
---|
2052 | case ROCMMP_KERNEL_GROUP_SEGMENT_FIXED_SIZE: |
---|
2053 | kernel.groupSegmentFixedSize = kParser. |
---|
2054 | parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2055 | break; |
---|
2056 | case ROCMMP_KERNEL_KERNARG_SEGMENT_ALIGN: |
---|
2057 | kernel.kernargSegmentAlign = kParser. |
---|
2058 | parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2059 | break; |
---|
2060 | case ROCMMP_KERNEL_KERNARG_SEGMENT_SIZE: |
---|
2061 | kernel.kernargSegmentSize = kParser. |
---|
2062 | parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2063 | break; |
---|
2064 | case ROCMMP_KERNEL_LANGUAGE: |
---|
2065 | kernel.language = kParser.parseValueString(); |
---|
2066 | break; |
---|
2067 | case ROCMMP_KERNEL_LANGUAGE_VERSION: |
---|
2068 | parseMsgPackValueTypedArrayForMap(kParser, kernel.langVersion, |
---|
2069 | 2, MSGPACK_WS_UNSIGNED); |
---|
2070 | break; |
---|
2071 | case ROCMMP_KERNEL_MAX_FLAT_WORKGROUP_SIZE: |
---|
2072 | kernel.maxFlatWorkGroupSize = kParser. |
---|
2073 | parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2074 | break; |
---|
2075 | case ROCMMP_KERNEL_NAME: |
---|
2076 | kernel.name = kParser.parseValueString(); |
---|
2077 | break; |
---|
2078 | case ROCMMP_KERNEL_PRIVATE_SEGMENT_FIXED_SIZE: |
---|
2079 | kernel.privateSegmentFixedSize = kParser. |
---|
2080 | parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2081 | break; |
---|
2082 | case ROCMMP_KERNEL_REQD_WORKGROUP_SIZE: |
---|
2083 | parseMsgPackValueTypedArrayForMap(kParser, kernel.reqdWorkGroupSize, |
---|
2084 | 3, MSGPACK_WS_UNSIGNED); |
---|
2085 | break; |
---|
2086 | case ROCMMP_KERNEL_SGPR_COUNT: |
---|
2087 | kernel.sgprsNum = kParser.parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2088 | break; |
---|
2089 | case ROCMMP_KERNEL_SGPR_SPILL_COUNT: |
---|
2090 | kernel.spilledSgprs = kParser.parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2091 | break; |
---|
2092 | case ROCMMP_KERNEL_SYMBOL: |
---|
2093 | kernel.symbolName = kParser.parseValueString(); |
---|
2094 | break; |
---|
2095 | case ROCMMP_KERNEL_VEC_TYPE_HINT: |
---|
2096 | kernel.vecTypeHint = kParser.parseValueString(); |
---|
2097 | break; |
---|
2098 | case ROCMMP_KERNEL_VGPR_COUNT: |
---|
2099 | kernel.vgprsNum = kParser.parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2100 | break; |
---|
2101 | case ROCMMP_KERNEL_VGPR_SPILL_COUNT: |
---|
2102 | kernel.spilledVgprs = kParser.parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2103 | break; |
---|
2104 | case ROCMMP_KERNEL_WAVEFRONT_SIZE: |
---|
2105 | kernel.wavefrontSize = kParser.parseValueInteger(MSGPACK_WS_UNSIGNED); |
---|
2106 | break; |
---|
2107 | case ROCMMP_KERNEL_WORKGROUP_SIZE_HINT: |
---|
2108 | parseMsgPackValueTypedArrayForMap(kParser, kernel.workGroupSizeHint, |
---|
2109 | 3, MSGPACK_WS_UNSIGNED); |
---|
2110 | break; |
---|
2111 | default: |
---|
2112 | kParser.skipValue(); |
---|
2113 | break; |
---|
2114 | } |
---|
2115 | } |
---|
2116 | } |
---|
2117 | |
---|
2118 | static void parseROCmMetadataMsgPack(size_t metadataSize, const cxbyte* metadata, |
---|
2119 | ROCmMetadata& metadataInfo) |
---|
2120 | { |
---|
2121 | // init metadata info object |
---|
2122 | metadataInfo.kernels.clear(); |
---|
2123 | metadataInfo.printfInfos.clear(); |
---|
2124 | metadataInfo.version[0] = metadataInfo.version[1] = 0; |
---|
2125 | |
---|
2126 | std::vector<ROCmKernelMetadata>& kernels = metadataInfo.kernels; |
---|
2127 | |
---|
2128 | MsgPackMapParser mainMap(metadata, metadata+metadataSize); |
---|
2129 | while (mainMap.haveElements()) |
---|
2130 | { |
---|
2131 | const CString name = mainMap.parseKeyString(); |
---|
2132 | if (name == "amdhsa.version") |
---|
2133 | parseMsgPackValueTypedArrayForMap(mainMap, metadataInfo.version, |
---|
2134 | 2, MSGPACK_WS_UNSIGNED); |
---|
2135 | else if (name == "amdhsa.kernels") |
---|
2136 | { |
---|
2137 | MsgPackArrayParser kernelsParser = mainMap.parseValueArray(); |
---|
2138 | while (kernelsParser.haveElements()) |
---|
2139 | { |
---|
2140 | ROCmKernelMetadata kernel{}; |
---|
2141 | parseROCmMetadataKernelMsgPack(kernelsParser, kernel); |
---|
2142 | } |
---|
2143 | } |
---|
2144 | else |
---|
2145 | mainMap.skipValue(); |
---|
2146 | } |
---|
2147 | } |
---|
2148 | |
---|
2149 | void ROCmMetadata::parseMsgPack(size_t metadataSize, const cxbyte* metadata) |
---|
2150 | { |
---|
2151 | parseROCmMetadataMsgPack(metadataSize, metadata, *this); |
---|
2152 | } |
---|
2153 | |
---|
2154 | /* |
---|
2155 | * ROCm binary reader and generator |
---|
2156 | */ |
---|
2157 | |
---|
2158 | /* TODO: add support for various kernel code offset (now only 256 is supported) */ |
---|
2159 | |
---|
2160 | ROCmBinary::ROCmBinary(size_t binaryCodeSize, cxbyte* binaryCode, Flags creationFlags) |
---|
2161 | : ElfBinary64(binaryCodeSize, binaryCode, creationFlags), |
---|
2162 | regionsNum(0), codeSize(0), code(nullptr), |
---|
2163 | globalDataSize(0), globalData(nullptr), metadataSize(0), metadata(nullptr), |
---|
2164 | newBinFormat(false) |
---|
2165 | { |
---|
2166 | cxuint textIndex = SHN_UNDEF; |
---|
2167 | try |
---|
2168 | { textIndex = getSectionIndex(".text"); } |
---|
2169 | catch(const Exception& ex) |
---|
2170 | { } // ignore failed |
---|
2171 | uint64_t codeOffset = 0; |
---|
2172 | // find '.text' section |
---|
2173 | if (textIndex!=SHN_UNDEF) |
---|
2174 | { |
---|
2175 | code = getSectionContent(textIndex); |
---|
2176 | const Elf64_Shdr& textShdr = getSectionHeader(textIndex); |
---|
2177 | codeSize = ULEV(textShdr.sh_size); |
---|
2178 | codeOffset = ULEV(textShdr.sh_offset); |
---|
2179 | } |
---|
2180 | |
---|
2181 | cxuint rodataIndex = SHN_UNDEF; |
---|
2182 | try |
---|
2183 | { rodataIndex = getSectionIndex(".rodata"); } |
---|
2184 | catch(const Exception& ex) |
---|
2185 | { } // ignore failed |
---|
2186 | // find '.text' section |
---|
2187 | if (rodataIndex!=SHN_UNDEF) |
---|
2188 | { |
---|
2189 | globalData = getSectionContent(rodataIndex); |
---|
2190 | const Elf64_Shdr& rodataShdr = getSectionHeader(rodataIndex); |
---|
2191 | globalDataSize = ULEV(rodataShdr.sh_size); |
---|
2192 | } |
---|
2193 | |
---|
2194 | cxuint gpuConfigIndex = SHN_UNDEF; |
---|
2195 | try |
---|
2196 | { gpuConfigIndex = getSectionIndex(".AMDGPU.config"); } |
---|
2197 | catch(const Exception& ex) |
---|
2198 | { } // ignore failed |
---|
2199 | newBinFormat = (gpuConfigIndex == SHN_UNDEF); |
---|
2200 | |
---|
2201 | cxuint relaDynIndex = SHN_UNDEF; |
---|
2202 | try |
---|
2203 | { relaDynIndex = getSectionIndex(".rela.dyn"); } |
---|
2204 | catch(const Exception& ex) |
---|
2205 | { } // ignore failed |
---|
2206 | |
---|
2207 | cxuint gotIndex = SHN_UNDEF; |
---|
2208 | try |
---|
