private void initGLAndCuda()
{
//Create render target control
m_renderControl = new OpenTK.GLControl(GraphicsMode.Default, 1, 0, GraphicsContextFlags.Default);
m_renderControl.Dock = DockStyle.Fill;
m_renderControl.BackColor = Color.White;
m_renderControl.BorderStyle = BorderStyle.FixedSingle;
m_renderControl.KeyDown += new KeyEventHandler(m_renderControl_KeyDown);
m_renderControl.MouseMove += new MouseEventHandler(m_renderControl_MouseMove);
m_renderControl.MouseDown += new MouseEventHandler(m_renderControl_MouseDown);
m_renderControl.SizeChanged += new EventHandler(m_renderControl_SizeChanged);
panel1.Controls.Add(m_renderControl);
Console.WriteLine(" OpenGL device is Available");
int deviceID = CudaContext.GetMaxGflopsDeviceId();
ctx = CudaContext.CreateOpenGLContext(deviceID, CUCtxFlags.BlockingSync);
string console = string.Format("CUDA device [{0}] has {1} Multi-Processors", ctx.GetDeviceName(), ctx.GetDeviceInfo().MultiProcessorCount);
Console.WriteLine(console);
CUmodule module = ctx.LoadModulePTX("kernel.ptx");
addForces_k = new CudaKernel("addForces_k", module, ctx);
advectVelocity_k = new CudaKernel("advectVelocity_k", module, ctx);
diffuseProject_k = new CudaKernel("diffuseProject_k", module, ctx);
updateVelocity_k = new CudaKernel("updateVelocity_k", module, ctx);
advectParticles_k = new CudaKernel("advectParticles_OGL", module, ctx);
hvfield = new cData[DS];
dvfield = new CudaPitchedDeviceVariable<cData>(DIM, DIM);
tPitch = dvfield.Pitch;
dvfield.CopyToDevice(hvfield);
vxfield = new CudaDeviceVariable<cData>(DS);
vyfield = new CudaDeviceVariable<cData>(DS);
// Create particle array
particles = new cData[DS];
initParticles(particles, DIM, DIM);
// TODO: update kernels to use the new unpadded memory layout for perf
// rather than the old FFTW-compatible layout
planr2c = new CudaFFTPlan2D(DIM, DIM, cufftType.R2C, Compatibility.FFTWPadding);
planc2r = new CudaFFTPlan2D(DIM, DIM, cufftType.C2R, Compatibility.FFTWPadding);
GL.GenBuffers(1, out vbo);
GL.BindBuffer(BufferTarget.ArrayBuffer, vbo);
GL.BufferData<cData>(BufferTarget.ArrayBuffer, new IntPtr(cData.SizeOf * DS), particles, BufferUsageHint.DynamicDraw);
int bsize;
GL.GetBufferParameter(BufferTarget.ArrayBuffer, BufferParameterName.BufferSize, out bsize);
if (bsize != DS * cData.SizeOf)
throw new Exception("Sizes don't match.");
GL.BindBuffer(BufferTarget.ArrayBuffer, 0);
cuda_vbo_resource = new CudaGraphicsInteropResourceCollection();
cuda_vbo_resource.Add(new CudaOpenGLBufferInteropResource(vbo, CUGraphicsRegisterFlags.None));
texref = new CudaTextureArray2D(advectVelocity_k, "texref", CUAddressMode.Wrap, CUFilterMode.Linear, 0, CUArrayFormat.Float, DIM, DIM, CudaArray2DNumChannels.Two);
stopwatch = new CudaStopWatch(CUEventFlags.Default);
reshape();
isInit = true;
display();
}
// General GPU Device CUDA Initialization
static int gpuDeviceInit(int devID)
{
int deviceCount = CudaContext.GetDeviceCount();
if (deviceCount == 0)
{
Console.Write("gpuDeviceInit() CUDA error: no devices supporting CUDA.\n");
Environment.Exit(-1);
}
if (devID < 0)
{
devID = 0;
}
if (devID > deviceCount - 1)
{
Console.Write("\n");
Console.Write(">> {0} CUDA capable GPU device(s) detected. <<\n", deviceCount);
Console.Write(">> gpuDeviceInit (-device={0}) is not a valid GPU device. <<\n", devID);
Console.Write("\n");
return(-devID);
}
if (CudaContext.GetDeviceComputeCapability(devID).Major < 1)
{
Console.Write("gpuDeviceInit(): GPU device does not support CUDA.\n");
Environment.Exit(-1);
}
ctx = new CudaContext(devID);
Console.Write("> gpuDeviceInit() CUDA device [{0}]: {1}\n", devID, ctx.GetDeviceName());
return(devID);
}
private int findGraphicsGPU(out string devName)
{
int nGraphicsGPU = 0;
int deviceCount = 0;
bool bFoundGraphics = false;
string firstGraphicsName = string.Empty, temp;
devName = string.Empty;
deviceCount = CudaContext.GetDeviceCount();
// This function call returns 0 if there are no CUDA capable devices.
if (deviceCount == 0)
{
Console.WriteLine("There are no device(s) supporting CUDA");
return(0);
}
else
{
Console.WriteLine("> Found " + deviceCount + " CUDA Capable Device(s)");
}
for (int dev = 0; dev < deviceCount; dev++)
{
temp = CudaContext.GetDeviceName(dev);
bool bGraphics = !temp.Contains("Tesla");
StringBuilder sb = new StringBuilder();
sb.Append("> ");
if (bGraphics)
{
sb.Append("Graphics");
}
else
{
sb.Append("Compute");
}
sb.Append("\t\tGPU ").Append(dev).Append(": ").Append(CudaContext.GetDeviceName(dev));
Console.WriteLine(sb.ToString());
if (bGraphics)
{
if (!bFoundGraphics)
{
firstGraphicsName = temp;
}
nGraphicsGPU++;
}
}
if (nGraphicsGPU != 0)
{
devName = firstGraphicsName;
}
else
{
devName = "this hardware";
}
return(nGraphicsGPU);
}
static int Main()
{
try
{
int deviceCount = CudaContext.GetDeviceCount();
if (deviceCount == 0)
{
Console.Error.WriteLine("No CUDA devices detected. Sad face.");
return(-1);
}
Console.WriteLine($"{deviceCount} CUDA devices detected (first will be used)");
for (int i = 0; i < deviceCount; i++)
{
Console.WriteLine($"{i}: {CudaContext.GetDeviceName(i)}");
}
const int Count = 100;
using (var state = new CudaProcessor(deviceId: 0))
{
Console.WriteLine("Initializing kernel...");
var compileResult = state.LoadKernel(out string log);
if (compileResult != ManagedCuda.NVRTC.nvrtcResult.Success)
{
Console.Error.WriteLine(compileResult);
Console.Error.WriteLine(log);
return(-1);
}
Console.WriteLine(log);
Console.WriteLine("Initializing data...");
state.InitializeData(Count);
Console.WriteLine("Running kernel...");
state.CalculateAsync(2);
Console.WriteLine("Copying data back...");
state.CopyToHost();
Console.WriteLine("Waiting for completion...");
state.Synchronize();
Console.WriteLine("all done; showing some results");
var random = new Random(123456);
for (int i = 0; i < 20; i++)
{
//var rnd = random.Next(Count);
var multiply = state[i];
var adding = state.Get(i);
Console.WriteLine($"{i}: Multiplication:{nameof(multiply.X)}={multiply.X}, {nameof(multiply.Y)}={multiply.Y}");
Console.WriteLine($"{i}: Addition: {nameof(adding.X)}={adding.X}, {nameof(adding.Y)}={adding.Y}");
}
Console.WriteLine("Cleaning up...");
}
Console.WriteLine("All done; have a nice day");
}
catch (Exception ex)
{
Console.Error.WriteLine(ex.Message);
Console.ReadKey();
return(-1);
}
Console.ReadKey();
return(0);
}
// Initialization code to find the best CUDA Device
static int findCudaDevice(string[] args)
{
int devID = 0;
// If the command-line has a device number specified, use it
bool found = false;
foreach (var item in args)
{
if (item.Contains("device="))
{
found = true;
if (!int.TryParse(item, out devID))
{
Console.WriteLine("Invalid command line parameters");
Environment.Exit(-1);
}
if (devID < 0)
{
Console.WriteLine("Invalid command line parameters\n");
Environment.Exit(-1);
}
else
{
devID = gpuDeviceInit(devID);
if (devID < 0)
{
Console.WriteLine("exiting...\n");
ShrQATest.shrQAFinishExit(args, ShrQATest.eQAstatus.QA_FAILED);
Environment.Exit(-1);
}
}
}
}
if (!found)
{
// Otherwise pick the device with highest Gflops/s
devID = CudaContext.GetMaxGflopsDeviceId();
ctx = new CudaContext(devID, CUCtxFlags.SchedAuto);
Console.Write("> Using CUDA device [{0}]: {1}\n", devID, ctx.GetDeviceName());
}
return(devID);
}
private bool InitializeD3D()
{
HwndSource hwnd = new HwndSource(0, 0, 0, 0, 0, "null", IntPtr.Zero);
// Create the D3D object.
