Coverage for src/flag_gems/runtime/backend/_tsingmicro/heuristics_config_utils.py: 0%

1import torch 

2import triton 

3 

4_MIN_TILE_N = 64 

5_MAX_TILE_N_PER_ROW = 4096 

6_MAX_ONE_TILE_N = 2048 

7 

8 

9def simple_elementwise_blocksize_heur(args): 

10 return 1024 

11 

12 

13def argmax_heur_tile_k(args): 

14 MAX_TILE_K = 512 

15 NUM_SMS = torch.txda.get_device_properties( 

16 torch.txda.current_device() 

17 ).multi_processor_count 

18 

19 K = args["K"] 

20 M = args["M"] 

21 dtype = "fp32" if args["inp"].dtype == torch.float32 else "fp16" 

22 

23 if M == 64 and K == 512: 

24 return 64 if dtype == "fp32" else 128 

25 

26 if K <= 128: 

27 return 1 << (K.bit_length() - 1) if K > 0 else 1 

28 

29 tile_k = 64 

30 upper_bound = min(K, MAX_TILE_K) 

31 

32 while tile_k <= upper_bound: 

33 num_blocks = M * triton.cdiv(K, tile_k) 

34 num_waves = num_blocks / NUM_SMS 

35 

36 if num_waves > 1 and (tile_k * 2 <= upper_bound): 

37 tile_k *= 2 

38 else: 

39 break 

40 

41 return tile_k 

42 

43 

44def argmax_heur_tile_n_non_inner(args): 

45 n = args["N"] 

46 tile_k = args["TILE_K"] 

47 

48 if n <= 128: 

49 return n 

50 

51 target_tile = min(8192, n) 

52 tile_n = triton.next_power_of_2(target_tile) 

53 tile_n = max(64, min(tile_n, 4096)) 

54 

55 if tile_n * tile_k > 32768: 

56 tile_n = max(64, 32768 // tile_k) 

57 

58 return tile_n 

59 

60 

61def argmax_heur_one_tile_per_cta(args): 

62 return args["TILE_N"] >= args["N"] 

63 

64 

65def argmax_heur_num_warps_non_inner(args): 

66 # tile_n = args["TILE_N"] 

67 # dtype = "fp32" if args["inp"].dtype == torch.float32 else "fp16" 

68 

69 # if tile_n <= 32: 

70 # num_warps = 2 

71 # elif tile_n <= 64: 

72 # num_warps = 4 

73 # elif tile_n <= 128: 

74 # num_warps = 4 

75 # else: 

76 # num_warps = 8 

77 

78 # if dtype == "fp32": 

79 # num_warps = min(num_warps, 4) 

80 

81 # return num_warps 

82 

83 return 1 

84 

85 

86def argmax_heur_tile_n_inner(args): 

87 if args["N"] <= (32 * 1024): 

88 return triton.next_power_of_2(args["N"]) 

89 else: 

90 return 4096 

91 

92 

93def argmax_heur_num_warps_inner(args): 

94 # tile_size = args["TILE_N"] 

95 # if tile_size < 2048: 

96 # return 4 

97 # elif tile_size < 4096: 

98 # return 8 

99 # else: 

100 # return 16 

101 

102 return 1 

103 

104 

105def argmin_heur_block_m(args): 

106 return 16 if args["M"] < 4096 else 32 

107 

108 

109def argmin_heur_block_n(args): 

110 return min(16384, triton.next_power_of_2(args["N"])) 

111 

112 

113def bmm_heur_divisible_m(args): 

114 return args["M"] % args["TILE_M"] == 0 

115 

116 

117def bmm_heur_divisible_n(args): 

118 return args["N"] % args["TILE_N"] == 0 

119 

120 

121def bmm_heur_divisible_k(args): 

122 return args["K"] % args["TILE_K"] == 0 

123 

124 

125def baddbmm_heur_divisible_m(args): 

126 return args["M"] % args["TILE_M"] == 0 

127 

128 

129def baddbmm_heur_divisible_n(args): 

130 return args["N"] % args["TILE_N"] == 0 

131 

132 

133def baddbmm_heur_divisible_k(args): 

134 return args["K"] % args["TILE_K"] == 0 

135 

136 

137def dropout_heur_block(args): 

138 if args["N"] <= 512: 

139 return 512 

140 else: 

141 return 1024 

142 

143 

144def dropout_heur_num_warps(args): 

145 if args["N"] <= 512: 

146 return 4 

147 elif args["N"] <= 1024: 

