Line data Source code
1 : /**
2 : * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3 : * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4 : * CANN Open Software License Agreement Version 2.0 (the "License").
5 : * Please refer to the License for details. You may not use this file except in compliance with the License.
6 : * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7 : * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8 : * See LICENSE in the root of the software repository for the full text of the License.
9 : */
10 :
11 : /*!
12 : * \file kernel_utils_constants.h
13 : * \brief
14 : */
15 : #ifndef ASCENDC_MODULE_UTILS_CONSTANTS_H
16 : #define ASCENDC_MODULE_UTILS_CONSTANTS_H
17 : #include "utils/kernel_utils_ceil_oom_que.h"
18 :
19 : namespace AscendC {
20 : const int32_t DEFAULT_BLK_NUM = 8;
21 : const int32_t HALF_FACTOR = 2;
22 : const int32_t DEFAULT_BLK_STRIDE = 1;
23 : const uint8_t DEFAULT_REPEAT_STRIDE = 8;
24 : const uint64_t FULL_MASK = 0xffffffffffffffff;
25 : const uint64_t CONST_MASK_VALUE = 0x8000000000000000;
26 : const uint16_t MAX_HALF_MASK_LEN = 64;
27 : const int32_t DEFAULT_C0_SIZE = 32;
28 : const int32_t DEFAULT_BLOCK_SIZE = 256;
29 : const int32_t MAX_REPEAT_TIMES = 255;
30 : const int32_t MIN_REPEAT_TIMES = 0;
31 : const int32_t ONE_BYTE_BIT_SIZE = 8;
32 : const uint32_t MAX_SLICE_SIZE = 6 * 256;
33 : const uint16_t VALUE_512 = 512; // align with 512B / value range [0, 512]
34 :
35 : // int4b_t param
36 : constexpr uint32_t INT4_TWO = 2;
37 : constexpr uint32_t INT4_BIT_NUM = 4;
38 :
39 : #if (__NPU_ARCH__ == 5102)
40 : // int2b_t param
41 : constexpr uint32_t INT2_FOUR = 4;
42 : constexpr uint32_t INT2_BIT_NUM = 2;
43 : constexpr uint32_t INT1_BIT_NUM = 1;
44 : #endif
45 :
46 : #if defined(__NPU_ARCH__) && \
47 : ((__NPU_ARCH__ == 2103) || (__NPU_ARCH__ == 3003) || (__NPU_ARCH__ == 3103) || (__NPU_ARCH__ == 3113))
48 : // int3b_t param
49 : constexpr uint32_t INT3_BIT_NUM = 3;
50 : // int2b_t param
51 : constexpr uint32_t INT2_FOUR = 4;
52 : constexpr uint32_t INT2_BIT_NUM = 2;
53 : #endif
54 :
55 : #if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
56 : namespace ConstantsInternal {
57 : constexpr uint32_t ASCENDC_B4_TWO = 2;
58 : constexpr uint32_t ASCENDC_B4_BIT_NUM = 4;
59 : } // namespace ConstantsInternal
60 : #endif
61 :
62 : // AddDeqRelu param
63 : constexpr int32_t DEQ_SHIFT_LEFT_17_BIT = 131072;
64 : constexpr float DEQ_SHIFT_RIGHT_17_BIT = 1.0 / DEQ_SHIFT_LEFT_17_BIT;
65 : constexpr int8_t ADDDEQRELU_MASK_MODE_ONE = 1;
66 : constexpr int8_t ADDDEQRELU_MASK_MODE_TWO = 2;
67 :
68 : #if (__NPU_ARCH__ == 5102)
69 : const int32_t TOTAL_VEC_LOCAL_SIZE = 248 * 1024;
70 : const uint32_t TOTAL_UB_SIZE = 248 * 1024;
71 : const uint32_t TMP_UB_OFFSET = 248 * 1024;
72 : const uint32_t TOTAL_L1_SIZE = 1024 * 1024;
73 : const uint32_t SINGLE_MSG_SIZE = 64;
