1 | /* Copyright 2018 Canaan Inc.
|
---|
2 | *
|
---|
3 | * Licensed under the Apache License, Version 2.0 (the "License");
|
---|
4 | * you may not use this file except in compliance with the License.
|
---|
5 | * You may obtain a copy of the License at
|
---|
6 | *
|
---|
7 | * http://www.apache.org/licenses/LICENSE-2.0
|
---|
8 | *
|
---|
9 | * Unless required by applicable law or agreed to in writing, software
|
---|
10 | * distributed under the License is distributed on an "AS IS" BASIS,
|
---|
11 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
---|
12 | * See the License for the specific language governing permissions and
|
---|
13 | * limitations under the License.
|
---|
14 | */
|
---|
15 | #ifndef _KPU_H
|
---|
16 | #define _KPU_H
|
---|
17 |
|
---|
18 | #include <kendryte-k210.h>
|
---|
19 | #include <stdint.h>
|
---|
20 | #include "device.h"
|
---|
21 |
|
---|
22 | #ifdef __cplusplus
|
---|
23 | extern "C" {
|
---|
24 | #endif
|
---|
25 |
|
---|
26 | #define kpu_matmul_begin kpu_conv2d_output
|
---|
27 | #define IOMEM 0x40000000
|
---|
28 |
|
---|
29 | typedef int (*plic_irq_callback_t)(void *ctx);
|
---|
30 |
|
---|
31 | typedef struct
|
---|
32 | {
|
---|
33 | union
|
---|
34 | {
|
---|
35 | uint64_t reg;
|
---|
36 | struct
|
---|
37 | {
|
---|
38 | uint64_t int_en : 1;
|
---|
39 | uint64_t ram_flag : 1;
|
---|
40 | uint64_t full_add : 1;
|
---|
41 | uint64_t depth_wise_layer : 1;
|
---|
42 | uint64_t reserved : 60;
|
---|
43 | } data;
|
---|
44 | } interrupt_enabe;
|
---|
45 |
|
---|
46 | union
|
---|
47 | {
|
---|
48 | uint64_t reg;
|
---|
49 | struct
|
---|
50 | {
|
---|
51 | uint64_t image_src_addr : 15;
|
---|
52 | uint64_t reserved0 : 17;
|
---|
53 | uint64_t image_dst_addr : 15;
|
---|
54 | uint64_t reserved1 : 17;
|
---|
55 | } data;
|
---|
56 | } image_addr;
|
---|
57 |
|
---|
58 | union
|
---|
59 | {
|
---|
60 | uint64_t reg;
|
---|
61 | struct
|
---|
62 | {
|
---|
63 | uint64_t i_ch_num : 10;
|
---|
64 | uint64_t reserved0 : 22;
|
---|
65 | uint64_t o_ch_num : 10;
|
---|
66 | uint64_t reserved1 : 6;
|
---|
67 | uint64_t o_ch_num_coef : 10;
|
---|
68 | uint64_t reserved2 : 6;
|
---|
69 | } data;
|
---|
70 | } image_channel_num;
|
---|
71 |
|
---|
72 | union
|
---|
73 | {
|
---|
74 | uint64_t reg;
|
---|
75 | struct
|
---|
76 | {
|
---|
77 | uint64_t i_row_wid : 10;
|
---|
78 | uint64_t i_col_high : 9;
|
---|
79 | uint64_t reserved0 : 13;
|
---|
80 | uint64_t o_row_wid : 10;
|
---|
81 | uint64_t o_col_high : 9;
|
---|
82 | uint64_t reserved1 : 13;
|
---|
83 | } data;
|
---|
84 | } image_size;
|
---|
85 |
|
---|
86 | union
|
---|
87 | {
|
---|
88 | uint64_t reg;
|
---|
89 | struct
|
---|
90 | {
|
---|
91 | uint64_t kernel_type : 3;
|
---|
92 | uint64_t pad_type : 1;
|
---|
93 | uint64_t pool_type : 4;
|
---|
94 | uint64_t first_stride : 1;
|
---|
95 | uint64_t bypass_conv : 1;
|
---|
96 | uint64_t load_para : 1;
|
---|
97 | uint64_t reserved0 : 5;
|
---|
98 | uint64_t dma_burst_size : 8;
|
---|
99 | uint64_t pad_value : 8;
|
---|
100 | uint64_t bwsx_base_addr : 32;
|
---|
101 | } data;
|
---|
102 | } kernel_pool_type_cfg;
|
---|
103 |
|
---|
104 | union
|
---|
105 | {
|
---|
106 | uint64_t reg;
|
---|
107 | struct
|
---|
108 | {
|
---|
109 | uint64_t load_coor : 1;
|
---|
110 | uint64_t load_time : 6;
|
---|
111 | uint64_t reserved0 : 8;
|
---|
112 | uint64_t para_size : 17;
|
---|
113 | uint64_t para_start_addr : 32;
|
---|
114 | } data;
|
---|
115 | } kernel_load_cfg;
|
---|
116 |
|
---|
117 | union
|
---|
118 | {
|
---|
119 | uint64_t reg;
|
---|
120 | struct
|
---|
121 | {
|
---|
122 | uint64_t coef_column_offset : 4;
|
---|
123 | uint64_t coef_row_offset : 12;
|
---|
124 | uint64_t reserved0 : 48;
|
---|
