Merge pull request #4498 from mseminatore/win_tidy
blas_server_win32.c pass to clean up code
This commit is contained in:
Martin Kroeker
GitHub
commit
5266998b9f
|
|
@ -219,6 +219,7 @@ In chronological order:
|
|||
|
||||
* Mark Seminatore <https://github.com/mseminatore>
|
||||
* [2023-11-09] Improve Windows threading performance scaling
|
||||
* [2024-02-09] Introduce MT_TRACE facility and improve code consistency
|
||||
|
||||
* Dirreke <https://github.com/mseminatore>
|
||||
* [2024-01-16] Add basic support for the CSKY architecture
|
||||
|
|
|
|||
|
|
@ -48,6 +48,12 @@
|
|||
#endif
|
||||
#endif
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
# define MT_TRACE(...) fprintf(stderr, __VA_ARGS__)
|
||||
#else
|
||||
# define MT_TRACE(...)
|
||||
#endif
|
||||
|
||||
/* This is a thread implementation for Win32 lazy implementation */
|
||||
|
||||
/* Thread server common information */
|
||||
|
|
@ -68,16 +74,9 @@ static HANDLE blas_threads [MAX_CPU_NUMBER];
|
|||
static DWORD blas_threads_id[MAX_CPU_NUMBER];
|
||||
static volatile int thread_target; // target num of live threads, volatile for cross-thread reads
|
||||
|
||||
#if defined (__GNUC__) && (__GNUC__ < 6)
|
||||
#define WIN_CAS(dest, exch, comp) __sync_val_compare_and_swap(dest, comp, exch)
|
||||
#else
|
||||
#if defined(_WIN64)
|
||||
#define WIN_CAS(dest, exch, comp) InterlockedCompareExchange64(dest, exch, comp)
|
||||
#else
|
||||
#define WIN_CAS(dest, exch, comp) InterlockedCompareExchange(dest, exch, comp)
|
||||
#endif
|
||||
#endif
|
||||
|
||||
//
|
||||
// Legacy code path
|
||||
//
|
||||
static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb) {
|
||||
|
||||
if (!(mode & BLAS_COMPLEX)) {
|
||||
|
|
@ -201,9 +200,9 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
|
|||
}
|
||||
}
|
||||
|
||||
/* This is a main routine of threads. Each thread waits until job is */
|
||||
/* queued. */
|
||||
|
||||
//
|
||||
// This is a main routine of threads. Each thread waits until job is queued.
|
||||
//
|
||||
static DWORD WINAPI blas_thread_server(void *arg) {
|
||||
|
||||
/* Thread identifier */
|
||||
|
|
@ -215,31 +214,24 @@ static DWORD WINAPI blas_thread_server(void *arg){
|
|||
/* Each server needs each buffer */
|
||||
buffer = blas_memory_alloc(2);
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Server[%2ld] Thread is started!\n", cpu);
|
||||
#endif
|
||||
MT_TRACE("Server[%2ld] Thread is started!\n", cpu);
|
||||
|
||||
while (1) {
|
||||
|
||||
/* Waiting for Queue */
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Server[%2ld] Waiting for Queue.\n", cpu);
|
||||
#endif
|
||||
MT_TRACE("Server[%2ld] Waiting for Queue.\n", cpu);
|
||||
|
||||
// event raised when work is added to the queue
|
||||
WaitForSingleObject(kickoff_event, INFINITE);
|
||||
|
||||
if (cpu > thread_target - 2)
|
||||
{
|
||||
//printf("thread [%d] exiting.\n", cpu);
|
||||
if (cpu > thread_target - 2) {
|
||||
//MT_TRACE("thread [%d] exiting.\n", cpu);
|
||||
break; // excess thread, so worker thread exits
|
||||
}
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Server[%2ld] Got it.\n", cpu);
|
||||
#endif
|
||||
MT_TRACE("Server[%2ld] Got it.\n", cpu);
|
||||
|
||||
#if 1
|
||||
EnterCriticalSection(&queue_lock);
|
||||
|
||||
queue = work_queue;
|
||||
|
|
@ -247,19 +239,6 @@ static DWORD WINAPI blas_thread_server(void *arg){
|
|||
work_queue = work_queue->next;
|
||||
|
||||
LeaveCriticalSection(&queue_lock);
|
||||
#else
|
||||
volatile blas_queue_t* queue_next;
|
||||
|
||||
INT_PTR prev_value;
|
||||
do {
|
||||