74 : const uint32_t CACHE_LINE_SIZE = 64;
75 : const uint32_t TOTAL_L0C_SIZE = 256 * 1024;
76 : const uint32_t VECTOR_REG_WIDTH = 256;
77 : const uint32_t VECTOR_REG_WIDTH_2XVL = 512;
78 : const uint32_t ONE_BLOCK_SIZE = 32;
79 : #elif defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 1001) || (__NPU_ARCH__ == 2002))
80 : const int32_t TOTAL_VEC_LOCAL_SIZE = 248 * 1024;
81 : const uint32_t TOTAL_UB_SIZE = 256 * 1024;
82 :
83 : const uint32_t TMP_UB_OFFSET = 248 * 1024;
84 : const uint32_t TOTAL_L1_SIZE = 1024 * 1024;
85 : const uint32_t TOTAL_L0C_SIZE = 256 * 1024;
86 : #elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2103)
87 : const int32_t TOTAL_VEC_LOCAL_SIZE = 112 * 1024;
88 : const uint32_t TOTAL_UB_SIZE = 120 * 1024;
89 : const uint32_t TMP_UB_OFFSET = 112 * 1024;
90 : const uint32_t TOTAL_L1_SIZE = 1024 * 1024;
91 : const uint32_t VECTOR_REG_WIDTH = 128;
92 : const uint32_t ONE_BLOCK_SIZE = 32;
93 : const uint32_t TOTAL_L0C_SIZE = 64 * 1024;
94 : #elif (__NPU_ARCH__ == 2201)
95 : const int32_t TOTAL_VEC_LOCAL_SIZE = 184 * 1024;
96 : const uint32_t TOTAL_UB_SIZE = 192 * 1024;
97 : const uint32_t TMP_UB_OFFSET = 184 * 1024;
98 : #ifndef KFC_L1_RESERVER_SIZE
99 : #define KFC_L1_RESERVER_SIZE 128
100 : #endif
101 : const uint32_t TOTAL_L1_SIZE = 512 * 1024 - KFC_L1_RESERVER_SIZE;
102 : const uint32_t SINGLE_MSG_SIZE = 64;
103 : const uint32_t CACHE_LINE_SIZE = 64;
104 : const uint32_t TOTAL_L0C_SIZE = 128 * 1024;
105 : #elif (__NPU_ARCH__ == 3002)
106 : const int32_t TOTAL_VEC_LOCAL_SIZE = 184 * 1024;
107 : const uint32_t TOTAL_UB_SIZE = 248 * 1024;
108 : const uint32_t TMP_UB_OFFSET = 248 * 1024;
109 : const uint32_t TOTAL_L1_SIZE = 1024 * 1024;
110 : const uint32_t SINGLE_MSG_SIZE = 64;
111 : const uint32_t CACHE_LINE_SIZE = 64;
112 : const uint32_t TOTAL_L0C_SIZE = 128 * 1024;
113 : const uint32_t VECTOR_REG_WIDTH = 256;
114 : #elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3003)
115 : const int32_t TOTAL_VEC_LOCAL_SIZE = 118 * 1024;
116 : const uint32_t TOTAL_UB_SIZE = 118 * 1024;
117 : const uint32_t TMP_UB_OFFSET = 118 * 1024;
118 : const uint32_t TOTAL_L1_SIZE = 1024 * 1024;
119 : const uint32_t VECTOR_REG_WIDTH = 256;
120 : const uint32_t ONE_BLOCK_SIZE = 32;
121 : const uint32_t TOTAL_L0C_SIZE = 128 * 1024;
122 : #elif (__NPU_ARCH__ == 3102)
123 : const int32_t TOTAL_VEC_LOCAL_SIZE = 184 * 1024;
124 : const uint32_t TOTAL_UB_SIZE = 256 * 1024;
125 : const uint32_t TMP_UB_OFFSET = 248 * 1024;
126 : const uint32_t TOTAL_L1_SIZE = 1024 * 1024;
127 : const uint32_t SINGLE_MSG_SIZE = 64;
128 : const uint32_t CACHE_LINE_SIZE = 64;
129 : const uint32_t TOTAL_L0C_SIZE = 128 * 1024;
130 : const uint32_t VECTOR_REG_WIDTH = 256;
131 : #elif (__NPU_ARCH__ == 3510)
132 : const int32_t TOTAL_VEC_LOCAL_SIZE = 248 * 1024;
133 : const uint32_t TOTAL_UB_SIZE = 248 * 1024;
134 : const uint32_t TMP_UB_OFFSET = 248 * 1024;
