Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 1 | /* |
Rajasekaran Kalidoss | 2fa2bf0 | 2024-04-22 09:11:02 +0200 | [diff] [blame] | 2 | * SPDX-FileCopyrightText: Copyright 2019-2024 Arm Limited and/or its affiliates <open-source-office@arm.com> |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 3 | * SPDX-License-Identifier: Apache-2.0 |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the License); you may |
| 6 | * not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at |
| 8 | * |
| 9 | * www.apache.org/licenses/LICENSE-2.0 |
| 10 | * |
| 11 | * Unless required by applicable law or agreed to in writing, software |
| 12 | * distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| 13 | * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | * See the License for the specific language governing permissions and |
| 15 | * limitations under the License. |
| 16 | */ |
| 17 | |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 18 | /****************************************************************************** |
| 19 | * Includes |
| 20 | ******************************************************************************/ |
| 21 | |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 22 | #include "ethosu_driver.h" |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 23 | #include "ethosu_device.h" |
Anton Moberg | 6eab40b | 2021-07-07 11:43:51 +0200 | [diff] [blame] | 24 | #include "ethosu_log.h" |
Per Åstrand | 25d78c0 | 2020-04-21 14:19:44 +0200 | [diff] [blame] | 25 | |
Kristofer Jonsson | 1c0e7ae | 2022-05-12 11:35:33 +0200 | [diff] [blame] | 26 | #ifdef ETHOSU55 |
| 27 | #include "ethosu_config_u55.h" |
| 28 | #else |
| 29 | #include "ethosu_config_u65.h" |
| 30 | #endif |
| 31 | |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 32 | #include <assert.h> |
Per Åstrand | 25d78c0 | 2020-04-21 14:19:44 +0200 | [diff] [blame] | 33 | #include <cmsis_compiler.h> |
Per Åstrand | 14ccfee | 2020-09-25 10:40:20 +0200 | [diff] [blame] | 34 | #include <inttypes.h> |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 35 | #include <stdbool.h> |
Bhavik Patel | bf7ae63 | 2020-06-11 21:00:16 +0200 | [diff] [blame] | 36 | #include <stddef.h> |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 37 | #include <stdio.h> |
| 38 | #include <stdlib.h> |
| 39 | |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 40 | /****************************************************************************** |
| 41 | * Defines |
| 42 | ******************************************************************************/ |
| 43 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 44 | #define UNUSED(x) ((void)x) |
| 45 | |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 46 | #define BYTES_IN_32_BITS 4 |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 47 | #define MASK_16_BYTE_ALIGN (0xF) |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 48 | #define OPTIMIZER_CONFIG_LENGTH_32_BIT_WORD 2 |
| 49 | #define DRIVER_ACTION_LENGTH_32_BIT_WORD 1 |
| 50 | #define ETHOSU_FOURCC ('1' << 24 | 'P' << 16 | 'O' << 8 | 'C') // "Custom Operator Payload 1" |
| 51 | |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 52 | #define FAST_MEMORY_BASE_ADDR_INDEX 2 |
| 53 | |
| 54 | /****************************************************************************** |
| 55 | * Types |
| 56 | ******************************************************************************/ |
| 57 | |
| 58 | // Driver actions |
| 59 | enum DRIVER_ACTION_e |
| 60 | { |
| 61 | RESERVED = 0, |
| 62 | OPTIMIZER_CONFIG = 1, |
| 63 | COMMAND_STREAM = 2, |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 64 | NOP = 5, |
| 65 | }; |
| 66 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 67 | // Custom operator payload data struct |
| 68 | struct cop_data_s |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 69 | { |
| 70 | union |
| 71 | { |
| 72 | // Driver action data |
| 73 | struct |
| 74 | { |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 75 | uint8_t driver_action_command; // (valid values in DRIVER_ACTION_e) |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 76 | uint8_t reserved; |
| 77 | |
| 78 | // Driver action data |
| 79 | union |
| 80 | { |
| 81 | // DA_CMD_OPT_CFG |
| 82 | struct |
| 83 | { |
| 84 | uint16_t rel_nbr : 4; |
| 85 | uint16_t patch_nbr : 4; |
| 86 | uint16_t opt_cfg_reserved : 8; |
| 87 | }; |
| 88 | |
| 89 | // DA_CMD_CMSTRM |
| 90 | struct |
| 91 | { |
| 92 | uint16_t length; |
| 93 | }; |
| 94 | |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 95 | uint16_t driver_action_data; |
| 96 | }; |
| 97 | }; |
| 98 | |
| 99 | uint32_t word; |
| 100 | }; |
| 101 | }; |
| 102 | |
| 103 | // optimizer config struct |
| 104 | struct opt_cfg_s |
| 105 | { |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 106 | struct cop_data_s da_data; |
| 107 | uint32_t cfg; |
| 108 | uint32_t id; |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 109 | }; |
| 110 | |
| 111 | /****************************************************************************** |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 112 | * Variables |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 113 | ******************************************************************************/ |
| 114 | |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 115 | // Registered drivers linked list HEAD |
| 116 | static struct ethosu_driver *registered_drivers = NULL; |
| 117 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 118 | /****************************************************************************** |
| 119 | * Weak functions - Cache |
| 120 | * |
