Merge pull request #2988 from xiegengxin/smp-asum

Improve the performance of dasum and sasum when SMP is defined

kernel/x86_64/dasum.c

@@ -58,21 +58,19 @@ static FLOAT dasum_kernel(BLASLONG n, FLOAT *x1)
 }
 
 #endif
-
-FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+static FLOAT asum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
 {
-    BLASLONG i=0;
+    BLASLONG i = 0;
     FLOAT sumf = 0.0;
+    
+    if (n <= 0 || inc_x <= 0) return (sumf);
 
-    if (n <= 0 || inc_x <= 0) return(sumf);
-
-    if ( inc_x == 1 ) {
+    if (inc_x == 1) {
         sumf = dasum_kernel(n, x);
-    } 
+    }
     else {
         n *= inc_x;
-       
-        while(i < n) {
+        while (i < n) {
             sumf += ABS_K(x[i]);
             i += inc_x;
         }
@@ -80,3 +78,53 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
     return(sumf);
 }
 
+#if defined(SMP)
+static int asum_thread_function(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *dummy3, BLASLONG dummy4, FLOAT *result, BLASLONG dummy5)
+{
+    *(FLOAT *)result = asum_compute(n, x, inc_x);
+    return 0;
+}
+
+extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n, BLASLONG k, void *alpha, void *a, BLASLONG lda, void *b, BLASLONG ldb, void *c, BLASLONG ldc, int (*function)(), int nthreads);
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+#if defined(SMP)
+    int nthreads;
+    FLOAT dummy_alpha;
+    FLOAT * dummy_b;
+#endif
+    FLOAT sumf = 0.0;
+
+#if defined(SMP)
+    int num_cpu = num_cpu_avail(1);
+    if (n <= 100000 || inc_x <= 0)
+        nthreads = 1;
+    else 
+	    nthreads = num_cpu < n/100000 ? num_cpu : n/100000;
+
+    if (nthreads == 1) {
+        sumf = asum_compute(n, x, inc_x);
+    } else {
+        int mode, i;
+        char result[MAX_CPU_NUMBER * sizeof(double) *2];
+        FLOAT *ptr;
+#if !defined(DOUBLE)
+        mode = BLAS_SINGLE | BLAS_REAL;
+#else
+        mode = BLAS_DOUBLE | BLAS_REAL;
+#endif
+        blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, x, inc_x, dummy_b, 0, result, 0, (void *)asum_thread_function, nthreads);
+        ptr = (FLOAT *)result;
+        for (i = 0; i < nthreads; i++) {
+            sumf += (*ptr);
+            ptr = (FLOAT *)(((char *)ptr) + sizeof(double) *2);
+        }
+    }
+#else
+    sumf = asum_compute(n, x, inc_x);
+#endif
+    return(sumf);
+}
+

kernel/x86_64/sasum.c

@@ -67,24 +67,71 @@ static FLOAT sasum_kernel(BLASLONG n, FLOAT *x1)
 
 #endif
 
-FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+static FLOAT asum_compute(BLASLONG n, FLOAT * x, BLASLONG inc_x)
 {
-    BLASLONG i=0;
+    BLASLONG i = 0;
     FLOAT sumf = 0.0;
+    
+    if (n <= 0 || inc_x <= 0) return (sumf);
 
-    if (n <= 0 || inc_x <= 0) return(sumf);
-
-    if ( inc_x == 1 ) {
+    if (inc_x == 1) {
         sumf = sasum_kernel(n, x);
     }
     else {
-
         n *= inc_x;
         while(i < n) {
             sumf += ABS_K(x[i]);
             i += inc_x;
         }
+    }
+    return (sumf);
+}
 
+#if defined(SMP)
+static int asum_thread_function(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *dummy3, BLASLONG dummy4, FLOAT *result, BLASLONG dummy5)
+{
+    *(FLOAT *)result = asum_compute(n, x, inc_x);
+    return 0;
+}
+
+extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n, BLASLONG k, void * alpha, void *a, BLASLONG lda, void *b, BLASLONG ldb, void *c, BLASLONG ldc, int(*function)(), int nthreads);
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+#if defined(SMP)
+    int nthreads;
+    FLOAT dummy_alpha;
+#endif
+    FLOAT sumf = 0.0;
+
+#if defined(SMP)
+    int num_cpu = num_cpu_avail(1);
+    if (n <= 100000 || inc_x <= 0)
+        nthreads = 1;
+    else
+        nthreads = num_cpu < n/100000 ? num_cpu : n/100000;
+    if (nthreads == 1) {
+        sumf = asum_compute(n, x, inc_x);
     }
+    else {
+        int mode, i;
+        char result[MAX_CPU_NUMBER * sizeof(double) *2];
+        FLOAT * ptr;
+#if !defined(DOUBLE)
+        mode = BLAS_SINGLE | BLAS_REAL;
+#else
+        mode = BLAS_DOUBLE | BLAS_REAL;
+#endif
+        blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, x, inc_x, NULL, 0, result, 0, (void *)asum_thread_function, nthreads);
+        ptr = (FLOAT *)result;
+        for (i = 0; i < nthreads; i++) {
+            sumf += (*ptr);
+            ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
+        }
+    }
+#else
+    sumf = asum_compute(n, x, inc_x);
+#endif
     return(sumf);
 }