135 : const uint32_t TOTAL_L1_SIZE = 512 * 1024;
136 : const uint32_t SINGLE_MSG_SIZE = 64;
137 : const uint32_t CACHE_LINE_SIZE = 64;
138 : const uint32_t TOTAL_L0C_SIZE = 256 * 1024;
139 : const uint32_t VECTOR_REG_WIDTH = 256;
140 : const uint32_t VECTOR_REG_WIDTH_2XVL = 512;
141 : const uint32_t ONE_BLOCK_SIZE = 32;
142 : #elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3113)
143 : const int32_t TOTAL_VEC_LOCAL_SIZE = 118 * 1024;
144 : const uint32_t TOTAL_UB_SIZE = 118 * 1024;
145 : const uint32_t TMP_UB_OFFSET = 118 * 1024;
146 : const uint32_t TOTAL_L1_SIZE = 512 * 1024;
147 : const uint32_t TOTAL_L0C_SIZE = 64 * 1024;
148 : const uint32_t VECTOR_REG_WIDTH = 256;
149 : const uint32_t ONE_BLOCK_SIZE = 32;
150 : #endif
151 :
152 : #ifndef __NPU_HOST__
153 :
154 : #ifndef ASCC_PARAM_BLOCK_CUBE
155 : #define ASCC_PARAM_BLOCK_CUBE
156 : const int32_t BLOCK_CUBE = 16;
157 : #endif
158 :
159 : #ifndef ASCC_PARAM_ONE_BLK_SIZE
160 : #define ASCC_PARAM_ONE_BLK_SIZE
161 : const uint16_t ONE_BLK_SIZE = 32;
162 : #endif
163 :
164 : #ifndef ASCC_PARAM_CUBE_MAX_SIZE
165 : #define ASCC_PARAM_CUBE_MAX_SIZE
166 : const int32_t CUBE_MAX_SIZE = 256;
167 : #endif
168 :
169 : #else // #ifdef __NPU_HOST__
170 :
171 : #ifndef ASCC_PARAM_BLOCK_CUBE
172 : #define ASCC_PARAM_BLOCK_CUBE
173 : constexpr uint32_t BLOCK_CUBE = 16;
174 : #endif
175 :
176 : #ifndef ASCC_PARAM_ONE_BLK_SIZE
177 : #define ASCC_PARAM_ONE_BLK_SIZE
178 : constexpr uint32_t ONE_BLK_SIZE = 32;
179 : #endif
180 :
181 : #ifndef ASCC_PARAM_CUBE_MAX_SIZE
182 : #define ASCC_PARAM_CUBE_MAX_SIZE
183 : constexpr int32_t CUBE_MAX_SIZE = 256;
184 : #endif
185 :
186 : #endif // __NPU_HOST__
187 : const uint8_t PAD_SIZE = 4;
188 : const uint8_t MRG_SORT_ELEMENT_LEN = 4;
189 : const uint8_t DEFAULT_DATA_COPY_NBURST = 1;
190 : const uint8_t DEFAULT_DATA_COPY_STRIDE = 0;
191 : const int32_t BYTE_PER_FRACTAL = 512;
192 : const int32_t SRC_BURST_LEN_SIZE_ELE = 16;
193 : const int32_t SRC_GAP_SIZE_BYTE = 32;
194 : const int32_t DST_BURST_LEN_SIZE_ELE = 256;
195 : const int32_t VREDUCE_PER_REP_OUTPUT = 2;
196 : const uint16_t ONE_PARAM_SIZE = 8;
197 : #if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
198 : const uint16_t AIV_CORE_NUM = 72;
199 : #else
200 : const uint16_t AIV_CORE_NUM = 50;
201 : #endif
202 : const uint16_t DUMP_MSG_HEAD_SIZE = 24;
203 : #if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3103)
204 : const int32_t ONE_REPEAT_BYTE_SIZE = 128;
205 : #else
206 : const int32_t ONE_REPEAT_BYTE_SIZE = 256;
207 : #endif
208 : const int32_t FULL_MASK_LEN = 128;
209 : const int32_t HLAF_MASK_LEN = 64;
210 : const int32_t DEFAULT_REDUCE_DST_REP_SRIDE = 1;
211 : const uint8_t B64_BYTE_SIZE = 8;
212 : const uint8_t B32_BYTE_SIZE = 4;
213 : const uint8_t B16_BYTE_SIZE = 2;
214 : const uint8_t B8_BYTE_SIZE = 1;
215 : const uint8_t B32_DATA_NUM_PER_BLOCK = 8;
