编译的加速代码的性能差异来自ghci和shell

时间:2014-12-18 07:51:09

标签: performance haskell profiling ghci accelerate-haskell

问题

您好,我正在使用加速库创建一个应用程序,允许用户以交互方式调用处理图像的函数,这就是我使用ghc api进行扩展和扩展ghci的原因。

问题是当从shell运行已编译的可执行文件时,计算在100ms(略小于80)下完成,而在ghci中运行相同的编译代码则需要超过100ms(平均大于140)光洁度。

资源

示例代码+执行日志: https://gist.github.com/zgredzik/15a437c87d3d8d03b8fc

描述

首先:测试是在编译CUDA内核之后运行的(编译本身又增加了2秒,但事实并非如此)。

从shell运行已编译的可执行文件时,计算在10ms内完成。 (shell first runsecond shell run传递了不同的参数,以确保数据不会缓存在任何地方。)

当尝试从ghci运行相同的代码并摆弄输入数据时,计算时间超过100毫秒。我理解解释的代码比编译的代码慢,但是我在ghci会话中加载相同的编译代码并调用相同的顶级绑定(packedFunction)。我已明确键入它以确保它是专门的(与使用SPECIALIZED编译指示相同的结果)。

但是如果我在ghci中运行main函数(即使在连续调用之间用:set args更改输入数据),计算的确需要不到10毫秒。

使用Main.hs

编译ghc -o main Main.hs -O2 -dynamic -threaded

我想知道开销来自哪里。有没有人对为什么会这样做有任何建议?


remdezx发布的示例的简化版:

{-# LANGUAGE OverloadedStrings #-}

module Main where

import Data.Array.Accelerate as A
import Data.Array.Accelerate.CUDA as C
import Data.Time.Clock       (diffUTCTime, getCurrentTime)

main :: IO ()
main = do
    start <- getCurrentTime
    print $ C.run $ A.maximum $ A.map (+1) $ A.use (fromList (Z:.1000000) [1..1000000] :: Vector Double)
    end   <- getCurrentTime
    print $ diffUTCTime end start

当我编译并执行时,需要 0,09s 才能完成。

$ ghc -O2 Main.hs -o main -threaded
[1 of 1] Compiling Main             ( Main.hs, Main.o )
Linking main ...
$ ./main
Array (Z) [1000001.0]
0.092906s

但是当我预编译它并在解释器中运行它需要 0,25s

$ ghc -O2 Main.hs -c -dynamic
$ ghci Main
ghci> main
Array (Z) [1000001.0]
0.258224s

1 个答案:

答案 0 :(得分:3)

