src/init.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // brotensor runtime: init(), default-device policy, DeviceScope, sync. | ||
| 2 | // | ||
| 3 | // The CPU backend self-registers from a static-init object in | ||
| 4 | // src/cpu/register.cpp. init() probes CUDA / Metal if the corresponding | ||
| 5 | // backend was compiled in. When a backend isn't built, | ||
| 6 | // BROTENSOR_HAS_CUDA / BROTENSOR_HAS_METAL are not defined so the probe | ||
| 7 | // branches compile out. | ||
| 8 | |||
| 9 | #include <brotensor/runtime.h> | ||
| 10 | #include <brotensor/tensor.h> | ||
| 11 | #include <brotensor/detail/dispatch.h> | ||
| 12 | #include <brotensor/detail/cpu/thread_pool.h> | ||
| 13 | |||
| 14 | #include <atomic> | ||
| 15 | #include <cstdlib> | ||
| 16 | #include <cstring> | ||
| 17 | #include <mutex> | ||
| 18 | #include <optional> | ||
| 19 | #include <stdexcept> | ||
| 20 | #include <string> | ||
| 21 | #include <vector> | ||
| 22 | |||
| 23 | #if defined(BROTENSOR_HAS_CUDA) | ||
| 24 | // Defined in src/cuda/init.cu. | ||
| 25 | extern "C" void brotensor_probe_and_register_cuda(); | ||
| 26 | #endif | ||
| 27 | |||
| 28 | #if defined(BROTENSOR_HAS_METAL) | ||
| 29 | // Defined in src/metal/init.mm. | ||
| 30 | extern "C" void brotensor_probe_and_register_metal(); | ||
| 31 | #endif | ||
| 32 | |||
| 33 | namespace brotensor { | ||
| 34 | |||
| 35 | namespace { | ||
| 36 | |||
| 37 | 154 | std::mutex& init_mutex() { | |
| 38 |
2/4✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 154 times.
|
154 | static std::mutex m; |
| 39 | 154 | return m; | |
| 40 | } | ||
| 41 | |||
| 42 | 464 | std::atomic<bool>& init_done_flag() { | |
| 43 | static std::atomic<bool> f{false}; | ||
| 44 | 464 | return f; | |
| 45 | } | ||
| 46 | |||
| 47 | // Global default device. Initialised lazily on first read. | ||
| 48 | 164 | std::atomic<Device>& global_default() { | |
| 49 |
3/4✓ Branch 0 taken 154 times.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 154 times.
|
164 | static std::atomic<Device> d{Device::CPU}; |
| 50 | 164 | return d; | |
| 51 | } | ||
| 52 | |||
| 53 | 333 | std::atomic<bool>& global_default_set_flag() { | |
| 54 | static std::atomic<bool> f{false}; | ||
| 55 | 333 | return f; | |
| 56 | } | ||
| 57 | |||
| 58 | // Thread-local DeviceScope override stack — we only need the *current* value | ||
| 59 | // since DeviceScope ctor saves the previous on the local stack frame. | ||
| 60 | thread_local std::optional<Device> tls_scope_override; | ||
| 61 | |||
| 62 | 169 | Device pick_default_from_available() { | |
| 63 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 169 times.
|
169 | if (detail::is_registered(Device::CUDA)) return Device::CUDA; |
| 64 |
2/2✓ Branch 0 taken 154 times.
✓ Branch 1 taken 15 times.
|
169 | if (detail::is_registered(Device::Metal)) return Device::Metal; |
| 65 | 15 | return Device::CPU; | |
| 66 | 169 | } | |
| 67 | |||
| 68 | 154 | std::optional<Device> parse_env_device() { | |
| 69 | 154 | const char* env = std::getenv("BROTENSOR_DEFAULT_DEVICE"); | |
| 70 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 154 times.