| 121 | * Default NOP operations. Override if available on the targeted device. |
| 122 | ******************************************************************************/ |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 123 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 124 | /* |
| 125 | * Flush/clean the data cache by address and size. Passing NULL as p argument |
| 126 | * expects the whole cache to be flushed. |
| 127 | */ |
| 128 | void __attribute__((weak)) ethosu_flush_dcache(uint32_t *p, size_t bytes) |
| 129 | { |
| 130 | UNUSED(p); |
| 131 | UNUSED(bytes); |
| 132 | } |
| 133 | |
| 134 | /* |
| 135 | * Invalidate the data cache by address and size. Passing NULL as p argument |
| 136 | * expects the whole cache to be invalidated. |
| 137 | */ |
| 138 | void __attribute__((weak)) ethosu_invalidate_dcache(uint32_t *p, size_t bytes) |
| 139 | { |
| 140 | UNUSED(p); |
| 141 | UNUSED(bytes); |
| 142 | } |
| 143 | |
| 144 | /****************************************************************************** |
| 145 | * Weak functions - Semaphore/Mutex for multi NPU |
| 146 | * |
| 147 | * Following section handles the minimal sempahore and mutex implementation in |
| 148 | * case of baremetal applications. Weak symbols will be overridden by RTOS |
| 149 | * definitions and implement true thread-safety (in application layer). |
| 150 | ******************************************************************************/ |
| 151 | |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 152 | struct ethosu_semaphore_t |
| 153 | { |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 154 | uint8_t count; |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 155 | }; |
| 156 | |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 157 | static void *ethosu_mutex; |
| 158 | static void *ethosu_semaphore; |
| 159 | |
Anton Moberg | 9f346ab | 2021-05-21 17:20:21 +0200 | [diff] [blame] | 160 | void *__attribute__((weak)) ethosu_mutex_create(void) |
| 161 | { |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 162 | static uint8_t mutex_placeholder; |
| 163 | return &mutex_placeholder; |
Anton Moberg | 9f346ab | 2021-05-21 17:20:21 +0200 | [diff] [blame] | 164 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 165 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 166 | void __attribute__((weak)) ethosu_mutex_destroy(void *mutex) |
| 167 | { |
| 168 | UNUSED(mutex); |
| 169 | } |
| 170 | |
Ledion Daja | c6505f3 | 2022-04-20 09:55:21 +0200 | [diff] [blame] | 171 | int __attribute__((weak)) ethosu_mutex_lock(void *mutex) |
Anton Moberg | 61ec36b | 2021-04-30 17:10:48 +0200 | [diff] [blame] | 172 | { |
| 173 | UNUSED(mutex); |
Ledion Daja | c6505f3 | 2022-04-20 09:55:21 +0200 | [diff] [blame] | 174 | return 0; |
Anton Moberg | 61ec36b | 2021-04-30 17:10:48 +0200 | [diff] [blame] | 175 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 176 | |
Ledion Daja | c6505f3 | 2022-04-20 09:55:21 +0200 | [diff] [blame] | 177 | int __attribute__((weak)) ethosu_mutex_unlock(void *mutex) |
Anton Moberg | 61ec36b | 2021-04-30 17:10:48 +0200 | [diff] [blame] | 178 | { |
| 179 | UNUSED(mutex); |
Ledion Daja | c6505f3 | 2022-04-20 09:55:21 +0200 | [diff] [blame] | 180 | return 0; |
Anton Moberg | 61ec36b | 2021-04-30 17:10:48 +0200 | [diff] [blame] | 181 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 182 | |
| 183 | // Baremetal implementation of creating a semaphore |
| 184 | void *__attribute__((weak)) ethosu_semaphore_create(void) |
| 185 | { |
| 186 | struct ethosu_semaphore_t *sem = malloc(sizeof(*sem)); |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 187 | if (sem != NULL) |
| 188 | { |
| 189 | sem->count = 0; |
| 190 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 191 | return sem; |
| 192 | } |
| 193 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 194 | void __attribute__((weak)) ethosu_semaphore_destroy(void *sem) |
| 195 | { |
| 196 | free((struct ethosu_semaphore_t *)sem); |
| 197 | } |
| 198 | |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 199 | // Baremetal simulation of waiting/sleeping for and then taking a semaphore using intrisics |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 200 | int __attribute__((weak)) ethosu_semaphore_take(void *sem, uint64_t timeout) |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 201 | { |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 202 | UNUSED(timeout); |
| 203 | // Baremetal pseudo-example on how to trigger a timeout: |
| 204 | // if (timeout != ETHOSU_SEMAPHORE_WAIT_FOREVER) { |
| 205 | // setup_a_timer_to_call_SEV_after_time(timeout); |
| 206 | // } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 207 | struct ethosu_semaphore_t *s = sem; |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 208 | while (s->count == 0) |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 209 | { |
| 210 | __WFE(); |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 211 | // Baremetal pseudo-example check if timeout triggered: |
| 212 | // if (SEV_timer_triggered()) { |
| 213 | // return -1; |
| 214 | // } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 215 | } |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 216 | s->count--; |
Ledion Daja | c6505f3 | 2022-04-20 09:55:21 +0200 | [diff] [blame] | 217 | return 0; |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 218 | } |
| 219 | |
| 220 | // Baremetal simulation of giving a semaphore and waking up processes using intrinsics |
Ledion Daja | c6505f3 | 2022-04-20 09:55:21 +0200 | [diff] [blame] | 221 | int __attribute__((weak)) ethosu_semaphore_give(void *sem) |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 222 | { |
