FullyConnected.cpp 13 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438
  1. #include "FullyConnected.h"
  2. #include <stdio.h>
  3. #include "Error.h"
  4. #include "arm_nnfunctions.h"
  5. #include "Test.h"
  6. #include "stdio.h"
  7. void printPattern(char *s,Client::AnyPattern<q7_t> pat)
  8. {
  9. q7_t *p=pat.ptr();
  10. printf("%s\n",s);
  11. for(int i=0;i < pat.nbSamples(); i++)
  12. {
  13. printf("0x%02x\n",p[i]);
  14. }
  15. printf("----\n");
  16. }
  17. void FullyConnected::test_fully_connected_tflite_s8()
  18. {
  19. q7_t *inp=input.ptr();
  20. q31_t *biasp=bias.ptr();
  21. q7_t *weightp=weight.ptr();
  22. q7_t *outp=output.ptr();
  23. q7_t *refp=ref.ptr();
  24. q15_t *tempp=temp.ptr();
  25. arm_fully_connected_s8((int8_t*)inp
  26. ,(const int8_t*)weightp
  27. ,colDim
  28. ,rowDim
  29. ,nb_batches
  30. ,input_offset
  31. ,filter_offset
  32. ,output_mult
  33. ,output_shift
  34. ,output_offset
  35. ,(const int32_t*)biasp
  36. ,(int8_t*)outp
  37. ,act_min
  38. ,act_max
  39. ,tempp
  40. );
  41. ASSERT_EQ(ref,output);
  42. }
  43. void FullyConnected::setUp(Testing::testID_t id,std::vector<Testing::param_t>& params,Client::PatternMgr *mgr)
  44. {
  45. nb_batches = 1;
  46. switch(id)
  47. {
  48. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_1:
  49. output_mult = 1073741824;
  50. output_shift = -1;
  51. filter_offset = 1;
  52. input_offset = 1;
  53. output_offset = -1;
  54. act_min =-128;
  55. act_max= 127;
  56. input.reload(FullyConnected::INPUT1_S8_ID,mgr);
  57. bias.reload(FullyConnected::BIAS1_S8_ID,mgr);
  58. weight.reload(FullyConnected::WEIGHT1_S8_ID,mgr);
  59. ref.reload(FullyConnected::REF1_S8_ID,mgr);
  60. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  61. temp.create(input.nbSamples(),FullyConnected::TEMP_S16_ID,mgr);
  62. colDim=input.nbSamples();
  63. rowDim=output.nbSamples();
  64. break;
  65. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_2:
  66. output_mult = 1073741824;
  67. output_shift = 1;
  68. filter_offset = 1;
  69. input_offset = 1;
  70. output_offset = -1;
  71. act_min =-128;
  72. act_max= 127;
  73. input.reload(FullyConnected::INPUT2_S8_ID,mgr);
  74. bias.reload(FullyConnected::BIAS2_S8_ID,mgr);
  75. weight.reload(FullyConnected::WEIGHT2_S8_ID,mgr);
  76. ref.reload(FullyConnected::REF2_S8_ID,mgr);
  77. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  78. temp.create(input.nbSamples(),FullyConnected::TEMP_S16_ID,mgr);
  79. colDim=input.nbSamples();
  80. rowDim=output.nbSamples();
  81. break;
  82. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_3:
  83. output_mult = 1073741824;
  84. output_shift = 2;
  85. filter_offset = 1;
  86. input_offset = 1;
  87. output_offset = -1;
  88. act_min =-1;
  89. act_max= 127;
  90. input.reload(FullyConnected::INPUT3_S8_ID,mgr);
  91. bias.reload(FullyConnected::BIAS3_S8_ID,mgr);
  92. weight.reload(FullyConnected::WEIGHT3_S8_ID,mgr);
  93. ref.reload(FullyConnected::REF3_S8_ID,mgr);
  94. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  95. temp.create(input.nbSamples(),FullyConnected::TEMP_S16_ID,mgr);
  96. colDim=input.nbSamples();
  97. rowDim=output.nbSamples();
  98. break;
  99. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_4:
  100. output_mult = 1073741824;
  101. output_shift = 1;
  102. filter_offset = 1;
  103. input_offset = 1;
  104. output_offset = -1;
  105. act_min =-128;
  106. act_max= 127;
  107. input.reload(FullyConnected::INPUT4_S8_ID,mgr);
  108. bias.reload(FullyConnected::BIAS4_S8_ID,mgr);
  109. weight.reload(FullyConnected::WEIGHT4_S8_ID,mgr);
  110. ref.reload(FullyConnected::REF4_S8_ID,mgr);
