/* 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).
#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:
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;
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)
{
xcount = a->dim[0].ubound + 1 - a->dim[0].lbound;
}
- assert(count == b->dim[0].ubound + 1 - b->dim[0].lbound);
+ if (count != b->dim[0].ubound + 1 - b->dim[0].lbound)
+ runtime_error ("dimension of array B incorrect in MATMUL intrinsic");
if (GFC_DESCRIPTOR_RANK (b) == 1)
{
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;
}
else if (rxstride == 1 && aystride == 1 && bxstride == 1)
{
- const GFC_REAL_16 *restrict abase_x;
- const GFC_REAL_16 *restrict bbase_y;
- GFC_REAL_16 *restrict dest_y;
- GFC_REAL_16 s;
+ if (GFC_DESCRIPTOR_RANK (a) != 1)
+ {
+ const GFC_REAL_16 *restrict abase_x;
+ const GFC_REAL_16 *restrict bbase_y;
+ GFC_REAL_16 *restrict dest_y;
+ GFC_REAL_16 s;
- for (y = 0; y < ycount; y++)
+ 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_16) 0;
+ for (n = 0; n < count; n++)
+ s += abase_x[n] * bbase_y[n];
+ dest_y[x] = s;
+ }
+ }
+ }
+ else
{
- bbase_y = &bbase[y*bystride];
- dest_y = &dest[y*rystride];
- for (x = 0; x < xcount; x++)
+ const GFC_REAL_16 *restrict bbase_y;
+ GFC_REAL_16 s;
+
+ for (y = 0; y < ycount; y++)
{
- abase_x = &abase[x*axstride];
+ bbase_y = &bbase[y*bystride];
s = (GFC_REAL_16) 0;
for (n = 0; n < count; n++)
- s += abase_x[n] * bbase_y[n];
- dest_y[x] = s;
+ s += abase[n*axstride] * bbase_y[n];
+ dest[y*rystride] = s;
}
}
}
/* 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;