Index: liboil/simdpack/multsum.c
===================================================================
RCS file: /cvs/liboil/liboil/liboil/simdpack/multsum.c,v
retrieving revision 1.12
diff -u -p -r1.12 multsum.c
--- liboil/simdpack/multsum.c	13 Dec 2005 20:13:44 -0000	1.12
+++ liboil/simdpack/multsum.c	17 May 2006 18:00:54 -0000
@@ -55,3 +55,47 @@ static void multsum_f32_unroll2 (float *
 }
 OIL_DEFINE_IMPL (multsum_f32_unroll2, multsum_f32);
 
+
+static void multsum_f64_unroll8 (double *dest, double *src1, int sstr1,
+    double *src2, int sstr2, int n)
+{
+  int i = 0;
+  double sum = 0;
+
+  while(i<n-7) {
+    sum += (OIL_GET(src1,0, double) * OIL_GET(src2,0, double)) +
+           (OIL_GET(src1,sstr1, double) * OIL_GET(src2,sstr2, double)) +
+           (OIL_GET(src1,2*sstr1, double) * OIL_GET(src2,2*sstr2, double)) +
+           (OIL_GET(src1,3*sstr1, double) * OIL_GET(src2,3*sstr2, double)) +
+           (OIL_GET(src1,4*sstr1, double) * OIL_GET(src2,4*sstr2, double)) +
+           (OIL_GET(src1,5*sstr1, double) * OIL_GET(src2,5*sstr2, double)) +
+           (OIL_GET(src1,6*sstr1, double) * OIL_GET(src2,6*sstr2, double)) +
+           (OIL_GET(src1,7*sstr1, double) * OIL_GET(src2,7*sstr2, double));
+    OIL_INCREMENT (src1, sstr1*8);
+    OIL_INCREMENT (src2, sstr2*8);
+    i+=8;
+  }
+  while(i<n-3) {
+    sum += (OIL_GET(src1,0, double) * OIL_GET(src2,0, double)) +
+           (OIL_GET(src1,sstr1, double) * OIL_GET(src2,sstr2, double)) +
+           (OIL_GET(src1,2*sstr1, double) * OIL_GET(src2,2*sstr2, double)) +
+           (OIL_GET(src1,3*sstr1, double) * OIL_GET(src2,3*sstr2, double));
+    OIL_INCREMENT (src1, sstr1*4);
+    OIL_INCREMENT (src2, sstr2*4);
+    i+=4;
+  }
+  while(i<n-1) {
+    sum += (OIL_GET(src1,0, double) * OIL_GET(src2,0, double)) +
+           (OIL_GET(src1,sstr1, double) * OIL_GET(src2,sstr2, double));
+    OIL_INCREMENT (src1, sstr1*2);
+    OIL_INCREMENT (src2, sstr2*2);
+    i+=2;
+  }
+  if (i<n) {
+    sum += OIL_GET(src1,0, double) * OIL_GET(src2,0, double);
+  }
+
+  *dest = sum;
+}
+OIL_DEFINE_IMPL (multsum_f64_unroll8, multsum_f64);
+
Index: liboil/sse/Makefile.am
===================================================================
RCS file: /cvs/liboil/liboil/liboil/sse/Makefile.am,v
retrieving revision 1.15
diff -u -p -r1.15 Makefile.am
--- liboil/sse/Makefile.am	23 Dec 2005 22:40:34 -0000	1.15
+++ liboil/sse/Makefile.am	17 May 2006 18:00:54 -0000
@@ -9,6 +9,7 @@ libsse_la_SOURCES = \
 	copy_sse.c \
 	math_sse.c \
 	math_sse_unroll2.c \
+	multsum_sse.c \
 	sad8x8_sse.c \
 	splat_sse.c
 
Index: liboil/sse/multsum_sse.c
===================================================================
RCS file: liboil/sse/multsum_sse.c
diff -N liboil/sse/multsum_sse.c
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ liboil/sse/multsum_sse.c	17 May 2006 18:00:54 -0000
@@ -0,0 +1,65 @@
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+#include <liboil/liboilclasses.h>
+#include <liboil/liboilfunction.h>
+#include <emmintrin.h>
+
+#define MULTSUM_SSE2_NSTRIDED(i) { \
+  t1 = _mm_load_pd(&OIL_GET(src1, i, double)); \
+  t2 = _mm_load_pd(&OIL_GET(src2, i, double)); \
+  t1 = _mm_mul_pd(t1,t2); \
+  sum.reg = _mm_add_pd(sum.reg,t1); \
+}
+#define MULTSUM_SSE2_NSTRIDEDP(i) { \
+  t1 = _mm_load_pd(&OIL_GET(src1, i*sstr1, double)); \
+  t2 = _mm_loadl_pd(t2, &OIL_GET(src2, i*sstr2, double)); \
+  t2 = _mm_loadh_pd(t2, &OIL_GET(src2, (i+1)*sstr2, double)); \
+  t1 = _mm_mul_pd(t1,t2); \
+  sum.reg = _mm_add_pd(sum.reg,t1); \
+}
+#define MULTSUM_SSE2_STRIDED(i) { \
+  t1 = _mm_loadl_pd(t1, &OIL_GET(src1, i*sstr1, double)); \
+  t1 = _mm_loadh_pd(t1, &OIL_GET(src1, (i+1)*sstr1, double)); \
+  t2 = _mm_loadl_pd(t2, &OIL_GET(src2, i*sstr2, double)); \
+  t2 = _mm_loadh_pd(t2, &OIL_GET(src2, (i+1)*sstr2, double)); \
+  t1 = _mm_mul_pd(t1,t2); \
+  sum.reg = _mm_add_pd(sum.reg,t1); \
+}
+
+
+static void
+multsum_f64_sse2_unroll4(double *dest,
+     const double *src1, int sstr1,
+     const double *src2, int sstr2,
+     int n)
+{
+  __m128d t1, t2;
+  union {
+    __m128d reg;
+    double vals[2];
+  } sum;
+  int i = 0;
+
+  sum.reg = _mm_setzero_pd();
+  while (i < n-3) {
+    MULTSUM_SSE2_STRIDED(0);
+    MULTSUM_SSE2_STRIDED(2);
+
+    OIL_INCREMENT(src1, 4*sstr1);
+    OIL_INCREMENT(src2, 4*sstr2);
+    i += 4;
+  }
+  while (i < n-1) {
+    MULTSUM_SSE2_STRIDED(0);
+
+    OIL_INCREMENT(src1, 2*sstr1);
+    OIL_INCREMENT(src2, 2*sstr2);
+    i+=2;
+  }
+  *dest = sum.vals[0] + sum.vals[1];
+  if (i < n) {
+    *dest += (OIL_GET(src1,0,double)*OIL_GET(src2,0,double));
+  }
+}
+OIL_DEFINE_IMPL_FULL (multsum_f64_sse2_unroll4, multsum_f64, OIL_IMPL_FLAG_SSE2);