216 : const uint8_t B16_DATA_NUM_PER_BLOCK = 16;
217 : #if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3103)
218 : const int32_t B16_DATA_NUM_PER_REPEAT = 64;
219 : const int32_t B32_DATA_NUM_PER_REPEAT = 32;
220 : #else
221 : const int32_t B16_DATA_NUM_PER_REPEAT = 128;
222 : const int32_t B32_DATA_NUM_PER_REPEAT = 64;
223 : #endif
224 :
225 : #if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
226 : const uint32_t B64_DATA_NUM_PER_REPEAT = 32;
227 : const uint32_t B4_BYTE_SIZE_PER_REPEAT = 64;
228 : const uint32_t L1_DUMP_UB_SIZE = TOTAL_UB_SIZE - 32 * 1024;
229 : #endif
230 :
231 : const int32_t BLOCK_STRIDE_POS_IN_SM = 16;
232 : const int32_t PLD_BUFFER_SIZE = 2;
233 : const uint8_t FIXPIPE_DEQ_TENSOR_SIZE = 16;
234 : const uint8_t SET_DATA_EXP_ZERO = 0;
235 : const uint8_t SET_DATA_EXP_ONE = 1;
236 : const uint8_t SET_DATA_EXP_TWO = 2;
237 : const uint8_t SET_DATA_EXP_THREE = 3;
238 : const uint8_t VDEQ_TENSOR_SIZE = 16;
239 : // workspace system reserve 16MB
240 : #if (__NPU_ARCH__ == 1001)
241 : constexpr size_t RESERVED_WORKSPACE = 2 * 1024 * 1024;
242 : #elif (__NPU_ARCH__ == 2002)
243 : constexpr size_t RESERVED_WORKSPACE = 2 * 1024 * 1024;
244 : #elif (__NPU_ARCH__ == 2201)
245 : constexpr size_t RESERVED_WORKSPACE = 16 * 1024 * 1024;
246 : #elif (__NPU_ARCH__ == 3002)
247 : constexpr size_t RESERVED_WORKSPACE = 16 * 1024 * 1024;
248 : #elif (__NPU_ARCH__ == 3102)
249 : constexpr size_t RESERVED_WORKSPACE = 16 * 1024 * 1024;
250 : #elif defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
251 : constexpr size_t RESERVED_WORKSPACE = 16 * 1024 * 1024;
252 : #elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 2103)
253 : constexpr size_t RESERVED_WORKSPACE = 16 * 1024 * 1024;
254 : #elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3003)
255 : constexpr size_t RESERVED_WORKSPACE = 0;
256 : #elif defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3103) || (__NPU_ARCH__ == 3113))
257 : constexpr size_t RESERVED_WORKSPACE = 0;
258 : #endif
259 :
260 : // nchwconv address list size
261 : const int32_t NCHW_CONV_ADDR_LIST_SIZE = 16;
262 : const int32_t VA_REG_ARRAY_LEN = 8;
263 : const uint8_t CONV2D_IMG_SIZE = 2;
264 : const uint8_t CONV2D_KERNEL_SIZE = 2;
265 : const uint8_t CONV2D_STRIDE = 2;
266 : const uint8_t CONV2D_PAD = 4;
267 : const uint8_t CONV2D_DILATION = 2;
268 : const int32_t K_MAX_DIM = 8;
269 :
270 : const uint32_t TWO_OF_STACK_BUFFER = 2;
271 : const uint32_t THREE_OF_STACK_BUFFER = 3;
272 : const uint32_t HALF_REPEAT_SIZE = ONE_REPEAT_BYTE_SIZE / B16_BYTE_SIZE;
273 : const uint32_t FLOAT_REPEAT_SIZE = ONE_REPEAT_BYTE_SIZE / B32_BYTE_SIZE;
274 : const uint32_t ONE_REPEAT_FLOAT_SIZE = ONE_REPEAT_BYTE_SIZE / B32_BYTE_SIZE;
275 : const uint32_t ONE_REPEAT_HALF_SIZE = ONE_REPEAT_BYTE_SIZE / B16_BYTE_SIZE;
276 : const uint32_t MAX_REPEAT_FLOAT_SIZE = ONE_REPEAT_FLOAT_SIZE * MAX_REPEAT_TIMES;
