static int _testTaskTimerCallStopTwice(void) {
TaskTimer t = newTaskTimer(1);
startTimingTask(t, 0);
_testSleep();
stopTiming(t);
stopTiming(t);
assertIntEquals(t->currentTask, -1);
// Recorded time should be at most 1ms off
assertDoubleEquals(t->totalTaskTimes[0], SLEEP_DURATION_MS, MAX_TIMER_TOLERANCE_MS);
freeTaskTimer(t);
return 0;
}
bool Sobel::runHalideGPU(Image input, Image output, const Params& params)
{
#if ENABLE_HALIDE
// Create halide buffers
buffer_t inputBuffer = createHalideBuffer(input);
buffer_t outputBuffer = createHalideBuffer(output);
reportStatus("Running Halide GPU filter");
// Warm-up run
inputBuffer.host_dirty = true;
halide_sobel_gpu(&inputBuffer, &outputBuffer);
halide_dev_sync(NULL);
// Timed runs
startTiming();
for (int i = 0; i < params.iterations; i++)
{
halide_sobel_gpu(&inputBuffer, &outputBuffer);
}
halide_dev_sync(NULL);
stopTiming();
halide_copy_to_host(NULL, &outputBuffer);
halide_release(NULL);
return outputResults(input, output, params);
#else
reportStatus("Halide not enabled during build.");
return false;
#endif
}
bool Bilateral::runHalideCPU(Image input, Image output, const Params& params)
{
#if ENABLE_HALIDE
// Create halide buffers
buffer_t inputBuffer = createHalideBuffer(input);
buffer_t outputBuffer = createHalideBuffer(output);
reportStatus("Running Halide CPU filter");
// Warm-up run
halide_bilateral_cpu(&inputBuffer, &outputBuffer);
// Timed runs
startTiming();
for (int i = 0; i < params.iterations; i++)
{
halide_bilateral_cpu(&inputBuffer, &outputBuffer);
}
stopTiming();
halide_release(NULL);
return outputResults(input, output, params);
#else
reportStatus("Halide not enabled during build.");
return false;
#endif
}
void minorCheneyCopyGC (GC_state s) {
size_t bytesAllocated;
size_t bytesCopied;
struct rusage ru_start;
if (DEBUG_GENERATIONAL)
fprintf (stderr, "minorGC nursery = "FMTPTR" frontier = "FMTPTR"\n",
(uintptr_t)s->heap.nursery, (uintptr_t)s->frontier);
assert (invariantForGC (s));
bytesAllocated = s->frontier - s->heap.nursery;
if (bytesAllocated == 0)
return;
s->cumulativeStatistics.bytesAllocated += bytesAllocated;
if (not s->canMinor) {
s->heap.oldGenSize += bytesAllocated;
bytesCopied = 0;
} else {
if (detailedGCTime (s))
startTiming (&ru_start);
s->cumulativeStatistics.numMinorGCs++;
s->forwardState.amInMinorGC = TRUE;
if (DEBUG_GENERATIONAL or s->controls.messages) {
fprintf (stderr,
"[GC: Starting minor Cheney-copy;]\n");
fprintf (stderr,
"[GC:\tfrom nursery at "FMTPTR" of size %s bytes.]\n",
(uintptr_t)(s->heap.nursery),
uintmaxToCommaString(bytesAllocated));
}
s->forwardState.toStart = s->heap.start + s->heap.oldGenSize;
assert (isFrontierAligned (s, s->forwardState.toStart));
s->forwardState.toLimit = s->forwardState.toStart + bytesAllocated;
assert (invariantForGC (s));
s->forwardState.back = s->forwardState.toStart;
/* Forward all globals. Would like to avoid doing this once all
* the globals have been assigned.
