static void CleanupResources()
{
// Free device memory
if (d_A != null)
{
d_A.Dispose();
}
if (d_B != null)
{
d_B.Dispose();
}
if (d_C != null)
{
d_C.Dispose();
}
if (ctx != null)
{
ctx.Dispose();
}
// Free host memory
// We have a GC for that :-)
}
public void Dispose()
{
baseImageBuffer.Dispose();
nextImageBuffer.Dispose();
pointsBuffer.Dispose();
ctx.Dispose();
}
public void RunExperiment(Grammar grm, bool nvrtc = true, bool remote = false)
{
Initialize(grm); /// --- (1) gp, grammar, population, individual
// Evolution of generations
for (int gen = 0; gen < GENCOUNT; gen++)
{
CreateSourceCodes(PARALLELISM_LEVEL); /// --- (2) create source
PTXcompile(nvrtc, remote); /// --- (3) ptx compile
JITcompile(); /// --- (4) JIT compile
tic();
CreateKernelObjects(); /// ---
LaunchKernels(); /// --- (5) launch kernel
WaitCompleteAndReadBack(); /// ---
GPoperations(); /// --- (6) gp operations: selection, XO, mutation
SampleCollector.Collect("other", toc());
}
ctx.Dispose();
}
void OnDestroy()
{
d_idata.Dispose();
d_odata.Dispose();
d_result_data.Dispose();
ctx.Dispose();
}
private void CleanupResources()
{
// Free device memory
if (d_A != null)
{
d_A?.Dispose();
}
if (d_B != null)
{
d_B?.Dispose();
}
// d_C?.Dispose();
if (C != null)
{
C?.Dispose();
}
if (ctx != null)
{
ctx?.Dispose();
}
// Free host memory
// We have a GC for that :-)
}
static void CleanupResources()
{
// Free device memory
if (dev_v0 != null)
{
dev_v0.Dispose();
}
if (dev_v1 != null)
{
dev_v1.Dispose();
}
if (dev_v2 != null)
{
dev_v2.Dispose();
}
if (dev_da != null)
{
dev_da.Dispose();
}
if (dev_db != null)
{
dev_db.Dispose();
}
if (dev_dc != null)
{
dev_dc.Dispose();
}
if (dev_dOut != null)
{
dev_dOut.Dispose();
}
if (dev_height != null)
{
dev_height.Dispose();
}
if (dev_width != null)
{
dev_width.Dispose();
}
if (ctx != null)
{
ctx.Dispose();
}
// Free host memory
// We have a GC for that :-)
}
public virtual void Dispose()
{
DevicePositions.Dispose();
DevicePersonalBestValues.Dispose();
DeviceVelocities.Dispose();
DevicePersonalBests.Dispose();
_phis1.Dispose();
_phis2.Dispose();
Ctx.Dispose();
}
protected virtual void Dispose(bool disposing)
{
if (disposing && !_disposed)
{
_blas.Dispose();
_cuda.Dispose();
//if(_solver.IsValueCreated)
// _solver.Value.Dispose();
_numerics.Dispose();
_disposed = true;
}
}
public void DisposeCuda()
{
CudaContext context = Context;
if (context != null)
{
Context = null;
try {
context.Dispose();
} catch {
// if this fails, it will probably just mask the actual error further up the stack (e.g. OOM)
}
}
}
protected virtual void Dispose(bool disposing)
{
if (!isDisposed)
{
if (disposing)
{
}
prefixScan.Dispose();
// Unloading every single kernel will cause an error.
context.UnloadModule(module);
context.Dispose();
}
isDisposed = true;
}
//Clean up before closing
private void Form1_FormClosing(object sender, FormClosingEventArgs e)
{
isRunning = false;
isInit = false;
cuda_vbo_resource.Dispose();
texref.Dispose();
dvfield.Dispose();
vxfield.Dispose();
vyfield.Dispose();
planc2r.Dispose();
planr2c.Dispose();
GL.BindBuffer(BufferTarget.ArrayBuffer, 0);
GL.DeleteBuffers(1, ref vbo);
stopwatch.Dispose();
ctx.Dispose();
}
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 void Main(string[] args)
{
int SIGNAL_SIZE = 50;
int FILTER_KERNEL_SIZE = 11;
Console.WriteLine("[simpleCUFFT] is starting...");
var assembly = Assembly.GetExecutingAssembly();
var resourceName = "simpleCUFFT.simpleCUFFTKernel.ptx";
CudaContext ctx = new CudaContext(0);
CudaKernel ComplexPointwiseMulAndScale;
string[] liste = assembly.GetManifestResourceNames();
using (Stream stream = assembly.GetManifestResourceStream(resourceName))
{
ComplexPointwiseMulAndScale = ctx.LoadKernelPTX(stream, "ComplexPointwiseMulAndScale");
}
// Allocate host memory for the signal
cuFloatComplex[] h_signal = new cuFloatComplex[SIGNAL_SIZE]; //we use cuFloatComplex for complex multiplaction in reference host code...
