[multiple changes]

2016-11-15  Jerry DeLisle  <jvdelisle@gcc.gnu.org>
	    Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR libgfortran/51119
	* Makefile.am: Add new optimization flags matmul.
	* Makefile.in: Regenerate.
	* m4/matmul.m4: For the case of all strides = 1, implement a
	fast blocked matrix multiply. Fix some whitespace.
	* generated/matmul_c10.c: Regenerate.
	* generated/matmul_c16.c: Regenerate.
	* generated/matmul_c4.c: Regenerate.
	* generated/matmul_c8.c: Regenerate.
	* generated/matmul_i1.c: Regenerate.
	* generated/matmul_i16.c: Regenerate.
	* generated/matmul_i2.c: Regenerate.
	* generated/matmul_i4.c: Regenerate.
	* generated/matmul_i8.c: Regenerate.
	* generated/matmul_r10.c: Regenerate.
	* generated/matmul_r16.c: Regenerate.
	* generated/matmul_r4.c: Regenerate.
	* generated/matmul_r8.c: Regenerate.

2016-11-15  Thomas Koenig  <tkoenig@gcc.gnu.org>

	PR libgfortran/51119
	* gfortran.dg/matmul_12.f90: New test case.

From-SVN: r242462

1 files changed, 315 insertions, 81 deletions

diff --git a/libgfortran/generated/matmul_r4.c b/libgfortran/generated/matmul_r4.c
index ff3b93ff4d4..cf3ffa35232 100644
--- a/libgfortran/generated/matmul_r4.c
+++ b/libgfortran/generated/matmul_r4.c
@@ -32,7 +32,7 @@ see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
 #if defined (HAVE_GFC_REAL_4)
 
 /* 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
+   passed to us by the front-end, in which case we call it for large
    matrices.  */
 
 typedef void (*blas_call)(const char *, const char *, const int *, const int *,
@@ -99,7 +99,7 @@ matmul_r4 (gfc_array_r4 * const restrict retarray,
 
    o One-dimensional argument B is implicitly treated as a column matrix
      dimensioned [count, 1], so ycount=1.
-  */
+*/
 
   if (retarray->base_addr == NULL)
     {
@@ -127,47 +127,47 @@ matmul_r4 (gfc_array_r4 * const restrict retarray,
 	= xmallocarray (size0 ((array_t *) retarray), sizeof (GFC_REAL_4));
       retarray->offset = 0;
     }
-    else if (unlikely (compile_options.bounds_check))
-      {
-	index_type ret_extent, arg_extent;
-
-	if (GFC_DESCRIPTOR_RANK (a) == 1)
-	  {
-	    arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
-	    ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
-	    if (arg_extent != ret_extent)
-	      runtime_error ("Incorrect extent in return array in"
-			     " MATMUL intrinsic: is %ld, should be %ld",
-			     (long int) ret_extent, (long int) arg_extent);
-	  }
-	else if (GFC_DESCRIPTOR_RANK (b) == 1)
-	  {
-	    arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
-	    ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
-	    if (arg_extent != ret_extent)
-	      runtime_error ("Incorrect extent in return array in"
-			     " MATMUL intrinsic: is %ld, should be %ld",
-			     (long int) ret_extent, (long int) arg_extent);	    
-	  }
-	else
-	  {
-	    arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
-	    ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
-	    if (arg_extent != ret_extent)
-	      runtime_error ("Incorrect extent in return array in"
-			     " MATMUL intrinsic for dimension 1:"
-			     " is %ld, should be %ld",
-			     (long int) ret_extent, (long int) arg_extent);
-
-	    arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
-	    ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
-	    if (arg_extent != ret_extent)
-	      runtime_error ("Incorrect extent in return array in"
-			     " MATMUL intrinsic for dimension 2:"
-			     " is %ld, should be %ld",
-			     (long int) ret_extent, (long int) arg_extent);
-	  }
-      }
+  else if (unlikely (compile_options.bounds_check))
+    {
+      index_type ret_extent, arg_extent;
+
+      if (GFC_DESCRIPTOR_RANK (a) == 1)
+	{
+	  arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
+	  ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
+	  if (arg_extent != ret_extent)
+	    runtime_error ("Incorrect extent in return array in"
+			   " MATMUL intrinsic: is %ld, should be %ld",
+			   (long int) ret_extent, (long int) arg_extent);
+	}
+      else if (GFC_DESCRIPTOR_RANK (b) == 1)
+	{
+	  arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
+	  ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
+	  if (arg_extent != ret_extent)
+	    runtime_error ("Incorrect extent in return array in"
+			   " MATMUL intrinsic: is %ld, should be %ld",
+			   (long int) ret_extent, (long int) arg_extent);
+	}
+      else
+	{
+	  arg_extent = GFC_DESCRIPTOR_EXTENT(a,0);
+	  ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,0);
+	  if (arg_extent != ret_extent)
+	    runtime_error ("Incorrect extent in return array in"
+			   " MATMUL intrinsic for dimension 1:"
+			   " is %ld, should be %ld",
+			   (long int) ret_extent, (long int) arg_extent);
+
+	  arg_extent = GFC_DESCRIPTOR_EXTENT(b,1);
+	  ret_extent = GFC_DESCRIPTOR_EXTENT(retarray,1);
+	  if (arg_extent != ret_extent)
+	    runtime_error ("Incorrect extent in return array in"
+			   " MATMUL intrinsic for dimension 2:"
+			   " is %ld, should be %ld",
+			   (long int) ret_extent, (long int) arg_extent);
+	}
+    }
 