2209 | { gotIndex = getSectionIndex(".got"); } |
---|
2210 | catch(const Exception& ex) |
---|
2211 | { } // ignore failed |
---|
2212 | |
---|
2213 | // counts regions (symbol or kernel) |
---|
2214 | regionsNum = 0; |
---|
2215 | const size_t symbolsNum = getSymbolsNum(); |
---|
2216 | for (size_t i = 0; i < symbolsNum; i++) |
---|
2217 | { |
---|
2218 | // count regions number |
---|
2219 | const Elf64_Sym& sym = getSymbol(i); |
---|
2220 | const cxbyte symType = ELF64_ST_TYPE(sym.st_info); |
---|
2221 | const cxbyte bind = ELF64_ST_BIND(sym.st_info); |
---|
2222 | if (ULEV(sym.st_shndx)==textIndex && |
---|
2223 | (symType==STT_GNU_IFUNC || (symType==STT_FUNC && !newBinFormat) || |
---|
2224 | (bind==STB_GLOBAL && symType==STT_OBJECT))) |
---|
2225 | regionsNum++; |
---|
2226 | } |
---|
2227 | if (code==nullptr && regionsNum!=0) |
---|
2228 | throw BinException("No code if regions number is not zero"); |
---|
2229 | regions.reset(new ROCmRegion[regionsNum]); |
---|
2230 | size_t j = 0; |
---|
2231 | typedef std::pair<uint64_t, size_t> RegionOffsetEntry; |
---|
2232 | std::unique_ptr<RegionOffsetEntry[]> symOffsets(new RegionOffsetEntry[regionsNum]); |
---|
2233 | |
---|
2234 | // get regions info |
---|
2235 | for (size_t i = 0; i < symbolsNum; i++) |
---|
2236 | { |
---|
2237 | const Elf64_Sym& sym = getSymbol(i); |
---|
2238 | if (ULEV(sym.st_shndx)!=textIndex) |
---|
2239 | continue; // if not in '.text' section |
---|
2240 | const size_t value = ULEV(sym.st_value); |
---|
2241 | if (value < codeOffset) |
---|
2242 | throw BinException("Region offset is too small!"); |
---|
2243 | const size_t size = ULEV(sym.st_size); |
---|
2244 | |
---|
2245 | const cxbyte symType = ELF64_ST_TYPE(sym.st_info); |
---|
2246 | const cxbyte bind = ELF64_ST_BIND(sym.st_info); |
---|
2247 | if (symType==STT_GNU_IFUNC || symType==STT_FUNC || |
---|
2248 | (bind==STB_GLOBAL && symType==STT_OBJECT)) |
---|
2249 | { |
---|
2250 | ROCmRegionType type = ROCmRegionType::DATA; |
---|
2251 | // if kernel |
---|
2252 | if (symType==STT_GNU_IFUNC) |
---|
2253 | type = ROCmRegionType::KERNEL; |
---|
2254 | // if function kernel |
---|
2255 | else if (symType==STT_FUNC) |
---|
2256 | { |
---|
2257 | if (newBinFormat) |
---|
2258 | continue; |
---|
2259 | type = ROCmRegionType::FKERNEL; |
---|
2260 | } |
---|
2261 | symOffsets[j] = std::make_pair(value, j); |
---|
2262 | if (type!=ROCmRegionType::DATA && value+0x100 > codeOffset+codeSize) |
---|
2263 | throw BinException("Kernel or code offset is too big!"); |
---|
2264 | regions[j++] = { getSymbolName(i), size, value, type }; |
---|
2265 | } |
---|
2266 | } |
---|
2267 | // sort regions by offset |
---|
2268 | std::sort(symOffsets.get(), symOffsets.get()+regionsNum, |
---|
2269 | [](const RegionOffsetEntry& a, const RegionOffsetEntry& b) |
---|
2270 | { return a.first < b.first; }); |
---|
2271 | // checking distance between regions |
---|
2272 | for (size_t i = 1; i <= regionsNum; i++) |
---|
2273 | { |
---|
2274 | size_t end = (i<regionsNum) ? symOffsets[i].first : codeOffset+codeSize; |
---|
2275 | ROCmRegion& region = regions[symOffsets[i-1].second]; |
---|
2276 | if (region.type==ROCmRegionType::KERNEL && symOffsets[i-1].first+0x100 > end) |
---|
2277 | throw BinException("Kernel size is too small!"); |
---|
2278 | |
---|
2279 | const size_t regSize = end - symOffsets[i-1].first; |
---|
2280 | if (region.size==0) |
---|
2281 | region.size = regSize; |
---|
2282 | else |
---|
2283 | region.size = std::min(regSize, region.size); |
---|
2284 | } |
---|
2285 | |
---|
2286 | // load got symbols |
---|
2287 | if (relaDynIndex != SHN_UNDEF && gotIndex != SHN_UNDEF) |
---|
2288 | { |
---|
2289 | const Elf64_Shdr& relaShdr = getSectionHeader(relaDynIndex); |
---|
2290 | const Elf64_Shdr& gotShdr = getSectionHeader(gotIndex); |
---|
2291 | |
---|
2292 | size_t relaEntrySize = ULEV(relaShdr.sh_entsize); |
---|
2293 | if (relaEntrySize==0) |
---|
2294 | relaEntrySize = sizeof(Elf64_Rela); |
---|
2295 | const size_t relaEntriesNum = ULEV(relaShdr.sh_size)/relaEntrySize; |
---|
2296 | const size_t gotEntriesNum = ULEV(gotShdr.sh_size) >> 3; |
---|
2297 | if (gotEntriesNum != relaEntriesNum) |
---|
2298 | throw BinException("RelaDyn entries number and GOT entries " |
---|
2299 | "number doesn't match!"); |
---|
2300 | |
---|
2301 | // initialize GOT symbols table |
---|
2302 | gotSymbols.resize(gotEntriesNum); |
---|
2303 | const cxbyte* relaDyn = getSectionContent(relaDynIndex); |
---|
2304 | for (size_t i = 0; i < relaEntriesNum; i++) |
---|
2305 | { |
---|
2306 | const Elf64_Rela& rela = *reinterpret_cast<const Elf64_Rela*>( |
---|
2307 | relaDyn + relaEntrySize*i); |
---|
2308 | // check rela entry fields |
---|
2309 | if (ULEV(rela.r_offset) != ULEV(gotShdr.sh_offset) + i*8) |
---|
2310 | throw BinException("Wrong dyn relocation offset"); |
---|
2311 | if (ULEV(rela.r_addend) != 0ULL) |
---|
2312 | throw BinException("Wrong dyn relocation addend"); |
---|
2313 | size_t symIndex = ELF64_R_SYM(ULEV(rela.r_info)); |
---|
2314 | if (symIndex >= getDynSymbolsNum()) |
---|
2315 | throw BinException("Dyn relocation symbol index out of range"); |
---|
2316 | // just set in gotSymbols |
---|
2317 | gotSymbols[i] = symIndex; |
---|
2318 | } |
---|
2319 | } |
---|
2320 | |
---|
2321 | // get metadata |
---|
2322 | const size_t notesSize = getNotesSize(); |
---|
2323 | const cxbyte* noteContent = (const cxbyte*)getNotes(); |
---|
2324 | |
---|
2325 | for (size_t offset = 0; offset < notesSize; ) |
---|
2326 | { |
---|
2327 | const Elf64_Nhdr* nhdr = (const Elf64_Nhdr*)(noteContent + offset); |
---|
2328 | size_t namesz = ULEV(nhdr->n_namesz); |
---|
2329 | size_t descsz = ULEV(nhdr->n_descsz); |
---|
2330 | if (usumGt(offset, namesz+descsz, notesSize)) |
---|
2331 | throw BinException("Note offset+size out of range"); |
---|
2332 | |
---|
2333 | if (namesz==4 && |
---|
2334 | ::strcmp((const char*)noteContent+offset+ sizeof(Elf64_Nhdr), "AMD")==0) |
---|
2335 | { |
---|
2336 | const uint32_t noteType = ULEV(nhdr->n_type); |
---|
2337 | if (noteType == 0xa) |
---|
2338 | { |
---|
2339 | metadata = (char*)(noteContent+offset+sizeof(Elf64_Nhdr) + 4); |
---|
2340 | metadataSize = descsz; |
---|
2341 | } |
---|
2342 | else if (noteType == 0xb) |
---|
2343 | target.assign((char*)(noteContent+offset+sizeof(Elf64_Nhdr) + 4), descsz); |
---|
2344 | } |
---|
2345 | size_t align = (((namesz+descsz)&3)!=0) ? 4-((namesz+descsz)&3) : 0; |
---|
2346 | offset += sizeof(Elf64_Nhdr) + namesz + descsz + align; |
---|
2347 | } |
---|
2348 | |
---|
2349 | if (hasRegionMap()) |
---|
2350 | { |
---|
2351 | // create region map |
---|
2352 | regionsMap.resize(regionsNum); |
---|
2353 | for (size_t i = 0; i < regionsNum; i++) |
