From 99e5f103393d554b0d2725303682a35d343a09b6 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Fri, 18 Jan 2019 15:16:17 +0100
Subject: moved to subdirectory

---
 test/monniaux/mod_int_mat/Makefile      |  46 +++++++++++
 test/monniaux/mod_int_mat/int_mat.c     | 138 ++++++++++++++++++++++++++++++++
 test/monniaux/mod_int_mat/int_mat_run.c |  94 ++++++++++++++++++++++
 test/monniaux/mod_int_mat/modint.h      |  41 ++++++++++
 4 files changed, 319 insertions(+)
 create mode 100644 test/monniaux/mod_int_mat/Makefile
 create mode 100644 test/monniaux/mod_int_mat/int_mat.c
 create mode 100644 test/monniaux/mod_int_mat/int_mat_run.c
 create mode 100644 test/monniaux/mod_int_mat/modint.h

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/Makefile b/test/monniaux/mod_int_mat/Makefile
new file mode 100644
index 00000000..be534653
--- /dev/null
+++ b/test/monniaux/mod_int_mat/Makefile
@@ -0,0 +1,46 @@
+CFLAGS=-Wall -O3
+K1C_CC=k1-mbr-gcc
+K1C_CFLAGS=-Wall -O3 -std=c99
+K1C_CCOMP=../../ccomp
+K1C_CCOMPFLAGS=-Wall -O3 -D__thread= -D__int128=int
+
+PRODUCTS=int_mat.host int_mat.gcc.k1c.out int_mat.ccomp.k1c.out int_mat.ccomp.k1c.s int_mat.gcc.k1c.s
+
+all:	$(PRODUCTS)
+
+%.gcc.k1c.s: %.c
+	$(K1C_CC) $(K1C_CFLAGS) -S $< -o $@
+
+%.gcc.k1c.o: %.gcc.k1c.s
+	$(K1C_CC) $(K1C_CFLAGS) -c $< -o $@
+
+%.gcc.k1c.o: %.gcc.k1c.c
+	$(K1C_CC) $(K1C_CFLAGS) -c $< -o $@
+
+%.ccomp.k1c.s: %.c
+	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) -S $< -o $@
+
+%.ccomp.k1c.o: %.ccomp.k1c.s
+	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) -c $< -o $@
+
+%.ccomp.k1c.o: %.ccomp.k1c.c
+	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) -c $< -o $@
+
+int_mat.host: int_mat.c int_mat_run.c modint.h
+	$(CC) $(CFLAGS) int_mat.c int_mat_run.c -o $@
+
+int_mat.gcc.k1c.s int_mat.ccomp.k1c.s: modint.h
+
+int_mat.gcc.k1c: int_mat.gcc.k1c.o int_mat_run.gcc.k1c.o
+	$(K1C_CC) $(K1C_CFLAGS) $+ -o $@
+
+int_mat.ccomp.k1c: int_mat.ccomp.k1c.o int_mat_run.gcc.k1c.o
+	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) $+ -o $@
+
+%.k1c.out: %.k1c
+	k1-cluster --cycle-based -- $< | tee $@
+
+clean:
+	$(RM) -f $(PRODUCTS) int_mat.gcc.k1c.o int_mat.ccomp.k1c.o
+
+.PHONY: clean
diff --git a/test/monniaux/mod_int_mat/int_mat.c b/test/monniaux/mod_int_mat/int_mat.c
new file mode 100644
index 00000000..d3e14e26
--- /dev/null
+++ b/test/monniaux/mod_int_mat/int_mat.c
@@ -0,0 +1,138 @@
+#include "modint.h"
+
+void modint_mat_mul1(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b) {
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned k=0; k<p; k++) {
+      c[i*stride_c+k] = 0;
+    }
+  }
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned k=0; k<p; k++) {
+      for(unsigned j=0; j<n; j++) {
+	c[i*stride_c+k] += a[i*stride_a+j] * b[j*stride_b+k]; 
+      }
+    }
+  }
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned k=0; k<p; k++) {
+      c[i*stride_c+k] %= MODULUS;
+    }
+  }  
+}
+
+void modint_mat_mul2(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b) {
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned k=0; k<p; k++) {
+      modint total = 0;
+      for(unsigned j=0; j<n; j++) {
+	total += a[i*stride_a + j] * b[j*stride_b + k];
+      }
+      c[i*stride_c+k] = total % MODULUS;
+    }
+  }
+}
+
+void modint_mat_mul3(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b) {
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned k=0; k<p; k++) {
+      modint total0 = 0, total1 = 0;
+      unsigned j;
+      for(j=0; j+1<n; j+=2) {
+	total0 += a[i*stride_a + j] * b[j*stride_b + k];
+	total1 += a[i*stride_a + (j+1)] * b[(j+1)*stride_b + k];
+      }
+      if (j < n) {
+	total0 += a[i*stride_a + j] * b[j*stride_b + k];
+      }
+      c[i*stride_c+k] = (total0+total1) % MODULUS;
+    }
+  }
+}
+
+void modint_mat_mul4(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 j=0; j<n; j++) {
