X-Git-Url: http://git.sourceforge.jp/view?p=pf3gnuchains%2Fgcc-fork.git;a=blobdiff_plain;f=libgfortran%2Fgenerated%2Fmatmul_r16.c;h=ec760f2d3d8ce2268c310ea8de51b1c3d51a5d36;hp=f120e7fdc56673a772c27b621062cded1933ec4b;hb=4e8e57b0ce67551ca61b7883e73586ba805f0a61;hpb=dd5b9961bbe31eeb8b7e3721be32b0d456c61f69

diff --git a/libgfortran/generated/matmul_r16.c b/libgfortran/generated/matmul_r16.c
index f120e7fdc56..ec760f2d3d8 100644
--- a/libgfortran/generated/matmul_r16.c
+++ b/libgfortran/generated/matmul_r16.c
@@ -36,6 +36,16 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_REAL_16)
 
+/* Prototype for the BLAS ?gemm subroutine, a pointer to which can be
+   passed to us by the front-end, in which case we'll call it for large
+   matrices.  */
+
+typedef void (*blas_call)(const char *, const char *, const int *, const int *,
+                          const int *, const GFC_REAL_16 *, const GFC_REAL_16 *,
+                          const int *, const GFC_REAL_16 *, const int *,
+                          const GFC_REAL_16 *, GFC_REAL_16 *, const int *,
+                          int, int);
+
 /* The order of loops is different in the case of plain matrix
    multiplication C=MATMUL(A,B), and in the frequent special case where
    the argument A is the temporary result of a TRANSPOSE intrinsic:
@@ -56,18 +66,24 @@ Boston, MA 02110-1301, USA.  */
        DO I=1,M
          S = 0
          DO K=1,COUNT
-           S = S+A(I,K)+B(K,J)
+           S = S+A(I,K)*B(K,J)
          C(I,J) = S
    ENDIF
 */
 
+/* If try_blas is set to a nonzero value, then the matmul function will
+   see if there is a way to perform the matrix multiplication by a call
+   to the BLAS gemm function.  */
+
 extern void matmul_r16 (gfc_array_r16 * const restrict retarray, 
-	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b);
+	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm);
 export_proto(matmul_r16);
 
 void
 matmul_r16 (gfc_array_r16 * const restrict retarray, 
-	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b)
+	gfc_array_r16 * const restrict a, gfc_array_r16 * const restrict b, int try_blas,
+	int blas_limit, blas_call gemm)
 {
   const GFC_REAL_16 * restrict abase;
   const GFC_REAL_16 * restrict bbase;
@@ -177,6 +193,31 @@ matmul_r16 (gfc_array_r16 * const restrict retarray,
   bbase = b->data;
   dest = retarray->data;
 
+
+  /* Now that everything is set up, we're performing the multiplication
+     itself.  */
+
+#define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+
+  if (try_blas && rxstride == 1 && (axstride == 1 || aystride == 1)
+      && (bxstride == 1 || bystride == 1)
+      && (((float) xcount) * ((float) ycount) * ((float) count)
+          > POW3(blas_limit)))
+  {
+    const int m = xcount, n = ycount, k = count, ldc = rystride;
+    const GFC_REAL_16 one = 1, zero = 0;
+    const int lda = (axstride == 1) ? aystride : axstride,
+              ldb = (bxstride == 1) ? bystride : bxstride;
+
+    if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
+      {
+        assert (gemm != NULL);
+        gemm (axstride == 1 ? "N" : "T", bxstride == 1 ? "N" : "T", &m, &n, &k,
+              &one, abase, &lda, bbase, &ldb, &zero, dest, &ldc, 1, 1);
+        return;
+      }
+  }
+
   if (rxstride == 1 && axstride == 1 && bxstride == 1)
     {
       const GFC_REAL_16 * restrict bbase_y;
@@ -258,6 +299,20 @@ matmul_r16 (gfc_array_r16 * const restrict retarray,
 	    /* dest[x,y] += a[x,n] * b[n,y] */
 	    dest[x*rxstride + y*rystride] += abase[x*axstride + n*aystride] * bbase[n*bxstride + y*bystride];
     }
+  else if (GFC_DESCRIPTOR_RANK (a) == 1)
+    {
+      const GFC_REAL_16 *restrict bbase_y;
+      GFC_REAL_16 s;
+
+      for (y = 0; y < ycount; y++)
+	{
+	  bbase_y = &bbase[y*bystride];
+	  s = (GFC_REAL_16) 0;
+	  for (n = 0; n < count; n++)
+	    s += abase[n*axstride] * bbase_y[n*bxstride];
+	  dest[y*rxstride] = s;
+	}
+    }
   else
     {
       const GFC_REAL_16 *restrict abase_x;