148 return 8 

149 else: 

150 return 16 

151 

152 

153def exponential_heur_block(args): 

154 if args["N"] <= 512: 

155 return 512 

156 else: 

157 return 1024 

158 

159 

160def exponential_heur_num_warps(args): 

161 if args["N"] <= 512: 

162 return 4 

163 elif args["N"] <= 1024: 

164 return 8 

165 else: 

166 return 16 

167 

168 

169def gather_heur_block_m(args): 

170 return min(4, triton.next_power_of_2(triton.cdiv(args["N"], 2048))) 

171 

172 

173def gather_heur_block_n(args): 

174 return min(2048, triton.next_power_of_2(args["N"])) 

175 

176 

177def index_select_heur_block_m(args): 

178 return min(4, triton.next_power_of_2(triton.cdiv(256, args["N"]))) 

179 

180 

181def index_select_heur_block_n(args): 

182 m = min(triton.next_power_of_2(triton.cdiv(args["N"], 16)), 512) 

183 return max(m, 16) 

184 

185 

186def mm_heur_even_k(args): 

187 return args["K"] % (args["BLOCK_K"] * args["SPLIT_K"]) == 0 

188 

189 

190def rand_heur_block(args): 

191 if args["N"] <= 512: 

192 return 512 

193 else: 

194 return 1024 

195 

196 

197def rand_heur_num_warps(args): 

198 if args["N"] <= 512: 

199 return 4 

200 elif args["N"] <= 1024: 

201 return 8 

202 else: 

203 return 16 

204 

205 

206def randn_heur_block(args): 

207 if args["N"] <= 512: 

208 return 512 

209 else: 

210 return 1024 

211 

212 

213def randn_heur_num_warps(args): 

214 if args["N"] <= 512: 

215 return 4 

216 elif args["N"] <= 1024: 

217 return 8 

218 else: 

219 return 16 

220 

221 

222def softmax_heur_tile_k(args): 

223 MAX_TILE_K = 8192 

224 NUM_SMS = torch.txda.get_device_properties( 

225 torch.txda.current_device() 

226 ).multi_processor_count 

227 tile_k = 1 

228 upper_bound = min(args["K"], MAX_TILE_K) 

229 while tile_k <= upper_bound: 

230 num_blocks = args["M"] * triton.cdiv(args["K"], tile_k) 

231 num_waves = num_blocks / NUM_SMS 

232 if (num_waves > 1) and (tile_k * 2 <= upper_bound): 

233 tile_k *= 2 

234 else: 

235 break 

236 return tile_k 

237 

238 

239def softmax_heur_tile_n_non_inner(args): 

240 return triton.cdiv(8192, args["TILE_K"]) 

241 

242 

243def softmax_heur_one_tile_per_cta(args): 

244 return args["TILE_N"] >= args["N"] 

245 

246 

247def softmax_heur_num_warps_non_inner(args): 

248 tile_size = args["TILE_N"] * args["TILE_K"] 

249 if tile_size < 2048: 

250 return 4 

251 elif tile_size < 4096: 

252 return 8 

253 else: 

254 return 16 

255 

256 

257def softmax_heur_tile_n_inner(args): 

258 if args["N"] <= (32 * 1024): 

259 return triton.next_power_of_2(args["N"]) 

260 else: 

261 return 4096 

262 

263 

264def softmax_heur_num_warps_inner(args): 

265 tile_size = args["TILE_N"] 

266 if tile_size < 2048: 

267 return 4 

268 elif tile_size < 4096: 

269 return 8 

270 else: 

271 return 16 

272 

273 

274def softmax_heur_tile_n_bwd_non_inner(args): 

275 return max(1, 1024 // args["TILE_K"]) 

276 

277 

278def softmax_heur_tile_m(args): 

279 return max(1, 1024 // args["TILE_N"]) 

280 

281 

282def uniform_heur_block(args): 

283 if args["N"] <= 512: 

284 return 512 

285 else: 

286 return 1024 

287 

288 

289def uniform_heur_num_warps(args): 

290 if args["N"] <= 512: 

291 return 4 

292 elif args["N"] <= 1024: 

293 return 8 

294 else: 

295 return 16 

296 

297 

298def var_mean_heur_block_n(args): 

299 return triton.next_power_of_2(args["BLOCK_NUM"]) 

300 

301 

302def upsample_nearest1d_SAME_L(args): 

303 return args["OL"] == args["IL"] 

