OpenCL：MQL5中的内部实现测试

Sceptic Philozoff 2012.06.01 00:03 #641

fyords: 在报告中，右击 "查看"，新窗口 "查询 "按钮，日志的计时正确，而且更容易阅读（对我来说）。

谢谢，我没有意识到你在开报告。

收益甚至更大。

我很惊讶：这是一张不到80美元的预算卡!因此，NVidia在驱动上做了一些认真的工作。

Dmitriy Parfenovich 2012.06.01 06:42 #642

而这里是新的结果。

2012.06.01 09:25:23     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     ========================================
2012.06.01 09:32:25     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     CPU time = 421203 ms
2012.06.01 09:32:25     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     Соunt indicators = 16; Count history bars = 144000; Count pass = 12800
2012.06.01 09:32:25     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     Result on Cpu МахResult==1.2809 at 9448 pass
2012.06.01 09:32:25     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     -------------------------
2012.06.01 09:32:28     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     Device number = 0
2012.06.01 09:32:39     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     GPU time = 11263 ms
2012.06.01 09:32:39     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     CpuTime/GpuTime = 37.39705229512563
2012.06.01 09:32:39     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     Result on Gpu МахResult==1.2809 at 9448 pass
2012.06.01 09:32:39     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     ------------
2012.06.01 09:32:39     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     Device number = 1
2012.06.01 09:32:40     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     GPU time = 998 ms
2012.06.01 09:32:40     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     CpuTime/GpuTime = 422.0470941883768
2012.06.01 09:32:40     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     Result on Gpu МахResult==1.2809 at 9448 pass
2012.06.01 09:32:40     ParallelTester_00-01 x_new_cycle (EURUSD,D1)     ------------

我的理解是：1.纯粹的CPU，2。带有OpenCL的CPU, 3.使用OpenCL的GPU？

而且还是422。

Mathemat:

我很惊讶：这是一张成本低于80美元的经济型卡片！这是我的梦想。因此，NVidia在驱动上做了很大的努力。

我是多么惊讶啊，从泥土到财富。人们的印象是，NVidia读了这个论坛，做了类似的测试，发现了错误并修复了它们。

如果测试人员能选择在什么地方进行模拟，即不强迫写代码，那就非常好了。尽管如此，1秒（或11秒，如果视频不允许或无法使用）对7分钟的力量。

Renat Fatkhullin 2012.06.01 08:15 #643

为了在一个小的误差范围内进行测试，有一个很好的方法，那就是进行一系列的测量，并对结果进行平均，或放弃极值。但是，当然，最好是增加计算量以稳定结果。

现代操作系统和真正的多核处理器已经严重消除了通过GetTickCount 进行测量的散乱问题。我最初的评论只是针对 "GetTickCount的平均误差至少有几十毫秒 "这一错误的说法。

Andrey Dik 2012.06.01 09:30 #644

英特尔处理器的用户请注意!当安装nVidia驱动301.42时，CPU的OpenCL驱动会自动安装。

在注册表中，它看起来像这样。

[HKEY_LOCAL_MACHINE/SOFTWARE/Khronos/OpenCL/Vendors]。
"nvcuda.dll"=dword:00000000
"amdocl.dll"=dword:00000000
"amdocl64.dll"=dword:00000000
"IntelOpenCL64.dll"=dword:00000000

它们比英特尔的本地驱动程序（绿色突出显示）慢了大约1.5倍（红色突出显示）。

你可以删除相应的注册表值，但为了以防万一，要保存该分支。

Oleksandr Olefirov 2012.06.01 21:14 #645

亲爱的管理员。好久没上你们的论坛了，可能错过了这一点。

是否会有向云端需求交出显卡的实施方案？

MetaQuotes 2012.06.01 21:46 #646

ilovebtc:

亲爱的管理员。好久没上你们的论坛了，可能错过了这一点。

是否会为了云计算的需要而实施显卡互换？

几乎完成https://www.mql5.com/ru/forum/23/page15#comment_201948

19.MetaTester：增加了对测试代理中使用OpenCL程序的支持。

OpenCL 程序是用来在支持OpenCL 1.1或更高版本的显卡上进行计算的。现代显卡包含数百个小型专用处理器，可以同时对传入的数据流进行简单的数学运算。 OpenCL语言承担了这种并行计算的组织工作，并为某类任务提供了巨大的速度提升。