| 223 | struct ethosu_semaphore_t *s = sem; |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 224 | s->count++; |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 225 | __SEV(); |
Ledion Daja | c6505f3 | 2022-04-20 09:55:21 +0200 | [diff] [blame] | 226 | return 0; |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 227 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 228 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 229 | /****************************************************************************** |
| 230 | * Weak functions - Inference begin/end callbacks |
| 231 | ******************************************************************************/ |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 232 | |
Kristofer Jonsson | b3cde3c | 2022-01-27 17:30:15 +0100 | [diff] [blame] | 233 | void __attribute__((weak)) ethosu_inference_begin(struct ethosu_driver *drv, void *user_arg) |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 234 | { |
Kristofer Jonsson | b3cde3c | 2022-01-27 17:30:15 +0100 | [diff] [blame] | 235 | UNUSED(user_arg); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 236 | UNUSED(drv); |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 237 | } |
| 238 | |
Kristofer Jonsson | b3cde3c | 2022-01-27 17:30:15 +0100 | [diff] [blame] | 239 | void __attribute__((weak)) ethosu_inference_end(struct ethosu_driver *drv, void *user_arg) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 240 | { |
Kristofer Jonsson | b3cde3c | 2022-01-27 17:30:15 +0100 | [diff] [blame] | 241 | UNUSED(user_arg); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 242 | UNUSED(drv); |
| 243 | } |
| 244 | |
| 245 | /****************************************************************************** |
| 246 | * Static functions |
| 247 | ******************************************************************************/ |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 248 | static void ethosu_register_driver(struct ethosu_driver *drv) |
Jens Elofsson | 04961a4 | 2021-04-08 18:51:38 +0200 | [diff] [blame] | 249 | { |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 250 | ethosu_mutex_lock(ethosu_mutex); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 251 | drv->next = registered_drivers; |
| 252 | registered_drivers = drv; |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 253 | ethosu_mutex_unlock(ethosu_mutex); |
| 254 | |
| 255 | ethosu_semaphore_give(ethosu_semaphore); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 256 | |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 257 | LOG_INFO("New NPU driver registered (handle: 0x%p, NPU: 0x%p)", drv, drv->dev->reg); |
Jens Elofsson | 04961a4 | 2021-04-08 18:51:38 +0200 | [diff] [blame] | 258 | } |
| 259 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 260 | static int ethosu_deregister_driver(struct ethosu_driver *drv) |
| 261 | { |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 262 | struct ethosu_driver *curr; |
| 263 | struct ethosu_driver **prev; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 264 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 265 | ethosu_mutex_lock(ethosu_mutex); |
| 266 | curr = registered_drivers; |
| 267 | prev = ®istered_drivers; |
| 268 | |
| 269 | while (curr != NULL) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 270 | { |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 271 | if (curr == drv) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 272 | { |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 273 | *prev = curr->next; |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 274 | LOG_INFO("NPU driver handle %p deregistered.", drv); |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 275 | ethosu_semaphore_take(ethosu_semaphore, ETHOSU_SEMAPHORE_WAIT_FOREVER); |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 276 | break; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 277 | } |
| 278 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 279 | prev = &curr->next; |
| 280 | curr = curr->next; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 281 | } |
| 282 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 283 | ethosu_mutex_unlock(ethosu_mutex); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 284 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 285 | if (curr == NULL) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 286 | { |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 287 | LOG_ERR("No NPU driver handle registered at address %p.", drv); |
| 288 | return -1; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 289 | } |
| 290 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 291 | return 0; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 292 | } |
| 293 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 294 | static void ethosu_reset_job(struct ethosu_driver *drv) |
| 295 | { |
| 296 | memset(&drv->job, 0, sizeof(struct ethosu_job)); |
| 297 | } |
| 298 | |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 299 | static int handle_optimizer_config(struct ethosu_driver *drv, struct opt_cfg_s const *opt_cfg_p) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 300 | { |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 301 | LOG_INFO("Optimizer release nbr: %u patch: %u", opt_cfg_p->da_data.rel_nbr, opt_cfg_p->da_data.patch_nbr); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 302 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 303 | if (ethosu_dev_verify_optimizer_config(drv->dev, opt_cfg_p->cfg, opt_cfg_p->id) != true) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 304 | { |