125 | } data;
|
---|
126 | } kernel_offset;
|
---|
127 |
|
---|
128 | union
|
---|
129 | {
|
---|
130 | uint64_t reg;
|
---|
131 | struct
|
---|
132 | {
|
---|
133 | uint64_t channel_switch_addr : 15;
|
---|
134 | uint64_t reserved : 1;
|
---|
135 | uint64_t row_switch_addr : 4;
|
---|
136 | uint64_t coef_size : 8;
|
---|
137 | uint64_t coef_group : 3;
|
---|
138 | uint64_t load_act : 1;
|
---|
139 | uint64_t active_addr : 32;
|
---|
140 | } data;
|
---|
141 | } kernel_calc_type_cfg;
|
---|
142 |
|
---|
143 | union
|
---|
144 | {
|
---|
145 | uint64_t reg;
|
---|
146 | struct
|
---|
147 | {
|
---|
148 | uint64_t wb_channel_switch_addr : 15;
|
---|
149 | uint64_t reserved0 : 1;
|
---|
150 | uint64_t wb_row_switch_addr : 4;
|
---|
151 | uint64_t wb_group : 3;
|
---|
152 | uint64_t reserved1 : 41;
|
---|
153 | } data;
|
---|
154 | } write_back_cfg;
|
---|
155 |
|
---|
156 | union
|
---|
157 | {
|
---|
158 | uint64_t reg;
|
---|
159 | struct
|
---|
160 | {
|
---|
161 | uint64_t shr_w : 4;
|
---|
162 | uint64_t shr_x : 4;
|
---|
163 | uint64_t arg_w : 24;
|
---|
164 | uint64_t arg_x : 24;
|
---|
165 | uint64_t reserved0 : 8;
|
---|
166 | } data;
|
---|
167 | } conv_value;
|
---|
168 |
|
---|
169 | union
|
---|
170 | {
|
---|
171 | uint64_t reg;
|
---|
172 | struct
|
---|
173 | {
|
---|
174 | uint64_t arg_add : 40;
|
---|
175 | uint64_t reserved : 24;
|
---|
176 | } data;
|
---|
177 | } conv_value2;
|
---|
178 |
|
---|
179 | union
|
---|
180 | {
|
---|
181 | uint64_t reg;
|
---|
182 | struct
|
---|
183 | {
|
---|
184 | uint64_t send_data_out : 1;
|
---|
185 | uint64_t reserved : 15;
|
---|
186 | uint64_t channel_byte_num : 16;
|
---|
187 | uint64_t dma_total_byte : 32;
|
---|
188 | } data;
|
---|
189 | } dma_parameter;
|
---|
190 | } kpu_layer_argument_t;
|
---|
191 |
|
---|
192 | typedef struct
|
---|
193 | {
|
---|
194 | union
|
---|
195 | {
|
---|
196 | uint64_t reg;
|
---|
197 | struct
|
---|
198 | {
|
---|
199 | uint64_t shift_number : 8;
|
---|
200 | uint64_t y_mul : 16;
|
---|
201 | uint64_t x_start : 36;
|
---|
202 | } data;
|
---|
203 | } activate_para[16];
|
---|
204 |
|
---|
205 | union
|
---|
206 | {
|
---|
207 | uint64_t reg;
|
---|
208 | struct
|
---|
209 | {
|
---|
210 | uint8_t result_bias[8];
|
---|
211 | } data;
|
---|
212 | } activate_para_bias0;
|
---|
213 |
|
---|
214 | union
|
---|
215 | {
|
---|
216 | uint64_t reg;
|
---|
217 | struct
|
---|
218 | {
|
---|
219 | uint8_t result_bias[8];
|
---|
220 | } data;
|
---|
221 | } activate_para_bias1;
|
---|
222 | } kpu_activate_table_t;
|
---|
223 |
|
---|
224 | typedef struct
|
---|
225 | {
|
---|
226 | union
|
---|
227 | {
|
---|
228 | uint64_t reg;
|
---|
229 | struct
|
---|
230 | {
|
---|
231 | uint64_t norm_mul : 24;
|
---|
232 | uint64_t norm_add : 32;
|
---|
233 | uint64_t norm_shift : 4;
|
---|
234 | } data;
|
---|
235 | } batchnorm;
|
---|
236 | } kpu_batchnorm_argument_t;
|
---|
237 |
|
---|
238 | typedef struct
|
---|
239 | {
|
---|
240 | union
|
---|
241 | {
|
---|
242 | uint64_t reg;
|
---|
243 | struct
|
---|
244 | {
|
---|
245 | uint16_t weight[9];
|
---|
246 | } data;
|
---|
247 | } weights;
|
---|
248 | } kpu_weights_kernel_16_3x3_t;
|
---|
249 |
|
---|
250 | typedef struct
|
---|
251 | {
|
---|
252 | uint64_t calc_done_int : 1;
|
---|
253 | uint64_t layer_cfg_almost_empty_int : 1;
|
---|
254 | uint64_t layer_cfg_almost_full_int : 1;
|
---|
255 | uint64_t reserved : 61;
|
---|
256 | } kpu_config_interrupt_t;
|
---|
257 |
|
---|
258 | typedef struct
|
---|
259 | {
|
---|
260 | uint64_t fifo_full_threshold : 4;
|
---|
261 | uint64_t fifo_empty_threshold : 4;
|
---|
262 | uint64_t reserved : 56;
|
---|
263 | } kpu_config_fifo_threshold_t;
|
---|
264 |
|
---|
