konjac / division Goto Github PK

执行这里的测试，乘加运算fmadd的性能明显过低。

MIC理论峰值单精度是1T，陈老师之前测得是700+G。

**************** Test for FMADD ****************
fmadd_intrin flops = 62.070177 Gflops
fmadd_intrin flops = 153.658040 Gflops
fmadd_autovec flops = 3.188641 Gflops
fmadd_autovec flops = 3.175300 Gflops
**************** Test for DIV ****************
division_cpu flops = 2.377036 Gflops
division_cpu flops = 2.408519 Gflops
division_intrin flops = 10.930253 Gflops
division_intrin flops = 10.779101 Gflops
division_autovec flops = 0.874387 Gflops
division_autovec flops = 0.806637 Gflops
newdiv_autovec flops = 34.460859 Gflops
newdiv_autovec flops = 34.163176 Gflops
newdiv_intrin flops = 35.127830 Gflops
newdiv_intrin flops = 31.881515 Gflops

不查表是查表的3倍

代码在branch yesx下面

精度请参考程序 https://github.com/konjac/division/blob/yesx/yesx/test.cpp
使用快速平方根倒数得到一个接近的参考值，然后使用牛顿迭代法
自己测了若干数据：
牛顿迭代发的次数，float迭代2次，double迭代3次收敛

速度请参考程序 https://github.com/konjac/division/blob/yesx/gputest-yesx/divisionflops.cu
time = 544.192322 ms

对比的程序为https://github.com/konjac/division/blob/yesx/gputest/divisionflops.cu
time = 748.940674 ms

ARCH = sm_21
实验环境
Device 0: "GeForce GT 630"
CUDA Driver Version / Runtime Version 5.5 / 5.0
CUDA Capability Major/Minor version number: 2.1
Total amount of global memory: 1024 MBytes (1073283072 bytes)
( 2) Multiprocessors x ( 48) CUDA Cores/MP: 96 CUDA Cores
GPU Clock rate: 1620 MHz (1.62 GHz)
Memory Clock rate: 667 Mhz
Memory Bus Width: 128-bit
L2 Cache Size: 131072 bytes
Max Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536,65535), 3D=(2048,2048,2048)
Max Layered Texture Size (dim) x layers 1D=(16384) x 2048, 2D=(16384,16384) x 2048
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 32768
Warp size: 32
Maximum number of threads per multiprocessor: 1536
Maximum number of threads per block: 1024
Maximum sizes of each dimension of a block: 1024 x 1024 x 64
Maximum sizes of each dimension of a grid: 65535 x 65535 x 65535
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 1 copy engine(s)
Run time limit on kernels: Yes
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): No
Device PCI Bus ID / PCI location ID: 1 / 0
Compute Mode:
< Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 5.5, CUDA Runtime Version = 5.0, NumDevs = 1, Device0 = GeForce GT 630

除法精度测试比较：
（分别为CPU除法，MIC函数库除法，陈老师的方法with 5 iterations，叶树雄的方法）

0.840188 / 0.394383 =
CPU Standard: 2.130386
MIC Standard: 2.130386
MIC Iterative: 2.130386
MIC FISR: 2.089152

0.783099 / 0.798440 =
CPU Standard: 0.980787
MIC Standard: 0.980787
MIC Iterative: 0.980787
MIC FISR: 0.980659

0.911647 / 0.197551 =
CPU Standard: 4.614736
MIC Standard: 4.614736
MIC Iterative: 4.614732
MIC FISR: 4.174708

0.335223 / 0.768230 =
CPU Standard: 0.436358
MIC Standard: 0.436358
MIC Iterative: 0.436358
MIC FISR: 0.436254

0.277775 / 0.553970 =
CPU Standard: 0.501426
MIC Standard: 0.501426
MIC Iterative: 0.501426
MIC FISR: 0.499256

0.477397 / 0.628871 =
CPU Standard: 0.759134
MIC Standard: 0.759134
MIC Iterative: 0.759134
MIC FISR: 0.757703

0.364784 / 0.513401 =
CPU Standard: 0.710526
MIC Standard: 0.710526
MIC Iterative: 0.710526
MIC FISR: 0.705917

0.952230 / 0.916195 =
CPU Standard: 1.039331
MIC Standard: 1.039331
MIC Iterative: 1.039331
MIC FISR: 1.039327

lei-april commented a day ago

对长度为1024的double型数组进行2^15次随机访问，
在使用prefetch指令的情况下，L1 cache的命中率大约提升1%（96.6% -> 97.9%）

又测试了几遍，似乎prefetch没有引入明显的性能提升。

把那天组会讲过的性能数据贴这里吧，以后如果需要方便查

https://github.com/konjac/division/raw/d7e842d0cd96557f44a359693c901e34a1ea8817/doc/algorithm-description.pdf

https://github.com/konjac/division/raw/master/doc/algorithm-description.pdf