| 305 | return -1; |
| 306 | } |
| 307 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 308 | return 0; |
| 309 | } |
| 310 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 311 | static int handle_command_stream(struct ethosu_driver *drv, const uint8_t *cmd_stream, const int cms_length) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 312 | { |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 313 | uint32_t cms_bytes = cms_length * BYTES_IN_32_BITS; |
| 314 | ptrdiff_t cmd_stream_ptr = (ptrdiff_t)cmd_stream; |
| 315 | |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 316 | LOG_INFO("handle_command_stream: cmd_stream=%p, cms_length %d", cmd_stream, cms_length); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 317 | |
| 318 | if (0 != ((ptrdiff_t)cmd_stream & MASK_16_BYTE_ALIGN)) |
| 319 | { |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 320 | LOG_ERR("Command stream addr %p not aligned to 16 bytes", cmd_stream); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 321 | return -1; |
| 322 | } |
| 323 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 324 | // Verify 16 byte alignment for base address' |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 325 | for (int i = 0; i < drv->job.num_base_addr; i++) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 326 | { |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 327 | if (0 != (drv->job.base_addr[i] & MASK_16_BYTE_ALIGN)) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 328 | { |
Rajasekaran Kalidoss | 2fa2bf0 | 2024-04-22 09:11:02 +0200 | [diff] [blame] | 329 | LOG_ERR("Base addr %d: 0x%" PRIx64 "not aligned to 16 bytes", i, drv->job.base_addr[i]); |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 330 | return -1; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 331 | } |
| 332 | } |
| 333 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 334 | // Flush the cache if available on CPU. |
| 335 | // The upcasting to uin32_t* is ok since the pointer never is dereferenced. |
| 336 | // The base_addr_size is null if invoking from prior to invoke_V2, in that case |
| 337 | // the whole cache is being flushed. |
| 338 | |
| 339 | if (drv->job.base_addr_size != NULL) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 340 | { |
| 341 | ethosu_flush_dcache((uint32_t *)cmd_stream_ptr, cms_bytes); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 342 | for (int i = 0; i < drv->job.num_base_addr; i++) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 343 | { |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 344 | ethosu_flush_dcache((uint32_t *)(uintptr_t)drv->job.base_addr[i], drv->job.base_addr_size[i]); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 345 | } |
| 346 | } |
| 347 | else |
| 348 | { |
| 349 | ethosu_flush_dcache(NULL, 0); |
| 350 | } |
| 351 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 352 | // Request power gating disabled during inference run |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 353 | if (ethosu_request_power(drv)) |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 354 | { |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 355 | LOG_ERR("Failed to request power"); |
| 356 | return -1; |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 357 | } |
| 358 | |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 359 | drv->job.state = ETHOSU_JOB_RUNNING; |
| 360 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 361 | // Inference begin callback |
| 362 | ethosu_inference_begin(drv, drv->job.user_arg); |
| 363 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 364 | // Execute the command stream |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 365 | ethosu_dev_run_command_stream(drv->dev, cmd_stream, cms_bytes, drv->job.base_addr, drv->job.num_base_addr); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 366 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 367 | return 0; |
| 368 | } |
| 369 | |
| 370 | /****************************************************************************** |
| 371 | * Weak functions - Interrupt handler |
| 372 | ******************************************************************************/ |
| 373 | void __attribute__((weak)) ethosu_irq_handler(struct ethosu_driver *drv) |
| 374 | { |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 375 | LOG_DEBUG("Got interrupt from Ethos-U"); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 376 | |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 377 | // Prevent race condition where interrupt triggered after a timeout waiting |
| 378 | // for semaphore, but before NPU is reset. |
| 379 | if (drv->job.result == ETHOSU_JOB_RESULT_TIMEOUT) |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 380 | { |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 381 | return; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 382 | } |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 383 | |
| 384 | drv->job.state = ETHOSU_JOB_DONE; |
| 385 | drv->job.result = ethosu_dev_handle_interrupt(drv->dev) ? ETHOSU_JOB_RESULT_OK : ETHOSU_JOB_RESULT_ERROR; |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 386 | ethosu_semaphore_give(drv->semaphore); |
| 387 | } |
| 388 | |
| 389 | /****************************************************************************** |
| 390 | * Functions API |
| 391 | ******************************************************************************/ |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 392 | |
