diff options
author | Marat Dukhan <maratek@google.com> | 2020-03-05 13:38:29 -0800 |
---|---|---|
committer | Marat Dukhan <maratek@google.com> | 2020-03-05 13:38:29 -0800 |
commit | ef23a4af82e34a4785c56b5fbae631efbcb77aea (patch) | |
tree | 1e84c34157bb96ce9f0dfced1ae2324d86481d1d | |
parent | efa3c028d0f7b0414f0738a706f8187430425477 (diff) | |
download | pthreadpool-ef23a4af82e34a4785c56b5fbae631efbcb77aea.tar.gz |
PTHREADPOOL_FLAG_YIELD_WORKERS flag to bypass spin-wait
Makes it possible to signal the last operation in a sequence of computations,
so pthreadpool workers don't spin in vain.
-rw-r--r-- | include/pthreadpool.h | 9 | ||||
-rw-r--r-- | src/threadpool-pthreads.c | 28 |
2 files changed, 25 insertions, 12 deletions
diff --git a/include/pthreadpool.h b/include/pthreadpool.h index d15bf3f..30471e6 100644 --- a/include/pthreadpool.h +++ b/include/pthreadpool.h @@ -32,6 +32,15 @@ typedef void (*pthreadpool_task_6d_tile_2d_t)(void*, size_t, size_t, size_t, siz */ #define PTHREADPOOL_FLAG_DISABLE_DENORMALS 0x00000001 +/** + * Yield worker threads to the system scheduler after the operation is finished. + * + * Force workers to use kernel wait (instead of active spin-wait by default) for new commands after this command is + * processed. This flag affects only the immediate next operation on this thread pool. To make the thread pool always + * use kernel wait, pass this flag to all parallelization functions. + */ +#define PTHREADPOOL_FLAG_YIELD_WORKERS 0x00000002 + #ifdef __cplusplus extern "C" { #endif diff --git a/src/threadpool-pthreads.c b/src/threadpool-pthreads.c index 4bbf427..955d0b9 100644 --- a/src/threadpool-pthreads.c +++ b/src/threadpool-pthreads.c @@ -335,7 +335,8 @@ static void thread_parallelize_1d(struct pthreadpool* threadpool, struct thread_ static uint32_t wait_for_new_command( struct pthreadpool* threadpool, - uint32_t last_command) + uint32_t last_command, + uint32_t last_flags) { uint32_t command = atomic_load_explicit(&threadpool->command, memory_order_relaxed); if (command != last_command) { @@ -343,19 +344,21 @@ static uint32_t wait_for_new_command( return command; } - /* Spin-wait loop */ - for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) { - /* This fence serves as a sleep instruction */ - atomic_thread_fence(memory_order_acquire); - - command = atomic_load_explicit(&threadpool->command, memory_order_relaxed); - if (command != last_command) { + if ((last_flags & PTHREADPOOL_FLAG_YIELD_WORKERS) == 0) { + /* Spin-wait loop */ + for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) { + /* This fence serves as a sleep instruction */ atomic_thread_fence(memory_order_acquire); - return command; + + command = atomic_load_explicit(&threadpool->command, memory_order_relaxed); + if (command != last_command) { + atomic_thread_fence(memory_order_acquire); + return command; + } } } - /* Spin-wait timed out, fall back to mutex/futex wait */ + /* Spin-wait disabled or timed out, fall back to mutex/futex wait */ #if PTHREADPOOL_USE_FUTEX do { futex_wait(&threadpool->command, last_command); @@ -381,14 +384,15 @@ static void* thread_main(void* arg) { struct pthreadpool* threadpool = ((struct pthreadpool*) (thread - thread->thread_number)) - 1; uint32_t last_command = threadpool_command_init; struct fpu_state saved_fpu_state = { 0 }; + uint32_t flags = 0; /* Check in */ checkin_worker_thread(threadpool); /* Monitor new commands and act accordingly */ for (;;) { - uint32_t command = wait_for_new_command(threadpool, last_command); - const uint32_t flags = atomic_load_explicit(&threadpool->flags, memory_order_relaxed); + uint32_t command = wait_for_new_command(threadpool, last_command, flags); + flags = atomic_load_explicit(&threadpool->flags, memory_order_relaxed); /* Process command */ switch (command & THREADPOOL_COMMAND_MASK) { |