diff options
Diffstat (limited to 'result')
| -rw-r--r-- | result/ldc_2d_benchmark.K2200.data (renamed from result/ldc_2d_benchmark.k2200.data) | 0 | ||||
| -rw-r--r-- | result/ldc_2d_benchmark.K2200.log (renamed from result/ldc_2d_benchmark.k2200.log) | 0 | ||||
| -rw-r--r-- | result/ldc_2d_benchmark.P100.data | 1 | ||||
| -rw-r--r-- | result/ldc_2d_benchmark.P100.log | 156 | ||||
| -rw-r--r-- | result/ldc_3d_benchmark.K2200.data (renamed from result/ldc_3d_benchmark.k2200.data) | 0 | ||||
| -rw-r--r-- | result/ldc_3d_benchmark.K2200.log (renamed from result/ldc_3d_benchmark.k2200.log) | 0 | ||||
| -rw-r--r-- | result/ldc_3d_benchmark.P100.data | 1 | ||||
| -rw-r--r-- | result/ldc_3d_benchmark.P100.log | 144 |
8 files changed, 302 insertions, 0 deletions
diff --git a/result/ldc_2d_benchmark.k2200.data b/result/ldc_2d_benchmark.K2200.data index 8079473..8079473 100644 --- a/result/ldc_2d_benchmark.k2200.data +++ b/result/ldc_2d_benchmark.K2200.data diff --git a/result/ldc_2d_benchmark.k2200.log b/result/ldc_2d_benchmark.K2200.log index b88df82..b88df82 100644 --- a/result/ldc_2d_benchmark.k2200.log +++ b/result/ldc_2d_benchmark.K2200.log diff --git a/result/ldc_2d_benchmark.P100.data b/result/ldc_2d_benchmark.P100.data new file mode 100644 index 0000000..4552a35 --- /dev/null +++ b/result/ldc_2d_benchmark.P100.data @@ -0,0 +1 @@ +[((32, (16, 1), 'single', True), [14, 14, 14, 14, 14, 14, 15, 14, 15, 14]), ((32, (16, 1), 'single', False), [15, 15, 15, 15, 15, 15, 15, 15, 15, 15]), ((32, (16, 1), 'double', True), [15, 15, 15, 15, 15, 15, 15, 15, 15, 15]), ((32, (16, 1), 'double', False), [15, 15, 15, 15, 15, 15, 15, 15, 15, 15]), ((32, (32, 1), 'single', True), [14, 15, 15, 15, 15, 15, 15, 15, 15, 15]), ((32, (32, 1), 'single', False), [15, 15, 15, 15, 15, 15, 15, 15, 15, 15]), ((32, (32, 1), 'double', True), [15, 15, 15, 15, 15, 15, 15, 15, 15, 15]), ((32, (32, 1), 'double', False), [15, 15, 15, 15, 15, 15, 15, 15, 15, 15]), ((64, (16, 1), 'single', True), [59, 59, 61, 59, 56, 56, 56, 58, 60, 61]), ((64, (16, 1), 'single', False), [58, 59, 60, 60, 60, 59, 59, 59, 59, 59]), ((64, (16, 1), 'double', True), [58, 59, 59, 60, 60, 59, 60, 59, 59, 60]), ((64, (16, 1), 'double', False), [56, 56, 56, 57, 57, 57, 57, 57, 57, 56]), ((64, (32, 1), 'single', True), [59, 60, 59, 60, 61, 60, 61, 60, 61, 61]), ((64, (32, 1), 'single', False), [59, 59, 60, 60, 60, 60, 60, 59, 60, 60]), ((64, (32, 1), 'double', True), [58, 60, 60, 59, 60, 60, 59, 60, 60, 59]), ((64, (32, 1), 'double', False), [57, 60, 59, 60, 61, 60, 59, 60, 60, 60]), ((64, (64, 1), 'single', True), [59, 60, 60, 60, 60, 59, 60, 61, 60, 61]), ((64, (64, 1), 'single', False), [58, 58, 59, 59, 59, 58, 56, 59, 56, 58]), ((64, (64, 1), 'double', True), [58, 60, 59, 60, 60, 61, 61, 60, 60, 60]), ((64, (64, 1), 'double', False), [59, 59, 59, 59, 59, 59, 60, 59, 60, 59]), ((96, (16, 1), 'single', True), [132, 132, 133, 131, 134, 134, 135, 134, 134, 134]), ((96, (16, 1), 'single', False), [132, 131, 132, 132, 135, 134, 132, 133, 133, 133]), ((96, (16, 1), 'double', True), [132, 133, 130, 133, 134, 134, 135, 133, 134, 133]), ((96, (16, 1), 'double', False), [126, 125, 128, 128, 126, 126, 123, 127, 126, 127]), ((96, (24, 1), 'single', True), [122, 124, 125, 126, 132, 135, 133, 128, 133, 133]), ((96, (24, 1), 'single', False), [132, 133, 135, 135, 134, 135, 135, 133, 132, 133]), ((96, (24, 1), 'double', True), [132, 135, 133, 135, 135, 135, 132, 135, 134, 135]), ((96, (24, 1), 'double', False), [133, 133, 134, 131, 135, 131, 136, 136, 133, 134]), ((96, (32, 1), 'single', True), [133, 136, 135, 135, 133, 137, 131, 135, 135, 134]), ((96, (32, 1), 'single', False), [132, 132, 134, 133, 133, 134, 134, 134, 132, 135]), ((96, (32, 1), 'double', True), [132, 135, 134, 136, 135, 135, 136, 134, 136, 135]), ((96, (32, 1), 'double', False), [129, 133, 134, 136, 137, 135, 136, 132, 134, 136]), ((96, (48, 1), 'single', True), [129, 133, 131, 132, 133, 134, 135, 136, 134, 135]), ((96, (48, 1), 'single', False), [133, 135, 135, 136, 135, 135, 136, 135, 135, 135]), ((96, (48, 1), 'double', True), [134, 136, 136, 135, 137, 137, 136, 136, 135, 137]), ((96, (48, 1), 'double', False), [132, 134, 134, 134, 132, 133, 134, 135, 134, 133]), ((96, (96, 1), 'single', True), [135, 137, 135, 137, 137, 136, 138, 137, 137, 136]), ((96, (96, 1), 'single', False), [134, 134, 136, 135, 137, 138, 137, 136, 135, 137]), ((96, (96, 1), 'double', True), [134, 135, 136, 137, 136, 137, 138, 138, 137, 139]), ((96, (96, 1), 'double', False), [134, 136, 137, 136, 136, 137, 135, 137, 138, 135]), ((128, (16, 1), 'single', True), [238, 234, 238, 243, 243, 244, 242, 244, 240, 238]), ((128, (16, 1), 'single', False), [218, 218, 222, 222, 222, 221, 221, 222, 221, 220]), ((128, (16, 1), 'double', True), [227, 229, 229, 229, 223, 216, 218, 219, 212, 220]), ((128, (16, 1), 'double', False), [201, 205, 204, 204, 205, 205, 205, 204, 205, 204]), ((128, (32, 1), 'single', True), [214, 215, 214, 217, 219, 217, 217, 221, 230, 224]), ((128, (32, 1), 'single', False), [236, 239, 240, 236, 235, 236, 235, 236, 238, 240]), ((128, (32, 1), 'double', True), [237, 241, 240, 241, 241, 237, 241, 241, 242, 243]), ((128, (32, 1), 'double', False), [239, 244, 243, 241, 244, 243, 244, 244, 243, 241]), ((128, (64, 1), 'single', True), [239, 242, 241, 243, 242, 242, 242, 242, 242, 236]), ((128, (64, 1), 'single', False), [237, 241, 241, 239, 239, 240, 242, 241, 241, 239]), ((128, (64, 1), 'double', True), [240, 240, 241, 243, 240, 242, 242, 244, 240, 245]), ((128, (64, 1), 'double', False), [234, 237, 231, 237, 236, 241, 240, 239, 242, 240]), ((128, (128, 1), 'single', True), [236, 241, 238, 237, 240, 243, 242, 240, 240, 242]), ((128, (128, 1), 'single', False), [238, 239, 235, 242, 235, 240, 238, 239, 240, 238]), ((128, (128, 1), 'double', True), [238, 238, 241, 239, 238, 239, 239, 238, 237, 235]), ((128, (128, 1), 'double', False), [239, 236, 242, 238, 243, 238, 237, 238, 234, 241]), ((256, (16, 1), 'single', True), [937, 948, 946, 955, 960, 957, 963, 960, 956, 954]), ((256, (16, 1), 'single', False), [926, 932, 943, 929, 931, 929, 933, 928, 930, 944]), ((256, (16, 1), 'double', True), [944, 958, 957, 955, 960, 957, 960, 955, 966, 963]), ((256, (16, 1), 'double', False), [902, 901, 898, 905, 916, 919, 911, 907, 904, 900]), ((256, (32, 1), 'single', True), [960, 954, 974, 979, 981, 982, 978, 975, 970, 