Aleagpu 如何让Alea更快?
在Alea中实现各种ML算法之后,我尝试在Alea中对一些简单但重要的例程进行基准测试。我很惊讶地发现,Alea“比同等的cuBLAS调用sgeam做同样的事情所花费的时间大约要长3倍。如果我在做一些更复杂的事情,比如矩阵乘法,我必须处理共享内存,这是可以理解的,但下面只是简单的数组转换Aleagpu 如何让Alea更快?,aleagpu,Aleagpu,在Alea中实现各种ML算法之后,我尝试在Alea中对一些简单但重要的例程进行基准测试。我很惊讶地发现,Alea“比同等的cuBLAS调用sgeam做同样的事情所花费的时间大约要长3倍。如果我在做一些更复杂的事情,比如矩阵乘法,我必须处理共享内存,这是可以理解的,但下面只是简单的数组转换 let dmat = createRandomUniformMatrix 100 1000 1.0f 0.0f let dmat2 = createRandomUniformMatrix 100 1000 1.
let dmat = createRandomUniformMatrix 100 1000 1.0f 0.0f
let dmat2 = createRandomUniformMatrix 100 1000 1.0f 0.0f
let rmat = createEmptyMatrixLike dmat
let m = new DeviceUnaryTransformModule<float32> <@ fun x -> x*2.0f @>
#time
//4.85s/100k
for i=1 to 100000 do
m.Apply(dmat, rmat) |> ignore
#time
#time
//1.8s/100k
for i=1 to 100000 do
sgeam2 nT nT 2.0f dmat 0.0f dmat2 rmat |> ignore
#time
让dmat=createRandomUniformMatrix 100 1000 1.0f 0.0f
设dmat2=createRandomUniformMatrix 100 1000 1.0f 0.0f
设rmat=CreateEmptyMatrix类dmat
设m=新设备一元转换模块x*2.0f@>
#时间
//4.85s/100k
对于i=1到100000 do
m、 应用(dmat,rmat)|>忽略
#时间
#时间
//1.8s/100k
对于i=1到100000 do
sgeam2 nT nT 2.0f dmat 0.0f dmat2 rmat |>忽略
#时间
DeviceUnaryTransformModule转换模块的内核与基本转换示例中的内核相同,唯一的区别是,之后它不再收集到主机,而是将数据保留在设备上
另外,Unbound的reduce对我来说效果很差,事实上效果很差,我使用它的方式肯定有错误。它大约比使用sgeamv两次求和矩阵慢20倍
let makeReduce (op:Expr<'T -> 'T -> 'T>) =
let compileReductionKernel (op:Expr<'T -> 'T -> 'T>) =
worker.LoadProgram(
DeviceReduceImpl.DeviceReduce(op, worker.Device.Arch, PlatformUtil.Instance.ProcessBitness).Template
)
let prog = compileReductionKernel op
let runReduceProgram (sumProg : Program<DeviceReduceImpl.IDeviceReduceFactory<'A>>) (x: DeviceMemory<'A>) =
sumProg.Entry.Create(blob, x.Length)
.Reduce(None, x.Ptr, x.Length)
let reduceProg (x: DeviceMemory<'T>) = runReduceProgram prog x
reduceProg
let sumReduce: DeviceMemory<float32> -> float32 = makeReduce <@ fun (a:float32) b -> a + b @>
#time
//3.5s/10k
for i=1 to 10000 do
sumReduce dmat.dArray |> ignore
#time
让makeReduce(op:Expr'T>)=
让编译器还原内核(op:Expr'T>)=
worker.LoadProgram(
DeviceReduceImpl.DeviceReduce(op,worker.Device.Arch,PlatformUtil.Instance.ProcessBitness).Template
)
设prog=compileReductionKernel op
让runReduceProgram(sumProg:Program)=
sumProg.Entry.Create(blob,x.Length)
.减少(无,x.Ptr,x.长度)
让reduceProg(x:DeviceMemory我认为您的测试代码中存在一些问题:
在您的映射模块中,您应该预加载GPUModule。GPUModule在第一次启动时是JIT编译的。因此,实际上您的计时测量包括GPU代码编译时间
在映射模块中,无论是Alea代码还是cublas代码,都应该同步工作者(同步CUDA上下文)。CUDA编程是异步风格的。因此,当您启动内核时,它会立即返回,而不等待内核完成。如果您不同步工作进程,实际上您是在测量内核启动时间,而不是内核执行时间。哪个Alea gpu的启动时间会比本机C代码慢,因为它会执行一些封送处理还有一些与内核启动时间相关的问题,我将在下面的示例代码中向您展示
您的reduce测试实际上每次都加载reduce模块!也就是说,每次执行reduce时,您都会测量时间,包括GPU编译时间!建议您将GPU模块或程序的实例设置为长寿命,因为它们代表已编译的GPU代码
因此,我按照您的用法做了一个测试。这里我首先列出了完整的测试代码:
#r @"packages\Alea.CUDA.2.1.2.3274\lib\net40\Alea.CUDA.dll"
#r @"packages\Alea.CUDA.IL.2.1.2.3274\lib\net40\Alea.CUDA.IL.dll"
#r @"packages\Alea.CUDA.Unbound.2.1.2.3274\lib\net40\Alea.CUDA.Unbound.dll"
#r "System.Configuration"
open System.IO
Alea.CUDA.Settings.Instance.Resource.AssemblyPath <- Path.Combine(@"packages\Alea.CUDA.2.1.2.3274", "private")
Alea.CUDA.Settings.Instance.Resource.Path <- Path.GetTempPath()
open Alea.CUDA
open Alea.CUDA.Utilities
open Alea.CUDA.CULib