  111. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  112. temp.create(input.nbSamples(),FullyConnected::TEMP_S16_ID,mgr);
  113. colDim=input.nbSamples();
  114. rowDim=output.nbSamples();
  115. break;
  116. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_5:
  117. output_mult = 1073741824;
  118. output_shift = 1;
  119. filter_offset = 1;
  120. input_offset = 1;
  121. output_offset = -1;
  122. act_min =-128;
  123. act_max= 127;
  124. input.reload(FullyConnected::INPUT5_S8_ID,mgr);
  125. bias.reload(FullyConnected::BIAS5_S8_ID,mgr);
  126. weight.reload(FullyConnected::WEIGHT5_S8_ID,mgr);
  127. ref.reload(FullyConnected::REF5_S8_ID,mgr);
  128. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  129. temp.create(input.nbSamples(),FullyConnected::TEMP_S16_ID,mgr);
  130. colDim=input.nbSamples();
  131. rowDim=output.nbSamples();
  132. break;
  133. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_6:
  134. output_mult = 1073741824;
  135. output_shift = -1;
  136. filter_offset = 1;
  137. input_offset = 1;
  138. output_offset = -1;
  139. act_min =-128;
  140. act_max= 127;
  141. nb_batches=9;
  142. colDim=6;
  143. rowDim=1;
  144. input.reload(FullyConnected::INPUT6_S8_ID,mgr);
  145. bias.reload(FullyConnected::BIAS6_S8_ID,mgr);
  146. weight.reload(FullyConnected::WEIGHT6_S8_ID,mgr);
  147. ref.reload(FullyConnected::REF6_S8_ID,mgr);
  148. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  149. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  150. break;
  151. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_7:
  152. output_mult = 1073741824;
  153. output_shift = -1;
  154. filter_offset = 1;
  155. input_offset = 1;
  156. output_offset = -1;
  157. act_min =-128;
  158. act_max= 127;
  159. nb_batches=8;
  160. colDim=8;
  161. rowDim=1;
  162. input.reload(FullyConnected::INPUT7_S8_ID,mgr);
  163. bias.reload(FullyConnected::BIAS7_S8_ID,mgr);
  164. weight.reload(FullyConnected::WEIGHT7_S8_ID,mgr);
  165. ref.reload(FullyConnected::REF7_S8_ID,mgr);
  166. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  167. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  168. break;
  169. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_8:
  170. output_mult = 1073741824;
  171. output_shift = -1;
  172. filter_offset = 1;
  173. input_offset = 1;
  174. output_offset = -1;
  175. act_min =-128;
  176. act_max= 127;
  177. nb_batches=4;
  178. colDim=10;
  179. rowDim=1;
  180. input.reload(FullyConnected::INPUT8_S8_ID,mgr);
  181. bias.reload(FullyConnected::BIAS8_S8_ID,mgr);
  182. weight.reload(FullyConnected::WEIGHT8_S8_ID,mgr);
  183. ref.reload(FullyConnected::REF8_S8_ID,mgr);
  184. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  185. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  186. break;
  187. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_9:
  188. output_mult = 1073741824;
  189. output_shift = -1;
  190. filter_offset = 1;
  191. input_offset = 1;
  192. output_offset = -1;
  193. act_min =-128;
  194. act_max= 127;
  195. nb_batches=9;
  196. colDim=6;
  197. rowDim=1;
  198. input.reload(FullyConnected::INPUT9_S8_ID,mgr);
  199. bias.reload(FullyConnected::BIAS9_S8_ID,mgr);
  200. weight.reload(FullyConnected::WEIGHT9_S8_ID,mgr);
  201. ref.reload(FullyConnected::REF9_S8_ID,mgr);
  202. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  203. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  204. break;
  205. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_10:
  206. output_mult = 1073741824;
  207. output_shift = -1;
  208. filter_offset = 1;
  209. input_offset = 1;
  210. output_offset = -1;
  211. act_min =-128;
  212. act_max= 127;
  213. nb_batches=4;
  214. colDim=10;
  215. rowDim=1;