测试结束后，我们将在分布式网络中启用OpenCL。

Sceptic Philozoff 2012.06.02 03:22 #647

fyords: 我的理解是：1.纯粹的CPU，2.带有OpenCL的CPU, 3.使用OpenCL的GPU？

是的，这是正确的。

你是否介意运行所附的脚本并公布结果？这真的很有趣。

不要害怕大量的数字。它们只是用来检查计算的正确性。

该脚本还运行了所有的设备。主要任务是将两个大矩阵相乘。

设置只能在代码中改变--这一行中矩阵的线性大小_size。

#define _size       2000

只有当你的内存用完时才会改变它们。一个标志是在离散的GPU上运行时数组数字的差异：如果数字的差异超过10^(-4)，那就是一个明显的错误。但你似乎有足够的记忆。

附加的文件：

vect_v2_all_devices.mq5 17 kb

Oleksandr Olefirov 2012.06.02 05:16 #648

例如，我有一块Radeon 6930显卡，它有1280个流处理器。它将如何显示在代理人名单中？作为1个设备，或所有1280。

它本身的速度是10个处理器的数倍，而且奖金不是为了增加1个设备。

Dmitriy Parfenovich 2012.06.02 06:38 #649

Mathemat:

你是否介意运行所附脚本并显示结果？这真的很有趣。

不，这不是一个无聊的问题。我自己也对它感到好奇。我在设置中没有改变任何东西。

2012.06.02 09:28:27     vect_v2_all_devices (EURUSD,D1) =======================================
2012.06.02 09:28:27     vect_v2_all_devices (EURUSD,D1) OCL martices mul:         ROWS1 = 2000; COLSROWS = 2000; COLS2 = 2000
2012.06.02 09:30:31     vect_v2_all_devices (EURUSD,D1) CPUTime = 124.504
2012.06.02 09:30:31     vect_v2_all_devices (EURUSD,D1) ---------------
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) read = 4000000 elements
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) Device = 0: time = 2.824 sec.
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) CPUTime / GPUTotalTime = 44.088
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 1968,1939 ) = -5.27639246;    thirdCPU[ 1968,1939 ] = -5.27639246;    buf[ 1968,1939 ] = -5.27639198
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 585,810 ) = 3.74615073;    thirdCPU[ 585,810 ] = 3.74615073;    buf[ 585,810 ] = 3.74614906
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 1131,1732 ) = -4.46934557;    thirdCPU[ 1131,1732 ] = -4.46934557;    buf[ 1131,1732 ] = -4.46934605
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 587,999 ) = -4.46048546;    thirdCPU[ 587,999 ] = -4.46048546;    buf[ 587,999 ] = -4.46048260
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 983,1903 ) = 3.42076445;    thirdCPU[ 983,1903 ] = 3.42076445;    buf[ 983,1903 ] = 3.42076564
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 1927,313 ) = 5.62960339;    thirdCPU[ 1927,313 ] = 5.62960339;    buf[ 1927,313 ] = 5.62960196
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 355,1897 ) = 5.86679220;    thirdCPU[ 355,1897 ] = 5.86679220;    buf[ 355,1897 ] = 5.86678505
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 1455,1651 ) = -3.67937088;    thirdCPU[ 1455,1651 ] = -3.67937088;    buf[ 1455,1651 ] = -3.67936754
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 1207,856 ) = 1.30920172;    thirdCPU[ 1207,856 ] = 1.30920172;    buf[ 1207,856 ] = 1.30920100
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) sum( 1699,575 ) = 2.55669522;    thirdCPU[ 1699,575 ] = 2.55669522;    buf[ 1699,575 ] = 2.55669498
2012.06.02 09:30:38     vect_v2_all_devices (EURUSD,D1) ________________________
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) read = 4000000 elements
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) Device = 1: time = 1.514 sec.
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) CPUTime / GPUTotalTime = 82.235
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 407,514 ) = -3.69270682;    thirdCPU[ 407,514 ] = -3.69270682;    buf[ 407,514 ] = -3.69270515
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 1421,1902 ) = -7.43944120;    thirdCPU[ 1421,1902 ] = -7.43944120;    buf[ 1421,1902 ] = -7.43943167
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 1197,1072 ) = -1.49989450;    thirdCPU[ 1197,1072 ] = -1.49989450;    buf[ 1197,1072 ] = -1.49989557
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 1249,1056 ) = -0.22817086;    thirdCPU[ 1249,1056 ] = -0.22817086;    buf[ 1249,1056 ] = -0.22817032
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 385,1856 ) = 3.88903213;    thirdCPU[ 385,1856 ] = 3.88903213;    buf[ 385,1856 ] = 3.88902068
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 952,488 ) = 0.37963703;    thirdCPU[ 952,488 ] = 0.37963703;    buf[ 952,488 ] = 0.37963703
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 345,1572 ) = 2.28500485;    thirdCPU[ 345,1572 ] = 2.28500485;    buf[ 345,1572 ] = 2.28500390
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 1928,468 ) = -1.35805547;    thirdCPU[ 1928,468 ] = -1.35805547;    buf[ 1928,468 ] = -1.35805535
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 1881,1968 ) = -3.12033391;    thirdCPU[ 1881,1968 ] = -3.12033391;    buf[ 1881,1968 ] = -3.12033129
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) sum( 1454,575 ) = 5.97233009;    thirdCPU[ 1454,575 ] = 5.97233009;    buf[ 1454,575 ] = 5.97232151
2012.06.02 09:30:42     vect_v2_all_devices (EURUSD,D1) ________________________

