Update benchmark results of LDC @ Tesla P100

4 files changed, 616 insertions, 302 deletions

diff --git a/result/ldc_2d_benchmark_P100.log b/result/ldc_2d_benchmark_P100.log
index 3bacbe6..a8cc599 100644
--- a/result/ldc_2d_benchmark_P100.log
+++ b/result/ldc_2d_benchmark_P100.log
@@ -1,156 +1,314 @@
-(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
+(32, (32, 1), 'single', False, True): ~13 MLUPS
+(32, (32, 1), 'single', True, True): ~13 MLUPS
+(32, (32, 1), 'double', False, True): ~13 MLUPS
+(32, (32, 1), 'double', True, True): ~13 MLUPS
+(32, (16, 1), 'single', False, True): ~13 MLUPS
+(32, (16, 1), 'single', True, True): ~13 MLUPS
+(32, (16, 1), 'double', False, True): ~13 MLUPS
+(32, (16, 1), 'double', True, True): ~13 MLUPS
+(256, (256, 1), 'single', False, True): ~852 MLUPS
+(256, (256, 1), 'single', True, True): ~862 MLUPS
+(256, (256, 1), 'double', False, True): ~847 MLUPS
+(256, (256, 1), 'double', True, True): ~856 MLUPS
+(256, (32, 1), 'single', False, True): ~862 MLUPS
+(256, (32, 1), 'single', True, True): ~859 MLUPS
+(256, (32, 1), 'double', False, True): ~851 MLUPS
+(256, (32, 1), 'double', True, True): ~848 MLUPS
+(256, (64, 1), 'single', False, True): ~813 MLUPS
+(256, (64, 1), 'single', True, True): ~856 MLUPS
+(256, (64, 1), 'double', False, True): ~849 MLUPS
+(256, (64, 1), 'double', True, True): ~857 MLUPS
+(256, (128, 1), 'single', False, True): ~952 MLUPS
+(256, (128, 1), 'single', True, True): ~953 MLUPS
+(256, (128, 1), 'double', False, True): ~938 MLUPS
+(256, (128, 1), 'double', True, True): ~964 MLUPS
+(256, (16, 1), 'single', False, True): ~950 MLUPS
+(256, (16, 1), 'single', True, True): ~939 MLUPS
+(256, (16, 1), 'double', False, True): ~935 MLUPS
+(256, (16, 1), 'double', True, True): ~913 MLUPS
+(64, (32, 1), 'single', False, True): ~56 MLUPS
+(64, (32, 1), 'single', True, True): ~59 MLUPS
+(64, (32, 1), 'double', False, True): ~58 MLUPS
+(64, (32, 1), 'double', True, True): ~58 MLUPS
+(64, (64, 1), 'single', False, True): ~58 MLUPS
+(64, (64, 1), 'single', True, True): ~59 MLUPS
+(64, (64, 1), 'double', False, True): ~57 MLUPS
+(64, (64, 1), 'double', True, True): ~59 MLUPS
+(64, (16, 1), 'single', False, True): ~59 MLUPS
+(64, (16, 1), 'single', True, True): ~59 MLUPS
+(64, (16, 1), 'double', False, True): ~59 MLUPS
+(64, (16, 1), 'double', True, True): ~58 MLUPS
+(128, (32, 1), 'single', False, True): ~213 MLUPS
+(128, (32, 1), 'single', True, True): ~215 MLUPS
+(128, (32, 1), 'double', False, True): ~213 MLUPS
+(128, (32, 1), 'double', True, True): ~215 MLUPS
+(128, (64, 1), 'single', False, True): ~215 MLUPS
+(128, (64, 1), 'single', True, True): ~215 MLUPS
+(128, (64, 1), 'double', False, True): ~215 MLUPS
+(128, (64, 1), 'double', True, True): ~217 MLUPS
+(128, (128, 1), 'single', False, True): ~216 MLUPS
+(128, (128, 1), 'single', True, True): ~215 MLUPS
+(128, (128, 1), 'double', False, True): ~213 MLUPS
+(128, (128, 1), 'double', True, True): ~214 MLUPS
+(128, (16, 1), 'single', False, True): ~214 MLUPS
+(128, (16, 1), 'single', True, True): ~215 MLUPS
+(128, (16, 1), 'double', False, True): ~213 MLUPS
