protected void LoadNextField()
{
// Keep t1 timestep as new t0. Update mapping on device side.
_t0X = _t1X;
new CudaTextureArray2D(_advectParticlesKernel, "vX_t0", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t0X);
_t0Y = _t1Y;
new CudaTextureArray2D(_advectParticlesKernel, "vY_t0", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t0Y);
CurrentTime++;
// Load new t1.
LoaderNCF ncFile = RedSea.Singleton.GetLoaderNCF(CurrentTime);
ScalarField t1X = ncFile.LoadFieldSlice(_ensembleRanges[0]);
ScalarField t1Y = ncFile.LoadFieldSlice(_ensembleRanges[1]);
ncFile.Close();
// All members are above each other.
int vHeight = _height * _numMembers;
// vX, t=1
_t1X = new CudaArray2D(CUArrayFormat.Float, _width, vHeight, CudaArray2DNumChannels.One);
_t1X.CopyFromHostToThis(t1X.Data);
new CudaTextureArray2D(_advectParticlesKernel, "vX_t1", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t1X);
// vY, t=1
_t1Y = new CudaArray2D(CUArrayFormat.Float, _width, vHeight, CudaArray2DNumChannels.One);
_t1Y.CopyFromHostToThis(t1Y.Data);
new CudaTextureArray2D(_advectParticlesKernel, "vY_t1", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t1Y);
}
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);
}
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);
}
/// <summary>
/// Setup as empty map with only one value at 1.
/// </summary>
/// <param name="pos"></param>
/// <param name="fieldEnsemble"></param>
/// <param name="startTime"></param>
/// <param name="endTime"></param>
public void SetupPoint(Int2 pos, int startTime)
{
// ~~~~~~~~~~~~~~ Copy relevant data ~~~~~~~~~~~~~~ \\
// Count up when advection was executed.
CurrentTime = startTime;
_startTime = startTime;
// ~~~~~~~~~~~~ Load ensemble ~~~~~~~~~~~~ \\
// Load fields first to get the grid size.
//Loader ncFile = new Loader(RedSea.Singleton.DataFolder + (_startTime + 1) + RedSea.Singleton.FileName);
//ScalarField t0X = ncFile.LoadFieldSlice(_ensembleRanges[0]);
//ScalarField t0Y = ncFile.LoadFieldSlice(_ensembleRanges[1]);
//ncFile.Close();
LoaderNCF ncFile = RedSea.Singleton.GetLoaderNCF(_startTime);
ScalarField t1X = ncFile.LoadFieldSlice(_ensembleRanges[0]);
ScalarField t1Y = ncFile.LoadFieldSlice(_ensembleRanges[1]);
ncFile.Close();
// ~~~~~~~~~~~~~~ Copy relevant data ~~~~~~~~~~~~~~ \\
// Keep for plane creation and size reference.
_ensembleGrid = t1X.Grid as RectlinearGrid;
// Mapper for binding the SlimDX texture to CUDA easily.
_cudaDxMapper = new CudaGraphicsInteropResourceCollection();
// Tell CUDA which value is a border.
_texInvalidValue = t1X.InvalidValue ?? float.MaxValue;
// ~~~~~~~~~~~~ Fill CUDA resources ~~~~~~~~~~~~ \\
// All members are above each other.
int vHeight = _height * _numMembers;
//// vX, t=0
//_t0X = new CudaArray2D(CUArrayFormat.Float, _width, vHeight, CudaArray2DNumChannels.One);
//_t0X.CopyFromHostToThis<float>(t0X.Data);
//new CudaTextureArray2D(_advectParticlesKernel, "vX_t0", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t0X);
//// vY, t=0
//_t0Y = new CudaArray2D(CUArrayFormat.Float, _width, vHeight, CudaArray2DNumChannels.One);
//_t0Y.CopyFromHostToThis<float>(t0Y.Data);
//new CudaTextureArray2D(_advectParticlesKernel, "vY_t0", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t0Y);
// vX, t=1
_t1X = new CudaArray2D(CUArrayFormat.Float, _width, vHeight, CudaArray2DNumChannels.One);
_t1X.CopyFromHostToThis <float>(t1X.Data);
new CudaTextureArray2D(_advectParticlesKernel, "vX_t1", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t1X);
// vY, t=1
_t1Y = new CudaArray2D(CUArrayFormat.Float, _width, vHeight, CudaArray2DNumChannels.One);
_t1Y.CopyFromHostToThis <float>(t1Y.Data);
new CudaTextureArray2D(_advectParticlesKernel, "vY_t1", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, _t1Y);
// ~~~~~~~~~~~~~ Create texture ~~~~~~~~~~~~~~~~~~~~ \\
// Create texture. Completely zero, except for one point.
Texture2DDescription desc = new Texture2DDescription
{
ArraySize = 1,
BindFlags = BindFlags.ShaderResource,
CpuAccessFlags = CpuAccessFlags.None,
Format = Format.R32_Float,
Width = _width,
Height = _height,
MipLevels = 1,
OptionFlags = ResourceOptionFlags.None,
SampleDescription = new SampleDescription(1, 0),
Usage = ResourceUsage.Default
};
// Put field data into stream/rectangle object
float[] zeros = new float[_width * _height];
Array.Clear(zeros, 0, zeros.Length);
// Fill the empty texture.
DataRectangle texData = new DataRectangle(_width * sizeof(float), new DataStream(zeros, true, true));
_pongFlowMap = new CudaDeviceVariable <float>(_width * _height);//new Texture2D(_device, desc, texData);
// Magically, copy to device happens here.
_pongFlowMap = zeros;
// Add one pixel for integration.
zeros[pos.X + pos.Y * _width] = 1;
texData = new DataRectangle(_width * sizeof(float), new DataStream(zeros, true, true));
// Create texture.
FlowMap = new Texture2D(_device, desc, texData);
// ~~~~~~~~~ Make textures mappable to CUDA ~~~~~~~~~~ \\
_cudaDxMapper.Add(new CudaDirectXInteropResource(FlowMap.ComPointer, CUGraphicsRegisterFlags.None, CudaContext.DirectXVersion.D3D11));
_cudaDxMapper.MapAllResources();
CudaArray2D lastFlowMap = _cudaDxMapper[0].GetMappedArray2D(0, 0);
new CudaTextureArray2D(_advectParticlesKernel, "flowMap", CUAddressMode.Wrap, CUFilterMode.Linear, CUTexRefSetFlags.None, lastFlowMap);
_cudaDxMapper.UnmapAllResources();
}