304 

305 

306def upsample_nearest1d_USE_INT32_IDX(args): 

307 return args["N"] * args["C"] * args["OL"] <= (2**31 - 1) # INT32 MAX 

308 

309 

310def upsample_nearest2d_SAME_H(args): 

311 return args["OH"] == args["IH"] 

312 

313 

314def upsample_nearest2d_SAME_W(args): 

315 return args["OW"] == args["IW"] 

316 

317 

318def upsample_nearest2d_USE_INT32_IDX(args): 

319 return args["N"] * args["C"] * args["OH"] * args["OW"] <= (2**31 - 1) # INT32 MAX 

320 

321 

322def batch_norm_heur_block_m(args): 

323 return min(2048, triton.next_power_of_2(args["batch_dim"])) 

324 

325 

326def batch_norm_heur_block_n(args): 

327 # A maximum of 16384 elements are loaded at once. 

328 BLOCK_M = batch_norm_heur_block_m(args) 

329 BLOCK_N = triton.next_power_of_2(args["spatial_dim"]) 

330 return min(BLOCK_N, max(1, 2**14 // BLOCK_M)) 

331 

332 

333def vdot_heur_block_size(args): 

334 n = args["n_elements"] 

335 if n < 1024: 

336 return 32 

337 elif n < 8192: 

338 return 256 

339 else: 

340 return 1024 

341 

342 

343def mean_heur_tile_k(args): 

344 MAX_TILE_K = 512 

345 NUM_SMS = torch.txda.get_device_properties( 

346 torch.txda.current_device() 

347 ).multi_processor_count 

348 tile_k = 1 

349 upper_bound = min(args["K"], MAX_TILE_K) 

350 max_tile_k_allowed_by_tile_n = max(1, _MAX_TILE_N_PER_ROW // _MIN_TILE_N) 

351 upper_bound = min(upper_bound, max_tile_k_allowed_by_tile_n) 

352 while tile_k <= upper_bound: 

353 num_blocks = args["M"] * triton.cdiv(args["K"], tile_k) 

354 num_waves = num_blocks / NUM_SMS 

355 if (num_waves > 1) and (tile_k * 2 <= upper_bound): 

356 tile_k *= 2 

357 else: 

358 break 

359 return tile_k 

360 

361 

362def mean_heur_tile_n_non_inner(args): 

363 tile_k = args.get("TILE_K", 1) 

364 limit_by_k = max(1, _MAX_TILE_N_PER_ROW // tile_k) 

365 N = args.get("N", 1) 

366 desired = min(max(N, _MIN_TILE_N), limit_by_k) 

367 desired = min(desired, _MAX_ONE_TILE_N, limit_by_k) 

368 tile_n = triton.next_power_of_2(desired) 

369 if tile_n > limit_by_k: 

370 tile_n = limit_by_k 

371 tile_n = max(tile_n, _MIN_TILE_N) 

372 return tile_n 

373 

374 

375def mean_heur_one_tile_per_cta(args): 

376 return args["TILE_N"] >= args["N"] 

