Kshitij Sisodia | c22e80e | 2022-03-14 09:26:48 +0000 | [diff] [blame] | 1 | /* |
Richard Burton | f32a86a | 2022-11-15 11:46:11 +0000 | [diff] [blame] | 2 | * SPDX-FileCopyrightText: Copyright 2022 Arm Limited and/or its affiliates <open-source-office@arm.com> |
Kshitij Sisodia | c22e80e | 2022-03-14 09:26:48 +0000 | [diff] [blame] | 3 | * SPDX-License-Identifier: Apache-2.0 |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | * you may not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at |
| 8 | * |
| 9 | * http://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, |
| 13 | * WITHOUT 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 | |
| 18 | #include "ethosu_cpu_cache.h" |
| 19 | |
| 20 | #include "RTE_Components.h" /* For CPU related defintiions */ |
| 21 | #include "ethosu_driver.h" /* Arm Ethos-U driver header */ |
| 22 | #include "log_macros.h" /* Logging macros */ |
| 23 | |
Kshitij Sisodia | de54e16 | 2022-12-20 16:09:15 +0000 | [diff] [blame] | 24 | /** Structure to maintain data cache states. */ |
| 25 | typedef struct _cpu_cache_state { |
| 26 | uint32_t dcache_invalidated : 1; |
| 27 | uint32_t dcache_cleaned : 1; |
| 28 | } cpu_cache_state; |
| 29 | |
| 30 | /** Static CPU cache state object. |
| 31 | * @note This logic around flipping these states is based on the driver |
| 32 | * calling the functions in this sequence: |
| 33 | * |
| 34 | * Cache flush (ethosu_flush_dcache) |
| 35 | * ↓ |
| 36 | * Start inference (ethosu_inference_begin) |
| 37 | * ↓ |
| 38 | * Inference (ethosu_dev_run_command_stream) |
| 39 | * ↓ |
| 40 | * End inference (ethosu_inference_end) |
| 41 | * ↓ |
| 42 | * Cache invalidate (ethosu_dcache_invalidate) |
| 43 | **/ |
| 44 | static cpu_cache_state s_cache_state = {.dcache_cleaned = 0, .dcache_invalidated = 0}; |
| 45 | |
| 46 | /** |
| 47 | * @brief Gets the current CPU cache state. |
| 48 | * @return Pointer to the CPU cache state object. |
| 49 | */ |
| 50 | static cpu_cache_state* ethosu_get_cpu_cache_state(void) |
| 51 | { |
| 52 | return &s_cache_state; |
| 53 | } |
| 54 | |
| 55 | void ethosu_clear_cache_states(void) |
| 56 | { |
| 57 | cpu_cache_state* const state = ethosu_get_cpu_cache_state(); |
| 58 | trace("Clearing cache state members\n"); |
| 59 | state->dcache_invalidated = 0; |
| 60 | state->dcache_cleaned = 0; |
| 61 | } |
| 62 | |
Kshitij Sisodia | c22e80e | 2022-03-14 09:26:48 +0000 | [diff] [blame] | 63 | void ethosu_flush_dcache(uint32_t *p, size_t bytes) |
| 64 | { |
Kshitij Sisodia | c22e80e | 2022-03-14 09:26:48 +0000 | [diff] [blame] | 65 | UNUSED(p); |
| 66 | UNUSED(bytes); |
Kshitij Sisodia | de54e16 | 2022-12-20 16:09:15 +0000 | [diff] [blame] | 67 | #if defined (__DCACHE_PRESENT) && (__DCACHE_PRESENT == 1U) |
| 68 | cpu_cache_state* const state = ethosu_get_cpu_cache_state(); |
| 69 | if (SCB->CCR & SCB_CCR_DC_Msk) { |
| 70 | |
| 71 | /** |
| 72 | * @note We could choose to call the `SCB_CleanDCache_by_Addr` function |
| 73 | * here, but the sizes which this function is called for, can |
| 74 | * cause unnecessary delays. It's worth noting that this function |
| 75 | * is called from the Arm Ethos-U NPU drive repeatedly for each |
| 76 | * region it accesses. This could even be RO memory which does |
| 77 | * not need cache maintenance, along with parts of the input and |
| 78 | * output tensors which rightly need to be cleaned. Therefore, to |
| 79 | * reduce overhead of repeated calls for large memory sizes, we |
| 80 | * call the clean and invalidation functions for whole cache. |
| 81 | * |
| 82 | * If the neural network to be executed is completely falling |
| 83 | * onto the NPU, consider disabling the data cache altogether |
| 84 | * for the duration of the inference to further reduce the cache |
| 85 | * maintenance burden in these functions. |
| 86 | */ |
| 87 | |
| 88 | /** Clean the cache if it hasn't been cleaned already */ |
| 89 | if (!state->dcache_cleaned) { |
| 90 | trace("Cleaning data cache\n"); |
| 91 | SCB_CleanDCache(); |
| 92 | |
| 93 | /** Assert the cache cleaned state and clear the invalidation |
| 94 | * state. */ |
| 95 | state->dcache_cleaned = 1; |
| 96 | state->dcache_invalidated = 0; |
| 97 | } |
| 98 | } |
Kshitij Sisodia | c22e80e | 2022-03-14 09:26:48 +0000 | [diff] [blame] | 99 | #endif /* defined (__DCACHE_PRESENT) && (__DCACHE_PRESENT == 1U) */ |
| 100 | } |
| 101 | |
| 102 | void ethosu_invalidate_dcache(uint32_t *p, size_t bytes) |
| 103 | { |
Kshitij Sisodia | c22e80e | 2022-03-14 09:26:48 +0000 | [diff] [blame] | 104 | UNUSED(p); |
| 105 | UNUSED(bytes); |
Kshitij Sisodia | de54e16 | 2022-12-20 16:09:15 +0000 | [diff] [blame] | 106 | #if defined (__DCACHE_PRESENT) && (__DCACHE_PRESENT == 1U) |
| 107 | cpu_cache_state* const state = ethosu_get_cpu_cache_state(); |
| 108 | if (SCB->CCR & SCB_CCR_DC_Msk) { |
| 109 | /** |
| 110 | * See note in ethosu_flush_dcache function for why we clean the whole |
| 111 | * cache instead of calling it for specific addresses. |
| 112 | **/ |
| 113 | if (!state->dcache_invalidated) { |
| 114 | trace("Invalidating data cache\n"); |
| 115 | SCB_InvalidateDCache(); |
| 116 | |
| 117 | /** Assert the cache invalidation state and clear the clean |
| 118 | * state. */ |
| 119 | state->dcache_invalidated = 1; |
| 120 | state->dcache_cleaned = 0; |
| 121 | } |
| 122 | } |
Kshitij Sisodia | c22e80e | 2022-03-14 09:26:48 +0000 | [diff] [blame] | 123 | #endif /* defined (__DCACHE_PRESENT) && (__DCACHE_PRESENT == 1U) */ |
| 124 | } |