Anton Moberg | effc7aa | 2021-05-03 09:25:06 +0200 | [diff] [blame] | 393 | int ethosu_init(struct ethosu_driver *drv, |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 394 | void *const base_address, |
Anton Moberg | effc7aa | 2021-05-03 09:25:06 +0200 | [diff] [blame] | 395 | const void *fast_memory, |
| 396 | const size_t fast_memory_size, |
| 397 | uint32_t secure_enable, |
| 398 | uint32_t privilege_enable) |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 399 | { |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 400 | LOG_INFO("Initializing NPU: base_address=%p, fast_memory=%p, fast_memory_size=%zu, secure=%" PRIu32 |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 401 | ", privileged=%" PRIu32, |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 402 | base_address, |
| 403 | fast_memory, |
Per Åstrand | e6498f0 | 2020-11-09 15:33:12 +0100 | [diff] [blame] | 404 | fast_memory_size, |
| 405 | secure_enable, |
| 406 | privilege_enable); |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 407 | |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 408 | if (!ethosu_mutex) |
| 409 | { |
| 410 | ethosu_mutex = ethosu_mutex_create(); |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 411 | if (!ethosu_mutex) |
| 412 | { |
| 413 | LOG_ERR("Failed to create global driver mutex"); |
| 414 | return -1; |
| 415 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 416 | } |
| 417 | |
| 418 | if (!ethosu_semaphore) |
| 419 | { |
| 420 | ethosu_semaphore = ethosu_semaphore_create(); |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 421 | if (!ethosu_semaphore) |
| 422 | { |
| 423 | LOG_ERR("Failed to create global driver semaphore"); |
| 424 | return -1; |
| 425 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 426 | } |
| 427 | |
Rajasekaran Kalidoss | 2fa2bf0 | 2024-04-22 09:11:02 +0200 | [diff] [blame] | 428 | drv->fast_memory = (uint64_t)fast_memory; |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 429 | drv->fast_memory_size = fast_memory_size; |
| 430 | drv->power_request_counter = 0; |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 431 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 432 | // Initialize the device and set requested security state and privilege mode |
| 433 | drv->dev = ethosu_dev_init(base_address, secure_enable, privilege_enable); |
| 434 | |
| 435 | if (drv->dev == NULL) |
Bhavik Patel | dae5be0 | 2020-06-18 15:25:15 +0200 | [diff] [blame] | 436 | { |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 437 | LOG_ERR("Failed to initialize Ethos-U device"); |
Bhavik Patel | dae5be0 | 2020-06-18 15:25:15 +0200 | [diff] [blame] | 438 | return -1; |
| 439 | } |
| 440 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 441 | drv->semaphore = ethosu_semaphore_create(); |
| 442 | if (!drv->semaphore) |
| 443 | { |
| 444 | LOG_ERR("Failed to create driver semaphore"); |
| 445 | ethosu_dev_deinit(drv->dev); |
| 446 | drv->dev = NULL; |
| 447 | return -1; |
| 448 | } |
| 449 | |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 450 | ethosu_reset_job(drv); |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 451 | ethosu_register_driver(drv); |
| 452 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 453 | return 0; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 454 | } |
| 455 | |
Anton Moberg | 547ca53 | 2021-06-14 09:43:53 +0200 | [diff] [blame] | 456 | void ethosu_deinit(struct ethosu_driver *drv) |
| 457 | { |
| 458 | ethosu_deregister_driver(drv); |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 459 | ethosu_semaphore_destroy(drv->semaphore); |
| 460 | ethosu_dev_deinit(drv->dev); |
| 461 | drv->dev = NULL; |
Anton Moberg | 547ca53 | 2021-06-14 09:43:53 +0200 | [diff] [blame] | 462 | } |
| 463 | |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 464 | int ethosu_soft_reset(struct ethosu_driver *drv) |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 465 | { |
| 466 | // Soft reset the NPU |
| 467 | if (ethosu_dev_soft_reset(drv->dev) != ETHOSU_SUCCESS) |
| 468 | { |
| 469 | LOG_ERR("Failed to soft-reset NPU"); |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 470 | return -1; |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 471 | } |
| 472 | |
| 473 | // Update power and clock gating after the soft reset |
| 474 | ethosu_dev_set_clock_and_power(drv->dev, |
| 475 | drv->power_request_counter > 0 ? ETHOSU_CLOCK_Q_DISABLE : ETHOSU_CLOCK_Q_ENABLE, |
| 476 | drv->power_request_counter > 0 ? ETHOSU_POWER_Q_DISABLE : ETHOSU_POWER_Q_ENABLE); |
| 477 | |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 478 | return 0; |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 479 | } |
| 480 | |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 481 | int ethosu_request_power(struct ethosu_driver *drv) |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 482 | { |
| 483 | // Check if this is the first power request, increase counter |
| 484 | if (drv->power_request_counter++ == 0) |
| 485 | { |
| 486 | // Always reset to a known state. Changes to requested |
| 487 | // security state/privilege mode if necessary. |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 488 | if (ethosu_soft_reset(drv)) |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 489 | { |
| 490 | LOG_ERR("Failed to request power for Ethos-U"); |
| 491 | drv->power_request_counter--; |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 492 | return -1; |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 493 | } |
| 494 | } |
Kristofer Jonsson | 68b5c91 | 2022-11-14 10:52:38 +0100 | [diff] [blame] | 495 | return 0; |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 496 | } |
| 497 | |