---|
2354 | regionsMap[i] = std::make_pair(regions[i].regionName, i); |
---|
2355 | // sort region map |
---|
2356 | mapSort(regionsMap.begin(), regionsMap.end()); |
---|
2357 | } |
---|
2358 | |
---|
2359 | if ((creationFlags & ROCMBIN_CREATE_METADATAINFO) != 0 && |
---|
2360 | metadata != nullptr && metadataSize != 0) |
---|
2361 | { |
---|
2362 | metadataInfo.reset(new ROCmMetadata()); |
---|
2363 | parseROCmMetadata(metadataSize, metadata, *metadataInfo); |
---|
2364 | |
---|
2365 | if (hasKernelInfoMap()) |
---|
2366 | { |
---|
2367 | const std::vector<ROCmKernelMetadata>& kernels = metadataInfo->kernels; |
---|
2368 | kernelInfosMap.resize(kernels.size()); |
---|
2369 | for (size_t i = 0; i < kernelInfosMap.size(); i++) |
---|
2370 | kernelInfosMap[i] = std::make_pair(kernels[i].name, i); |
---|
2371 | // sort region map |
---|
2372 | mapSort(kernelInfosMap.begin(), kernelInfosMap.end()); |
---|
2373 | } |
---|
2374 | } |
---|
2375 | } |
---|
2376 | |
---|
2377 | /// determint GPU device from ROCm notes |
---|
2378 | GPUDeviceType ROCmBinary::determineGPUDeviceType(uint32_t& outArchMinor, |
---|
2379 | uint32_t& outArchStepping) const |
---|
2380 | { |
---|
2381 | uint32_t archMajor = 0; |
---|
2382 | uint32_t archMinor = 0; |
---|
2383 | uint32_t archStepping = 0; |
---|
2384 | |
---|
2385 | { |
---|
2386 | const cxbyte* noteContent = (const cxbyte*)getNotes(); |
---|
2387 | if (noteContent==nullptr) |
---|
2388 | throw BinException("Missing notes in inner binary!"); |
---|
2389 | size_t notesSize = getNotesSize(); |
---|
2390 | // find note about AMDGPU |
---|
2391 | for (size_t offset = 0; offset < notesSize; ) |
---|
2392 | { |
---|
2393 | const Elf64_Nhdr* nhdr = (const Elf64_Nhdr*)(noteContent + offset); |
---|
2394 | size_t namesz = ULEV(nhdr->n_namesz); |
---|
2395 | size_t descsz = ULEV(nhdr->n_descsz); |
---|
2396 | if (usumGt(offset, namesz+descsz, notesSize)) |
---|
2397 | throw BinException("Note offset+size out of range"); |
---|
2398 | if (ULEV(nhdr->n_type) == 0x3 && namesz==4 && descsz>=0x1a && |
---|
2399 | ::strcmp((const char*)noteContent+offset+sizeof(Elf64_Nhdr), "AMD")==0) |
---|
2400 | { // AMDGPU type |
---|
2401 | const uint32_t* content = (const uint32_t*) |
---|
2402 | (noteContent+offset+sizeof(Elf64_Nhdr) + 4); |
---|
2403 | archMajor = ULEV(content[1]); |
---|
2404 | archMinor = ULEV(content[2]); |
---|
2405 | archStepping = ULEV(content[3]); |
---|
2406 | } |
---|
2407 | size_t align = (((namesz+descsz)&3)!=0) ? 4-((namesz+descsz)&3) : 0; |
---|
2408 | offset += sizeof(Elf64_Nhdr) + namesz + descsz + align; |
---|
2409 | } |
---|
2410 | } |
---|
2411 | // determine device type |
---|
2412 | GPUDeviceType deviceType = getGPUDeviceTypeFromArchVersion(archMajor, archMinor, |
---|
2413 | archStepping); |
---|
2414 | outArchMinor = archMinor; |
---|
2415 | outArchStepping = archStepping; |
---|
2416 | return deviceType; |
---|
2417 | } |
---|
2418 | |
---|
2419 | const ROCmRegion& ROCmBinary::getRegion(const char* name) const |
---|
2420 | { |
---|
2421 | RegionMap::const_iterator it = binaryMapFind(regionsMap.begin(), |
---|
2422 | regionsMap.end(), name); |
---|
2423 | if (it == regionsMap.end()) |
---|
2424 | throw BinException("Can't find region name"); |
---|
2425 | return regions[it->second]; |
---|
2426 | } |
---|
2427 | |
---|
2428 | const ROCmKernelMetadata& ROCmBinary::getKernelInfo(const char* name) const |
---|
2429 | { |
---|
2430 | if (!hasMetadataInfo()) |
---|
2431 | throw BinException("Can't find kernel info name"); |
---|
2432 | RegionMap::const_iterator it = binaryMapFind(kernelInfosMap.begin(), |
---|
2433 | kernelInfosMap.end(), name); |
---|
2434 | if (it == kernelInfosMap.end()) |
---|
2435 | throw BinException("Can't find kernel info name"); |
---|
2436 | return metadataInfo->kernels[it->second]; |
---|
2437 | } |
---|
2438 | |
---|
2439 | // if ROCm binary |
---|
2440 | bool CLRX::isROCmBinary(size_t binarySize, const cxbyte* binary) |
---|
2441 | { |
---|
2442 | if (!isElfBinary(binarySize, binary)) |
---|
2443 | return false; |
---|
2444 | if (binary[EI_CLASS] != ELFCLASS64) |
---|
2445 | return false; |
---|
2446 | const Elf64_Ehdr* ehdr = reinterpret_cast<const Elf64_Ehdr*>(binary); |
---|
2447 | if (ULEV(ehdr->e_machine) != 0xe0) |
---|
2448 | return false; |
---|
2449 | return true; |
---|
2450 | } |
---|
2451 | |
---|
2452 | |
---|
2453 | void ROCmInput::addEmptyKernel(const char* kernelName) |
---|
2454 | { |
---|
2455 | symbols.push_back({ kernelName, 0, 0, ROCmRegionType::KERNEL }); |
---|
2456 | } |
---|
2457 | |
---|
2458 | /* |
---|
2459 | * ROCm YAML metadata generator |
---|
2460 | */ |
---|
2461 | |
---|
2462 | static const char* rocmValueKindNames[] = |
---|
2463 | { |
---|
2464 | "ByValue", "GlobalBuffer", "DynamicSharedPointer", "Sampler", "Image", "Pipe", "Queue", |
---|
2465 | "HiddenGlobalOffsetX", "HiddenGlobalOffsetY", "HiddenGlobalOffsetZ", "HiddenNone", |
---|
2466 | "HiddenPrintfBuffer", "HiddenDefaultQueue", "HiddenCompletionAction", |
---|
2467 | "HiddenMultiGridSyncArg" |
---|
2468 | }; |
---|
2469 | |
---|
2470 | static const char* rocmValueTypeNames[] = |
---|
2471 | { |
---|
2472 | "Struct", "I8", "U8", "I16", "U16", "F16", "I32", "U32", "F32", "I64", "U64", "F64" |
---|
2473 | }; |
---|
2474 | |
---|
2475 | static void genArrayValue(cxuint n, const cxuint* values, std::string& output) |
---|
2476 | { |
---|
2477 | char numBuf[24]; |
---|
2478 | output += "[ "; |
---|
2479 | for (cxuint i = 0; i < n; i++) |
---|
2480 | { |
---|
2481 | itocstrCStyle(values[i], numBuf, 24); |
---|
2482 | output += numBuf; |
---|
2483 | output += (i+1<n) ? ", " : " ]\n"; |
---|
2484 | } |
---|
2485 | } |
---|
2486 | |
---|
2487 | // helper for checking whether value is supplied |
---|
2488 | static inline bool hasValue(cxuint value) |
---|
2489 | { return value!=BINGEN_NOTSUPPLIED && value!=BINGEN_DEFAULT; } |
---|
2490 | |
---|
2491 | static inline bool hasValue(uint64_t value) |
---|
2492 | { return value!=BINGEN64_NOTSUPPLIED && value!=BINGEN64_DEFAULT; } |
---|
2493 | |
---|
2494 | // get escaped YAML string if needed, otherwise get this same string |
---|
2495 | static std::string escapeYAMLString(const CString& input) |
---|
2496 | { |
---|
2497 | bool toEscape = false; |
---|
2498 | const char* s; |
---|
2499 | for (s = input.c_str(); *s!=0; s++) |
---|
2500 | { |
---|
2501 | cxbyte c = *s; |
---|
2502 | if (c < 0x20 || c >= 0x80 || c=='*' || c=='&' || c=='!' || c=='@' || |
---|
2503 | c=='\'' || c=='\"') |
---|
2504 | toEscape = true; |
---|
2505 | } |
---|
2506 | // if spaces in begin and end |
---|
2507 | if (isSpace(input[0]) || isDigit(input[0]) || |
---|
2508 | (!input.empty() && isSpace(s[-1]))) |
---|
2509 | toEscape = true; |