976]), ((256, (32, 1), 'single', False), [952, 956, 970, 971, 960, 968, 966, 955, 967, 971]), ((256, (32, 1), 'double', True), [946, 946, 968, 972, 975, 972, 962, 974, 972, 970]), ((256, (32, 1), 'double', False), [955, 955, 962, 943, 958, 957, 962, 967, 965, 974]), ((256, (64, 1), 'single', True), [943, 960, 937, 966, 966, 964, 969, 970, 970, 964]), ((256, (64, 1), 'single', False), [953, 965, 963, 971, 966, 975, 974, 972, 974, 971]), ((256, (64, 1), 'double', True), [961, 974, 934, 968, 975, 977, 958, 975, 974, 974]), ((256, (64, 1), 'double', False), [947, 961, 962, 957, 964, 966, 951, 967, 969, 958]), ((256, (128, 1), 'single', True), [942, 953, 932, 962, 965, 960, 964, 963, 955, 963]), ((256, (128, 1), 'single', False), [944, 960, 917, 946, 833, 803, 902, 965, 958, 962]), ((256, (128, 1), 'double', True), [931, 950, 969, 973, 976, 969, 960, 951, 970, 958]), ((256, (128, 1), 'double', False), [948, 958, 966, 970, 955, 959, 960, 969, 962, 967]), ((256, (256, 1), 'single', True), [933, 964, 966, 956, 974, 973, 971, 974, 972, 971]), ((256, (256, 1), 'single', False), [953, 970, 970, 970, 966, 962, 971, 964, 969, 965]), ((256, (256, 1), 'double', True), [952, 966, 970, 968, 972, 972, 971, 974, 981, 976]), ((256, (256, 1), 'double', False), [937, 956, 956, 959, 961, 964, 966, 948, 966, 958]), ((512, (16, 1), 'single', True), [3758, 3817, 3803, 3814, 3816, 3811, 3804, 3808, 3807, 3830]), ((512, (16, 1), 'single', False), [3289, 3027, 2960, 2908, 2856, 2829, 2830, 2827, 2831, 2832]), ((512, (16, 1), 'double', True), [3252, 3265, 3278, 3277, 3277, 3277, 3275, 3275, 3274, 3278]), ((512, (16, 1), 'double', False), [1303, 1256, 1225, 1220, 1221, 1222, 1221, 1221, 1222, 1222]), ((512, (32, 1), 'single', True), [3768, 3788, 3814, 3816, 3829, 3788, 3707, 3770, 3809, 3773]), ((512, (32, 1), 'single', False), [3729, 3795, 2896, 3638, 3745, 3759, 3799, 3823, 3810, 3824]), ((512, (32, 1), 'double', True), [3270, 3294, 3295, 3285, 3279, 3283, 3288, 3284, 3281, 3283]), ((512, (32, 1), 'double', False), [2381, 2322, 2286, 2285, 2284, 2288, 2289, 2290, 2286, 2289]), ((512, (64, 1), 'single', True), [3718, 3773, 3809, 3809, 3794, 3815, 3784, 3837, 3829, 3823]), ((512, (64, 1), 'single', False), [3747, 3796, 3799, 3826, 3809, 3820, 3773, 3837, 3837, 3818]), ((512, (64, 1), 'double', True), [3282, 3303, 3299, 3304, 3303, 3306, 3293, 3302, 3298, 3304]), ((512, (64, 1), 'double', False), [2372, 2315, 2274, 2269, 2270, 2276, 2276, 2274, 2276, 2274]), ((512, (128, 1), 'single', True), [3667, 3760, 3805, 3790, 3766, 3763, 3756, 3821, 3770, 3793]), ((512, (128, 1), 'single', False), [2337, 3498, 3708, 3639, 3708, 3511, 2649, 3632, 3681, 3725]), ((512, (128, 1), 'double', True), [3275, 3298, 3298, 3297, 3300, 3298, 3301, 3296, 3296, 3295]), ((512, (128, 1), 'double', False), [2368, 2313, 2265, 2264, 2263, 2263, 2265, 2267, 2267, 2268]), ((512, (256, 1), 'single', True), [3684, 3785, 3654, 3725, 3774, 3816, 3799, 3819, 3790, 3800]), ((512, (256, 1), 'single', False), [3831, 3853, 3903, 3884, 3897, 3906, 3893, 3906, 3914, 3906]), ((512, (256, 1), 'double', True), [3270, 3290, 3291, 3292, 3290, 3285, 3290, 3291, 3292, 3290]), ((512, (256, 1), 'double', False), [2378, 2325, 2279, 2266, 2275, 2277, 2281, 2278, 2289, 2289]), ((512, (512, 1), 'single', True), [3762, 3819, 3827, 3839, 3825, 3844, 3822, 3863, 3856, 3846]), ((512, (512, 1), 'single', False), [3757, 3810, 3803, 3828, 3777, 3820, 3831, 3829, 3840, 3826]), ((512, (512, 1), 'double', True), [3272, 3294, 3288, 3292, 3296, 3295, 3292, 3292, 3290, 3292]), ((512, (512, 1), 'double', False), [2196, 2126, 2115, 2113, 2114, 2113, 2115, 2113, 2115, 2113]), ((1024, (16, 1), 'single', True), [6315, 6335, 6334, 6333, 6333, 6331, 6336, 6335, 6335, 6337]), ((1024, (16, 1), 'single', False), [3978, 3949, 3913, 3869, 3825, 3785, 3749, 3713, 3676, 3641]), ((1024, (16, 1), 'double', True), [3470, 3473, 3473, 3471, 3474, 3474, 3474, 3474, 3475, 3473]), ((1024, (16, 1), 'double', False), [1380, 1344, 1319, 1297, 1276, 1262, 1262, 1263, 1260, 1256]), ((1024, (32, 1), 'single', True), [6628, 6642, 6636, 6640, 6644, 6645, 6648, 6644, 6644, 6645]), ((1024, (32, 1), 'single', False), [6493, 6510, 6510, 6507, 6499, 6499, 6494, 6483, 6489, 6480]), ((1024, (32, 1), 'double', True), [3461, 3472, 3472, 3473, 3472, 3473, 3469, 3471, 3473, 3472]), ((1024, (32, 1), 'double', False), [2673, 2615, 2571, 2529, 2492, 2469, 2474, 2476, 2479, 2480]), ((1024, (64, 1), 'single', True), [6575, 6592, 6587, 6591, 6581, 6587, 6595, 6594, 6593, 6591]), ((1024, (64, 1), 'single', False), [6418, 6454, 6450, 6446, 6443, 6432, 6429, 6431, 6429, 6431]), ((1024, (64, 1), 'double', True), [3440, 3448, 3452, 3451, 3453, 3455, 3451, 3457, 3451, 3450]), ((1024, (64, 1), 'double', False), [2668, 2610, 2565, 2524, 2487, 2466, 2469, 2473, 2475, 2474]), ((1024, (128, 1), 'single', True), [6577, 6609, 6610, 6608, 6609, 6609, 6604, 6609, 6613, 6606]), ((1024, (128, 1), 'single', False), [6402, 6443, 6448, 6441, 6436, 6434, 6421, 6419, 6421, 6410]), ((1024, (128, 1), 'double', True), [3446, 3451, 3451, 3449, 3450, 3451, 3448, 3451, 3450, 3451]), ((1024, (128, 1), 'double', False), [2672, 2618, 2573, 2531, 2491, 2467, 2469, 2473, 2475, 2476]), ((1024, (256, 1), 'single', True), [6589, 6605, 6604, 6605, 6603, 6606, 6604, 6607, 6597, 6601]), ((1024, (256, 1), 'single', False), [6365, 6391, 6384, 6367, 6362, 6348, 6349, 6337, 6342, 6314]), ((1024, (256, 1), 'double', True), [3454, 3457, 3456, 3455, 3457, 3456, 3457, 3458, 3456, 3456]), ((1024, (256, 1), 'double', False), [2709, 2651, 2631, 2577, 2540, 2516, 2525, 2526, 2527, 2528]), ((1024, (512, 1), 'single', True), [6584, 6595, 6592, 6598, 6599, 6592, 6598, 6588, 6602, 6594]), ((1024, (512, 1), 'single', False), [6166, 6202, 6187, 6175, 6153, 6146, 6129, 6131, 6098, 6050]), ((1024, (512, 1), 'double', True), [3461, 3463, 3465, 3465, 3465, 3465, 3464, 3466, 3465, 3464]), ((1024, (512, 1), 'double', False), [2613, 2539, 2490, 2447, 2405, 2359, 2357, 2358, 2357, 2358]), ((1024, (1024, 1), 'single', True), [6536, 6547, 6545, 6551, 6551, 6552, 6550, 6546, 6549, 6547]), ((1024, (1024, 1), 'single', False), [5703, 5711, 5692, 5667, 5630, 5589, 5574, 5564, 5530, 5497]), ((1024, (1024, 1), 'double', True), [3417, 3421, 3420, 3420, 3421, 3421, 3422, 3422, 3421, 3423]), ((1024, (1024, 1), 'double', False), [2346, 2322, 2301, 2274, 2241, 2215, 2216, 2217, 2219, 2218]), ,((2048, (16, 