| 498 | void ethosu_release_power(struct ethosu_driver *drv) |
| 499 | { |
| 500 | if (drv->power_request_counter == 0) |
| 501 | { |
| 502 | LOG_WARN("No power request left to release, reference counter is 0"); |
| 503 | } |
| 504 | else |
| 505 | { |
| 506 | // Decrement ref counter and enable power gating if no requests remain |
| 507 | if (--drv->power_request_counter == 0) |
| 508 | { |
| 509 | ethosu_dev_set_clock_and_power(drv->dev, ETHOSU_CLOCK_Q_ENABLE, ETHOSU_POWER_Q_ENABLE); |
| 510 | } |
| 511 | } |
| 512 | } |
| 513 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 514 | void ethosu_get_driver_version(struct ethosu_driver_version *ver) |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 515 | { |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 516 | assert(ver != NULL); |
| 517 | ver->major = ETHOSU_DRIVER_VERSION_MAJOR; |
| 518 | ver->minor = ETHOSU_DRIVER_VERSION_MINOR; |
| 519 | ver->patch = ETHOSU_DRIVER_VERSION_PATCH; |
| 520 | } |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 521 | |
Jonny Svärd | a830f17 | 2021-06-07 16:57:00 +0200 | [diff] [blame] | 522 | void ethosu_get_hw_info(struct ethosu_driver *drv, struct ethosu_hw_info *hw) |
| 523 | { |
| 524 | assert(hw != NULL); |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 525 | ethosu_dev_get_hw_info(drv->dev, hw); |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 526 | } |
| 527 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 528 | int ethosu_wait(struct ethosu_driver *drv, bool block) |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 529 | { |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 530 | int ret = 0; |
| 531 | |
| 532 | switch (drv->job.state) |
| 533 | { |
| 534 | case ETHOSU_JOB_IDLE: |
| 535 | LOG_ERR("Inference job not running..."); |
| 536 | ret = -2; |
| 537 | break; |
| 538 | case ETHOSU_JOB_RUNNING: |
| 539 | if (!block) |
| 540 | { |
| 541 | // Inference still running, do not block |
| 542 | ret = 1; |
| 543 | break; |
| 544 | } |
| 545 | // fall through |
| 546 | case ETHOSU_JOB_DONE: |
| 547 | // Wait for interrupt in blocking mode. In non-blocking mode |
| 548 | // the interrupt has already triggered |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 549 | ret = ethosu_semaphore_take(drv->semaphore, ETHOSU_SEMAPHORE_WAIT_INFERENCE); |
| 550 | if (ret < 0) |
| 551 | { |
| 552 | drv->job.result = ETHOSU_JOB_RESULT_TIMEOUT; |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 553 | |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 554 | // There's a race where the NPU interrupt can have fired between semaphore |
| 555 | // timing out and setting the result above (checked in interrupt handler). |
| 556 | // By checking if the job state has been changed (only set to DONE by interrupt |
| 557 | // handler), we know if the interrupt handler has run, if so decrement the |
| 558 | // semaphore count by one (given in interrupt handler). |
| 559 | if (drv->job.state == ETHOSU_JOB_DONE) |
| 560 | { |
| 561 | drv->job.result = ETHOSU_JOB_RESULT_TIMEOUT; // Reset back to timeout |
| 562 | ethosu_semaphore_take(drv->semaphore, ETHOSU_SEMAPHORE_WAIT_INFERENCE); |
| 563 | } |
| 564 | } |
| 565 | |
| 566 | // Inference done callback - always called even in case of timeout |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 567 | ethosu_inference_end(drv, drv->job.user_arg); |
| 568 | |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 569 | // Release power gating disabled requirement |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 570 | ethosu_release_power(drv); |
| 571 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 572 | // Check NPU and interrupt status |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 573 | if (drv->job.result) |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 574 | { |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 575 | if (drv->job.result == ETHOSU_JOB_RESULT_ERROR) |
| 576 | { |
| 577 | LOG_ERR("NPU error(s) occured during inference."); |
| 578 | ethosu_dev_print_err_status(drv->dev); |
| 579 | } |
| 580 | else |
| 581 | { |
| 582 | LOG_ERR("NPU inference timed out."); |
| 583 | } |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 584 | |
| 585 | // Reset the NPU |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 586 | (void)ethosu_soft_reset(drv); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 587 | |
| 588 | ret = -1; |
| 589 | } |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 590 | else |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 591 | { |
| 592 | // Invalidate cache |
| 593 | if (drv->job.base_addr_size != NULL) |
| 594 | { |
| 595 | for (int i = 0; i < drv->job.num_base_addr; i++) |
| 596 | { |
| 597 | ethosu_invalidate_dcache((uint32_t *)(uintptr_t)drv->job.base_addr[i], drv->job.base_addr_size[i]); |
| 598 | } |
| 599 | } |
| 600 | else |
| 601 | { |
| 602 | ethosu_invalidate_dcache(NULL, 0); |
| 603 | } |
| 604 | |
| 605 | LOG_DEBUG("Inference finished successfully..."); |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 606 | ret = 0; |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 607 | } |
| 608 | |
| 609 | // Reset internal job (state resets to IDLE) |
| 610 | ethosu_reset_job(drv); |
| 611 | break; |
| 612 | |
| 613 | default: |
| 614 | LOG_ERR("Unexpected job state"); |
| 615 | ethosu_reset_job(drv); |
| 616 | ret = -1; |
| 617 | break; |
| 618 | } |
| 619 | |
| 620 | // Return inference job status |
| 621 | return ret; |
| 622 | } |
| 623 | |
| 624 | int ethosu_invoke_async(struct ethosu_driver *drv, |
| 625 | const void *custom_data_ptr, |