*/
foreachGlobalObjptr (s, forwardObjptrIfInNursery);
forwardInterGenerationalObjptrs (s);
foreachObjptrInRange (s, s->forwardState.toStart, &s->forwardState.back,
forwardObjptrIfInNursery, TRUE);
updateWeaksForCheneyCopy (s);
bytesCopied = s->forwardState.back - s->forwardState.toStart;
s->cumulativeStatistics.bytesCopiedMinor += bytesCopied;
s->heap.oldGenSize += bytesCopied;
s->lastMajorStatistics.numMinorGCs++;
if (detailedGCTime (s))
stopTiming (&ru_start, &s->cumulativeStatistics.ru_gcMinor);
if (DEBUG_GENERATIONAL or s->controls.messages)
fprintf (stderr,
"[GC: Finished minor Cheney-copy; copied %s bytes.]\n",
uintmaxToCommaString(bytesCopied));
}
}
void startTimingTask(TaskTimer taskTimer, const int taskId) {
if(taskId == taskTimer->currentTask) {
return;
}
stopTiming(taskTimer);
#if WINDOWS
QueryPerformanceCounter(&(taskTimer->startTime));
#elif UNIX
gettimeofday(taskTimer->startTime, NULL);
#endif
taskTimer->currentTask = taskId;
}
static int _testTaskTimerDuration(void) {
TaskTimer t = newTaskTimer(1);
assertIntEquals(t->currentTask, -1);
startTimingTask(t, 0);
_testSleep();
stopTiming(t);
assertIntEquals(t->currentTask, -1);
assertDoubleEquals(t->totalTaskTimes[0], SLEEP_DURATION_MS, MAX_TIMER_TOLERANCE_MS);
freeTaskTimer(t);
return 0;
}
void majorCheneyCopyGC (GC_state s) {
size_t bytesCopied;
struct rusage ru_start;
pointer toStart;
assert (s->secondaryHeap.size >= s->heap.oldGenSize);
if (detailedGCTime (s))
startTiming (&ru_start);
s->cumulativeStatistics.numCopyingGCs++;
s->forwardState.amInMinorGC = FALSE;
if (DEBUG or s->controls.messages) {
fprintf (stderr,
"[GC: Starting major Cheney-copy;]\n");
fprintf (stderr,
"[GC:\tfrom heap at "FMTPTR" of size %s bytes,]\n",
(uintptr_t)(s->heap.start),
uintmaxToCommaString(s->heap.size));
fprintf (stderr,
"[GC:\tto heap at "FMTPTR" of size %s bytes.]\n",
(uintptr_t)(s->secondaryHeap.start),
uintmaxToCommaString(s->secondaryHeap.size));
}
s->forwardState.toStart = s->secondaryHeap.start;
s->forwardState.toLimit = s->secondaryHeap.start + s->secondaryHeap.size;
assert (s->secondaryHeap.start != (pointer)NULL);
/* The next assert ensures there is enough space for the copy to
* succeed. It does not assert
* (s->secondaryHeap.size >= s->heap.size)
* because that is too strong.
*/
assert (s->secondaryHeap.size >= s->heap.oldGenSize);
toStart = alignFrontier (s, s->secondaryHeap.start);
s->forwardState.back = toStart;
foreachGlobalObjptr (s, forwardObjptr);
foreachObjptrInRange (s, toStart, &s->forwardState.back, forwardObjptr, TRUE);
updateWeaksForCheneyCopy (s);
s->secondaryHeap.oldGenSize = s->forwardState.back - s->secondaryHeap.start;
bytesCopied = s->secondaryHeap.oldGenSize;
s->cumulativeStatistics.bytesCopied += bytesCopied;
swapHeapsForCheneyCopy (s);
s->lastMajorStatistics.kind = GC_COPYING;