1), 'single', True), [6450, 6508, 6522, 6523, 6523, 6524, 6522, 6522, 6524, 6521]), ((2048, (16, 1), 'double', True), [3552, 3556, 3555, 3555, 3555, 3555, 3555, 3556, 3555, 3556]), ((2048, (32, 1), 'single', True), [6813, 6829, 6831, 6831, 6831, 6830, 6831, 6831, 5781, 6829]), ((2048, (32, 1), 'double', True), [3569, 3573, 3574, 3573, 3574, 3574, 3572, 3573, 3573, 3573]), ((2048, (64, 1), 'single', True), [6756, 6774, 6771, 6765, 6765, 6767, 6766, 6768, 6766, 6767]), ((2048, (64, 1), 'double', True), [3552, 3557, 3557, 3557, 3557, 3557, 3557, 3558, 3557, 3557]), ((2048, (128, 1), 'single', True), [6780, 6794, 6791, 6793, 6791, 6793, 6792, 6791, 6791, 6793]), ((2048, (128, 1), 'double', True), [3551, 3554, 3554, 3553, 3552, 3554, 3554, 3554, 3555, 3555]), ((2048, (256, 1), 'single', True), [6788, 6794, 6795, 6795, 6795, 6795, 6796, 6796, 6796, 6797]), ((2048, (256, 1), 'double', True), [3565, 3562, 3560, 3559, 3559, 3560, 3560, 3560, 3560, 3559]), ((2048, (512, 1), 'single', True), [6795, 6803, 6802, 6804, 6805, 6803, 6804, 6803, 6804, 6803]), ((2048, (512, 1), 'double', True), [3559, 3563, 3563, 3564, 3564, 3564, 3564, 3563, 3564, 3564]), ((2048, (1024, 1), 'single', True), [6740, 6745, 6746, 6747, 6748, 6745, 6734, 6746, 6745, 6746]), ((2048, (1024, 1), 'double', True), [3518, 3522, 3522, 3522, 3522, 3522, 3522, 3522, 3522, 3522]), ((4096, (16, 1), 'single', True), [6564, 6583, 6583, 6583, 6584, 6583, 6583, 6583, 6584, 6583]), ((4096, (16, 1), 'double', True), [3444, 3442, 3442, 3442, 3442, 3442, 3442, 3442, 3442, 3442]), ((4096, (32, 1), 'single', True), [6972, 6963, 6963, 6963, 6963, 6963, 6963, 6961, 6959, 6963]), ((4096, (32, 1), 'double', True), [3462, 3459, 3460, 3460, 3460, 3460, 3459, 3459, 3459, 3459]), ((4096, (64, 1), 'single', True), [6897, 6902, 6901, 6901, 6902, 6901, 6900, 6901, 6902, 6901]), ((4096, (64, 1), 'double', True), [3437, 3438, 3438, 3438, 3438, 3438, 3438, 3438, 3438, 3438]), ((4096, (128, 1), 'single', True), [6925, 6923, 6923, 6923, 6923, 6923, 6923, 6923, 6924, 6923]), ((4096, (128, 1), 'double', True), [3421, 3420, 3419, 3420, 3420, 3420, 3420, 3420, 3420, 3420]), ((4096, (256, 1), 'single', True), [6896, 6881, 6881, 6882, 6882, 6881, 6882, 6882, 6881, 6881]), ((4096, (256, 1), 'double', True), [3433, 3433, 3433, 3434, 3433, 3433, 3433, 3433, 3433, 3433]), ((4096, (512, 1), 'single', True), [6900, 6890, 6890, 6889, 6889, 6890, 6890, 6890, 6890, 6889]), ((4096, (512, 1), 'double', True), [3444, 3442, 3442, 3442, 3442, 3442, 3442, 3442, 3442, 3442]), ((4096, (1024, 1), 'single', True), [6872, 6861, 6861, 6861, 6861, 6860, 6860, 6861, 6861, 6861]), ((4096, (1024, 1), 'double', True), [3410, 3410, 3410, 3411, 3411, 3411, 3411, 3411, 3411, 3410])] diff --git a/result/ldc_2d_benchmark.P100.log b/result/ldc_2d_benchmark.P100.log new file mode 100644 index 0000000..3bacbe6 --- /dev/null +++ b/result/ldc_2d_benchmark.P100.log @@ -0,0 +1,156 @@ +(32, (16, 1), 'single', True): ~14 MLUPS +(32, (16, 1), 'single', False): ~15 MLUPS +(32, (16, 1), 'double', True): ~15 MLUPS +(32, (16, 1), 'double', False): ~15 MLUPS +(32, (32, 1), 'single', True): ~14 MLUPS +(32, (32, 1), 'single', False): ~15 MLUPS +(32, (32, 1), 'double', True): ~15 MLUPS +(32, (32, 1), 'double', False): ~15 MLUPS +(64, (16, 1), 'single', True): ~58 MLUPS +(64, (16, 1), 'single', False): ~59 MLUPS +(64, (16, 1), 'double', True): ~59 MLUPS +(64, (16, 1), 'double', False): ~56 MLUPS +(64, (32, 1), 'single', True): ~60 MLUPS +(64, (32, 1), 'single', False): ~59 MLUPS +(64, (32, 1), 'double', True): ~59 MLUPS +(64, (32, 1), 'double', False): ~59 MLUPS +(64, (64, 1), 'single', True): ~60 MLUPS +(64, (64, 1), 'single', False): ~58 MLUPS +(64, (64, 1), 'double', True): ~59 MLUPS +(64, (64, 1), 'double', False): ~59 MLUPS +(96, (16, 1), 'single', True): ~133 MLUPS +(96, (16, 1), 'single', False): ~132 MLUPS +(96, (16, 1), 'double', True): ~133 MLUPS +(96, (16, 1), 'double', False): ~126 MLUPS +(96, (24, 1), 'single', True): ~129 MLUPS +(96, (24, 1), 'single', False): ~133 MLUPS +(96, (24, 1), 'double', True): ~134 MLUPS +(96, (24, 1), 'double', False): ~133 MLUPS +(96, (32, 1), 'single', True): ~134 MLUPS +(96, (32, 1), 'single', False): ~133 MLUPS +(96, (32, 1), 'double', True): ~134 MLUPS +(96, (32, 1), 'double', False): ~134 MLUPS +(96, (48, 1), 'single', True): ~133 MLUPS +(96, (48, 1), 'single', False): ~135 MLUPS +(96, (48, 1), 'double', True): ~135 MLUPS +(96, (48, 1), 'double', False): ~133 MLUPS +(96, (96, 1), 'single', True): ~136 MLUPS +(96, (96, 1), 'single', False): ~135 MLUPS +(96, (96, 1), 'double', True): ~136 MLUPS +(96, (96, 1), 'double', False): ~136 MLUPS +(128, (16, 1), 'single', True): ~240 MLUPS +(128, (16, 1), 'single', False): ~220 MLUPS +(128, (16, 1), 'double', True): ~222 MLUPS +(128, (16, 1), 'double', False): ~204 MLUPS +(128, (32, 1), 'single', True): ~218 MLUPS +(128, (32, 1), 'single', False): ~237 MLUPS +(128, (32, 1), 'double', True): ~240 MLUPS +(128, (32, 1), 'double', False): ~242 MLUPS +(128, (64, 1), 'single', True): ~241 MLUPS +(128, (64, 1), 'single', False): ~240 MLUPS +(128, (64, 1), 'double', True): ~241 MLUPS +(128, (64, 1), 'double', False): ~237 MLUPS +(128, (128, 1), 'single', True): ~239 MLUPS +(128, (128, 1), 'single', False): ~238 MLUPS +(128, (128, 1), 'double', True): ~238 MLUPS +(128, (128, 1), 'double', False): ~238 MLUPS +(256, (16, 1), 'single', True): ~953 MLUPS +(256, (16, 1), 'single', False): ~932 MLUPS +(256, (16, 1), 'double', True): ~957 MLUPS +(256, (16, 1), 'double', False): ~906 MLUPS +(256, (32, 1), 'single', True): ~972 MLUPS +(256, (32, 1), 'single', False): ~963 MLUPS +(256, (32, 1), 'double', True): ~965 MLUPS +(256, (32, 1), 'double', False): ~959 MLUPS +(256, (64, 1), 'single', True): ~960 MLUPS +(256, (64, 1), 'single', False): ~968 MLUPS +(256, (64, 1), 'double', True): ~967 MLUPS +(256, (64, 1), 'double', False): ~960 MLUPS +(256, (128, 1), 'single', True): ~955 MLUPS +(256, (128, 1), 'single', False): ~919 MLUPS +(256, (128, 1), 'double', True): ~960 MLUPS +(256, (128, 1), 'double', False): ~961 MLUPS +(256, (256, 1), 'single', True): ~965 MLUPS +(256, (256, 1), 'single', False): ~966 MLUPS +(256, (256, 1), 'double', True): ~970 MLUPS +(256, (256, 1), 'double', False): ~957 MLUPS +(512, (16, 1), 'single', True): ~3806 MLUPS +(512, (16, 1), 'single', False): ~2918 MLUPS +(512, (16, 1), 'double', True): ~3272 MLUPS +(512, (16, 1), 'double', False): ~1233 MLUPS +(512, (32, 1), 'single', True): ~3786 MLUPS +(512, (32, 1), 'single', False): ~3681 MLUPS +(512, (32, 1), 'double', True): ~3284 MLUPS +(512, (32, 1), 'double', False): ~2300 MLUPS +(512, (64, 1), 'single', True): ~3799 MLUPS +(512, (64, 1), 'single', False): ~3806 MLUPS +(512, (64, 1), 'double', True): ~3299 MLUPS +(512, (64, 1), 'double', False): ~2287 MLUPS +(512, (128, 1), 'single', True): ~3769 MLUPS +(512, (128, 1), 'single', False): ~3408 MLUPS +(512, (128, 1), 'double', True): ~3295 MLUPS +(512, (128, 1), 'double', False): ~2280 MLUPS +(512, (256, 1), 'single', True): ~3764 MLUPS +(512, (256, 1), 'single', False): ~3889 MLUPS +(512, (256, 1), 'double', True): ~3288 MLUPS +(512, (256, 1), 'double', False): ~2293 MLUPS +(512, (512, 1), 'single', True): ~3830 MLUPS +(512, (512, 1), 'single', False): ~3812 MLUPS +(512, (512, 1), 'double', True): ~3290 MLUPS +(512, (512, 1), 'double', False): ~2123 MLUPS +(1024, (16, 1), 'single', True): ~6332 MLUPS +(1024, (16, 1), 'single', False): ~3809 MLUPS +(1024, (16, 1), 'double', True): ~3473 MLUPS +(1024, (16, 1), 'double', False): ~1291 MLUPS +(1024, (32, 1), 'single', True): ~6641 MLUPS +(1024, (32, 1), 'single', False): ~6496 MLUPS +(1024, (32, 1), 'double', True): ~3470 MLUPS +(1024, (32, 1), 'double', False): ~2525 MLUPS +(1024, (64, 1), 'single', True): ~6588 MLUPS +(1024, (64, 1), 'single', False): ~6436 MLUPS +(1024, (64, 1), 'double', True): ~3450 MLUPS +(1024, (64, 1), 'double', False): ~2521 MLUPS +(1024, (128, 1), 'single', True): ~6605 MLUPS +(1024, (128, 1), 'single', False): ~6427 MLUPS +(1024, (128, 1), 'double', True): ~3449 MLUPS +(1024, (128, 1), 'double', False): ~2524 MLUPS +(1024, (256, 1), 'single', True): ~6602 MLUPS +(1024, (256, 1), 'single', False): ~6355 MLUPS +(1024, (256, 1), 'double', True): ~3456 MLUPS +(1024, (256, 1), 'double', False): ~2573 MLUPS +(1024, (512, 1), 'single', True): ~6594 MLUPS +(1024, (512, 1), 'single', False): ~6143 MLUPS +(1024, (512, 1), 'double', True): ~3464 MLUPS +(1024, (512, 1), 'double', False): ~2428 MLUPS +(1024, (1024, 1), 'single', True): ~6547 MLUPS +(1024, (1024, 1), 'single', False): ~5615 MLUPS +(1024, (1024, 1), 'double', True): ~3420 MLUPS +(1024, (1024, 1), 'double', False): ~2256 MLUPS +(2048, (16, 1), 'single', True): ~6513 MLUPS +(2048, (16, 1), 'double', True): ~3555 MLUPS +(2048, (32, 1), 'single', True): ~6723 MLUPS +(2048, (32, 1), 'double', True): ~3572 MLUPS +(2048, (64, 1), 'single', True): ~6766 MLUPS +(2048, (64, 1), 'double', True): ~3556 MLUPS +(2048, (128, 1), 'single', True): ~6790 MLUPS +(2048, (128, 1), 'double', True): ~3553 MLUPS +(2048, (256, 1), 'single', True): ~6794 MLUPS +(2048, (256, 1), 'double', True): ~3560 MLUPS +(2048, (512, 1), 'single', True): ~6802 MLUPS +(2048, (512, 1), 'double', True): ~3563 MLUPS +(2048, (1024, 1), 'single', True): ~6744 MLUPS +(2048, (1024, 1), 'double', True): ~3521 MLUPS +(4096, (16, 1), 'single', True): ~6636 MLUPS +(4096, (16, 1), 'double', True): ~3432 MLUPS +(4096, (32, 1), 'single', True): ~6910 MLUPS +(4096, (32, 1), 'double', True): ~3440 MLUPS +(4096, (64, 1), 'single', True): ~6852 MLUPS +(4096, (64, 1), 'double', True): ~3415 MLUPS +(4096, (128, 1), 'single', True): ~6875 MLUPS +(4096, (128, 1), 'double', True): ~3420 MLUPS +(4096, (256, 1), 'single', True): ~6882 MLUPS +(4096, (256, 1), 'double', True): ~3432 MLUPS +(4096, (512, 1), 'single', True): ~6890 MLUPS +(4096, (512, 1), 'double', True): ~3442 MLUPS +(4096, (1024, 1), 'single', True): ~6862 MLUPS +(4096, (1024, 1), 'double', True): ~3410 MLUPS diff --git a/result/ldc_3d_benchmark.k2200.data b/result/ldc_3d_benchmark.K2200.data index 95150a7..95150a7 100644 --- a/result/ldc_3d_benchmark.k2200.data +++ b/result/ldc_3d_benchmark.K2200.data diff --git a/result/ldc_3d_benchmark.k2200.log b/result/ldc_3d_benchmark.K2200.log index 7626f87..7626f87 100644 --- a/result/ldc_3d_benchmark.k2200.log +++ b/result/ldc_3d_benchmark.K2200.log diff --git a/result/ldc_3d_benchmark.P100.data b/result/ldc_3d_benchmark.P100.data new file mode 100644 index 0000000..ca634eb --- /dev/null +++ b/result/ldc_3d_benchmark.P100.data @@ -0,0 +1 @@ +[((16, (16, 1, 1), 'symbolic.D3Q19', 'single', True), [59, 61, 61, 61, 61, 61, 61, 61, 61, 61]), ((16, (16, 1, 1), 'symbolic.D3Q19', 'single', False), [57, 58, 58, 57, 58, 58, 58, 58, 58, 58]), ((16, (16, 1, 1), 'symbolic.D3Q19', 'double', True), [59, 60, 60, 60, 61, 60, 60, 61, 61, 60]), ((16, (16, 1, 1), 'symbolic.D3Q19', 'double', False), [56, 57, 57, 58, 57, 58, 58, 58, 57, 58]), ((16, (16, 1, 1), 'symbolic.D3Q27', 'single', True), [60, 60, 59, 60, 60, 61, 61, 60, 61, 60]), ((16, (16, 1, 1), 'symbolic.D3Q27', 'single', False), [57, 58, 58, 57, 58, 58, 58, 58, 58, 58]), ((16, (16, 1, 1), 'symbolic.D3Q27', 'double', True), [59, 61, 61, 61, 61, 61, 61, 61, 61, 60]), ((16, (16, 1, 1), 'symbolic.D3Q27', 'double', False), [56, 57, 57, 57, 58, 57, 57, 57, 57, 57]), ((32, (16, 1, 1), 'symbolic.D3Q19', 'single', True), [473, 477, 482, 482, 481, 484, 485, 485, 486, 483]), ((32, (16, 1, 1), 'symbolic.D3Q19', 'single', False), [453, 461, 460, 462, 462, 465, 463, 466, 464, 464]), ((32, (16, 1, 1), 'symbolic.D3Q19', 'double', True), [475, 480, 480, 475, 484, 484, 483, 483, 483, 483]), ((32, (16, 1, 1), 'symbolic.D3Q19', 'double', False), [410, 413, 414, 413, 414, 413, 415, 413, 413, 412]), ((32, (16, 1, 1), 'symbolic.D3Q27', 'single', True), [474, 481, 479, 473, 483, 486, 485, 480, 485, 483]), ((32, (16, 1, 1), 'symbolic.D3Q27', 'single', False), [448, 449, 452, 457, 457, 450, 452, 456, 461, 461]), ((32, (16, 1, 1), 'symbolic.D3Q27', 'double', True), [467, 478, 476, 481, 480, 481, 479, 481, 473, 477]), ((32, (16, 1, 1), 'symbolic.D3Q27', 'double', False), [265, 265, 265, 265, 266, 267, 267, 268, 267, 268]), ((32, (32, 1, 1), 'symbolic.D3Q19', 'single', True), [479, 478, 475, 479, 484, 486, 485, 481, 485, 486]), ((32, (32, 1, 1), 'symbolic.D3Q19', 'single', False), [464, 477, 477, 475, 472, 479, 481, 482, 476, 479]), ((32, (32, 1, 1), 'symbolic.D3Q19', 'double', True), [475, 481, 480, 481, 483, 482, 477, 464, 481, 474]), ((32, (32, 1, 1), 'symbolic.D3Q19', 'double', False), [469, 473, 471, 477, 479, 474, 478, 476, 471, 477]), ((32, (32, 1, 1), 'symbolic.D3Q27', 'single', True), [476, 481, 471, 479, 484, 481, 481, 479, 472, 484]), ((32, (32, 1, 1), 'symbolic.D3Q27', 'single', False), [478, 