 
   if (GFC_DESCRIPTOR_RANK (retarray) == 1)
@@ -230,61 +230,294 @@ matmul_r4 (gfc_array_r4 * const restrict retarray,
   bbase = b->base_addr;
   dest = retarray->base_addr;
 
-
-  /* Now that everything is set up, we're performing the multiplication
+  /* Now that everything is set up, we perform the multiplication
      itself.  */
 
 #define POW3(x) (((float) (x)) * ((float) (x)) * ((float) (x)))
+#define min(a,b) ((a) <= (b) ? (a) : (b))
+#define max(a,b) ((a) >= (b) ? (a) : (b))
 
   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_4 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_4 * restrict bbase_y;
-      GFC_REAL_4 * restrict dest_y;
-      const GFC_REAL_4 * restrict abase_n;
-      GFC_REAL_4 bbase_yn;
+      const int m = xcount, n = ycount, k = count, ldc = rystride;
+      const GFC_REAL_4 one = 1, zero = 0;
+      const int lda = (axstride == 1) ? aystride : axstride,
+		ldb = (bxstride == 1) ? bystride : bxstride;
 
-      if (rystride == xcount)
-	memset (dest, 0, (sizeof (GFC_REAL_4) * xcount * ycount));
-      else
+      if (lda > 0 && ldb > 0 && ldc > 0 && m > 1 && n > 1 && k > 1)
 	{
-	  for (y = 0; y < ycount; y++)
-	    for (x = 0; x < xcount; x++)
-	      dest[x + y*rystride] = (GFC_REAL_4)0;
+	  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;
 	}
+    }
 
