| 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 "support/atomic.h" | ||
| 25 | #include "support/debug.h" | ||
| 26 | #include "support/dlb_common.h" | ||
| 27 | #include "talp/backend.h" | ||
| 28 | #include "talp/backend_manager.h" | ||
| 29 | #include "talp/regions.h" | ||
| 30 | #include "talp/sample.h" | ||
| 31 | #include "talp/talp.h" | ||
| 32 | #include "talp/talp_output.h" | ||
| 33 | #include "talp/talp_types.h" | ||
| 34 | |||
| 35 | |||
| 36 | extern __thread bool thread_is_observer; | ||
| 37 | extern __thread bool thread_is_known; | ||
| 38 | |||
| 39 | static const backend_api_t *gpu_backend_api = NULL; | ||
| 40 | static gpu_measurements_t gpu_sample = {0}; | ||
| 41 | |||
| 42 | // Called from talp core | ||
| 43 | 15 | int talp_gpu_init(const subprocess_descriptor_t *spd) { | |
| 44 | |||
| 45 | 15 | gpu_backend_api = talp_backend_manager_load_gpu_backend(spd->options.talp_gpu_backend); | |
| 46 |
1/2✓ Branch 0 taken 15 times.
✗ Branch 1 not taken.
|
15 | if (gpu_backend_api == NULL) { |
| 47 | 15 | debug_warning("GPU backend could not be loaded"); | |
| 48 | 15 | return DLB_ERR_UNKNOWN; | |
| 49 | } | ||
| 50 | |||
| 51 | int error; | ||
| 52 | |||
| 53 | /* If GPU component is not explicitly set, probe plugin first */ | ||
| 54 | ✗ | if (!(spd->options.talp & TALP_COMPONENT_GPU)) { | |
| 55 | ✗ | error = gpu_backend_api->probe(); | |
| 56 | ✗ | if (error == DLB_BACKEND_ERROR) { | |
| 57 | ✗ | debug_warning("HWC backend probe failed"); | |
| 58 | ✗ | return DLB_ERR_UNKNOWN; | |
| 59 | } | ||
| 60 | } | ||
| 61 | |||
| 62 | ✗ | error = gpu_backend_api->init(&core_api); | |
| 63 | ✗ | if (error == DLB_BACKEND_ERROR) { | |
| 64 | ✗ | debug_warning("GPU backend could not be initialized"); | |
| 65 | ✗ | return DLB_ERR_UNKNOWN; | |
| 66 | } | ||
| 67 | |||
| 68 | ✗ | error = gpu_backend_api->start(); | |
| 69 | ✗ | if (error == DLB_BACKEND_ERROR) { | |
| 70 | ✗ | debug_warning("GPU backend could not be started"); | |
| 71 | ✗ | gpu_backend_api->finalize(); | |
| 72 | ✗ | return DLB_ERR_UNKNOWN; | |
| 73 | } | ||
| 74 | |||
| 75 | ✗ | if (gpu_backend_api->capabilities.gpu_amd) { | |
| 76 | ✗ | talp_output_record_gpu_vendor(GPU_VENDOR_AMD); | |
| 77 | ✗ | } else if (gpu_backend_api->capabilities.gpu_nvidia) { | |
| 78 | ✗ | talp_output_record_gpu_vendor(GPU_VENDOR_NVIDIA); | |
| 79 | } | ||
| 80 | |||
| 81 | ✗ | return DLB_SUCCESS; | |
| 82 | } | ||
| 83 | |||
| 84 | |||
| 85 | // Called from talp core | ||
| 86 | ✗ | void talp_gpu_finalize(void) { | |
| 87 | |||
| 88 | ✗ | if (gpu_backend_api != NULL) { | |
| 89 | |||
| 90 | ✗ | gpu_backend_api->stop(); | |
| 91 | ✗ | gpu_backend_api->flush(); | |
| 92 | ✗ | gpu_backend_api->finalize(); | |
| 93 | |||
| 94 | ✗ | talp_backend_manager_unload_gpu_backend(); | |
| 95 | ✗ | gpu_backend_api = NULL; | |
| 96 | } | ||
| 97 | } | ||
| 98 | |||
| 99 | |||
