@Override
public Pointer getHostPointer(int offset) {
if (hostPointer == null) {
hostPointer = Pointer.to(asNio());
}
return hostPointer.withByteOffset(offset * getElementSize());
}
static FloatBuffer getUnifiedFloatBuffer(Pointer pinnedMemory, CUdeviceptr devicePtr, long size) {
JCudaDriver.cuMemHostAlloc(pinnedMemory, size, JCudaDriver.CU_MEMHOSTALLOC_DEVICEMAP);
final ByteBuffer byteBuffer = pinnedMemory.getByteBuffer(0, size);
byteBuffer.order(ByteOrder.nativeOrder());
JCudaDriver.cuMemHostGetDevicePointer(devicePtr, pinnedMemory, 0);
return byteBuffer.asFloatBuffer();
}
@Override
public Pointer getHostPointer() {
if (hostPointer == null) {
hostPointer = Pointer.to(asNio());
}
return hostPointer;
}
public void copyDtoH() {
if (System.getProperty("use_cuda").equals("true")) {
if (this.cPointer != null) {
// JCublas.cublasInit();
SimpleCuBlas.getData(this, this.cPointer, Pointer.to(this.data()));
}
}
}
public static CUdeviceptr allocateInput(float[] input) {
int typeSize = Sizeof.FLOAT;
Pointer ptr = Pointer.to(input);
int size = input.length;
CUdeviceptr dInput = new CUdeviceptr();
cuMemAlloc(dInput, size * Sizeof.FLOAT);
cuMemcpyHtoD(dInput, ptr, size * typeSize);
return dInput;
}
public static int[] getUnifiedIntArray(Pointer pinnedMemory, CUdeviceptr devicePtr, int size) {
int[] values = new int[size];
JCudaDriver.cuMemHostAlloc(pinnedMemory, size, JCudaDriver.CU_MEMHOSTALLOC_DEVICEMAP);
final ByteBuffer byteBuffer = pinnedMemory.getByteBuffer(0, size);
byteBuffer.order(ByteOrder.nativeOrder());
JCudaDriver.cuMemHostGetDevicePointer(devicePtr, pinnedMemory, 0);
return values;
}
private synchronized void unload() {
if (initialized) {
if (projectionArray != null) {
JCudaDriver.cuArrayDestroy(projectionArray);
}
int reconDimensionX = getGeometry().getReconDimensionX();
int reconDimensionY = getGeometry().getReconDimensionY();
int reconDimensionZ = getGeometry().getReconDimensionZ();
if ((projectionVolume != null) && (!largeVolumeMode)) {
// fetch data
int memorysize = reconDimensionX * reconDimensionY * reconDimensionZ * 4;
JCudaDriver.cuMemcpyDtoH(Pointer.to(h_volume), volumePointer, memorysize);
int width = projectionVolume.getSize()[0];
int height = projectionVolume.getSize()[1];
if (this.useVOImap) {
for (int k = 0; k < projectionVolume.getSize()[2]; k++) {
for (int j = 0; j < height; j++) {
for (int i = 0; i < width; i++) {
float value = h_volume[(((height * k) + j) * width) + i];
if (voiMap[i][j][k]) {
projectionVolume.setAtIndex(i, j, k, value);
} else {
projectionVolume.setAtIndex(i, j, k, 0);
}
}
}
}
} else {
for (int k = 0; k < projectionVolume.getSize()[2]; k++) {
for (int j = 0; j < height; j++) {
for (int i = 0; i < width; i++) {
float value = h_volume[(((height * k) + j) * width) + i];
projectionVolume.setAtIndex(i, j, k, value);
}
}
}
}
} else {
System.out.println("Check ProjectionVolume. It seems null.");
}
h_volume = null;
// free memory on device
JCudaDriver.cuMemFree(volumePointer);
// destory context
JCudaDriver.cuCtxDestroy(cuCtx);
reset();
initialized = false;
}
}
private synchronized void initProjectionData(Grid2D projection) {
initialize(projection);
if (projection != null) {
float[] proj = new float[projection.getWidth() * projection.getHeight()];
for (int i = 0; i < projection.getWidth(); i++) {
for (int j = 0; j < projection.getHeight(); j++) {
proj[(j * projection.getWidth()) + i] = projection.getPixelValue(i, j);
}
}
if (projectionArray == null) {
// Create the 2D array that will contain the
// projection data.