Random rand = new Random(0);
// Initialize the memory for the signal
for (int i = 0; i < SIGNAL_SIZE; ++i)
{
h_signal[i].real = (float)rand.NextDouble();
h_signal[i].imag = 0;
}
// Allocate host memory for the filter
cuFloatComplex[] h_filter_kernel = new cuFloatComplex[FILTER_KERNEL_SIZE];
// Initialize the memory for the filter
for (int i = 0; i < FILTER_KERNEL_SIZE; ++i)
{
h_filter_kernel[i].real = (float)rand.NextDouble();
h_filter_kernel[i].imag = 0;
}
// Pad signal and filter kernel
cuFloatComplex[] h_padded_signal = null;
cuFloatComplex[] h_padded_filter_kernel = null;
int new_size = PadData(h_signal, ref h_padded_signal, SIGNAL_SIZE,
h_filter_kernel, ref h_padded_filter_kernel, FILTER_KERNEL_SIZE);
int mem_size = (int)cuFloatComplex.SizeOf * new_size;
// Allocate device memory for signal
CudaDeviceVariable <cuFloatComplex> d_signal = new CudaDeviceVariable <cuFloatComplex>(new_size);
// Copy host memory to device
d_signal.CopyToDevice(h_padded_signal);
// Allocate device memory for filter kernel
CudaDeviceVariable <cuFloatComplex> d_filter_kernel = new CudaDeviceVariable <cuFloatComplex>(new_size);
// Copy host memory to device
d_filter_kernel.CopyToDevice(h_padded_filter_kernel);
// CUFFT plan simple API
CudaFFTPlan1D plan = new CudaFFTPlan1D(new_size, cufftType.C2C, 1);
// Transform signal and kernel
Console.WriteLine("Transforming signal cufftExecC2C");
plan.Exec(d_signal.DevicePointer, TransformDirection.Forward);
plan.Exec(d_filter_kernel.DevicePointer, TransformDirection.Forward);
// Multiply the coefficients together and normalize the result
Console.WriteLine("Launching ComplexPointwiseMulAndScale<<< >>>");
ComplexPointwiseMulAndScale.BlockDimensions = 256;
ComplexPointwiseMulAndScale.GridDimensions = 32;
ComplexPointwiseMulAndScale.Run(d_signal.DevicePointer, d_filter_kernel.DevicePointer, new_size, 1.0f / new_size);
// Transform signal back
Console.WriteLine("Transforming signal back cufftExecC2C");
plan.Exec(d_signal.DevicePointer, TransformDirection.Inverse);
// Copy device memory to host
cuFloatComplex[] h_convolved_signal = d_signal;
// Allocate host memory for the convolution result
cuFloatComplex[] h_convolved_signal_ref = new cuFloatComplex[SIGNAL_SIZE];
// Convolve on the host
Convolve(h_signal, SIGNAL_SIZE,
h_filter_kernel, FILTER_KERNEL_SIZE,
h_convolved_signal_ref);
// check result
bool bTestResult = sdkCompareL2fe(h_convolved_signal_ref, h_convolved_signal, 1e-5f);
//Destroy CUFFT context
plan.Dispose();
// cleanup memory
d_filter_kernel.Dispose();
d_signal.Dispose();
ctx.Dispose();
if (bTestResult)
{
Console.WriteLine("Test Passed");
}
else
{
Console.WriteLine("Test Failed");
}
}
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 { }
}
static void Main(string[] args)
{
int SIGNAL_SIZE = 50;
int FILTER_KERNEL_SIZE = 11;
Console.WriteLine("[simpleCUFFT] is starting...");
var assembly = Assembly.GetExecutingAssembly();
var resourceName = "simpleCUFFT.simpleCUFFTKernel.ptx";
CudaContext ctx = new CudaContext(0);
CudaKernel ComplexPointwiseMulAndScale;
string[] liste = assembly.GetManifestResourceNames();
using (Stream stream = assembly.GetManifestResourceStream(resourceName))
{
ComplexPointwiseMulAndScale = ctx.LoadKernelPTX(stream, "ComplexPointwiseMulAndScale");
}
// Allocate host memory for the signal
cuFloatComplex[] h_signal = new cuFloatComplex[SIGNAL_SIZE]; //we use cuFloatComplex for complex multiplaction in reference host code...