-      for (y = 0; y < ycount; y++)
+  if (rxstride == 1 && axstride == 1 && bxstride == 1)
+    {
+      /* This block of code implements a tuned matmul, derived from
+         Superscalar GEMM-based level 3 BLAS,  Beta version 0.1
+
+               Bo Kagstrom and Per Ling
+               Department of Computing Science
+               Umea University
+               S-901 87 Umea, Sweden
+
+	 from netlib.org, translated to C, and modified for matmul.m4.  */
+
+      const GFC_REAL_4 *a, *b;
+      GFC_REAL_4 *c;
+      const index_type m = xcount, n = ycount, k = count;
+
+      /* System generated locals */
+      index_type a_dim1, a_offset, b_dim1, b_offset, c_dim1, c_offset,
+		 i1, i2, i3, i4, i5, i6;
+
+      /* Local variables */
+      GFC_REAL_4 t1[65536], /* was [256][256] */
+		 f11, f12, f21, f22, f31, f32, f41, f42,
+		 f13, f14, f23, f24, f33, f34, f43, f44;
+      index_type i, j, l, ii, jj, ll;
+      index_type isec, jsec, lsec, uisec, ujsec, ulsec;
+
+      a = abase;
+      b = bbase;
+      c = retarray->base_addr;
+
+      /* Parameter adjustments */
+      c_dim1 = rystride;
+      c_offset = 1 + c_dim1;
+      c -= c_offset;
+      a_dim1 = aystride;
+      a_offset = 1 + a_dim1;
+      a -= a_offset;
+      b_dim1 = bystride;
+      b_offset = 1 + b_dim1;
+      b -= b_offset;
+
+      /* Early exit if possible */
+      if (m == 0 || n == 0 || k == 0)
+	return;
+
+      /* Empty c first.  */
+      for (j=1; j<=n; j++)
+	for (i=1; i<=m; i++)
+	  c[i + j * c_dim1] = (GFC_REAL_4)0;
+
+      /* Start turning the crank. */
+      i1 = n;
+      for (jj = 1; jj <= i1; jj += 512)
 	{
-	  bbase_y = bbase + y*bystride;
-	  dest_y = dest + y*rystride;
-	  for (n = 0; n < count; n++)
+	  /* Computing MIN */
+	  i2 = 512;
+	  i3 = n - jj + 1;
+	  jsec = min(i2,i3);
+	  ujsec = jsec - jsec % 4;
+	  i2 = k;
+	  for (ll = 1; ll <= i2; ll += 256)
 	    {
-	      abase_n = abase + n*aystride;
-	      bbase_yn = bbase_y[n];
-	      for (x = 0; x < xcount; x++)
+	      /* Computing MIN */
+	      i3 = 256;
+	      i4 = k - ll + 1;
+	      lsec = min(i3,i4);
+	      ulsec = lsec - lsec % 2;
+
+	      i3 = m;
+	      for (ii = 1; ii <= i3; ii += 256)
 		{
-		  dest_y[x] += abase_n[x] * bbase_yn;
+		  /* Computing MIN */
+		  i4 = 256;
+		  i5 = m - ii + 1;
+		  isec = min(i4,i5);
+		  uisec = isec - isec % 2;
+		  i4 = ll + ulsec - 1;
+		  for (l = ll; l <= i4; l += 2)
+		    {
+		      i5 = ii + uisec - 1;
+		      for (i = ii; i <= i5; i += 2)
+			{
+			  t1[l - ll + 1 + ((i - ii + 1) << 8) - 257] =
+					a[i + l * a_dim1];
+			  t1[l - ll + 2 + ((i - ii + 1) << 8) - 257] =
+					a[i + (l + 1) * a_dim1];
+			  t1[l - ll + 1 + ((i - ii + 2) << 8) - 257] =
+					a[i + 1 + l * a_dim1];
+			  t1[l - ll + 2 + ((i - ii + 2) << 8) - 257] =
+					a[i + 1 + (l + 1) * a_dim1];
+			}
+		      if (uisec < isec)
+			{
+			  t1[l - ll + 1 + (isec << 8) - 257] =
+				    a[ii + isec - 1 + l * a_dim1];
+			  t1[l - ll + 2 + (isec << 8) - 257] =
+				    a[ii + isec - 1 + (l + 1) * a_dim1];
+			}
+		    }
+		  if (ulsec < lsec)
+		    {