我只是不明白任何数字。你能解释一下吗？那么，至少在手指上：是好还是不好？它们在不同的设备之间是不同的，在小数点之后的5-6位数的行中已经是不同的地方。

我想我明白了：这是对重复操作的多重测试，最后的时间是每个设备的平均时间。对吗？

Sceptic Philozoff 2012.06.02 07:10 #650

fyords: 只是我从这些数字中什么都不明白。你能解释一下吗？就在你的手指上：它是好还是不好？它们在不同的设备之间是不同的，在小数点之后的5-6位数的行中已经是不同的地方。

sum( 407,514 ) = -3.69270682;    thirdCPU[ 407,514 ] = -3.69270682;    buf[ 407,514 ] = -3.69270515

这些只是检查数字。如果它们与0.00001重合，则一切正常。索引是随机选择的--这是一个随机检查，以确保计算的正确性。好吧，我们不打算在这里打印对所得矩阵的所有400万个元素进行全面检查的结果，对吗？

我想我明白了：这是对重复操作的多重测试，最后的时间是每个设备的平均时间。对吗？

不，这是两个大矩阵相乘的单一操作。

在性能数据方面：对于这块卡来说非常好。现在我的结果。设备（从下到上--初始化顺序）。

2012.06.02 05:49:25     OpenCL  CPU: GenuineIntel  Intel(R) Pentium(R) CPU G840 @ 2.80 GHz with OpenCL 1.2 (2 units, 2793 MHz, 8040 Mb, version 2.0 (sse2))
2012.06.02 05:49:25     OpenCL  GPU: Advanced Micro Devices, Inc. ATI RV770 with OpenCL 1.0 (10 units, 780 MHz, 512 Mb, version CAL 1.4.1720)
2012.06.02 05:49:25     OpenCL  CPU: Intel(R) Corporation  Intel(R) Pentium(R) CPU G840 @ 2.80 GHz with OpenCL 1.1 (2 units, 2800 MHz, 8040 Mb, version 1.1)