d3d = new Direct3D();
PresentParameters pp = new PresentParameters();
pp.BackBufferWidth = 512;
pp.BackBufferHeight = 512;
pp.BackBufferFormat = Format.Unknown;
pp.BackBufferCount = 0;
pp.Multisample = MultisampleType.None;
pp.MultisampleQuality = 0;
pp.SwapEffect = SwapEffect.Discard;
pp.DeviceWindowHandle = (IntPtr)0;
pp.Windowed = true;
pp.EnableAutoDepthStencil = false;
pp.AutoDepthStencilFormat = Format.Unknown;
pp.PresentationInterval = PresentInterval.Default;
bDeviceFound = false;
CUdevice[] cudaDevices = null;
for (g_iAdapter = 0; g_iAdapter < d3d.AdapterCount; g_iAdapter++)
{
device = new Device(d3d, d3d.Adapters[g_iAdapter].Adapter, DeviceType.Hardware, hwnd.Handle, CreateFlags.HardwareVertexProcessing | CreateFlags.Multithreaded, pp);
try
{
cudaDevices = CudaContext.GetDirectXDevices(device.ComPointer, CUd3dXDeviceList.All, CudaContext.DirectXVersion.D3D9);
bDeviceFound = cudaDevices.Length > 0;
infoLog.AppendText("> Display Device #" + d3d.Adapters[g_iAdapter].Adapter
+ ": \"" + d3d.Adapters[g_iAdapter].Details.Description + "\" supports Direct3D9 and CUDA.\n");
break;
}
catch (CudaException)
{
//No Cuda device found for this Direct3D9 device
infoLog.AppendText("> Display Device #" + d3d.Adapters[g_iAdapter].Adapter
+ ": \"" + d3d.Adapters[g_iAdapter].Details.Description + "\" supports Direct3D9 but not CUDA.\n");
}
}
// we check to make sure we have found a cuda-compatible D3D device to work on
if (!bDeviceFound)
{
infoLog.AppendText("No CUDA-compatible Direct3D9 device available");
if (device != null)
{
device.Dispose();
}
return(false);
}
ctx = new CudaContext(cudaDevices[0], device.ComPointer, CUCtxFlags.BlockingSync, CudaContext.DirectXVersion.D3D9);
deviceName.Text = "Device name: " + ctx.GetDeviceName();
// Set projection matrix
SlimDX.Matrix matProj = SlimDX.Matrix.OrthoOffCenterLH(0, 1, 1, 0, 0, 1);
device.SetTransform(TransformState.Projection, matProj);
// Turn off D3D lighting, since we are providing our own vertex colors
device.SetRenderState(RenderState.Lighting, false);
//Load kernels
CUmodule module = ctx.LoadModulePTX("kernel.ptx");
addForces_k = new CudaKernel("addForces_k", module, ctx);
advectVelocity_k = new CudaKernel("advectVelocity_k", module, ctx);
diffuseProject_k = new CudaKernel("diffuseProject_k", module, ctx);
updateVelocity_k = new CudaKernel("updateVelocity_k", module, ctx);
advectParticles_k = new CudaKernel("advectParticles_k", module, ctx);
d3dimage.Lock();
Surface surf = device.GetBackBuffer(0, 0);
d3dimage.SetBackBuffer(D3DResourceType.IDirect3DSurface9, surf.ComPointer);
d3dimage.Unlock();
surf.Dispose();
//Setup the "real" frame rate counter.
//The cuda counter only measures cuda runtime, not the overhead to actually
//show the result via DirectX and WPF.
realLastTick = Environment.TickCount;
return(true);
}
static int Main()
{
try
{
int deviceCount = CudaContext.GetDeviceCount();
if (deviceCount == 0)
{
Console.Error.WriteLine("No CUDA devices detected. Sad face.");
return(-1);
}
Console.WriteLine($"{deviceCount} CUDA devices detected (first will be used)");
for (int i = 0; i < deviceCount; i++)
{
Console.WriteLine($"{i}: {CudaContext.GetDeviceName(i)}");
}
const int Count = 1024 * 1024;
using (var state = new SomeState(deviceId: 0))
{
Console.WriteLine("Initializing kernel...");
string log;
var compileResult = state.LoadKernel(out log);
if (compileResult != ManagedCuda.NVRTC.nvrtcResult.Success)
{
Console.Error.WriteLine(compileResult);
Console.Error.WriteLine(log);
return(-1);
}
Console.WriteLine(log);
Console.WriteLine("Initializing data...");
state.InitializeData(Count);
Console.WriteLine("Running kernel...");
for (int i = 0; i < 8; i++)
{
state.MultiplyAsync(2);
}
Console.WriteLine("Copying data back...");
state.CopyToHost(); // note: usually you try to minimize how much you need to
// fetch from the device, as that can be a bottleneck; you should prefer fetching
// minimal aggregate data (counts, etc), or the required pages of data; fetching
// *all* the data works, but should be avoided when possible.