|
154 | if (!env) return std::nullopt; |
| 71 | ✗ | std::string s(env); | |
| 72 | ✗ | for (char& c : s) { | |
| 73 | ✗ | if (c >= 'A' && c <= 'Z') c = static_cast<char>(c + 32); | |
| 74 | } | ||
| 75 | ✗ | if (s == "cpu") return Device::cpu(); | |
| 76 | ✗ | if (s == "cuda" || s == "cuda:0") return Device::cuda(0); | |
| 77 | ✗ | if (s.rfind("cuda:", 0) == 0) { | |
| 78 | ✗ | int idx = std::atoi(s.c_str() + 5); | |
| 79 | ✗ | return Device::cuda(idx); | |
| 80 | } | ||
| 81 | ✗ | if (s == "metal" || s == "metal:0") return Device::metal(0); | |
| 82 | ✗ | if (s.rfind("metal:", 0) == 0) { | |
| 83 | ✗ | int idx = std::atoi(s.c_str() + 6); | |
| 84 | ✗ | return Device::metal(idx); | |
| 85 | } | ||
| 86 | ✗ | return std::nullopt; | |
| 87 | 154 | } | |
| 88 | |||
| 89 | } // namespace | ||
| 90 | |||
| 91 | 156 | void init() { | |
| 92 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 154 times.
|
156 | if (init_done_flag().load(std::memory_order_acquire)) return; |
| 93 | 154 | std::lock_guard<std::mutex> lock(init_mutex()); | |
| 94 |
2/4✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 154 times.
|
154 | if (init_done_flag().load(std::memory_order_relaxed)) return; |
| 95 | |||
| 96 | #if defined(BROTENSOR_HAS_CUDA) | ||
| 97 | try { brotensor_probe_and_register_cuda(); } catch (...) { /* no CUDA */ } | ||
| 98 | #endif | ||
| 99 | #if defined(BROTENSOR_HAS_METAL) | ||
| 100 |
1/4✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
154 | try { brotensor_probe_and_register_metal(); } catch (...) { /* no Metal */ } |
| 101 | #endif | ||
| 102 | |||
| 103 | // Determine default device once. | ||
| 104 |
2/4✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 154 times.
✗ Branch 3 not taken.
|
154 | if (!global_default_set_flag().load(std::memory_order_relaxed)) { |
| 105 |
2/4✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 154 times.
|
154 | if (auto envd = parse_env_device()) { |
| 106 | ✗ | if (detail::is_registered(*envd)) { | |
| 107 | ✗ | global_default().store(*envd, std::memory_order_relaxed); | |
| 108 | ✗ | } else { | |
| 109 | ✗ | global_default().store(pick_default_from_available(), | |
| 110 | std::memory_order_relaxed); | ||
| 111 | } | ||
| 112 | ✗ | } else { | |
| 113 |
2/4✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 154 times.
✗ Branch 3 not taken.
|
154 | global_default().store(pick_default_from_available(), |
| 114 | std::memory_order_relaxed); | ||
| 115 | } | ||
| 116 |
1/2✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
|
154 | global_default_set_flag().store(true, std::memory_order_relaxed); |
| 117 | 154 | } | |
| 118 | |||
| 119 |
1/2✓ Branch 0 taken 154 times.
✗ Branch 1 not taken.
|
154 | init_done_flag().store(true, std::memory_order_release); |
| 120 | 156 | } | |
| 121 | |||
| 122 | ✗ | void shutdown() { | |
| 123 | ✗ | detail::cpu::ThreadPool::instance().shutdown(); | |
| 124 | ✗ | } | |
| 125 | |||
| 126 | 73 | Device default_device() { | |
| 127 |
2/2✓ Branch 0 taken 50 times.
✓ Branch 1 taken 23 times.
|
73 | if (tls_scope_override.has_value()) return *tls_scope_override; |
| 128 | // Lazy default: if init hasn't run, fall back to whatever's registered. | ||
| 129 |
2/2✓ Branch 0 taken 8 times.