479, 482, 484, 484, 483, 485, 483, 482, 485]), ((32, (32, 1, 1), 'symbolic.D3Q27', 'double', True), [471, 483, 477, 484, 481, 483, 479, 479, 480, 483]), ((32, (32, 1, 1), 'symbolic.D3Q27', 'double', False), [423, 426, 427, 426, 426, 427, 427, 427, 426, 427]), ((48, (16, 1, 1), 'symbolic.D3Q19', 'single', True), [1579, 1625, 1609, 1620, 1602, 1620, 1607, 1613, 1614, 1610]), ((48, (16, 1, 1), 'symbolic.D3Q19', 'single', False), [1250, 1251, 1251, 1250, 1250, 1249, 1251, 1250, 1250, 1252]), ((48, (16, 1, 1), 'symbolic.D3Q19', 'double', True), [1418, 1428, 1431, 1427, 1432, 1428, 1427, 1430, 1431, 1432]), ((48, (16, 1, 1), 'symbolic.D3Q19', 'double', False), [514, 511, 510, 510, 511, 510, 511, 510, 511, 511]), ((48, (16, 1, 1), 'symbolic.D3Q27', 'single', True), [1585, 1603, 1605, 1609, 1603, 1590, 1605, 1599, 1605, 1604]), ((48, (16, 1, 1), 'symbolic.D3Q27', 'single', False), [776, 776, 777, 777, 777, 776, 777, 778, 778, 775]), ((48, (16, 1, 1), 'symbolic.D3Q27', 'double', True), [1007, 1009, 1010, 1011, 1008, 1013, 1011, 1011, 1010, 1012]), ((48, (16, 1, 1), 'symbolic.D3Q27', 'double', False), [322, 323, 323, 323, 323, 323, 323, 323, 323, 323]), ((48, (24, 1, 1), 'symbolic.D3Q19', 'single', True), [1598, 1606, 1618, 1602, 1615, 1610, 1590, 1622, 1610, 1624]), ((48, (24, 1, 1), 'symbolic.D3Q19', 'single', False), [1574, 1605, 1597, 1608, 1609, 1612, 1594, 1608, 1612, 1617]), ((48, (24, 1, 1), 'symbolic.D3Q19', 'double', True), [1392, 1403, 1405, 1402, 1404, 1404, 1405, 1404, 1402, 1403]), ((48, (24, 1, 1), 'symbolic.D3Q19', 'double', False), [718, 718, 718, 719, 719, 717, 719, 718, 718, 717]), ((48, (24, 1, 1), 'symbolic.D3Q27', 'single', True), [1624, 1623, 1645, 1604, 1640, 1599, 1648, 1649, 1651, 1660]), ((48, (24, 1, 1), 'symbolic.D3Q27', 'single', False), [1027, 1030, 1032, 1030, 1032, 1031, 1030, 1032, 1032, 1032]), ((48, (24, 1, 1), 'symbolic.D3Q27', 'double', True), [1003, 1008, 1008, 1008, 1006, 1007, 1007, 1006, 1007, 1007]), ((48, (24, 1, 1), 'symbolic.D3Q27', 'double', False), [453, 454, 455, 455, 455, 455, 455, 455, 455, 455]), ((48, (48, 1, 1), 'symbolic.D3Q19', 'single', True), [1606, 1646, 1589, 1626, 1623, 1632, 1597, 1631, 1617, 1645]), ((48, (48, 1, 1), 'symbolic.D3Q19', 'single', False), [1600, 1627, 1643, 1634, 1640, 1637, 1643, 1642, 1643, 1643]), ((48, (48, 1, 1), 'symbolic.D3Q19', 'double', True), [1412, 1420, 1417, 1419, 1420, 1417, 1416, 1419, 1421, 1417]), ((48, (48, 1, 1), 'symbolic.D3Q19', 'double', False), [724, 724, 726, 725, 725, 724, 725, 724, 724, 726]), ((48, (48, 1, 1), 'symbolic.D3Q27', 'single', True), [1589, 1613, 1619, 1620, 1607, 1622, 1625, 1628, 1623, 1625]), ((48, (48, 1, 1), 'symbolic.D3Q27', 'single', False), [1046, 1046, 1049, 1045, 1047, 1046, 1048, 1047, 1047, 1045]), ((48, (48, 1, 1), 'symbolic.D3Q27', 'double', True), [1018, 1023, 1023, 1022, 1023, 1020, 1024, 1023, 1022, 1024]), ((48, (48, 1, 1), 'symbolic.D3Q27', 'double', False), [462, 466, 464, 465, 466, 465, 465, 465, 466, 465]), ((64, (16, 1, 1), 'symbolic.D3Q19', 'single', True), [2652, 2669, 2671, 2671, 2672, 2668, 2671, 2672, 2671, 2667]), ((64, (16, 1, 1), 'symbolic.D3Q19', 'single', False), [1343, 1325, 1326, 1326, 1325, 1327, 1326, 1326, 1325, 1326]), ((64, (16, 1, 1), 'symbolic.D3Q19', 'double', True), [1555, 1559, 1559, 1558, 1558, 1559, 1560, 1559, 1558, 1559]), ((64, (16, 1, 1), 'symbolic.D3Q19', 'double', False), [525, 517, 517, 517, 517, 517, 517, 517, 517, 517]), ((64, (16, 1, 1), 'symbolic.D3Q27', 'single', True), [1790, 1795, 1792, 1796, 1795, 1794, 1794, 1793, 1794, 1796]), ((64, (16, 1, 1), 'symbolic.D3Q27', 'single', False), [821, 821, 821, 821, 821, 821, 821, 821, 821, 821]), ((64, (16, 1, 1), 'symbolic.D3Q27', 'double', True), [1115, 1117, 1117, 1117, 1117, 1117, 1117, 1117, 1117, 1118]), ((64, (16, 1, 1), 'symbolic.D3Q27', 'double', False), [327, 328, 328, 328, 328, 328, 328, 328, 328, 328]), ((64, (32, 1, 1), 'symbolic.D3Q19', 'single', True), [2808, 2819, 2819, 2819, 2816, 2818, 2819, 2822, 2816, 2816]), ((64, (32, 1, 1), 'symbolic.D3Q19', 'single', False), [2296, 2296, 2297, 2293, 2296, 2296, 2295, 2294, 2293, 2294]), ((64, (32, 1, 1), 'symbolic.D3Q19', 'double', True), [1556, 1561, 1562, 1562, 1562, 1561, 1560, 1561, 1561, 1560]), ((64, (32, 1, 1), 'symbolic.D3Q19', 'double', False), [965, 967, 968, 969, 968, 968, 967, 967, 967, 967]), ((64, (32, 1, 1), 'symbolic.D3Q27', 'single', True), [1897, 1900, 1905, 1899, 1905, 1906, 1904, 1903, 1904, 1903]), ((64, (32, 1, 1), 'symbolic.D3Q27', 'single', False), [1412, 1414, 1416, 1416, 1414, 1414, 1416, 1415, 1414, 1416]), ((64, (32, 1, 1), 'symbolic.D3Q27', 'double', True), [1110, 1113, 1114, 1114, 1113, 1114, 1114, 1114, 1115, 1113]), ((64, (32, 1, 1), 'symbolic.D3Q27', 'double', False), [607, 610, 611, 610, 610, 610, 611, 611, 611, 610]), ((64, (64, 1, 1), 'symbolic.D3Q19', 'single', True), [2726, 2739, 2740, 2735, 2740, 2739, 2736, 2740, 2742, 2737]), ((64, (64, 1, 1), 'symbolic.D3Q19', 'single', False), [2313, 2314, 2316, 2315, 2316, 2315, 2313, 2316, 2315, 2315]), ((64, (64, 1, 1), 'symbolic.D3Q19', 'double', True), [1558, 1562, 1562, 1562, 1561, 1562, 1560, 1561, 1562, 1561]), ((64, (64, 1, 1), 'symbolic.D3Q19', 'double', False), [969, 971, 971, 971, 971, 971, 971, 971, 971, 971]), ((64, (64, 1, 1), 'symbolic.D3Q27', 'single', True), [1881, 1883, 1882, 1884, 1884, 1881, 1885, 1884, 1884, 1883]), ((64, (64, 1, 1), 'symbolic.D3Q27', 'single', False), [1430, 1433, 1430, 1434, 1431, 1429, 1435, 1433, 1432, 1431]), ((64, (64, 1, 1), 'symbolic.D3Q27', 'double', True), [1114, 1116, 1117, 1117, 1117, 1117, 1117, 1116, 1116, 1117]), ((64, (64, 1, 1), 'symbolic.D3Q27', 'double', False), [618, 622, 622, 622, 622, 622, 622, 622, 623, 623]), ((96, (16, 1, 1), 'symbolic.D3Q19', 'single', True), [2512, 2515, 2515, 2516, 2516, 2516, 2516, 2515, 2514, 2515]), ((96, (16, 1, 1), 'symbolic.D3Q19', 'single', False), [1381, 1339, 1339, 1340, 1340, 1340, 1340, 1340, 1340, 1340]), ((96, (16, 1, 1), 'symbolic.D3Q19', 'double', True), [1536, 1538, 1538, 1539, 1538, 1539, 1539, 1539, 1538, 1539]), ((96, (16, 1, 1), 'symbolic.D3Q19', 'double', False), [526, 511, 511, 511, 511, 511, 511, 511, 511, 511]), ((96, (16, 1, 1), 'symbolic.D3Q27', 'single', True), [1720, 1722, 1723, 1723, 1723, 1723, 1723, 1723, 1723, 1722]), ((96, (16, 1, 1), 'symbolic.D3Q27', 'single', False), [838, 837, 837, 837, 837, 837, 837, 837, 837, 