我调查了accelerateaccelerate-cuda并添加了一些调试代码来测量ghci和编译优化版本下的时间。

结果如下,您可以看到堆栈跟踪和执行时间。

ghci run

$ ghc -O2 -dynamic -c -threaded Main.hs && ghci 
GHCi, version 7.8.3: http://www.haskell.org/ghc/  :? for help
…
Loading package ghc-prim ... linking ... done.
Loading package integer-gmp ... linking ... done.
Loading package base ... linking ... done.
Ok, modules loaded: Main.
Prelude Main> Loading package transformers-0.3.0.0 ... linking ... done.
…
Loading package array-0.5.0.0 ... linking ... done.
(...)
Loading package accelerate-cuda-0.15.0.0 ... linking ... done.
>>>>> run
>>>>> runAsyncIn.execute
>>>>>  runAsyncIn.seq ctx
<<<<<  runAsyncIn.seq ctx: 4.1609e-2 CPU  0.041493s TOTAL
>>>>>  runAsyncIn.seq a
<<<<<  runAsyncIn.seq a: 1.0e-6 CPU  0.000001s TOTAL
>>>>>  runAsyncIn.seq acc
>>>>>   convertAccWith True
<<<<<   convertAccWith: 0.0 CPU  0.000017s TOTAL
<<<<<  runAsyncIn.seq acc: 2.68e-4 CPU  0.000219s TOTAL
>>>>>  evalCUDA
>>>>>   push
<<<<<   push: 0.0 CPU  0.000002s TOTAL
>>>>>   evalStateT
>>>>>    runAsyncIn.compileAcc
>>>>>     compileOpenAcc
>>>>>      compileOpenAcc.traveuseAcc.Alet
>>>>>      compileOpenAcc.traveuseAcc.Use
>>>>>       compileOpenAcc.traveuseAcc.use3
>>>>>       compileOpenAcc.traveuseAcc.use1
<<<<<       compileOpenAcc.traveuseAcc.use1: 0.0 CPU  0.000001s TOTAL
>>>>>       compileOpenAcc.traveuseAcc.use2
>>>>>        compileOpenAcc.traveuseAcc.seq arr
<<<<<        compileOpenAcc.traveuseAcc.seq arr: 0.105716 CPU  0.105501s TOTAL
>>>>>        useArrayAsync
<<<<<        useArrayAsync: 1.234e-3 CPU  0.001505s TOTAL
<<<<<       compileOpenAcc.traveuseAcc.use2: 0.108012 CPU  0.108015s TOTAL
<<<<<       compileOpenAcc.traveuseAcc.use3: 0.108539 CPU  0.108663s TOTAL
<<<<<      compileOpenAcc.traveuseAcc.Use: 0.109375 CPU  0.109005s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Fold1
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 0.0 CPU  0.000001s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 0.0 CPU  0s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 0.0 CPU  0.000001s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 0.0 CPU  0s TOTAL
<<<<<      compileOpenAcc.traveuseAcc.Fold1: 2.059e-3 CPU  0.002384s TOTAL
<<<<<      compileOpenAcc.traveuseAcc.Alet: 0.111434 CPU  0.112034s TOTAL
<<<<<     compileOpenAcc: 0.11197 CPU  0.112615s TOTAL
<<<<<    runAsyncIn.compileAcc: 0.11197 CPU  0.112833s TOTAL
>>>>>    runAsyncIn.dumpStats
<<<<<    runAsyncIn.dumpStats: 2.0e-6 CPU  0.000001s TOTAL
>>>>>    runAsyncIn.executeAcc
>>>>>     executeAcc
<<<<<     executeAcc: 8.96e-4 CPU  0.00049s TOTAL
<<<<<    runAsyncIn.executeAcc: 9.36e-4 CPU  0.0007s TOTAL
>>>>>    runAsyncIn.collect
<<<<<    runAsyncIn.collect: 0.0 CPU  0.000027s TOTAL
<<<<<   evalStateT: 0.114156 CPU  0.115327s TOTAL
>>>>>   pop
<<<<<   pop: 0.0 CPU  0.000002s TOTAL
>>>>>   performGC
<<<<<   performGC: 5.7246e-2 CPU  0.057814s TOTAL
<<<<<  evalCUDA: 0.17295 CPU  0.173943s TOTAL
<<<<< runAsyncIn.execute: 0.215475 CPU  0.216563s TOTAL
<<<<< run: 0.215523 CPU  0.216771s TOTAL
Array (Z) [1000001.0]
0.217148s
Prelude Main> Leaving GHCi.

