| 49 | } |
| 50 | |
| 51 | int main(int argc, char** argv) { |
| 52 | try { |
| 53 | int device = argc > 1 ? atoi(argv[1]) : 0; |
| 54 | setDevice(device); |
| 55 | info(); |
| 56 | |
| 57 | printf( |
| 58 | "** ArrayFire Black-Scholes Example **\n" |
| 59 | "** by AccelerEyes **\n\n"); |
| 60 | |
| 61 | array GC1(4000, 1, C1); |
| 62 | array GC2(4000, 1, C2); |
| 63 | array GC3(4000, 1, C3); |
| 64 | array GC4(4000, 1, C4); |
| 65 | array GC5(4000, 1, C5); |
| 66 | |
| 67 | // Compile kernels |
| 68 | // Create GPU copies of the data |
| 69 | array Sg = GC1; |
| 70 | array Xg = GC2; |
| 71 | array Rg = GC3; |
| 72 | array Vg = GC4; |
| 73 | array Tg = GC5; |
| 74 | array Cg, Pg; |
| 75 | |
| 76 | // Warm up black scholes example |
| 77 | black_scholes(Cg, Pg, Sg, Xg, Rg, Vg, Tg); |
| 78 | eval(Cg, Pg); |
| 79 | printf("Warming up done\n"); |
| 80 | af::sync(); |
| 81 | |
| 82 | int iter = 1000; |
| 83 | for (int n = 50; n <= 500; n += 50) { |
| 84 | // Create GPU copies of the data |
| 85 | Sg = tile(GC1, n, 1); |
| 86 | Xg = tile(GC2, n, 1); |
| 87 | Rg = tile(GC3, n, 1); |
| 88 | Vg = tile(GC4, n, 1); |
| 89 | Tg = tile(GC5, n, 1); |
| 90 | af::eval(Sg, Xg, Rg, Vg, Tg); |
| 91 | |
| 92 | dim4 dims = Xg.dims(); |
| 93 | // Force compute on the GPU |
| 94 | af::sync(); |
| 95 | |
| 96 | timer::start(); |
| 97 | for (int i = 0; i < iter; i++) { |
| 98 | black_scholes(Cg, Pg, Sg, Xg, Rg, Vg, Tg); |
| 99 | eval(Cg, Pg); |
| 100 | } |
| 101 | af::sync(); |
| 102 | |
| 103 | double t = timer::stop() / iter; |
| 104 | printf("Input Data Size = %8d. Mean GPU Time: %0.6f ms\n", |
| 105 | (int)dims[0], 1000 * t); |
| 106 | } |
| 107 | } catch (af::exception& e) { |
| 108 | fprintf(stderr, "%s\n", e.what()); |