+(128, (16, 1), 'double', True, True): ~216 MLUPS
+(512, (256, 1), 'single', False, True): ~3376 MLUPS
+(512, (256, 1), 'single', True, True): ~3407 MLUPS
+(512, (256, 1), 'double', False, True): ~2145 MLUPS
+(512, (256, 1), 'double', True, True): ~3339 MLUPS
+(512, (32, 1), 'single', False, True): ~3418 MLUPS
+(512, (32, 1), 'single', True, True): ~3402 MLUPS
+(512, (32, 1), 'double', False, True): ~2126 MLUPS
+(512, (32, 1), 'double', True, True): ~3359 MLUPS
+(512, (64, 1), 'single', False, True): ~3628 MLUPS
+(512, (64, 1), 'single', True, True): ~3414 MLUPS
+(512, (64, 1), 'double', False, True): ~2102 MLUPS
+(512, (64, 1), 'double', True, True): ~3351 MLUPS
+(512, (128, 1), 'single', False, True): ~3392 MLUPS
+(512, (128, 1), 'single', True, True): ~3410 MLUPS
+(512, (128, 1), 'double', False, True): ~1960 MLUPS
+(512, (128, 1), 'double', True, True): ~3325 MLUPS
+(512, (512, 1), 'single', False, True): ~3234 MLUPS
+(512, (512, 1), 'single', True, True): ~3408 MLUPS
+(512, (512, 1), 'double', False, True): ~2029 MLUPS
+(512, (512, 1), 'double', True, True): ~3331 MLUPS
+(512, (16, 1), 'single', False, True): ~3183 MLUPS
+(512, (16, 1), 'single', True, True): ~3441 MLUPS
+(512, (16, 1), 'double', False, True): ~1137 MLUPS
+(512, (16, 1), 'double', True, True): ~3305 MLUPS
+(1024, (256, 1), 'single', False, True): ~6302 MLUPS
+(1024, (256, 1), 'single', True, True): ~6621 MLUPS
+(1024, (256, 1), 'double', False, True): ~2637 MLUPS
+(1024, (256, 1), 'double', True, True): ~3481 MLUPS
+(1024, (32, 1), 'single', False, True): ~6603 MLUPS
+(1024, (32, 1), 'single', True, True): ~6728 MLUPS
+(1024, (32, 1), 'double', False, True): ~2608 MLUPS
+(1024, (32, 1), 'double', True, True): ~3491 MLUPS
+(1024, (64, 1), 'single', False, True): ~6552 MLUPS
+(1024, (64, 1), 'single', True, True): ~6699 MLUPS
+(1024, (64, 1), 'double', False, True): ~2604 MLUPS
+(1024, (64, 1), 'double', True, True): ~3485 MLUPS
+(1024, (128, 1), 'single', False, True): ~6478 MLUPS
+(1024, (128, 1), 'single', True, True): ~6696 MLUPS
+(1024, (128, 1), 'double', False, True): ~2606 MLUPS
+(1024, (128, 1), 'double', True, True): ~3459 MLUPS
+(1024, (512, 1), 'single', False, True): ~6184 MLUPS
+(1024, (512, 1), 'single', True, True): ~6712 MLUPS
+(1024, (512, 1), 'double', False, True): ~2573 MLUPS
+(1024, (512, 1), 'double', True, True): ~3492 MLUPS
+(1024, (1024, 1), 'single', False, True): ~5705 MLUPS
+(1024, (1024, 1), 'single', True, True): ~6647 MLUPS
+(1024, (1024, 1), 'double', False, True): ~2338 MLUPS
+(1024, (1024, 1), 'double', True, True): ~3347 MLUPS
+(1024, (16, 1), 'single', False, True): ~3873 MLUPS
+(1024, (16, 1), 'single', True, True): ~6460 MLUPS
+(1024, (16, 1), 'double', False, True): ~1333 MLUPS
+(1024, (16, 1), 'double', True, True): ~3443 MLUPS
+(2048, (256, 1), 'single', False, True): ~6762 MLUPS
+(2048, (256, 1), 'single', True, True): ~6884 MLUPS
+(2048, (256, 1), 'double', False, True): ~2814 MLUPS
+(2048, (256, 1), 'double', True, True): ~3570 MLUPS
+(2048, (32, 1), 'single', False, True): ~6957 MLUPS