265 | typedef struct
|
---|
266 | {
|
---|
267 | uint64_t dma_fifo_flush_n : 1;
|
---|
268 | uint64_t gs_fifo_flush_n : 1;
|
---|
269 | uint64_t cfg_fifo_flush_n : 1;
|
---|
270 | uint64_t cmd_fifo_flush_n : 1;
|
---|
271 | uint64_t resp_fifo_flush_n : 1;
|
---|
272 | uint64_t reserved : 59;
|
---|
273 | } kpu_config_fifo_ctrl_t;
|
---|
274 |
|
---|
275 | typedef struct
|
---|
276 | {
|
---|
277 | uint64_t eight_bit_mode : 1;
|
---|
278 | uint64_t reserved : 63;
|
---|
279 | } kpu_config_eight_bit_mode_t;
|
---|
280 |
|
---|
281 | typedef struct
|
---|
282 | {
|
---|
283 | volatile uint64_t layer_argument_fifo;
|
---|
284 |
|
---|
285 | volatile union
|
---|
286 | {
|
---|
287 | uint64_t reg;
|
---|
288 | kpu_config_interrupt_t data;
|
---|
289 | } interrupt_status;
|
---|
290 |
|
---|
291 | volatile union
|
---|
292 | {
|
---|
293 | uint64_t reg;
|
---|
294 | kpu_config_interrupt_t data;
|
---|
295 | } interrupt_raw;
|
---|
296 |
|
---|
297 | volatile union
|
---|
298 | {
|
---|
299 | uint64_t reg;
|
---|
300 | kpu_config_interrupt_t data;
|
---|
301 | } interrupt_mask;
|
---|
302 |
|
---|
303 | volatile union
|
---|
304 | {
|
---|
305 | uint64_t reg;
|
---|
306 | kpu_config_interrupt_t data;
|
---|
307 | } interrupt_clear;
|
---|
308 |
|
---|
309 | volatile union
|
---|
310 | {
|
---|
311 | uint64_t reg;
|
---|
312 | kpu_config_fifo_threshold_t data;
|
---|
313 | } fifo_threshold;
|
---|
314 |
|
---|
315 | volatile uint64_t fifo_data_out;
|
---|
316 |
|
---|
317 | volatile union
|
---|
318 | {
|
---|
319 | uint64_t reg;
|
---|
320 | kpu_config_fifo_ctrl_t data;
|
---|
321 | } fifo_ctrl;
|
---|
322 |
|
---|
323 | volatile union
|
---|
324 | {
|
---|
325 | uint64_t reg;
|
---|
326 | kpu_config_eight_bit_mode_t data;
|
---|
327 | } eight_bit_mode;
|
---|
328 | } kpu_config_t;
|
---|
329 |
|
---|
330 | #define dmac_channel_number_t int
|
---|
331 |
|
---|
332 | typedef struct
|
---|
333 | {
|
---|
334 | kpu_layer_argument_t *layers;
|
---|
335 | kpu_layer_argument_t *remain_layers;
|
---|
336 | plic_irq_callback_t callback;
|
---|
337 | void *ctx;
|
---|
338 | uint64_t *src;
|
---|
339 | uint64_t *dst;
|
---|
340 | uint32_t src_length;
|
---|
341 | uint32_t dst_length;
|
---|
342 | uint32_t layers_length;
|
---|
343 | uint32_t remain_layers_length;
|
---|
344 | dmac_channel_number_t dma_ch;
|
---|
345 | uint32_t eight_bit_mode;
|
---|
346 | float output_scale;
|
---|
347 | float output_bias;
|
---|
348 | float input_scale;
|
---|
349 | float input_bias;
|
---|
350 | } kpu_task_t;
|
---|
351 |
|
---|
352 | typedef struct
|
---|
353 | {
|
---|
354 | uint32_t version;
|
---|
355 | uint32_t flags;
|
---|
356 | uint32_t arch;
|
---|
357 | uint32_t layers_length;
|
---|
358 | uint32_t max_start_address;
|
---|
359 | uint32_t main_mem_usage;
|
---|
360 | uint32_t output_count;
|
---|
361 | } kpu_kmodel_header_t;
|
---|
362 |
|
---|
363 | typedef struct
|
---|
364 | {
|
---|
365 | uint32_t version;
|
---|
366 | uint32_t flags;
|
---|
367 | uint32_t layers_length;
|
---|
368 | uint32_t max_start_address;
|
---|
369 | uint32_t layers_argument_start;
|
---|
370 | } kpu_model_header_t;
|
---|
371 |
|
---|
372 | typedef struct
|
---|
373 | {
|
---|
374 | uint32_t address;
|
---|
375 | uint32_t size;
|
---|
376 | } kpu_model_output_t;
|
---|
377 |
|
---|
378 | typedef enum
|
---|
379 | {
|
---|
380 | KL_INVALID = 0,
|
---|
381 | KL_ADD,
|
---|
382 | KL_QUANTIZED_ADD,
|
---|
383 | KL_GLOBAL_MAX_POOL2D,
|
---|
384 | KL_QUANTIZED_GLOBAL_MAX_POOL2D,
|
---|
385 | KL_GLOBAL_AVERAGE_POOL2D,
|
---|
386 | KL_QUANTIZED_GLOBAL_AVERAGE_POOL2D,
|
---|
387 | KL_MAX_POOL2D,
|
---|
388 | KL_QUANTIZED_MAX_POOL2D,
|
---|
389 | KL_AVERAGE_POOL2D,
|
---|
390 | KL_QUANTIZED_AVERAGE_POOL2D,