Console.WriteLine("Waiting for completion...");
state.Synchronize();
Console.WriteLine("all done; showing some results");
var random = new Random(123456);
for (int i = 0; i < 20; i++)
{
var record = state[random.Next(Count)];
Console.WriteLine($"{i}: {nameof(record.Id)}={record.Id}, {nameof(record.Value)}={record.Value}");
}
Console.WriteLine("Cleaning up...");
}
Console.WriteLine("All done; have a nice day");
return(0);
} catch (Exception ex)
{
Console.Error.WriteLine(ex.Message);
return(-1);
}
}
static void Main(string[] args)
{
try
{
if (args.Length > 0)
{
deviceID = int.Parse(args[0]);
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Device ID parse error");
}
try
{
if (args.Length > 1)
{
port = int.Parse(args[1]);
Comms.ConnectToMaster(port);
}
else
{
TEST = true;
Logger.CopyToConsole = true;
CGraph.ShowCycles = true;
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Master connection error");
}
try
{
if (args.Length > 3)
{
gpuCount = int.Parse(args[3]);
fastCuda = gpuCount <= (Environment.ProcessorCount / 2);
if (fastCuda)
{
Logger.Log(LogLevel.Info, "Using single GPU blocking mode");
}
}
}
catch
{
}
if (TEST)
{
currentJob = nextJob = new Job()
{
jobID = 0,
k0 = 0xf4956dc403730b01L,
k1 = 0xe6d45de39c2a5a3eL,
k2 = 0xcbf626a8afee35f6L,
k3 = 0x4307b94b1a0c9980L,
pre_pow = TestPrePow,
timestamp = DateTime.Now
};
}
else
{
currentJob = nextJob = new Job()
{
jobID = 0,
k0 = 0xf4956dc403730b01L,
k1 = 0xe6d45de39c2a5a3eL,
k2 = 0xcbf626a8afee35f6L,
k3 = 0x4307b94b1a0c9980L,
pre_pow = TestPrePow,
timestamp = DateTime.Now
};
if (!Comms.IsConnected())
{
Console.WriteLine("Master connection failed, aborting");
Logger.Log(LogLevel.Error, "No master connection, exitting!");
return;
}
if (deviceID < 0)
{
int devCnt = CudaContext.GetDeviceCount();
GpuDevicesMessage gpum = new GpuDevicesMessage()
{
devices = new List <GpuDevice>(devCnt)
};
for (int i = 0; i < devCnt; i++)
{
string name = CudaContext.GetDeviceName(i);
var info = CudaContext.GetDeviceInfo(i);
gpum.devices.Add(new GpuDevice()
{
deviceID = i, name = name, memory = info.TotalGlobalMemory
});
}
//Console.WriteLine(devCnt);
Comms.gpuMsg = gpum;
Comms.SetEvent();
//Console.WriteLine("event fired");
Task.Delay(1000).Wait();
//Console.WriteLine("closing");
Comms.Close();
return;
}
}
try
{
var assembly = Assembly.GetEntryAssembly();
var resourceStream = assembly.GetManifestResourceStream("CudaSolver.kernel_x64.ptx");
ctx = new CudaContext(deviceID, !fastCuda ? (CUCtxFlags.BlockingSync | CUCtxFlags.MapHost) : CUCtxFlags.MapHost);
meanSeedA = ctx.LoadKernelPTX(resourceStream, "FluffySeed2A");
meanSeedA.BlockDimensions = 128;
meanSeedA.GridDimensions = 2048;
meanSeedA.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanSeedB = ctx.LoadKernelPTX(resourceStream, "FluffySeed2B");
meanSeedB.BlockDimensions = 128;
meanSeedB.GridDimensions = 2048;
meanSeedB.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanSeedB_4 = ctx.LoadKernelPTX(resourceStream, "FluffySeed2B");
meanSeedB_4.BlockDimensions = 128;
meanSeedB_4.GridDimensions = 1024;
meanSeedB_4.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRound = ctx.LoadKernelPTX(resourceStream, "FluffyRound");
meanRound.BlockDimensions = 512;
meanRound.GridDimensions = 4096;
meanRound.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRound_2 = ctx.LoadKernelPTX(resourceStream, "FluffyRound");
meanRound_2.BlockDimensions = 512;
meanRound_2.GridDimensions = 2048;
meanRound_2.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRoundJoin = ctx.LoadKernelPTX(resourceStream, "FluffyRound_J");
meanRoundJoin.BlockDimensions = 512;
meanRoundJoin.GridDimensions = 4096;
meanRoundJoin.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanTail = ctx.LoadKernelPTX(resourceStream, "FluffyTail");
meanTail.BlockDimensions = 1024;
meanTail.GridDimensions = 4096;
meanTail.PreferredSharedMemoryCarveout = CUshared_carveout.MaxL1;
meanRecover = ctx.LoadKernelPTX(resourceStream, "FluffyRecovery");
meanRecover.BlockDimensions = 256;
meanRecover.GridDimensions = 2048;
meanRecover.PreferredSharedMemoryCarveout = CUshared_carveout.MaxL1;
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Unable to create kernels: " + ex.Message);
Task.Delay(500).Wait();
Comms.Close();
return;
}
try
{
d_buffer = new CudaDeviceVariable <ulong>(BUFFER_SIZE_U32);
d_bufferMid = new CudaDeviceVariable <ulong>(d_buffer.DevicePointer + (BUFFER_SIZE_B * 8));
d_bufferB = new CudaDeviceVariable <ulong>(d_buffer.DevicePointer + (BUFFER_SIZE_A * 8));
d_indexesA = new CudaDeviceVariable <uint>(INDEX_SIZE * 2);
d_indexesB = new CudaDeviceVariable <uint>(INDEX_SIZE * 2);
Array.Clear(h_indexesA, 0, h_indexesA.Length);
Array.Clear(h_indexesB, 0, h_indexesA.Length);
d_indexesA = h_indexesA;
d_indexesB = h_indexesB;
streamPrimary = new CudaStream(CUStreamFlags.NonBlocking);
streamSecondary = new CudaStream(CUStreamFlags.NonBlocking);
}
catch (Exception ex)
{
Task.Delay(200).Wait();
Logger.Log(LogLevel.Error, $"Out of video memory! Only {ctx.GetFreeDeviceMemorySize()} free");
Task.Delay(500).Wait();
Comms.Close();
return;
}
try
{
AllocateHostMemory(true, ref h_a, ref hAligned_a, 1024 * 1024 * 32);
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Unable to create pinned memory.");
Task.Delay(500).Wait();
Comms.Close();
return;
}
int loopCnt = 0;
while (!Comms.IsTerminated)
{
try
{
if (!TEST && (Comms.nextJob.pre_pow == null || Comms.nextJob.pre_pow == "" || Comms.nextJob.pre_pow == TestPrePow))
{
Logger.Log(LogLevel.Info, string.Format("Waiting for job...."));
Task.Delay(1000).Wait();
continue;
}
if (!TEST && ((currentJob.pre_pow != Comms.nextJob.pre_pow) || (currentJob.origin != Comms.nextJob.origin)))
{
currentJob = Comms.nextJob;
currentJob.timestamp = DateTime.Now;
}
if (!TEST && (currentJob.timestamp.AddMinutes(30) < DateTime.Now) && Comms.lastIncoming.AddMinutes(30) < DateTime.Now)
{
Logger.Log(LogLevel.Info, string.Format("Job too old..."));
Task.Delay(1000).Wait();
continue;
}
// test runs only once
if (TEST && loopCnt++ > 100)
{
Comms.IsTerminated = true;
}
Solution s;
while (graphSolutions.TryDequeue(out s))
{
meanRecover.SetConstantVariable <ulong>("recovery", s.GetUlongEdges());