| 100 | // Called from GPU backend plugin: CPU enters GPU runtime | ||
| 101 | ✗ | void talp_gpu_enter_runtime(void) { | |
| 102 | |||
| 103 | /* Observer and unknown threads may call GPU offload functions, but TALP must ignore them */ | ||
| 104 | ✗ | if (unlikely(thread_is_observer || !thread_is_known)) return; | |
| 105 | |||
| 106 | ✗ | spd_enter_dlb(thread_spd); | |
| 107 | ✗ | talp_info_t *talp_info = thread_spd->talp_info; | |
| 108 | ✗ | if (talp_info) { | |
| 109 | /* Update sample */ | ||
| 110 | ✗ | talp_sample_update(talp_info); | |
| 111 | |||
| 112 | /* Into Sync call -> not_useful_gpu */ | ||
| 113 | ✗ | talp_sample_set_state(talp_info, TALP_STATE_NOT_USEFUL_GPU); | |
| 114 | } | ||
| 115 | } | ||
| 116 | |||
| 117 | |||
| 118 | // Called from GPU backend plugin: CPU exits GPU runtime | ||
| 119 | ✗ | void talp_gpu_exit_runtime(void) { | |
| 120 | |||
| 121 | /* Observer and unknown threads may call GPU offload functions, but TALP must ignore them */ | ||
| 122 | ✗ | if (unlikely(thread_is_observer || !thread_is_known)) return; | |
| 123 | |||
| 124 | ✗ | spd_enter_dlb(thread_spd); | |
| 125 | ✗ | talp_info_t *talp_info = thread_spd->talp_info; | |
| 126 | ✗ | if (talp_info) { | |
| 127 | /* Update sample */ | ||
| 128 | ✗ | talp_sample_update(talp_info); | |
| 129 | |||
| 130 | /* Add statistic */ | ||
| 131 | ✗ | talp_sample_t *sample = talp_sample_get(talp_info); | |
| 132 | ✗ | DLB_ATOMIC_ADD_RLX(&sample->stats.num_gpu_runtime_calls, 1); | |
| 133 | |||
| 134 | /* Out of Sync call -> useful */ | ||
| 135 | ✗ | talp_sample_set_state(talp_info, TALP_STATE_USEFUL); | |
| 136 | |||
| 137 | /* Only when needed, update all regions */ | ||
| 138 | ✗ | if (talp_info->flags.external_profiler | |
| 139 | ✗ | && talp_sample_is_main()) { | |
| 140 | ✗ | talp_aggregate_samples_to_regions(talp_info); | |
| 141 | } | ||
| 142 | } | ||
| 143 | } | ||
| 144 | |||
| 145 | |||
| 146 | // Called from GPU backend plugin: flush GPU data | ||
| 147 | ✗ | void talp_gpu_submit(const gpu_measurements_t *measurements) { | |
| 148 | |||
| 149 | ✗ | gpu_sample.useful_time += measurements->useful_time; | |
| 150 | ✗ | gpu_sample.communication_time += measurements->communication_time; | |
| 151 | ✗ | gpu_sample.inactive_time += measurements->inactive_time; | |
| 152 | } | ||
| 153 | |||
| 154 | // called from core | ||
| 155 | ✗ | void talp_gpu_collect(gpu_measurements_t *out) { | |
| 156 | |||
| 157 | ✗ | fatal_cond(thread_is_observer, "Observer thread collecting GPU metrics. Please report bug."); | |
| 158 | ✗ | fatal_cond(!thread_is_known, "Unknown thread collecting GPU metrics. Please report bug."); | |
| 159 | |||
| 160 | // flush GPU, may cause callback to talp_gpu_submit | ||
| 161 | ✗ | gpu_backend_api->flush(); | |
| 162 | |||
| 163 | ✗ | *out = gpu_sample; | |
| 164 | ✗ | gpu_sample = (const gpu_measurements_t){0}; | |
| 165 | } | ||
| 166 |