moved to subdirectory

4 files changed, 319 insertions, 0 deletions

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);