if (detailedGCTime (s))
stopTiming (&ru_start, &s->cumulativeStatistics.ru_gcCopying);
if (DEBUG or s->controls.messages)
fprintf (stderr,
"[GC: Finished major Cheney-copy; copied %s bytes.]\n",
uintmaxToCommaString(bytesCopied));
}
static int _testTaskTimerDurationMultipleTimes(void) {
TaskTimer t = newTaskTimer(1);
int i;
for(i = 0; i < 5; i++) {
assertIntEquals(t->currentTask, -1);
startTimingTask(t, 0);
_testSleep();
stopTiming(t);
assertIntEquals(t->currentTask, -1);
}
assertDoubleEquals(t->totalTaskTimes[0], 5.0 * SLEEP_DURATION_MS, MAX_TIMER_TOLERANCE_MS * 5.0);
freeTaskTimer(t);
return 0;
}
/*========================================================================
readYUV DEFINITION
======================================================================*/
void readYUV(int id, int xSize, int ySize, unsigned char *y, unsigned char *u, unsigned char *v) {
if( ftell(ptfile[id])/(xSize*ySize + xSize*ySize/2) >=NB_FRAME){
rewind(ptfile[id]);
}
if(id == 1 && ftell(ptfile[id])%(FPS*(xSize*ySize + xSize*ySize/2)) == 0){
unsigned int time = 0;
time = stopTiming(0);
printf("\nMain: %d frames in %d us - %f fps\n", FPS ,time, ((float)FPS)/(float)time*1000000);
startTiming(0);
}
fread(y, sizeof(char), xSize * ySize, ptfile[id]);
fread(u, sizeof(char), xSize * ySize / 4, ptfile[id]);
fread(v, sizeof(char), xSize * ySize / 4, ptfile[id]);
}
//单击 ”定时“ 处理
void ToolGlobal::pbn_clock_clicked()
{
if (! clockState) //如果 clockState 是 关闭 状态
{
pbn_clock_state->move(gbx_clock->x()+73, pbn_clock_state->y()); //设置 按钮位置
pbn_clock->move(gbx_clock->pos()); //...........
pbn_clock_state ->setText(tr("OFF")); //设置 按钮文字
pbn_clock ->setText(tr("关闭")); //...........
pbn_clock_state ->setToolTip(tr("关闭")); //设置 按钮提示
pbn_clock ->setToolTip(tr("关闭")); //...........
clockState = true; //设置为 “开启” 状态
// timing = new Timing(parentWidget()); //创建 “定时器设置” 窗口 并显示
timing = new Timing(gbx_clock->x()-42, 113, 225, 135, parentWidget()); //创建 “定时器设置” 窗口 并显示
timing ->show(); //.........................
connect(timing, SIGNAL(timing_pbnOk_click()), this, SLOT(startTiming())); //关联 定时器 “确定” 按钮
connect(timing, SIGNAL(timing_pbnCancle_click()), this, SLOT(stopTiming())); //关联 定时器 “取消” 按钮
}
else //如果 clockState 是 开启 状态
{
if (timing) //如果存在 “定时器设置” 窗口
{
timing ->close(); //关闭 “定时器设置” 窗口
}
if (lcdNumber) //如果存在 “LCD显示器”
{
delete lcdNumber; //销毁 “LCD显示器”
lcdNumber = 0; //指针赋0值
timer_default ->stop(); //终止 “1秒倒计时"
timer_target ->stop(); //终止 "自定义倒计时”
}
pbn_clock_state->move(gbx_clock->x()+13, pbn_clock_state->y()); //设置 按钮位置
pbn_clock->move(gbx_clock->x()+46, pbn_clock->y()); //...........
pbn_clock_state ->setText(tr("ON")); //设置 按钮文字
pbn_clock ->setText(tr("定时")); //...........
pbn_clock_state ->setToolTip(tr("开启魔音")); //设置 按钮提示
pbn_clock ->setToolTip(tr("开启魔音")); //...........