+	total += *pa_i_j * *pb_j_k;
+	pa_i_j ++;
+	pb_j_k += stride_b;
+      }
+      pc_i[k] = total % MODULUS;
+    }
+    pa_i += stride_a;
+    pc_i += stride_c;
+  }
+}
+
+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;
+  return next % MODULUS;
+}
+
+void modint_mat_random(unsigned m,
+		       unsigned n,
+		       modint *a, unsigned stride_a) {
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned j=0; j<n; j++) {
+      a[i*stride_a + j] = modint_random();
+    }
+  }
+}
+
+bool modint_mat_equal(unsigned m,
+		      unsigned n,
+		      const modint *a, unsigned stride_a,
+		      const modint *b, unsigned stride_b) {
+  for(unsigned i=0; i<m; i++) {
+    for(unsigned j=0; j<n; j++) {
+      if (a[i*stride_a + j] != b[i*stride_b + j]) return false;
+    }
+  }
+  return true;
+}
diff --git a/test/monniaux/mod_int_mat/int_mat_run.c b/test/monniaux/mod_int_mat/int_mat_run.c
new file mode 100644
index 00000000..9d5c0c57
--- /dev/null
+++ b/test/monniaux/mod_int_mat/int_mat_run.c
@@ -0,0 +1,94 @@
+#include <stdbool.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <inttypes.h>
+#include "modint.h"
+
+typedef uint64_t cycle_t;
+
+#ifdef __K1C__
+#include <mppa_bare_runtime/k1c/registers.h>
+static inline void cycle_count_config(void)
+{
+        /* config pmc for cycle count */
+        uint64_t pmc_value = __builtin_k1_get(K1_SFR_PMC);
+
+        pmc_value &= ~(0xfULL);
+        __builtin_k1_set(K1_SFR_PMC, pmc_value);
+}
+
+static inline uint64_t get_cycle(void)
+{
+        return __builtin_k1_get(K1_SFR_PM0);
+}
+#else
+static inline void cycle_count_config(void) { }
+#ifdef  __x86_64__
+#include <x86intrin.h>
+static inline cycle_t get_cycle(void) { return __rdtsc(); }
+#else
+static inline cycle_t get_cycle(void) { return 0; }
+#endif
+#endif
+
+int main() {
+  const unsigned m = 40, n = 21, p = 30;
+  cycle_count_config();
+  modint *a = malloc(sizeof(modint) * m * n);
+  modint_mat_random(m, n, a, n);
+  modint *b = malloc(sizeof(modint) * n * p);
+  modint_mat_random(n, p, b, p);
+  
+  modint *c1 = malloc(sizeof(modint) * m * p);
+  cycle_t c1_time = get_cycle();
+  modint_mat_mul1(m, n, p, c1, p, a, n, b, p);
+  c1_time = get_cycle()-c1_time;
+  
+  modint *c2 = malloc(sizeof(modint) * m * p);
+  cycle_t c2_time = get_cycle();
+  modint_mat_mul2(m, n, p, c2, p, a, n, b, p);
+  c2_time = get_cycle()-c2_time;
+  
+  modint *c3 = malloc(sizeof(modint) * m * p);
+  cycle_t c3_time = get_cycle();
+  modint_mat_mul3(m, n, p, c3, p, a, n, b, p);
+  c3_time = get_cycle()-c3_time;
+  
+  modint *c4 = malloc(sizeof(modint) * m * p);
+  cycle_t c4_time = get_cycle();
+  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"
+	 "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,
+	 c5_time);
+  
+  free(a);
+  free(b);
+  free(c1);
+  free(c2);
+  free(c3);
+  free(c4);
+  free(c5);
+  return 0;
+}
diff --git a/test/monniaux/mod_int_mat/modint.h b/test/monniaux/mod_int_mat/modint.h
new file mode 100644
index 00000000..5295258b
--- /dev/null
+++ b/test/monniaux/mod_int_mat/modint.h
@@ -0,0 +1,41 @@
+#include <stdint.h>
+#include <stdbool.h>
+
+typedef uint32_t modint;
+#define MODULUS 257
+
+void modint_mat_mul1(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b);
+
+void modint_mat_mul2(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b);
+
+void modint_mat_mul3(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     const modint *a, unsigned stride_a,
+		     const modint *b, unsigned stride_b);
+
+void modint_mat_mul4(unsigned m, unsigned n, unsigned p,
+		     modint * restrict c, unsigned stride_c,
+		     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,
+		       unsigned n,
+		       modint *a, unsigned stride_a);
+
+bool modint_mat_equal(unsigned m,
+		      unsigned n,
+		      const modint *a, unsigned stride_a,
+		      const modint *b, unsigned stride_b);
-- 
cgit 