|
---|
391 | KL_QUANTIZE,
|
---|
392 | KL_DEQUANTIZE,
|
---|
393 | KL_REQUANTIZE,
|
---|
394 | KL_L2_NORMALIZATION,
|
---|
395 | KL_SOFTMAX,
|
---|
396 | KL_CONCAT,
|
---|
397 | KL_QUANTIZED_CONCAT,
|
---|
398 | KL_FULLY_CONNECTED,
|
---|
399 | KL_QUANTIZED_FULLY_CONNECTED,
|
---|
400 | KL_TENSORFLOW_FLATTEN,
|
---|
401 | KL_QUANTIZED_TENSORFLOW_FLATTEN,
|
---|
402 | KL_RESIZE_NEAREST_NEIGHBOR,
|
---|
403 | KL_QUANTIZED_RESIZE_NEAREST_NEIGHBOR,
|
---|
404 | KL_CHANNELWISE_DEQUANTIZE,
|
---|
405 | KL_LOGISTIC,
|
---|
406 | KL_K210_CONV = 10240,
|
---|
407 | KL_K210_ADD_PADDING,
|
---|
408 | KL_K210_REMOVE_PADDING,
|
---|
409 | KL_K210_UPLOAD
|
---|
410 | } kpu_model_layer_type_t;
|
---|
411 |
|
---|
412 | typedef struct
|
---|
413 | {
|
---|
414 | uint32_t type;
|
---|
415 | uint32_t body_size;
|
---|
416 | } kpu_model_layer_header_t;
|
---|
417 |
|
---|
418 | typedef enum
|
---|
419 | {
|
---|
420 | KLF_NONE = 0,
|
---|
421 | KLF_MAIN_MEM_OUT = 1
|
---|
422 | } kpu_model_layer_flags_t;
|
---|
423 |
|
---|
424 | typedef enum
|
---|
425 | {
|
---|
426 | KLP_SAME = 0,
|
---|
427 | KLP_VALID = 1
|
---|
428 | } kpu_model_padding_t;
|
---|
429 |
|
---|
430 | typedef enum
|
---|
431 | {
|
---|
432 | KLA_LINEAR = 0,
|
---|
433 | KLA_RELU = 1,
|
---|
434 | KLA_RELU6 = 2
|
---|
435 | } kpu_model_activation_t;
|
---|
436 |
|
---|
437 | typedef struct
|
---|
438 | {
|
---|
439 | float scale;
|
---|
440 | float bias;
|
---|
441 | } kpu_model_quant_param_t;
|
---|
442 |
|
---|
443 | typedef struct
|
---|
444 | {
|
---|
445 | uint32_t width;
|
---|
446 | uint32_t height;
|
---|
447 | uint32_t channels;
|
---|
448 | } kpu_model_shape_t;
|
---|
449 |
|
---|
450 | typedef struct
|
---|
451 | {
|
---|
452 | uint32_t start;
|
---|
453 | uint32_t size;
|
---|
454 | } kpu_model_memory_range_t;
|
---|
455 |
|
---|
456 | typedef struct
|
---|
457 | {
|
---|
458 | uint32_t flags;
|
---|
459 | uint32_t main_mem_out_address;
|
---|
460 | uint32_t layer_offset;
|
---|
461 | uint32_t weights_offset;
|
---|
462 | uint32_t bn_offset;
|
---|
463 | uint32_t act_offset;
|
---|
464 | } kpu_model_conv_layer_argument_t;
|
---|
465 |
|
---|
466 | typedef struct
|
---|
467 | {
|
---|
468 | uint32_t flags;
|
---|
469 | uint32_t main_mem_in_a_address;
|
---|
470 | uint32_t main_mem_in_b_address;
|
---|
471 | uint32_t main_mem_out_address;
|
---|
472 | uint32_t count;
|
---|
473 | } kpu_model_add_layer_argument_t;
|
---|
474 |
|
---|
475 | typedef struct
|
---|
476 | {
|
---|
477 | uint32_t flags;
|
---|
478 | uint32_t main_mem_in_a_address;
|
---|
479 | uint32_t main_mem_in_b_address;
|
---|
480 | uint32_t main_mem_out_address;
|
---|
481 | uint32_t count;
|
---|
482 | int32_t in_a_offset;
|
---|
483 | int32_t in_a_mul;
|
---|
484 | int32_t in_a_shift;
|
---|
485 | int32_t in_b_offset;
|
---|
486 | int32_t in_b_mul;
|
---|
487 | int32_t in_b_shift;
|
---|
488 | int32_t out_offset;
|
---|
489 | int32_t out_mul;
|
---|
490 | int32_t out_shift;
|
---|
491 | } kpu_model_quant_add_layer_argument_t;
|
---|
492 |
|
---|
493 | typedef struct
|
---|
494 | {
|
---|
495 | uint32_t flags;
|
---|
496 | uint32_t main_mem_in_address;
|
---|
497 | uint32_t main_mem_out_address;
|
---|
498 | uint32_t kernel_size;
|
---|
499 | uint32_t channels;
|
---|
500 | } kpu_model_gap2d_layer_argument_t;
|
---|
501 |
|
---|
502 | typedef struct
|
---|
503 | {
|
---|
504 | uint32_t flags;
|
---|
505 | uint32_t main_mem_in_address;
|
---|
506 | uint32_t main_mem_out_address;
|
---|
507 | kpu_model_shape_t in_shape;
|
---|
508 | kpu_model_shape_t out_shape;
|
---|
509 | uint32_t kernel_width;
|
---|
510 | uint32_t kernel_height;
|
---|
511 | uint32_t stride_width;
|
---|
512 | uint32_t stride_height;