377 

378 

379HEURISTICS_CONFIGS = { 

380 "argmax_non_inner": { 

381 "TILE_K": argmax_heur_tile_k, 

382 "TILE_N": argmax_heur_tile_n_non_inner, 

383 "ONE_TILE_PER_CTA": argmax_heur_one_tile_per_cta, 

384 "num_warps": argmax_heur_num_warps_non_inner, 

385 }, 

386 "argmax_inner": { 

387 "TILE_N": argmax_heur_tile_n_inner, 

388 "ONE_TILE_PER_CTA": argmax_heur_one_tile_per_cta, 

389 "num_warps": argmax_heur_num_warps_inner, 

390 }, 

391 "argmin": { 

392 "BLOCK_M": argmin_heur_block_m, 

393 "BLOCK_N": argmin_heur_block_n, 

394 }, 

395 "bmm": { 

396 "DIVISIBLE_M": bmm_heur_divisible_m, 

397 "DIVISIBLE_N": bmm_heur_divisible_n, 

398 "DIVISIBLE_K": bmm_heur_divisible_k, 

399 }, 

400 "baddbmm": { 

401 "DIVISIBLE_M": baddbmm_heur_divisible_m, 

402 "DIVISIBLE_N": baddbmm_heur_divisible_n, 

403 "DIVISIBLE_K": baddbmm_heur_divisible_k, 

404 }, 

405 "dropout": { 

406 "BLOCK": dropout_heur_block, 

407 "num_warps": dropout_heur_num_warps, 

408 }, 

409 "exponential_": { 

410 "BLOCK": exponential_heur_block, 

411 "num_warps": exponential_heur_num_warps, 

412 }, 

413 "gather": { 

414 "BLOCK_M": gather_heur_block_m, 

415 "BLOCK_N": gather_heur_block_n, 

416 }, 

417 "index_select": { 

418 "BLOCK_M": index_select_heur_block_m, 

419 "BLOCK_N": index_select_heur_block_n, 

420 }, 

421 "mm": { 

422 "EVEN_K": mm_heur_even_k, 

423 }, 

424 "rand": { 

425 "BLOCK": rand_heur_block, 

426 "num_warps": rand_heur_num_warps, 

427 }, 

428 "randn": { 

429 "BLOCK": randn_heur_block, 

430 "num_warps": randn_heur_num_warps, 

431 }, 

432 "softmax_non_inner": { 

433 "TILE_K": softmax_heur_tile_k, 

434 "TILE_N": softmax_heur_tile_n_non_inner, 

435 "ONE_TILE_PER_CTA": softmax_heur_one_tile_per_cta, 

436 "num_warps": softmax_heur_num_warps_non_inner, 

437 }, 

438 "mean_non_inner": { 

439 "TILE_K": mean_heur_tile_k, 

440 "TILE_N": mean_heur_tile_n_non_inner, 

441 "ONE_TILE_PER_CTA": mean_heur_one_tile_per_cta, 

442 "num_warps": softmax_heur_num_warps_non_inner, 

443 }, 

444 "softmax_inner": { 

445 "TILE_N": softmax_heur_tile_n_inner, 

446 "ONE_TILE_PER_CTA": softmax_heur_one_tile_per_cta, 

447 "num_warps": softmax_heur_num_warps_inner, 

448 }, 

449 "softmax_backward_non_inner": { 

450 "TILE_N": softmax_heur_tile_n_bwd_non_inner, 

451 "ONE_TILE_PER_CTA": softmax_heur_one_tile_per_cta, 

452 }, 

453 "softmax_backward_inner": { 

454 "TILE_M": softmax_heur_tile_m, 

455 "ONE_TILE_PER_CTA": softmax_heur_one_tile_per_cta, 

456 }, 

457 "uniform": { 

458 "BLOCK": uniform_heur_block, 

459 "num_warps": uniform_heur_num_warps, 

460 }, 

461 "upsample_nearest1d": { 

462 "SAME_L": upsample_nearest1d_SAME_L, 

463 "USE_INT32_IDX": upsample_nearest1d_USE_INT32_IDX, 

464 }, 

465 "upsample_nearest2d": { 

466 "SAME_H": upsample_nearest2d_SAME_H, 

467 "SAME_W": upsample_nearest2d_SAME_W, 

468 "USE_INT32_IDX": upsample_nearest2d_USE_INT32_IDX, 

469 }, 

470 "var_mean": { 

471 "BLOCK_N": var_mean_heur_block_n, 

472 }, 

473 "batch_norm": { 

474 "BLOCK_M": batch_norm_heur_block_m, 

475 "BLOCK_N": batch_norm_heur_block_n, 

476 }, 

477 "vdot": { 

478 "BLOCK_SIZE": vdot_heur_block_size, 

479 }, 

480 "mha_block_128": { 

481 "BLOCK_M": lambda args: 128, 

482 "BLOCK_N": lambda args: 32, 

483 "num_warps": lambda args: 4, 

484 "num_stages": lambda args: 3, 

485 }, 

486 "mha_block_64": { 

487 "BLOCK_M": lambda args: 64, 

488 "BLOCK_N": lambda args: 64, 

489 "num_warps": lambda args: 4, 

490 "num_stages": lambda args: 3, 

491 }, 

492 "mha_block_32": { 

493 "BLOCK_M": lambda args: 32, 

494 "BLOCK_N": lambda args: 64, 

495 "num_warps": lambda args: 4, 

496 "num_stages": lambda args: 3, 

497 }, 

498 "mha_block_16": { 

499 "BLOCK_M": lambda args: 16, 

500 "BLOCK_N": lambda args: 64, 

501 "num_warps": lambda args: 4, 

502 "num_stages": lambda args: 3, 

503 }, 

504 "elementwise_generic": { 

505 "BLOCK_SIZE": simple_elementwise_blocksize_heur, 

506 "num_warps": lambda args: 8, 

507 }, 

508}