277 : const uint32_t MAX_REPEAT_HALF_SIZE = ONE_REPEAT_HALF_SIZE * MAX_REPEAT_TIMES;
278 : const uint32_t ONE_BLK_HALF_NUM = ONE_BLK_SIZE / B16_BYTE_SIZE;
279 : const uint32_t ONE_BLK_FLOAT_NUM = ONE_BLK_SIZE / B32_BYTE_SIZE;
280 : #if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
281 : namespace ConstantsInternal {
282 : const uint32_t ONE_BLK_FP4_NUM = 64;
283 : const uint32_t ONE_BLK_B2_NUM = 128;
284 : } // namespace ConstantsInternal
285 : #endif
286 : const uint32_t BRCB_BROADCAST_NUMBER = 8;
287 : const uint32_t BRCB_MAX_REPEAT_SIZE = BRCB_BROADCAST_NUMBER * MAX_REPEAT_TIMES;
288 : const int32_t MIN_BLOCK_LEN = 1;
289 : const uint32_t PAIR_REDUCE_REPEAT_STRIDE_LEN = 128;
290 : const uint32_t PAIR_REDUCE_SUM_MERGES = 2;
291 : const uint32_t TWO_HUNDRED_FIFTY_TWO_REPEAT = 252;
292 : const uint32_t TWO_HUNDRED_FIFTY_TWO_REPEAT_BYTE_SIZE = TWO_HUNDRED_FIFTY_TWO_REPEAT * ONE_REPEAT_BYTE_SIZE;
293 : const uint32_t REDUCEV2_MODE_SEVEN = 7;
294 : const uint32_t DROPOUT_MODE_BYTE_MISALIGN = 1;
295 : const uint32_t DROPOUT_MODE_BYTE_ALIGN = 2;
296 : const uint32_t DROPOUT_MODE_BIT_ALIGN = 3;
297 : const uint32_t DROPOUT_MODE_BIT_MISALIGN = 4;
298 : const uint32_t REDUCEV2_MODE_ONE = 1;
299 : const uint32_t REDUCEV2_MODE_TWO = 2;
300 : const uint32_t REDUCEV2_MODE_THREE = 3;
301 :
302 : // 4dTrans param size
303 : const int32_t B8_TMP_ELE_LEN = 1024;
304 : const int32_t B16_TMP_ELE_LEN = 256;
305 : const int32_t B32_TMP_ELE_LEN = 128;
306 : const int32_t B8_TRANS_LEN = 1024;
307 : const int32_t B8_TRANS_FRACTAL = 512;
308 : const int32_t B8_TRANS_ROW = 32;
309 : const int32_t B8_COPY_COL = 32;
310 :
311 : // load3dPro config
312 : const uint64_t LOAD_M_START_POSITION = 48;
313 : const uint64_t LOAD_K_START_POSITION = 32;
314 : const uint64_t LOAD_M_EXTENSION = 16;
315 : const uint64_t LOAD_DILATION_FILTER_H = 40;
316 : const uint64_t LOAD_DILATION_FILTER_W = 32;
317 : const uint64_t LOAD_FILTER_H = 24;
318 : const uint64_t LOAD_FILTER_W = 16;
319 : const uint64_t LOAD_STRIDE_H = 8;
320 :
321 : #if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
322 : namespace Internal {
323 : constexpr int32_t TSCM_CROSS_SYNC_ID_MAX = 11;
324 : }
325 : #endif
326 :
327 : #if defined(ASCENDC_CPU_DEBUG) && ASCENDC_CPU_DEBUG == 1
328 : // param check size
329 : const int32_t MAX_BLOCK_COUNT = 4095;
330 : const int32_t MIN_BLOCK_COUNT = 1;
331 : const int32_t MAX_BLOCK_LEN = 65535;
332 :
333 : const int32_t MAX_16BITS_STRIDE = 65535;
334 : const int32_t MAX_8BITS_STRIDE = 255;
335 : const int32_t MIN_BLOCK_NUM = 1;
336 : const int32_t MAX_PROPOSAL_MODE_NUM = 5;
337 : const int32_t MIN_PROPOSAL_MODE_NUM = 0;
338 :
339 : // load2d param size
340 : const int32_t MAX_LOAD2D_START_INDEX = 65535;
341 : const int32_t MIN_LOAD2D_START_INDEX = 0;
342 : const int32_t MAX_LOAD2D_SID = 15;