|
---|
513 | uint32_t padding_width;
|
---|
514 | uint32_t padding_height;
|
---|
515 | } kpu_model_quant_max_pool2d_layer_argument_t;
|
---|
516 |
|
---|
517 | typedef struct
|
---|
518 | {
|
---|
519 | uint32_t flags;
|
---|
520 | uint32_t main_mem_in_address;
|
---|
521 | uint32_t main_mem_out_address;
|
---|
522 | kpu_model_shape_t in_shape;
|
---|
523 | kpu_model_shape_t out_shape;
|
---|
524 | uint32_t kernel_width;
|
---|
525 | uint32_t kernel_height;
|
---|
526 | uint32_t stride_width;
|
---|
527 | uint32_t stride_height;
|
---|
528 | uint32_t padding_width;
|
---|
529 | uint32_t padding_height;
|
---|
530 | kpu_model_activation_t act;
|
---|
531 | } kpu_model_ave_pool2d_layer_argument_t;
|
---|
532 |
|
---|
533 | typedef struct
|
---|
534 | {
|
---|
535 | uint32_t flags;
|
---|
536 | uint32_t main_mem_in_address;
|
---|
537 | uint32_t mem_out_address;
|
---|
538 | uint32_t count;
|
---|
539 | kpu_model_quant_param_t quant_param;
|
---|
540 | } kpu_model_quantize_layer_argument_t;
|
---|
541 |
|
---|
542 | typedef struct
|
---|
543 | {
|
---|
544 | uint32_t flags;
|
---|
545 | uint32_t main_mem_in_address;
|
---|
546 | uint32_t main_mem_out_address;
|
---|
547 | uint32_t count;
|
---|
548 | kpu_model_quant_param_t quant_param;
|
---|
549 | } kpu_model_dequantize_layer_argument_t;
|
---|
550 |
|
---|
551 | typedef struct
|
---|
552 | {
|
---|
553 | uint32_t flags;
|
---|
554 | uint32_t main_mem_in_address;
|
---|
555 | uint32_t main_mem_out_address;
|
---|
556 | uint32_t count;
|
---|
557 | uint8_t table[256];
|
---|
558 | } kpu_model_requantize_layer_argument_t;
|
---|
559 |
|
---|
560 | typedef struct
|
---|
561 | {
|
---|
562 | uint32_t flags;
|
---|
563 | uint32_t main_mem_in_address;
|
---|
564 | uint32_t kpu_mem_out_address;
|
---|
565 | uint32_t channels;
|
---|
566 | } kpu_model_add_padding_layer_argument_t;
|
---|
567 |
|
---|
568 | typedef struct
|
---|
569 | {
|
---|
570 | uint32_t flags;
|
---|
571 | uint32_t main_mem_in_address;
|
---|
572 | uint32_t main_mem_out_address;
|
---|
573 | uint32_t channels;
|
---|
574 | } kpu_model_remove_padding_layer_argument_t;
|
---|
575 |
|
---|
576 | typedef struct
|
---|
577 | {
|
---|
578 | uint32_t flags;
|
---|
579 | uint32_t main_mem_in_address;
|
---|
580 | uint32_t kpu_mem_out_address;
|
---|
581 | uint32_t width;
|
---|
582 | uint32_t height;
|
---|
583 | uint32_t channels;
|
---|
584 | } kpu_model_upload_layer_argument_t;
|
---|
585 |
|
---|
586 | typedef struct
|
---|
587 | {
|
---|
588 | uint32_t flags;
|
---|
589 | uint32_t main_mem_in_address;
|
---|
590 | uint32_t main_mem_out_address;
|
---|
591 | uint32_t channels;
|
---|
592 | } kpu_model_l2_norm_layer_argument_t;
|
---|
593 |
|
---|
594 | typedef struct
|
---|
595 | {
|
---|
596 | uint32_t flags;
|
---|
597 | uint32_t main_mem_in_address;
|
---|
598 | uint32_t main_mem_out_address;
|
---|
599 | uint32_t channels;
|
---|
600 | } kpu_model_softmax_layer_argument_t;
|
---|
601 |
|
---|
602 | typedef struct
|
---|
603 | {
|
---|
604 | uint32_t flags;
|
---|
605 | uint32_t main_mem_out_address;
|
---|
606 | uint32_t input_count;
|
---|
607 | kpu_model_memory_range_t inputs_mem[0];
|
---|
608 | } kpu_model_concat_layer_argument_t;
|
---|
609 |
|
---|
610 | typedef struct
|
---|
611 | {
|
---|
612 | uint32_t flags;
|
---|
613 | uint32_t main_mem_in_address;
|
---|
614 | uint32_t main_mem_out_address;
|
---|
615 | uint32_t in_channels;
|
---|
616 | uint32_t out_channels;
|
---|
617 | kpu_model_activation_t act;
|
---|
618 | float weights[0];
|
---|
619 | } kpu_model_fully_connected_layer_argument_t;
|
---|
620 |
|
---|
621 | typedef struct
|
---|
622 | {
|
---|
623 | uint32_t flags;
|
---|
624 | uint32_t main_mem_in_address;