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRecover.RunAsync(streamPrimary.Stream, s.job.k0, s.job.k1, s.job.k2, s.job.k3, d_indexesB.DevicePointer);
streamPrimary.Synchronize();
s.nonces = new uint[40];
d_indexesB.CopyToHost(s.nonces, 0, 0, 40 * 4);
s.nonces = s.nonces.OrderBy(n => n).ToArray();
lock (Comms.graphSolutionsOut)
{
Comms.graphSolutionsOut.Enqueue(s);
}
Comms.SetEvent();
}
uint[] count;
do
{
if (!TEST && ((currentJob.pre_pow != Comms.nextJob.pre_pow) || (currentJob.origin != Comms.nextJob.origin)))
{
currentJob = Comms.nextJob;
currentJob.timestamp = DateTime.Now;
}
currentJob = currentJob.Next();
Logger.Log(LogLevel.Debug, string.Format("GPU NV{4}:Trimming #{4}: {0} {1} {2} {3}", currentJob.k0, currentJob.k1, currentJob.k2, currentJob.k3, currentJob.jobID, deviceID));
timer.Restart();
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanSeedA.RunAsync(streamPrimary.Stream, currentJob.k0, currentJob.k1, currentJob.k2, currentJob.k3, d_bufferMid.DevicePointer, d_indexesB.DevicePointer);
meanSeedB_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, 0);
meanSeedB_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer + ((BUFFER_SIZE_A * 8) / 4) * 1, d_indexesB.DevicePointer, d_indexesA.DevicePointer, 16);
meanSeedB_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer + ((BUFFER_SIZE_A * 8) / 4) * 2, d_indexesB.DevicePointer, d_indexesA.DevicePointer, 32);
meanSeedB_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer + ((BUFFER_SIZE_A * 8) / 4) * 3, d_indexesB.DevicePointer, d_indexesA.DevicePointer, 48);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRound_2.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer + ((BUFFER_SIZE_A * 8) / 4) * 2, d_bufferB.DevicePointer, d_indexesA.DevicePointer + (2048 * 4), d_indexesB.DevicePointer + (4096 * 4), DUCK_EDGES_A, DUCK_EDGES_B / 2);
meanRound_2.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer, d_bufferB.DevicePointer - (BUFFER_SIZE_B * 8), d_indexesA.DevicePointer, d_indexesB.DevicePointer, DUCK_EDGES_A, DUCK_EDGES_B / 2);
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanRoundJoin.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer - (BUFFER_SIZE_B * 8), d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 2);
//d_indexesA.MemsetAsync(0, streamPrimary.Stream);
//meanRound.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_B, DUCK_EDGES_B / 2);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer, d_bufferB.DevicePointer, d_indexesA.DevicePointer, d_indexesB.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 2);
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 2);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer, d_bufferB.DevicePointer, d_indexesA.DevicePointer, d_indexesB.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 4);
for (int i = 0; i < trimRounds; i++)
{
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_B / 4, DUCK_EDGES_B / 4);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer, d_bufferB.DevicePointer, d_indexesA.DevicePointer, d_indexesB.DevicePointer, DUCK_EDGES_B / 4, DUCK_EDGES_B / 4);
}
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanTail.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer);
ctx.Synchronize();
streamPrimary.Synchronize();
count = new uint[2];
d_indexesA.CopyToHost(count, 0, 0, 8);
if (count[0] > 4194304)
{
// trouble
count[0] = 4194304;
// log
}
hAligned_a.AsyncCopyFromDevice(d_buffer.DevicePointer, 0, 0, count[0] * 8, streamPrimary.Stream);
streamPrimary.Synchronize();
System.Runtime.InteropServices.Marshal.Copy(hAligned_a.PinnedHostPointer, h_a, 0, ((int)count[0] * 8) / sizeof(int));
timer.Stop();
currentJob.solvedAt = DateTime.Now;
currentJob.trimTime = timer.ElapsedMilliseconds;
//Console.WriteLine("Trimmed in {0}ms to {1} edges", timer.ElapsedMilliseconds, count[0]);
Logger.Log(LogLevel.Info, string.Format("GPU NV{2}: Trimmed in {0}ms to {1} edges, h {3}", timer.ElapsedMilliseconds, count[0], deviceID, currentJob.height));
}while((currentJob.height != Comms.nextJob.height) && (!Comms.IsTerminated) && (!TEST));
if (TEST)
{
//Console.WriteLine("Trimmed in {0}ms to {1} edges", timer.ElapsedMilliseconds, count[0]);
CGraph cg = FinderBag.GetFinder();
if (cg == null)
{
continue;
}
cg.SetEdges(h_a, (int)count[0]);
cg.SetHeader(currentJob);
//currentJob = currentJob.Next();
Task.Factory.StartNew(() =>
{
Stopwatch sw = new Stopwatch();
sw.Start();
if (count[0] < 200000)
{
try
{
if (findersInFlight++ < 3)
{
Stopwatch cycleTime = new Stopwatch();
cycleTime.Start();
cg.FindSolutions(graphSolutions);
cycleTime.Stop();
AdjustTrims(cycleTime.ElapsedMilliseconds);
if (graphSolutions.Count > 0)
{
solutions++;
}
}
else
{
Logger.Log(LogLevel.Warning, "CPU overloaded!");
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Cycle finder error" + ex.Message);
}
finally
{
findersInFlight--;
FinderBag.ReturnFinder(cg);
}
}
sw.Stop();
if (++trims % 50 == 0)
{
Console.ForegroundColor = ConsoleColor.Green;
Console.WriteLine("SOLS: {0}/{1} - RATE: {2:F1}", solutions, trims, (float)trims / solutions);
Console.ResetColor();
}
//Console.WriteLine("Finder completed in {0}ms on {1} edges with {2} solution(s)", sw.ElapsedMilliseconds, count[0], graphSolutions.Count);
//Console.WriteLine("Duped edges: {0}", cg.dupes);
Logger.Log(LogLevel.Info, string.Format("Finder completed in {0}ms on {1} edges with {2} solution(s) and {3} dupes", sw.ElapsedMilliseconds, count[0], graphSolutions.Count, cg.dupes));
});
//h_indexesA = d_indexesA;
//h_indexesB = d_indexesB;
//var sumA = h_indexesA.Sum(e => e);
//var sumB = h_indexesB.Sum(e => e);
;
}
else
{
CGraph cg = FinderBag.GetFinder();
cg.SetEdges(h_a, (int)count[0]);
cg.SetHeader(currentJob);
Task.Factory.StartNew(() =>
{
if (count[0] < 200000)
{
try
{
if (findersInFlight++ < 3)
{
Stopwatch cycleTime = new Stopwatch();
cycleTime.Start();
cg.FindSolutions(graphSolutions);
cycleTime.Stop();
AdjustTrims(cycleTime.ElapsedMilliseconds);
if (graphSolutions.Count > 0)
{
solutions++;
}
}
else
{
Logger.Log(LogLevel.Warning, "CPU overloaded!");
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Cycle finder crashed: " + ex.Message);
}
finally
{
findersInFlight--;
FinderBag.ReturnFinder(cg);
}
}
});
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Critical error in main cuda loop " + ex.Message);