343 : const int32_t MIN_LOAD2D_SID = 0;
344 :
345 : // load3dv1 param size
346 : const int32_t MAX_LOAD3D_PAD = 255;
347 : const int32_t MIN_LOAD3D_PAD = 0;
348 : const int32_t MAX_LOAD3D_L1 = 32767;
349 : const int32_t MIN_LOAD3D_L1 = 1;
350 : const int32_t MAX_LOAD3D_C1_IDX = 4095;
351 : const int32_t MIN_LOAD3D_C1_IDX = 0;
352 : const int32_t MAX_LOAD3D_LEFT_TOP = 32767;
353 : const int32_t MIN_LOAD3D_LEFT_TOP = -255;
354 : const int32_t MAX_LOAD3D_STRIDE = 63;
355 : const int32_t MIN_LOAD3D_STRIDE = 1;
356 : const int32_t MAX_LOAD3D_FILTER = 255;
357 : const int32_t MIN_LOAD3D_FILTER = 1;
358 : const int32_t MIN_LOAD3D_FETCH_FILTER = 0;
359 : const int32_t MAX_LOAD3D_FETCH_FILTER = 254;
360 : const int32_t MIN_LOAD3D_DILATION_FILTER = 1;
361 : const int32_t MAX_LOAD3D_JUMP_STRIDE = 127;
362 : const int32_t MIN_LOAD3D_JUMP_STRIDE = 1;
363 : const int32_t MAX_LOAD3D_REPEAT_MODE = 1;
364 : const int32_t MIN_LOAD3D_REPEAT_MODE = 0;
365 : const int32_t MIN_LOAD3D_REPEAT_TIMES = 1;
366 : const int32_t MAX_LOAD3D_CSIZE = 1;
367 : const int32_t MIN_LOAD3D_CSIZE = 0;
368 :
369 : // load3dv2 param size
370 : const int32_t MAX_LOAD3D_CHANNEL_SIZE = 65535;
371 : const int32_t MIN_LOAD3D_CHANNEL_SIZE = 1;
372 : const int32_t MAX_LOAD3D_EXTENSION = 65535;
373 : const int32_t MIN_LOAD3D_EXTENSION = 1;
374 : const int32_t MAX_LOAD3D_START_PT = 65535;
375 : const int32_t MIN_LOAD3D_START_PT = 0;
376 : const int32_t KEXTENSION_HALF = 16;
377 : const int32_t MEXTENSION_HALF = 16;
378 : const int32_t KSTARTPT_HALF = 16;
379 : const int32_t MSTARTPT_HALF = 16;
380 : const int32_t KEXTENSION_B8 = 32;
381 : const int32_t MEXTENSION_B8 = 16;
382 : const int32_t KSTARTPT_B8 = 32;
383 : const int32_t MSTARTPT_B8 = 16;
384 :
385 : // loadImageToLocal param size
386 : constexpr int32_t MAX_LOADIMANG_L1_HORSIZE = 4095;
387 : constexpr int32_t MIN_LOADIMANG_L1_HORSIZE = 1;
388 : constexpr int32_t MAX_LOADIMANG_L1_VERSIZE = 4095;
389 : constexpr int32_t MIN_LOADIMANG_L1_VERSIZE = 0;
390 : constexpr int32_t MAX_LOADIMANG_L1_HWSTART = 4095;
391 : constexpr int32_t MIN_LOADIMANG_L1_HWSTART = 0;
392 : constexpr int32_t MAX_LOADIMANG_L1_SHORRES = 65535;
393 : constexpr int32_t MIN_LOADIMANG_L1_SHORRES = 1;
394 : constexpr int32_t MIN_LOADIMANG_L1_PADSIZE = 0;
395 :
396 : // mmad param size
397 : const int32_t MAX_M_K_N_SIZE = 4095;
398 : const int32_t MIN_M_K_N_SIZE = 0;
399 :
400 : // mrgsort4 param size
401 : const int32_t MAX_SORT_ELE_LEN = 4095;
402 : const int32_t MIN_SORT_ELE_LEN = 0;
403 : const int32_t MIN_SORT_REPEAT_TIMES = 1;
404 : namespace Internal {
405 : constexpr uint8_t REGION_PROPOSAL_ELEMENT_NUM = 8;
406 : }
407 :
408 : template <typename T>
409 : std::string ScalarToString(T scalarValue);
410 : template <>
411 4 : inline std::string ScalarToString(half scalarValue)
412 : {
413 4 : return std::to_string(scalarValue.ToFloat());
414 : }
