private void InitializeCUFFT()
{
g_hvfield = new float2[DS];
g_dvfield = new CudaPitchedDeviceVariable <float2>(DIM, DIM);
g_tPitch = g_dvfield.Pitch; //Store pitch in g_tPitch to keep consistency to the C++ code
g_dvfield.CopyToDevice(g_hvfield);
// Temporary complex velocity field data
g_vxfield = new CudaDeviceVariable <float2>(PDS);
g_vyfield = new CudaDeviceVariable <float2>(PDS);
texref = new CudaTextureArray2D(advectVelocity_k, "texref", CUAddressMode.Wrap, CUFilterMode.Linear, 0, CUArrayFormat.Float, DIM, DIM, CudaArray2DNumChannels.Two);
g_particles = new float2[DS];
initParticles(g_particles, DIM, DIM);
g_planr2c = new CudaFFTPlan2D(DIM, DIM, cufftType.R2C, Compatibility.PADDING);
g_planc2r = new CudaFFTPlan2D(DIM, DIM, cufftType.C2R, Compatibility.PADDING);
}
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);
}
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();
}