+(2048, (32, 1), 'single', True, True): ~6922 MLUPS
+(2048, (32, 1), 'double', False, True): ~2714 MLUPS
+(2048, (32, 1), 'double', True, True): ~3585 MLUPS
+(2048, (64, 1), 'single', False, True): ~6865 MLUPS
+(2048, (64, 1), 'single', True, True): ~6833 MLUPS
+(2048, (64, 1), 'double', False, True): ~2644 MLUPS
+(2048, (64, 1), 'double', True, True): ~3570 MLUPS
+(2048, (128, 1), 'single', False, True): ~6809 MLUPS
+(2048, (128, 1), 'single', True, True): ~6889 MLUPS
+(2048, (128, 1), 'double', False, True): ~2719 MLUPS
+(2048, (128, 1), 'double', True, True): ~3570 MLUPS
+(2048, (512, 1), 'single', False, True): ~6585 MLUPS
+(2048, (512, 1), 'single', True, True): ~6832 MLUPS
+(2048, (512, 1), 'double', False, True): ~2732 MLUPS
+(2048, (512, 1), 'double', True, True): ~3573 MLUPS
+(2048, (1024, 1), 'single', False, True): ~6269 MLUPS
+(2048, (1024, 1), 'single', True, True): ~6811 MLUPS
+(2048, (1024, 1), 'double', False, True): ~2467 MLUPS
+(2048, (1024, 1), 'double', True, True): ~3534 MLUPS
+(2048, (16, 1), 'single', False, True): ~3905 MLUPS
+(2048, (16, 1), 'single', True, True): ~6693 MLUPS
+(2048, (16, 1), 'double', False, True): ~1367 MLUPS
+(2048, (16, 1), 'double', True, True): ~3578 MLUPS
+(800, (200, 1), 'single', True, False): ~6033 MLUPS
+(800, (200, 1), 'single', True, True): ~6013 MLUPS
+(800, (200, 1), 'double', True, False): ~3216 MLUPS
+(800, (200, 1), 'double', True, True): ~3273 MLUPS
+(800, (10, 1), 'single', True, False): ~4696 MLUPS
+(800, (10, 1), 'single', True, True): ~4724 MLUPS
+(800, (10, 1), 'double', True, False): ~3052 MLUPS
+(800, (10, 1), 'double', True, True): ~3031 MLUPS
+(800, (100, 1), 'single', True, False): ~5990 MLUPS
+(800, (100, 1), 'single', True, True): ~5927 MLUPS
+(800, (100, 1), 'double', True, False): ~3144 MLUPS
+(800, (100, 1), 'double', True, True): ~3216 MLUPS
+(800, (50, 1), 'single', True, False): ~5927 MLUPS
+(800, (50, 1), 'single', True, True): ~5897 MLUPS
+(800, (50, 1), 'double', True, False): ~3141 MLUPS
+(800, (50, 1), 'double', True, True): ~3156 MLUPS
+(100, (10, 1), 'single', True, False): ~144 MLUPS
+(100, (10, 1), 'single', True, True): ~143 MLUPS
+(100, (10, 1), 'double', True, False): ~142 MLUPS
+(100, (10, 1), 'double', True, True): ~140 MLUPS
+(100, (100, 1), 'single', True, False): ~143 MLUPS
+(100, (100, 1), 'single', True, True): ~144 MLUPS
+(100, (100, 1), 'double', True, False): ~140 MLUPS
+(100, (100, 1), 'double', True, True): ~144 MLUPS
+(100, (50, 1), 'single', True, False): ~141 MLUPS
+(100, (50, 1), 'single', True, True): ~142 MLUPS
+(100, (50, 1), 'double', True, False): ~147 MLUPS
+(100, (50, 1), 'double', True, True): ~146 MLUPS
+(200, (200, 1), 'single', True, False): ~576 MLUPS
+(200, (200, 1), 'single', True, True): ~576 MLUPS
+(200, (200, 1), 'double', True, False): ~573 MLUPS
+(200, (200, 1), 'double', True, True): ~580 MLUPS
+(200, (10, 1), 'single', True, False): ~582 MLUPS
+(200, (10, 1), 'single', True, True): ~581 MLUPS
+(200, (10, 1), 'double', True, False): ~580 MLUPS
+(200, (10, 1), 'double', True, True): ~583 MLUPS
