Introduced callback to Pthread, Win32 and OpenMP backend

a, *sb

@@ -0,0 +1,193 @@
+commit 3db0922de7600746113d79773d645b1b1740127c (HEAD -> Threading_Callback)
+Author: shivammonaka <[email protected]>
+Date:   Mon Apr 1 16:02:03 2024 +0530
+
+    Changes to win32
+
+diff --git a/driver/others/blas_server_win32.c b/driver/others/blas_server_win32.c
+index 49d876fc7..9a4916cae 100644
+--- a/driver/others/blas_server_win32.c
++++ b/driver/others/blas_server_win32.c
+@@ -214,7 +214,6 @@ static DWORD WINAPI blas_thread_server(void *arg) {
+ 
+   /* Thread identifier */
+   BLASLONG  cpu = (BLASLONG)arg;
+-
+   blas_queue_t	*queue;
+ 
+   MT_TRACE("Server[%2ld] Thread is started!\n", cpu);
+@@ -243,11 +242,8 @@ static DWORD WINAPI blas_thread_server(void *arg) {
+ 
+     LeaveCriticalSection(&queue_lock);
+ 
+-    if (queue) {
+-
+-      exec_threads(cpu, queue, 0);
+-	  }
+-
++    exec_threads(cpu, queue, 0);
++    
+     MT_TRACE("Server[%2ld] Finished!\n", cpu);
+ 	
+ 	  queue->finished = 1;
+@@ -579,84 +575,90 @@ static void adjust_thread_buffers() {
+ static void exec_threads(int cpu, blas_queue_t *queue, int buf_index)
+ {
+   
+-  void *buffer, *sa, *sb;
++    if (queue) {
++      void *buffer, *sa, *sb;
++      
++      buffer = blas_thread_buffer[cpu];
++      sa = queue -> sa;
++      sb = queue -> sb;
+ 
+-  buffer = blas_thread_buffer[cpu];
+-  sa = queue -> sa;
+-  sb = queue -> sb;
++      int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine;
+ 
+-  int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine;
++      #ifdef CONSISTENT_FPCSR
++        __asm__ __volatile__ ("ldmxcsr %0" : : "m" (queue -> sse_mode));
++        __asm__ __volatile__ ("fldcw %0"   : : "m" (queue -> x87_mode));
++      #endif
+ 
+-  #ifdef CONSISTENT_FPCSR
+-    __asm__ __volatile__ ("ldmxcsr %0" : : "m" (queue -> sse_mode));
+-    __asm__ __volatile__ ("fldcw %0"   : : "m" (queue -> x87_mode));
+-  #endif
++      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);
+ 
+-  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);
++      // fprintf(stderr, "queue start[%ld]!!!\n", cpu);
+ 
+-  // fprintf(stderr, "queue start[%ld]!!!\n", cpu);
++      #ifdef MONITOR
++        main_status[cpu] = MAIN_RUNNING1;
++      #endif
+ 
+-  #ifdef MONITOR
+-    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)) {
++    #ifdef EXPRECISION
++    if ((queue -> mode & BLAS_PREC) == BLAS_XDOUBLE) {
++      sb = (void *)(((BLASLONG)sa + ((XGEMM_P * XGEMM_Q * sizeof(xdouble)
++          + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
++    } else
++    #endif
++      if ((queue -> mode & BLAS_PREC) == BLAS_DOUBLE) {
++    #ifdef BUILD_DOUBLE
++        sb = (void *)(((BLASLONG)sa + ((DGEMM_P * DGEMM_Q * sizeof(double)
++            + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
++    #endif
++      } else if ((queue -> mode & BLAS_PREC) == BLAS_SINGLE) {
++    #ifdef BUILD_SINGLE
++        sb = (void *)(((BLASLONG)sa + ((SGEMM_P * SGEMM_Q * sizeof(float)
++            + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
++    #endif
++      } else {
++            /* Other types in future */
++      }
++    } else {
++    #ifdef EXPRECISION
++    if ((queue -> mode & BLAS_PREC) == BLAS_XDOUBLE){
++      sb = (void *)(((BLASLONG)sa + ((XGEMM_P * XGEMM_Q * 2 * sizeof(xdouble)
++          + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
++    } else
++    #endif
++      if ((queue -> mode & BLAS_PREC) == BLAS_DOUBLE){
++    #ifdef BUILD_COMPLEX16
++        sb = (void *)(((BLASLONG)sa + ((ZGEMM_P * ZGEMM_Q * 2 * sizeof(double)
++            + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
++    #endif
++      } else if ((queue -> mode & BLAS_PREC) == BLAS_SINGLE) {
++    #ifdef BUILD_COMPLEX
++        sb = (void *)(((BLASLONG)sa + ((CGEMM_P * CGEMM_Q * 2 * sizeof(float)
++            + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
++    #endif
++      } else {
++            /* Other types in future */
++      }
++    }
++        queue->sb=sb;
++      }
+ 
+-  if (sb == NULL) {
+-    if (!(queue -> mode & BLAS_COMPLEX)) {
+-#ifdef EXPRECISION
+-if ((queue -> mode & BLAS_PREC) == BLAS_XDOUBLE) {
+-  sb = (void *)(((BLASLONG)sa + ((XGEMM_P * XGEMM_Q * sizeof(xdouble)
+-      + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
+-} else
+-#endif
+-  if ((queue -> mode & BLAS_PREC) == BLAS_DOUBLE) {
+-#ifdef BUILD_DOUBLE
+-    sb = (void *)(((BLASLONG)sa + ((DGEMM_P * DGEMM_Q * sizeof(double)
+-        + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
+-#endif
+-  } else if ((queue -> mode & BLAS_PREC) == BLAS_SINGLE) {
+-#ifdef BUILD_SINGLE
+-    sb = (void *)(((BLASLONG)sa + ((SGEMM_P * SGEMM_Q * sizeof(float)
+-        + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
+-#endif
+-  } else {
+-        /* Other types in future */
+-  }
+-} else {
+-#ifdef EXPRECISION
+-if ((queue -> mode & BLAS_PREC) == BLAS_XDOUBLE){
+-  sb = (void *)(((BLASLONG)sa + ((XGEMM_P * XGEMM_Q * 2 * sizeof(xdouble)
+-      + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
+-} else
+-#endif
+-  if ((queue -> mode & BLAS_PREC) == BLAS_DOUBLE){
+-#ifdef BUILD_COMPLEX16
+-    sb = (void *)(((BLASLONG)sa + ((ZGEMM_P * ZGEMM_Q * 2 * sizeof(double)
+-        + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
+-#endif
+-  } else if ((queue -> mode & BLAS_PREC) == BLAS_SINGLE) {
+-#ifdef BUILD_COMPLEX
+-    sb = (void *)(((BLASLONG)sa + ((CGEMM_P * CGEMM_Q * 2 * sizeof(float)
+-        + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
+-#endif
+-  } else {
+-        /* Other types in future */
+-  }
+-}
+-    queue->sb=sb;
+-  }
++      #ifdef MONITOR
++        main_status[cpu] = MAIN_RUNNING2;
++      #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);
++      }
+ 
+-  #ifdef MONITOR
+-    main_status[cpu] = MAIN_RUNNING2;
+-  #endif
++    } else {
++  		continue; //if queue == NULL
+ 
+-  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);
+-  }
++	  }
+ 
+ }
+\ No newline at end of file

