DotCompute RC2, bu kavramlar\u0131 .NET geli\u015ftiricileri i\u00e7in eri\u015filebilir hale getirirken, ayn\u0131 zamanda d\u00fc\u015f\u00fck seviyeli API’lerin sundu\u011fu g\u00fcc\u00fc ve esnekli\u011fi de korur. B\u00f6ylece, hem h\u0131zl\u0131 prototipleme yapabilir hem de kritik performans gereksinimleri olan uygulamalar i\u00e7in detayl\u0131 optimizasyonlar uygulayabilirsiniz. K\u00fct\u00fcphane, .NET’in tan\u0131d\u0131k s\u00f6zdizimi ve geli\u015ftirme ortam\u0131 ile GPU g\u00fcc\u00fcn\u00fc birle\u015ftirerek, .NET ekosistemindeki y\u00fcksek performansl\u0131 hesaplama eksikli\u011fini giderir.<\/p>\n
DotCompute RC2 ile \u0130lk Ad\u0131mlar: Basit Bir Vekt\u00f6r Toplama Nas\u0131l Yap\u0131l\u0131r?<\/h2>\n
\u015eimdi DotCompute RC2’yi kullanarak pratik bir \u00f6rnek \u00fczerinde \u00e7al\u0131\u015fal\u0131m: iki vekt\u00f6r\u00fcn eleman baz\u0131nda toplanmas\u0131. Bu basit ama temel \u00f6rnek, k\u00fct\u00fcphanenin nas\u0131l kuruldu\u011funu, bir kernel’\u0131n nas\u0131l yaz\u0131ld\u0131\u011f\u0131n\u0131 ve GPU \u00fczerinde nas\u0131l \u00e7al\u0131\u015ft\u0131r\u0131ld\u0131\u011f\u0131n\u0131 anlamak i\u00e7in harika bir ba\u015flang\u0131\u00e7 noktas\u0131d\u0131r. Uygulamam\u0131z\u0131 Visual Studio’da yeni bir Konsol Uygulamas\u0131 (.NET Core veya .NET 6+) olarak olu\u015fturdu\u011funuzu varsayal\u0131m.<\/p>\n
Ad\u0131m 1: NuGet Paketlerini Y\u00fckleme<\/h3>\n
\u0130lk olarak, projemize DotCompute RC2 NuGet paketlerini eklememiz gerekiyor. Temel olarak <\/code><\/p>\n GPU'da \u00e7al\u0131\u015facak olan kernel'\u0131m\u0131z\u0131 OpenCL C dilinde yazaca\u011f\u0131z. Bu kernel, her bir i\u015f par\u00e7ac\u0131\u011f\u0131na bir indeks atayacak ve bu indekse kar\u015f\u0131l\u0131k gelen vekt\u00f6r elemanlar\u0131n\u0131 toplayacakt\u0131r.<\/p>\n <\/code><\/p>\n Yukar\u0131daki kernel kodu, \u015eimdi s\u0131ra geldi bu kernel'\u0131 C# uygulamam\u0131zdan \u00e7al\u0131\u015ft\u0131rmaya. A\u015fa\u011f\u0131daki kod blo\u011fu, DotCompute RC2'nin nas\u0131l ba\u015flat\u0131ld\u0131\u011f\u0131n\u0131, cihazlar\u0131n nas\u0131l se\u00e7ildi\u011fini, bellek buffer'lar\u0131n\u0131n nas\u0131l olu\u015fturuldu\u011funu ve kernel'\u0131n nas\u0131l y\u00fcr\u00fct\u00fcld\u00fc\u011f\u00fcn\u00fc g\u00f6steriyor.<\/p>\n <\/code><\/p>\n Bu \u00f6rnek, DotCompute RC2'nin temel i\u015f ak\u0131\u015f\u0131n\u0131 net bir \u015fekilde ortaya koymaktad\u0131r. Bir compute context olu\u015fturarak ba\u015flar, uygun bir cihaz se\u00e7er, kernel kodunu derler, veri buffer'lar\u0131n\u0131 haz\u0131rlar, kernel parametrelerini ayarlar ve son olarak kernel'\u0131 \u00e7al\u0131\u015ft\u0131r\u0131r. Sonu\u00e7lar geri okunduktan sonra do\u011frulama yap\u0131l\u0131r. Bu ad\u0131mlar, daha karma\u015f\u0131k GPU tabanl\u0131 uygulamalar i\u00e7in de temel bir \u015fablon sunar.