+(200, (100, 1), 'single', True, False): ~584 MLUPS
+(200, (100, 1), 'single', True, True): ~587 MLUPS
+(200, (100, 1), 'double', True, False): ~580 MLUPS
+(200, (100, 1), 'double', True, True): ~579 MLUPS
+(200, (50, 1), 'single', True, False): ~583 MLUPS
+(200, (50, 1), 'single', True, True): ~588 MLUPS
+(200, (50, 1), 'double', True, False): ~578 MLUPS
+(200, (50, 1), 'double', True, True): ~573 MLUPS
+(1000, (200, 1), 'single', True, False): ~6442 MLUPS
+(1000, (200, 1), 'single', True, True): ~6252 MLUPS
+(1000, (200, 1), 'double', True, False): ~3312 MLUPS
+(1000, (200, 1), 'double', True, True): ~3317 MLUPS
+(1000, (10, 1), 'single', True, False): ~4990 MLUPS
+(1000, (10, 1), 'single', True, True): ~4749 MLUPS
+(1000, (10, 1), 'double', True, False): ~3130 MLUPS
+(1000, (10, 1), 'double', True, True): ~3126 MLUPS
+(1000, (100, 1), 'single', True, False): ~6348 MLUPS
+(1000, (100, 1), 'single', True, True): ~6218 MLUPS
+(1000, (100, 1), 'double', True, False): ~3327 MLUPS
+(1000, (100, 1), 'double', True, True): ~3287 MLUPS
+(1000, (50, 1), 'single', True, False): ~6318 MLUPS
+(1000, (50, 1), 'single', True, True): ~6095 MLUPS
+(1000, (50, 1), 'double', True, False): ~3242 MLUPS
+(1000, (50, 1), 'double', True, True): ~3269 MLUPS
+(400, (200, 1), 'single', True, False): ~2295 MLUPS
+(400, (200, 1), 'single', True, True): ~2252 MLUPS
+(400, (200, 1), 'double', True, False): ~2197 MLUPS
+(400, (200, 1), 'double', True, True): ~2204 MLUPS
+(400, (10, 1), 'single', True, False): ~2278 MLUPS
+(400, (10, 1), 'single', True, True): ~2325 MLUPS
+(400, (10, 1), 'double', True, False): ~2215 MLUPS
+(400, (10, 1), 'double', True, True): ~2306 MLUPS
+(400, (100, 1), 'single', True, False): ~2319 MLUPS
+(400, (100, 1), 'single', True, True): ~2307 MLUPS
+(400, (100, 1), 'double', True, False): ~2298 MLUPS
+(400, (100, 1), 'double', True, True): ~2296 MLUPS
+(400, (50, 1), 'single', True, False): ~2307 MLUPS
+(400, (50, 1), 'single', True, True): ~2318 MLUPS
+(400, (50, 1), 'double', True, False): ~2245 MLUPS
+(400, (50, 1), 'double', True, True): ~2240 MLUPS
+(50, (10, 1), 'single', True, False): ~35 MLUPS
+(50, (10, 1), 'single', True, True): ~35 MLUPS
+(50, (10, 1), 'double', True, False): ~35 MLUPS
+(50, (10, 1), 'double', True, True): ~35 MLUPS
+(50, (50, 1), 'single', True, False): ~35 MLUPS
+(50, (50, 1), 'single', True, True): ~35 MLUPS
+(50, (50, 1), 'double', True, False): ~33 MLUPS
+(50, (50, 1), 'double', True, True): ~36 MLUPS
+(600, (30, 1), 'single', True, False): ~5074 MLUPS
+(600, (30, 1), 'single', True, True): ~5089 MLUPS
+(600, (30, 1), 'double', True, False): ~3260 MLUPS
+(600, (30, 1), 'double', True, True): ~3258 MLUPS
+(600, (200, 1), 'single', True, False): ~5065 MLUPS
+(600, (200, 1), 'single', True, True): ~5038 MLUPS
+(600, (200, 1), 'double', True, False): ~3298 MLUPS
+(600, (200, 1), 'double', True, True): ~3274 MLUPS
+(600, (10, 1), 'single', True, False): ~4640 MLUPS
+(600, (10, 1), 'single', True, True): ~4594 MLUPS
+(600, (10, 1), 'double', True, False): ~3103 MLUPS
+(600, (10, 1), 'double', True, True): ~3075 MLUPS