Random rand = new Random(0);
// Initialize the memory for the signal
for (int i = 0; i < SIGNAL_SIZE; ++i)
{
h_signal[i].real = (float)rand.NextDouble();
h_signal[i].imag = 0;
}
// Allocate host memory for the filter
cuFloatComplex[] h_filter_kernel = new cuFloatComplex[FILTER_KERNEL_SIZE];
// Initialize the memory for the filter
for (int i = 0; i < FILTER_KERNEL_SIZE; ++i)
{
h_filter_kernel[i].real = (float)rand.NextDouble();
h_filter_kernel[i].imag = 0;
}
// Pad signal and filter kernel
cuFloatComplex[] h_padded_signal = null;
cuFloatComplex[] h_padded_filter_kernel = null;
int new_size = PadData(h_signal, ref h_padded_signal, SIGNAL_SIZE,
h_filter_kernel, ref h_padded_filter_kernel, FILTER_KERNEL_SIZE);
int mem_size = (int)cuFloatComplex.SizeOf * new_size;
// Allocate device memory for signal
CudaDeviceVariable<cuFloatComplex> d_signal = new CudaDeviceVariable<cuFloatComplex>(new_size);
// Copy host memory to device
d_signal.CopyToDevice(h_padded_signal);
// Allocate device memory for filter kernel
CudaDeviceVariable<cuFloatComplex> d_filter_kernel = new CudaDeviceVariable<cuFloatComplex>(new_size);
// Copy host memory to device
d_filter_kernel.CopyToDevice(h_padded_filter_kernel);
// CUFFT plan simple API
CudaFFTPlan1D plan = new CudaFFTPlan1D(new_size, cufftType.C2C, 1);
// Transform signal and kernel
Console.WriteLine("Transforming signal cufftExecC2C");
plan.Exec(d_signal.DevicePointer, TransformDirection.Forward);
plan.Exec(d_filter_kernel.DevicePointer, TransformDirection.Forward);
// Multiply the coefficients together and normalize the result
Console.WriteLine("Launching ComplexPointwiseMulAndScale<<< >>>");
ComplexPointwiseMulAndScale.BlockDimensions = 256;
ComplexPointwiseMulAndScale.GridDimensions = 32;
ComplexPointwiseMulAndScale.Run(d_signal.DevicePointer, d_filter_kernel.DevicePointer, new_size, 1.0f / new_size);
// Transform signal back
Console.WriteLine("Transforming signal back cufftExecC2C");
plan.Exec(d_signal.DevicePointer, TransformDirection.Inverse);
// Copy device memory to host
cuFloatComplex[] h_convolved_signal = d_signal;
// Allocate host memory for the convolution result
cuFloatComplex[] h_convolved_signal_ref = new cuFloatComplex[SIGNAL_SIZE];
// Convolve on the host
Convolve(h_signal, SIGNAL_SIZE,
h_filter_kernel, FILTER_KERNEL_SIZE,
h_convolved_signal_ref);
// check result
bool bTestResult = sdkCompareL2fe(h_convolved_signal_ref, h_convolved_signal, 1e-5f);
//Destroy CUFFT context
plan.Dispose();
// cleanup memory
d_filter_kernel.Dispose();
d_signal.Dispose();
ctx.Dispose();
if (bTestResult)
{
Console.WriteLine("Test Passed");
}
else
{
Console.WriteLine("Test Failed");
}
}
public RtRaster <double> RunGpuInterpolation(YieldReportData dataPoints, IGeometry dataBoundary)
{
try
{
int cellSize = 1;
var bounds = new RTBounds(dataBoundary).ToMeters();
var xMax = bounds.MaxX;
var xMin = bounds.MinX;
var yMax = bounds.MaxY;
var yMin = bounds.MinY;
var xDist = (int)(xMax - xMin + 1);
var yDist = (int)(yMax - yMin + 1);
//var cost = (xDist * yDist * dataPoints.Count) / (Environment.ProcessorCount * cellSize * cellSize);
var nCols = xDist / cellSize;
var nRows = yDist / cellSize;
CollectionStats yld = dataPoints.YieldData.Yield;
//create host side arrays for input data and boundary and output data
double[] h_datax = yld.DataPoints.Select(pt => RtPoint <double> .LatLonToMeters(pt)).Select(ptz => ptz.X).ToArray();
double[] h_datay = yld.DataPoints.Select(pt => RtPoint <double> .LatLonToMeters(pt)).Select(ptz => ptz.Y).ToArray();
double[] h_dataz = yld.DataPoints.Select(pt => RtPoint <double> .LatLonToMeters(pt)).Select(ptz => ptz.Z).ToArray();
var dataSize = sizeof(double) * h_datax.Length;
