gcc/fortran/

[pf3gnuchains/gcc-fork.git] / libgfortran / generated / matmul_i16.c

diff --git a/libgfortran/generated/matmul_i16.c b/libgfortran/generated/matmul_i16.c
index a5a40b4..eacc47f 100644 (file)
--- a/libgfortran/generated/matmul_i16.c
+++ b/libgfortran/generated/matmul_i16.c
@@ -36,16 +36,29 @@ Boston, MA 02110-1301, USA.  */
 
 #if defined (HAVE_GFC_INTEGER_16)
 
-/* 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_i16 (gfc_array_i16 * const restrict retarray, 
@@ -204,7 +217,28 @@ matmul_i16 (gfc_array_i16 * const restrict retarray,
            }
        }
     }
-  else
+  else if (rxstride == 1 && aystride == 1 && bxstride == 1)
+    {
+      const GFC_INTEGER_16 *restrict abase_x;
+      const GFC_INTEGER_16 *restrict bbase_y;
+      GFC_INTEGER_16 *restrict dest_y;
+      GFC_INTEGER_16 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_INTEGER_16) 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 +250,27 @@ matmul_i16 (gfc_array_i16 * 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_INTEGER_16 *restrict abase_x;
+      const GFC_INTEGER_16 *restrict bbase_y;
+      GFC_INTEGER_16 *restrict dest_y;
+      GFC_INTEGER_16 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_INTEGER_16) 0;
+             for (n = 0; n < count; n++)
+               s += abase_x[n*aystride] * bbase_y[n*bxstride];
+             dest_y[x*rxstride] = s;
+           }
+       }
+    }
 }
 
 #endif