Task.Delay(5000).Wait();
}
}
// clean up
try
{
Task.Delay(500).Wait();
Comms.Close();
d_buffer.Dispose();
d_indexesA.Dispose();
d_indexesB.Dispose();
streamPrimary.Dispose();
streamSecondary.Dispose();
hAligned_a.Dispose();
if (ctx != null)
{
ctx.Dispose();
}
}
catch { }
}
// Initialization code to find the best CUDA Device
static int findCudaDevice(string[] args)
{
int devID = 0;
// If the command-line has a device number specified, use it
bool found = false;
foreach (var item in args)
{
if (item.Contains("device="))
{
found = true;
if (!int.TryParse(item, out devID))
{
Console.WriteLine("Invalid command line parameters");
Environment.Exit(-1);
}
if (devID < 0)
{
Console.WriteLine("Invalid command line parameters\n");
Environment.Exit(-1);
}
else
{
devID = gpuDeviceInit(devID);
if (devID < 0)
{
Console.WriteLine("exiting...\n");
ShrQATest.shrQAFinishExit(args, ShrQATest.eQAstatus.QA_FAILED);
Environment.Exit(-1);
}
}
}
}
if (!found)
{
// Otherwise pick the device with highest Gflops/s
devID = CudaContext.GetMaxGflopsDeviceId();
ctx = new CudaContext(devID, CUCtxFlags.SchedAuto);
Console.Write("> Using CUDA device [{0}]: {1}\n", devID, ctx.GetDeviceName());
}
return devID;
}
// General GPU Device CUDA Initialization
static int gpuDeviceInit(int devID)
{
int deviceCount = CudaContext.GetDeviceCount();
if (deviceCount == 0)
{
Console.Write("gpuDeviceInit() CUDA error: no devices supporting CUDA.\n");
Environment.Exit(-1);
}
if (devID < 0)
devID = 0;
if (devID > deviceCount - 1)
{
Console.Write("\n");
Console.Write(">> {0} CUDA capable GPU device(s) detected. <<\n", deviceCount);
Console.Write(">> gpuDeviceInit (-device={0}) is not a valid GPU device. <<\n", devID);
Console.Write("\n");
return -devID;
}
if (CudaContext.GetDeviceComputeCapability(devID).Major < 1)
{
Console.Write("gpuDeviceInit(): GPU device does not support CUDA.\n");
Environment.Exit(-1);
}
ctx = new CudaContext(devID);
Console.Write("> gpuDeviceInit() CUDA device [{0}]: {1}\n", devID, ctx.GetDeviceName());
return devID;
}
static void Main(string[] args)
{
try
{
if (args.Length == 1 && args[0].ToLower().Contains("fidelity"))
{
string[] fseg = args[0].Split(':');
deviceID = int.Parse(fseg[1]);
nonce = Int64.Parse(fseg[2]) - 1;
range = int.Parse(fseg[3]);
QTEST = true;
}
else
{
if (args.Length > 0)
{
deviceID = int.Parse(args[0]);
}
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Device ID parse error: " + ex.Message);
}
try
{
if (args.Length > 0)
{
deviceID = int.Parse(args[0]);
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Device ID parse error");
}
try
{
if (args.Length > 1)
{
port = int.Parse(args[1]);
Comms.ConnectToMaster(port);
}
else
{
TEST = true;
Logger.CopyToConsole = true;
CGraph.ShowCycles = true;
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Master connection error");
}
try
{
if (args.Length > 3)
{
gpuCount = int.Parse(args[3]);
fastCuda = gpuCount <= (Environment.ProcessorCount / 2);
if (fastCuda)
{
Logger.Log(LogLevel.Info, "Using single GPU blocking mode");
}
}
}
catch
{
}
if (TEST)
{
currentJob = nextJob = new Job()
{
jobID = 0,
k0 = 0xf4956dc403730b01L,
k1 = 0xe6d45de39c2a5a3eL,
k2 = 0xcbf626a8afee35f6L,
k3 = 0x4307b94b1a0c9980L,
pre_pow = TestPrePow,
timestamp = DateTime.Now
};
}
else
{
currentJob = nextJob = new Job()
{
jobID = 0,
k0 = 0xf4956dc403730b01L,
k1 = 0xe6d45de39c2a5a3eL,
k2 = 0xcbf626a8afee35f6L,
k3 = 0x4307b94b1a0c9980L,
pre_pow = TestPrePow,
timestamp = DateTime.Now
};
if (!Comms.IsConnected())
{
Console.WriteLine("Master connection failed, aborting");
Logger.Log(LogLevel.Error, "No master connection, exitting!");
return;
}
if (deviceID < 0)
{
int devCnt = CudaContext.GetDeviceCount();
GpuDevicesMessage gpum = new GpuDevicesMessage()
{
devices = new List <GpuDevice>(devCnt)
};
for (int i = 0; i < devCnt; i++)
{
string name = CudaContext.GetDeviceName(i);
var info = CudaContext.GetDeviceInfo(i);
gpum.devices.Add(new GpuDevice()
{
deviceID = i, name = name, memory = info.TotalGlobalMemory
});
}
//Console.WriteLine(devCnt);
Comms.gpuMsg = gpum;
Comms.SetEvent();
//Console.WriteLine("event fired");
Task.Delay(1000).Wait();
//Console.WriteLine("closing");
Comms.Close();
return;
}
}
try
{
var assembly = Assembly.GetEntryAssembly();
var resourceStream = assembly.GetManifestResourceStream("CudaSolver.kernel_x64.ptx");
ctx = new CudaContext(deviceID, /*!fastCuda ? (CUCtxFlags.BlockingSync | CUCtxFlags.MapHost) :*/ CUCtxFlags.MapHost);
string pow = new StreamReader(resourceStream).ReadToEnd();
//pow = File.ReadAllText(@"kernel_x64.ptx");
Turing = ctx.GetDeviceInfo().MaxSharedMemoryPerMultiprocessor == 65536;
using (var s = GenerateStreamFromString(pow))
{
if (!Turing)
{
meanSeedA = ctx.LoadKernelPTX(s, "FluffySeed4K", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)40 });
meanSeedA.BlockDimensions = 512;
meanSeedA.GridDimensions = 1024;
meanSeedA.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRound = ctx.LoadKernelPTX(s, "FluffyRound_A2", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)40 });
meanRound.BlockDimensions = 512;
meanRound.GridDimensions = 4096;
meanRound.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRound_4 = ctx.LoadKernelPTX(s, "FluffyRound_A1", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)32 });
meanRound_4.BlockDimensions = 1024;
meanRound_4.GridDimensions = 1024;
meanRound_4.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRoundJoin = ctx.LoadKernelPTX(s, "FluffyRound_A3", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)32 });
meanRoundJoin.BlockDimensions = 1024;
meanRoundJoin.GridDimensions = 4096;
meanRoundJoin.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanTail = ctx.LoadKernelPTX(s, "FluffyTail");
meanTail.BlockDimensions = 1024;
meanTail.GridDimensions = 4096;
meanTail.PreferredSharedMemoryCarveout = CUshared_carveout.MaxL1;
meanRecover = ctx.LoadKernelPTX(s, "FluffyRecovery");
meanRecover.BlockDimensions = 256;
meanRecover.GridDimensions = 2048;
meanRecover.PreferredSharedMemoryCarveout = CUshared_carveout.MaxL1;
}
else
{
meanSeedA = ctx.LoadKernelPTX(s, "FluffySeed4K", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)64 });
meanSeedA.BlockDimensions = 512;
meanSeedA.GridDimensions = 1024;