<\/p>\n G\u00f6r\u00fcnt\u00fc i\u015fleme, GPU'lar\u0131n paralel hesaplama g\u00fcc\u00fcnden en \u00e7ok faydalan\u0131lan alanlardan biridir. Her pikselin ba\u011f\u0131ms\u0131z olarak veya kom\u015fu piksellerle ili\u015fkili olarak i\u015flenebildi\u011fi bu t\u00fcr g\u00f6revlerde GPU'lar, CPU'lara g\u00f6re kat kat daha h\u0131zl\u0131 sonu\u00e7lar \u00fcretebilir. Basit bir g\u00f6r\u00fcnt\u00fc \u00fczerinde gri tonlama filtresi uygulamay\u0131 ele alal\u0131m. Bu i\u015flem, her pikselin RGB de\u011ferlerini al\u0131p, belirli bir form\u00fcle g\u00f6re tek bir gri tonlama de\u011ferine d\u00f6n\u00fc\u015ft\u00fcrmekten ibarettir.<\/p>\n Geleneksel olarak, bu i\u015flemi CPU \u00fczerinde yapmak, g\u00f6r\u00fcnt\u00fcn\u00fcn her bir pikseli i\u00e7in d\u00f6ng\u00fclerle tek tek i\u015flemeyi gerektirir. Y\u00fcksek \u00e7\u00f6z\u00fcn\u00fcrl\u00fckl\u00fc bir g\u00f6r\u00fcnt\u00fcde (\u00f6rne\u011fin 4K), milyonlarca pikselin her biri i\u00e7in bu i\u015flemi yapmak \u00f6nemli zaman alabilir. Ancak bir GPU, binlerce pikseli ayn\u0131 anda i\u015fleyebilir, bu da i\u015flem s\u00fcresini dramatik bir \u015fekilde k\u0131salt\u0131r. DotCompute RC2 sayesinde, bu t\u00fcr bir paralel i\u015flemi .NET uygulaman\u0131zda kolayca ger\u00e7ekle\u015ftirebilirsiniz.<\/p>\n A\u015fa\u011f\u0131daki OpenCL kernel'\u0131, bir g\u00f6r\u00fcnt\u00fcn\u00fcn piksel verilerini (RGBA format\u0131nda varsay\u0131yoruz) al\u0131p, her pikseli gri tonlamaya \u00e7evirir. Gri tonlama form\u00fcl\u00fc genellikle <\/code><\/p>\n Bu kernel'da \u015eimdi bu kernel'\u0131 .NET uygulamam\u0131zdan nas\u0131l \u00e7a\u011f\u0131raca\u011f\u0131m\u0131z\u0131 g\u00f6relim. Basitlik ad\u0131na, bir g\u00f6r\u00fcnt\u00fcden piksel verilerini do\u011frudan bir byte dizisine okudu\u011fumuzu ve ard\u0131ndan bu diziyi DotCompute RC2'ye aktard\u0131\u011f\u0131m\u0131z\u0131 varsayal\u0131m. Ger\u00e7ek bir uygulamada, <\/code><\/p>\n Bu \u00f6rnekte, DotCompute RC2, temel GPU h\u0131zland\u0131rmas\u0131n\u0131 basit bir \u015fekilde sa\u011flasa da, ger\u00e7ekten y\u00fcksek performansl\u0131 uygulamalar geli\u015ftirmek i\u00e7in baz\u0131 ileri d\u00fczey teknikleri ve optimizasyon stratejilerini anlamak ve uygulamak kritik \u00f6neme sahiptir. GPU'lar g\u00fc\u00e7l\u00fcd\u00fcr, ancak onlar\u0131n mimarisini anlamadan rastgele kod yazmak, beklenen performans art\u0131\u015f\u0131n\u0131 sa\u011flamayabilir.<\/p>\n GPU'lar\u0131n en b\u00fcy\u00fck darbo\u011fazlar\u0131ndan biri, CPU ve GPU aras\u0131ndaki veri transferidir (PCIe veri yolu \u00fczerinden). Bu transferler, genellikle kernel \u00e7al\u0131\u015ft\u0131rma s\u00fcresinden daha uzun s\u00fcrebilir. Bu nedenle:<\/p>\n Kernel kodunuzun kendisi, GPU performans\u0131n\u0131n temelidir. \u0130yi optimize edilmi\u015f bir kernel, k\u00f6t\u00fc yaz\u0131lm\u0131\u015f bir kernel'dan kat kat daha h\u0131zl\u0131 \u00e7al\u0131\u015fabilir:<\/p>\n CPU'nuzun GPU'nun i\u015fini bitirmesini beklemesini engellemek i\u00e7in asenkron programlama tekniklerini kullan\u0131n:<\/p>\n <\/code><\/p>\nDotCompute.Core<\/code> ve kullanmak istedi\u011fimiz backend paketleri gereklidir. Bu \u00f6rnekte OpenCL backend’ini kullanaca\u011f\u0131z, ancak sisteminizde NVIDIA GPU varsa DotCompute.Cuda<\/code>‘y\u0131 da ekleyebilirsiniz.<\/p>\n\n Install-Package DotCompute.Core\n Install-Package DotCompute.OpenCL\n <\/pre>\nAd\u0131m 2: Kernel Kodunu Yazma<\/h3>\n