From 2a1db308610355d66c6baff004702fd00816e25a Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Fri, 18 Jan 2019 16:09:13 +0100
Subject: adjust path

---
 test/monniaux/mod_int_mat/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/Makefile b/test/monniaux/mod_int_mat/Makefile
index be534653..9a90a4ae 100644
--- a/test/monniaux/mod_int_mat/Makefile
+++ b/test/monniaux/mod_int_mat/Makefile
@@ -1,7 +1,7 @@
 CFLAGS=-Wall -O3
 K1C_CC=k1-mbr-gcc
 K1C_CFLAGS=-Wall -O3 -std=c99
-K1C_CCOMP=../../ccomp
+K1C_CCOMP=../../../ccomp
 K1C_CCOMPFLAGS=-Wall -O3 -D__thread= -D__int128=int
 
 PRODUCTS=int_mat.host int_mat.gcc.k1c.out int_mat.ccomp.k1c.out int_mat.ccomp.k1c.s int_mat.gcc.k1c.s
-- 
cgit 


From a873e3fb7164db2be4641b244a63895dfc0660dd Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Fri, 18 Jan 2019 17:17:23 +0100
Subject: loop transformation

---
 test/monniaux/mod_int_mat/int_mat.c     | 33 +++++++++++++++++++++++++++++++++
 test/monniaux/mod_int_mat/int_mat_run.c | 14 ++++++++++++--
 test/monniaux/mod_int_mat/modint.h      |  5 +++++
 3 files changed, 50 insertions(+), 2 deletions(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/int_mat.c b/test/monniaux/mod_int_mat/int_mat.c
index d3e14e26..58f968c1 100644
--- a/test/monniaux/mod_int_mat/int_mat.c
+++ b/test/monniaux/mod_int_mat/int_mat.c
@@ -109,6 +109,39 @@ void modint_mat_mul5(unsigned m, unsigned n, unsigned p,
   }
 }
 
+void modint_mat_mul6(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;
+      unsigned j2=0, n2=n/2;
+      if (n2 > 0) {
+	do {
+	  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;
+	  j2++;
+	} while (j2 < n2);
+      }
+      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/mod_int_mat/int_mat_run.c b/test/monniaux/mod_int_mat/int_mat_run.c
index 9d5c0c57..20d564cc 100644
--- a/test/monniaux/mod_int_mat/int_mat_run.c
+++ b/test/monniaux/mod_int_mat/int_mat_run.c
@@ -64,24 +64,33 @@ int main() {
   modint_mat_mul5(m, n, p, c5, p, a, n, b, p);
   c5_time = get_cycle()-c5_time;
   
+  modint *c6 = malloc(sizeof(modint) * m * p);
+  cycle_t c6_time = get_cycle();
+  modint_mat_mul6(m, n, p, c6, p, a, n, b, p);
+  c6_time = get_cycle()-c6_time;
+  
   printf("c1==c2: %s\n"
 	 "c1==c3: %s\n"
 	 "c1==c4: %s\n"
 	 "c1==c5: %s\n"
+	 "c1==c6: %s\n"
 	 "c1_time = %" PRIu64 "\n"
 	 "c2_time = %" PRIu64 "\n"
 	 "c3_time = %" PRIu64 "\n"
 	 "c4_time = %" PRIu64 "\n"
-	 "c5_time = %" PRIu64 "\n",
+	 "c5_time = %" PRIu64 "\n"
+	 "c6_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",
+	 modint_mat_equal(m, n, c1, p, c6, p)?"true":"false",
 	 c1_time,
 	 c2_time,
 	 c3_time,
 	 c4_time,
-	 c5_time);
+	 c5_time,
+	 c6_time);
   