Console.WriteLine("Performing interpolation on {0} data points of {1} bytes", h_datax.Length, dataSize);
var cols = Enumerable.Range(0, nCols - 1).ToArray();
var coords = Enumerable.Range(0, nRows).SelectMany(row => cols.Select(col => Tuple.Create(row, col))).ToArray();
double[] h_outputx = coords.Select(coord => xMin + (coord.Item2 * cellSize)).ToArray();
double[] h_outputy = coords.Select(coord => yMin + (coord.Item1 * cellSize)).ToArray();
double[] h_outputz = coords.Select(coord => 0.0).ToArray();
var outputSize = sizeof(double) * h_outputx.Length;
Console.WriteLine("Output Raster has {0} raster points of {1} bytes", h_outputx.Length, outputSize);
double2[] h_boundCoords = dataBoundary.Coordinates.Select(coord => new double2(coord.X, coord.Y)).ToArray();
using (CudaDeviceVariable <double> d_datax = h_datax)
using (CudaDeviceVariable <double> d_datay = h_datay)
using (CudaDeviceVariable <double> d_dataz = h_dataz)
using (CudaDeviceVariable <double> d_outputx = h_outputx)
using (CudaDeviceVariable <double> d_outputy = h_outputy)
using (CudaDeviceVariable <double> d_outputz = h_outputz)
{
const int ThreadsPerBlock = 512;
InterpolateKernel.BlockDimensions = ThreadsPerBlock;
InterpolateKernel.GridDimensions = (h_outputx.Length + ThreadsPerBlock - 1) / ThreadsPerBlock;
Console.WriteLine("Invoke Kernel with Grid Dimensions = {0}", InterpolateKernel.GridDimensions);
Stopwatch timer = new Stopwatch();
timer.Start();
InterpolateKernel.Run(d_datax.DevicePointer, d_datay.DevicePointer, d_dataz.DevicePointer,
d_outputx.DevicePointer, d_outputy.DevicePointer, d_outputz.DevicePointer,
h_outputx.Length, h_datax.Length);
timer.Stop();
var elapsed = timer.ElapsedTicks;
Console.WriteLine("Gpu Interpolation took {0} ms", elapsed / (Stopwatch.Frequency / 1000));
h_outputz = d_outputz;
}
if (h_outputz.All(pt => pt != 0.0))
{
Console.WriteLine("Kernel succeeded");
}
else
{
Console.WriteLine("Kernel failed");
}
var raster = new RtRaster <double>(nCols, nRows, xMin, yMin, cellSize * 2);
var noDataValue = 0.0f;
var rasterBand = RtRasterBand <double> .CreateRasterBand(typeof(double), nCols, nRows, noDataValue, h_outputz);
raster.SetRasterBand(0, rasterBand);
return(raster);
}
catch (Exception ex)
{
ctx.Dispose();
Console.WriteLine(ex.Message + "\nStackTrace: " + ex.StackTrace);
//+
// ex.InnerException == null ? "" :
// String.Format("\n\t{0}\n\tStackTrace: {1}", ex.InnerException.Message,
// ex.InnerException.StackTrace)
return(null);
}
}
static void Main(string[] args)
{
const int nx = 2048;
const int ny = 2048;
// shifts applied to x and y data
const int x_shift = 5;
const int y_shift = 7;
ShrQATest.shrQAStart(args);
if ((nx % TILE_DIM != 0) || (ny % TILE_DIM != 0))
{
Console.Write("nx and ny must be multiples of TILE_DIM\n");
ShrQATest.shrQAFinishExit(args, ShrQATest.eQAstatus.QA_WAIVED);
}
// execution configuration parameters
dim3 grid = new dim3(nx / TILE_DIM, ny / TILE_DIM, 1);
dim3 threads = new dim3(TILE_DIM, TILE_DIM, 1);
// This will pick the best possible CUDA capable device
int devID = findCudaDevice(args);
//Load Kernel image from resources
string resName;
if (IntPtr.Size == 8)
{
resName = "simplePitchLinearTexture_x64.ptx";
}
else
{
resName = "simplePitchLinearTexture.ptx";
}
string resNamespace = "simplePitchLinearTexture";
string resource = resNamespace + "." + resName;
Stream stream = Assembly.GetExecutingAssembly().GetManifestResourceStream(resource);
if (stream == null)
{
throw new ArgumentException("Kernel not found in resources.");
}
byte[] kernels = new byte[stream.Length];
int bytesToRead = (int)stream.Length;
while (bytesToRead > 0)
{
bytesToRead -= stream.Read(kernels, (int)stream.Position, bytesToRead);
}
CudaKernel PLKernel = ctx.LoadKernelPTX(kernels, "shiftPitchLinear");