已编译的代码

$ ghc -O2 -threaded Main.hs && ./Main
[1 of 1] Compiling Main             ( Main.hs, Main.o )
Linking Main ...
>>>>> run
>>>>> runAsyncIn.execute
>>>>>  runAsyncIn.seq ctx
<<<<<  runAsyncIn.seq ctx: 4.0639e-2 CPU  0.041498s TOTAL
>>>>>  runAsyncIn.seq a
<<<<<  runAsyncIn.seq a: 1.0e-6 CPU  0.000001s TOTAL
>>>>>  runAsyncIn.seq acc
>>>>>   convertAccWith True
<<<<<   convertAccWith: 1.2e-5 CPU  0.000005s TOTAL
<<<<<  runAsyncIn.seq acc: 1.15e-4 CPU  0.000061s TOTAL
>>>>>  evalCUDA
>>>>>   push
<<<<<   push: 2.0e-6 CPU  0.000002s TOTAL
>>>>>   evalStateT
>>>>>    runAsyncIn.compileAcc
>>>>>     compileOpenAcc
>>>>>      compileOpenAcc.traveuseAcc.Alet
>>>>>      compileOpenAcc.traveuseAcc.Use
>>>>>       compileOpenAcc.traveuseAcc.use3
>>>>>       compileOpenAcc.traveuseAcc.use1
<<<<<       compileOpenAcc.traveuseAcc.use1: 0.0 CPU  0.000001s TOTAL
>>>>>       compileOpenAcc.traveuseAcc.use2
>>>>>        compileOpenAcc.traveuseAcc.seq arr
<<<<<        compileOpenAcc.traveuseAcc.seq arr: 3.6651e-2 CPU  0.03712s TOTAL
>>>>>        useArrayAsync
<<<<<        useArrayAsync: 1.427e-3 CPU  0.001427s TOTAL
<<<<<       compileOpenAcc.traveuseAcc.use2: 3.8776e-2 CPU  0.039152s TOTAL
<<<<<       compileOpenAcc.traveuseAcc.use3: 3.8794e-2 CPU  0.039207s TOTAL
<<<<<      compileOpenAcc.traveuseAcc.Use: 3.8808e-2 CPU  0.03923s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Fold1
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 2.0e-6 CPU  0.000001s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 2.0e-6 CPU  0.000001s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 0.0 CPU  0.000001s TOTAL
>>>>>      compileOpenAcc.traveuseAcc.Avar
<<<<<      compileOpenAcc.traveuseAcc.Avar: 0.0 CPU  0.000001s TOTAL
<<<<<      compileOpenAcc.traveuseAcc.Fold1: 1.342e-3 CPU  0.001284s TOTAL
<<<<<      compileOpenAcc.traveuseAcc.Alet: 4.0197e-2 CPU  0.040578s TOTAL
<<<<<     compileOpenAcc: 4.0248e-2 CPU  0.040895s TOTAL
<<<<<    runAsyncIn.compileAcc: 4.0834e-2 CPU  0.04103s TOTAL
>>>>>    runAsyncIn.dumpStats
<<<<<    runAsyncIn.dumpStats: 0.0 CPU  0s TOTAL
>>>>>    runAsyncIn.executeAcc
>>>>>     executeAcc
<<<<<     executeAcc: 2.87e-4 CPU  0.000403s TOTAL
<<<<<    runAsyncIn.executeAcc: 2.87e-4 CPU  0.000488s TOTAL
>>>>>    runAsyncIn.collect
<<<<<    runAsyncIn.collect: 9.2e-5 CPU  0.000049s TOTAL
<<<<<   evalStateT: 4.1213e-2 CPU  0.041739s TOTAL
>>>>>   pop
<<<<<   pop: 0.0 CPU  0.000002s TOTAL
>>>>>   performGC
<<<<<   performGC: 9.41e-4 CPU  0.000861s TOTAL
<<<<<  evalCUDA: 4.3308e-2 CPU  0.042893s TOTAL
<<<<< runAsyncIn.execute: 8.5154e-2 CPU  0.084815s TOTAL
<<<<< run: 8.5372e-2 CPU  0.085035s TOTAL
Array (Z) [1000001.0]
0.085169s

正如我们所看到的,有两个主要问题:评估fromList (Z:.1000000) [1..1000000] :: Vector Double,其中 69 ms 额外需要ghci(106ms - 37ms),而performGC调用需要<强> 57 ms 额外(58 ms - 1 ms)。这两个总结了ghci下的执行和编译版本之间的区别。

我想,在编译程序中,RTS以与ghci不同的方式管理内存,因此分配和gc可以更快。我们也可以只测试这部分评估下面的代码(它根本不需要CUDA):

import Data.Array.Accelerate.Array.Sugar
import Data.Time.Clock                   (diffUTCTime, getCurrentTime)
import System.Mem                        (performGC)


main :: IO ()
main = do
    measure $ seq (fromList (Z:.1000000) [1..1000000] :: Vector Double) $ return ()
    measure $ performGC

measure action = do
    start <- getCurrentTime
    action
    end   <- getCurrentTime
    print $ diffUTCTime end start

<强>结果:

  • 评估向量在ghci和 0.035162s 下需要 0.121653s 编译版
  • performGC在ghci下使用 0.044876s 编译版本中 0.00031s

这可能是另一个问题,但也许有人知道:我们能否以某种方式调整垃圾收集器以便在ghci下更快地工作?