   free(a);
   free(b);
@@ -90,5 +99,6 @@ int main() {
   free(c3);
   free(c4);
   free(c5);
+  free(c6);
   return 0;
 }
diff --git a/test/monniaux/mod_int_mat/modint.h b/test/monniaux/mod_int_mat/modint.h
index 5295258b..92005455 100644
--- a/test/monniaux/mod_int_mat/modint.h
+++ b/test/monniaux/mod_int_mat/modint.h
@@ -29,6 +29,11 @@ void modint_mat_mul5(unsigned m, unsigned n, unsigned p,
 		     const modint *a, unsigned stride_a,
 		     const modint *b, unsigned stride_b);
 
+void modint_mat_mul6(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,
-- 
cgit 


From 017d67d668d47fc67038226939653889814cbcac Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Fri, 18 Jan 2019 17:30:07 +0100
Subject: more on matrices

---
 test/monniaux/mod_int_mat/int_mat.c     | 53 +++++++++++++++++++++++++++++++--
 test/monniaux/mod_int_mat/int_mat_run.c | 15 ++++++++--
 test/monniaux/mod_int_mat/modint.h      |  5 ++++
 3 files changed, 68 insertions(+), 5 deletions(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/int_mat.c b/test/monniaux/mod_int_mat/int_mat.c
index 58f968c1..0e51d7ba 100644
--- a/test/monniaux/mod_int_mat/int_mat.c
+++ b/test/monniaux/mod_int_mat/int_mat.c
@@ -122,13 +122,12 @@ void modint_mat_mul6(unsigned m, unsigned n, unsigned p,
       unsigned j2=0, n2=n/2;
       if (n2 > 0) {
 	do {
-	  modint p0 = *pa_i_j * *pb_j_k;
+	  total += *pa_i_j * *pb_j_k;
 	  pa_i_j ++;
 	  pb_j_k += stride_b;
-	  modint p1 = *pa_i_j * *pb_j_k;
+	  total += *pa_i_j * *pb_j_k;
 	  pa_i_j ++;
 	  pb_j_k += stride_b;
-	  total += p0 + p1;
 	  j2++;
 	} while (j2 < n2);
       }
@@ -142,6 +141,54 @@ void modint_mat_mul6(unsigned m, unsigned n, unsigned p,
   }
 }
 
+void modint_mat_mul7(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;
+      {
+	unsigned j4=0, n4=n/4;
+	if (n4 > 0) {
+	  do {
+	    total += *pa_i_j * *pb_j_k;
+	    pa_i_j ++;
+	    pb_j_k += stride_b;
+	    total += *pa_i_j * *pb_j_k;
+	    pa_i_j ++;
+	    pb_j_k += stride_b;
+	    total += *pa_i_j * *pb_j_k;
+	    pa_i_j ++;
+	    pb_j_k += stride_b;
+	    total += *pa_i_j * *pb_j_k;
+	    pa_i_j ++;
+	    pb_j_k += stride_b;
+	    j4++;
+	  } while (j4 < n4);
+	}
+      }
+      {
+	unsigned j4=0, n4=n%4;
+	if (n4 > 0) {
+	  do {
+	    total += *pa_i_j * *pb_j_k;
+	    pa_i_j ++;
+	    pb_j_k += stride_b;
+	    j4++;
+	  } while (j4 < n4);
+	}
+      }
+      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/mod_int_mat/int_mat_run.c b/test/monniaux/mod_int_mat/int_mat_run.c
index 20d564cc..42cb54fb 100644
--- a/test/monniaux/mod_int_mat/int_mat_run.c
+++ b/test/monniaux/mod_int_mat/int_mat_run.c
@@ -69,28 +69,39 @@ int main() {
   modint_mat_mul6(m, n, p, c6, p, a, n, b, p);
   c6_time = get_cycle()-c6_time;
   