open Alea.CUDA.Unbound
open Microsoft.FSharp.Quotations
type MapModule(target, op:Expr<float32 -> float32>) =
inherit GPUModule(target)
[<Kernel;ReflectedDefinition>]
member this.Kernel (C:deviceptr<float32>) (A:deviceptr<float32>) (B:deviceptr<float32>) (n:int) =
let start = blockIdx.x * blockDim.x + threadIdx.x
let stride = gridDim.x * blockDim.x
let mutable i = start
while i < n do
C.[i] <- __eval(op) A.[i] + __eval(op) B.[i]
i <- i + stride
member this.Apply(C:deviceptr<float32>, A:deviceptr<float32>, B:deviceptr<float32>, n:int) =
let lp = LaunchParam(64, 256)
this.GPULaunch <@ this.Kernel @> lp C A B n
let inline mapTemplate (op:Expr<'T -> 'T>) = cuda {
let! kernel =
<@ fun (C:deviceptr<'T>) (A:deviceptr<'T>) (B:deviceptr<'T>) (n:int) ->
let start = blockIdx.x * blockDim.x + threadIdx.x
let stride = gridDim.x * blockDim.x
let mutable i = start
while i < n do
C.[i] <- (%op) A.[i] + (%op) B.[i]
i <- i + stride @>
|> Compiler.DefineKernel
return Entry(fun program ->
let worker = program.Worker
let kernel = program.Apply kernel
let lp = LaunchParam(64, 256)
let run C A B n =
kernel.Launch lp C A B n
run ) }
let test1 (worker:Worker) m n sync iters =
let n = m * n
use m = new MapModule(GPUModuleTarget.Worker(worker), <@ fun x -> x * 2.0f @>)
let rng = System.Random(42)
use A = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use B = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use C = worker.Malloc<float32>(n)
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to iters do
m.Apply(C.Ptr, A.Ptr, B.Ptr, n)
if sync then worker.Synchronize()
timer.Stop()
printfn "%f ms / %d %s (no pre-load module)" timer.Elapsed.TotalMilliseconds iters (if sync then "sync" else "nosync")
let test2 (worker:Worker) m n sync iters =
let n = m * n
use m = new MapModule(GPUModuleTarget.Worker(worker), <@ fun x -> x * 2.0f @>)
// we pre-load the module, this will JIT compile the GPU code
m.GPUForceLoad()
let rng = System.Random(42)
use A = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use B = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use C = worker.Malloc<float32>(n)
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to iters do
m.Apply(C.Ptr, A.Ptr, B.Ptr, n)
if sync then worker.Synchronize()
timer.Stop()
printfn "%f ms / %d %s (pre-loaded module)" timer.Elapsed.TotalMilliseconds iters (if sync then "sync" else "nosync")
let test3 (worker:Worker) m n sync iters =
let n = m * n
use m = new MapModule(GPUModuleTarget.Worker(worker), <@ fun x -> x * 2.0f @>)
// we pre-load the module, this will JIT compile the GPU code
m.GPUForceLoad()
let rng = System.Random(42)
use A = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use B = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use C = worker.Malloc<float32>(n)
// since the worker is running in a background thread
// each cuda api will switch to that thread
// use eval() to avoid the many thread switching
worker.Eval <| fun _ ->
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to iters do
m.Apply(C.Ptr, A.Ptr, B.Ptr, n)
if sync then worker.Synchronize()
timer.Stop()
printfn "%f ms / %d %s (pre-loaded module + worker.eval)" timer.Elapsed.TotalMilliseconds iters (if sync then "sync" else "nosync")
