| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /*********************************************************************************/ | ||
| 2 | /* Copyright 2009-2026 Barcelona Supercomputing Center */ | ||
| 3 | /* */ | ||
| 4 | /* This file is part of the DLB library. */ | ||
| 5 | /* */ | ||
| 6 | /* DLB is free software: you can redistribute it and/or modify */ | ||
| 7 | /* it under the terms of the GNU Lesser General Public License as published by */ | ||
| 8 | /* the Free Software Foundation, either version 3 of the License, or */ | ||
| 9 | /* (at your option) any later version. */ | ||
| 10 | /* */ | ||
| 11 | /* DLB is distributed in the hope that it will be useful, */ | ||
| 12 | /* but WITHOUT ANY WARRANTY; without even the implied warranty of */ | ||
| 13 | /* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the */ | ||
| 14 | /* GNU Lesser General Public License for more details. */ | ||
| 15 | /* */ | ||
| 16 | /* You should have received a copy of the GNU Lesser General Public License */ | ||
| 17 | /* along with DLB. If not, see <https://www.gnu.org/licenses/>. */ | ||
| 18 | /*********************************************************************************/ | ||
| 19 | |||
| 20 | #include "talp/talp_gpu.h" | ||
| 21 | |||
| 22 | #include "apis/dlb_errors.h" | ||
| 23 | #include "LB_core/spd.h" | ||
| 24 | #include "LB_core/thread_ctx.h" | ||
| 25 | #include "support/debug.h" | ||
| 26 | #include "support/dlb_common.h" | ||
| 27 | #include "support/gpu_mask_utils.h" | ||
| 28 | #include "talp/backend.h" | ||
| 29 | #include "talp/backend_manager.h" | ||
| 30 | #include "talp/regions.h" | ||
| 31 | #include "talp/sample.h" | ||
| 32 | #include "talp/talp.h" | ||
| 33 | #include "talp/talp_output.h" | ||
| 34 | #include "talp/talp_types.h" | ||
| 35 | |||
| 36 | #include <string.h> | ||
| 37 | |||
| 38 | |||
| 39 | static const backend_api_t *gpu_backend_api = NULL; | ||
| 40 | static gpu_device_entry_t *devices = NULL; | ||
| 41 | static size_t num_devices = 0; | ||
| 42 | |||
| 43 | // Called from talp core | ||
| 44 | 20 | int talp_gpu_init(const subprocess_descriptor_t *spd) { | |
| 45 | |||
| 46 | 20 | gpu_backend_api = talp_backend_manager_load_gpu_backend(spd->options.talp_gpu_backend); | |
| 47 |
1/2✓ Branch 0 taken 20 times.
✗ Branch 1 not taken.
|
20 | if (gpu_backend_api == NULL) { |
| 48 | 20 | debug_warning("GPU backend could not be loaded"); | |
| 49 | 20 | return DLB_ERR_UNKNOWN; | |
| 50 | } | ||
| 51 | |||
| 52 | int error; | ||
| 53 | |||
| 54 | /* If GPU component is not explicitly set, probe plugin first */ | ||
| 55 | ✗ | if (!(spd->options.talp & TALP_COMPONENT_GPU)) { | |
| 56 | ✗ | error = gpu_backend_api->probe(); | |
| 57 | ✗ | if (error == DLB_BACKEND_ERROR) { | |
| 58 | ✗ | debug_warning("HWC backend probe failed"); | |
| 59 | ✗ | return DLB_ERR_UNKNOWN; | |
| 60 | } | ||
| 61 | } | ||
| 62 | |||
| 63 | ✗ | error = gpu_backend_api->init(&core_api); | |
| 64 | ✗ | if (error == DLB_BACKEND_ERROR) { | |
| 65 | ✗ | debug_warning("GPU backend could not be initialized"); | |
| 66 | ✗ | return DLB_ERR_UNKNOWN; | |
| 67 | } | ||
| 68 | |||
| 69 | ✗ | error = gpu_backend_api->start(); | |
| 70 | ✗ | if (error == DLB_BACKEND_ERROR) { | |
| 71 | ✗ | debug_warning("GPU backend could not be started"); | |
| 72 | ✗ | gpu_backend_api->finalize(); | |
| 73 | ✗ | return DLB_ERR_UNKNOWN; | |
| 74 | } | ||
| 75 | |||
| 76 | ✗ | if (gpu_backend_api->capabilities.gpu_amd) { | |
| 77 | ✗ | talp_output_record_gpu_vendor(GPU_VENDOR_AMD); | |
| 78 | ✗ | } else if (gpu_backend_api->capabilities.gpu_nvidia) { | |
| 79 | ✗ | talp_output_record_gpu_vendor(GPU_VENDOR_NVIDIA); | |
| 80 | } | ||
| 81 | |||
| 82 | ✗ | error = gpu_backend_api->gpu.get_devices(NULL, 0, &num_devices); | |
| 83 | ✗ | if (error == DLB_BACKEND_ERROR || num_devices == 0) { | |
| 84 | ✗ | debug_warning("GPU backend could not obtain number of devices"); | |