clockState = false; //设置为 “关闭” 状态
}
}
bool TreeModel::dropMimeData(const QMimeData *mimeData,
Qt::DropAction action, int row, int column,
const QModelIndex &parent)
{
if (action == Qt::IgnoreAction)
return true;
if (action != Qt::MoveAction || column > 0 ||
!mimeData || !mimeData->hasFormat(MimeType))
return false;
if (TaskItem *item = itemForIndex(parent)) {
emit stopTiming();
QByteArray xmlData = qUncompress(mimeData->data(MimeType));
QXmlStreamReader reader(xmlData);
if (row == -1)
row = parent.isValid() ? parent.row()
: rootItem->childCount();
beginInsertRows(parent, row, row);
readTasks(&reader, item);
endInsertRows();
return true;
}
return false;
}
int main (int argc, char **argv) {
int N;
double * values;
char type = argc > 2 ? argv[2][0] : 'A';
if (argc < 2) {
fprintf (stderr, "Usage: csort SIZE [ TYPE [SEED]] \n");
exit (1);
}
N = atoi (argv[1]);
srand (argc > 3 ? atol (argv[3]) : 42);
if (type == 'A')
values = randomVector(N, 0.0, 1.0);
else
values = randomVector(N, 1.0, 20.0);
startTiming();
sort (values, N);
stopTiming();
if (N < 20)
printArr(values, N);
return 0;
}
// ----------------------------------------------------------------------------
// gpuNUFFT_gpu: NUFFT
//
// Inverse gpuNUFFT implementation - interpolation from uniform grid data onto
// nonuniform k-space data based on optimized
// gpuNUFFT kernel with minimal oversampling
// ratio (see Beatty et al.)
//
// Basic steps: - apodization correction
// - zero padding with osf
// - FFT
// - convolution and resampling
//
// parameters:
// * data : output kspace data
// * data_count : number of samples on trajectory
// * n_coils : number of channels or coils
// * crds : coordinates on trajectory, passed as SoA
// * imdata : input image data
// * imdata_count : number of image data points
// * grid_width : size of grid
// * kernel : precomputed convolution kernel as lookup table
// * kernel_count : number of kernel lookup table entries
// * sectors : mapping of data indices according to each sector
// * sector_count : number of sectors
// * sector_centers: coordinates (x,y,z) of sector centers
// * sector_width : width of sector
// * im_width : dimension of image
// * osr : oversampling ratio
// * gpuNUFFT_out : enum indicating how far gpuNUFFT has to be processed
//
void gpuNUFFT::GpuNUFFTOperator::performForwardGpuNUFFT(gpuNUFFT::Array<DType2> imgData,gpuNUFFT::Array<CufftType>& kspaceData, GpuNUFFTOutput gpuNUFFTOut)
{
if (DEBUG)
{
std::cout << "performing forward gpuNUFFT!!!" << std::endl;
std::cout << "dataCount: " << kspaceData.count() << " chnCount: " << kspaceData.dim.channels << std::endl;
std::cout << "imgCount: " << imgData.count() << " gridWidth: " << this->getGridWidth() << std::endl;
}
showMemoryInfo();
if (debugTiming)
startTiming();
int data_count = (int)this->kSpaceTraj.count();
int n_coils = (int)kspaceData.dim.channels;
IndType imdata_count = this->imgDims.count();
int sector_count = (int)this->gridSectorDims.count();
//cuda mem allocation
DType2 *imdata_d;
CufftType *data_d;
if (DEBUG)
printf("allocate and copy imdata of size %d...\n",imdata_count);
allocateDeviceMem<DType2>(&imdata_d,imdata_count);
if (DEBUG)
printf("allocate and copy data of size %d...\n",data_count);
allocateDeviceMem<CufftType>(&data_d,data_count);
initDeviceMemory(n_coils);
if (debugTiming)
printf("Memory allocation: %.2f ms\n",stopTiming());
int err;
//iterate over coils and compute result
for (int coil_it = 0; coil_it < n_coils; coil_it++)