let test4 (worker:Worker) m n sync iters =
use program = worker.LoadProgram(mapTemplate <@ fun x -> x * 2.0f @>)
let n = m * n
let rng = System.Random(42)
use A = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use B = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use C = worker.Malloc<float32>(n)
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to iters do
program.Run C.Ptr A.Ptr B.Ptr n
if sync then worker.Synchronize()
timer.Stop()
printfn "%f ms / %d %s (template usage)" timer.Elapsed.TotalMilliseconds iters (if sync then "sync" else "nosync")
let test5 (worker:Worker) m n sync iters =
use program = worker.LoadProgram(mapTemplate <@ fun x -> x * 2.0f @>)
let n = m * n
let rng = System.Random(42)
use A = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use B = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use C = worker.Malloc<float32>(n)
worker.Eval <| fun _ ->
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to iters do
program.Run C.Ptr A.Ptr B.Ptr n
if sync then worker.Synchronize()
timer.Stop()
printfn "%f ms / %d %s (template usage + worker.Eval)" timer.Elapsed.TotalMilliseconds iters (if sync then "sync" else "nosync")
let test6 (worker:Worker) m n sync iters =
use cublas = new CUBLAS(worker)
let rng = System.Random(42)
use dmat1 = worker.Malloc(Array.init (m * n) (fun _ -> rng.NextDouble() |> float32))
use dmat2 = worker.Malloc(Array.init (m * n) (fun _ -> rng.NextDouble() |> float32))
use dmatr = worker.Malloc<float32>(m * n)
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to iters do
cublas.Sgeam(cublasOperation_t.CUBLAS_OP_N, cublasOperation_t.CUBLAS_OP_N, m, n, 2.0f, dmat1.Ptr, m, 2.0f, dmat2.Ptr, m, dmatr.Ptr, m)
if sync then worker.Synchronize()
timer.Stop()
printfn "%f ms / %d %s (cublas)" timer.Elapsed.TotalMilliseconds iters (if sync then "sync" else "nosync")
let test7 (worker:Worker) m n sync iters =
use cublas = new CUBLAS(worker)
let rng = System.Random(42)
use dmat1 = worker.Malloc(Array.init (m * n) (fun _ -> rng.NextDouble() |> float32))
use dmat2 = worker.Malloc(Array.init (m * n) (fun _ -> rng.NextDouble() |> float32))
use dmatr = worker.Malloc<float32>(m * n)
worker.Eval <| fun _ ->
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to iters do
cublas.Sgeam(cublasOperation_t.CUBLAS_OP_N, cublasOperation_t.CUBLAS_OP_N, m, n, 2.0f, dmat1.Ptr, m, 2.0f, dmat2.Ptr, m, dmatr.Ptr, m)
if sync then worker.Synchronize()
timer.Stop()
printfn "%f ms / %d %s (cublas + worker.eval)" timer.Elapsed.TotalMilliseconds iters (if sync then "sync" else "nosync")
let test worker m n sync iters =
test6 worker m n sync iters
test7 worker m n sync iters
test1 worker m n sync iters
test2 worker m n sync iters
test3 worker m n sync iters
test4 worker m n sync iters
test5 worker m n sync iters
let testReduce1 (worker:Worker) n iters =
let rng = System.Random(42)
use input = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use reduceModule = new DeviceReduceModule<float32>(GPUModuleTarget.Worker(worker), <@ (+) @>)
// JIT compile and load GPU code for this module
reduceModule.GPUForceLoad()
// create a reducer which will allocate temp memory for maxNum=n