+(600, (100, 1), 'single', True, False): ~5065 MLUPS
+(600, (100, 1), 'single', True, True): ~5065 MLUPS
+(600, (100, 1), 'double', True, False): ~3280 MLUPS
+(600, (100, 1), 'double', True, True): ~3267 MLUPS
+(600, (50, 1), 'single', True, False): ~5173 MLUPS
+(600, (50, 1), 'single', True, True): ~5148 MLUPS
+(600, (50, 1), 'double', True, False): ~3271 MLUPS
+(600, (50, 1), 'double', True, True): ~3288 MLUPS
+(800, (256, 1), 'single', True, True): ~6014 MLUPS
+(800, (256, 1), 'double', True, True): ~3320 MLUPS
+(800, (32, 1), 'single', True, True): ~6203 MLUPS
+(800, (32, 1), 'double', True, True): ~3331 MLUPS
+(800, (64, 1), 'single', True, True): ~6256 MLUPS
+(800, (64, 1), 'double', True, True): ~3348 MLUPS
+(800, (128, 1), 'single', True, True): ~6359 MLUPS
+(800, (128, 1), 'double', True, True): ~3421 MLUPS
+(800, (512, 1), 'single', True, True): ~6022 MLUPS
+(800, (512, 1), 'double', True, True): ~3304 MLUPS
+(800, (16, 1), 'single', True, True): ~5725 MLUPS
+(800, (16, 1), 'double', True, True): ~3190 MLUPS
+(100, (32, 1), 'single', True, True): ~147 MLUPS
+(100, (32, 1), 'double', True, True): ~145 MLUPS
+(100, (64, 1), 'single', True, True): ~145 MLUPS
+(100, (64, 1), 'double', True, True): ~146 MLUPS
+(100, (128, 1), 'single', True, True): ~145 MLUPS
+(100, (128, 1), 'double', True, True): ~146 MLUPS
+(100, (16, 1), 'single', True, True): ~148 MLUPS
+(100, (16, 1), 'double', True, True): ~146 MLUPS
+(200, (256, 1), 'single', True, True): ~583 MLUPS
+(200, (256, 1), 'double', True, True): ~573 MLUPS
+(200, (32, 1), 'single', True, True): ~580 MLUPS
+(200, (32, 1), 'double', True, True): ~577 MLUPS
+(200, (64, 1), 'single', True, True): ~579 MLUPS
+(200, (64, 1), 'double', True, True): ~581 MLUPS
+(200, (128, 1), 'single', True, True): ~579 MLUPS
+(200, (128, 1), 'double', True, True): ~579 MLUPS
+(200, (16, 1), 'single', True, True): ~574 MLUPS
+(200, (16, 1), 'double', True, True): ~584 MLUPS
+(1000, (256, 1), 'single', True, True): ~6403 MLUPS
+(1000, (256, 1), 'double', True, True): ~3343 MLUPS
+(1000, (32, 1), 'single', True, True): ~6414 MLUPS
+(1000, (32, 1), 'double', True, True): ~3362 MLUPS
+(1000, (64, 1), 'single', True, True): ~6300 MLUPS
+(1000, (64, 1), 'double', True, True): ~3379 MLUPS
+(1000, (128, 1), 'single', True, True): ~6342 MLUPS
+(1000, (128, 1), 'double', True, True): ~3385 MLUPS
+(1000, (512, 1), 'single', True, True): ~6401 MLUPS
+(1000, (512, 1), 'double', True, True): ~3360 MLUPS
+(1000, (1024, 1), 'single', True, True): ~6395 MLUPS
+(1000, (1024, 1), 'double', True, True): ~3301 MLUPS
+(1000, (16, 1), 'single', True, True): ~5847 MLUPS
+(1000, (16, 1), 'double', True, True): ~3201 MLUPS
+(400, (256, 1), 'single', True, True): ~2077 MLUPS
+(400, (256, 1), 'double', True, True): ~2031 MLUPS
+(400, (32, 1), 'single', True, True): ~2078 MLUPS
+(400, (32, 1), 'double', True, True): ~2059 MLUPS
+(400, (64, 1), 'single', True, True): ~2066 MLUPS
+(400, (64, 1), 'double', True, True): ~2043 MLUPS
+(400, (128, 1), 'single', True, True): ~2082 MLUPS
+(400, (128, 1), 'double', True, True): ~2072 MLUPS