|
---|
625 | uint32_t main_mem_out_address;
|
---|
626 | kpu_model_shape_t shape;
|
---|
627 | } kpu_model_tf_flatten_layer_argument_t;
|
---|
628 |
|
---|
629 | typedef struct
|
---|
630 | {
|
---|
631 | uint32_t flags;
|
---|
632 | uint32_t main_mem_in_address;
|
---|
633 | uint32_t main_mem_out_address;
|
---|
634 | kpu_model_shape_t in_shape;
|
---|
635 | uint32_t out_width;
|
---|
636 | uint32_t out_height;
|
---|
637 | uint32_t align_corners;
|
---|
638 | } kpu_model_resize_nearest_neighbor_layer_argument_t;
|
---|
639 |
|
---|
640 | typedef struct
|
---|
641 | {
|
---|
642 | uint32_t flags;
|
---|
643 | uint32_t main_mem_in_address;
|
---|
644 | uint32_t main_mem_out_address;
|
---|
645 | kpu_model_shape_t in_shape;
|
---|
646 | uint32_t out_width;
|
---|
647 | uint32_t out_height;
|
---|
648 | uint32_t align_corners;
|
---|
649 | } kpu_model_quant_resize_nearest_neighbor_layer_argument_t;
|
---|
650 |
|
---|
651 | typedef struct
|
---|
652 | {
|
---|
653 | uint32_t flags;
|
---|
654 | uint32_t main_mem_in_address;
|
---|
655 | uint32_t main_mem_out_address;
|
---|
656 | uint32_t channels;
|
---|
657 | uint32_t channel_size;
|
---|
658 | kpu_model_quant_param_t quant_params[0];
|
---|
659 | } kpu_model_channelwise_dequant_argument_t;
|
---|
660 |
|
---|
661 | typedef struct
|
---|
662 | {
|
---|
663 | uint32_t flags;
|
---|
664 | uint32_t main_mem_in_address;
|
---|
665 | uint32_t main_mem_out_address;
|
---|
666 | uint32_t channels;
|
---|
667 | } kpu_model_logistic_layer_argument_t;
|
---|
668 |
|
---|
669 | typedef void (*kpu_done_callback_t)(void *userdata);
|
---|
670 |
|
---|
671 | typedef struct
|
---|
672 | {
|
---|
673 | int is_nncase;
|
---|
674 |
|
---|
675 | union
|
---|
676 | {
|
---|
677 | struct
|
---|
678 | {
|
---|
679 | const uint8_t *model_buffer;
|
---|
680 | uint8_t *main_buffer;
|
---|
681 | uint32_t output_count;
|
---|
682 | const kpu_model_output_t *outputs;
|
---|
683 | const kpu_model_layer_header_t *layer_headers;
|
---|
684 | const uint8_t *body_start;
|
---|
685 | uint32_t layers_length;
|
---|
686 | volatile uint32_t current_layer;
|
---|
687 | const uint8_t *volatile current_body;
|
---|
688 | dmac_channel_number_t dma_ch;
|
---|
689 | kpu_done_callback_t done_callback;
|
---|
690 | void *userdata;
|
---|
691 | };
|
---|
692 |
|
---|
693 | struct
|
---|
694 | {
|
---|
695 | void* nncase_ctx;
|
---|
696 | };
|
---|
697 | };
|
---|
698 | } kpu_model_context_t;
|
---|
699 |
|
---|
700 | typedef struct
|
---|
701 | {
|
---|
702 | uint32_t weigths_offset;
|
---|
703 | uint32_t bn_offset;
|
---|
704 | uint32_t act_offset;
|
---|
705 | float input_scale;
|
---|
706 | float input_bias;
|
---|
707 | float output_scale;
|
---|
708 | float output_bias;
|
---|
709 | } kpu_model_layer_metadata_t;
|
---|
710 |
|
---|
711 | typedef struct _quantize_param
|
---|
712 | {
|
---|
713 | float scale;
|
---|
714 | float bias;
|
---|
715 | } quantize_param_t;
|
---|
716 |
|
---|
717 | extern volatile kpu_config_t *const kpu;
|
---|
718 |
|
---|
719 | /**
|
---|
720 | * @brief Modle complier init kpu handler
|
---|
721 | *
|
---|
722 | * @param[in] task Kpu handler
|
---|
723 | *
|
---|
724 | * @return Kpu handler
|
---|
725 | */
|
---|
726 | extern kpu_task_t *kpu_task_init(kpu_task_t *task);
|
---|
727 |
|
---|
728 | /**
|
---|
729 | * @brief Kpu run for AI
|
---|
730 | *
|
---|
731 | * @param[in] task Kpu handler
|
---|
732 | * @param[in] dma_ch DMA for kpu
|
---|
733 | * @param[in] src The picture data
|
---|
734 | * @param[in] dest The result of kpu
|
---|
735 | * @param[in] callback The callback of kpu
|
---|
736 | *
|
---|
737 | * @return result
|
---|
738 | * - 0 Success
|
---|
739 | * - Other Fail.Kpu is busy.