let reduce = reduceModule.Create(n)
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to 10000 do
reduce.Reduce(input.Ptr, n) |> ignore
timer.Stop()
printfn "%f ms / %d (pre-load gpu code)" timer.Elapsed.TotalMilliseconds iters
let testReduce2 (worker:Worker) n iters =
let rng = System.Random(42)
use input = worker.Malloc(Array.init n (fun _ -> rng.NextDouble() |> float32))
use reduceModule = new DeviceReduceModule<float32>(GPUModuleTarget.Worker(worker), <@ (+) @>)
// JIT compile and load GPU code for this module
reduceModule.GPUForceLoad()
// create a reducer which will allocate temp memory for maxNum=n
let reduce = reduceModule.Create(n)
worker.Eval <| fun _ ->
let timer = System.Diagnostics.Stopwatch.StartNew()
for i = 1 to 10000 do
reduce.Reduce(input.Ptr, n) |> ignore
timer.Stop()
printfn "%f ms / %d (pre-load gpu code and avoid thread switching)" timer.Elapsed.TotalMilliseconds iters
let testReduce worker n iters =
testReduce1 worker n iters
testReduce2 worker n iters
let workerDefault = Worker.Default
let workerNoThread = Worker.CreateOnCurrentThread(Device.Default)
另外,您可能会注意到,我运行了两次此测试,第一次,没有预加载模块的测试使用304毫秒,但第二次,没有预加载模块的测试只使用29毫秒。原因是,我们使用LLVM p/Invoke编译内核。而这些p/Invoke函数是惰性函数,因此在fir时会进行一些初始化使用它之后,它会变得更快
现在,我们同步worker,它实际上测量了实际内核执行时间,现在它们是相似的。我在这里创建的内核非常简单,但它对矩阵A和B都有作用:
> test workerDefault 10000 10000 true 100;;
843.695000 ms / 100 sync (cublas)
841.452400 ms / 100 sync (cublas + worker.eval)
919.244900 ms / 100 sync (no pre-load module)
912.348000 ms / 100 sync (pre-loaded module)
908.909000 ms / 100 sync (pre-loaded module + worker.eval)
914.834100 ms / 100 sync (template usage)
914.170100 ms / 100 sync (template usage + worker.Eval)
现在,如果我们在threadless worker上测试它们,它们会有点快,因为没有线程切换:
> test workerNoThread 10000 10000 true 100;;
842.132100 ms / 100 sync (cublas)
841.627200 ms / 100 sync (cublas + worker.eval)
918.007800 ms / 100 sync (no pre-load module)
908.575900 ms / 100 sync (pre-loaded module)
908.770100 ms / 100 sync (pre-loaded module + worker.eval)
913.405300 ms / 100 sync (template usage)
913.942600 ms / 100 sync (template usage + worker.Eval)
下面是关于减少的测试:
> testReduce workerDefault 10000000 100;;
7691.335300 ms / 100 (pre-load gpu code)
6448.782500 ms / 100 (pre-load gpu code and avoid thread switching)
val it : unit = ()
> testReduce workerNoThread 10000000 100;;
6467.105300 ms / 100 (pre-load gpu code)
6426.296900 ms / 100 (pre-load gpu code and avoid thread switching)
val it : unit = ()
请注意,在这个缩减测试中,有一个内存收集(memcpyDtoH)对于每次缩减,都要将结果从设备复制到主机。而此内存复制API调用会自动同步工作进程,因为如果内核未完成,该值将毫无意义。因此,如果要将性能与C代码进行比较,还应将结果标量从设备复制到主机。虽然这只是一次CUDA API调用,但如果您进行了多次迭代(在本例中为100次),那么它将在那里累积一些时间
希望这能回答您的问题。顺便说一句,当前的alea gpu使用CUDA 6.0的nvidia llvm编译器。我们将发布一个新版本,在CUDA 7.5发布时将其升级到CUDA 7.5。根据我现在观察到的情况,这次升级提高了生成内核的性能。这是一个很好的答案。我实际上知道CUDA有异步内核默认情况下,UCH在C++中,但我假设阿莱亚中的不同。这是因为我不知道异步内核执行有点像函数语言中的懒惰评估。我认为这是另外一回事。谢谢。
> testReduce workerDefault 10000000 100;;
7691.335300 ms / 100 (pre-load gpu code)
6448.782500 ms / 100 (pre-load gpu code and avoid thread switching)
val it : unit = ()
> testReduce workerNoThread 10000000 100;;
6467.105300 ms / 100 (pre-load gpu code)
6426.296900 ms / 100 (pre-load gpu code and avoid thread switching)
val it : unit = ()