meanSeedA.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRound = ctx.LoadKernelPTX(s, "FluffyRound_C2", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)32 });
meanRound.BlockDimensions = 1024;
meanRound.GridDimensions = 4096;
meanRound.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRound_4 = ctx.LoadKernelPTX(s, "FluffyRound_C1", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)64 });
meanRound_4.BlockDimensions = 1024;
meanRound_4.GridDimensions = 1024;
meanRound_4.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanRoundJoin = ctx.LoadKernelPTX(s, "FluffyRound_C3", new CUJITOption[] { CUJITOption.MaxRegisters }, new object[] { (uint)32 });
meanRoundJoin.BlockDimensions = 1024;
meanRoundJoin.GridDimensions = 4096;
meanRoundJoin.PreferredSharedMemoryCarveout = CUshared_carveout.MaxShared;
meanTail = ctx.LoadKernelPTX(s, "FluffyTail");
meanTail.BlockDimensions = 1024;
meanTail.GridDimensions = 4096;
meanTail.PreferredSharedMemoryCarveout = CUshared_carveout.MaxL1;
meanRecover = ctx.LoadKernelPTX(s, "FluffyRecovery");
meanRecover.BlockDimensions = 256;
meanRecover.GridDimensions = 2048;
meanRecover.PreferredSharedMemoryCarveout = CUshared_carveout.MaxL1;
}
}
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Unable to create kernels: " + ex.Message);
Task.Delay(500).Wait();
Comms.Close();
return;
}
try
{
d_buffer = new CudaDeviceVariable <ulong>(BUFFER_SIZE_U32 * (temp ? 8 : 1));
d_bufferMid = new CudaDeviceVariable <ulong>(d_buffer.DevicePointer + (BUFFER_SIZE_B * 2));
d_bufferB = new CudaDeviceVariable <ulong>(d_buffer.DevicePointer + (BUFFER_SIZE_B * 8));
d_indexesA = new CudaDeviceVariable <uint>(INDEX_SIZE);
d_indexesB = new CudaDeviceVariable <uint>(INDEX_SIZE);
d_aux = new CudaDeviceVariable <uint>(512);
Array.Clear(h_indexesA, 0, h_indexesA.Length);
Array.Clear(h_indexesB, 0, h_indexesA.Length);
d_indexesA = h_indexesA;
d_indexesB = h_indexesB;
streamPrimary = new CudaStream(CUStreamFlags.NonBlocking);
}
catch (Exception ex)
{
Task.Delay(200).Wait();
Logger.Log(LogLevel.Error, $"Mem alloc exception. Out of video memory? {ctx.GetFreeDeviceMemorySize()} free");
Task.Delay(500).Wait();
Comms.Close();
return;
}
try
{
AllocateHostMemory(true, ref h_a, ref hAligned_a, 1024 * 1024 * 32);
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Unable to create pinned memory.");
Task.Delay(500).Wait();
Comms.Close();
return;
}
int loopCnt = 0;
while (!Comms.IsTerminated)
{
try
{
if (!TEST && (Comms.nextJob.pre_pow == null || Comms.nextJob.pre_pow == "" || Comms.nextJob.pre_pow == TestPrePow))
{
Logger.Log(LogLevel.Info, string.Format("Waiting for job...."));
Task.Delay(1000).Wait();
continue;
}
if (!TEST && ((currentJob.pre_pow != Comms.nextJob.pre_pow) || (currentJob.origin != Comms.nextJob.origin)))
{
currentJob = Comms.nextJob;
currentJob.timestamp = DateTime.Now;
}
if (!TEST && (currentJob.timestamp.AddMinutes(30) < DateTime.Now) && Comms.lastIncoming.AddMinutes(30) < DateTime.Now)
{
Logger.Log(LogLevel.Info, string.Format("Job too old..."));
Task.Delay(1000).Wait();
continue;
}
// test runs only once
if (TEST && ++loopCnt >= range)
{
Comms.IsTerminated = true;
}
Solution s;
while (graphSolutions.TryDequeue(out s))
{
meanRecover.SetConstantVariable <ulong>("recovery", s.GetUlongEdges());
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRecover.RunAsync(streamPrimary.Stream, s.job.k0, s.job.k1, s.job.k2, s.job.k3, d_indexesB.DevicePointer);
streamPrimary.Synchronize();
s.nonces = new uint[32];
d_indexesB.CopyToHost(s.nonces, 0, 0, 32 * 4);
s.nonces = s.nonces.OrderBy(n => n).ToArray();
//fidelity = (32-cycles_found / graphs_searched) * 32
solutions++;
s.fidelity = ((double)solutions / (double)trims) * 32.0;
//Console.WriteLine(s.fidelity.ToString("0.000"));
if (Comms.IsConnected())
{
Comms.graphSolutionsOut.Enqueue(s);
Comms.SetEvent();
}
if (QTEST)
{
Console.ForegroundColor = ConsoleColor.Red;
Console.WriteLine($"Solution for nonce {s.job.nonce}: {string.Join(' ', s.nonces)}");
Console.ResetColor();
}
}
if (QTEST)
{
currentJob = currentJob.NextSequential(ref nonce);
Console.WriteLine($"Nonce: {nonce} K0: {currentJob.k0:X} K1: {currentJob.k1:X} K2: {currentJob.k2:X} K3: {currentJob.k3:X}");
}
else
{
currentJob = currentJob.Next();
}
Logger.Log(LogLevel.Debug, string.Format("GPU NV{4}:Trimming #{4}: {0} {1} {2} {3}", currentJob.k0, currentJob.k1, currentJob.k2, currentJob.k3, currentJob.jobID, deviceID));
timer.Restart();
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
d_aux.MemsetAsync(0, streamPrimary.Stream);
meanSeedA.RunAsync(streamPrimary.Stream, currentJob.k0, currentJob.k1, currentJob.k2, currentJob.k3, d_bufferMid.DevicePointer, d_indexesB.DevicePointer, 0);
meanSeedA.RunAsync(streamPrimary.Stream, currentJob.k0, currentJob.k1, currentJob.k2, currentJob.k3, d_bufferMid.DevicePointer + ((BUFFER_SIZE_A * 8) / 4 / 4) * 1, d_indexesB.DevicePointer + (4096 * 4), EDGE_SEG);
meanSeedA.RunAsync(streamPrimary.Stream, currentJob.k0, currentJob.k1, currentJob.k2, currentJob.k3, d_bufferMid.DevicePointer + ((BUFFER_SIZE_A * 8) / 4 / 4) * 2, d_indexesB.DevicePointer + (4096 * 8), EDGE_SEG * 2);
meanSeedA.RunAsync(streamPrimary.Stream, currentJob.k0, currentJob.k1, currentJob.k2, currentJob.k3, d_bufferMid.DevicePointer + ((BUFFER_SIZE_A * 8) / 4 / 4) * 3, d_indexesB.DevicePointer + (4096 * 12), EDGE_SEG * 3);
meanRound_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_A / 4, DUCK_EDGES_B / 4, 0);
meanRound_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer + ((BUFFER_SIZE_B * 8) / 4) * 1, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_A / 4, DUCK_EDGES_B / 4, 1024);
meanRound_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer + ((BUFFER_SIZE_B * 8) / 4) * 2, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_A / 4, DUCK_EDGES_B / 4, 2048);
meanRound_4.RunAsync(streamPrimary.Stream, d_bufferMid.DevicePointer, d_buffer.DevicePointer + ((BUFFER_SIZE_B * 8) / 4) * 3, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_A / 4, DUCK_EDGES_B / 4, 3072);
//streamPrimary.Synchronize();
//h_indexesA = d_indexesA;
//h_indexesB = d_indexesB;
//var sumA = h_indexesA.Sum(e => e);
//var sumB = h_indexesB.Sum(e => e);
//streamPrimary.Synchronize();
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRoundJoin.RunAsync(streamPrimary.Stream,