  216. input.reload(FullyConnected::INPUT10_S8_ID,mgr);
  217. bias.reload(FullyConnected::BIAS10_S8_ID,mgr);
  218. weight.reload(FullyConnected::WEIGHT10_S8_ID,mgr);
  219. ref.reload(FullyConnected::REF10_S8_ID,mgr);
  220. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  221. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  222. break;
  223. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_11:
  224. output_mult = 1073741824;
  225. output_shift = -1;
  226. filter_offset = 1;
  227. input_offset = 1;
  228. output_offset = -1;
  229. act_min =-128;
  230. act_max= 127;
  231. nb_batches=8;
  232. colDim=8;
  233. rowDim=1;
  234. input.reload(FullyConnected::INPUT11_S8_ID,mgr);
  235. bias.reload(FullyConnected::BIAS11_S8_ID,mgr);
  236. weight.reload(FullyConnected::WEIGHT11_S8_ID,mgr);
  237. ref.reload(FullyConnected::REF11_S8_ID,mgr);
  238. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  239. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  240. break;
  241. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_12:
  242. output_mult = 1073741824;
  243. output_shift = 1;
  244. filter_offset = 0;
  245. input_offset = 0;
  246. output_offset = 0;
  247. act_min =-128;
  248. act_max= 127;
  249. nb_batches=9;
  250. colDim=8;
  251. rowDim=4;
  252. input.reload(FullyConnected::INPUT12_S8_ID,mgr);
  253. bias.reload(FullyConnected::BIAS12_S8_ID,mgr);
  254. weight.reload(FullyConnected::WEIGHT12_S8_ID,mgr);
  255. ref.reload(FullyConnected::REF12_S8_ID,mgr);
  256. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  257. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  258. break;
  259. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_13:
  260. output_mult = 1077969154;
  261. output_shift = 2;
  262. filter_offset = 0;
  263. input_offset = 0;
  264. output_offset = 1;
  265. act_min =-128;
  266. act_max= 127;
  267. nb_batches=8;
  268. colDim=8;
  269. rowDim=5;
  270. input.reload(FullyConnected::INPUT13_S8_ID,mgr);
  271. bias.reload(FullyConnected::BIAS13_S8_ID,mgr);
  272. weight.reload(FullyConnected::WEIGHT13_S8_ID,mgr);
  273. ref.reload(FullyConnected::REF13_S8_ID,mgr);
  274. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  275. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  276. break;
  277. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_14:
  278. output_mult = 1073741824;
  279. output_shift = 1;
  280. filter_offset = 1;
  281. input_offset = 1;
  282. output_offset = -1;
  283. act_min =-128;
  284. act_max= 127;
  285. nb_batches=4;
  286. colDim=7;
  287. rowDim=3;
  288. input.reload(FullyConnected::INPUT14_S8_ID,mgr);
  289. bias.reload(FullyConnected::BIAS14_S8_ID,mgr);
  290. weight.reload(FullyConnected::WEIGHT14_S8_ID,mgr);
  291. ref.reload(FullyConnected::REF14_S8_ID,mgr);
  292. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  293. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  294. break;
  295. case FullyConnected::TEST_FULLY_CONNECTED_TFLITE_S8_15:
  296. output_mult = 1073741824;
  297. output_shift = 1;
  298. filter_offset = 1;
  299. input_offset = 1;
  300. output_offset = -1;
  301. act_min =-128;
  302. act_max= 127;
  303. nb_batches=8;
  304. colDim=7;
  305. rowDim=4;
  306. input.reload(FullyConnected::INPUT15_S8_ID,mgr);
  307. bias.reload(FullyConnected::BIAS15_S8_ID,mgr);
  308. weight.reload(FullyConnected::WEIGHT15_S8_ID,mgr);
  309. ref.reload(FullyConnected::REF15_S8_ID,mgr);
  310. output.create(ref.nbSamples(),FullyConnected::OUTPUT_S8_ID,mgr);
  311. temp.create(colDim,FullyConnected::TEMP_S16_ID,mgr);
  312. break;
  313. }
  314. }
  315. void FullyConnected::tearDown(Testing::testID_t id,Client::PatternMgr *mgr)
  316. {
  317. output.dump(mgr);
  318. }