Add shortcuts for (small) cases that do not need expensive buffer allocation

interface/ger.c

@@ -164,6 +164,11 @@ void CNAME(enum CBLAS_ORDER order,
   if (m == 0 || n == 0) return;
   if (alpha == 0.) return;
 
+  if (incx == 1 && incy == 1 && 1L*m*n <= 2048 *GEMM_MULTITHREAD_THRESHOLD) {
+    GER(m, n, 0, alpha, x, incx, y, incy, a, lda, buffer);
+    return;
+  }  
+
   IDEBUG_START;
 
   FUNCTION_PROFILE_START();

interface/spr.c

@@ -167,6 +167,26 @@ void CNAME(enum CBLAS_ORDER order,
 
   FUNCTION_PROFILE_START();
 
+  if (incx == 1 && n <100) {
+    blasint i;
+    if (uplo==0) {
+      for (i = 0; i < n; i++){
+        if (x[i] != ZERO) {
+          AXPYU_K(i + 1, 0, 0, alpha * x[i], x,     1, a, 1, NULL, 0);
+        }
+        a += i + 1;
+      }
+    } else { 
+      for (i = 0; i < n; i++){
+        if (x[i] != ZERO) {
+          AXPYU_K(n - i, 0, 0, alpha * x[i], x + i, 1, a, 1, NULL, 0);
+        }
+        a += n - i;
+      }
+    }
+    return;
+  }
+
   if (incx < 0 ) x -= (n - 1) * incx;
 
   buffer = (FLOAT *)blas_memory_alloc(1);

interface/spr2.c

@@ -168,6 +168,24 @@ void CNAME(enum CBLAS_ORDER order,
 
   if (alpha == ZERO) return;
 
+  if (incx == 1 && incy == 1 && n < 50) {
+    blasint i;
+    if (!uplo) {
+      for (i = 0; i < n; i++){
+        AXPYU_K(i + 1, 0, 0, alpha * x[i], y,     1, a, 1, NULL, 0);
+        AXPYU_K(i + 1, 0, 0, alpha * y[i], x,     1, a, 1, NULL, 0);
+        a += i + 1;
+      }
+    } else {
+      for (i = 0; i < n; i++){
+	AXPYU_K(n - i, 0, 0, alpha * x[i], y + i, 1, a, 1, NULL, 0);
+        AXPYU_K(n - i, 0, 0, alpha * y[i], x + i, 1, a, 1, NULL, 0);
+        a += n - i;
+      }
+    }
+    return;
+  }
+
   IDEBUG_START;
 
   FUNCTION_PROFILE_START();

interface/symv.c

@@ -170,6 +170,10 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha,
 
   if (alpha == ZERO) return;
 
+  if (incx == 1 && incy == 1 && n*n < 2304 *GEMM_MULTITHREAD_THRESHOLD) {
+    (symv[uplo])(n, n, alpha, a, lda, x, incx, y, incy, buffer);
+    return;
+  }
   IDEBUG_START;
 
   FUNCTION_PROFILE_START();

interface/syr2.c

@@ -170,6 +170,25 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha,
 
   IDEBUG_START;
 
+  if (incx == 1 && incy == 1 && n < 100) {
+    blasint i;
+    if (!uplo) {
+      for (i = 0; i < n; i++){
+        AXPYU_K(i + 1, 0, 0, alpha * x[i], y,     1, a, 1, NULL, 0);
+        AXPYU_K(i + 1, 0, 0, alpha * y[i], x,     1, a, 1, NULL, 0);
+        a += lda;
+      }
+    } else {
+      for (i = 0; i < n; i++){
+        AXPYU_K(n - i, 0, 0, alpha * x[i], y + i, 1, a, 1, NULL, 0);
+        AXPYU_K(n - i, 0, 0, alpha * y[i], x + i, 1, a, 1, NULL, 0);
+        a += 1 + lda;
+      }
+    }
+    return;
+  }
+
+	  
   FUNCTION_PROFILE_START();
 
   if (incx < 0 ) x -= (n - 1) * incx;

interface/zsyr.c

@@ -172,6 +172,32 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, int n, FLOAT alpha, FLO
 
   if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
 
+  if (incx == 1 && incy == 1 && n < 50) {
+    blasint i;
+    if (!uplo) {
+      for (i = 0; i < n; i++){
+        if ((x[i * 2 + 0] != ZERO) || (x[i * 2 + 1] != ZERO)) {
+          AXPYU_K(i + 1, 0, 0,
+              alpha_r * x[i * 2 + 0] - alpha_i * x[i * 2 + 1],
+              alpha_i * x[i * 2 + 0] + alpha_r * x[i * 2 + 1],
+              x,         1, a, 1, NULL, 0);
+        }
+        a += lda;
+      }
+    } else {
+      for (i = 0; i < n; i++){
+        if ((x[i * 2 + 0] != ZERO) || (x[i * 2 + 1] != ZERO)) {
+          AXPYU_K(m - i, 0, 0,
+              alpha_r * x[i * 2 + 0] - alpha_i * x[i * 2 + 1],
+              alpha_i * x[i * 2 + 0] + alpha_r * x[i * 2 + 1],
+              x + i * 2, 1, a, 1, NULL, 0);
+        }
+        a += 2 + lda;
+      }
+    }
+    return;
+  }
+
   IDEBUG_START;
 
   FUNCTION_PROFILE_START();