CudaKernel ArrayKernel = ctx.LoadKernelPTX(kernels, "shiftArray");
CudaStopWatch stopwatch = new CudaStopWatch();
// ----------------------------------
// Host allocation and initialization
// ----------------------------------
float[] h_idata = new float[nx * ny];
float[] h_odata = new float[nx * ny];
float[] gold = new float[nx * ny];
for (int i = 0; i < nx * ny; ++i)
{
h_idata[i] = (float)i;
}
// ------------------------
// Device memory allocation
// ------------------------
// Pitch linear input data
CudaPitchedDeviceVariable <float> d_idataPL = new CudaPitchedDeviceVariable <float>(nx, ny);
// Array input data
CudaArray2D d_idataArray = new CudaArray2D(CUArrayFormat.Float, nx, ny, CudaArray2DNumChannels.One);
// Pitch linear output data
CudaPitchedDeviceVariable <float> d_odata = new CudaPitchedDeviceVariable <float>(nx, ny);
// ------------------------
// copy host data to device
// ------------------------
// Pitch linear
d_idataPL.CopyToDevice(h_idata);
// Array
d_idataArray.CopyFromHostToThis <float>(h_idata);
// ----------------------
// Bind texture to memory
// ----------------------
// Pitch linear
CudaTextureLinearPitched2D <float> texRefPL = new CudaTextureLinearPitched2D <float>(PLKernel, "texRefPL", CUAddressMode.Wrap, CUFilterMode.Point, CUTexRefSetFlags.NormalizedCoordinates, CUArrayFormat.Float, d_idataPL);
CudaTextureArray2D texRefArray = new CudaTextureArray2D(ArrayKernel, "texRefArray", CUAddressMode.Wrap, CUFilterMode.Point, CUTexRefSetFlags.NormalizedCoordinates, d_idataArray);
// ---------------------
// reference calculation
// ---------------------
for (int j = 0; j < ny; j++)
{
int jshift = (j + y_shift) % ny;
for (int i = 0; i < nx; i++)
{
int ishift = (i + x_shift) % nx;
gold[j * nx + i] = h_idata[jshift * nx + ishift];
}
}
// ----------------
// shiftPitchLinear
// ----------------
ctx.ClearMemory(d_odata.DevicePointer, 0, d_odata.TotalSizeInBytes);
PLKernel.BlockDimensions = threads;
PLKernel.GridDimensions = grid;
stopwatch.Start();
for (int i = 0; i < NUM_REPS; i++)
{
PLKernel.Run(d_odata.DevicePointer, (int)(d_odata.Pitch / sizeof(float)), nx, ny, x_shift, y_shift);
}
stopwatch.Stop();
stopwatch.StopEvent.Synchronize();
float timePL = stopwatch.GetElapsedTime();
// check results
d_odata.CopyToHost(h_odata);
bool res = cutComparef(gold, h_odata);
bool success = true;
if (res == false)
{
Console.Write("*** shiftPitchLinear failed ***\n");
success = false;
}
// ----------
// shiftArray
// ----------
ctx.ClearMemory(d_odata.DevicePointer, 0, d_odata.TotalSizeInBytes);
ArrayKernel.BlockDimensions = threads;
ArrayKernel.GridDimensions = grid;
stopwatch.Start();
for (int i = 0; i < NUM_REPS; i++)
{
ArrayKernel.Run(d_odata.DevicePointer, (int)(d_odata.Pitch / sizeof(float)), nx, ny, x_shift, y_shift);
}
stopwatch.Stop();
stopwatch.StopEvent.Synchronize();
float timeArray = stopwatch.GetElapsedTime();
// check results
d_odata.CopyToHost(h_odata);
res = cutComparef(gold, h_odata);
if (res == false)
{
Console.Write("*** shiftArray failed ***\n");
success = false;
}
float bandwidthPL = 2.0f * 1000.0f * nx * ny * sizeof(float) / (1e+9f) / (timePL / NUM_REPS);
float bandwidthArray = 2.0f * 1000.0f * nx * ny * sizeof(float) / (1e+9f) / (timeArray / NUM_REPS);
Console.Write("\nBandwidth (GB/s) for pitch linear: {0}; for array: {1}\n",
bandwidthPL, bandwidthArray);
float fetchRatePL = nx * ny / 1e+6f / (timePL / (1000.0f * NUM_REPS));
float fetchRateArray = nx * ny / 1e+6f / (timeArray / (1000.0f * NUM_REPS));
Console.Write("\nTexture fetch rate (Mpix/s) for pitch linear: {0}; for array: {1}\n\n",
fetchRatePL, fetchRateArray);
// cleanup
texRefPL.Dispose();
texRefArray.Dispose();
d_idataPL.Dispose();
d_idataArray.Dispose();
d_odata.Dispose();
stopwatch.Dispose();