|
---|
740 | */
|
---|
741 | int kpu_run(kpu_task_t *task, dmac_channel_number_t dma_ch, const void *src, void *dest, plic_irq_callback_t callback);
|
---|
742 |
|
---|
743 | /**
|
---|
744 | * @brief Get kpu result buf
|
---|
745 | *
|
---|
746 | * @param[in] task Kpu handler
|
---|
747 | *
|
---|
748 | * @return Kpu result buf
|
---|
749 | */
|
---|
750 | uint8_t *kpu_get_output_buf(kpu_task_t *task);
|
---|
751 |
|
---|
752 | /**
|
---|
753 | * @brief Release kpu output buf
|
---|
754 | *
|
---|
755 | * @param[in] output_buf Kpu output buf
|
---|
756 | *
|
---|
757 | */
|
---|
758 | void kpu_release_output_buf(uint8_t *output_buf);
|
---|
759 |
|
---|
760 | /**
|
---|
761 | * @brief Kpu run for AI
|
---|
762 | *
|
---|
763 | * @param[in] task Kpu handler
|
---|
764 | *
|
---|
765 | * @return result
|
---|
766 | * - 0 Success
|
---|
767 | * - Other Fail.Kpu is busy.
|
---|
768 | */
|
---|
769 | int kpu_start(kpu_task_t *task);
|
---|
770 |
|
---|
771 | /**
|
---|
772 | * @brief Initialize kpu handler
|
---|
773 | *
|
---|
774 | * @param[in] task Kpu handler
|
---|
775 | *
|
---|
776 | * @return result
|
---|
777 | * - 0 Success
|
---|
778 | * - Other Fail.
|
---|
779 | */
|
---|
780 | int kpu_single_task_init(kpu_task_t *task);
|
---|
781 |
|
---|
782 | /**
|
---|
783 | * @brief Uninitialize kpu handler
|
---|
784 | *
|
---|
785 | * @param[in] task Kpu handler
|
---|
786 | *
|
---|
787 | * @return result
|
---|
788 | * - 0 Success
|
---|
789 | * - Other Fail.
|
---|
790 | */
|
---|
791 | int kpu_single_task_deinit(kpu_task_t *task);
|
---|
792 |
|
---|
793 | /**
|
---|
794 | * @brief Load kmodel and init kpu task
|
---|
795 | *
|
---|
796 | * @param[in] task Kpu handler
|
---|
797 | * @param[in] buffer Kmodel
|
---|
798 | * @param[in] meta Test data
|
---|
799 | *
|
---|
800 | * @return result
|
---|
801 | * - 0 Success
|
---|
802 | * - Other Fail.
|
---|
803 | */
|
---|
804 | int kpu_model_load_from_buffer(kpu_task_t *task, uint8_t *buffer, kpu_model_layer_metadata_t **meta);
|
---|
805 |
|
---|
806 | /**
|
---|
807 | * @brief Kpu initialize
|
---|
808 | *
|
---|
809 | * @param[in] eight_bit_mode 0:16bit mode 1:8bit mode
|
---|
810 | * @param[in] callback Callback of kpu
|
---|
811 | * @param[in] userdata Data of callback
|
---|
812 | *
|
---|
813 | */
|
---|
814 | void kpu_init(int eight_bit_mode, plic_irq_callback_t callback, void *userdata);
|
---|
815 |
|
---|
816 | /**
|
---|
817 | * @brief Kpu input data by dma
|
---|
818 | *
|
---|
819 | * @param[in] layer Kpu task layer
|
---|
820 | * @param[in] src Image data
|
---|
821 | * @param[in] dma_ch Dmac channel
|
---|
822 | * @param[in] callback Dmac complete callback
|
---|
823 | * @param[in] userdata Data of callback
|
---|
824 | *
|
---|
825 | */
|
---|
826 | void kpu_input_dma(const kpu_layer_argument_t *layer, const uint8_t *src, dmac_channel_number_t dma_ch, plic_irq_callback_t callback, void *userdata);
|
---|
827 |
|
---|
828 | /**
|
---|
829 | * @brief Kpu input data by cpu
|
---|
830 | *
|
---|
831 | * @param[in] layer Kpu task layer
|
---|
832 | * @param[in] src Image data
|
---|
833 | * @param[in] width Image width
|
---|
834 | * @param[in] height Image heigth
|
---|
835 | * @param[in] channels Color channel, RGB is 3
|
---|
836 | *
|
---|
837 | */
|
---|
838 | void kpu_input_with_padding(kpu_layer_argument_t *layer, const uint8_t *src, int width, int height, int channels);
|
---|
839 |
|
---|
840 | /**
|
---|
841 | * @brief Kpu run only one layer
|
---|
842 | *
|
---|
843 | * @param[in] layer Kpu task layer
|
---|
844 | *
|
---|
845 | */
|
---|
846 | void kpu_conv2d(kpu_layer_argument_t *layer);
|
---|
847 |
|
---|
848 | /**
|
---|
849 | * @brief Kpu run only one layer then get the result by dma
|
---|
850 | *
|
---|
851 | * @param[in] layer Kpu task layer
|
---|
852 | * @param[in] dma_ch Dmac channel
|
---|
853 | * @param[in] dest Result
|
---|
854 | * @param[in] callback Dmac complete callback
|
---|
855 | * @param[in] userdata Data of callback
|
---|
856 | *
|
---|
857 | */
|
---|
858 | void kpu_conv2d_output(kpu_layer_argument_t *layer, dmac_channel_number_t dma_ch, uint8_t *dest, plic_irq_callback_t callback, void *userdata);
|
---|
859 |
|