\n const string vectorAddKernel = @\"\n __kernel void VectorAdd(__global const float* a,\n __global const float* b,\n __global float* result,\n const int n)\n {\n int gid = get_global_id(0); \/\/ \u0130\u015f par\u00e7ac\u0131\u011f\u0131n\u0131n global indeksini al\n\n if (gid < n) {\n result[gid] = a[gid] + b[gid]; \/\/ \u0130lgili elemanlar\u0131 topla\n }\n }\n \";\n <\/pre>\na<\/code> ve b<\/code> adl\u0131 iki girdi vekt\u00f6r\u00fcn\u00fc al\u0131p, sonu\u00e7lar\u0131 result<\/code> vekt\u00f6r\u00fcne yaz\u0131yor. n<\/code> ise vekt\u00f6rlerin boyutunu belirtiyor. get_global_id(0)<\/code>, her bir i\u015f par\u00e7ac\u0131\u011f\u0131n\u0131n kendi benzersiz indeksini almas\u0131n\u0131 sa\u011fl\u0131yor.<\/p>\nAd\u0131m 3: C# Host Kodu Yazma<\/h3>\n
\n using DotCompute;\n using DotCompute.OpenCL;\n using System;\n using System.Linq;\n\n public class Program\n {\n public static void Main(string[] args)\n {\n const int N = 1_000_000; \/\/ Vekt\u00f6r boyutu\n float[] a = Enumerable.Range(0, N).Select(i => (float)i).ToArray();\n float[] b = Enumerable.Range(0, N).Select(i => (float)(i * 2)).ToArray();\n float[] result = new float[N];\n\n using var computeContext = new ComputeContext(new OpenCLBackend()); \/\/ OpenCL backend ile bir context olu\u015ftur\n \n \/\/ Mevcut cihazlar\u0131 listele ve ilkini se\u00e7\n var device = computeContext.Devices.FirstOrDefault(d => d.DeviceType == ComputeDeviceType.GPU);\n if (device == null)\n {\n Console.WriteLine(\"GPU cihaz\u0131 bulunamad\u0131, varsay\u0131lan CPU kullan\u0131lacak.\");\n device = computeContext.Devices.FirstOrDefault(d => d.DeviceType == ComputeDeviceType.CPU); \/\/ GPU yoksa CPU dene\n if (device == null)\n {\n Console.WriteLine(\"Hesaplama cihaz\u0131 bulunamad\u0131. \u00c7\u0131k\u0131l\u0131yor.\");\n return;\n }\n }\n Console.WriteLine($\"Kullan\u0131lan Cihaz: {device.Name} ({device.DeviceType})\");\n\n using var computeDevice = computeContext.CreateDevice(device);\n using var queue = computeDevice.CreateCommandQueue();\n\n \/\/ Giri\u015f ve \u00c7\u0131k\u0131\u015f bellek buffer'lar\u0131 olu\u015ftur\n using var bufferA = computeDevice.CreateBuffer(a.Length * sizeof(float), ComputeBufferFlags.ReadWrite);\n using var bufferB = computeDevice.CreateBuffer(b.Length * sizeof(float), ComputeBufferFlags.ReadWrite);\n using var bufferResult = computeDevice.CreateBuffer(N * sizeof(float), ComputeBufferFlags.ReadWrite);\n\n \/\/ Verileri CPU'dan GPU'ya kopyala\n queue.WriteBuffer(bufferA, a);\n queue.WriteBuffer(bufferB, b);\n\n \/\/ Kernel'\u0131 derle ve parametrelerini ayarla\n using var program = computeDevice.CreateProgram(vectorAddKernel);\n using var kernel = program.CreateKernel(\"VectorAdd\");\n kernel.SetArgument(0, bufferA);\n kernel.SetArgument(1, bufferB);\n kernel.SetArgument(2, bufferResult);\n kernel.SetArgument(3, N);\n\n \/\/ Kernel'\u0131 \u00e7al\u0131\u015ft\u0131r\n queue.Execute(kernel, N); \/\/ N global boyutunda \u00e7al\u0131\u015ft\u0131r\n queue.Finish(); \/\/ T\u00fcm i\u015flemlerin tamamlanmas\u0131n\u0131 bekle\n\n \/\/ Sonu\u00e7lar\u0131 GPU'dan CPU'ya geri kopyala\n queue.ReadBuffer(bufferResult, result);\n\n \/\/ Sonu\u00e7lar\u0131 do\u011frula\n Console.WriteLine($\"\u0130lk 5 sonu\u00e7: {string.Join(\", \", result.Take(5))}\");\n Console.WriteLine($\"Sonu\u00e7 do\u011frulama (a[0]+b[0]): {a[0] + b[0]}\");\n Console.WriteLine($\"Sonu\u00e7 do\u011frulama (a[N-1]+b[N-1]): {a[N - 1] + b[N - 1]}\");\n \n \/\/ K\u00fc\u00e7\u00fck bir do\u011frulama yapal\u0131m\n bool success = true;\n for (int i = 0; i < N; i++)\n {\n if (result[i] != a[i] + b[i])\n {\n success = false;\n Console.WriteLine($\"Hata! Index {i}: Beklenen {a[i] + b[i]}, Al\u0131nan {result[i]}\");\n break;\n }\n }\n\n if (success)\n {\n Console.WriteLine(\"Vekt\u00f6r toplama i\u015flemi ba\u015far\u0131yla tamamland\u0131 ve do\u011fruland\u0131.\");\n }\n else\n {\n Console.WriteLine(\"Vekt\u00f6r toplama i\u015fleminde hatalar bulundu.\");\n }\n }\n }\n <\/pre>\nComputeBufferFlags.Pinned<\/code> gibi bayraklar\u0131 kullanarak pinli (sayfa kilitli) bellekten faydalanabilir, bu da CPU ile GPU aras\u0131ndaki veri transfer h\u0131z\u0131n\u0131 art\u0131rabilir.\n <\/div>\nGer\u00e7ek D\u00fcnya Senaryosu: G\u00f6r\u00fcnt\u00fc \u0130\u015flemede DotCompute RC2 Nas\u0131l Kullan\u0131l\u0131r?<\/h3>\n
Ad\u0131m 1: Gri Tonlama Kernel'\u0131 Yazma<\/h3>\n
(R*0.299 + G*0.587 + B*0.114)<\/code> \u015feklindedir.<\/p>\n\n const string grayscaleKernel = @\"\n __kernel void Grayscale(__global unsigned char* inputPixels,\n __global unsigned char* outputPixels,\n const int width,\n const int height)\n {\n int x = get_global_id(0); \/\/ X koordinat\u0131\n int y = get_global_id(1); \/\/ Y koordinat\u0131\n\n if (x < width && y < height)\n {\n int index = (y * width + x) * 4; \/\/ RGBA oldu\u011fu i\u00e7in 4 ile \u00e7arp\u0131yoruz\n\n unsigned char r = inputPixels[index + 0];\n unsigned char g = inputPixels[index + 1];\n unsigned char b = inputPixels[index + 2];\n \/\/ alpha kanal\u0131n\u0131 koruyoruz: inputPixels[index + 3];\n\n \/\/ Gri tonlama form\u00fcl\u00fc\n unsigned char gray = (unsigned char)(0.299f * r + 0.587f * g + 0.114f * b);\n\n outputPixels[index + 0] = gray;\n outputPixels[index + 1] = gray;\n outputPixels[index + 2] = gray;\n outputPixels[index + 3] = inputPixels[index + 3]; \/\/ Alpha kanal\u0131n\u0131 koru\n }\n }\n \";\n <\/pre>\nget_global_id(0)<\/code> ve get_global_id(1)<\/code> kullan\u0131larak 2 boyutlu bir \u00e7al\u0131\u015fma alan\u0131 tan\u0131ml\u0131yoruz, bu da bir g\u00f6r\u00fcnt\u00fc matrisine do\u011frudan e\u015flenebilir. Her i\u015f par\u00e7ac\u0131\u011f\u0131, g\u00f6r\u00fcnt\u00fcn\u00fcn belirli bir pikselinin koordinatlar\u0131na kar\u015f\u0131l\u0131k gelir.<\/p>\nAd\u0131m 2: C# Host Kodu ile Entegrasyon<\/h3>\n