| 85 | ✗ | gpu_backend_api->stop(); | |
| 86 | ✗ | gpu_backend_api->finalize(); | |
| 87 | } | ||
| 88 | |||
| 89 | ✗ | devices = malloc(sizeof(*devices) * num_devices); | |
| 90 | ✗ | error = gpu_backend_api->gpu.get_devices(devices, num_devices, NULL); | |
| 91 | ✗ | if (error == DLB_BACKEND_ERROR) { | |
| 92 | ✗ | debug_warning("GPU backend could not obtain device unique IDs"); | |
| 93 | ✗ | gpu_backend_api->stop(); | |
| 94 | ✗ | gpu_backend_api->finalize(); | |
| 95 | } | ||
| 96 | |||
| 97 | ✗ | if (num_devices > MAX_NODE_GPUS) { | |
| 98 | ✗ | warning("%zu devices registered within a node but MAX_NODE_GPUS=%d. " | |
| 99 | "If you think this is an error, please report bug.", | ||
| 100 | num_devices, MAX_NODE_GPUS); | ||
| 101 | } | ||
| 102 | |||
| 103 | ✗ | return DLB_SUCCESS; | |
| 104 | } | ||
| 105 | |||
| 106 | |||
| 107 | // Called from talp core | ||
| 108 | ✗ | void talp_gpu_finalize(void) { | |
| 109 | |||
| 110 | ✗ | if (gpu_backend_api != NULL) { | |
| 111 | |||
| 112 | ✗ | gpu_backend_api->stop(); | |
| 113 | ✗ | gpu_backend_api->finalize(); | |
| 114 | |||
| 115 | ✗ | free(devices); | |
| 116 | ✗ | devices = NULL; | |
| 117 | ✗ | num_devices = 0; | |
| 118 | |||
| 119 | ✗ | talp_backend_manager_unload_gpu_backend(); | |
| 120 | ✗ | gpu_backend_api = NULL; | |
| 121 | } | ||
| 122 | } | ||
| 123 | |||
| 124 | |||
| 125 | ✗ | uint64_t talp_gpu_local_to_unique_id(uint32_t local_id) { | |
| 126 | |||
| 127 | ✗ | for (size_t i = 0; i < num_devices; ++i) { | |
| 128 | ✗ | if (devices[i].local_id == local_id) { | |
| 129 | ✗ | return devices[i].node_unique_id; | |
| 130 | } | ||
| 131 | } | ||
| 132 | |||
| 133 | ✗ | return ULLONG_MAX; | |
| 134 | } | ||
| 135 | |||
| 136 | |||
| 137 | // Called from GPU backend plugin: CPU enters GPU runtime | ||
| 138 | ✗ | void talp_gpu_enter_runtime(void) { | |
| 139 | |||
| 140 | /* Observer and unknown threads may call GPU offload functions, but TALP must ignore them */ | ||
| 141 | ✗ | if (unlikely(!thread_is_profiled())) return; | |
| 142 | |||
| 143 | ✗ | talp_info_t *talp_info = thread_spd->talp_info; | |
| 144 | ✗ | if (talp_info) { | |
| 145 | /* Update sample */ | ||
| 146 | ✗ | talp_sample_update(talp_info); | |
| 147 | |||
| 148 | /* Into Sync call -> not_useful_gpu */ | ||
| 149 | ✗ | talp_sample_set_state(talp_info, TALP_STATE_NOT_USEFUL_GPU); | |
| 150 | } | ||
| 151 | } | ||
| 152 | |||
| 153 | |||
| 154 | // Called from GPU backend plugin: CPU exits GPU runtime | ||
| 155 | ✗ | void talp_gpu_exit_runtime(void) { | |
| 156 | |||
| 157 | /* Observer and unknown threads may call GPU offload functions, but TALP must ignore them */ | ||
| 158 | ✗ | if (unlikely(!thread_is_profiled())) return; | |
| 159 | |||
| 160 | ✗ | talp_info_t *talp_info = thread_spd->talp_info; | |
| 161 | ✗ | if (talp_info) { | |
| 162 | /* Update sample */ | ||
| 163 | ✗ | talp_sample_update(talp_info); | |
| 164 | |||
| 165 | /* Add statistic */ | ||
| 166 | ✗ | talp_sample_t *sample = talp_sample_get(talp_info); | |
| 167 | ✗ | ++sample->stats.num_gpu_runtime_calls; | |
| 168 | |||
| 169 | /* Out of Sync call -> useful */ | ||
| 170 | ✗ | talp_sample_set_state(talp_info, TALP_STATE_USEFUL); | |
| 171 | |||
| 172 | /* Only when needed, update all regions */ | ||
| 173 | ✗ | if (talp_info->flags.external_profiler | |
| 174 | ✗ | && thread_is_main_sequential()) { | |
| 175 | ✗ | talp_aggregate_samples_to_regions(talp_info); | |
| 176 | } | ||
| 177 | } | ||
| 178 | } | ||
| 179 | |||
| 180 | |||
| 181 | // Called from core | ||
| 182 | ✗ | void talp_gpu_collect(gpu_timers_t *out, size_t capacity, uint64_t *out_mask) { | |
| 183 | |||
| 184 | ✗ | ensure(thread_is_main(), "Non-main thread collecting GPU measurements. Please report bug."); | |
| 185 | |||
| 186 | ✗ | gpu_backend_api->gpu.collect(out, capacity, out_mask); | |
| 187 | } | ||
| 188 |