---|
2510 | |
---|
2511 | if (toEscape) |
---|
2512 | { |
---|
2513 | std::string out = "'"; |
---|
2514 | out += escapeStringCStyle(s-input.c_str(), input.c_str()); |
---|
2515 | out += "'"; |
---|
2516 | return out; |
---|
2517 | } |
---|
2518 | return input.c_str(); |
---|
2519 | } |
---|
2520 | |
---|
2521 | static std::string escapePrintfFormat(const std::string& fmt) |
---|
2522 | { |
---|
2523 | std::string out; |
---|
2524 | out.reserve(fmt.size()); |
---|
2525 | for (char c: fmt) |
---|
2526 | if (c!=':') |
---|
2527 | out.push_back(c); |
---|
2528 | else |
---|
2529 | out += "\\72"; |
---|
2530 | return out; |
---|
2531 | } |
---|
2532 | |
---|
2533 | static void generateROCmMetadata(const ROCmMetadata& mdInfo, |
---|
2534 | const ROCmKernelConfig** kconfigs, std::string& output) |
---|
2535 | { |
---|
2536 | output.clear(); |
---|
2537 | char numBuf[24]; |
---|
2538 | output += "---\n"; |
---|
2539 | // version |
---|
2540 | output += "Version: "; |
---|
2541 | if (hasValue(mdInfo.version[0])) |
---|
2542 | genArrayValue(2, mdInfo.version, output); |
---|
2543 | else // default |
---|
2544 | output += "[ 1, 0 ]\n"; |
---|
2545 | if (!mdInfo.printfInfos.empty()) |
---|
2546 | output += "Printf: \n"; |
---|
2547 | // check print ids uniquness |
---|
2548 | { |
---|
2549 | std::unordered_set<cxuint> printfIds; |
---|
2550 | for (const ROCmPrintfInfo& printfInfo: mdInfo.printfInfos) |
---|
2551 | if (printfInfo.id!=BINGEN_DEFAULT) |
---|
2552 | if (!printfIds.insert(printfInfo.id).second) |
---|
2553 | throw BinGenException("Duplicate of printf id"); |
---|
2554 | // printfs |
---|
2555 | uint32_t freePrintfId = 1; |
---|
2556 | for (const ROCmPrintfInfo& printfInfo: mdInfo.printfInfos) |
---|
2557 | { |
---|
2558 | // skip used printfids; |
---|
2559 | uint32_t printfId = printfInfo.id; |
---|
2560 | if (printfId == BINGEN_DEFAULT) |
---|
2561 | { |
---|
2562 | // skip used printfids |
---|
2563 | for (; printfIds.find(freePrintfId) != printfIds.end(); ++freePrintfId); |
---|
2564 | // just use this free printfid |
---|
2565 | printfId = freePrintfId++; |
---|
2566 | } |
---|
2567 | |
---|
2568 | output += " - '"; |
---|
2569 | itocstrCStyle(printfId, numBuf, 24); |
---|
2570 | output += numBuf; |
---|
2571 | output += ':'; |
---|
2572 | itocstrCStyle(printfInfo.argSizes.size(), numBuf, 24); |
---|
2573 | output += numBuf; |
---|
2574 | output += ':'; |
---|
2575 | for (size_t argSize: printfInfo.argSizes) |
---|
2576 | { |
---|
2577 | itocstrCStyle(argSize, numBuf, 24); |
---|
2578 | output += numBuf; |
---|
2579 | output += ':'; |
---|
2580 | } |
---|
2581 | // printf format |
---|
2582 | std::string escapedFmt = escapeStringCStyle(printfInfo.format); |
---|
2583 | escapedFmt = escapePrintfFormat(escapedFmt); |
---|
2584 | output += escapedFmt; |
---|
2585 | output += "'\n"; |
---|
2586 | } |
---|
2587 | } |
---|
2588 | |
---|
2589 | if (!mdInfo.kernels.empty()) |
---|
2590 | output += "Kernels: \n"; |
---|
2591 | // kernels |
---|
2592 | for (size_t i = 0; i < mdInfo.kernels.size(); i++) |
---|
2593 | { |
---|
2594 | const ROCmKernelMetadata& kernel = mdInfo.kernels[i]; |
---|
2595 | output += " - Name: "; |
---|
2596 | output.append(kernel.name.c_str(), kernel.name.size()); |
---|
2597 | output += "\n SymbolName: "; |
---|
2598 | if (!kernel.symbolName.empty()) |
---|
2599 | output += escapeYAMLString(kernel.symbolName); |
---|
2600 | else |
---|
2601 | { |
---|
2602 | // default is kernel name + '@kd' |
---|
2603 | std::string symName = kernel.name.c_str(); |
---|
2604 | symName += "@kd"; |
---|
2605 | output += escapeYAMLString(symName); |
---|
2606 | } |
---|
2607 | output += "\n"; |
---|
2608 | if (!kernel.language.empty()) |
---|
2609 | { |
---|
2610 | output += " Language: "; |
---|
2611 | output += escapeYAMLString(kernel.language); |
---|
2612 | output += "\n"; |
---|
2613 | } |
---|
2614 | if (kernel.langVersion[0] != BINGEN_NOTSUPPLIED) |
---|
2615 | { |
---|
2616 | output += " LanguageVersion: "; |
---|
2617 | genArrayValue(2, kernel.langVersion, output); |
---|
2618 | } |
---|
2619 | // kernel attributes |
---|
2620 | if (kernel.reqdWorkGroupSize[0] != 0 || kernel.reqdWorkGroupSize[1] != 0 || |
---|
2621 | kernel.reqdWorkGroupSize[2] != 0 || |
---|
2622 | kernel.workGroupSizeHint[0] != 0 || kernel.workGroupSizeHint[1] != 0 || |
---|
2623 | kernel.workGroupSizeHint[2] != 0 || |
---|
2624 | !kernel.vecTypeHint.empty() || !kernel.runtimeHandle.empty()) |
---|
2625 | { |
---|
2626 | output += " Attrs: \n"; |
---|
2627 | if (kernel.workGroupSizeHint[0] != 0 || kernel.workGroupSizeHint[1] != 0 || |
---|
2628 | kernel.workGroupSizeHint[2] != 0) |
---|
2629 | { |
---|
2630 | output += " WorkGroupSizeHint: "; |
---|
2631 | genArrayValue(3, kernel.workGroupSizeHint, output); |
---|
2632 | } |
---|
2633 | if (kernel.reqdWorkGroupSize[0] != 0 || kernel.reqdWorkGroupSize[1] != 0 || |
---|
2634 | kernel.reqdWorkGroupSize[2] != 0) |
---|
2635 | { |
---|
2636 | output += " ReqdWorkGroupSize: "; |
---|
2637 | genArrayValue(3, kernel.reqdWorkGroupSize, output); |
---|
2638 | } |
---|
2639 | if (!kernel.vecTypeHint.empty()) |
---|
2640 | { |
---|
2641 | output += " VecTypeHint: "; |
---|
2642 | output += escapeYAMLString(kernel.vecTypeHint); |
---|
2643 | output += "\n"; |
---|
2644 | } |
---|
2645 | if (!kernel.runtimeHandle.empty()) |
---|
2646 | { |
---|
2647 | output += " RuntimeHandle: "; |
---|
2648 | output += escapeYAMLString(kernel.runtimeHandle); |
---|
2649 | output += "\n"; |
---|
2650 | } |
---|
2651 | } |
---|
2652 | // kernel arguments |
---|
2653 | if (!kernel.argInfos.empty()) |
---|
2654 | output += " Args: \n"; |
---|
2655 | for (const ROCmKernelArgInfo& argInfo: kernel.argInfos) |
---|
2656 | { |
---|
2657 | output += " - "; |
---|
2658 | if (!argInfo.name.empty()) |
---|
2659 | { |
---|
2660 | output += "Name: "; |
---|
2661 | output += escapeYAMLString(argInfo.name); |
---|
2662 | output += "\n "; |
---|
2663 | } |
---|
2664 | if (!argInfo.typeName.empty()) |
---|
2665 | { |
---|
2666 | output += "TypeName: "; |
---|
2667 | output += escapeYAMLString(argInfo.typeName); |
---|
2668 | output += "\n "; |
---|
2669 | } |
---|
2670 | output += "Size: "; |
---|
2671 | itocstrCStyle(argInfo.size, numBuf, 24); |
---|
2672 | output += numBuf; |
---|
2673 | output += "\n Align: "; |
---|
2674 | itocstrCStyle(argInfo.align, numBuf, 24); |
---|
2675 | output += numBuf; |
---|
2676 | output += "\n ValueKind: "; |
---|
2677 | |
---|
2678 | if (argInfo.valueKind > ROCmValueKind::MAX_VALUE) |
---|
2679 | throw BinGenException("Unknown ValueKind"); |
---|
2680 | output += rocmValueKindNames[cxuint(argInfo.valueKind)]; |
---|
2681 | |
---|
2682 | if (argInfo.valueType > ROCmValueType::MAX_VALUE) |
