floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

clean up whitespace

This commit is contained in:
Mark Seminatore 2024-01-29 22:33:47 -08:00
parent 6bd7c54af5
commit 0d7fe5ea61
1 changed files with 85 additions and 107 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

120
driver/others/blas_server_win32.c
View File

@ -72,19 +72,9 @@ static HANDLE blas_threads [MAX_CPU_NUMBER];
static DWORD blas_threads_id[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 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
static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
if (!(mode & BLAS_COMPLEX)){ if (!(mode & BLAS_COMPLEX)) {
#ifdef EXPRECISION #ifdef EXPRECISION
if ((mode & BLAS_PREC) == BLAS_XDOUBLE){ if ((mode & BLAS_PREC) == BLAS_XDOUBLE){
/* REAL / Extended Double */ /* REAL / Extended Double */
@ -99,7 +89,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
args -> c, args -> ldc, sb); args -> c, args -> ldc, sb);
} else } else
#endif #endif
if ((mode & BLAS_PREC) == BLAS_DOUBLE){ if ((mode & BLAS_PREC) == BLAS_DOUBLE) {
/* REAL / Double */ /* REAL / Double */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double,
double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG,
@ -110,7 +100,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
args -> a, args -> lda, args -> a, args -> lda,
args -> b, args -> ldb, args -> b, args -> ldb,
args -> c, args -> ldc, sb); args -> c, args -> ldc, sb);
} else if ((mode & BLAS_PREC) == BLAS_SINGLE){ } else if ((mode & BLAS_PREC) == BLAS_SINGLE) {
/* REAL / Single */ /* REAL / Single */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float,
float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG,
@ -122,7 +112,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
args -> b, args -> ldb, args -> b, args -> ldb,
args -> c, args -> ldc, sb); args -> c, args -> ldc, sb);
#ifdef BUILD_BFLOAT16 #ifdef BUILD_BFLOAT16
} else if ((mode & BLAS_PREC) == BLAS_BFLOAT16){ } else if ((mode & BLAS_PREC) == BLAS_BFLOAT16) {
/* REAL / BFLOAT16 */ /* REAL / BFLOAT16 */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, bfloat16, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, bfloat16,
bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG,
@ -133,7 +123,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
args -> a, args -> lda, args -> a, args -> lda,
args -> b, args -> ldb, args -> b, args -> ldb,
args -> c, args -> ldc, sb); args -> c, args -> ldc, sb);
} else if ((mode & BLAS_PREC) == BLAS_STOBF16){ } else if ((mode & BLAS_PREC) == BLAS_STOBF16) {
/* REAL / BLAS_STOBF16 */ /* REAL / BLAS_STOBF16 */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float,
float *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG, bfloat16 *, BLASLONG,
@ -144,7 +134,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
args -> a, args -> lda, args -> a, args -> lda,
args -> b, args -> ldb, args -> b, args -> ldb,
args -> c, args -> ldc, sb); args -> c, args -> ldc, sb);
} else if ((mode & BLAS_PREC) == BLAS_DTOBF16){ } else if ((mode & BLAS_PREC) == BLAS_DTOBF16) {
/* REAL / BLAS_DTOBF16 */ /* REAL / BLAS_DTOBF16 */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double,
double *, BLASLONG, bfloat16 *, BLASLONG, double *, BLASLONG, bfloat16 *, BLASLONG,
@ -161,7 +151,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
} }
} else { } else {
#ifdef EXPRECISION #ifdef EXPRECISION
if ((mode & BLAS_PREC) == BLAS_XDOUBLE){ if ((mode & BLAS_PREC) == BLAS_XDOUBLE) {
/* COMPLEX / Extended Double */ /* COMPLEX / Extended Double */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble,
xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG,
@ -175,7 +165,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
args -> c, args -> ldc, sb); args -> c, args -> ldc, sb);
} else } else
#endif #endif
if ((mode & BLAS_PREC) == BLAS_DOUBLE){ if ((mode & BLAS_PREC) == BLAS_DOUBLE) {
/* COMPLEX / Double */ /* COMPLEX / Double */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, double, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, double,
double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG,
@ -205,10 +195,9 @@ 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 */ // This is a main routine of threads. Each thread waits until job is
/* queued. */ // queued.