System.Drawing.Bitmap<\/code> veya ImageSharp<\/code> gibi k\u00fct\u00fcphanelerle g\u00f6r\u00fcnt\u00fc verilerini okuyup yazman\u0131z gerekecektir.<\/p>\n\n using DotCompute;\n using DotCompute.OpenCL;\n using System;\n using System.Diagnostics;\n using System.Linq;\n\n public class ImageProcessingExample\n {\n \/\/ ... (grayscaleKernel string tan\u0131mlamas\u0131 yukar\u0131daki gibi) ...\n\n public static void RunGrayscaleExample()\n {\n \/\/ \u00d6rnek bir 1024x768 boyutunda RGBA piksel verisi olu\u015ftural\u0131m (siyah bir g\u00f6r\u00fcnt\u00fc)\n const int width = 1024;\n const int height = 768;\n const int channels = 4; \/\/ RGBA\n int imageSize = width * height * channels;\n byte[] inputPixels = new byte[imageSize];\n byte[] outputPixels = new byte[imageSize];\n\n \/\/ Test i\u00e7in her pikseli rastgele bir renge boyayal\u0131m (sadece k\u0131rm\u0131z\u0131 ve ye\u015fili aktif edelim)\n var random = new Random();\n for (int i = 0; i < imageSize; i += channels)\n {\n inputPixels[i + 0] = (byte)random.Next(256); \/\/ R\n inputPixels[i + 1] = (byte)random.Next(256); \/\/ G\n inputPixels[i + 2] = 0; \/\/ B\n inputPixels[i + 3] = 255; \/\/ A\n }\n\n using var computeContext = new ComputeContext(new OpenCLBackend());\n var device = computeContext.Devices.FirstOrDefault(d => d.DeviceType == ComputeDeviceType.GPU);\n if (device == null)\n {\n Console.WriteLine(\"GPU cihaz\u0131 bulunamad\u0131. G\u00f6r\u00fcnt\u00fc i\u015fleme \u00f6rne\u011fi atlan\u0131yor.\");\n return;\n }\n Console.WriteLine($\"G\u00f6r\u00fcnt\u00fc \u0130\u015fleme \u0130\u00e7in Kullan\u0131lan Cihaz: {device.Name} ({device.DeviceType})\");\n\n using var computeDevice = computeContext.CreateDevice(device);\n using var queue = computeDevice.CreateCommandQueue();\n\n using var inputBuffer = computeDevice.CreateBuffer(imageSize, ComputeBufferFlags.ReadWrite);\n using var outputBuffer = computeDevice.CreateBuffer(imageSize, ComputeBufferFlags.ReadWrite);\n\n \/\/ Verileri GPU'ya kopyala\n queue.WriteBuffer(inputBuffer, inputPixels);\n\n using var program = computeDevice.CreateProgram(grayscaleKernel);\n using var kernel = program.CreateKernel(\"Grayscale\");\n kernel.SetArgument(0, inputBuffer);\n kernel.SetArgument(1, outputBuffer);\n kernel.SetArgument(2, width);\n kernel.SetArgument(3, height);\n\n \/\/ Kernel'\u0131 2 boyutlu bir global boyutla \u00e7al\u0131\u015ft\u0131r\n \/\/ Her piksel i\u00e7in bir i\u015f par\u00e7ac\u0131\u011f\u0131\n queue.Execute(kernel, new long[] { width, height }); \n queue.Finish();\n\n \/\/ Sonu\u00e7lar\u0131 GPU'dan CPU'ya geri kopyala\n queue.ReadBuffer(outputBuffer, outputPixels);\n\n Console.WriteLine(\"Gri tonlama i\u015flemi ba\u015far\u0131yla tamamland\u0131.