d_buffer.DevicePointer,
d_buffer.DevicePointer + ((BUFFER_SIZE_B * 8) / 4) * 1,
d_buffer.DevicePointer + ((BUFFER_SIZE_B * 8) / 4) * 2,
d_buffer.DevicePointer + ((BUFFER_SIZE_B * 8) / 4) * 3,
d_bufferB.DevicePointer,
d_indexesA.DevicePointer,
d_indexesB.DevicePointer, DUCK_EDGES_B / 4, DUCK_EDGES_B / 2);
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 2, 0, d_aux.DevicePointer);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer, d_bufferB.DevicePointer, d_indexesA.DevicePointer, d_indexesB.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 2, 1, d_aux.DevicePointer);
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 2, 2, d_aux.DevicePointer);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer, d_bufferB.DevicePointer, d_indexesA.DevicePointer, d_indexesB.DevicePointer, DUCK_EDGES_B / 2, DUCK_EDGES_B / 4, 3, d_aux.DevicePointer);
for (int i = 0; i < (TEST ? 80 : trimRounds); i++)
//for (int i = 0; i < 85; i++)
{
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer, DUCK_EDGES_B / 4, DUCK_EDGES_B / 4, i * 2 + 4, d_aux.DevicePointer);
d_indexesB.MemsetAsync(0, streamPrimary.Stream);
meanRound.RunAsync(streamPrimary.Stream, d_buffer.DevicePointer, d_bufferB.DevicePointer, d_indexesA.DevicePointer, d_indexesB.DevicePointer, DUCK_EDGES_B / 4, DUCK_EDGES_B / 4, i * 2 + 5, d_aux.DevicePointer);
}
d_indexesA.MemsetAsync(0, streamPrimary.Stream);
meanTail.RunAsync(streamPrimary.Stream, d_bufferB.DevicePointer, d_buffer.DevicePointer, d_indexesB.DevicePointer, d_indexesA.DevicePointer);
Task.Delay((int)lastTrimMs).Wait();
streamPrimary.Synchronize();
uint[] count = new uint[2];
d_indexesA.CopyToHost(count, 0, 0, 8);
if (count[0] > 131071)
{
// trouble
count[0] = 131071;
// log
}
hAligned_a.AsyncCopyFromDevice(d_buffer.DevicePointer, 0, 0, count[0] * 8, streamPrimary.Stream);
streamPrimary.Synchronize();
System.Runtime.InteropServices.Marshal.Copy(hAligned_a.PinnedHostPointer, h_a, 0, ((int)count[0] * 8) / sizeof(int));
trims++;
timer.Stop();
lastTrimMs = (long)Math.Min(Math.Max((float)timer.ElapsedMilliseconds * 0.9f, 50), 500);
currentJob.solvedAt = DateTime.Now;
currentJob.trimTime = timer.ElapsedMilliseconds;
//Console.WriteLine("Trimmed in {0}ms to {1} edges", timer.ElapsedMilliseconds, count[0]);
Logger.Log(LogLevel.Info, string.Format("GPU NV{2}: Trimmed in {0}ms to {1} edges", timer.ElapsedMilliseconds, count[0], deviceID));
FinderBag.RunFinder(TEST, ref trims, count[0], h_a, currentJob, graphSolutions, timer);
if (trims % 50 == 0 && TEST)
{
Console.ForegroundColor = ConsoleColor.Green;
Console.WriteLine("SOLS: {0}/{1} - RATE: {2:F1}", solutions, trims, (float)trims / solutions);
Console.ResetColor();
}
/*
* if (TEST)
* {
* //Console.WriteLine("Trimmed in {0}ms to {1} edges", timer.ElapsedMilliseconds, count[0]);
*
* CGraph cg = FinderBag.GetFinder();
* cg.SetEdges(h_a, (int)count[0]);
* cg.SetHeader(currentJob);
*
* //currentJob = currentJob.Next();
*
* Task.Factory.StartNew(() =>
* {
* Stopwatch sw = new Stopwatch();
* sw.Start();
*
* if (count[0] < 131071)
* {
* try
* {
* if (findersInFlight++ < 3)
* {
* Stopwatch cycleTime = new Stopwatch();
* cycleTime.Start();
* cg.FindSolutions(graphSolutions);
* cycleTime.Stop();
* AdjustTrims(cycleTime.ElapsedMilliseconds);
* //if (graphSolutions.Count > 0) solutions++;
* }
* else
* Logger.Log(LogLevel.Warning, "CPU overloaded!");
* }
* catch (Exception ex)
* {
* Logger.Log(LogLevel.Error, "Cycle finder error" + ex.Message);
* }
* finally
* {
* FinderBag.ReturnFinder(cg);
* findersInFlight--;
* }
* }
*
* sw.Stop();
*
* if (trims % 50 == 0)
* {
* Console.ForegroundColor = ConsoleColor.Green;
* Console.WriteLine("SOLS: {0}/{1} - RATE: {2:F1}", solutions, trims, (float)trims/solutions );
* Console.ResetColor();
* }
* //Console.WriteLine("Finder completed in {0}ms on {1} edges with {2} solution(s)", sw.ElapsedMilliseconds, count[0], graphSolutions.Count);
* //Console.WriteLine("Duped edges: {0}", cg.dupes);
* if (!QTEST)
* Logger.Log(LogLevel.Info, string.Format("Finder completed in {0}ms on {1} edges with {2} solution(s) and {3} dupes", sw.ElapsedMilliseconds, count[0], graphSolutions.Count, cg.dupes));
* });
*
* //h_indexesA = d_indexesA;
* //h_indexesB = d_indexesB;
*
* //var sumA = h_indexesA.Sum(e => e);
* //var sumB = h_indexesB.Sum(e => e);
*
* ;
* }
* else
* {
* CGraph cg = FinderBag.GetFinder();
* cg.SetEdges(h_a, (int)count[0]);
* cg.SetHeader(currentJob);
*
* Task.Factory.StartNew(() =>
* {
* if (count[0] < 131071)
* {
* try
* {
* if (findersInFlight++ < 3)
* {
* Stopwatch cycleTime = new Stopwatch();
* cycleTime.Start();
* cg.FindSolutions(graphSolutions);
* cycleTime.Stop();
* AdjustTrims(cycleTime.ElapsedMilliseconds);
* }
* else
* Logger.Log(LogLevel.Warning, "CPU overloaded!");
* }
* catch (Exception ex)
* {
* Logger.Log(LogLevel.Warning, "Cycle finder crashed: " + ex.Message);
* }
* finally
* {
* FinderBag.ReturnFinder(cg);
* findersInFlight--;
* }
* }
* });
* }
*
*/
}
catch (Exception ex)
{
Logger.Log(LogLevel.Error, "Critical error in main cuda loop " + ex.Message);
Task.Delay(500).Wait();
break;
}
}
// clean up
try
{
Task.Delay(500).Wait();
Comms.Close();
d_buffer.Dispose();
d_indexesA.Dispose();
d_indexesB.Dispose();
d_aux.Dispose();
streamPrimary.Dispose();
streamSecondary.Dispose();
hAligned_a.Dispose();
if (ctx != null)
{
ctx.Dispose();
}
}
catch { }
}
static int Main()
{
Console.Write("Press enter to test your GPU");
var scr = Console.ReadLine();
try
{
int deviceCount = CudaContext.GetDeviceCount();
if (deviceCount == 0)
{
Console.Error.WriteLine("No CUDA devices detected. Sad face.");
return(-1);
}
Console.WriteLine($"{deviceCount} CUDA devices detected (first will be used)");
for (int i = 0; i < deviceCount; i++)
{
Console.WriteLine($"{i}: {CudaContext.GetDeviceName(i)}");
}
bool GoOn = true;
while (GoOn)
{
//for (int z = 100000; z < 100000*3000; z = z*2)
//{
long z = 50000000;
for (int a = 2; a < 3; a++)
{
Console.WriteLine($"GridDim:{a} n:{z / 1000000} MM");
using (var myGPU = new GPU(deviceId: 0, _count: 1024 * a))
{
// Console.WriteLine("Initializing kernel...");
string log;
var compileResult = myGPU.LoadKernel(out log);
if (compileResult != ManagedCuda.NVRTC.nvrtcResult.Success)
{