queue = (volatile blas_queue_t*)work_queue;
|
||||
if (!queue)
|
||||
break;
|
||||
|
||||
queue_next = (volatile blas_queue_t*)queue->next;
|
||||
prev_value = WIN_CAS((INT_PTR*)&work_queue, (INT_PTR)queue_next, (INT_PTR)queue);
|
||||
} while (prev_value != queue);
|
||||
#endif
|
||||
|
||||
if (queue) {
|
||||
int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine;
|
||||
|
|
@ -272,10 +251,8 @@ static DWORD WINAPI blas_thread_server(void *arg){
|
|||
__asm__ __volatile__ ("fldcw %0" : : "m" (queue -> x87_mode));
|
||||
#endif
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Server[%2ld] Started. Mode = 0x%03x M = %3ld N=%3ld K=%3ld\n",
|
||||
MT_TRACE("Server[%2ld] Started. Mode = 0x%03x M = %3ld N=%3ld K=%3ld\n",
|
||||
cpu, queue->mode, queue-> args ->m, queue->args->n, queue->args->k);
|
||||
#endif
|
||||
|
||||
// fprintf(stderr, "queue start[%ld]!!!\n", cpu);
|
||||
|
||||
|
|
@ -283,7 +260,8 @@ static DWORD WINAPI blas_thread_server(void *arg){
|
|||
main_status[cpu] = MAIN_RUNNING1;
|
||||
#endif
|
||||
|
||||
if (sa == NULL) sa = (void *)((BLASLONG)buffer + GEMM_OFFSET_A);
|
||||
if (sa == NULL)
|
||||
sa = (void *)((BLASLONG)buffer + GEMM_OFFSET_A);
|
||||
|
||||
if (sb == NULL) {
|
||||
if (!(queue -> mode & BLAS_COMPLEX)) {
|
||||
|
|
@ -335,7 +313,6 @@ static DWORD WINAPI blas_thread_server(void *arg){
|
|||
#endif
|
||||
|
||||
if (!(queue -> mode & BLAS_LEGACY)) {
|
||||
|
||||
(routine)(queue -> args, queue -> range_m, queue -> range_n, sa, sb, queue -> position);
|
||||
} else {
|
||||
legacy_exec(routine, queue -> mode, queue -> args, sb);
|
||||
|
|
@ -344,26 +321,23 @@ static DWORD WINAPI blas_thread_server(void *arg){
|
|||
continue; //if queue == NULL
|
||||
}
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Server[%2ld] Finished!\n", cpu);
|
||||
#endif
|
||||
MT_TRACE("Server[%2ld] Finished!\n", cpu);
|
||||
|
||||
queue->finished = 1;
|
||||
|
||||
}
|
||||
|
||||
/* Shutdown procedure */
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Server[%2ld] Shutdown!\n", cpu);
|
||||
#endif
|
||||
MT_TRACE("Server[%2ld] Shutdown!\n", cpu);
|
||||
|
||||
blas_memory_free(buffer);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
/* Initializing routine */
|
||||
//
|
||||
// Initializing routine
|
||||
//
|
||||
int blas_thread_init(void) {
|
||||
BLASLONG i;
|
||||
|
||||
|
|
@ -371,10 +345,7 @@ int blas_thread_init(void){
|
|||
|
||||
LOCK_COMMAND(&server_lock);
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Initializing Thread(Num. threads = %d)\n",
|
||||
blas_cpu_number);
|
||||
#endif
|
||||
MT_TRACE("Initializing Thread(Num. threads = %d)\n", blas_cpu_number);
|
||||
|
||||
if (!blas_server_avail) {
|
||||
// create the kickoff Event
|
||||
|
|
@ -385,7 +356,7 @@ int blas_thread_init(void){
|
|||
InitializeCriticalSection(&queue_lock);
|
||||
|
||||
for(i = 0; i < blas_cpu_number - 1; i++) {
|
||||
//printf("thread_init: creating thread [%d]\n", i);
|
||||
//MT_TRACE("thread_init: creating thread [%d]\n", i);
|
||||
|
||||
blas_threads[i] = CreateThread(NULL, 0,
|
||||
blas_thread_server, (void *)i,
|
||||
|
|
@ -400,14 +371,11 @@ int blas_thread_init(void){
|
|||
return 0;
|
||||
}
|
||||
|
||||
/*
|
||||
User can call one of two routines.
|
||||
|
||||
exec_blas_async ... immediately returns after jobs are queued.
|
||||
|
||||
exec_blas ... returns after jobs are finished.
|
||||
*/
|
||||
|
||||
//
|
||||
// User can call one of two routines.
|
||||
// exec_blas_async ... immediately returns after jobs are queued.
|
||||
// exec_blas ... returns after jobs are finished.