| 626 | const int custom_data_size, |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 627 | uint64_t *const base_addr, |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 628 | const size_t *base_addr_size, |
| 629 | const int num_base_addr, |
| 630 | void *user_arg) |
| 631 | { |
| 632 | |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 633 | const struct cop_data_s *data_ptr = custom_data_ptr; |
Kristofer Jonsson | 24455ee | 2022-02-08 13:33:22 +0100 | [diff] [blame] | 634 | const struct cop_data_s *data_end = (struct cop_data_s *)((ptrdiff_t)custom_data_ptr + custom_data_size); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 635 | |
| 636 | // Make sure an inference is not already running |
| 637 | if (drv->job.state != ETHOSU_JOB_IDLE) |
| 638 | { |
| 639 | LOG_ERR("Inference already running, or waiting to be cleared..."); |
| 640 | return -1; |
| 641 | } |
| 642 | |
| 643 | drv->job.state = ETHOSU_JOB_IDLE; |
| 644 | drv->job.custom_data_ptr = custom_data_ptr; |
| 645 | drv->job.custom_data_size = custom_data_size; |
| 646 | drv->job.base_addr = base_addr; |
| 647 | drv->job.base_addr_size = base_addr_size; |
| 648 | drv->job.num_base_addr = num_base_addr; |
| 649 | drv->job.user_arg = user_arg; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 650 | |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 651 | // First word in custom_data_ptr should contain "Custom Operator Payload 1" |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 652 | if (data_ptr->word != ETHOSU_FOURCC) |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 653 | { |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 654 | LOG_ERR("Custom Operator Payload: %" PRIu32 " is not correct, expected %x", data_ptr->word, ETHOSU_FOURCC); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 655 | goto err; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 656 | } |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 657 | |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 658 | // Custom data length must be a multiple of 32 bits |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 659 | if ((custom_data_size % BYTES_IN_32_BITS) != 0) |
| 660 | { |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 661 | LOG_ERR("custom_data_size=0x%x not a multiple of 4", (unsigned)custom_data_size); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 662 | goto err; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 663 | } |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 664 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 665 | data_ptr++; |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 666 | |
| 667 | // Adjust base address to fast memory area |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 668 | if (drv->fast_memory != 0 && num_base_addr >= FAST_MEMORY_BASE_ADDR_INDEX) |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 669 | { |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 670 | |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 671 | if (base_addr_size != NULL && base_addr_size[FAST_MEMORY_BASE_ADDR_INDEX] > drv->fast_memory_size) |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 672 | { |
Rajasekaran Kalidoss | 2fa2bf0 | 2024-04-22 09:11:02 +0200 | [diff] [blame] | 673 | LOG_ERR("Fast memory area too small. fast_memory_size=%zu, base_addr_size=%zu", |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 674 | drv->fast_memory_size, |
Kristofer Jonsson | 4c94b30 | 2020-11-06 10:33:21 +0100 | [diff] [blame] | 675 | base_addr_size[FAST_MEMORY_BASE_ADDR_INDEX]); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 676 | goto err; |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 677 | } |
Kristofer Jonsson | 4c94b30 | 2020-11-06 10:33:21 +0100 | [diff] [blame] | 678 | |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 679 | base_addr[FAST_MEMORY_BASE_ADDR_INDEX] = drv->fast_memory; |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 680 | } |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 681 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 682 | // Parse Custom Operator Payload data |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 683 | while (data_ptr < data_end) |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 684 | { |
| 685 | switch (data_ptr->driver_action_command) |
| 686 | { |
| 687 | case OPTIMIZER_CONFIG: |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 688 | LOG_DEBUG("OPTIMIZER_CONFIG"); |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 689 | struct opt_cfg_s const *opt_cfg_p = (const struct opt_cfg_s *)data_ptr; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 690 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 691 | if (handle_optimizer_config(drv, opt_cfg_p) < 0) |
| 692 | { |
| 693 | goto err; |
| 694 | } |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 695 | data_ptr += DRIVER_ACTION_LENGTH_32_BIT_WORD + OPTIMIZER_CONFIG_LENGTH_32_BIT_WORD; |
| 696 | break; |
| 697 | case COMMAND_STREAM: |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 698 | // Vela only supports putting one COMMAND_STREAM per op |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 699 | LOG_DEBUG("COMMAND_STREAM"); |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 700 | const uint8_t *command_stream = (const uint8_t *)(data_ptr + 1); |
| 701 | int cms_length = (data_ptr->reserved << 16) | data_ptr->length; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 702 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 703 | if (handle_command_stream(drv, command_stream, cms_length) < 0) |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 704 | { |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 705 | goto err; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 706 | } |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 707 | data_ptr += DRIVER_ACTION_LENGTH_32_BIT_WORD + cms_length; |