✓ Branch 1 taken 15 times.
|
23 | if (!global_default_set_flag().load(std::memory_order_acquire)) { |
| 130 | 15 | return pick_default_from_available(); | |
| 131 | } | ||
| 132 | 8 | return global_default().load(std::memory_order_acquire); | |
| 133 | 73 | } | |
| 134 | |||
| 135 | 8 | Dtype compute_dtype() { | |
| 136 | 8 | return compute_dtype(default_device()); | |
| 137 | } | ||
| 138 | |||
| 139 | 8 | Dtype compute_dtype(Device d) { | |
| 140 | 8 | return d.is_cpu() ? Dtype::FP32 : Dtype::FP16; | |
| 141 | } | ||
| 142 | |||
| 143 | 2 | void set_default_device(Device d) { | |
| 144 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
2 | if (!detail::is_registered(d)) { |
| 145 | ✗ | std::string m = "brotensor: set_default_device: backend "; | |
| 146 | ✗ | m += device_name(d); | |
| 147 | ✗ | m += " is not available"; | |
| 148 | ✗ | throw std::runtime_error(m); | |
| 149 | ✗ | } | |
| 150 | 2 | global_default().store(d, std::memory_order_release); | |
| 151 | 2 | global_default_set_flag().store(true, std::memory_order_release); | |
| 152 | 2 | } | |
| 153 | |||
| 154 | 2916 | std::vector<Device> available_devices() { | |
| 155 | 2916 | std::vector<Device> out; | |
| 156 |
2/4✓ Branch 0 taken 2916 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2916 times.
✗ Branch 3 not taken.
|
2916 | if (detail::is_registered(Device::CPU)) { |
| 157 |
1/2✓ Branch 0 taken 2916 times.
✗ Branch 1 not taken.
|
2916 | out.push_back(Device::CPU); |
| 158 | 2916 | } | |
| 159 |
2/4✓ Branch 0 taken 2916 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2916 times.
|
2916 | if (detail::is_registered(Device::CUDA)) { |
| 160 | ✗ | int count = cuda_device_count(); | |
| 161 | ✗ | if (count <= 0) count = 1; | |
| 162 | ✗ | for (int i = 0; i < count; ++i) { | |
| 163 | ✗ | out.push_back(Device::cuda(i)); | |
| 164 | ✗ | } | |
| 165 | ✗ | } | |
| 166 |
2/4✓ Branch 0 taken 2916 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2916 times.
✗ Branch 3 not taken.
|
2916 | if (detail::is_registered(Device::Metal)) { |
| 167 |
1/2✓ Branch 0 taken 2916 times.
✗ Branch 1 not taken.
|
2916 | out.push_back(Device::Metal); |
| 168 | 2916 | } | |
| 169 | 2916 | return out; | |
| 170 | 2916 | } | |
| 171 | |||
| 172 | 430 | bool is_available(Device d) { | |
| 173 | 430 | return detail::is_registered(d); | |
| 174 | } | ||
| 175 | |||
| 176 | // ─── DeviceScope ─────────────────────────────────────────────────────────── | ||
| 177 | |||
| 178 | // Track whether tls_scope_override held a value before this scope was | ||
| 179 | // pushed. The Device prev_ field on the scope object stores the prior value | ||
| 180 | // when one existed; the parallel bool stack tracks "had a value". | ||
| 181 | namespace { | ||
| 182 | thread_local std::vector<bool> tls_scope_had_prev; | ||
| 183 | } // namespace | ||
| 184 | |||
| 185 | 16 | DeviceScope::DeviceScope(Device d) { | |
| 186 |
1/2✓ Branch 0 taken 8 times.
✗ Branch 1 not taken.
|
8 | if (!detail::is_registered(d)) { |
| 187 | ✗ | std::string m = "brotensor: DeviceScope: backend "; | |
| 188 | ✗ | m += device_name(d); | |
| 189 | ✗ | m += " is not available"; | |
| 190 | ✗ | throw std::runtime_error(m); | |
| 191 | ✗ | } | |
| 192 |
2/2✓ Branch 0 taken 1 time.