cblas.h

@@ -26,6 +26,11 @@ char* openblas_get_config(void);
 /*Get the CPU corename on runtime.*/
 char* openblas_get_corename(void);
 
+/*Set the threading backend to a custom callback.*/
+typedef void (*openblas_dojob_callback)(int thread_num, void *jobdata, int dojob_data);
+typedef void (*openblas_threads_callback)(int sync, openblas_dojob_callback dojob, int numjobs, size_t jobdata_elsize, void *jobdata, int dojob_data);
+void openblas_set_threads_callback_function(openblas_threads_callback callback);
+
 #ifdef OPENBLAS_OS_LINUX
 /* Sets thread affinity for OpenBLAS threads. `thread_idx` is in [0, openblas_get_num_threads()-1]. */
 int openblas_setaffinity(int thread_idx, size_t cpusetsize, cpu_set_t* cpu_set);

common_interface.h

@@ -47,6 +47,11 @@ int    BLASFUNC(xerbla)(char *, blasint *info, blasint);
 
 void    openblas_set_num_threads_(int *);
 
+/*Set the threading backend to a custom callback.*/
+typedef void (*openblas_dojob_callback)(int thread_num, void *jobdata, int dojob_data);
+typedef void (*openblas_threads_callback)(int sync, openblas_dojob_callback dojob, int numjobs, size_t jobdata_elsize, void *jobdata, int dojob_data);
+extern openblas_threads_callback openblas_threads_callback_;
+
 FLOATRET  BLASFUNC(sdot)  (blasint *, float  *, blasint *, float  *, blasint *);
 FLOATRET  BLASFUNC(sdsdot)(blasint *, float  *,        float  *, blasint *, float  *, blasint *);
 

driver/others/CMakeLists.txt

@@ -25,6 +25,7 @@ if (USE_THREAD)
     ${BLAS_SERVER}
     divtable.c # TODO: Makefile has -UDOUBLE
     blas_l1_thread.c
+    blas_server_callback.c
   )
 
   if (NOT NO_AFFINITY)

driver/others/Makefile

@@ -6,7 +6,7 @@ COMMONOBJS	 = memory.$(SUFFIX) xerbla.$(SUFFIX) c_abs.$(SUFFIX) z_abs.$(SUFFIX)
 #COMMONOBJS	+= slamch.$(SUFFIX) slamc3.$(SUFFIX) dlamch.$(SUFFIX)  dlamc3.$(SUFFIX)
 
 ifdef SMP
-COMMONOBJS	+= blas_server.$(SUFFIX) divtable.$(SUFFIX) blasL1thread.$(SUFFIX)
+COMMONOBJS	+= blas_server.$(SUFFIX) divtable.$(SUFFIX) blasL1thread.$(SUFFIX) blas_server_callback.$(SUFFIX)
 ifneq ($(NO_AFFINITY), 1)
 COMMONOBJS	+= init.$(SUFFIX)
 endif
@@ -140,6 +140,9 @@ memory.$(SUFFIX) : $(MEMORY) ../../common.h ../../param.h
 blas_server.$(SUFFIX) : $(BLAS_SERVER) ../../common.h ../../common_thread.h ../../param.h
 	$(CC) $(CFLAGS) -c $< -o $(@F)
 
+blas_server_callback.$(SUFFIX) : blas_server_callback.c ../../common.h
+	$(CC) $(CFLAGS) -c $< -o $(@F)
+
 openblas_set_num_threads.$(SUFFIX) : openblas_set_num_threads.c
 	$(CC) $(CFLAGS) -c $< -o $(@F)