+		      i4 = ii + isec - 1;
+		      for (i = ii; i<= i4; ++i)
+			{
+			  t1[lsec + ((i - ii + 1) << 8) - 257] =
+				    a[i + (ll + lsec - 1) * a_dim1];
+			}
+		    }
+
+		  uisec = isec - isec % 4;
+		  i4 = jj + ujsec - 1;
+		  for (j = jj; j <= i4; j += 4)
+		    {
+		      i5 = ii + uisec - 1;
+		      for (i = ii; i <= i5; i += 4)
+			{
+			  f11 = c[i + j * c_dim1];
+			  f21 = c[i + 1 + j * c_dim1];
+			  f12 = c[i + (j + 1) * c_dim1];
+			  f22 = c[i + 1 + (j + 1) * c_dim1];
+			  f13 = c[i + (j + 2) * c_dim1];
+			  f23 = c[i + 1 + (j + 2) * c_dim1];
+			  f14 = c[i + (j + 3) * c_dim1];
+			  f24 = c[i + 1 + (j + 3) * c_dim1];
+			  f31 = c[i + 2 + j * c_dim1];
+			  f41 = c[i + 3 + j * c_dim1];
+			  f32 = c[i + 2 + (j + 1) * c_dim1];
+			  f42 = c[i + 3 + (j + 1) * c_dim1];
+			  f33 = c[i + 2 + (j + 2) * c_dim1];
+			  f43 = c[i + 3 + (j + 2) * c_dim1];
+			  f34 = c[i + 2 + (j + 3) * c_dim1];
+			  f44 = c[i + 3 + (j + 3) * c_dim1];
+			  i6 = ll + lsec - 1;
+			  for (l = ll; l <= i6; ++l)
+			    {
+			      f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
+				      * b[l + j * b_dim1];
+			      f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
+				      * b[l + j * b_dim1];
+			      f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
+				      * b[l + (j + 1) * b_dim1];
+			      f22 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
+				      * b[l + (j + 1) * b_dim1];
+			      f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
+				      * b[l + (j + 2) * b_dim1];
+			      f23 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
+				      * b[l + (j + 2) * b_dim1];
+			      f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) - 257]
+				      * b[l + (j + 3) * b_dim1];
+			      f24 += t1[l - ll + 1 + ((i - ii + 2) << 8) - 257]
+				      * b[l + (j + 3) * b_dim1];
+			      f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
+				      * b[l + j * b_dim1];
+			      f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
+				      * b[l + j * b_dim1];
+			      f32 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
+				      * b[l + (j + 1) * b_dim1];
+			      f42 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
+				      * b[l + (j + 1) * b_dim1];
+			      f33 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
+				      * b[l + (j + 2) * b_dim1];
+			      f43 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
+				      * b[l + (j + 2) * b_dim1];
+			      f34 += t1[l - ll + 1 + ((i - ii + 3) << 8) - 257]
+				      * b[l + (j + 3) * b_dim1];
+			      f44 += t1[l - ll + 1 + ((i - ii + 4) << 8) - 257]
+				      * b[l + (j + 3) * b_dim1];
+			    }
+			  c[i + j * c_dim1] = f11;
+			  c[i + 1 + j * c_dim1] = f21;
+			  c[i + (j + 1) * c_dim1] = f12;
+			  c[i + 1 + (j + 1) * c_dim1] = f22;
+			  c[i + (j + 2) * c_dim1] = f13;
+			  c[i + 1 + (j + 2) * c_dim1] = f23;
+			  c[i + (j + 3) * c_dim1] = f14;
+			  c[i + 1 + (j + 3) * c_dim1] = f24;
+			  c[i + 2 + j * c_dim1] = f31;
+			  c[i + 3 + j * c_dim1] = f41;
+			  c[i + 2 + (j + 1) * c_dim1] = f32;
