2006-06-06 Janne Blomqvist <jb@gcc.gnu.org>

[pf3gnuchains/gcc-fork.git] / libgfortran / generated / matmul_r10.c

diff --git a/libgfortran/generated/matmul_r10.c b/libgfortran/generated/matmul_r10.c
index b0ebbed..e90ac57 100644 (file)
--- a/libgfortran/generated/matmul_r10.c
+++ b/libgfortran/generated/matmul_r10.c
@@ -1,5 +1,5 @@
 /* Implementation of the MATMUL intrinsic
-   Copyright 2002, 2005 Free Software Foundation, Inc.
+   Copyright 2002, 2005, 2006 Free Software Foundation, Inc.
    Contributed by Paul Brook <paul@nowt.org>
 
 This file is part of the GNU Fortran 95 runtime library (libgfortran).
@@ -36,16 +36,29 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_REAL_10)
 
-/* This is a C version of the following fortran pseudo-code. The key
-   point is the loop order -- we access all arrays column-first, which
-   improves the performance enough to boost galgel spec score by 50%.
+/* 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:
+   C=MATMUL(TRANSPOSE(A),B).  Transposed temporaries are detected by
+   looking at their strides.
+
+   The equivalent Fortran pseudo-code is:
 
    DIMENSION A(M,COUNT), B(COUNT,N), C(M,N)
-   C = 0
-   DO J=1,N
-     DO K=1,COUNT
+   IF (.NOT.IS_TRANSPOSED(A)) THEN
+     C = 0
+     DO J=1,N
+       DO K=1,COUNT
+         DO I=1,M
+           C(I,J) = C(I,J)+A(I,K)*B(K,J)
+   ELSE
+     DO J=1,N
        DO I=1,M
-         C(I,J) = C(I,J)+A(I,K)*B(K,J)
+         S = 0
+         DO K=1,COUNT
+           S = S+A(I,K)+B(K,J)
+         C(I,J) = S
+   ENDIF
 */
 
 extern void matmul_r10 (gfc_array_r10 * const restrict retarray, 
@@ -107,15 +120,6 @@ matmul_r10 (gfc_array_r10 * const restrict retarray,
       retarray->offset = 0;
     }
 
-  if (retarray->dim[0].stride == 0)
-    retarray->dim[0].stride = 1;
-
-  /* This prevents constifying the input arguments.  */
-  if (a->dim[0].stride == 0)
-    a->dim[0].stride = 1;
-  if (b->dim[0].stride == 0)
-    b->dim[0].stride = 1;
-
 
   if (GFC_DESCRIPTOR_RANK (retarray) == 1)
     {
@@ -180,8 +184,8 @@ matmul_r10 (gfc_array_r10 * const restrict retarray,
       const GFC_REAL_10 * restrict abase_n;
       GFC_REAL_10 bbase_yn;
 
-      if (rystride == ycount)
-       memset (dest, 0, (sizeof (GFC_REAL_10) * size0((array_t *) retarray)));
+      if (rystride == xcount)
+       memset (dest, 0, (sizeof (GFC_REAL_10) * xcount * ycount));
       else
        {
          for (y = 0; y < ycount; y++)
@@ -204,7 +208,28 @@ matmul_r10 (gfc_array_r10 * const restrict retarray,
            }
        }
     }
-  else
+  else if (rxstride == 1 && aystride == 1 && bxstride == 1)
+    {
+      const GFC_REAL_10 *restrict abase_x;
+      const GFC_REAL_10 *restrict bbase_y;
+      GFC_REAL_10 *restrict dest_y;
+      GFC_REAL_10 s;
+
+      for (y = 0; y < ycount; y++)
+       {
+         bbase_y = &bbase[y*bystride];
+         dest_y = &dest[y*rystride];
+         for (x = 0; x < xcount; x++)
+           {
+             abase_x = &abase[x*axstride];
+             s = (GFC_REAL_10) 0;
+             for (n = 0; n < count; n++)
+               s += abase_x[n] * bbase_y[n];
+             dest_y[x] = s;
+           }
+       }
+    }
+  else if (axstride < aystride)
     {
       for (y = 0; y < ycount; y++)
        for (x = 0; x < xcount; x++)
@@ -216,6 +241,27 @@ matmul_r10 (gfc_array_r10 * 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
+    {
+      const GFC_REAL_10 *restrict abase_x;
+      const GFC_REAL_10 *restrict bbase_y;
+      GFC_REAL_10 *restrict dest_y;
+      GFC_REAL_10 s;
+
+      for (y = 0; y < ycount; y++)
+       {
+         bbase_y = &bbase[y*bystride];
+         dest_y = &dest[y*rystride];
+         for (x = 0; x < xcount; x++)
+           {
+             abase_x = &abase[x*axstride];
+             s = (GFC_REAL_10) 0;
+             for (n = 0; n < count; n++)
+               s += abase_x[n*aystride] * bbase_y[n*bxstride];
+             dest_y[x*rxstride] = s;
+           }
+       }
+    }
 }
 
 #endif