---|
2683 | throw BinGenException("Unknown ValueType"); |
---|
2684 | output += "\n ValueType: "; |
---|
2685 | output += rocmValueTypeNames[cxuint(argInfo.valueType)]; |
---|
2686 | output += "\n"; |
---|
2687 | |
---|
2688 | if (argInfo.valueKind == ROCmValueKind::DYN_SHARED_PTR) |
---|
2689 | { |
---|
2690 | output += " PointeeAlign: "; |
---|
2691 | itocstrCStyle(argInfo.pointeeAlign, numBuf, 24); |
---|
2692 | output += numBuf; |
---|
2693 | output += "\n"; |
---|
2694 | } |
---|
2695 | if (argInfo.valueKind == ROCmValueKind::DYN_SHARED_PTR || |
---|
2696 | argInfo.valueKind == ROCmValueKind::GLOBAL_BUFFER) |
---|
2697 | { |
---|
2698 | if (argInfo.addressSpace > ROCmAddressSpace::MAX_VALUE || |
---|
2699 | argInfo.addressSpace == ROCmAddressSpace::NONE) |
---|
2700 | throw BinGenException("Unknown AddressSpace"); |
---|
2701 | output += " AddrSpaceQual: "; |
---|
2702 | output += rocmAddrSpaceTypesTbl[cxuint(argInfo.addressSpace)-1]; |
---|
2703 | output += "\n"; |
---|
2704 | } |
---|
2705 | if (argInfo.valueKind == ROCmValueKind::IMAGE || |
---|
2706 | argInfo.valueKind == ROCmValueKind::PIPE) |
---|
2707 | { |
---|
2708 | if (argInfo.accessQual> ROCmAccessQual::MAX_VALUE) |
---|
2709 | throw BinGenException("Unknown AccessQualifier"); |
---|
2710 | output += " AccQual: "; |
---|
2711 | output += rocmAccessQualifierTbl[cxuint(argInfo.accessQual)]; |
---|
2712 | output += "\n"; |
---|
2713 | } |
---|
2714 | if (argInfo.valueKind == ROCmValueKind::GLOBAL_BUFFER || |
---|
2715 | argInfo.valueKind == ROCmValueKind::IMAGE || |
---|
2716 | argInfo.valueKind == ROCmValueKind::PIPE) |
---|
2717 | { |
---|
2718 | if (argInfo.actualAccessQual> ROCmAccessQual::MAX_VALUE) |
---|
2719 | throw BinGenException("Unknown ActualAccessQualifier"); |
---|
2720 | output += " ActualAccQual: "; |
---|
2721 | output += rocmAccessQualifierTbl[cxuint(argInfo.actualAccessQual)]; |
---|
2722 | output += "\n"; |
---|
2723 | } |
---|
2724 | if (argInfo.isConst) |
---|
2725 | output += " IsConst: true\n"; |
---|
2726 | if (argInfo.isRestrict) |
---|
2727 | output += " IsRestrict: true\n"; |
---|
2728 | if (argInfo.isVolatile) |
---|
2729 | output += " IsVolatile: true\n"; |
---|
2730 | if (argInfo.isPipe) |
---|
2731 | output += " IsPipe: true\n"; |
---|
2732 | } |
---|
2733 | |
---|
2734 | // kernel code properties |
---|
2735 | const ROCmKernelConfig& kconfig = *kconfigs[i]; |
---|
2736 | |
---|
2737 | output += " CodeProps: \n"; |
---|
2738 | output += " KernargSegmentSize: "; |
---|
2739 | itocstrCStyle(hasValue(kernel.kernargSegmentSize) ? |
---|
2740 | kernel.kernargSegmentSize : ULEV(kconfig.kernargSegmentSize), |
---|
2741 | numBuf, 24); |
---|
2742 | output += numBuf; |
---|
2743 | output += "\n GroupSegmentFixedSize: "; |
---|
2744 | itocstrCStyle(hasValue(kernel.groupSegmentFixedSize) ? |
---|
2745 | kernel.groupSegmentFixedSize : |
---|
2746 | uint64_t(ULEV(kconfig.workgroupGroupSegmentSize)), |
---|
2747 | numBuf, 24); |
---|
2748 | output += numBuf; |
---|
2749 | output += "\n PrivateSegmentFixedSize: "; |
---|
2750 | itocstrCStyle(hasValue(kernel.privateSegmentFixedSize) ? |
---|
2751 | kernel.privateSegmentFixedSize : |
---|
2752 | uint64_t(ULEV(kconfig.workitemPrivateSegmentSize)), |
---|
2753 | numBuf, 24); |
---|
2754 | output += numBuf; |
---|
2755 | output += "\n KernargSegmentAlign: "; |
---|
2756 | itocstrCStyle(hasValue(kernel.kernargSegmentAlign) ? |
---|
2757 | kernel.kernargSegmentAlign : |
---|
2758 | uint64_t(1ULL<<kconfig.kernargSegmentAlignment), |
---|
2759 | numBuf, 24); |
---|
2760 | output += numBuf; |
---|
2761 | output += "\n WavefrontSize: "; |
---|
2762 | itocstrCStyle(hasValue(kernel.wavefrontSize) ? kernel.wavefrontSize : |
---|
2763 | cxuint(1U<<kconfig.wavefrontSize), numBuf, 24); |
---|
2764 | output += numBuf; |
---|
2765 | output += "\n NumSGPRs: "; |
---|
2766 | itocstrCStyle(hasValue(kernel.sgprsNum) ? kernel.sgprsNum : |
---|
2767 | cxuint(ULEV(kconfig.wavefrontSgprCount)), numBuf, 24); |
---|
2768 | output += numBuf; |
---|
2769 | output += "\n NumVGPRs: "; |
---|
2770 | itocstrCStyle(hasValue(kernel.vgprsNum) ? kernel.vgprsNum : |
---|
2771 | cxuint(ULEV(kconfig.workitemVgprCount)), numBuf, 24); |
---|
2772 | output += numBuf; |
---|
2773 | // spilled registers |
---|
2774 | if (hasValue(kernel.spilledSgprs)) |
---|
2775 | { |
---|
2776 | output += "\n NumSpilledSGPRs: "; |
---|
2777 | itocstrCStyle(kernel.spilledSgprs, numBuf, 24); |
---|
2778 | output += numBuf; |
---|
2779 | } |
---|
2780 | if (hasValue(kernel.spilledVgprs)) |
---|
2781 | { |
---|
2782 | output += "\n NumSpilledVGPRs: "; |
---|
2783 | itocstrCStyle(kernel.spilledVgprs, numBuf, 24); |
---|
2784 | output += numBuf; |
---|
2785 | } |
---|
2786 | output += "\n MaxFlatWorkGroupSize: "; |
---|
2787 | itocstrCStyle(hasValue(kernel.maxFlatWorkGroupSize) ? |
---|
2788 | kernel.maxFlatWorkGroupSize : uint64_t(256), numBuf, 24); |
---|
2789 | output += numBuf; |
---|
2790 | output += "\n"; |
---|
2791 | if (kernel.fixedWorkGroupSize[0] != 0 || kernel.fixedWorkGroupSize[1] != 0 || |
---|
2792 | kernel.fixedWorkGroupSize[2] != 0) |
---|
2793 | { |
---|
2794 | output += " FixedWorkGroupSize: "; |
---|
2795 | genArrayValue(3, kernel.fixedWorkGroupSize, output); |
---|
2796 | } |
---|
2797 | } |
---|
2798 | output += "...\n"; |
---|
2799 | } |
---|
2800 | |
---|
2801 | /* ROCm section generators */ |
---|
2802 | |
---|
2803 | class CLRX_INTERNAL ROCmGotGen: public ElfRegionContent |
---|
2804 | { |
---|
2805 | private: |
---|
2806 | const ROCmInput* input; |
---|
2807 | public: |
---|
2808 | explicit ROCmGotGen(const ROCmInput* _input) : input(_input) |
---|
2809 | { } |
---|
2810 | |
---|
2811 | void operator()(FastOutputBuffer& fob) const |
---|
2812 | { |
---|
2813 | fob.fill(input->gotSymbols.size()*8, 0); |
---|
2814 | } |
---|
2815 | }; |
---|
2816 | |
---|
2817 | class CLRX_INTERNAL ROCmRelaDynGen: public ElfRegionContent |
---|
2818 | { |
---|
2819 | private: |
---|
2820 | size_t gotOffset; |
---|
2821 | const ROCmInput* input; |
---|
2822 | public: |
---|
2823 | explicit ROCmRelaDynGen(const ROCmInput* _input) : gotOffset(0), input(_input) |
---|
2824 | { } |
---|
2825 | |
---|
2826 | void setGotOffset(size_t _gotOffset) |
---|
2827 | { gotOffset = _gotOffset; } |
---|
2828 | |
---|
2829 | void operator()(FastOutputBuffer& fob) const |
---|
2830 | { |
---|
2831 | for (size_t i = 0; i < input->gotSymbols.size(); i++) |
---|
2832 | { |
---|
2833 | size_t symIndex = input->gotSymbols[i]; |
---|
2834 | Elf64_Rela rela{}; |
---|
2835 | SLEV(rela.r_offset, gotOffset + 8*i); |
---|
2836 | SLEV(rela.r_info, ELF64_R_INFO(symIndex + 1, 3)); |
---|
2837 | rela.r_addend = 0; |
---|
2838 | fob.writeObject(rela); |
---|