\");\n\n \/\/ OutputPixels dizisi \u015fimdi gri tonlamal\u0131 g\u00f6r\u00fcnt\u00fc verilerini i\u00e7eriyor.\n \/\/ Bu veriyi bir dosyaya kaydedebilir veya g\u00f6r\u00fcnt\u00fcleyicide g\u00f6sterebilirsiniz.\n \/\/ \u00d6rne\u011fin, ilk pikselin (0,0) gri tonlama de\u011ferini kontrol edelim.\n Console.WriteLine($\"Orijinal Piksel (0,0) RGBA: ({inputPixels[0]}, {inputPixels[1]}, {inputPixels[2]}, {inputPixels[3]})\");\n Console.WriteLine($\"Gri Tonlama Piksel (0,0) RGBA: ({outputPixels[0]}, {outputPixels[1]}, {outputPixels[2]}, {outputPixels[3]})\");\n \n \/\/ Performans kar\u015f\u0131la\u015ft\u0131rmas\u0131 i\u00e7in CPU \u00fczerinde ayn\u0131 i\u015flemi yapal\u0131m\n byte[] cpuOutputPixels = new byte[imageSize];\n var stopwatch = Stopwatch.StartNew();\n for (int y = 0; y < height; y++)\n {\n for (int x = 0; x < width; x++)\n {\n int index = (y * width + x) * 4;\n byte r = inputPixels[index + 0];\n byte g = inputPixels[index + 1];\n byte b = inputPixels[index + 2];\n byte a = inputPixels[index + 3];\n\n byte gray = (byte)(0.299f * r + 0.587f * g + 0.114f * b);\n\n cpuOutputPixels[index + 0] = gray;\n cpuOutputPixels[index + 1] = gray;\n cpuOutputPixels[index + 2] = gray;\n cpuOutputPixels[index + 3] = a;\n }\n }\n stopwatch.Stop();\n Console.WriteLine($\"CPU \u00fczerinde gri tonlama s\u00fcresi: {stopwatch.ElapsedMilliseconds} ms\");\n\n \/\/ GPU s\u00fcresini \u00f6l\u00e7mek i\u00e7in CommandQueue.Execute'un hemen \u00f6ncesi ve sonras\u0131na ekleme yapabilirsiniz\n \/\/ Bu \u00f6rnekte basitle\u015ftirilmi\u015f bir genel s\u00fcre \u00f6l\u00e7\u00fcm\u00fc yap\u0131lm\u0131\u015ft\u0131r.\n \/\/ Daha hassas \u00f6l\u00e7\u00fcmler i\u00e7in DotCompute.Profiling<\/code> ara\u00e7lar\u0131 kullan\u0131labilir.\n }\n }\n <\/pre>\nqueue.Execute(kernel, new long[] { width, height });<\/code> sat\u0131r\u0131, kernel'\u0131n iki boyutlu bir i\u015f alan\u0131nda \u00e7al\u0131\u015ft\u0131r\u0131ld\u0131\u011f\u0131n\u0131 belirtir. Her pikselin i\u015flenmesi i\u00e7in bir i\u015f par\u00e7ac\u0131\u011f\u0131 atan\u0131r. Bu yakla\u015f\u0131m, g\u00f6r\u00fcnt\u00fc i\u015fleme, matris manip\u00fclasyonlar\u0131 ve di\u011fer grid tabanl\u0131 hesaplamalar i\u00e7in son derece etkilidir. G\u00f6r\u00fcld\u00fc\u011f\u00fc gibi, DotCompute RC2 ile GPU h\u0131zland\u0131rmal\u0131 g\u00f6r\u00fcnt\u00fc i\u015fleme yapmak, sadece birka\u00e7 sat\u0131r ek C# ve kernel kodu ile m\u00fcmk\u00fcn olmaktad\u0131r.<\/p>\n__local<\/code> bellek kullan\u0131m\u0131n\u0131 optimize ederek veya veri eri\u015fim modellerinizi GPU'nun bellek mimarisine uygun hale getirerek \u00f6nemli h\u0131z art\u0131\u015flar\u0131 elde edebilirsiniz. Ayr\u0131ca, kernel kodunuzda branch prediction ve warp\/wavefront divergence gibi konulara dikkat etmek de \u00f6nemlidir.\n <\/div>\nPerformans Optimizasyonu ve \u0130leri D\u00fczey Teknikler: DotCompute RC2'yi Maksimum Verimle Nas\u0131l Kullan\u0131rs\u0131n\u0131z?<\/h2>\n