| 708 | break; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 709 | case NOP: |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 710 | LOG_DEBUG("NOP"); |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 711 | data_ptr += DRIVER_ACTION_LENGTH_32_BIT_WORD; |
| 712 | break; |
| 713 | default: |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 714 | LOG_ERR("UNSUPPORTED driver_action_command: %u", data_ptr->driver_action_command); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 715 | goto err; |
Bhavik Patel | e645fed | 2020-06-12 14:46:47 +0200 | [diff] [blame] | 716 | break; |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 717 | } |
| 718 | } |
Jonny Svärd | 136810f | 2021-10-13 16:04:26 +0200 | [diff] [blame] | 719 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 720 | return 0; |
| 721 | err: |
| 722 | LOG_ERR("Failed to invoke inference."); |
| 723 | ethosu_reset_job(drv); |
| 724 | return -1; |
| 725 | } |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 726 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 727 | int ethosu_invoke_v3(struct ethosu_driver *drv, |
| 728 | const void *custom_data_ptr, |
| 729 | const int custom_data_size, |
Ledion Daja | ecdce6c | 2023-01-16 15:39:29 +0100 | [diff] [blame] | 730 | uint64_t *const base_addr, |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 731 | const size_t *base_addr_size, |
| 732 | const int num_base_addr, |
| 733 | void *user_arg) |
| 734 | { |
| 735 | if (ethosu_invoke_async( |
| 736 | drv, custom_data_ptr, custom_data_size, base_addr, base_addr_size, num_base_addr, user_arg) < 0) |
Bhavik Patel | 5f8dad1 | 2020-09-30 09:06:52 +0200 | [diff] [blame] | 737 | { |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 738 | return -1; |
Bhavik Patel | 5f8dad1 | 2020-09-30 09:06:52 +0200 | [diff] [blame] | 739 | } |
Kristofer Jonsson | 2b201c3 | 2020-09-02 16:42:43 +0200 | [diff] [blame] | 740 | |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 741 | return ethosu_wait(drv, true); |
Kristofer Jonsson | 49bdee8 | 2020-04-06 13:21:21 +0200 | [diff] [blame] | 742 | } |
| 743 | |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 744 | struct ethosu_driver *ethosu_reserve_driver(void) |
| 745 | { |
Anton Moberg | df386e0 | 2021-02-02 11:26:48 +0100 | [diff] [blame] | 746 | struct ethosu_driver *drv = NULL; |
| 747 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 748 | LOG_INFO("Acquiring NPU driver handle"); |
Jonny Svärd | a2732ec | 2023-12-18 17:19:15 +0100 | [diff] [blame] | 749 | ethosu_semaphore_take(ethosu_semaphore, ETHOSU_SEMAPHORE_WAIT_FOREVER); // This is meant to block until available |
Anton Moberg | df386e0 | 2021-02-02 11:26:48 +0100 | [diff] [blame] | 750 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 751 | ethosu_mutex_lock(ethosu_mutex); |
| 752 | drv = registered_drivers; |
| 753 | |
| 754 | while (drv != NULL) |
| 755 | { |
| 756 | if (!drv->reserved) |
Anton Moberg | df386e0 | 2021-02-02 11:26:48 +0100 | [diff] [blame] | 757 | { |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 758 | drv->reserved = true; |
| 759 | LOG_DEBUG("NPU driver handle %p reserved", drv); |
Anton Moberg | df386e0 | 2021-02-02 11:26:48 +0100 | [diff] [blame] | 760 | break; |
| 761 | } |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 762 | drv = drv->next; |
| 763 | } |
| 764 | ethosu_mutex_unlock(ethosu_mutex); |
Anton Moberg | df386e0 | 2021-02-02 11:26:48 +0100 | [diff] [blame] | 765 | |
Jonny Svärd | c964424 | 2023-01-09 15:54:24 +0100 | [diff] [blame] | 766 | if (!drv) |
| 767 | { |
| 768 | LOG_ERR("No NPU driver handle available, but semaphore taken"); |
| 769 | } |
Anton Moberg | df386e0 | 2021-02-02 11:26:48 +0100 | [diff] [blame] | 770 | |
| 771 | return drv; |
| 772 | } |
| 773 | |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 774 | void ethosu_release_driver(struct ethosu_driver *drv) |
| 775 | { |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 776 | ethosu_mutex_lock(ethosu_mutex); |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 777 | if (drv != NULL && drv->reserved) |
| 778 | { |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 779 | if (drv->job.state == ETHOSU_JOB_RUNNING || drv->job.state == ETHOSU_JOB_DONE) |
| 780 | { |
| 781 | // Give the inference one shot to complete or force kill the job |
| 782 | if (ethosu_wait(drv, false) == 1) |
| 783 | { |
| 784 | // Still running, soft reset the NPU and reset driver |
Jonny Svärd | 301399d | 2022-04-26 18:31:24 +0200 | [diff] [blame] | 785 | drv->power_request_counter = 0; |
| 786 | ethosu_soft_reset(drv); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 787 | ethosu_reset_job(drv); |
Jonny Svärd | 1a3bb92 | 2022-02-25 16:28:21 +0100 | [diff] [blame] | 788 | } |
| 789 | } |
| 790 | |
Anton Moberg | 61da4d3 | 2020-12-22 16:00:31 +0100 | [diff] [blame] | 791 | drv->reserved = false; |
Kristofer Jonsson | 089a347 | 2021-11-12 12:52:07 +0100 | [diff] [blame] | 792 | LOG_DEBUG("NPU driver handle %p released", drv); |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 793 | ethosu_semaphore_give(ethosu_semaphore); |
Anton Moberg | 8d65b6f | 2020-12-21 09:37:18 +0100 | [diff] [blame] | 794 | } |
Anton Moberg | dfed5fd | 2021-03-11 14:41:11 +0100 | [diff] [blame] | 795 | ethosu_mutex_unlock(ethosu_mutex); |
Anton Moberg | 8d65b6f | 2020-12-21 09:37:18 +0100 | [diff] [blame] | 796 | } |