opt

diff --git a/mat.c b/mat.c
index 4236f35064a1e854864aded9742908898f2a44c3..65eb0f21220a4ea9cd76428c5ce8897b9bf1621f 100644 (file)
--- a/mat.c
+++ b/mat.c
@@ -183,9 +183,6 @@ void mat_mul_strassen(mat_t* m, mat_t* m0, mat_t* m1, int n_min)
        mat_t* p2 = mat_alloc(0, 0, n, NULL);
        mat_t* p3 = mat_alloc(0, 0, n, NULL);
        mat_t* p4 = mat_alloc(0, 0, n, NULL);
-       mat_t* p5 = mat_alloc(0, 0, n, NULL);
-       mat_t* p6 = mat_alloc(0, 0, n, NULL);
-       mat_t* p7 = mat_alloc(0, 0, n, NULL);
 
        // tmp mat t0
        mat_t* t0 = mat_alloc(0, 0, n, NULL);
@@ -208,42 +205,47 @@ void mat_mul_strassen(mat_t* m, mat_t* m0, mat_t* m1, int n_min)
        // t = p3 + p4
        mat_add(t, p3,  p4);
 
+       mat_sub(r, p4,  p2);
+       mat_free(p2);
+       mat_free(p4);
+
+       mat_sub(u, p1,  p3);
+       mat_free(p1);
+       mat_free(p3);
+
        // tmp mat t1
        mat_t* t1 = mat_alloc(0, 0, n, NULL);
 
        //p5 = (a + d) * (e + h)
+       mat_t* p5 = mat_alloc(0, 0, n, NULL);
        mat_add(t0, a,  d);
        mat_add(t1, e,  h);
        mat_mul_strassen(p5, t0, t1, n_min);
        //p6 = (b - d) * (g + h)
+       mat_t* p6 = mat_alloc(0, 0, n, NULL);
        mat_sub(t0, b,  d);
        mat_add(t1, g,  h);
        mat_mul_strassen(p6, t0, t1, n_min);
        // r = p5 + p4 - p2 + p6
-       mat_add(r, p5,  p4);
-       mat_sub(r, r,   p2);
+       mat_add(r, r,   p5);
        mat_add(r, r,   p6);
 
+       mat_free(p6);
 
        //p7 = (a - c) * (e + f)
+       mat_t* p7 = mat_alloc(0, 0, n, NULL);
        mat_sub(t0, a,  c);
        mat_add(t1, e,  f);
        mat_mul_strassen(p7, t0, t1, n_min);
        // u = p5 + p1 -p3 -p7
-       mat_add(u, p5,  p1);
-       mat_sub(u, u,   p3);
+       mat_add(u, u,   p5);
        mat_sub(u, u,   p7);
 
        // free unused mats
        mat_free(t0);
        mat_free(t1);
 
-       mat_free(p1);
-       mat_free(p2);
-       mat_free(p3);
-       mat_free(p4);
        mat_free(p5);
-       mat_free(p6);
        mat_free(p7);
 
        mat_free(a);
@@ -333,6 +335,8 @@ int main(int argc, char* argv[])
        mat_fill(m0);
        mat_fill(m1);
 
+       printf("%s(), g_buf_size_max: %d\n", __func__, g_buf_size_max);
+
        switch (flag) {
                case 0:
                        mat_mul(m2, m0, m1);