1. Bellek Y\u00f6netimi ve Veri Transferi Optimizasyonu<\/h3>\n
\n
ComputeBufferFlags.Pinned<\/code> bayra\u011f\u0131n\u0131 kullanarak CPU belle\u011fini pinleyebilirsiniz. Pinli bellek, i\u015fletim sistemi taraf\u0131ndan takas alan\u0131 olarak kullan\u0131lmaz ve do\u011frudan GPU'ya DMA (Do\u011frudan Bellek Eri\u015fimi) ile daha h\u0131zl\u0131 transfer edilebilir. Bu, \u00f6zellikle b\u00fcy\u00fck veri setleriyle \u00e7al\u0131\u015f\u0131rken \u00f6nemli bir h\u0131z art\u0131\u015f\u0131 sa\u011flar.<\/li>\nCommandQueue<\/code> \u00fczerindeki ReadBuffer<\/code> ve WriteBuffer<\/code> metotlar\u0131n\u0131n asenkron versiyonlar\u0131n\u0131 kullanarak CPU'nun veri transferi s\u0131ras\u0131nda di\u011fer i\u015fleri yapmaya devam etmesini sa\u011flay\u0131n. Bu, genel uygulama yan\u0131t verme s\u00fcresini art\u0131r\u0131r.<\/li>\n<\/ul>\n2. Kernel Optimizasyonu<\/h3>\n
\n
local_size<\/code>), GPU'nun mimarisine g\u00f6re optimal olarak ayarlanmal\u0131d\u0131r. Genellikle 32, 64, 128 veya 256 gibi 2'nin katlar\u0131 olan boyutlar tercih edilir. Do\u011fru boyutu bulmak i\u00e7in deneme yan\u0131lma veya GPU'nun kendi ara\u00e7lar\u0131n\u0131 kullanmak faydal\u0131 olabilir. Bu, GPU'nun \"warp\" veya \"wavefront\" ad\u0131 verilen temel i\u015f birimlerini verimli bir \u015fekilde kullanmas\u0131n\u0131 sa\u011flar.<\/li>\nif-else<\/code> dallar\u0131, bir i\u015f grubundaki i\u015f par\u00e7ac\u0131klar\u0131n\u0131n farkl\u0131 yollara gitmesine neden olabilir. Bu \"dallanma farkl\u0131l\u0131\u011f\u0131\" (branch divergence), performans\u0131 ciddi \u015fekilde d\u00fc\u015f\u00fcrebilir \u00e7\u00fcnk\u00fc farkl\u0131 yollara giden i\u015f par\u00e7ac\u0131klar\u0131 s\u0131rayla i\u015flenmek zorunda kal\u0131r. M\u00fcmk\u00fcn oldu\u011funca dallanmadan ka\u00e7\u0131n\u0131n veya dallanmalar\u0131 minimize edin.<\/li>\nComputeContext<\/code>, ComputeDevice<\/code> ve CommandQueue<\/code> nesnelerini tekrar tekrar olu\u015fturmaktan ka\u00e7\u0131n\u0131n. Genellikle, bu nesneleri bir kez ba\u015flat\u0131p gerekti\u011finde yeniden kullanmak en iyisidir.<\/li>\n<\/ul>\n3. Asenkron \u0130\u015flemler ve E\u015fzamanl\u0131l\u0131k<\/h3>\n
\n
\n \/\/ Asenkron bellek yazma \u00f6rne\u011fi\n IComputeEvent writeEvent = queue.WriteBufferAsync(inputBuffer, inputData);\n\n \/\/ CPU'da ba\u015fka i\u015fler yap...\n\n \/\/ Kernel \u00e7al\u0131\u015ft\u0131rma\n IComputeEvent kernelEvent = queue.ExecuteAsync(kernel, globalSize, localSize, new [] { writeEvent });\n\n \/\/ Sonu\u00e7lar\u0131 okurken kernel'\u0131n bitmesini bekle\n queue.ReadBufferAsync(outputBuffer, outputData, new [] { kernelEvent });\n queue.Finish(); \/\/ T\u00fcm bekleyen asenkron i\u015flemlerin tamamlanmas\u0131n\u0131 bekle\n <\/pre>\n
![\"Gri](\"path\/to\/grayscale_image.png\")
\n