ctx.Dispose();
ShrQATest.shrQAFinishExit(args, (success == true) ? ShrQATest.eQAstatus.QA_PASSED : ShrQATest.eQAstatus.QA_FAILED);
}
void OnDestroy()
{
ctx.Dispose();
}
private void myWindow_Closing(object sender, System.ComponentModel.CancelEventArgs e)
{
//Stop render loop before closing
if (frameTimer != null)
{
frameTimer.Tick -= new EventHandler(frameTimer_Tick);
frameTimer.Stop();
}
//Cleanup
if (graphicsres != null)
{
graphicsres.Dispose();
}
if (g_mparticles != null)
{
g_mparticles.Dispose();
}
if (stopwatch != null)
{
stopwatch.Dispose();
}
if (texref != null)
{
texref.Dispose();
}
if (g_dvfield != null)
{
g_dvfield.Dispose();
}
if (g_vxfield != null)
{
g_vxfield.Dispose();
}
if (g_vyfield != null)
{
g_vyfield.Dispose();
}
if (g_planc2r != null)
{
g_planc2r.Dispose();
}
if (g_planr2c != null)
{
g_planr2c.Dispose();
}
if (g_pVB != null)
{
g_pVB.Dispose();
}
if (g_pTexture != null)
{
g_pTexture.Dispose();
}
if (device != null)
{
device.Dispose();
}
if (d3d != null)
{
d3d.Dispose();
}
if (ctx != null)
{
ctx.Dispose();
}
}
public override void Dispose()
{
cuBlas.Dispose();
cudaContext.Dispose();
}
static void Main(string[] args)
{
string filename = "vectorAdd_kernel.cu"; //we assume the file is in the same folder...
string fileToCompile = File.ReadAllText(filename);
CudaRuntimeCompiler rtc = new CudaRuntimeCompiler(fileToCompile, "vectorAdd_kernel");
rtc.Compile(args);
string log = rtc.GetLogAsString();
Console.WriteLine(log);
byte[] ptx = rtc.GetPTX();
rtc.Dispose();
CudaContext ctx = new CudaContext(0);
CudaKernel vectorAdd = ctx.LoadKernelPTX(ptx, "vectorAdd");
// Print the vector length to be used, and compute its size
int numElements = 50000;
SizeT size = numElements * sizeof(float);
Console.WriteLine("[Vector addition of {0} elements]", numElements);
// Allocate the host input vector A
float[] h_A = new float[numElements];
// Allocate the host input vector B
float[] h_B = new float[numElements];
// Allocate the host output vector C
float[] h_C = new float[numElements];
Random rand = new Random(0);
// Initialize the host input vectors
for (int i = 0; i < numElements; ++i)
{
h_A[i] = (float)rand.NextDouble();
h_B[i] = (float)rand.NextDouble();
}
Console.WriteLine("Allocate and copy input data from the host memory to the CUDA device\n");
// Allocate the device input vector A and copy to device
CudaDeviceVariable<float> d_A = h_A;
// Allocate the device input vector B and copy to device
CudaDeviceVariable<float> d_B = h_B;
// Allocate the device output vector C
CudaDeviceVariable<float> d_C = new CudaDeviceVariable<float>(numElements);
// Launch the Vector Add CUDA Kernel
int threadsPerBlock = 256;
int blocksPerGrid = (numElements + threadsPerBlock - 1) / threadsPerBlock;
Console.WriteLine("CUDA kernel launch with {0} blocks of {1} threads\n", blocksPerGrid, threadsPerBlock);
vectorAdd.BlockDimensions = new dim3(threadsPerBlock,1, 1);
vectorAdd.GridDimensions = new dim3(blocksPerGrid, 1, 1);
vectorAdd.Run(d_A.DevicePointer, d_B.DevicePointer, d_C.DevicePointer, numElements);
// Copy the device result vector in device memory to the host result vector
// in host memory.
Console.WriteLine("Copy output data from the CUDA device to the host memory\n");
d_C.CopyToHost(h_C);
// Verify that the result vector is correct
for (int i = 0; i < numElements; ++i)
{
if (Math.Abs(h_A[i] + h_B[i] - h_C[i]) > 1e-5)
{
Console.WriteLine("Result verification failed at element {0}!\n", i);
return;
}
}
Console.WriteLine("Test PASSED\n");
// Free device global memory
d_A.Dispose();
d_B.Dispose();
d_C.Dispose();
ctx.Dispose();
Console.WriteLine("Done\n");
}
static void Main(string[] args)
{
string filename = "vectorAdd_kernel.cu"; //we assume the file is in the same folder...