static DWORD WINAPI blas_thread_server(void *arg) {
static DWORD WINAPI blas_thread_server(void *arg){
/* Thread identifier */ /* Thread identifier */
BLASLONG cpu = (BLASLONG)arg; BLASLONG cpu = (BLASLONG)arg;
@ -221,7 +210,7 @@ static DWORD WINAPI blas_thread_server(void *arg){
MT_TRACE("Server[%2ld] Thread is started!\n", cpu); MT_TRACE("Server[%2ld] Thread is started!\n", cpu);
while (1){ while (1) {
/* Waiting for Queue */ /* Waiting for Queue */
@ -230,15 +219,13 @@ static DWORD WINAPI blas_thread_server(void *arg){
// event raised when work is added to the queue // event raised when work is added to the queue
WaitForSingleObject(kickoff_event, INFINITE); WaitForSingleObject(kickoff_event, INFINITE);
if (cpu > thread_target - 2) if (cpu > thread_target - 2) {
{
//MT_TRACE("thread [%d] exiting.\n", cpu); //MT_TRACE("thread [%d] exiting.\n", cpu);
break; // excess thread, so worker thread exits break; // excess thread, so worker thread exits
} }
MT_TRACE("Server[%2ld] Got it.\n", cpu); MT_TRACE("Server[%2ld] Got it.\n", cpu);
#if 1
EnterCriticalSection(&queue_lock); EnterCriticalSection(&queue_lock);
queue = work_queue; queue = work_queue;
@ -246,19 +233,6 @@ static DWORD WINAPI blas_thread_server(void *arg){
work_queue = work_queue->next; work_queue = work_queue->next;
LeaveCriticalSection(&queue_lock); 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) { if (queue) {
int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine; int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine;
@ -266,31 +240,32 @@ static DWORD WINAPI blas_thread_server(void *arg){
sa = queue -> sa; sa = queue -> sa;
sb = queue -> sb; sb = queue -> sb;
#ifdef CONSISTENT_FPCSR #ifdef CONSISTENT_FPCSR
__asm__ __volatile__ ("ldmxcsr %0" : : "m" (queue -> sse_mode)); __asm__ __volatile__ ("ldmxcsr %0" : : "m" (queue -> sse_mode));
__asm__ __volatile__ ("fldcw %0" : : "m" (queue -> x87_mode)); __asm__ __volatile__ ("fldcw %0" : : "m" (queue -> x87_mode));
#endif #endif
MT_TRACE("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); cpu, queue->mode, queue-> args ->m, queue->args->n, queue->args->k);
// fprintf(stderr, "queue start[%ld]!!!\n", cpu); // fprintf(stderr, "queue start[%ld]!!!\n", cpu);
#ifdef MONITOR #ifdef MONITOR
main_status[cpu] = MAIN_RUNNING1; main_status[cpu] = MAIN_RUNNING1;
#endif #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 (sb == NULL) {
if (!(queue -> mode & BLAS_COMPLEX)){ if (!(queue -> mode & BLAS_COMPLEX)) {
#ifdef EXPRECISION #ifdef EXPRECISION
if ((queue -> mode & BLAS_PREC) == BLAS_XDOUBLE){ if ((queue -> mode & BLAS_PREC) == BLAS_XDOUBLE) {
sb = (void *)(((BLASLONG)sa + ((XGEMM_P * XGEMM_Q * sizeof(xdouble) sb = (void *)(((BLASLONG)sa + ((XGEMM_P * XGEMM_Q * sizeof(xdouble)
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B); + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
} else } else
#endif #endif
if ((queue -> mode & BLAS_PREC) == BLAS_DOUBLE){ if ((queue -> mode & BLAS_PREC) == BLAS_DOUBLE) {
#ifdef BUILD_DOUBLE #ifdef BUILD_DOUBLE
sb = (void *)(((BLASLONG)sa + ((DGEMM_P * DGEMM_Q * sizeof(double) sb = (void *)(((BLASLONG)sa + ((DGEMM_P * DGEMM_Q * sizeof(double)
+ GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B); + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
@ -327,17 +302,16 @@ static DWORD WINAPI blas_thread_server(void *arg){
queue->sb=sb; queue->sb=sb;
} }
#ifdef MONITOR #ifdef MONITOR
main_status[cpu] = MAIN_RUNNING2; main_status[cpu] = MAIN_RUNNING2;
#endif #endif
if (!(queue -> mode & BLAS_LEGACY)) { if (!(queue -> mode & BLAS_LEGACY)) {
(routine)(queue -> args, queue -> range_m, queue -> range_n, sa, sb, queue -> position); (routine)(queue -> args, queue -> range_m, queue -> range_n, sa, sb, queue -> position);
} else { } else {
legacy_exec(routine, queue -> mode, queue -> args, sb); legacy_exec(routine, queue -> mode, queue -> args, sb);
} }
}else{ } else {
continue; //if queue == NULL continue; //if queue == NULL
} }
@ -353,10 +327,12 @@ static DWORD WINAPI blas_thread_server(void *arg){
blas_memory_free(buffer); blas_memory_free(buffer);
return 0; return 0;
} }
/* Initializing routine */ //
int blas_thread_init(void){ // Initializing routine
//
int blas_thread_init(void) {
BLASLONG i; BLASLONG i;
if (blas_server_avail || (blas_cpu_number <= 1)) return 0; if (blas_server_avail || (blas_cpu_number <= 1)) return 0;
@ -365,7 +341,7 @@ int blas_thread_init(void){
MT_TRACE("Initializing Thread(Num. threads = %d)\n", blas_cpu_number); MT_TRACE("Initializing Thread(Num. threads = %d)\n", blas_cpu_number);
if (!blas_server_avail){ if (!blas_server_avail) {
// create the kickoff Event // create the kickoff Event
kickoff_event = CreateEvent(NULL, TRUE, FALSE, NULL); kickoff_event = CreateEvent(NULL, TRUE, FALSE, NULL);
@ -373,7 +349,7 @@ int blas_thread_init(void){
InitializeCriticalSection(&queue_lock); InitializeCriticalSection(&queue_lock);
for(i = 0; i < blas_cpu_number - 1; i++){ for(i = 0; i < blas_cpu_number - 1; i++) {
//MT_TRACE("thread_init: creating thread [%d]\n", i); //MT_TRACE("thread_init: creating thread [%d]\n", i);
blas_threads[i] = CreateThread(NULL, 0, blas_threads[i] = CreateThread(NULL, 0,
@ -391,13 +367,10 @@ int blas_thread_init(void){
/* /*
User can call one of two routines. User can call one of two routines.