即先用英特尔的OCL引擎的英特尔CPU，然后是我的恐龙HD 4870，然后又是一块石头，但用AMD的引擎。脚本。

2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) ________________________
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 1477,98 ) = -5.84002066;    thirdCPU[ 1477,98 ] = -5.84002066;    buf[ 1477,98 ] = -5.84001255
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 1339,1186 ) = 0.59214997;    thirdCPU[ 1339,1186 ] = 0.59214997;    buf[ 1339,1186 ] = 0.59215009
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 1410,1861 ) = -0.27033439;    thirdCPU[ 1410,1861 ] = -0.27033439;    buf[ 1410,1861 ] = -0.27033412
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 1282,459 ) = -0.87189484;    thirdCPU[ 1282,459 ] = -0.87189484;    buf[ 1282,459 ] = -0.87189591
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 710,1645 ) = 4.86117268;    thirdCPU[ 710,1645 ] = 4.86117268;    buf[ 710,1645 ] = 4.86116362
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 526,938 ) = 0.94805324;    thirdCPU[ 526,938 ] = 0.94805324;    buf[ 526,938 ] = 0.94805157
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 914,489 ) = 5.58242941;    thirdCPU[ 914,489 ] = 5.58242941;    buf[ 914,489 ] = 5.58243275
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 811,257 ) = -1.11584055;    thirdCPU[ 811,257 ] = -1.11584055;    buf[ 811,257 ] = -1.11583853
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 318,498 ) = 1.62952971;    thirdCPU[ 318,498 ] = 1.62952971;    buf[ 318,498 ] = 1.62952805
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) sum( 648,1434 ) = -5.57316303;    thirdCPU[ 648,1434 ] = -5.57316303;    buf[ 648,1434 ] = -5.57315731
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) CPUTime / GPUTotalTime = 29.879
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) Device = 2: time = 3.105 sec.
2012.06.02 07:38:19     vect_v2_all_devices (EURUSD,H1) read = 4000000 elements
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) ________________________
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 684,439 ) = 0.21124490;    thirdCPU[ 684,439 ] = 0.21124490;    buf[ 684,439 ] = 0.21124732
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 795,204 ) = -1.68047857;    thirdCPU[ 795,204 ] = -1.68047857;    buf[ 795,204 ] = -1.68047154
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 579,1503 ) = 2.46559286;    thirdCPU[ 579,1503 ] = 2.46559286;    buf[ 579,1503 ] = 2.46558809
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 675,1504 ) = 0.44935751;    thirdCPU[ 675,1504 ] = 0.44935751;    buf[ 675,1504 ] = 0.44935691
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 1251,1415 ) = -2.85569835;    thirdCPU[ 1251,1415 ] = -2.85569835;    buf[ 1251,1415 ] = -2.85569715
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 204,1755 ) = 0.31420049;    thirdCPU[ 204,1755 ] = 0.31420049;    buf[ 204,1755 ] = 0.31420231
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 1999,74 ) = -2.22978306;    thirdCPU[ 1999,74 ] = -2.22978306;    buf[ 1999,74 ] = -2.22977948
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 436,657 ) = 0.59192652;    thirdCPU[ 436,657 ] = 0.59192652;    buf[ 436,657 ] = 0.59192693
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 967,922 ) = -4.91348410;    thirdCPU[ 967,922 ] = -4.91348410;    buf[ 967,922 ] = -4.91348314
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) sum( 1489,1175 ) = -2.48868656;    thirdCPU[ 1489,1175 ] = -2.48868656;    buf[ 1489,1175 ] = -2.48868561
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) CPUTime / GPUTotalTime = 179.795
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) Device = 1: time = 0.516 sec.
2012.06.02 07:38:15     vect_v2_all_devices (EURUSD,H1) read = 4000000 elements
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) ________________________
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 303,1215 ) = -7.46387863;    thirdCPU[ 303,1215 ] = -7.46387863;    buf[ 303,1215 ] = -7.46388054
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 1173,1406 ) = -5.64940453;    thirdCPU[ 1173,1406 ] = -5.64940453;    buf[ 1173,1406 ] = -5.64940882
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 1617,1405 ) = -0.98162729;    thirdCPU[ 1617,1405 ] = -0.98162729;    buf[ 1617,1405 ] = -0.98162866
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 760,1003 ) = -0.97699410;    thirdCPU[ 760,1003 ] = -0.97699410;    buf[ 760,1003 ] = -0.97699606
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 679,793 ) = -5.41226530;    thirdCPU[ 679,793 ] = -5.41226530;    buf[ 679,793 ] = -5.41227150
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 1345,1865 ) = 0.95630527;    thirdCPU[ 1345,1865 ] = 0.95630527;    buf[ 1345,1865 ] = 0.95630503
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 1289,1659 ) = -3.82919979;    thirdCPU[ 1289,1659 ] = -3.82919979;    buf[ 1289,1659 ] = -3.82920074
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 1216,1759 ) = 4.87398672;    thirdCPU[ 1216,1759 ] = 4.87398672;    buf[ 1216,1759 ] = 4.87398672
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 1268,1060 ) = 2.78621030;    thirdCPU[ 1268,1060 ] = 2.78621030;    buf[ 1268,1060 ] = 2.78621435
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) sum( 1686,577 ) = -4.36586094;    thirdCPU[ 1686,577 ] = -4.36586094;    buf[ 1686,577 ] = -4.36585188
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) CPUTime / GPUTotalTime = 22.783
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) Device = 0: time = 4.072 sec.
2012.06.02 07:38:14     vect_v2_all_devices (EURUSD,H1) read = 4000000 elements
2012.06.02 07:38:10     vect_v2_all_devices (EURUSD,H1) ---------------
2012.06.02 07:38:10     vect_v2_all_devices (EURUSD,H1) CPUTime = 92.774
2012.06.02 07:36:37     vect_v2_all_devices (EURUSD,H1) OCL martices mul:         ROWS1 = 2000; COLSROWS = 2000; COLS2 = 2000
2012.06.02 07:36:37     vect_v2_all_devices (EURUSD,H1) =======================================

Документация по MQL5: Основы языка / Переменные / Создание и уничтожение объектов

www.mql5.com

Основы языка / Переменные / Создание и уничтожение объектов - Документация по MQL5

OpenCL：MQL5中的内部实现测试 - 页 65