string fileToCompile = File.ReadAllText(filename);
CudaRuntimeCompiler rtc = new CudaRuntimeCompiler(fileToCompile, "vectorAdd_kernel");
rtc.Compile(args);
string log = rtc.GetLogAsString();
Console.WriteLine(log);
byte[] ptx = rtc.GetPTX();
rtc.Dispose();
CudaContext ctx = new CudaContext(0);
CudaKernel vectorAdd = ctx.LoadKernelPTX(ptx, "vectorAdd");
// Print the vector length to be used, and compute its size
int numElements = 50000;
SizeT size = numElements * sizeof(float);
Console.WriteLine("[Vector addition of {0} elements]", numElements);
// Allocate the host input vector A
float[] h_A = new float[numElements];
// Allocate the host input vector B
float[] h_B = new float[numElements];
// Allocate the host output vector C
float[] h_C = new float[numElements];
Random rand = new Random(0);
// Initialize the host input vectors
for (int i = 0; i < numElements; ++i)
{
h_A[i] = (float)rand.NextDouble();
h_B[i] = (float)rand.NextDouble();
}
Console.WriteLine("Allocate and copy input data from the host memory to the CUDA device\n");
// Allocate the device input vector A and copy to device
CudaDeviceVariable <float> d_A = h_A;
// Allocate the device input vector B and copy to device
CudaDeviceVariable <float> d_B = h_B;
// Allocate the device output vector C
CudaDeviceVariable <float> d_C = new CudaDeviceVariable <float>(numElements);
// Launch the Vector Add CUDA Kernel
int threadsPerBlock = 256;
int blocksPerGrid = (numElements + threadsPerBlock - 1) / threadsPerBlock;
Console.WriteLine("CUDA kernel launch with {0} blocks of {1} threads\n", blocksPerGrid, threadsPerBlock);
vectorAdd.BlockDimensions = new dim3(threadsPerBlock, 1, 1);
vectorAdd.GridDimensions = new dim3(blocksPerGrid, 1, 1);
vectorAdd.Run(d_A.DevicePointer, d_B.DevicePointer, d_C.DevicePointer, numElements);
// Copy the device result vector in device memory to the host result vector
// in host memory.
Console.WriteLine("Copy output data from the CUDA device to the host memory\n");
d_C.CopyToHost(h_C);
// Verify that the result vector is correct
for (int i = 0; i < numElements; ++i)
{
if (Math.Abs(h_A[i] + h_B[i] - h_C[i]) > 1e-5)
{
Console.WriteLine("Result verification failed at element {0}!\n", i);
return;
}
}
Console.WriteLine("Test PASSED\n");
// Free device global memory
d_A.Dispose();
d_B.Dispose();
d_C.Dispose();
ctx.Dispose();
Console.WriteLine("Done\n");
}
private void Form1_FormClosing(object sender, FormClosingEventArgs e)
{
isRunning = false;
//Cleanup
if (graphicsres != null)
{
graphicsres.Dispose();
}
if (g_mparticles != null)
{
g_mparticles.Dispose();
}
if (stopwatch != null)
{
stopwatch.Dispose();
}
if (texref != null)
{
texref.Dispose();
}
if (g_dvfield != null)
{
g_dvfield.Dispose();
}
if (g_vxfield != null)
{
g_vxfield.Dispose();
}
if (g_vyfield != null)
{
g_vyfield.Dispose();
}
if (g_planc2r != null)
{
g_planc2r.Dispose();
}
if (g_planr2c != null)
{
g_planr2c.Dispose();
}
if (g_pVB != null)
{
g_pVB.Dispose();
}
if (g_pTexture != null)
{
g_pTexture.Dispose();
}
if (device != null)
{
device.Dispose();
}
if (d3d != null)
{
d3d.Dispose();
}
if (ctx != null)
{
ctx.Dispose();
}
}
public void cuFFTreconstruct()
{
CudaContext ctx = new CudaContext(0);
ManagedCuda.BasicTypes.CUmodule cumodule = ctx.LoadModule("kernel.ptx");
CudaKernel cuKernel = new CudaKernel("cu_ArrayInversion", cumodule, ctx);
float2[] fData = new float2[Resolution * Resolution];
float2[] result = new float2[Resolution * Resolution];
FFTData2D = new float[Resolution, Resolution, 2];
CudaDeviceVariable <float2> devData = new CudaDeviceVariable <float2>(Resolution * Resolution);
CudaDeviceVariable <float2> copy_devData = new CudaDeviceVariable <float2>(Resolution * Resolution);
int i, j;
Random rnd = new Random();
double avrg = 0.0;
for (i = 0; i < Resolution; i++)
{
for (j = 0; j < Resolution; j++)
{
fData[i * Resolution + j].x = i + j * 2;
avrg += fData[i * Resolution + j].x;
fData[i * Resolution + j].y = 0.0f;
}
}
avrg = avrg / (double)(Resolution * Resolution);
for (i = 0; i < Resolution; i++)
{
for (j = 0; j < Resolution; j++)
{
fData[(i * Resolution + j)].x = fData[(i * Resolution + j)].x - (float)avrg;
}
}
devData.CopyToDevice(fData);
CudaFFTPlan1D plan1D = new CudaFFTPlan1D(Resolution, cufftType.C2C, Resolution);
plan1D.Exec(devData.DevicePointer, TransformDirection.Forward);
cuKernel.GridDimensions = new ManagedCuda.VectorTypes.dim3(Resolution / cuda_blockNum, Resolution, 1);
cuKernel.BlockDimensions = new ManagedCuda.VectorTypes.dim3(cuda_blockNum, 1, 1);
cuKernel.Run(devData.DevicePointer, copy_devData.DevicePointer, Resolution);
copy_devData.CopyToHost(result);
for (i = 0; i < Resolution; i++)
{
for (j = 0; j < Resolution; j++)
{
FFTData2D[i, j, 0] = result[i * Resolution + j].x;
FFTData2D[i, j, 1] = result[i * Resolution + j].y;
}
}
//Clean up
devData.Dispose();
copy_devData.Dispose();
plan1D.Dispose();
CudaContext.ProfilerStop();
ctx.Dispose();
}
/// <summary>
/// Performs application-defined tasks associated with freeing, releasing, or resetting unmanaged resources.