+			  c[i + 3 + (j + 1) * c_dim1] = f42;
+			  c[i + 2 + (j + 2) * c_dim1] = f33;
+			  c[i + 3 + (j + 2) * c_dim1] = f43;
+			  c[i + 2 + (j + 3) * c_dim1] = f34;
+			  c[i + 3 + (j + 3) * c_dim1] = f44;
+			}
+		      if (uisec < isec)
+			{
+			  i5 = ii + isec - 1;
+			  for (i = ii + uisec; i <= i5; ++i)
+			    {
+			      f11 = c[i + j * c_dim1];
+			      f12 = c[i + (j + 1) * c_dim1];
+			      f13 = c[i + (j + 2) * c_dim1];
+			      f14 = c[i + (j + 3) * c_dim1];
+			      i6 = ll + lsec - 1;
+			      for (l = ll; l <= i6; ++l)
+				{
+				  f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
+					  257] * b[l + j * b_dim1];
+				  f12 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
+					  257] * b[l + (j + 1) * b_dim1];
+				  f13 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
+					  257] * b[l + (j + 2) * b_dim1];
+				  f14 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
+					  257] * b[l + (j + 3) * b_dim1];
+				}
+			      c[i + j * c_dim1] = f11;
+			      c[i + (j + 1) * c_dim1] = f12;
+			      c[i + (j + 2) * c_dim1] = f13;
+			      c[i + (j + 3) * c_dim1] = f14;
+			    }
+			}
+		    }
+		  if (ujsec < jsec)
+		    {
+		      i4 = jj + jsec - 1;
+		      for (j = jj + ujsec; j <= i4; ++j)
+			{
+			  i5 = ii + uisec - 1;
+			  for (i = ii; i <= i5; i += 4)
+			    {
+			      f11 = c[i + j * c_dim1];
+			      f21 = c[i + 1 + j * c_dim1];
+			      f31 = c[i + 2 + j * c_dim1];
+			      f41 = c[i + 3 + j * c_dim1];
+			      i6 = ll + lsec - 1;
+			      for (l = ll; l <= i6; ++l)
+				{
+				  f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
+					  257] * b[l + j * b_dim1];
+				  f21 += t1[l - ll + 1 + ((i - ii + 2) << 8) -
+					  257] * b[l + j * b_dim1];
+				  f31 += t1[l - ll + 1 + ((i - ii + 3) << 8) -
+					  257] * b[l + j * b_dim1];
+				  f41 += t1[l - ll + 1 + ((i - ii + 4) << 8) -
+					  257] * b[l + j * b_dim1];
+				}
+			      c[i + j * c_dim1] = f11;
+			      c[i + 1 + j * c_dim1] = f21;
+			      c[i + 2 + j * c_dim1] = f31;
+			      c[i + 3 + j * c_dim1] = f41;
+			    }
+			  i5 = ii + isec - 1;
+			  for (i = ii + uisec; i <= i5; ++i)
+			    {
+			      f11 = c[i + j * c_dim1];
+			      i6 = ll + lsec - 1;
+			      for (l = ll; l <= i6; ++l)
+				{
+				  f11 += t1[l - ll + 1 + ((i - ii + 1) << 8) -
+					  257] * b[l + j * b_dim1];
+				}
+			      c[i + j * c_dim1] = f11;
+			    }
+			}
+		    }
 		}
 	    }
 	}
+      return;
     }
   else if (rxstride == 1 && aystride == 1 && bxstride == 1)
     {
@@ -334,7 +567,9 @@ matmul_r4 (gfc_array_r4 * const restrict retarray,
 	for (n = 0; n < count; n++)
 	  for (x = 0; x < xcount; x++)
 	    /* dest[x,y] += a[x,n] * b[n,y] */
-	    dest[x*rxstride + y*rystride] += abase[x*axstride + n*aystride] * bbase[n*bxstride + y*bystride];
+	    dest[x*rxstride + y*rystride] +=
+					abase[x*axstride + n*aystride] *
+					bbase[n*bxstride + y*bystride];
     }
   else if (GFC_DESCRIPTOR_RANK (a) == 1)
     {
@@ -372,5 +607,4 @@ matmul_r4 (gfc_array_r4 * const restrict retarray,
 	}
     }
 }
-
 #endif