✓ Branch 1 taken 7 times.
|
8 | if (tls_scope_override.has_value()) { |
| 193 | 1 | prev_ = *tls_scope_override; | |
| 194 | 1 | tls_scope_had_prev.push_back(true); | |
| 195 | 1 | } else { | |
| 196 | 7 | prev_ = Device::CPU; // unused when had_prev=false | |
| 197 | 7 | tls_scope_had_prev.push_back(false); | |
| 198 | } | ||
| 199 | 8 | tls_scope_override = d; | |
| 200 | 16 | } | |
| 201 | |||
| 202 | 16 | DeviceScope::~DeviceScope() { | |
| 203 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 8 times.
✗ Branch 3 not taken.
|
8 | bool had = tls_scope_had_prev.empty() ? false : tls_scope_had_prev.back(); |
| 204 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 8 times.
✗ Branch 3 not taken.
|
8 | if (!tls_scope_had_prev.empty()) tls_scope_had_prev.pop_back(); |
| 205 |
3/4✓ Branch 0 taken 1 time.
✓ Branch 1 taken 7 times.
✓ Branch 2 taken 1 time.
✗ Branch 3 not taken.
|
8 | if (had) tls_scope_override = prev_; |
| 206 | 7 | else tls_scope_override.reset(); | |
| 207 | 16 | } | |
| 208 | |||
| 209 | // ─── sync ────────────────────────────────────────────────────────────────── | ||
| 210 | |||
| 211 | 11 | void sync(Device d) { | |
| 212 |
1/2✓ Branch 0 taken 11 times.
✗ Branch 1 not taken.
|
11 | if (!detail::is_registered(d)) { |
| 213 | ✗ | std::string m = "brotensor: sync: backend "; | |
| 214 | ✗ | m += device_name(d); | |
| 215 | ✗ | m += " is not available"; | |
| 216 | ✗ | throw std::runtime_error(m); | |
| 217 | ✗ | } | |
| 218 | 11 | detail::alloc_for(d).sync(d.index); | |
| 219 | 11 | } | |
| 220 | |||
| 221 | 2914 | void sync_all() { | |
| 222 |
2/2✓ Branch 0 taken 2914 times.
✓ Branch 1 taken 5828 times.
|
8742 | for (Device d : available_devices()) { |
| 223 |
2/4✓ Branch 0 taken 5828 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 5828 times.
✗ Branch 3 not taken.
|
5828 | detail::alloc_for(d).sync(d.index); |
| 224 | } | ||
| 225 | 2914 | } | |
| 226 | |||
| 227 | // ─── memory ──────────────────────────────────────────────────────────────── | ||
| 228 | |||
| 229 | ✗ | bool device_mem_info(Device d, std::size_t& free_bytes, | |
| 230 | std::size_t& total_bytes) { | ||
| 231 | ✗ | if (!detail::is_registered(d)) return false; | |
| 232 | ✗ | const auto fn = detail::alloc_for(d).mem_info; | |
| 233 | ✗ | if (fn == nullptr) return false; | |
| 234 | ✗ | return fn(&free_bytes, &total_bytes, d.index); | |
| 235 | ✗ | } | |
| 236 | |||
| 237 | ✗ | bool device_mem_trim(Device d, std::size_t keep_bytes) { | |
| 238 | ✗ | if (!detail::is_registered(d)) return false; | |
| 239 | ✗ | const auto fn = detail::alloc_for(d).mem_trim; | |
| 240 | ✗ | if (fn == nullptr) return false; | |
| 241 | ✗ | return fn(keep_bytes, d.index); | |
| 242 | ✗ | } | |
| 243 | |||
| 244 | ✗ | std::string device_product_name(Device d) { | |
| 245 | ✗ | if (!detail::is_registered(d)) return {}; | |
| 246 | ✗ | const auto fn = detail::alloc_for(d).device_name; | |
| 247 | ✗ | if (fn == nullptr) return {}; | |
| 248 | ✗ | char buf[256] = {0}; | |
| 249 | ✗ | if (!fn(buf, sizeof(buf), d.index)) return {}; | |
| 250 | ✗ | return std::string(buf); | |
| 251 | ✗ | } | |
| 252 | |||
| 253 | } // namespace brotensor | ||
| 254 |