2839 | } |
---|
2840 | } |
---|
2841 | }; |
---|
2842 | |
---|
2843 | /* |
---|
2844 | * ROCm Binary Generator |
---|
2845 | */ |
---|
2846 | |
---|
2847 | ROCmBinGenerator::ROCmBinGenerator() : manageable(false), input(nullptr) |
---|
2848 | { } |
---|
2849 | |
---|
2850 | ROCmBinGenerator::ROCmBinGenerator(const ROCmInput* rocmInput) |
---|
2851 | : manageable(false), input(rocmInput), rocmGotGen(nullptr), rocmRelaDynGen(nullptr) |
---|
2852 | { } |
---|
2853 | |
---|
2854 | ROCmBinGenerator::ROCmBinGenerator(GPUDeviceType deviceType, |
---|
2855 | uint32_t archMinor, uint32_t archStepping, size_t codeSize, const cxbyte* code, |
---|
2856 | size_t globalDataSize, const cxbyte* globalData, |
---|
2857 | const std::vector<ROCmSymbolInput>& symbols) : |
---|
2858 | rocmGotGen(nullptr), rocmRelaDynGen(nullptr) |
---|
2859 | { |
---|
2860 | std::unique_ptr<ROCmInput> _input(new ROCmInput{}); |
---|
2861 | _input->deviceType = deviceType; |
---|
2862 | _input->archMinor = archMinor; |
---|
2863 | _input->archStepping = archStepping; |
---|
2864 | _input->eflags = 0; |
---|
2865 | _input->newBinFormat = false; |
---|
2866 | _input->globalDataSize = globalDataSize; |
---|
2867 | _input->globalData = globalData; |
---|
2868 | _input->symbols = symbols; |
---|
2869 | _input->codeSize = codeSize; |
---|
2870 | _input->code = code; |
---|
2871 | _input->commentSize = 0; |
---|
2872 | _input->comment = nullptr; |
---|
2873 | _input->target = ""; |
---|
2874 | _input->targetTripple = ""; |
---|
2875 | _input->metadataSize = 0; |
---|
2876 | _input->metadata = nullptr; |
---|
2877 | _input->useMetadataInfo = false; |
---|
2878 | _input->metadataInfo = ROCmMetadata{}; |
---|
2879 | input = _input.release(); |
---|
2880 | } |
---|
2881 | |
---|
2882 | ROCmBinGenerator::ROCmBinGenerator(GPUDeviceType deviceType, |
---|
2883 | uint32_t archMinor, uint32_t archStepping, size_t codeSize, const cxbyte* code, |
---|
2884 | size_t globalDataSize, const cxbyte* globalData, |
---|
2885 | std::vector<ROCmSymbolInput>&& symbols) : |
---|
2886 | rocmGotGen(nullptr), rocmRelaDynGen(nullptr) |
---|
2887 | { |
---|
2888 | std::unique_ptr<ROCmInput> _input(new ROCmInput{}); |
---|
2889 | _input->deviceType = deviceType; |
---|
2890 | _input->archMinor = archMinor; |
---|
2891 | _input->archStepping = archStepping; |
---|
2892 | _input->eflags = 0; |
---|
2893 | _input->newBinFormat = false; |
---|
2894 | _input->globalDataSize = globalDataSize; |
---|
2895 | _input->globalData = globalData; |
---|
2896 | _input->symbols = std::move(symbols); |
---|
2897 | _input->codeSize = codeSize; |
---|
2898 | _input->code = code; |
---|
2899 | _input->commentSize = 0; |
---|
2900 | _input->comment = nullptr; |
---|
2901 | _input->target = ""; |
---|
2902 | _input->targetTripple = ""; |
---|
2903 | _input->metadataSize = 0; |
---|
2904 | _input->metadata = nullptr; |
---|
2905 | _input->useMetadataInfo = false; |
---|
2906 | _input->metadataInfo = ROCmMetadata{}; |
---|
2907 | input = _input.release(); |
---|
2908 | } |
---|
2909 | |
---|
2910 | ROCmBinGenerator::~ROCmBinGenerator() |
---|
2911 | { |
---|
2912 | if (manageable) |
---|
2913 | delete input; |
---|
2914 | if (rocmGotGen!=nullptr) |
---|
2915 | delete (ROCmGotGen*)rocmGotGen; |
---|
2916 | if (rocmRelaDynGen!=nullptr) |
---|
2917 | delete (ROCmRelaDynGen*)rocmRelaDynGen; |
---|
2918 | } |
---|
2919 | |
---|
2920 | void ROCmBinGenerator::setInput(const ROCmInput* input) |
---|
2921 | { |
---|
2922 | if (manageable) |
---|
2923 | delete input; |
---|
2924 | manageable = false; |
---|
2925 | this->input = input; |
---|
2926 | } |
---|
2927 | |
---|
2928 | // ELF notes contents |
---|
2929 | static const cxbyte noteDescType1[8] = |
---|
2930 | { 2, 0, 0, 0, 1, 0, 0, 0 }; |
---|
2931 | |
---|
2932 | static const cxbyte noteDescType3[27] = |
---|
2933 | { 4, 0, 7, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
---|
2934 | 'A', 'M', 'D', 0, 'A', 'M', 'D', 'G', 'P', 'U', 0 }; |
---|
2935 | |
---|
2936 | static inline void addMainSectionToTable(cxuint& sectionsNum, uint16_t* builtinTable, |
---|
2937 | cxuint elfSectId) |
---|
2938 | { builtinTable[elfSectId - ELFSECTID_START] = sectionsNum++; } |
---|
2939 | |
---|
2940 | void ROCmBinGenerator::prepareBinaryGen() |
---|
2941 | { |
---|
2942 | AMDGPUArchVersion amdGpuArchValues = getGPUArchVersion(input->deviceType, |
---|
2943 | GPUArchVersionTable::ROCM); |
---|
2944 | if (input->archMinor!=UINT32_MAX) |
---|
2945 | amdGpuArchValues.minor = input->archMinor; |
---|
2946 | if (input->archStepping!=UINT32_MAX) |
---|
2947 | amdGpuArchValues.stepping = input->archStepping; |
---|
2948 | |
---|
2949 | comment = "CLRX ROCmBinGenerator " CLRX_VERSION; |
---|
2950 | commentSize = ::strlen(comment); |
---|
2951 | if (input->comment!=nullptr) |
---|
2952 | { |
---|
2953 | // if comment, store comment section |
---|
2954 | comment = input->comment; |
---|
2955 | commentSize = input->commentSize; |
---|
2956 | if (commentSize==0) |
---|
2957 | commentSize = ::strlen(comment); |
---|
2958 | } |
---|
2959 | |
---|
2960 | uint32_t eflags = input->newBinFormat ? 2 : 0; |
---|
2961 | if (input->eflags != BINGEN_DEFAULT) |
---|
2962 | eflags = input->eflags; |
---|
2963 | |
---|
2964 | std::fill(mainBuiltinSectTable, |
---|
2965 | mainBuiltinSectTable + ROCMSECTID_MAX-ELFSECTID_START+1, SHN_UNDEF); |
---|
2966 | mainSectionsNum = 1; |
---|
2967 | |
---|
2968 | // generate main builtin section table (for section id translation) |
---|
2969 | if (input->newBinFormat) |
---|
2970 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ROCMSECTID_NOTE); |
---|
2971 | if (input->globalData != nullptr) |
---|
2972 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_RODATA); |
---|
2973 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_DYNSYM); |
---|
2974 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ROCMSECTID_HASH); |
---|
2975 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_DYNSTR); |
---|
2976 | if (!input->gotSymbols.empty()) |
---|
2977 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ROCMSECTID_RELADYN); |
---|
2978 | const cxuint execProgHeaderRegionIndex = mainSectionsNum; |
---|
2979 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_TEXT); |
---|
2980 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ROCMSECTID_DYNAMIC); |
---|
2981 | if (!input->gotSymbols.empty()) |
---|
2982 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ROCMSECTID_GOT); |
---|
2983 | if (!input->newBinFormat) |
---|
2984 | { |
---|
2985 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ROCMSECTID_NOTE); |
---|
2986 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ROCMSECTID_GPUCONFIG); |
---|
2987 | } |
---|