+  modint *c7 = malloc(sizeof(modint) * m * p);
+  cycle_t c7_time = get_cycle();
+  modint_mat_mul7(m, n, p, c7, p, a, n, b, p);
+  c7_time = get_cycle()-c7_time;
+  
   printf("c1==c2: %s\n"
 	 "c1==c3: %s\n"
 	 "c1==c4: %s\n"
 	 "c1==c5: %s\n"
 	 "c1==c6: %s\n"
+	 "c1==c7: %s\n"
 	 "c1_time = %" PRIu64 "\n"
 	 "c2_time = %" PRIu64 "\n"
 	 "c3_time = %" PRIu64 "\n"
 	 "c4_time = %" PRIu64 "\n"
 	 "c5_time = %" PRIu64 "\n"
-	 "c6_time = %" PRIu64 "\n",
+	 "c6_time = %" PRIu64 "\n"
+	 "c7_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",
 	 modint_mat_equal(m, n, c1, p, c6, p)?"true":"false",
+	 modint_mat_equal(m, n, c1, p, c7, p)?"true":"false",
+	 
 	 c1_time,
 	 c2_time,
 	 c3_time,
 	 c4_time,
 	 c5_time,
-	 c6_time);
+	 c6_time,
+	 c7_time);
   
   free(a);
   free(b);
diff --git a/test/monniaux/mod_int_mat/modint.h b/test/monniaux/mod_int_mat/modint.h
index 92005455..15c70a15 100644
--- a/test/monniaux/mod_int_mat/modint.h
+++ b/test/monniaux/mod_int_mat/modint.h
@@ -34,6 +34,11 @@ void modint_mat_mul6(unsigned m, unsigned n, unsigned p,
 		     const modint *a, unsigned stride_a,
 		     const modint *b, unsigned stride_b);
 
+void modint_mat_mul7(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,
-- 
cgit 


From 91c23dece6e83c1b6566530ecf9467e091934474 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Fri, 18 Jan 2019 21:33:13 +0100
Subject: fix Makefile

---
 test/monniaux/mod_int_mat/Makefile | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/Makefile b/test/monniaux/mod_int_mat/Makefile
index 9a90a4ae..912d016b 100644
--- a/test/monniaux/mod_int_mat/Makefile
+++ b/test/monniaux/mod_int_mat/Makefile
@@ -4,7 +4,7 @@ K1C_CFLAGS=-Wall -O3 -std=c99
 K1C_CCOMP=../../../ccomp
 K1C_CCOMPFLAGS=-Wall -O3 -D__thread= -D__int128=int
 
-PRODUCTS=int_mat.host int_mat.gcc.k1c.out int_mat.ccomp.k1c.out int_mat.ccomp.k1c.s int_mat.gcc.k1c.s
+PRODUCTS=int_mat.host int_mat.gcc.k1c.out int_mat.ccomp.k1c.out int_mat.ccomp.k1c.s int_mat.gcc.k1c.s int_mat.gcc.k1c int_mat.ccomp.k1c
 
 all:	$(PRODUCTS)
 
@@ -41,6 +41,6 @@ int_mat.ccomp.k1c: int_mat.ccomp.k1c.o int_mat_run.gcc.k1c.o
 	k1-cluster --cycle-based -- $< | tee $@
 
 clean:
-	$(RM) -f $(PRODUCTS) int_mat.gcc.k1c.o int_mat.ccomp.k1c.o
+	$(RM) -f $(PRODUCTS) int_mat.gcc.k1c.o int_mat.ccomp.k1c.o int_mat_run.gcc.k1c.o
 
 .PHONY: clean
-- 
cgit 


From caa53a3dce35e53ed913557f30aef1f063cf67c0 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Sat, 19 Jan 2019 13:21:49 +0100
Subject: fixes in Makefile

---
 test/monniaux/mod_int_mat/Makefile | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/Makefile b/test/monniaux/mod_int_mat/Makefile
index 912d016b..b285c03c 100644
--- a/test/monniaux/mod_int_mat/Makefile
+++ b/test/monniaux/mod_int_mat/Makefile
@@ -14,22 +14,16 @@ all:	$(PRODUCTS)
 %.gcc.k1c.o: %.gcc.k1c.s
 	$(K1C_CC) $(K1C_CFLAGS) -c $< -o $@
 
-%.gcc.k1c.o: %.gcc.k1c.c
-	$(K1C_CC) $(K1C_CFLAGS) -c $< -o $@
-
 %.ccomp.k1c.s: %.c
 	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) -S $< -o $@
 
 %.ccomp.k1c.o: %.ccomp.k1c.s
 	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) -c $< -o $@
 