{
int data_coil_offset = coil_it * data_count;
int im_coil_offset = coil_it * (int)imdata_count;
if (this->applySensData())
// perform automatically "repeating" of input image in case
// of existing sensitivity data
copyToDevice(imgData.data,imdata_d,imdata_count);
else
copyToDevice(imgData.data + im_coil_offset,imdata_d,imdata_count);
//reset temp arrays
cudaMemset(gdata_d,0, sizeof(CufftType)*gi_host->grid_width_dim);
cudaMemset(data_d,0, sizeof(CufftType)*data_count);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at thread synchronization 1: %s\n",cudaGetErrorString(cudaGetLastError()));
if (this->applySensData())
{
copyToDevice(this->sens.data + im_coil_offset, sens_d,imdata_count);
performSensMul(imdata_d,sens_d,gi_host,false);
}
// apodization Correction
if (n_coils > 1 && deapo_d != NULL)
performForwardDeapodization(imdata_d,deapo_d,gi_host);
else
performForwardDeapodization(imdata_d,gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at thread synchronization 2: %s\n",cudaGetErrorString(cudaGetLastError()));
// resize by oversampling factor and zero pad
performPadding(imdata_d,gdata_d,gi_host);
if (debugTiming)
startTiming();
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at thread synchronization 3: %s\n",cudaGetErrorString(cudaGetLastError()));
// shift image to get correct zero frequency position
performFFTShift(gdata_d,INVERSE,getGridDims(),gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at thread synchronization 4: %s\n",cudaGetErrorString(cudaGetLastError()));
// eventually free imdata_d
// Forward FFT to kspace domain
if (err=pt2CufftExec(fft_plan, gdata_d, gdata_d, CUFFT_FORWARD) != CUFFT_SUCCESS)
{
fprintf(stderr,"cufft has failed with err %i \n",err);
showMemoryInfo(true,stderr);
}
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at thread synchronization 5: %s\n",cudaGetErrorString(cudaGetLastError()));
performFFTShift(gdata_d,FORWARD,getGridDims(),gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at thread synchronization 6: %s\n",cudaGetErrorString(cudaGetLastError()));
if (debugTiming)
printf("FFT (incl. shift): %.2f ms\n",stopTiming());
if (debugTiming)
startTiming();
// convolution and resampling to non-standard trajectory
forwardConvolution(data_d,crds_d,gdata_d,NULL,sectors_d,sector_centers_d,gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at thread synchronization 7: %s\n",cudaGetErrorString(cudaGetLastError()));
if (debugTiming)
printf("Forward Convolution: %.2f ms\n",stopTiming());
performFFTScaling(data_d,gi_host->data_count,gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error: at thread synchronization 8: %s\n",cudaGetErrorString(cudaGetLastError()));
//write result in correct order back into output array
writeOrderedGPU(data_sorted_d,data_indices_d,data_d,(int)this->kSpaceTraj.count());
copyFromDevice(data_sorted_d,kspaceData.data + data_coil_offset,data_count);
}//iterate over coils
freeTotalDeviceMemory(data_d,imdata_d,NULL);
freeDeviceMemory(n_coils);
if ((cudaThreadSynchronize() != cudaSuccess))
fprintf(stderr,"error in performForwardGpuNUFFT function: %s\n",cudaGetErrorString(cudaGetLastError()));
free(gi_host);
}
// ----------------------------------------------------------------------------
// performGpuNUFFTAdj: NUFFT^H
//
// GpuNUFFT implementation - interpolation from nonuniform k-space data onto
// oversampled grid based on optimized gpuNUFFT kernel
// with minimal oversampling ratio (see Beatty et al.)