2988 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_COMMENT); |
---|
2989 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_SYMTAB); |
---|
2990 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_SHSTRTAB); |
---|
2991 | addMainSectionToTable(mainSectionsNum, mainBuiltinSectTable, ELFSECTID_STRTAB); |
---|
2992 | |
---|
2993 | elfBinGen64.reset(new ElfBinaryGen64({ 0U, 0U, 0x40, 0, ET_DYN, 0xe0, EV_CURRENT, |
---|
2994 | cxuint(input->newBinFormat ? execProgHeaderRegionIndex : UINT_MAX), 0, eflags }, |
---|
2995 | true, true, true, PHREGION_FILESTART)); |
---|
2996 | |
---|
2997 | static const int32_t dynTags[] = { |
---|
2998 | DT_SYMTAB, DT_SYMENT, DT_STRTAB, DT_STRSZ, DT_HASH }; |
---|
2999 | elfBinGen64->addDynamics(sizeof(dynTags)/sizeof(int32_t), dynTags); |
---|
3000 | |
---|
3001 | // elf program headers |
---|
3002 | elfBinGen64->addProgramHeader({ PT_PHDR, PF_R, 0, 1, |
---|
3003 | true, Elf64Types::nobase, Elf64Types::nobase, 0 }); |
---|
3004 | elfBinGen64->addProgramHeader({ PT_LOAD, PF_R, PHREGION_FILESTART, |
---|
3005 | execProgHeaderRegionIndex, |
---|
3006 | true, Elf64Types::nobase, Elf64Types::nobase, 0, 0x1000 }); |
---|
3007 | elfBinGen64->addProgramHeader({ PT_LOAD, PF_R|PF_X, execProgHeaderRegionIndex, 1, |
---|
3008 | true, Elf64Types::nobase, Elf64Types::nobase, 0 }); |
---|
3009 | elfBinGen64->addProgramHeader({ PT_LOAD, PF_R|PF_W, execProgHeaderRegionIndex+1, |
---|
3010 | cxuint(1 + (!input->gotSymbols.empty())), |
---|
3011 | true, Elf64Types::nobase, Elf64Types::nobase, 0 }); |
---|
3012 | elfBinGen64->addProgramHeader({ PT_DYNAMIC, PF_R|PF_W, execProgHeaderRegionIndex+1, 1, |
---|
3013 | true, Elf64Types::nobase, Elf64Types::nobase, 0, 8 }); |
---|
3014 | elfBinGen64->addProgramHeader({ PT_GNU_RELRO, PF_R, execProgHeaderRegionIndex+1, |
---|
3015 | cxuint(1 + (!input->gotSymbols.empty())), |
---|
3016 | true, Elf64Types::nobase, Elf64Types::nobase, 0, 1 }); |
---|
3017 | elfBinGen64->addProgramHeader({ PT_GNU_STACK, PF_R|PF_W, PHREGION_FILESTART, 0, |
---|
3018 | true, 0, 0, 0 }); |
---|
3019 | |
---|
3020 | if (input->newBinFormat) |
---|
3021 | // program header for note (new binary format) |
---|
3022 | elfBinGen64->addProgramHeader({ PT_NOTE, PF_R, 1, 1, true, |
---|
3023 | Elf64Types::nobase, Elf64Types::nobase, 0, 4 }); |
---|
3024 | |
---|
3025 | target = input->target.c_str(); |
---|
3026 | if (target.empty() && !input->targetTripple.empty()) |
---|
3027 | { |
---|
3028 | target = input->targetTripple.c_str(); |
---|
3029 | char dbuf[20]; |
---|
3030 | snprintf(dbuf, 20, "-gfx%u%u%u", amdGpuArchValues.major, amdGpuArchValues.minor, |
---|
3031 | amdGpuArchValues.stepping); |
---|
3032 | target += dbuf; |
---|
3033 | } |
---|
3034 | // elf notes |
---|
3035 | elfBinGen64->addNote({"AMD", sizeof noteDescType1, noteDescType1, 1U}); |
---|
3036 | noteBuf.reset(new cxbyte[0x1b]); |
---|
3037 | ::memcpy(noteBuf.get(), noteDescType3, 0x1b); |
---|
3038 | SULEV(*(uint32_t*)(noteBuf.get()+4), amdGpuArchValues.major); |
---|
3039 | SULEV(*(uint32_t*)(noteBuf.get()+8), amdGpuArchValues.minor); |
---|
3040 | SULEV(*(uint32_t*)(noteBuf.get()+12), amdGpuArchValues.stepping); |
---|
3041 | elfBinGen64->addNote({"AMD", 0x1b, noteBuf.get(), 3U}); |
---|
3042 | if (!target.empty()) |
---|
3043 | elfBinGen64->addNote({"AMD", target.size(), (const cxbyte*)target.c_str(), 0xbU}); |
---|
3044 | |
---|
3045 | metadataSize = input->metadataSize; |
---|
3046 | metadata = input->metadata; |
---|
3047 | if (input->useMetadataInfo) |
---|
3048 | { |
---|
3049 | // generate ROCm metadata |
---|
3050 | std::vector<std::pair<CString, size_t> > symbolIndices(input->symbols.size()); |
---|
3051 | // create sorted indices of symbols by its name |
---|
3052 | for (size_t k = 0; k < input->symbols.size(); k++) |
---|
3053 | symbolIndices[k] = std::make_pair(input->symbols[k].symbolName, k); |
---|
3054 | mapSort(symbolIndices.begin(), symbolIndices.end()); |
---|
3055 | |
---|
3056 | const size_t mdKernelsNum = input->metadataInfo.kernels.size(); |
---|
3057 | std::unique_ptr<const ROCmKernelConfig*[]> kernelConfigPtrs( |
---|
3058 | new const ROCmKernelConfig*[mdKernelsNum]); |
---|
3059 | // generate ROCm kernel config pointers |
---|
3060 | for (size_t k = 0; k < mdKernelsNum; k++) |
---|
3061 | { |
---|
3062 | auto it = binaryMapFind(symbolIndices.begin(), symbolIndices.end(), |
---|
3063 | input->metadataInfo.kernels[k].name); |
---|
3064 | if (it == symbolIndices.end() || |
---|
3065 | (input->symbols[it->second].type != ROCmRegionType::FKERNEL && |
---|
3066 | input->symbols[it->second].type != ROCmRegionType::KERNEL)) |
---|
3067 | throw BinGenException("Kernel in metadata doesn't exists in code"); |
---|
3068 | kernelConfigPtrs[k] = reinterpret_cast<const ROCmKernelConfig*>( |
---|
3069 | input->code + input->symbols[it->second].offset); |
---|
3070 | } |
---|
3071 | // just generate ROCm metadata from info |
---|
3072 | generateROCmMetadata(input->metadataInfo, kernelConfigPtrs.get(), metadataStr); |
---|
3073 | metadataSize = metadataStr.size(); |
---|
3074 | metadata = metadataStr.c_str(); |
---|
3075 | } |
---|
3076 | |
---|
3077 | if (metadataSize != 0) |
---|
3078 | elfBinGen64->addNote({"AMD", metadataSize, (const cxbyte*)metadata, 0xaU}); |
---|
3079 | |
---|
3080 | /// region and sections |
---|
3081 | elfBinGen64->addRegion(ElfRegion64::programHeaderTable()); |
---|
3082 | if (input->newBinFormat) |
---|
3083 | elfBinGen64->addRegion(ElfRegion64::noteSection()); |
---|
3084 | if (input->globalData != nullptr) |
---|
3085 | elfBinGen64->addRegion(ElfRegion64(input->globalDataSize, input->globalData, 4, |
---|
3086 | ".rodata", SHT_PROGBITS, SHF_ALLOC, 0, 0, Elf64Types::nobase)); |
---|
3087 | |
---|
3088 | elfBinGen64->addRegion(ElfRegion64(0, (const cxbyte*)nullptr, 8, |
---|
3089 | ".dynsym", SHT_DYNSYM, SHF_ALLOC, 0, BINGEN_DEFAULT, Elf64Types::nobase)); |
---|
3090 | elfBinGen64->addRegion(ElfRegion64(0, (const cxbyte*)nullptr, 4, |
---|
3091 | ".hash", SHT_HASH, SHF_ALLOC, |
---|
3092 | mainBuiltinSectTable[ELFSECTID_DYNSYM-ELFSECTID_START], 0, |
---|
3093 | Elf64Types::nobase)); |
---|
3094 | elfBinGen64->addRegion(ElfRegion64(0, (const cxbyte*)nullptr, 1, ".dynstr", SHT_STRTAB, |
---|
3095 | SHF_ALLOC, 0, 0, Elf64Types::nobase)); |
---|
3096 | if (!input->gotSymbols.empty()) |
---|
3097 | { |
---|
3098 | ROCmRelaDynGen* sgen = new ROCmRelaDynGen(input); |
---|
3099 | rocmRelaDynGen = (void*)sgen; |
---|
3100 | elfBinGen64->addRegion(ElfRegion64(input->gotSymbols.size()*sizeof(Elf64_Rela), |
---|
3101 | sgen, 8, ".rela.dyn", SHT_RELA, SHF_ALLOC, |
---|
3102 | mainBuiltinSectTable[ELFSECTID_DYNSYM-ELFSECTID_START], 0, |
---|
3103 | Elf64Types::nobase, sizeof(Elf64_Rela))); |
---|
3104 | } |