837]), ((96, (16, 1, 1), 'symbolic.D3Q27', 'double', True), [1073, 1073, 1074, 1074, 1073, 1073, 1074, 1074, 1074, 1074]), ((96, (16, 1, 1), 'symbolic.D3Q27', 'double', False), [321, 323, 323, 323, 323, 323, 323, 323, 323, 323]), ((96, (24, 1, 1), 'symbolic.D3Q19', 'single', True), [2576, 2578, 2578, 2576, 2578, 2579, 2578, 2580, 2578, 2580]), ((96, (24, 1, 1), 'symbolic.D3Q19', 'single', False), [1990, 1953, 1953, 1953, 1953, 1954, 1954, 1953, 1953, 1953]), ((96, (24, 1, 1), 'symbolic.D3Q19', 'double', True), [1511, 1511, 1512, 1512, 1512, 1511, 1512, 1512, 1511, 1512]), ((96, (24, 1, 1), 'symbolic.D3Q19', 'double', False), [770, 757, 757, 757, 757, 757, 757, 757, 757, 757]), ((96, (24, 1, 1), 'symbolic.D3Q27', 'single', True), [1761, 1763, 1762, 1764, 1764, 1763, 1762, 1763, 1762, 1762]), ((96, (24, 1, 1), 'symbolic.D3Q27', 'single', False), [1218, 1217, 1217, 1218, 1217, 1218, 1217, 1217, 1218, 1218]), ((96, (24, 1, 1), 'symbolic.D3Q27', 'double', True), [1059, 1060, 1060, 1060, 1060, 1060, 1059, 1060, 1060, 1060]), ((96, (24, 1, 1), 'symbolic.D3Q27', 'double', False), [473, 478, 478, 478, 478, 477, 478, 478, 478, 478]), ((96, (32, 1, 1), 'symbolic.D3Q19', 'single', True), [2750, 2752, 2752, 2753, 2752, 2752, 2750, 2755, 2753, 2753]), ((96, (32, 1, 1), 'symbolic.D3Q19', 'single', False), [2512, 2495, 2497, 2496, 2496, 2496, 2497, 2496, 2495, 2496]), ((96, (32, 1, 1), 'symbolic.D3Q19', 'double', True), [1562, 1563, 1564, 1564, 1563, 1563, 1564, 1563, 1564, 1563]), ((96, (32, 1, 1), 'symbolic.D3Q19', 'double', False), [1001, 995, 995, 995, 995, 995, 995, 995, 995, 995]), ((96, (32, 1, 1), 'symbolic.D3Q27', 'single', True), [1899, 1900, 1900, 1901, 1899, 1900, 1900, 1899, 1899, 1900]), ((96, (32, 1, 1), 'symbolic.D3Q27', 'single', False), [1561, 1561, 1561, 1561, 1561, 1561, 1560, 1559, 1561, 1560]), ((96, (32, 1, 1), 'symbolic.D3Q27', 'double', True), [1095, 1095, 1096, 1095, 1095, 1096, 1096, 1096, 1095, 1095]), ((96, (32, 1, 1), 'symbolic.D3Q27', 'double', False), [621, 628, 628, 628, 628, 628, 628, 628, 628, 628]), ((96, (48, 1, 1), 'symbolic.D3Q19', 'single', True), [2534, 2538, 2534, 2537, 2538, 2535, 2535, 2536, 2535, 2536]), ((96, (48, 1, 1), 'symbolic.D3Q19', 'single', False), [1988, 1954, 1954, 1954, 1954, 1954, 1954, 1953, 1954, 1953]), ((96, (48, 1, 1), 'symbolic.D3Q19', 'double', True), [1547, 1548, 1548, 1548, 1548, 1548, 1548, 1548, 1548, 1548]), ((96, (48, 1, 1), 'symbolic.D3Q19', 'double', False), [764, 755, 755, 755, 755, 755, 755, 755, 755, 755]), ((96, (48, 1, 1), 'symbolic.D3Q27', 'single', True), [1813, 1814, 1815, 1814, 1814, 1814, 1815, 1815, 1816, 1815]), ((96, (48, 1, 1), 'symbolic.D3Q27', 'single', False), [1216, 1215, 1216, 1216, 1216, 1215, 1215, 1215, 1215, 1216]), ((96, (48, 1, 1), 'symbolic.D3Q27', 'double', True), [1084, 1084, 1084, 1084, 1084, 1084, 1084, 1084, 1085, 1084]), ((96, (48, 1, 1), 'symbolic.D3Q27', 'double', False), [472, 478, 478, 478, 479, 478, 478, 478, 478, 478]), ((96, (96, 1, 1), 'symbolic.D3Q19', 'single', True), [2677, 2683, 2682, 2680, 2682, 2681, 2683, 2683, 2682, 2682]), ((96, (96, 1, 1), 'symbolic.D3Q19', 'single', False), [2508, 2490, 2490, 2491, 2491, 2491, 2490, 2491, 2490, 2486]), ((96, (96, 1, 1), 'symbolic.D3Q19', 'double', True), [1563, 1564, 1564, 1564, 1564, 1564, 1564, 1564, 1564, 1564]), ((96, (96, 1, 1), 'symbolic.D3Q19', 'double', False), [1004, 995, 995, 995, 995, 995, 995, 995, 995, 995]), ((96, (96, 1, 1), 'symbolic.D3Q27', 'single', True), [1900, 1902, 1901, 1901, 1901, 1901, 1901, 1901, 1902, 1901]), ((96, (96, 1, 1), 'symbolic.D3Q27', 'single', False), [1560, 1559, 1560, 1560, 1560, 1560, 1559, 1560, 1560, 1560]), ((96, (96, 1, 1), 'symbolic.D3Q27', 'double', True), [1096, 1097, 1097, 1097, 1097, 1097, 1097, 1097, 1097, 1096]), ((96, (96, 1, 1), 'symbolic.D3Q27', 'double', False), [607, 614, 614, 614, 614, 613, 614, 613, 613, 614]), ((128, (16, 1, 1), 'symbolic.D3Q19', 'single', True), [2804, 2815, 2819, 2820, 2819, 2819, 2820, 2820, 2820, 2821]), ((128, (16, 1, 1), 'symbolic.D3Q19', 'single', False), [1390, 1327, 1328, 1328, 1328, 1328, 1328, 1328, 1328, 1328]), ((128, (16, 1, 1), 'symbolic.D3Q19', 'double', True), [1623, 1625, 1624, 1624, 1624, 1624, 1625, 1625, 1624, 1623]), ((128, (16, 1, 1), 'symbolic.D3Q19', 'double', False), [521, 501, 501, 501, 501, 501, 501, 501, 500, 501]), ((128, (16, 1, 1), 'symbolic.D3Q27', 'single', True), [1950, 1952, 1952, 1952, 1952, 1952, 1952, 1952, 1952, 1952]), ((128, (16, 1, 1), 'symbolic.D3Q27', 'single', False), [827, 825, 826, 826, 826, 826, 826, 826, 826, 826]), ((128, (16, 1, 1), 'symbolic.D3Q27', 'double', True), [1135, 1135, 1135, 1135, 1135, 1135, 1135, 1135, 1135, 1135]), ((128, (16, 1, 1), 'symbolic.D3Q27', 'double', False), [315, 319, 319, 319, 319, 319, 319, 319, 319, 319]), ((128, (32, 1, 1), 'symbolic.D3Q19', 'single', True), [3003, 3000, 2996, 2995, 2996, 2995, 2996, 2995, 2995, 2995]), ((128, (32, 1, 1), 'symbolic.D3Q19', 'single', False), [2603, 2552, 2554, 2554, 2554, 2554, 2554, 2554, 2554, 2553]), ((128, (32, 1, 1), 'symbolic.D3Q19', 'double', True), [1633, 1633, 1633, 1633, 1633, 1633, 1633, 1633, 1633, 1633]), ((128, (32, 1, 1), 'symbolic.D3Q19', 'double', False), [1001, 986, 986, 986, 986, 986, 986, 986, 986, 986]), ((128, (32, 1, 1), 'symbolic.D3Q27', 'single', True), [2095, 2097, 2096, 2096, 2096, 2095, 2096, 2096, 2096, 2096]), ((128, (32, 1, 1), 'symbolic.D3Q27', 'single', False), [1573, 1572, 1572, 1573, 1573, 1573, 1573, 1573, 1573, 1572]), ((128, (32, 1, 1), 'symbolic.D3Q27', 'double', True), [1146, 1146, 1146, 1147, 1147, 1146, 1146, 1147, 1147, 1147]), ((128, (32, 1, 1), 'symbolic.D3Q27', 'double', False), [614, 626, 626, 626, 626, 627, 626, 625, 627, 627]), ((128, (64, 1, 1), 'symbolic.D3Q19', 'single', True), [2921, 2923, 2926, 2929, 2930, 2928, 2929, 2929, 2929, 2929]), ((128, (64, 1, 1), 'symbolic.D3Q19', 'single', False), [2596, 2546, 2548, 2549, 2550, 2549, 2550, 2549, 2549, 2549]), ((128, (64, 1, 1), 'symbolic.D3Q19', 'double', True), [1636, 1636, 1637, 1636, 1636, 1636, 1636, 1636, 1636, 1636]), ((128, (64, 1, 1), 'symbolic.D3Q19', 'double', False), [992, 982, 982, 982, 982, 982, 979, 980, 982, 982]), ((128, (64, 1, 1), 'symbolic.D3Q27', 'single', True), [2095, 2096, 2095, 2096, 2096, 2095, 2096, 2096, 2096, 2096]), ((128, (64, 1, 1), 'symbolic.D3Q27', 'single', False), [1566, 1564, 1564, 1564, 1563, 1564, 1565, 1564, 1564, 