c++ - Improving asynchronous execution in CUDA -

-

i writing programme performs large simulations on gpu using cuda api. in order accelerate performance, tried run kernels simultaneously , asynchronously copy result host memory again. code looks this:

#define nstreams   8 #define blockdimx  16 #define blockdimy  16  void domainupdate(float* domain_cpu,       // pointer domain on host                   float* domain_gpu,       // pointer domain on device                   const unsigned int dimx,                   const unsigned int dimy,                   const unsigned int dimz) {     dim3 blocks((dimx + blockdimx - 1) / blockdimx, (dimy + blockdimy - 1) / blockdimy);     dim3 threads(blockdimx, blockdimy);      (unsigned int ii = 0; ii < nstreams; ++ii) {          updatedomain3d<<<blocks,threads, 0, streams[ii]>>>(domain_gpu,                                                            dimx, 0,  dimx - 1, // dimx, minx, maxx                                                            dimy, 0,  dimy - 1, // dimy, miny, maxy                                                            dimz, dimz * ii / nstreams,  dimz * (ii + 1) / nstreams - 1); // dimz, minz, maxz          unsigned int offset = dimx * dimy * dimz * ii / nstreams;         cudamemcpyasync(domain_cpu + offset ,                         domain_gpu+ offset ,                         sizeof(float) * dimx * dimy * dimz / nstreams,                         cudamemcpydevicetohost, streams[ii]);     }      cudadevicesynchronize(); }

all in simple for-loop, looping on streams (8 in case) , dividing work. deal faster (up 30% performance gain), although maybe less had hoped. analysed typical cycle in nvidia's compute visual profiler, , execution looks this:

cuda api trace in compute visual profiler

as can seen in picture, kernels overlap, although never more 2 kernels running @ same time. tried same thing different numbers of streams , different sizes of simulation domain, case.

so question is: there way encourage/force gpu scheduler run more 2 things @ same time? or limitation dependent on gpu device cannot represented in code?

my system specifications are: 64-bit windows 7, , geforce gtx 670 graphics card (that's kepler architecture, compute capability 3.0).

kernels overlap if gpu has resources left run second kernel. once gpu loaded, there no gain running more kernels in parallel, driver not that.


Popular posts from this blog

How to calculate SNR of signals in MATLAB? -

-
i've encountered simple, yet fundamental problem on calculating snr: there several signals : s1 = original , pure clean signal without noise. n1 = white gaussian noise going added s1. s2 = s1 + n1 (the noisy signal before performing noise-reduction algorithm) s3 = noise-reduced signal (after performing noise reduction algorithm) n2 = s3 - s1 (the amount of noise after performing noise-reduction algorithm) now want compare snr before , after performing noise-reduction algorithm. which signals should consider snr_before!? s1/n1 or s2/n1 ? which signals should consider snr_after!? s3/n1 or s3/n2 ? which commands or functions should use in matlab in order compute snr_before , snr_after? thanks billion putting time on helping me. what know calculating snr before : snrbeforenoise = mean( signal .^ 2 ) / mean( noise .^ 2 ); and snr after : residual_noise = signal - noise_reduced_signal; snr_after = mean( signal .^ 2 ) / mean( residual_noise .^ 2...
Read more

c# - Attempting to upload to FTP: System.Net.WebException: System error -

-
i have api takes in xml , uploads files based on information in xml. uploads on schedule (also xml), , have tested surrounding , know works. i getting error 40% of time on first file attempt upload in each time cycle (time cycle = 45 minutes files, 30 minutes others). here code upload: try { loggerftp.log("uploading file: " + filename, false); // create request. ftpwebrequest request = (ftpwebrequest)webrequest.create(appsettingsftp.ftpurl + @"/" + filename); request.method = webrequestmethods.ftp.uploadfile; request.timeout = 6000000; //set 100 minutes //request.timeout = -1; //set infinite // add login credentials. request.credentials = new networkcredential(appsettingsftp.ftplogin, appsettingsftp.ftppassword); // grab file contents. streamreader sourcestream = new streamreader(appsettingsftp.uploadfiledirectory + filename); byte[] filecontents = encoding.utf8.getbytes(sourcestream.readtoend()); sourcest...
Read more

ios - UISlider customization: how to properly add shadow to custom knob image -

-
Image
i need place custom knob image shadow uislider. far have image of 38x38, of there region defines shadow of knob. now while using image, when slide knob extreme ends actual knob not able cover end points, hence end points visible under transparent region of knob image this code using customize uislider uiimage *slidebtimg = [uiimage imagenamed:sliderknobimg]; [self.ratingslider setthumbimage:slidebtimg forstate:uicontrolstatenormal]; [self.ratingslider setthumbimage:slidebtimg forstate:uicontrolstatehighlighted]; how can implement custom knob shadow beneath it, covering end points?? look inside uislider.h , there methods let customize slider knob , track needed. (ios 2.0 , above) // lets subclass lay out track , thumb needed - (cgrect)minimumvalueimagerectforbounds:(cgrect)bounds; - (cgrect)maximumvalueimagerectforbounds:(cgrect)bounds; - (cgrect)trackrectforbounds:(cgrect)bounds; - (cgrect)thumbrectforbounds:(cgrect)bounds trackrect:(cgrect)rect value:...
Read more