exec_blas_async ... immediately returns after jobs are queued. exec_blas_async ... immediately returns after jobs are queued.
exec_blas ... returns after jobs are finished. exec_blas ... returns after jobs are finished.
*/ */
int exec_blas_async(BLASLONG pos, blas_queue_t *queue) {
int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
#if defined(SMP_SERVER) #if defined(SMP_SERVER)
// Handle lazy re-init of the thread-pool after a POSIX fork // Handle lazy re-init of the thread-pool after a POSIX fork
@ -447,12 +420,16 @@ int exec_blas_async(BLASLONG pos, blas_queue_t *queue){
return 0; return 0;
} }
int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue){ //
// Join. Wait for all queued tasks to complete
//
int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue) {
MT_TRACE("Synchronization Waiting.\n"); MT_TRACE("Synchronization Waiting.\n");
while (num){ while (num) {
MT_TRACE("Waiting Queue ..\n"); MT_TRACE("Waiting Queue ..\n");
while (!queue->finished) while (!queue->finished)
YIELDING; YIELDING;
@ -474,8 +451,10 @@ int exec_blas_async_wait(BLASLONG num, blas_queue_t *queue){
return 0; return 0;
} }
/* Execute Threads */ //
int exec_blas(BLASLONG num, blas_queue_t *queue){ // Execute Threads
//
int exec_blas(BLASLONG num, blas_queue_t *queue) {
#if defined(SMP_SERVER) && defined(OS_CYGWIN_NT) #if defined(SMP_SERVER) && defined(OS_CYGWIN_NT)
// Handle lazy re-init of the thread-pool after a POSIX fork // Handle lazy re-init of the thread-pool after a POSIX fork
@ -507,9 +486,8 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
return 0; return 0;
} }
/* Shutdown procedure, but user don't have to call this routine. The */ // Shutdown procedure, but user don't have to call this routine. The
/* kernel automatically kill threads. */ // kernel automatically kill threads.
int BLASFUNC(blas_thread_shutdown)(void){ int BLASFUNC(blas_thread_shutdown)(void){
int i; int i;
@ -518,9 +496,9 @@ int BLASFUNC(blas_thread_shutdown)(void){
LOCK_COMMAND(&server_lock); LOCK_COMMAND(&server_lock);
if (blas_server_avail){ if (blas_server_avail) {
for(i = 0; i < blas_num_threads - 1; i++){ for(i = 0; i < blas_num_threads - 1; i++) {
// Could also just use WaitForMultipleObjects // Could also just use WaitForMultipleObjects
DWORD wait_thread_value = WaitForSingleObject(blas_threads[i], 50); DWORD wait_thread_value = WaitForSingleObject(blas_threads[i], 50);
@ -586,7 +564,7 @@ void goto_set_num_threads(int num_threads)
thread_target = num_threads; thread_target = num_threads;
//increased_threads = 1; //increased_threads = 1;
if (!blas_server_avail){ if (!blas_server_avail) {
// create the kickoff Event // create the kickoff Event
kickoff_event = CreateEvent(NULL, TRUE, FALSE, NULL); kickoff_event = CreateEvent(NULL, TRUE, FALSE, NULL);
@ -595,7 +573,7 @@ void goto_set_num_threads(int num_threads)
blas_server_avail = 1; blas_server_avail = 1;
} }
for(i = (blas_num_threads > 0) ? blas_num_threads - 1 : 0; i < num_threads - 1; i++){ for(i = (blas_num_threads > 0) ? blas_num_threads - 1 : 0; i < num_threads - 1; i++) {
//MT_TRACE("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_threads[i] = CreateThread(NULL, 0,