---|
3105 | // '.text' with alignment=4096 |
---|
3106 | elfBinGen64->addRegion(ElfRegion64(input->codeSize, (const cxbyte*)input->code, |
---|
3107 | 0x1000, ".text", SHT_PROGBITS, SHF_ALLOC|SHF_EXECINSTR, 0, 0, |
---|
3108 | Elf64Types::nobase, 0, false, 256)); |
---|
3109 | elfBinGen64->addRegion(ElfRegion64(0, (const cxbyte*)nullptr, 0x1000, |
---|
3110 | ".dynamic", SHT_DYNAMIC, SHF_ALLOC|SHF_WRITE, |
---|
3111 | mainBuiltinSectTable[ELFSECTID_DYNSTR-ELFSECTID_START], 0, |
---|
3112 | Elf64Types::nobase, 0, false, 8)); |
---|
3113 | if (!input->gotSymbols.empty()) |
---|
3114 | { |
---|
3115 | ROCmGotGen* sgen = new ROCmGotGen(input); |
---|
3116 | rocmGotGen = (void*)sgen; |
---|
3117 | elfBinGen64->addRegion(ElfRegion64(input->gotSymbols.size()*8, sgen, |
---|
3118 | 8, ".got", SHT_PROGBITS, |
---|
3119 | SHF_ALLOC|SHF_WRITE, 0, 0, Elf64Types::nobase)); |
---|
3120 | } |
---|
3121 | if (!input->newBinFormat) |
---|
3122 | { |
---|
3123 | elfBinGen64->addRegion(ElfRegion64::noteSection()); |
---|
3124 | elfBinGen64->addRegion(ElfRegion64(0, (const cxbyte*)nullptr, 1, |
---|
3125 | ".AMDGPU.config", SHT_PROGBITS, 0)); |
---|
3126 | } |
---|
3127 | elfBinGen64->addRegion(ElfRegion64(commentSize, (const cxbyte*)comment, 1, ".comment", |
---|
3128 | SHT_PROGBITS, SHF_MERGE|SHF_STRINGS, 0, 0, 0, 1)); |
---|
3129 | elfBinGen64->addRegion(ElfRegion64(0, (const cxbyte*)nullptr, 8, |
---|
3130 | ".symtab", SHT_SYMTAB, 0, 0, BINGEN_DEFAULT)); |
---|
3131 | elfBinGen64->addRegion(ElfRegion64::shstrtabSection()); |
---|
3132 | elfBinGen64->addRegion(ElfRegion64::strtabSection()); |
---|
3133 | elfBinGen64->addRegion(ElfRegion64::sectionHeaderTable()); |
---|
3134 | |
---|
3135 | /* extra sections */ |
---|
3136 | for (const BinSection& section: input->extraSections) |
---|
3137 | elfBinGen64->addRegion(ElfRegion64(section, mainBuiltinSectTable, |
---|
3138 | ROCMSECTID_MAX, mainSectionsNum)); |
---|
3139 | updateSymbols(); |
---|
3140 | binarySize = elfBinGen64->countSize(); |
---|
3141 | |
---|
3142 | if (rocmRelaDynGen != nullptr) |
---|
3143 | ((ROCmRelaDynGen*)rocmRelaDynGen)->setGotOffset( |
---|
3144 | elfBinGen64->getRegionOffset( |
---|
3145 | mainBuiltinSectTable[ROCMSECTID_GOT - ELFSECTID_START])); |
---|
3146 | } |
---|
3147 | |
---|
3148 | void ROCmBinGenerator::updateSymbols() |
---|
3149 | { |
---|
3150 | elfBinGen64->clearSymbols(); |
---|
3151 | elfBinGen64->clearDynSymbols(); |
---|
3152 | // add symbols (kernels, function kernels and data symbols) |
---|
3153 | elfBinGen64->addSymbol(ElfSymbol64("_DYNAMIC", |
---|
3154 | mainBuiltinSectTable[ROCMSECTID_DYNAMIC-ELFSECTID_START], |
---|
3155 | ELF64_ST_INFO(STB_LOCAL, STT_NOTYPE), STV_HIDDEN, true, 0, 0)); |
---|
3156 | const uint16_t textSectIndex = mainBuiltinSectTable[ELFSECTID_TEXT-ELFSECTID_START]; |
---|
3157 | for (const ROCmSymbolInput& symbol: input->symbols) |
---|
3158 | { |
---|
3159 | ElfSymbol64 elfsym; |
---|
3160 | switch (symbol.type) |
---|
3161 | { |
---|
3162 | case ROCmRegionType::KERNEL: |
---|
3163 | elfsym = ElfSymbol64(symbol.symbolName.c_str(), textSectIndex, |
---|
3164 | ELF64_ST_INFO(STB_GLOBAL, STT_GNU_IFUNC), 0, true, |
---|
3165 | symbol.offset, symbol.size); |
---|
3166 | break; |
---|
3167 | case ROCmRegionType::FKERNEL: |
---|
3168 | elfsym = ElfSymbol64(symbol.symbolName.c_str(), textSectIndex, |
---|
3169 | ELF64_ST_INFO(STB_GLOBAL, STT_FUNC), 0, true, |
---|
3170 | symbol.offset, symbol.size); |
---|
3171 | break; |
---|
3172 | case ROCmRegionType::DATA: |
---|
3173 | elfsym = ElfSymbol64(symbol.symbolName.c_str(), textSectIndex, |
---|
3174 | ELF64_ST_INFO(STB_GLOBAL, STT_OBJECT), 0, true, |
---|
3175 | symbol.offset, symbol.size); |
---|
3176 | break; |
---|
3177 | default: |
---|
3178 | break; |
---|
3179 | } |
---|
3180 | // add to symbols and dynamic symbols table |
---|
3181 | elfBinGen64->addSymbol(elfsym); |
---|
3182 | elfBinGen64->addDynSymbol(elfsym); |
---|
3183 | } |
---|
3184 | /* extra symbols */ |
---|
3185 | for (const BinSymbol& symbol: input->extraSymbols) |
---|
3186 | { |
---|
3187 | ElfSymbol64 sym(symbol, mainBuiltinSectTable, |
---|
3188 | ROCMSECTID_MAX, mainSectionsNum); |
---|
3189 | elfBinGen64->addSymbol(sym); |
---|
3190 | elfBinGen64->addDynSymbol(sym); |
---|
3191 | } |
---|
3192 | } |
---|
3193 | |
---|
3194 | void ROCmBinGenerator::generateInternal(std::ostream* osPtr, std::vector<char>* vPtr, |
---|
3195 | Array<cxbyte>* aPtr) |
---|
3196 | { |
---|
3197 | if (elfBinGen64 == nullptr) |
---|
3198 | prepareBinaryGen(); |
---|
3199 | /**** |
---|
3200 | * prepare for write binary to output |
---|
3201 | ****/ |
---|
3202 | std::unique_ptr<std::ostream> outStreamHolder; |
---|
3203 | std::ostream* os = nullptr; |
---|
3204 | if (aPtr != nullptr) |
---|
3205 | { |
---|
3206 | aPtr->resize(binarySize); |
---|
3207 | outStreamHolder.reset( |
---|
3208 | new ArrayOStream(binarySize, reinterpret_cast<char*>(aPtr->data()))); |
---|
3209 | os = outStreamHolder.get(); |
---|
3210 | } |
---|
3211 | else if (vPtr != nullptr) |
---|
3212 | { |
---|
3213 | vPtr->resize(binarySize); |
---|
3214 | outStreamHolder.reset(new VectorOStream(*vPtr)); |
---|
3215 | os = outStreamHolder.get(); |
---|
3216 | } |
---|
3217 | else // from argument |
---|
3218 | os = osPtr; |
---|
3219 | |
---|
3220 | const std::ios::iostate oldExceptions = os->exceptions(); |
---|
3221 | try |
---|
3222 | { |
---|
3223 | os->exceptions(std::ios::failbit | std::ios::badbit); |
---|
3224 | /**** |
---|
3225 | * write binary to output |
---|
3226 | ****/ |
---|
3227 | FastOutputBuffer bos(256, *os); |
---|
3228 | elfBinGen64->generate(bos); |
---|
3229 | assert(bos.getWritten() == binarySize); |
---|
3230 | |
---|
3231 | if (rocmGotGen != nullptr) |
---|
3232 | { |
---|
3233 | delete (ROCmGotGen*)rocmGotGen; |
---|
3234 | rocmGotGen = nullptr; |
---|
3235 | } |
---|
3236 | if (rocmRelaDynGen != nullptr) |
---|
3237 | { |
---|
3238 | delete (ROCmGotGen*)rocmRelaDynGen; |
---|
3239 | rocmRelaDynGen = nullptr; |
---|
3240 | } |
---|
3241 | } |
---|
3242 | catch(...) |
---|
3243 | { |
---|
3244 | os->exceptions(oldExceptions); |
---|
3245 | throw; |
---|
3246 | } |
---|
3247 | os->exceptions(oldExceptions); |
---|
3248 | } |
---|
3249 | |
---|
3250 | void ROCmBinGenerator::generate(Array<cxbyte>& array) |
---|
3251 | { |
---|
3252 | generateInternal(nullptr, nullptr, &array); |
---|
3253 | } |
---|
3254 | |
---|
3255 | void ROCmBinGenerator::generate(std::ostream& os) |
---|
3256 | { |
---|
3257 | generateInternal(&os, nullptr, nullptr); |
---|
3258 | } |
---|
3259 | |
---|
3260 | void ROCmBinGenerator::generate(std::vector<char>& v) |
---|
3261 | { |
---|
3262 | generateInternal(nullptr, &v, nullptr); |
---|
3263 | } |
---|