-%.ccomp.k1c.o: %.ccomp.k1c.c
-	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) -c $< -o $@
-
 int_mat.host: int_mat.c int_mat_run.c modint.h
 	$(CC) $(CFLAGS) int_mat.c int_mat_run.c -o $@
 
-int_mat.gcc.k1c.s int_mat.ccomp.k1c.s: modint.h
+int_mat.gcc.k1c.s int_mat.ccomp.k1c.s int_mat_run.gcc.k1c.s: modint.h
 
 int_mat.gcc.k1c: int_mat.gcc.k1c.o int_mat_run.gcc.k1c.o
 	$(K1C_CC) $(K1C_CFLAGS) $+ -o $@
-- 
cgit 


From 6c5df17a302e99dec62aa749ffc30745cc7679a4 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Sat, 19 Jan 2019 13:33:57 +0100
Subject: quicksort

---
 test/monniaux/mod_int_mat/int_mat_run.c | 28 +---------------------------
 1 file changed, 1 insertion(+), 27 deletions(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/int_mat_run.c b/test/monniaux/mod_int_mat/int_mat_run.c
index 42cb54fb..f3955345 100644
--- a/test/monniaux/mod_int_mat/int_mat_run.c
+++ b/test/monniaux/mod_int_mat/int_mat_run.c
@@ -3,33 +3,7 @@
 #include <stdio.h>
 #include <inttypes.h>
 #include "modint.h"
-
-typedef uint64_t cycle_t;
-
-#ifdef __K1C__
-#include <mppa_bare_runtime/k1c/registers.h>
-static inline void cycle_count_config(void)
-{
-        /* config pmc for cycle count */
-        uint64_t pmc_value = __builtin_k1_get(K1_SFR_PMC);
-
-        pmc_value &= ~(0xfULL);
-        __builtin_k1_set(K1_SFR_PMC, pmc_value);
-}
-
-static inline uint64_t get_cycle(void)
-{
-        return __builtin_k1_get(K1_SFR_PM0);
-}
-#else
-static inline void cycle_count_config(void) { }
-#ifdef  __x86_64__
-#include <x86intrin.h>
-static inline cycle_t get_cycle(void) { return __rdtsc(); }
-#else
-static inline cycle_t get_cycle(void) { return 0; }
-#endif
-#endif
+#include "../cycles.h"
 
 int main() {
   const unsigned m = 40, n = 21, p = 30;
-- 
cgit 


From d70a9e55c3595cf7ee84ca9f3b1d5e272a5e3999 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Sat, 19 Jan 2019 13:42:40 +0100
Subject: use a prime in PRNG

---
 test/monniaux/mod_int_mat/int_mat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/int_mat.c b/test/monniaux/mod_int_mat/int_mat.c
index 0e51d7ba..fc763811 100644
--- a/test/monniaux/mod_int_mat/int_mat.c
+++ b/test/monniaux/mod_int_mat/int_mat.c
@@ -191,7 +191,7 @@ void modint_mat_mul7(unsigned m, unsigned n, unsigned p,
 
 modint modint_random(void) {
   static uint64_t next = 1325997111;
-  next = next * 1103515245 + 12345;
+  next = next * 1103515249 + 12345;
   return next % MODULUS;
 }
 
-- 
cgit 


From d8bef07031a1928e7e071c47f3d3702d4d9a4ca3 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Wed, 30 Jan 2019 16:08:49 +0100
Subject: remove preprocessor directives about SIZE_TYPE, now useless (long is
 now 8 bytes like in gcc)

---
 test/monniaux/mod_int_mat/int_mat.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/int_mat.c b/test/monniaux/mod_int_mat/int_mat.c
index fc763811..167c91fd 100644
--- a/test/monniaux/mod_int_mat/int_mat.c
+++ b/test/monniaux/mod_int_mat/int_mat.c
@@ -190,7 +190,7 @@ void modint_mat_mul7(unsigned m, unsigned n, unsigned p,
 }
 
 modint modint_random(void) {
-  static uint64_t next = 1325997111;
+  static uint32_t next = 1325997111;
   next = next * 1103515249 + 12345;
   return next % MODULUS;
 }
-- 
cgit 


From d48a2f396dede39eae20fcfe551bb11832360a5b Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Wed, 30 Jan 2019 17:13:15 +0100
Subject: remove cruft dealing with __int128 and __thread