1564]), ((128, (64, 1, 1), 'symbolic.D3Q27', 'double', True), [1149, 1149, 1149, 1149, 1149, 1149, 1149, 1149, 1149, 1149]), ((128, (64, 1, 1), 'symbolic.D3Q27', 'double', False), [611, 626, 627, 626, 626, 627, 627, 627, 627, 627]), ((128, (128, 1, 1), 'symbolic.D3Q19', 'single', True), [2943, 2945, 2947, 2946, 2948, 2949, 2949, 2949, 2948, 2946]), ((128, (128, 1, 1), 'symbolic.D3Q19', 'single', False), [2603, 2553, 2554, 2555, 2555, 2555, 2555, 2555, 2555, 2555]), ((128, (128, 1, 1), 'symbolic.D3Q19', 'double', True), [1632, 1633, 1633, 1633, 1633, 1633, 1633, 1633, 1633, 1633]), ((128, (128, 1, 1), 'symbolic.D3Q19', 'double', False), [1030, 1007, 1007, 1007, 1007, 1007, 1007, 1007, 1007, 1007]), ((128, (128, 1, 1), 'symbolic.D3Q27', 'single', True), [2105, 2105, 2105, 2103, 2104, 2105, 2104, 2105, 2105, 2106]), ((128, (128, 1, 1), 'symbolic.D3Q27', 'single', False), [1573, 1571, 1571, 1572, 1572, 1572, 1571, 1572, 1572, 1571]), ((128, (128, 1, 1), 'symbolic.D3Q27', 'double', True), [1151, 1151, 1151, 1151, 1152, 1151, 1151, 1151, 1151, 1151]), ((128, (128, 1, 1), 'symbolic.D3Q27', 'double', False), [627, 652, 653, 653, 653, 653, 653, 653, 653, 653])]
\ No newline at end of file diff --git a/result/ldc_3d_benchmark.P100.log b/result/ldc_3d_benchmark.P100.log new file mode 100644 index 0000000..0f3cb16 --- /dev/null +++ b/result/ldc_3d_benchmark.P100.log @@ -0,0 +1,144 @@ +(16, (16, 1, 1), 'symbolic.D3Q19', 'single', True): ~60 MLUPS +(16, (16, 1, 1), 'symbolic.D3Q19', 'single', False): ~57 MLUPS +(16, (16, 1, 1), 'symbolic.D3Q19', 'double', True): ~60 MLUPS +(16, (16, 1, 1), 'symbolic.D3Q19', 'double', False): ~57 MLUPS +(16, (16, 1, 1), 'symbolic.D3Q27', 'single', True): ~60 MLUPS +(16, (16, 1, 1), 'symbolic.D3Q27', 'single', False): ~57 MLUPS +(16, (16, 1, 1), 'symbolic.D3Q27', 'double', True): ~60 MLUPS +(16, (16, 1, 1), 'symbolic.D3Q27', 'double', False): ~57 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q19', 'single', True): ~481 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q19', 'single', False): ~462 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q19', 'double', True): ~481 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q19', 'double', False): ~413 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q27', 'single', True): ~480 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q27', 'single', False): ~454 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q27', 'double', True): ~477 MLUPS +(32, (16, 1, 1), 'symbolic.D3Q27', 'double', False): ~266 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q19', 'single', True): ~481 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q19', 'single', False): ~476 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q19', 'double', True): ~477 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q19', 'double', False): ~474 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q27', 'single', True): ~478 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q27', 'single', False): ~482 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q27', 'double', True): ~480 MLUPS +(32, (32, 1, 1), 'symbolic.D3Q27', 'double', False): ~426 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q19', 'single', True): ~1609 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q19', 'single', False): ~1250 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q19', 'double', True): ~1428 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q19', 'double', False): ~510 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q27', 'single', True): ~1600 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q27', 'single', False): ~776 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q27', 'double', True): ~1010 MLUPS +(48, (16, 1, 1), 'symbolic.D3Q27', 'double', False): ~322 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q19', 'single', True): ~1609 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q19', 'single', False): ~1603 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q19', 'double', True): ~1402 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q19', 'double', False): ~718 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q27', 'single', True): ~1634 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q27', 'single', False): ~1030 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q27', 'double', True): ~1006 MLUPS +(48, (24, 1, 1), 'symbolic.D3Q27', 'double', False): ~454 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q19', 'single', True): ~1621 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q19', 'single', False): ~1635 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q19', 'double', True): ~1417 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q19', 'double', False): ~724 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q27', 'single', True): ~1617 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q27', 'single', False): ~1046 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q27', 'double', True): ~1022 MLUPS +(48, (48, 1, 1), 'symbolic.D3Q27', 'double', False): ~464 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q19', 'single', True): ~2668 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q19', 'single', False): ~1327 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q19', 'double', True): ~1558 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q19', 'double', False): ~517 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q27', 'single', True): ~1793 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q27', 'single', False): ~821 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q27', 'double', True): ~1116 MLUPS +(64, (16, 1, 1), 'symbolic.D3Q27', 'double', False): ~327 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q19', 'single', True): ~2817 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q19', 'single', False): ~2295 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q19', 'double', True): ~1560 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q19', 'double', False): ~967 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q27', 'single', True): ~1902 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q27', 'single', False): ~1414 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q27', 'double', True): ~1113 MLUPS +(64, (32, 1, 1), 'symbolic.D3Q27', 'double', False): ~610 