415 : #if defined(__NPU_ARCH__) && \
416 : ((__NPU_ARCH__ == 2201) || (__NPU_ARCH__ == 3002) || (__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
417 : template <>
418 : inline std::string ScalarToString(bfloat16_t scalarValue)
419 : {
420 : return std::to_string(scalarValue.ToFloat());
421 : }
422 : #endif
423 : template <typename T>
424 : uint64_t GetScalarBitcode(T scalarValue);
425 : // deq tensor ptr could not be passed by cce instructions, so pass ptr to model by this function
426 : void SetModelDeqTensor(void* deqTensor);
427 : #if __NPU_ARCH__ == 3002
428 : void SetEleSrcPara(uint64_t baseAddr);
429 : #endif
430 : #if __NPU_ARCH__ == 2002
431 : void SetVbiSrc0Param(half* vbiSrc0Ptr, int32_t vbiSrc0Size);
432 : void SetUnzipCompressedLen(uint32_t compressedLength);
433 : #endif
434 : void SetModelBiasTensor(void* biasTensor);
435 : void SetIndexMatrix(void* indexMatrix);
436 :
437 : // src0 of gatherb instr could not be accessed by cce instructions, so pass ptr to model by this function
438 : void SetModelGatherbSrc0Tensor(uint64_t src0, const uint32_t length);
439 :
440 : // dst0 of scatter instr could not be accessed by cce instructions, so pass ptr to model by this function
441 : void SetModelScatterDst0Tensor(uint64_t dst0, const uint32_t length);
442 :
443 : #endif // ASCENDC_CPU_DEBUG
444 :
445 : #if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
446 : template <typename T>
447 : struct GetPadValueType {
448 : using Type = T;
449 : };
450 :
451 : // To support FP8 datacopypad, pad type needs transfer to b8
452 : template <>
453 : struct GetPadValueType<fp8_e5m2_t> {
454 : using Type = uint8_t;
455 : };
456 :
457 : template <>
458 : struct GetPadValueType<fp8_e4m3fn_t> {
459 : using Type = uint8_t;
460 : };
461 :
462 : template <>
463 : struct GetPadValueType<fp8_e8m0_t> {
464 : using Type = uint8_t;
465 : };
466 :
467 : template <>
468 : struct GetPadValueType<hifloat8_t> {
469 : using Type = uint8_t;
470 : };
471 :
472 : // To support FP4 datacopypad, pad type needs transfer to b8
473 : template <>
474 : struct GetPadValueType<fp4x2_e1m2_t> {
475 : using Type = uint8_t;
476 : };
477 :
478 : template <>
479 : struct GetPadValueType<fp4x2_e2m1_t> {
480 : using Type = uint8_t;
481 : };
482 : #endif
483 :
484 : template <bool condition, class T1, class T2>
485 : struct Conditional {
486 : using type = T1;
487 : };
488 :
489 : template <class T1, class T2>
490 : struct Conditional<false, T1, T2> {
491 : using type = T2;
492 : };
493 :
494 : template <int bitNum, bool sign = true>
495 : struct IntegerSubType {
496 : static int const kBits = bitNum;
497 : static bool const kSigned = sign;
498 :
499 : using T = typename Conditional<kSigned, int8_t, uint8_t>::type;
500 : using Storage = uint8_t;
501 :
502 : static Storage const mask = Storage(((static_cast<uint64_t>(1)) << static_cast<uint32_t>(kBits)) - 1);