projectionArray = new CUarray();
CUDA_ARRAY_DESCRIPTOR ad = new CUDA_ARRAY_DESCRIPTOR();
ad.Format = CUarray_format.CU_AD_FORMAT_FLOAT;
ad.Width = projection.getWidth();
ad.Height = projection.getHeight();
ad.NumChannels = 1; // projection.getNChannels();
JCudaDriver.cuArrayCreate(projectionArray, ad);
}
// Copy the projection data to the array
CUDA_MEMCPY2D copy2 = new CUDA_MEMCPY2D();
copy2.srcMemoryType = CUmemorytype.CU_MEMORYTYPE_HOST;
copy2.srcHost = Pointer.to(proj);
copy2.srcPitch = projection.getWidth() * Sizeof.FLOAT;
copy2.dstMemoryType = CUmemorytype.CU_MEMORYTYPE_ARRAY;
copy2.dstArray = projectionArray;
copy2.WidthInBytes = projection.getWidth() * Sizeof.FLOAT;
copy2.Height = projection.getHeight();
JCudaDriver.cuMemcpy2D(copy2);
// Obtain the texture reference from the module,
// set its parameters and assign the projection
// array as its reference.
projectionTex = new CUtexref();
JCudaDriver.cuModuleGetTexRef(projectionTex, module, "gTex2D");
JCudaDriver.cuTexRefSetFilterMode(projectionTex, CUfilter_mode.CU_TR_FILTER_MODE_LINEAR);
JCudaDriver.cuTexRefSetAddressMode(projectionTex, 0, CUaddress_mode.CU_TR_ADDRESS_MODE_CLAMP);
JCudaDriver.cuTexRefSetFlags(projectionTex, JCudaDriver.CU_TRSF_READ_AS_INTEGER);
JCudaDriver.cuTexRefSetFormat(projectionTex, CUarray_format.CU_AD_FORMAT_FLOAT, 4);
JCudaDriver.cuTexRefSetArray(
projectionTex, projectionArray, JCudaDriver.CU_TRSA_OVERRIDE_FORMAT);
// Set the texture references as parameters for the function call
JCudaDriver.cuParamSetTexRef(function, JCudaDriver.CU_PARAM_TR_DEFAULT, projectionTex);
} else {
System.out.println("Projection was null!!");
}
}
private synchronized void initProjectionMatrix(int projectionNumber) {
// load projection Matrix for current Projection.
SimpleMatrix pMat = getGeometry().getProjectionMatrix(projectionNumber).computeP();
float[] pMatFloat = new float[pMat.getCols() * pMat.getRows()];
for (int j = 0; j < pMat.getRows(); j++) {
for (int i = 0; i < pMat.getCols(); i++) {
pMatFloat[(j * pMat.getCols()) + i] = (float) pMat.getElement(j, i);
}
}
JCudaDriver.cuMemcpyHtoD(
projectionMatrix, Pointer.to(pMatFloat), Sizeof.FLOAT * pMatFloat.length);
}
@Override
public void assign(int[] indices, double[] data, boolean contiguous, int inc) {
if (indices.length != data.length)
throw new IllegalArgumentException("Indices and data length must be the same");
if (indices.length > length())
throw new IllegalArgumentException(
"More elements than space to assign. This buffer is of length "
+ length()
+ " where the indices are of length "
+ data.length);
if (contiguous) {
int offset = indices[0];
Pointer p = Pointer.to(data);
set(offset, data.length, p, inc);
} else throw new UnsupportedOperationException("Non contiguous is not supported");
}
private static Pointer pointerTo(double value) {
return Pointer.to(new double[] {value});
}
public void cudaRun() {
try {
while (projectionsAvailable.size() > 0) {
Thread.sleep(CONRAD.INVERSE_SPEEDUP);
if (showStatus) {
float status = (float) (1.0 / projections.size());
if (largeVolumeMode) {
IJ.showStatus("Streaming Projections to CUDA Buffer");
} else {
IJ.showStatus("Backprojecting with CUDA");
}
IJ.showProgress(status);
}
if (!largeVolumeMode) {
workOnProjectionData();
} else {
checkProjectionData();
}
}
System.out.println("large Volume " + largeVolumeMode);
if (largeVolumeMode) {
// we have collected all projections.