---
 test/monniaux/mod_int_mat/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/Makefile b/test/monniaux/mod_int_mat/Makefile
index b285c03c..f904c1e4 100644
--- a/test/monniaux/mod_int_mat/Makefile
+++ b/test/monniaux/mod_int_mat/Makefile
@@ -2,7 +2,7 @@ CFLAGS=-Wall -O3
 K1C_CC=k1-mbr-gcc
 K1C_CFLAGS=-Wall -O3 -std=c99
 K1C_CCOMP=../../../ccomp
-K1C_CCOMPFLAGS=-Wall -O3 -D__thread= -D__int128=int
+K1C_CCOMPFLAGS=-Wall -O3
 
 PRODUCTS=int_mat.host int_mat.gcc.k1c.out int_mat.ccomp.k1c.out int_mat.ccomp.k1c.s int_mat.gcc.k1c.s int_mat.gcc.k1c int_mat.ccomp.k1c
 
-- 
cgit 


From 4ba61956bec43ad004e37d1ad33c145daddf0ea2 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Thu, 31 Jan 2019 14:26:03 +0100
Subject: bigger matrix

---
 test/monniaux/mod_int_mat/int_mat_run.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/int_mat_run.c b/test/monniaux/mod_int_mat/int_mat_run.c
index f3955345..9ea1edd0 100644
--- a/test/monniaux/mod_int_mat/int_mat_run.c
+++ b/test/monniaux/mod_int_mat/int_mat_run.c
@@ -6,7 +6,7 @@
 #include "../cycles.h"
 
 int main() {
-  const unsigned m = 40, n = 21, p = 30;
+  const unsigned m = 60, n = 31, p = 50;
   cycle_count_config();
   modint *a = malloc(sizeof(modint) * m * n);
   modint_mat_random(m, n, a, n);
-- 
cgit 


From 18540b654ac649d56bfd7261bdc99c7506aaf602 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Mon, 18 Feb 2019 11:41:24 +0100
Subject: forgot to deallocate block -g changes performance

---
 test/monniaux/mod_int_mat/int_mat_run.c | 1 +
 1 file changed, 1 insertion(+)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/int_mat_run.c b/test/monniaux/mod_int_mat/int_mat_run.c
index 9ea1edd0..860c248d 100644
--- a/test/monniaux/mod_int_mat/int_mat_run.c
+++ b/test/monniaux/mod_int_mat/int_mat_run.c
@@ -85,5 +85,6 @@ int main() {
   free(c4);
   free(c5);
   free(c6);
+  free(c7);
   return 0;
 }
-- 
cgit 


From c0558ea2fd66679eeca136b41c4378ebebb9b3a0 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Fri, 29 Mar 2019 18:30:07 +0100
Subject: use C99 mode

---
 test/monniaux/mod_int_mat/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/Makefile b/test/monniaux/mod_int_mat/Makefile
index f904c1e4..d1365b34 100644
--- a/test/monniaux/mod_int_mat/Makefile
+++ b/test/monniaux/mod_int_mat/Makefile
@@ -1,4 +1,4 @@
-CFLAGS=-Wall -O3
+CFLAGS=-Wall -O3 -std=c99
 K1C_CC=k1-mbr-gcc
 K1C_CFLAGS=-Wall -O3 -std=c99
 K1C_CCOMP=../../../ccomp
-- 
cgit 


From a789b94e986375a713263d7d7584ceaac01084f2 Mon Sep 17 00:00:00 2001
From: David Monniaux <david.monniaux@univ-grenoble-alpes.fr>
Date: Tue, 28 May 2019 15:13:46 +0200
Subject: adaptation pour k1c-cos

---
 test/monniaux/mod_int_mat/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'test/monniaux/mod_int_mat')

diff --git a/test/monniaux/mod_int_mat/Makefile b/test/monniaux/mod_int_mat/Makefile
index d1365b34..08b97b67 100644
--- a/test/monniaux/mod_int_mat/Makefile
+++ b/test/monniaux/mod_int_mat/Makefile
@@ -1,5 +1,5 @@
 CFLAGS=-Wall -O3 -std=c99
-K1C_CC=k1-mbr-gcc
+K1C_CC=k1-cos-gcc
 K1C_CFLAGS=-Wall -O3 -std=c99
 K1C_CCOMP=../../../ccomp
 K1C_CCOMPFLAGS=-Wall -O3
-- 
cgit