+(400, (16, 1), 'single', True, True): ~2092 MLUPS
+(400, (16, 1), 'double', True, True): ~2272 MLUPS
+(50, (32, 1), 'single', True, True): ~32 MLUPS
+(50, (32, 1), 'double', True, True): ~32 MLUPS
+(50, (64, 1), 'single', True, True): ~32 MLUPS
+(50, (64, 1), 'double', True, True): ~32 MLUPS
+(50, (128, 1), 'single', True, True): ~32 MLUPS
+(50, (128, 1), 'double', True, True): ~32 MLUPS
+(50, (16, 1), 'single', True, True): ~32 MLUPS
+(50, (16, 1), 'double', True, True): ~32 MLUPS
+(600, (256, 1), 'single', True, True): ~4592 MLUPS
+(600, (256, 1), 'double', True, True): ~3299 MLUPS
+(600, (32, 1), 'single', True, True): ~4678 MLUPS
+(600, (32, 1), 'double', True, True): ~3330 MLUPS
+(600, (64, 1), 'single', True, True): ~4588 MLUPS
+(600, (64, 1), 'double', True, True): ~3310 MLUPS
+(600, (128, 1), 'single', True, True): ~4594 MLUPS
+(600, (128, 1), 'double', True, True): ~3313 MLUPS
+(600, (512, 1), 'single', True, True): ~4517 MLUPS
+(600, (512, 1), 'double', True, True): ~3200 MLUPS
+(600, (16, 1), 'single', True, True): ~4661 MLUPS
+(600, (16, 1), 'double', True, True): ~3260 MLUPS
diff --git a/result/ldc_2d_benchmark_P100.py b/result/ldc_2d_benchmark_P100.py
index 50f5f9d..683d22d 100644
--- a/result/ldc_2d_benchmark_P100.py
+++ b/result/ldc_2d_benchmark_P100.py
@@ -1 +1 @@
-ldc_2d_p100 = [((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])]
+ldc_2d_p100 = [((32, (32, 1), 'single', False, True), [13, 14, 14, 14, 14]), ((32, (32, 1), 'single', True, True), [13, 14, 14, 14, 14]), ((32, (32, 1), 'double', False, True), [13, 13, 13, 14, 14]), ((32, (32, 1), 'double', True, True), [13, 14, 13, 13, 14]), ((32, (16, 1), 'single', False, True), [13, 14, 14, 14, 14]), ((32, (16, 1), 'single', True, True), [13, 13, 13, 13, 13]), ((32, (16, 1), 'double', False, True), [13, 14, 14, 13, 14]), ((32, (16, 1), 'double', True, True), [13, 14, 14, 13, 14]), ((256, (256, 1), 'single', False, True), [840, 857, 850, 857, 857]), ((256, (256, 1), 'single', True, True), [853, 860, 867, 866, 864]), ((256, (256, 1), 'double', False, True), [839, 849, 850, 848, 852]), ((256, (256, 1), 'double', True, True), [840, 862, 860, 858, 862]), ((256, (32, 1), 'single', False, True), [856, 866, 851, 868, 871]), ((256, (32, 1), 'single', True, True), [853, 863, 863, 856, 861]), ((256, (32, 1), 'double', False, True), [840, 857, 855, 853, 854]), ((256, (32, 1), 'double', True, True), [832, 851, 849, 856, 856]), ((256, (64, 1), 'single', False, True), [822, 840, 831, 791, 783]), ((256, (64, 1), 'single', True, True), [848, 859, 856, 859, 860]), ((256, (64, 1), 'double', False, True), [836, 850, 852, 856, 852]), ((256, (64, 1), 'double', True, True), [847, 863, 860, 859, 857]), ((256, (128, 1), 'single', False, True), [942, 947, 953, 957, 962]), ((256, (128, 1), 'single', True, True), [937, 960, 950, 953, 965]), ((256, (128, 1), 'double', False, True), [916, 947, 954, 946, 931]), ((256, (128, 1), 'double', True, True), [940, 969, 965, 969, 980]), ((256, (16, 1), 'single', False, True), [938, 960, 935, 963, 955]), ((256, (16, 1), 'single', True, True), [926, 