Console.Error.WriteLine(compileResult);
Console.Error.WriteLine(log);
Console.ReadLine();
return(-1);
}
//Console.WriteLine(log);
//Tests.Test_0(Count, myGPU);
Tests.Test_1(myGPU, 1, z);
//Tests.Test_2(z, myGPU);
}
Console.WriteLine("Cleaning up...");
}
// Console.ReadKey();
//}
Console.WriteLine("All done; ESC to exit any key to repeat :)");
var k = Console.ReadKey();
GoOn = k.Key.ToString() != Keys.Escape.ToString();
}
} catch (Exception ex)
{
Console.Error.WriteLine(ex.Message);
}
return(0);
}
private void initGLAndCuda()
{
//Create render target control
m_renderControl = new OpenTK.GLControl(GraphicsMode.Default, 1, 0, GraphicsContextFlags.Default);
m_renderControl.Dock = DockStyle.Fill;
m_renderControl.BackColor = Color.White;
m_renderControl.BorderStyle = BorderStyle.FixedSingle;
m_renderControl.KeyDown += new KeyEventHandler(m_renderControl_KeyDown);
m_renderControl.MouseMove += new MouseEventHandler(m_renderControl_MouseMove);
m_renderControl.MouseDown += new MouseEventHandler(m_renderControl_MouseDown);
m_renderControl.SizeChanged += new EventHandler(m_renderControl_SizeChanged);
panel1.Controls.Add(m_renderControl);
Console.WriteLine(" OpenGL device is Available");
int deviceID = CudaContext.GetMaxGflopsDeviceId();
ctx = CudaContext.CreateOpenGLContext(deviceID, CUCtxFlags.BlockingSync);
string console = string.Format("CUDA device [{0}] has {1} Multi-Processors", ctx.GetDeviceName(), ctx.GetDeviceInfo().MultiProcessorCount);
Console.WriteLine(console);
CUmodule module = ctx.LoadModulePTX("kernel.ptx");
addForces_k = new CudaKernel("addForces_k", module, ctx);
advectVelocity_k = new CudaKernel("advectVelocity_k", module, ctx);
diffuseProject_k = new CudaKernel("diffuseProject_k", module, ctx);
updateVelocity_k = new CudaKernel("updateVelocity_k", module, ctx);
advectParticles_k = new CudaKernel("advectParticles_OGL", module, ctx);
hvfield = new cData[DS];
dvfield = new CudaPitchedDeviceVariable <cData>(DIM, DIM);
tPitch = dvfield.Pitch;
dvfield.CopyToDevice(hvfield);
vxfield = new CudaDeviceVariable <cData>(DS);
vyfield = new CudaDeviceVariable <cData>(DS);
// Create particle array
particles = new cData[DS];
initParticles(particles, DIM, DIM);
// TODO: update kernels to use the new unpadded memory layout for perf
// rather than the old FFTW-compatible layout
planr2c = new CudaFFTPlan2D(DIM, DIM, cufftType.R2C, Compatibility.FFTWPadding);
planc2r = new CudaFFTPlan2D(DIM, DIM, cufftType.C2R, Compatibility.FFTWPadding);
GL.GenBuffers(1, out vbo);
GL.BindBuffer(BufferTarget.ArrayBuffer, vbo);
GL.BufferData <cData>(BufferTarget.ArrayBuffer, new IntPtr(cData.SizeOf * DS), particles, BufferUsageHint.DynamicDraw);
int bsize;
GL.GetBufferParameter(BufferTarget.ArrayBuffer, BufferParameterName.BufferSize, out bsize);
if (bsize != DS * cData.SizeOf)
{
throw new Exception("Sizes don't match.");
}
GL.BindBuffer(BufferTarget.ArrayBuffer, 0);
cuda_vbo_resource = new CudaGraphicsInteropResourceCollection();
cuda_vbo_resource.Add(new CudaOpenGLBufferInteropResource(vbo, CUGraphicsRegisterFlags.None));
texref = new CudaTextureArray2D(advectVelocity_k, "texref", CUAddressMode.Wrap, CUFilterMode.Linear, 0, CUArrayFormat.Float, DIM, DIM, CudaArray2DNumChannels.Two);
stopwatch = new CudaStopWatch(CUEventFlags.Default);
reshape();
isInit = true;
display();
}
private bool InitializeD3D()
{
HwndSource hwnd = new HwndSource(0, 0, 0, 0, 0, "null", IntPtr.Zero);
// Create the D3D object.
d3d = new Direct3DEx();
PresentParameters pp = new PresentParameters();
pp.BackBufferWidth = 512;
pp.BackBufferHeight = 512;
pp.BackBufferFormat = Format.Unknown;
pp.BackBufferCount = 0;
pp.Multisample = MultisampleType.None;
pp.MultisampleQuality = 0;
pp.SwapEffect = SwapEffect.Discard;
pp.DeviceWindowHandle = (IntPtr)0;
pp.Windowed = true;
pp.EnableAutoDepthStencil = false;
pp.AutoDepthStencilFormat = Format.Unknown;
pp.PresentationInterval = PresentInterval.Default;
bDeviceFound = false;
CUdevice[] cudaDevices = null;
for (g_iAdapter = 0; g_iAdapter < d3d.AdapterCount; g_iAdapter++)
{
device = new DeviceEx(d3d, d3d.Adapters[g_iAdapter].Adapter, DeviceType.Hardware, hwnd.Handle, CreateFlags.HardwareVertexProcessing | CreateFlags.Multithreaded, pp);
try
{
cudaDevices = CudaContext.GetDirectXDevices(device.ComPointer, CUd3dXDeviceList.All, CudaContext.DirectXVersion.D3D9);
bDeviceFound = cudaDevices.Length > 0;
infoLog.AppendText("> Display Device #" + d3d.Adapters[g_iAdapter].Adapter
+ ": \"" + d3d.Adapters[g_iAdapter].Details.Description + "\" supports Direct3D9 and CUDA.\n");
break;
}
catch (CudaException)
{
//No Cuda device found for this Direct3D9 device
infoLog.AppendText("> Display Device #" + d3d.Adapters[g_iAdapter].Adapter
+ ": \"" + d3d.Adapters[g_iAdapter].Details.Description + "\" supports Direct3D9 but not CUDA.\n");
}
}
// we check to make sure we have found a cuda-compatible D3D device to work on
if (!bDeviceFound)
{
infoLog.AppendText("No CUDA-compatible Direct3D9 device available");
if (device != null)
device.Dispose();
return false;
}
ctx = new CudaContext(cudaDevices[0], device.ComPointer, CUCtxFlags.BlockingSync, CudaContext.DirectXVersion.D3D9);
deviceName.Text = "Device name: " + ctx.GetDeviceName();
// Set projection matrix
SlimDX.Matrix matProj = SlimDX.Matrix.OrthoOffCenterLH(0, 1, 1, 0, 0, 1);
device.SetTransform(TransformState.Projection, matProj);
// Turn off D3D lighting, since we are providing our own vertex colors
device.SetRenderState(RenderState.Lighting, false);
//Load kernels
CUmodule module = ctx.LoadModulePTX("kernel.ptx");
addForces_k = new CudaKernel("addForces_k", module, ctx);
advectVelocity_k = new CudaKernel("advectVelocity_k", module, ctx);
diffuseProject_k = new CudaKernel("diffuseProject_k", module, ctx);
updateVelocity_k = new CudaKernel("updateVelocity_k", module, ctx);
advectParticles_k = new CudaKernel("advectParticles_k", module, ctx);
d3dimage.Lock();
Surface surf = device.GetBackBuffer(0, 0);
d3dimage.SetBackBuffer(D3DResourceType.IDirect3DSurface9, surf.ComPointer);
d3dimage.Unlock();
surf.Dispose();
//Setup the "real" frame rate counter.
//The cuda counter only measures cuda runtime, not the overhead to actually
//show the result via DirectX and WPF.
realLastTick = Environment.TickCount;
return true;
}