---|
860 | /**
|
---|
861 | * @brief Kpu pooling
|
---|
862 | *
|
---|
863 | * @param[in] src Source
|
---|
864 | * @param[in] src_param Source param
|
---|
865 | * @param[in] kernel_size Kernel size, 7*7 is 49
|
---|
866 | * @param[in] channels Channels
|
---|
867 | * @param[in] dest Dest
|
---|
868 | * @param[in] dest_param Dest param
|
---|
869 | *
|
---|
870 | */
|
---|
871 | void kpu_global_average_pool(const uint8_t *src, const quantize_param_t *src_param, int kernel_size, int channels, uint8_t *dest, const quantize_param_t *dest_param);
|
---|
872 |
|
---|
873 | /**
|
---|
874 | * @brief Kpu pooling
|
---|
875 | *
|
---|
876 | * @param[in] src Source
|
---|
877 | * @param[in] src_param Source param
|
---|
878 | * @param[in] kernel_size Kernel size, 7*7 is 49
|
---|
879 | * @param[in] channels Channels
|
---|
880 | * @param[in] dest Dest
|
---|
881 | *
|
---|
882 | */
|
---|
883 | void kpu_global_average_pool_float(const uint8_t *src, const quantize_param_t *src_param, int kernel_size, int channels, float *dest);
|
---|
884 |
|
---|
885 | /**
|
---|
886 | * @brief Kpu fullly connected by cpu
|
---|
887 | *
|
---|
888 | * @param[in] src Source
|
---|
889 | * @param[in] weights Weight
|
---|
890 | * @param[in] biases Biases
|
---|
891 | * @param[in] dest Dest
|
---|
892 | * @param[in] input_channels Input channels
|
---|
893 | * @param[in] output_channels Output channels
|
---|
894 | *
|
---|
895 | */
|
---|
896 | void kpu_fully_connected(const float *src, const float *weights, const float *biases, float *dest, int input_channels, int output_channels);
|
---|
897 |
|
---|
898 | /**
|
---|
899 | * @brief Kpu matrix multiplication
|
---|
900 | *
|
---|
901 | * @param[in] src Source
|
---|
902 | * @param[in] channels Channels
|
---|
903 | * @param[in] dest Dest
|
---|
904 | * @param[in] dest_param Dest param
|
---|
905 | *
|
---|
906 | */
|
---|
907 | void kpu_matmul_end(const uint8_t *src, int channels, float *dest, const quantize_param_t *dest_param);
|
---|
908 |
|
---|
909 | /**
|
---|
910 | * @brief Kpu dequantize
|
---|
911 | *
|
---|
912 | * @param[in] src Source
|
---|
913 | * @param[in] src_param Source param
|
---|
914 | * @param[in] count Dequantize count
|
---|
915 | * @param[in] dest Dest
|
---|
916 | *
|
---|
917 | */
|
---|
918 | void kpu_dequantize(const uint8_t *src, const quantize_param_t *src_param, size_t count, float *dest);
|
---|
919 |
|
---|
920 | /**
|
---|
921 | * @brief Kpu load kmodel
|
---|
922 | *
|
---|
923 | * @param[in] ctx Kmodel object
|
---|
924 | * @param[in] buffer Kmodel buffer
|
---|
925 | *
|
---|
926 | * @return result
|
---|
927 | * - 0 Success
|
---|
928 | * - Other Fail.
|
---|
929 | */
|
---|
930 | int kpu_load_kmodel(kpu_model_context_t *ctx, const uint8_t *buffer);
|
---|
931 |
|
---|
932 | /**
|
---|
933 | * @brief Kpu free kmodel buffer
|
---|
934 | *
|
---|
935 | * @param[in] ctx kmodel object
|
---|
936 | *
|
---|
937 | */
|
---|
938 | void kpu_model_free(kpu_model_context_t *ctx);
|
---|
939 |
|
---|
940 | /**
|
---|
941 | * @brief Kpu load kmodel
|
---|
942 | *
|
---|
943 | * @param[in] ctx Kmodel object
|
---|
944 | * @param[in] index Output index
|
---|
945 | * @param[in] data Output data
|
---|
946 | * @param[in] size Output data size
|
---|
947 | *
|
---|
948 | * @return result
|
---|
949 | * - 0 Success
|
---|
950 | * - Other Fail.
|
---|
951 | */
|
---|
952 | int kpu_get_output(kpu_model_context_t *ctx, uint32_t index, uint8_t **data, size_t *size);
|
---|
953 |
|
---|
954 | /**
|
---|
955 | * @brief Kpu run kmodel
|
---|
956 | *
|
---|
957 | * @param[in] ctx Kmodel object
|
---|
958 | * @param[in] src Source data
|
---|
959 | * @param[in] dma_ch Dma channel
|
---|
960 | * @param[in] done_callback Kpu complete callback
|
---|
961 | * @param[in] userdata Data of callback
|
---|
962 | *
|
---|
963 | * @return result
|
---|
964 | * - 0 Success
|
---|
965 | * - Other Fail.
|
---|
966 | */
|
---|
967 | int kpu_run_kmodel(kpu_model_context_t *ctx, const uint8_t *src, dmac_channel_number_t dma_ch, kpu_done_callback_t done_callback, void *userdata);
|
---|
968 |
|
---|
969 | #ifdef __cplusplus
|
---|
970 | }
|
---|
971 | #endif
|
---|
972 |
|
---|
973 | #endif
|
---|