951, 938, 948, 933]), ((256, (16, 1), 'double', False, True), [929, 937, 944, 942, 927]), ((256, (16, 1), 'double', True, True), [894, 919, 916, 918, 918]), ((64, (32, 1), 'single', False, True), [56, 57, 57, 57, 57]), ((64, (32, 1), 'single', True, True), [59, 60, 59, 60, 60]), ((64, (32, 1), 'double', False, True), [57, 58, 59, 59, 59]), ((64, (32, 1), 'double', True, True), [58, 59, 59, 59, 59]), ((64, (64, 1), 'single', False, True), [58, 58, 59, 59, 59]), ((64, (64, 1), 'single', True, True), [58, 59, 60, 58, 60]), ((64, (64, 1), 'double', False, True), [54, 56, 59, 59, 59]), ((64, (64, 1), 'double', True, True), [59, 60, 59, 59, 59]), ((64, (16, 1), 'single', False, True), [59, 60, 58, 60, 60]), ((64, (16, 1), 'single', True, True), [59, 60, 59, 60, 61]), ((64, (16, 1), 'double', False, True), [58, 59, 59, 59, 60]), ((64, (16, 1), 'double', True, True), [58, 59, 60, 60, 57]), ((128, (32, 1), 'single', False, True), [213, 216, 211, 214, 215]), ((128, (32, 1), 'single', True, True), [213, 215, 216, 216, 218]), ((128, (32, 1), 'double', False, True), [211, 214, 213, 214, 215]), ((128, (32, 1), 'double', True, True), [213, 217, 216, 216, 217]), ((128, (64, 1), 'single', False, True), [213, 217, 217, 216, 216]), ((128, (64, 1), 'single', True, True), [213, 217, 217, 216, 215]), ((128, (64, 1), 'double', False, True), [214, 213, 216, 217, 218]), ((128, (64, 1), 'double', True, True), [215, 216, 218, 218, 218]), ((128, (128, 1), 'single', False, True), [214, 217, 218, 217, 217]), ((128, (128, 1), 'single', True, True), [213, 216, 215, 217, 218]), ((128, (128, 1), 'double', False, True), [210, 214, 215, 215, 215]), ((128, (128, 1), 'double', True, True), [212, 215, 215, 215, 214]), ((128, (16, 1), 'single', False, True), [212, 214, 215, 214, 215]), ((128, (16, 1), 'single', True, True), [214, 215, 216, 217, 217]), ((128, (16, 1), 'double', False, True), [212, 214, 214, 214, 212]), ((128, (16, 1), 'double', True, True), [215, 215, 217, 218, 219]), ((512, (256, 1), 'single', False, True), [3342, 3382, 3385, 3387, 3386]), ((512, (256, 1), 'single', True, True), [3382, 3411, 3429, 3429, 3385]), ((512, (256, 1), 'double', False, True), [2254, 2187, 2107, 2089, 2089]), ((512, (256, 1), 'double', True, True), [3307, 3344, 3347, 3348, 3350]), ((512, (32, 1), 'single', False, True), [3390, 3418, 3406, 3434, 3446]), ((512, (32, 1), 'single', True, True), [3360, 3409, 3415, 3419, 3407]), ((512, (32, 1), 'double', False, True), [2238, 2160, 2087, 2073, 2075]), ((512, (32, 1), 'double', True, True), [3329, 3366, 3366, 3369, 3368]), ((512, (64, 1), 'single', False, True), [3586, 3598, 3639, 3652, 3669]), ((512, (64, 1), 'single', True, True), [3356, 3425, 3425, 3435, 3429]), ((512, (64, 1), 'double', False, True), [2204, 2138, 2063, 2046, 2059]), ((512, (64, 1), 'double', True, True), [3313, 3359, 3364, 3358, 3365]), ((512, (128, 1), 'single', False, True), [3328, 3408, 3410, 3410, 3408]), ((512, (128, 1), 'single', True, True), [3368, 3431, 3412, 3425, 3418]), ((512, (128, 1), 'double', False, True), [2060, 1997, 1925, 1909, 1910]), ((512, (128, 