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q19', 'single', True): ~2737 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q19', 'single', False): ~2314 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q19', 'double', True): ~1561 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q19', 'double', False): ~970 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q27', 'single', True): ~1883 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q27', 'single', False): ~1431 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q27', 'double', True): ~1116 MLUPS +(64, (64, 1, 1), 'symbolic.D3Q27', 'double', False): ~621 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q19', 'single', True): ~2515 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q19', 'single', False): ~1343 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q19', 'double', True): ~1538 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q19', 'double', False): ~512 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q27', 'single', True): ~1722 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q27', 'single', False): ~837 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q27', 'double', True): ~1073 MLUPS +(96, (16, 1, 1), 'symbolic.D3Q27', 'double', False): ~322 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q19', 'single', True): ~2578 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q19', 'single', False): ~1956 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q19', 'double', True): ~1511 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q19', 'double', False): ~758 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q27', 'single', True): ~1762 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q27', 'single', False): ~1217 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q27', 'double', True): ~1059 MLUPS +(96, (24, 1, 1), 'symbolic.D3Q27', 'double', False): ~477 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q19', 'single', True): ~2752 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q19', 'single', False): ~2497 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q19', 'double', True): ~1563 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q19', 'double', False): ~995 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q27', 'single', True): ~1899 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q27', 'single', False): ~1560 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q27', 'double', True): ~1095 MLUPS +(96, (32, 1, 1), 'symbolic.D3Q27', 'double', False): ~627 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q19', 'single', True): ~2535 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q19', 'single', False): ~1957 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q19', 'double', True): ~1547 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q19', 'double', False): ~755 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q27', 'single', True): ~1814 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q27', 'single', False): ~1215 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q27', 'double', True): ~1084 MLUPS +(96, (48, 1, 1), 'symbolic.D3Q27', 'double', False): ~477 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q19', 'single', True): ~2681 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q19', 'single', False): ~2491 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q19', 'double', True): ~1563 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q19', 'double', False): ~995 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q27', 'single', True): ~1901 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q27', 'single', False): ~1559 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q27', 'double', True): ~1096 MLUPS +(96, (96, 1, 1), 'symbolic.D3Q27', 'double', False): ~613 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q19', 'single', True): ~2817 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q19', 'single', False): ~1334 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q19', 'double', True): ~1624 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q19', 'double', False): ~502 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q27', 'single', True): ~1951 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q27', 'single', False): ~826 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q27', 'double', True): ~1135 MLUPS +(128, (16, 1, 1), 'symbolic.D3Q27', 'double', False): ~318 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q19', 'single', True): ~2996 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q19', 'single', False): ~2558 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q19', 'double', True): ~1633 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q19', 'double', False): ~987 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q27', 'single', True): ~2095 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q27', 'single', False): ~1572 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q27', 'double', True): ~1146 MLUPS +(128, (32, 1, 1), 'symbolic.D3Q27', 'double', False): ~625 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q19', 'single', True): ~2927 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q19', 'single', False): ~2553 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q19', 'double', True): ~1636 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q19', 'double', False): ~982 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q27', 'single', True): ~2095 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q27', 'single', False): ~1564 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q27', 'double', True): ~1149 MLUPS +(128, (64, 1, 1), 'symbolic.D3Q27', 'double', False): ~625 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q19', 'single', True): ~2947 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q19', 'single', False): ~2559 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q19', 'double', True): ~1632 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q19', 'double', False): ~1009 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q27', 'single', True): ~2104 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q27', 'single', False): ~1571 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q27', 'double', True): ~1151 MLUPS +(128, (128, 1, 1), 'symbolic.D3Q27', 'double', False): ~650 MLUPS |