//
// Basic steps: - density compensation
// - convolution with interpolation function
// - iFFT
// - cropping due to oversampling ratio
// - apodization correction
//
// parameters:
// * data : input kspace data
// * data_count : number of samples on trajectory
// * n_coils : number of channels or coils
// * crds : coordinate array on trajectory
// * imdata : output image data
// * imdata_count : number of image data points
// * grid_width : size of grid
// * kernel : precomputed convolution kernel as lookup table
// * kernel_count : number of kernel lookup table entries
// * sectors : mapping of start and end points of each sector
// * sector_count : number of sectors
// * sector_centers: coordinates (x,y,z) of sector centers
// * sector_width : width of sector
// * im_width : dimension of image
// * osr : oversampling ratio
// * do_comp : true, if density compensation has to be done
// * density_comp : densiy compensation array
// * gpuNUFFT_out : enum indicating how far gpuNUFFT has to be processed
//
void gpuNUFFT::GpuNUFFTOperator::performGpuNUFFTAdj(gpuNUFFT::Array<DType2> kspaceData, gpuNUFFT::Array<CufftType>& imgData, GpuNUFFTOutput gpuNUFFTOut)
{
if (DEBUG)
{
std::cout << "performing gpuNUFFT adjoint!!!" << std::endl;
std::cout << "dataCount: " << kspaceData.count() << " chnCount: " << kspaceData.dim.channels << std::endl;
std::cout << "imgCount: " << imgData.count() << " gridWidth: " << this->getGridWidth() << std::endl;
std::cout << "apply density comp: " << this->applyDensComp() << std::endl;
std::cout << "apply sens data: " << this->applySensData() << std::endl;
}
if (debugTiming)
startTiming();
showMemoryInfo();
int data_count = (int)this->kSpaceTraj.count();
int n_coils = (int)kspaceData.dim.channels;
IndType imdata_count = this->imgDims.count();
int sector_count = (int)this->gridSectorDims.count();
// select data ordered and leave it on gpu
DType2* data_d;
if (DEBUG)
printf("allocate data of size %d...\n",data_count);
allocateDeviceMem<DType2>(&data_d,data_count);
CufftType *imdata_d, *imdata_sum_d = NULL;
if (DEBUG)
printf("allocate and copy imdata of size %d...\n",imdata_count);
allocateDeviceMem<CufftType>(&imdata_d,imdata_count);
if (this->applySensData())
{
if (DEBUG)
printf("allocate and copy temp imdata of size %d...\n",imdata_count);
allocateDeviceMem<CufftType>(&imdata_sum_d,imdata_count);
cudaMemset(imdata_sum_d,0,imdata_count*sizeof(CufftType));
}
initDeviceMemory(n_coils);
int err;
if (debugTiming)
printf("Memory allocation: %.2f ms\n",stopTiming());
//iterate over coils and compute result
for (int coil_it = 0; coil_it < n_coils; coil_it++)
{
int data_coil_offset = coil_it * data_count;
int im_coil_offset = coil_it * (int)imdata_count;//gi_host->width_dim;
cudaMemset(gdata_d,0, sizeof(CufftType)*gi_host->grid_width_dim);
//copy coil data to device and select ordered
copyToDevice(kspaceData.data + data_coil_offset,data_d,data_count);
selectOrderedGPU(data_d,data_indices_d,data_sorted_d,data_count);
if (this->applyDensComp())
performDensityCompensation(data_sorted_d,density_comp_d,gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at adj thread synchronization 1: %s\n",cudaGetErrorString(cudaGetLastError()));
if (debugTiming)
startTiming();
adjConvolution(data_sorted_d,crds_d,gdata_d,NULL,sectors_d,sector_centers_d,gi_host);
if (debugTiming)
printf("Adjoint convolution: %.2f ms\n",stopTiming());
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
fprintf(stderr,"error at adj thread synchronization 2: %s\n",cudaGetErrorString(cudaGetLastError()));
if (gpuNUFFTOut == CONVOLUTION)