// now we can reconstruct subvolumes and stich them together.
int reconDimensionX = getGeometry().getReconDimensionX();
int reconDimensionY = getGeometry().getReconDimensionY();
int reconDimensionZ = getGeometry().getReconDimensionZ();
double voxelSpacingX = getGeometry().getVoxelSpacingX();
double voxelSpacingY = getGeometry().getVoxelSpacingY();
double voxelSpacingZ = getGeometry().getVoxelSpacingZ();
useVOImap = false;
initialize(projections.get(0));
double originalOffsetZ = offsetZ;
double originalReconDimZ = reconDimensionZ;
reconDimensionZ = subVolumeZ;
int memorysize = reconDimensionX * reconDimensionY * subVolumeZ * Sizeof.FLOAT;
int maxProjectionNumber = projections.size();
float all = nSteps * maxProjectionNumber * 2;
for (int n = 0; n < nSteps; n++) { // For each subvolume
// set all to 0;
Arrays.fill(h_volume, 0);
JCudaDriver.cuMemcpyHtoD(volumePointer, Pointer.to(h_volume), memorysize);
offsetZ = originalOffsetZ - (reconDimensionZ * voxelSpacingZ * n);
for (int p = 0; p < maxProjectionNumber; p++) { // For all projections
float currentStep = (n * maxProjectionNumber * 2) + p;
if (showStatus) {
IJ.showStatus("Backprojecting with CUDA");
IJ.showProgress(currentStep / all);
}
// System.out.println("Current: " + p);
try {
projectSingleProjection(p, reconDimensionZ);
} catch (Exception e) {
System.out.println("Backprojection of projection " + p + " was not successful.");
e.printStackTrace();
}
}
// Gather volume
JCudaDriver.cuMemcpyDtoH(Pointer.to(h_volume), volumePointer, memorysize);
// move data to ImagePlus;
if (projectionVolume != null) {
for (int k = 0; k < reconDimensionZ; k++) {
int index = (n * subVolumeZ) + k;
if (showStatus) {
float currentStep = (n * maxProjectionNumber * 2) + maxProjectionNumber + k;
IJ.showStatus("Fetching Volume from CUDA");
IJ.showProgress(currentStep / all);
}
if (index < originalReconDimZ) {
for (int j = 0; j < projectionVolume.getSize()[1]; j++) {
for (int i = 0; i < projectionVolume.getSize()[0]; i++) {
double[][] voxel = new double[4][1];
int idx =
(((projectionVolume.getSize()[1] * k) + j) * projectionVolume.getSize()[0])
+ i;
float value = h_volume[idx];
voxel[0][0] = (voxelSpacingX * i) - offsetX;
voxel[1][0] = (voxelSpacingY * j) - offsetY;
voxel[2][0] = (voxelSpacingZ * index) - originalOffsetZ;
// exception for the case "interestedInVolume == null" and largeVolume is
// enabled
if (interestedInVolume == null) {
projectionVolume.setAtIndex(i, j, index, value);
} else {
if (interestedInVolume.contains(voxel[0][0], voxel[1][0], voxel[2][0])) {
projectionVolume.setAtIndex(i, j, index, value);
} else {
projectionVolume.setAtIndex(i, j, index, 0);
}
}
}
}
}
}
}
}
}
} catch (InterruptedException e) {
e.printStackTrace();
}
if (showStatus) IJ.showProgress(1.0);
unload();
if (debug) System.out.println("Unloaded");
}
private synchronized void projectSingleProjection(int projectionNumber, int dimz) {
// load projection matrix
initProjectionMatrix(projectionNumber);
// load projection
Grid2D projection = (Grid2D) projections.get(projectionNumber).clone();
// Correct for constant part of distance weighting + For angular sampling
double D = getGeometry().getSourceToDetectorDistance();
NumericPointwiseOperators.multiplyBy(
projection, (float) (D * D * 2 * Math.PI / getGeometry().getNumProjectionMatrices()));
initProjectionData(projection);
if (!largeVolumeMode) {
projections.remove(projectionNumber);
}
// backproject for each slice
// CUDA Grids are only two dimensional!
