131
|
1 // { dg-do compile { target { i?86-*-* x86_64-*-* } && { ! *-*-solaris* } } }
|
|
2 // { dg-options "-Ofast -funroll-loops -fopenmp -march=knl" }
|
|
3 // { dg-final { scan-assembler-not "vmovaps" } }
|
|
4
|
|
5 #include <math.h>
|
|
6
|
|
7 #include <xmmintrin.h>
|
|
8
|
|
9 #define max(a, b) ( (a) > (b) ? (a) : (b) )
|
|
10
|
|
11 struct Sdata {
|
|
12 float w;
|
|
13 float s;
|
|
14 float r;
|
|
15 float t;
|
|
16 float v;
|
|
17 };
|
|
18 extern int N1, N2, N3;
|
|
19
|
|
20 #define func(p, up, down) ((p)*(up) + (1.0f-(p)) * (down))
|
|
21
|
|
22 void foo (Sdata *in, int idx, float *out)
|
|
23 {
|
|
24 float* y1 = (float*)_mm_malloc(sizeof(float) * N1,16);
|
|
25 float* y2 = (float*)_mm_malloc(sizeof(float) * N1,16);
|
|
26 float* y3 = (float*)_mm_malloc(sizeof(float) * N1,16);
|
|
27 float* y4 = (float*)_mm_malloc(sizeof(float) * N1,16);
|
|
28
|
|
29 for (int k = idx; k < idx + N3; k++) {
|
|
30 float x1 = in[k].r;
|
|
31 float x2 = in[k].s;
|
|
32 float x3 = in[k].w;
|
|
33 float x4 = in[k].v;
|
|
34 float x5 = in[k].t;
|
|
35 x5 /= N2;
|
|
36 float u = exp(x4 * sqrt(x5));
|
|
37 float d = exp(-x4 * sqrt(x5));
|
|
38 float a = exp(x1 * x5);
|
|
39 float m = exp(-x1 * x5);
|
|
40 float p = (a - d) / (u - d);
|
|
41 y2[0] = x2;
|
|
42 y3[0] = float(1.f);
|
|
43 for (int i = 1; i <= N2; i++) {
|
|
44 y2[i] = u * y2[i - 1];
|
|
45 y3[i] = d * y3[i - 1];
|
|
46 }
|
|
47 #pragma omp simd
|
|
48 for (int i = 0; i <= N2; i++) {
|
|
49 y1[i] =
|
|
50 max((x3 - y2[N2 - i] * y3[i]), float(0.f));
|
|
51 }
|
|
52 for (int i = N2 - 1; i >= 0; i--) {
|
|
53 #pragma omp simd
|
|
54 for (int j = 0; j <= i; j++) {
|
|
55 y4[j] = func(p,y1[j],y1[j+1]) * m;
|
|
56 }
|
|
57 #pragma omp simd
|
|
58 for (int j = 0; j <= i; j++) {
|
|
59 float t1 = y2[i - j] * y3[j];
|
|
60 float t2 = max(x3 - t1, float(0.f));
|
|
61 y1[j] = max(t2, y4[j]);
|
|
62 }
|
|
63 }
|
|
64 out[k] = y1[0];
|
|
65 }
|
|
66 _mm_free(y1);
|
|
67 _mm_free(y2);
|
|
68 _mm_free(y3);
|
|
69 _mm_free(y4);
|
|
70 }
|