1), 'double', True, True), [3307, 3322, 3332, 3331, 3334]), ((512, (512, 1), 'single', False, True), [3375, 3415, 2523, 3422, 3436]), ((512, (512, 1), 'single', True, True), [3369, 3409, 3409, 3420, 3433]), ((512, (512, 1), 'double', False, True), [2124, 2069, 2006, 1973, 1975]), ((512, (512, 1), 'double', True, True), [3301, 3323, 3346, 3342, 3343]), ((512, (16, 1), 'single', False, True), [3270, 3257, 3192, 3129, 3070]), ((512, (16, 1), 'single', True, True), [3409, 3444, 3453, 3453, 3446]), ((512, (16, 1), 'double', False, True), [1204, 1160, 1115, 1103, 1103]), ((512, (16, 1), 'double', True, True), [3281, 3310, 3311, 3311, 3316]), ((1024, (256, 1), 'single', False, True), [6327, 6343, 6308, 6283, 6253]), ((1024, (256, 1), 'single', True, True), [6601, 6627, 6626, 6627, 6624]), ((1024, (256, 1), 'double', False, True), [2729, 2683, 2631, 2597, 2549]), ((1024, (256, 1), 'double', True, True), [3475, 3482, 3482, 3483, 3483]), ((1024, (32, 1), 'single', False, True), [6582, 6622, 6614, 6603, 6598]), ((1024, (32, 1), 'single', True, True), [6711, 6732, 6732, 6735, 6731]), ((1024, (32, 1), 'double', False, True), [2701, 2658, 2613, 2561, 2507]), ((1024, (32, 1), 'double', True, True), [3489, 3493, 3490, 3492, 3491]), ((1024, (64, 1), 'single', False, True), [6530, 6565, 6564, 6558, 6546]), ((1024, (64, 1), 'single', True, True), [6688, 6698, 6704, 6703, 6703]), ((1024, (64, 1), 'double', False, True), [2696, 2655, 2609, 2557, 2503]), ((1024, (64, 1), 'double', True, True), [3483, 3485, 3487, 3487, 3486]), ((1024, (128, 1), 'single', False, True), [6460, 6494, 6486, 6481, 6473]), ((1024, (128, 1), 'single', True, True), [6686, 6701, 6697, 6699, 6697]), ((1024, (128, 1), 'double', False, True), [2697, 2656, 2612, 2561, 2506]), ((1024, (128, 1), 'double', True, True), [3457, 3460, 3460, 3460, 3460]), ((1024, (512, 1), 'single', False, True), [6188, 6205, 6191, 6177, 6159]), ((1024, (512, 1), 'single', True, True), [6697, 6707, 6718, 6721, 6717]), ((1024, (512, 1), 'double', False, True), [2665, 2650, 2596, 2514, 2443]), ((1024, (512, 1), 'double', True, True), [3489, 3494, 3493, 3494, 3494]), ((1024, (1024, 1), 'single', False, True), [5747, 5732, 5712, 5681, 5654]), ((1024, (1024, 1), 'single', True, True), [6636, 6647, 6652, 6650, 6651]), ((1024, (1024, 1), 'double', False, True), [2410, 2374, 2342, 2304, 2262]), ((1024, (1024, 1), 'double', True, True), [3344, 3349, 3349, 3348, 3349]), ((1024, (16, 1), 'single', False, True), [3949, 3914, 3876, 3834, 3793]), ((1024, (16, 1), 'single', True, True), [6451, 6461, 6462, 6464, 6462]), ((1024, (16, 1), 'double', False, True), [1385, 1358, 1335, 1309, 1282]), ((1024, (16, 1), 'double', True, True), [3441, 3444, 3445, 3444, 3443]), ((2048, (256, 1), 'single', False, True), [6674, 6714, 6776, 6827, 6822]), ((2048, (256, 1), 'single', True, True), [6879, 6886, 6886, 6886, 6886]), ((2048, (256, 1), 'double', False, True), [2845, 2826, 2817, 2806, 2778]), ((2048, (256, 1), 'double', True, True), [3569, 3571, 3570, 3571, 3571]), ((2048, (32, 1), 'single', False, True), [6954, 6961, 6959, 69