int[] zed = new int[1];
int reconDimensionZ = dimz;
double voxelSpacingX = getGeometry().getVoxelSpacingX();
double voxelSpacingY = getGeometry().getVoxelSpacingY();
double voxelSpacingZ = getGeometry().getVoxelSpacingZ();
zed[0] = reconDimensionZ;
Pointer dOut = Pointer.to(volumePointer);
Pointer pWidth = Pointer.to(new int[] {(int) lineOffset});
Pointer pZOffset = Pointer.to(zed);
float[] vsx = new float[] {(float) voxelSpacingX};
Pointer pvsx = Pointer.to(vsx);
Pointer pvsy = Pointer.to(new float[] {(float) voxelSpacingY});
Pointer pvsz = Pointer.to(new float[] {(float) voxelSpacingZ});
Pointer pox = Pointer.to(new float[] {(float) offsetX});
Pointer poy = Pointer.to(new float[] {(float) offsetY});
Pointer poz = Pointer.to(new float[] {(float) offsetZ});
int offset = 0;
// System.out.println(dimz + " " + zed[0] + " " + offsetZ + " " + voxelSpacingZ);
offset = CUDAUtil.align(offset, Sizeof.POINTER);
JCudaDriver.cuParamSetv(function, offset, dOut, Sizeof.POINTER);
offset += Sizeof.POINTER;
offset = CUDAUtil.align(offset, Sizeof.INT);
JCudaDriver.cuParamSetv(function, offset, pWidth, Sizeof.INT);
offset += Sizeof.INT;
offset = CUDAUtil.align(offset, Sizeof.INT);
JCudaDriver.cuParamSetv(function, offset, pZOffset, Sizeof.INT);
offset += Sizeof.INT;
offset = CUDAUtil.align(offset, Sizeof.FLOAT);
JCudaDriver.cuParamSetv(function, offset, pvsx, Sizeof.FLOAT);
offset += Sizeof.FLOAT;
offset = CUDAUtil.align(offset, Sizeof.FLOAT);
JCudaDriver.cuParamSetv(function, offset, pvsy, Sizeof.FLOAT);
offset += Sizeof.FLOAT;
offset = CUDAUtil.align(offset, Sizeof.FLOAT);
JCudaDriver.cuParamSetv(function, offset, pvsz, Sizeof.FLOAT);
offset += Sizeof.FLOAT;
offset = CUDAUtil.align(offset, Sizeof.FLOAT);
JCudaDriver.cuParamSetv(function, offset, pox, Sizeof.FLOAT);
offset += Sizeof.FLOAT;
offset = CUDAUtil.align(offset, Sizeof.FLOAT);
JCudaDriver.cuParamSetv(function, offset, poy, Sizeof.FLOAT);
offset += Sizeof.FLOAT;
offset = CUDAUtil.align(offset, Sizeof.FLOAT);
JCudaDriver.cuParamSetv(function, offset, poz, Sizeof.FLOAT);
offset += Sizeof.FLOAT;
JCudaDriver.cuParamSetSize(function, offset);
// Call the CUDA kernel, writing the results into the volume which is pointed at
JCudaDriver.cuFuncSetBlockShape(function, bpBlockSize[0], bpBlockSize[1], 1);
JCudaDriver.cuLaunchGrid(function, gridSize.x, gridSize.y);
JCudaDriver.cuCtxSynchronize();
}
protected void init() {
if (!initialized) {
largeVolumeMode = false;
int reconDimensionX = getGeometry().getReconDimensionX();
int reconDimensionY = getGeometry().getReconDimensionY();
int reconDimensionZ = getGeometry().getReconDimensionZ();
projections = new ImageGridBuffer();
projectionsAvailable = new ArrayList<Integer>();
projectionsDone = new ArrayList<Integer>();
// Initialize the JCudaDriver. Note that this has to be done from
// the same thread that will later use the JCudaDriver API.