503 : Storage storage;
504 : __aicore__ inline IntegerSubType() = default;
505 :
506 : __aicore__ inline IntegerSubType(uint32_t value) : storage(reinterpret_cast<Storage const&>(value) & mask) {}
507 :
508 76 : __aicore__ inline IntegerSubType(int32_t value) : storage(reinterpret_cast<Storage const&>(value) & mask) {}
509 :
510 8 : __aicore__ inline operator T() const
511 : {
512 8 : if (kSigned && ((storage & Storage(static_cast<uint64_t>(1) << static_cast<uint32_t>(kBits - 1))) != 0)) {
513 : // Sign extend
514 4 : return T(storage) | ~T(mask);
515 : }
516 4 : return T(storage);
517 : }
518 :
519 8 : __aicore__ inline bool operator==(IntegerSubType const& rhs) const { return storage == rhs.storage; }
520 :
521 : __aicore__ inline bool operator!=(IntegerSubType const& rhs) const { return storage != rhs.storage; }
522 :
523 28 : __aicore__ inline bool operator>(IntegerSubType const& rhs) const
524 : {
525 28 : bool lhsIsNeg = (this->storage & (static_cast<uint64_t>(1) << static_cast<uint32_t>(this->kBits - 1)));
526 28 : bool rhsIsNeg = (rhs.storage & (static_cast<uint64_t>(1) << static_cast<uint32_t>(rhs.kBits - 1)));
527 28 : if (kSigned && (lhsIsNeg != rhsIsNeg)) {
528 12 : return (!lhsIsNeg) && rhsIsNeg;
529 : }
530 16 : return this->storage > rhs.storage;
531 : }
532 :
533 32 : __aicore__ inline bool operator>=(IntegerSubType const& rhs) const
534 : {
535 32 : bool lhsIsNeg = (this->storage & (static_cast<uint64_t>(1) << static_cast<uint32_t>(this->kBits - 1)));
536 32 : bool rhsIsNeg = (rhs.storage & (static_cast<uint64_t>(1) << static_cast<uint32_t>(rhs.kBits - 1)));
537 32 : if (kSigned && (lhsIsNeg != rhsIsNeg)) {
538 12 : return (!lhsIsNeg) && rhsIsNeg;
539 : }
540 20 : return storage >= rhs.storage;
541 : }
542 :
543 16 : __aicore__ inline bool operator<(IntegerSubType const& rhs) const { return !(*this >= rhs); }
544 :
545 16 : __aicore__ inline bool operator<=(IntegerSubType const& rhs) const { return !(*this > rhs); }
546 : };
547 :
548 : using int4b_t = IntegerSubType<INT4_BIT_NUM, true>;
549 :
550 : #if (__NPU_ARCH__ == 5102)
551 : using int2b_t = IntegerSubType<INT2_BIT_NUM, true>;
552 : #endif
553 :
554 : #if defined(__NPU_ARCH__) && ((__NPU_ARCH__ == 3510) || (__NPU_ARCH__ == 5102))
555 : using mx_fp8_e5m2_t = struct {};
556 : using mx_fp8_e4m3_t = struct {};
557 : using mx_fp8_e8m0_t = struct {};
558 :
559 : template <typename T>
560 : struct GetDstType {
561 : using Type = T;
562 : };
563 :
564 : template <>
565 : struct GetDstType<mx_fp8_e5m2_t> {
566 : using Type = fp8_e5m2_t;
567 : };
568 :
569 : template <>
570 : struct GetDstType<mx_fp8_e4m3_t> {
571 : using Type = fp8_e4m3fn_t;
572 : };
573 :
574 : template <>
575 : struct GetDstType<mx_fp8_e8m0_t> {
576 : using Type = fp8_e8m0_t;
577 : };
578 : #endif
579 :
580 : } // namespace AscendC
581 : #endif // ASCENDC_MODULE_UTILS_CONSTANTS_H
|