/// </summary>
public void Dispose()
{
BlasHandles.Dispose();
CudaContext.Dispose();
this.MemoryAllocator.Dispose();
}
public List <float> hypotesis_long_save(List <double> xx, List <double> h, int N)
{
int n = (int)Math.Ceiling((double)(xx.Count() + 0.000000000001) / N);
double[] temp_data = new double[n * (N + h.Count - 1) - (n - 1) * (h.Count - 1)];
xx.CopyTo(temp_data, h.Count - 1);
List <double> x = temp_data.ToList();
//int N = 2000000;
string path = Path.GetDirectoryName(mv.plugins[0].filename);
CudaContext ctx = new CudaContext();
CudaKernel kernel = ctx.LoadKernel(path + "\\kernel.ptx", "ComplexMultCUDA");
kernel.GridDimensions = (int)Math.Ceiling((double)(N + h.Count - 1) / 1024);
kernel.BlockDimensions = 1024;
int blocks = (int)Math.Ceiling((double)(x.Count + h.Count - 1) / (N + h.Count - 1));
double[][] temp_y = new double[n][];
double[] temp_h = new double[N + h.Count - 1];
double[] temp_x = new double[N + h.Count - 1];
h.ToArray().CopyTo(temp_h, 0);
CudaDeviceVariable <double> d_x = null;
CudaDeviceVariable <double> d_h = new CudaDeviceVariable <double>(N + h.Count - 1);
CudaDeviceVariable <double2> d_H = new CudaDeviceVariable <double2>(N + h.Count - 1);
//CudaDeviceVariable<double> d_y = new CudaDeviceVariable<double>(N + h.Count - 1);
CudaFFTPlan1D planForward = new CudaFFTPlan1D(N + h.Count - 1, cufftType.D2Z, 1);
CudaFFTPlan1D planInverse = new CudaFFTPlan1D(N + h.Count - 1, cufftType.Z2D, 1);
try
{
d_h = temp_h;
planForward.Exec(d_h.DevicePointer, d_H.DevicePointer, TransformDirection.Forward);
}
catch (Exception exp)
{
mainView.log(exp, "CUDA error: Impulse response FFT", this);
return(null);
}
for (int g = 0; g < n; g++)
{
CudaDeviceVariable <double2> d_X = new CudaDeviceVariable <double2>(N + h.Count - 1);
int P = N + h.Count - 1;
//if (x.Count - P * g < P) P = x.Count - P * g;
int L = h.Count - 1;
if (g == 0)
{
L = 0;
}
x.CopyTo(P * g - L * g, temp_x, 0, P);
try
{
d_x = temp_x;
planForward.Exec(d_x.DevicePointer, d_X.DevicePointer);
kernel.Run(d_H.DevicePointer, d_X.DevicePointer, N + h.Count - 1);
planInverse.Exec(d_X.DevicePointer, d_x.DevicePointer);
}
catch (Exception exp)
{
mainView.log(exp, "Cuda error: kernel run cuda error", this);
}
temp_y[g] = d_x;
d_x.Dispose();
d_X.Dispose();
}
planForward.Dispose();
planInverse.Dispose();
d_x.Dispose();
d_h.Dispose();
d_H.Dispose();
ctx.Dispose();
return(OverlapSave(temp_y, h.Count, N + h.Count - 1).GetRange(h.Count / 2, xx.Count));
}