some unrolling

4 files changed, 47 insertions, 3 deletions

diff --git a/test/monniaux/Makefile b/test/monniaux/Makefile
index be534653..aa559699 100644
--- a/test/monniaux/Makefile
+++ b/test/monniaux/Makefile
@@ -1,6 +1,6 @@
 CFLAGS=-Wall -O3
 K1C_CC=k1-mbr-gcc
-K1C_CFLAGS=-Wall -O3 -std=c99
+K1C_CFLAGS=-Wall -O2 -std=c99
 K1C_CCOMP=../../ccomp
 K1C_CCOMPFLAGS=-Wall -O3 -D__thread= -D__int128=int
 
diff --git a/test/monniaux/int_mat.c b/test/monniaux/int_mat.c
index cc8c59e5..d3e14e26 100644
--- a/test/monniaux/int_mat.c
+++ b/test/monniaux/int_mat.c
@@ -80,6 +80,35 @@ void modint_mat_mul4(unsigned m, unsigned n, unsigned p,
   }
 }
 
+void modint_mat_mul5(unsigned m, unsigned n, unsigned p,
+		     modint * c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b) {
+  const modint *pa_i = a;
+  modint * pc_i = c;
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned k=0; k<p; k++) {
+      const modint *pb_j_k = b+k, *pa_i_j = pa_i;
+      modint total = 0;
+      for(unsigned j2=0, n2=n/2; j2<n2; j2++) {
+	modint p0 = *pa_i_j * *pb_j_k;
+	pa_i_j ++;
+	pb_j_k += stride_b;
+	modint p1 = *pa_i_j * *pb_j_k;
+	pa_i_j ++;
+	pb_j_k += stride_b;
+	total += p0 + p1;
+      }
+      if (n%2) {
+	total += *pa_i_j * *pb_j_k;
+      }
+      pc_i[k] = total % MODULUS;
+    }
+    pa_i += stride_a;
+    pc_i += stride_c;
+  }
+}
+
 modint modint_random(void) {
   static uint64_t next = 1325997111;
   next = next * 1103515245 + 12345;
diff --git a/test/monniaux/int_mat_run.c b/test/monniaux/int_mat_run.c
index 9367fae0..9d5c0c57 100644
--- a/test/monniaux/int_mat_run.c
+++ b/test/monniaux/int_mat_run.c
@@ -59,20 +59,29 @@ int main() {
   modint_mat_mul4(m, n, p, c4, p, a, n, b, p);
   c4_time = get_cycle()-c4_time;
   
+  modint *c5 = malloc(sizeof(modint) * m * p);
+  cycle_t c5_time = get_cycle();
+  modint_mat_mul5(m, n, p, c5, p, a, n, b, p);
+  c5_time = get_cycle()-c5_time;
+  
   printf("c1==c2: %s\n"
 	 "c1==c3: %s\n"
 	 "c1==c4: %s\n"
+	 "c1==c5: %s\n"
 	 "c1_time = %" PRIu64 "\n"
 	 "c2_time = %" PRIu64 "\n"
 	 "c3_time = %" PRIu64 "\n"
-	 "c4_time = %" PRIu64 "\n",
+	 "c4_time = %" PRIu64 "\n"
+	 "c5_time = %" PRIu64 "\n",
 	 modint_mat_equal(m, n, c1, p, c2, p)?"true":"false",
 	 modint_mat_equal(m, n, c1, p, c3, p)?"true":"false",
 	 modint_mat_equal(m, n, c1, p, c4, p)?"true":"false",
+	 modint_mat_equal(m, n, c1, p, c5, p)?"true":"false",
 	 c1_time,
 	 c2_time,
 	 c3_time,
-	 c4_time);
+	 c4_time,
+	 c5_time);
   
   free(a);
   free(b);
@@ -80,5 +89,6 @@ int main() {
   free(c2);
   free(c3);
   free(c4);
+  free(c5);
   return 0;
 }
diff --git a/test/monniaux/modint.h b/test/monniaux/modint.h
index d3a154c1..5295258b 100644
--- a/test/monniaux/modint.h
+++ b/test/monniaux/modint.h
@@ -24,6 +24,11 @@ void modint_mat_mul4(unsigned m, unsigned n, unsigned p,
 		     const modint *a, unsigned stride_a,
 		     const modint *b, unsigned stride_b);
 
+void modint_mat_mul5(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b);
+
 modint modint_random(void);
 
 void modint_mat_random(unsigned m,