{
if (DEBUG)
printf("stopping output after CONVOLUTION step\n");
//get output
copyFromDevice<CufftType>(gdata_d,imgData.data,gi_host->grid_width_dim);
if (DEBUG)
printf("test value at point zero: %f\n",(imgData.data)[0].x);
free(gi_host);
freeTotalDeviceMemory(data_d,imdata_d,imdata_sum_d,NULL);
freeDeviceMemory(n_coils);
return;
}
if ((cudaThreadSynchronize() != cudaSuccess))
fprintf(stderr,"error at adj thread synchronization 3: %s\n",cudaGetErrorString(cudaGetLastError()));
if (debugTiming)
startTiming();
performFFTShift(gdata_d,INVERSE,getGridDims(),gi_host);
//Inverse FFT
if (err=pt2CufftExec(fft_plan, gdata_d, gdata_d, CUFFT_INVERSE) != CUFFT_SUCCESS)
{
fprintf(stderr,"cufft has failed at adj with err %i \n",err);
showMemoryInfo(true,stderr);
}
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
fprintf(stderr,"error at adj thread synchronization 4: %s\n",cudaGetErrorString(cudaGetLastError()));
if (gpuNUFFTOut == FFT)
{
if (DEBUG)
printf("stopping output after FFT step\n");
//get output
copyFromDevice<CufftType>(gdata_d,imgData.data,gi_host->grid_width_dim);
free(gi_host);
freeTotalDeviceMemory(data_d,imdata_d,imdata_sum_d,NULL);
freeDeviceMemory(n_coils);
printf("last cuda error: %s\n", cudaGetErrorString(cudaGetLastError()));
return;
}
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at adj thread synchronization 5: %s\n",cudaGetErrorString(cudaGetLastError()));
performFFTShift(gdata_d,INVERSE,getGridDims(),gi_host);
if (debugTiming)
printf("iFFT (incl. shift) : %.2f ms\n",stopTiming());
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at adj thread synchronization 6: %s\n",cudaGetErrorString(cudaGetLastError()));
performCrop(gdata_d,imdata_d,gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at adj thread synchronization 7: %s\n",cudaGetErrorString(cudaGetLastError()));
//check if precomputed deapo function can be used
if (n_coils > 1 && deapo_d != NULL)
performDeapodization(imdata_d,deapo_d,gi_host);
else
performDeapodization(imdata_d,gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error at adj thread synchronization 8: %s\n",cudaGetErrorString(cudaGetLastError()));
performFFTScaling(imdata_d,gi_host->im_width_dim,gi_host);
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error: at adj thread synchronization 9: %s\n",cudaGetErrorString(cudaGetLastError()));
if (this->applySensData())
{
copyToDevice(this->sens.data + im_coil_offset, sens_d,imdata_count);
performSensMul(imdata_d,sens_d,gi_host,true);
performSensSum(imdata_d,imdata_sum_d,gi_host);
}
else
{
// get result per coil
// no summation is performed in absence of sensitity data
copyFromDevice<CufftType>(imdata_d,imgData.data+im_coil_offset,imdata_count);
}
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error: at adj thread synchronization 10: %s\n",cudaGetErrorString(cudaGetLastError()));
}//iterate over coils
if (this->applySensData())
{
// get result of combined coils
copyFromDevice<CufftType>(imdata_sum_d,imgData.data,imdata_count);
}
if (DEBUG && (cudaThreadSynchronize() != cudaSuccess))
printf("error: at adj thread synchronization 11: %s\n",cudaGetErrorString(cudaGetLastError()));
freeTotalDeviceMemory(data_d,imdata_d,imdata_sum_d,NULL);
freeDeviceMemory(n_coils);
if ((cudaThreadSynchronize() != cudaSuccess))
fprintf(stderr,"error in gpuNUFFT_gpu_adj function: %s\n",cudaGetErrorString(cudaGetLastError()));
free(gi_host);
}
void performGC (GC_state s,
size_t oldGenBytesRequested,
size_t nurseryBytesRequested,
bool forceMajor,
bool mayResize) {
uintmax_t gcTime;
bool stackTopOk;
size_t stackBytesRequested;