JCudaDriver.setExceptionsEnabled(true);
JCudaDriver.cuInit(0);
CUdevice dev = CUDAUtil.getBestDevice();
cuCtx = new CUcontext();
JCudaDriver.cuCtxCreate(cuCtx, 0, dev);
// check space on device:
int[] memory = new int[1];
int[] total = new int[1];
JCudaDriver.cuDeviceTotalMem(memory, dev);
JCudaDriver.cuMemGetInfo(memory, total);
int availableMemory = (int) (CUDAUtil.correctMemoryValue(memory[0]) / ((long) 1024 * 1024));
int requiredMemory =
(int)
(((((double) reconDimensionX)
* reconDimensionY
* ((double) reconDimensionZ)
* Sizeof.FLOAT)
+ (((double)
Configuration.getGlobalConfiguration()
.getGeometry()
.getDetectorHeight())
* Configuration.getGlobalConfiguration().getGeometry().getDetectorWidth()
* Sizeof.FLOAT))
/ (1024.0 * 1024));
if (debug) {
System.out.println("Total available Memory on CUDA card:" + availableMemory);
System.out.println("Required Memory on CUDA card:" + requiredMemory);
}
if (requiredMemory > availableMemory) {
nSteps = CUDAUtil.iDivUp(requiredMemory, (int) (availableMemory));
if (debug) System.out.println("Switching to large volume mode with nSteps = " + nSteps);
largeVolumeMode = true;
}
if (debug) {
CUdevprop prop = new CUdevprop();
JCudaDriver.cuDeviceGetProperties(prop, dev);
System.out.println(prop.toFormattedString());
}
// Load the CUBIN file containing the kernel
module = new CUmodule();
JCudaDriver.cuModuleLoad(module, "backprojectWithCuda.ptx");
// Obtain a function pointer to the kernel function. This function
// will later be called.
//
function = new CUfunction();
JCudaDriver.cuModuleGetFunction(function, module, "_Z17backprojectKernelPfiiffffff");
// create the reconstruction volume;
int memorysize = reconDimensionX * reconDimensionY * reconDimensionZ * Sizeof.FLOAT;
if (largeVolumeMode) {
subVolumeZ = CUDAUtil.iDivUp(reconDimensionZ, nSteps);
if (debug) System.out.println("SubVolumeZ: " + subVolumeZ);
h_volume = new float[reconDimensionX * reconDimensionY * subVolumeZ];
memorysize = reconDimensionX * reconDimensionY * subVolumeZ * Sizeof.FLOAT;
if (debug) System.out.println("Memory: " + memorysize);
} else {
h_volume = new float[reconDimensionX * reconDimensionY * reconDimensionZ];
}
// copy volume to device
volumePointer = new CUdeviceptr();
JCudaDriver.cuMemAlloc(volumePointer, memorysize);
JCudaDriver.cuMemcpyHtoD(volumePointer, Pointer.to(h_volume), memorysize);
// compute adapted volume size
// volume size in x = multiple of bpBlockSize[0]
// volume size in y = multiple of bpBlockSize[1]
int adaptedVolSize[] = new int[3];
if ((reconDimensionX % bpBlockSize[0]) == 0) {
adaptedVolSize[0] = reconDimensionX;
} else {
adaptedVolSize[0] = ((reconDimensionX / bpBlockSize[0]) + 1) * bpBlockSize[0];
}
if ((reconDimensionY % bpBlockSize[1]) == 0) {
adaptedVolSize[1] = reconDimensionY;
} else {
adaptedVolSize[1] = ((reconDimensionY / bpBlockSize[1]) + 1) * bpBlockSize[1];
}
adaptedVolSize[2] = reconDimensionZ;
int volStrideHost[] = new int[2];
// compute volstride and copy it to constant memory
volStrideHost[0] = adaptedVolSize[0];
volStrideHost[1] = adaptedVolSize[0] * adaptedVolSize[1];
volStride = new CUdeviceptr();
JCudaDriver.cuModuleGetGlobal(volStride, new int[1], module, "gVolStride");
JCudaDriver.cuMemcpyHtoD(volStride, Pointer.to(volStrideHost), Sizeof.INT * 2);
// Calculate new grid size
gridSize =
new dim3(
CUDAUtil.iDivUp(adaptedVolSize[0], bpBlockSize[0]),
CUDAUtil.iDivUp(adaptedVolSize[1], bpBlockSize[1]),
adaptedVolSize[2]);
// Obtain the global pointer to the view matrix from
// the module
projectionMatrix = new CUdeviceptr();
JCudaDriver.cuModuleGetGlobal(projectionMatrix, new int[1], module, "gProjMatrix");
initialized = true;
}
}