|
||||
//
|
||||
int exec_blas_async(BLASLONG pos, blas_queue_t *queue) {
|
||||
|
||||
#if defined(SMP_SERVER)
|
||||
|
|
@ -458,16 +426,16 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
|
|||
return 0;
|
||||
}
|
||||
|
||||
//
|
||||
// Join. Wait for all queued tasks to complete
|
||||
//
|
||||
int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue) {
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Synchronization Waiting.\n");
|
||||
#endif
|
||||
MT_TRACE("Synchronization Waiting.\n");
|
||||
|
||||
while (num) {
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Waiting Queue ..\n");
|
||||
#endif
|
||||
MT_TRACE("Waiting Queue ..\n");
|
||||
|
||||
while (!queue->finished)
|
||||
YIELDING;
|
||||
|
||||
|
|
@ -475,9 +443,8 @@ int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue){
|
|||
num--;
|
||||
}
|
||||
|
||||
#ifdef SMP_DEBUG
|
||||
fprintf(STDERR, "Completely Done.\n\n");
|
||||
#endif
|
||||
MT_TRACE("Completely Done.\n\n");
|
||||
|
||||
// if work was added to the queue after this batch we can't sleep the worker threads
|
||||
// by resetting the event
|
||||
EnterCriticalSection(&queue_lock);
|
||||
|
|
@ -490,7 +457,9 @@ int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue){
|
|||
return 0;
|
||||
}
|
||||
|
||||
/* Execute Threads */
|
||||
//
|
||||
// Execute Threads
|
||||
//
|
||||
int exec_blas(BLASLONG num, blas_queue_t *queue) {
|
||||
|
||||
#if defined(SMP_SERVER) && defined(OS_CYGWIN_NT)
|
||||
|
|
@ -504,28 +473,32 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
|
|||
|
||||
if ((num <= 0) || (queue == NULL)) return 0;
|
||||
|
||||
if ((num > 1) && queue -> next) exec_blas_async(1, queue -> next);
|
||||
if ((num > 1) && queue -> next)
|
||||
exec_blas_async(1, queue -> next);
|
||||
|
||||
routine = queue -> routine;
|
||||
|
||||
if (queue -> mode & BLAS_LEGACY) {
|
||||
legacy_exec(routine, queue -> mode, queue -> args, queue -> sb);
|
||||
} else
|
||||
} else {
|
||||
if (queue -> mode & BLAS_PTHREAD) {
|
||||
void (*pthreadcompat)(void *) = queue -> routine;
|
||||
(pthreadcompat)(queue -> args);
|
||||
} else
|
||||
(routine)(queue -> args, queue -> range_m, queue -> range_n,
|
||||
queue -> sa, queue -> sb, 0);
|
||||
}
|
||||
|
||||
if ((num > 1) && queue -> next) exec_blas_async_wait(num - 1, queue -> next);
|
||||
if ((num > 1) && queue -> next)
|
||||
exec_blas_async_wait(num - 1, queue -> next);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
/* Shutdown procedure, but user don't have to call this routine. The */
|
||||
/* kernel automatically kill threads. */
|
||||
|
||||
//
|
||||
// Shutdown procedure, but user don't have to call this routine. The
|
||||
// kernel automatically kill threads.
|
||||
//
|
||||
int BLASFUNC(blas_thread_shutdown)(void) {
|
||||
|
||||
int i;
|
||||
|
|
@ -558,6 +531,9 @@ int BLASFUNC(blas_thread_shutdown)(void){
|
|||
return 0;
|
||||
}
|
||||
|
||||
//
|
||||
// Legacy function to set numbef of threads
|
||||
//
|
||||
void goto_set_num_threads(int num_threads)
|
||||
{
|
||||
long i;
|
||||
|
|
@ -579,11 +555,11 @@ void goto_set_num_threads(int num_threads)
|
|||
SetEvent(kickoff_event);
|
||||
|
||||
for (i = num_threads - 1; i < blas_num_threads - 1; i++) {
|
||||
//printf("set_num_threads: waiting on thread [%d] to quit.\n", i);
|
||||
//MT_TRACE("set_num_threads: waiting on thread [%d] to quit.\n", i);
|
||||
|
||||
WaitForSingleObject(blas_threads[i], INFINITE);
|
||||
|
||||
//printf("set_num_threads: thread [%d] has quit.\n", i);
|
||||
//MT_TRACE("set_num_threads: thread [%d] has quit.\n", i);
|
||||
|
||||
CloseHandle(blas_threads[i]);
|
||||
}
|
||||
|
|
@ -612,7 +588,7 @@ void goto_set_num_threads(int num_threads)
|
|||
}
|
||||
|
||||
for (i = (blas_num_threads > 0) ? blas_num_threads - 1 : 0; i < num_threads - 1; i++) {
|
||||
//printf("set_num_threads: creating thread [%d]\n", i);
|
||||
//MT_TRACE("set_num_threads: creating thread [%d]\n", i);
|
||||
|
||||
blas_threads[i] = CreateThread(NULL, 0,
|
||||
blas_thread_server, (void *)i,
|
||||
|
|
@ -627,6 +603,9 @@ void goto_set_num_threads(int num_threads)
|
|||
blas_cpu_number = num_threads;
|
||||
}
|
||||
|
||||
//
|
||||
// Openblas function to set thread count
|
||||
//
|
||||
void openblas_set_num_threads(int num)
|
||||
{
|
||||
goto_set_num_threads(num);
|
||||
|
|
|
|||
Loading…
Reference in New Issue