struct rusage ru_start;
size_t totalBytesRequested;
enterGC (s);
s->cumulativeStatistics.numGCs++;
if (DEBUG or s->controls.messages) {
size_t nurserySize = s->heap.size - ((size_t)(s->heap.nursery - s->heap.start));
size_t nurseryUsed = (size_t)(s->frontier - s->heap.nursery);
fprintf (stderr,
"[GC: Starting gc #%s; requesting %s nursery bytes and %s old-gen bytes,]\n",
uintmaxToCommaString(s->cumulativeStatistics.numGCs),
uintmaxToCommaString(nurseryBytesRequested),
uintmaxToCommaString(oldGenBytesRequested));
fprintf (stderr,
"[GC:\theap at "FMTPTR" of size %s bytes (+ %s bytes card/cross map),]\n",
(uintptr_t)(s->heap.start),
uintmaxToCommaString(s->heap.size),
uintmaxToCommaString(s->heap.withMapsSize - s->heap.size));
fprintf (stderr,
"[GC:\twith old-gen of size %s bytes (%.1f%% of heap),]\n",
uintmaxToCommaString(s->heap.oldGenSize),
100.0 * ((double)(s->heap.oldGenSize) / (double)(s->heap.size)));
fprintf (stderr,
"[GC:\tand nursery of size %s bytes (%.1f%% of heap),]\n",
uintmaxToCommaString(nurserySize),
100.0 * ((double)(nurserySize) / (double)(s->heap.size)));
fprintf (stderr,
"[GC:\tand nursery using %s bytes (%.1f%% of heap, %.1f%% of nursery).]\n",
uintmaxToCommaString(nurseryUsed),
100.0 * ((double)(nurseryUsed) / (double)(s->heap.size)),
100.0 * ((double)(nurseryUsed) / (double)(nurserySize)));
}
assert (invariantForGC (s));
if (needGCTime (s))
startTiming (&ru_start);
minorGC (s);
stackTopOk = invariantForMutatorStack (s);
stackBytesRequested =
stackTopOk
? 0
: sizeofStackWithHeader (s, sizeofStackGrowReserved (s, getStackCurrent (s)));
totalBytesRequested =
oldGenBytesRequested
+ nurseryBytesRequested
+ stackBytesRequested;
if (forceMajor
or totalBytesRequested > s->heap.size - s->heap.oldGenSize)
majorGC (s, totalBytesRequested, mayResize);
setGCStateCurrentHeap (s, oldGenBytesRequested + stackBytesRequested,
nurseryBytesRequested);
assert (hasHeapBytesFree (s, oldGenBytesRequested + stackBytesRequested,
nurseryBytesRequested));
unless (stackTopOk)
growStackCurrent (s);
setGCStateCurrentThreadAndStack (s);
if (needGCTime (s)) {
gcTime = stopTiming (&ru_start, &s->cumulativeStatistics.ru_gc);
s->cumulativeStatistics.maxPauseTime =
max (s->cumulativeStatistics.maxPauseTime, gcTime);
} else
gcTime = 0; /* Assign gcTime to quell gcc warning. */
if (DEBUG or s->controls.messages) {
size_t nurserySize = s->heap.size - (size_t)(s->heap.nursery - s->heap.start);
fprintf (stderr,
"[GC: Finished gc #%s; time %s ms,]\n",
uintmaxToCommaString(s->cumulativeStatistics.numGCs),
uintmaxToCommaString(gcTime));
fprintf (stderr,
"[GC:\theap at "FMTPTR" of size %s bytes (+ %s bytes card/cross map),]\n",
(uintptr_t)(s->heap.start),
uintmaxToCommaString(s->heap.size),
uintmaxToCommaString(s->heap.withMapsSize - s->heap.size));
fprintf (stderr,
"[GC:\twith old-gen of size %s bytes (%.1f%% of heap),]\n",
uintmaxToCommaString(s->heap.oldGenSize),
100.0 * ((double)(s->heap.oldGenSize) / (double)(s->heap.size)));
fprintf (stderr,
"[GC:\tand nursery of size %s bytes (%.1f%% of heap).]\n",
uintmaxToCommaString(nurserySize),
100.0 * ((double)(nurserySize) / (double)(s->heap.size)));
}
/* Send a GC signal. */
if (s->signalsInfo.gcSignalHandled
and s->signalHandlerThread != BOGUS_OBJPTR) {
if (DEBUG_SIGNALS)
fprintf (stderr, "GC Signal pending.\n");
s->signalsInfo.gcSignalPending = TRUE;
unless (s->signalsInfo.amInSignalHandler)
s->signalsInfo.signalIsPending = TRUE;
}
if (DEBUG)
displayGCState (s, stderr);
assert (hasHeapBytesFree (s, oldGenBytesRequested, nurseryBytesRequested));
assert (invariantForGC (s));
leaveGC (s);
}
