c++ - CUDA invalid argument when trying to copy struct to device's memory (cudaMemcpy) -


i'm trying figure out how should create struct/class destined sent device, keep getting "invalid argument" cuda error. did small example shows error:

#include <iostream> #include <cstdio> using namespace std;  #define cuda_warn(xxx) \     { if (xxx != cudasuccess) cerr << "cuda error: " << \         cudageterrorstring(xxx) << ", @ line " << __line__ \         << endl; cudadevicesynchronize(); } while (0)  struct p {     double x,y;     __host__ __device__ void init(const double &a, const double &b) {         x = a; y = b; } };  int main(int argc, char **argv) {     p hp, hq, dp;     cout << "size of p: " << sizeof(p) << endl;     cuda_warn(cudamalloc((void**) &dp, sizeof(p)));     printf("dp: %p\n", &dp); // print dp's address on device     hp.init(1.2,-2.1);     hq.init(0.,0.);     cuda_warn(cudamemcpy(&dp, &hp, sizeof(p), cudamemcpyhosttodevice));     cuda_warn(cudamemcpy(&hq, &dp, sizeof(p), cudamemcpydevicetohost));     cout << "copy back: " << hq.x << "\t" << hq.y << endl;     dp.init(3.,3.);     cuda_warn(cudamemcpy(&hp, &dp, sizeof(p), cudamemcpydevicetohost));     cout << "copy new:  " << hp.x << "\t" << hp.y << endl;     return 0; }

i'm compiling (my card tesla c2050):

nvcc -arch sm_20 -o exec file.cu

the result i'm getting is:

size of p: 16 dp: 0x7fff82d4b7b0 cuda error: invalid argument, @ line 24 cuda error: invalid argument, @ line 25 copy back: 0    0 cuda error: invalid argument, @ line 28 copy new:  1.2  -2.1   ------------------ (program exited code: 0) press return continue

thanks guys if me on this!

====== after comments of @talonmies, @jackolantern, @robert crovella =======

thanks, guys! helped! based on comments, correct code , working. register final solution:

#include <iostream> #include <cstdio> using namespace std;  #define cuda_warn(xxx) \     { if (xxx != cudasuccess) cerr << "cuda error: " << \         cudageterrorstring(xxx) << ", @ line " << __line__ \         << endl; cudadevicesynchronize(); } while (0)  struct p {     double x,y;     __host__ __device__ void init(const double &a, const double &b) {         x = a; y = b; } };  /* included kernel function */ __global__ void dev_p_init(p *p, double a, double b) {     p->init(a,b); }  int main(int argc, char **argv) {     p hp, hq, *dp; //*changed*     cout << "size of p: " << sizeof(p) << endl;     cuda_warn(cudamalloc((void**) &dp, sizeof(p)));     printf("dp: %p\n", &dp); // print dp's address on device     hp.init(1.2,-2.1);     hq.init(0.,0.);     cuda_warn(cudamemcpy(dp, &hp, sizeof(p), cudamemcpyhosttodevice)); //*changed*     cuda_warn(cudamemcpy(&hq, dp, sizeof(p), cudamemcpydevicetohost)); //*changed*     cout << "copy back: " << hq.x << "\t" << hq.y << endl;     dev_p_init<<< 1, 1 >>>(dp,3., 3.); //*call kernel*     cuda_warn(cudamemcpy(&hp, dp, sizeof(p), cudamemcpydevicetohost)); //*changed*     cout << "copy new:  " << hp.x << "\t" << hp.y << endl;     return 0; }

and corrected output:

size of p: 16 dp: 0x7fff6fa2e498 copy back: 1.2  -2.1 copy new:  3    3   ------------------ (program exited code: 0) press return continue

as noticed @talonmies, &dp not valid device pointer. indeed, dp variable resides on host, address points host memory space. opposite that, when dp pointer, cudamalloc receive value parameter , value point device memory space.

this correct version of code:

#include <iostream> #include <cstdio> using namespace std;  #define cuda_warn(xxx) \     { if (xxx != cudasuccess) cerr << "cuda error: " << \     cudageterrorstring(xxx) << ", @ line " << __line__ \     << endl; cudadevicesynchronize(); } while (0)  struct p {     double x,y;     __host__ __device__ void init(const double &a, const double &b) {     x = a; y = b; } };  int main(int argc, char **argv) {     p *dp;     p hp, hq;     cuda_warn(cudamalloc((void**) &dp, sizeof(p)));     cuda_warn(cudamemcpy(dp, &hp, sizeof(p), cudamemcpyhosttodevice));     cuda_warn(cudamemcpy(&hq, dp, sizeof(p), cudamemcpydevicetohost));     cuda_warn(cudamemcpy(&hp, dp, sizeof(p), cudamemcpydevicetohost));      return 0; }

Comments

Post a Comment

Popular posts from this blog

javascript - RequestAnimationFrame not working when exiting fullscreen switching space on Safari -

request animation frame stops working when exiting fullscreen clicking on original space's safari window. fine if fullscreen mode canceled escape key or calling cancelfullscreen(). steps reproduce: open https://dl.dropboxusercontent.com/u/769042/prezi/safari-fullscreen.html click "draw raf", kittie appears click "fullscreen", go fullscreen click "draw raf", kittie appears go space original safari was, showing "click exit fullscreen mode", click anywhere, out fullscreen click "draw raf", nothing happens what handling click calling: window.requestanimationframe(draw); which draws on canvas context: function draw() { ctx.drawimage(img, math.random()*500|0, math.random()*400|0, 100, 100); } i checked .hidden , .visibilitystate, updated correctly. tested on osx 10.9.3, safari 7.0.4 (9537.76.4). has workaround/solution other switching old settimeout? this sounds webkit bug 88940 : using reque...
Read more

jsf - How to ajax update an item in the footer of a PrimeFaces dataTable? -

this visual of table have: +---------+------------+-------------------+ | header1 | header2 | header3 | +---------+------------+-------------------+ | row | input | calc'ed output | | row b | input b | calc'ed output b | | etc.. | etc.. | etc.. | +---------+------------+-------------------+ | total: total calc'ed output | +------------------------------------------+ and stripped-down code: <p:datatable id="mytable" value="#{mybean.somelist}" var="currentitem"> <p:column headertext="header1"> <h:outputtext value="#{currentitem.name}" /> </p:column> <p:column headertext="header2"> <pe:inputnumber value="#{currentitem.inputval}"> <p:ajax event="change" listener="#{mybean.changelistener}" update="outputval outputtotal...
Read more

django - CSRF verification failed. Request aborted. CSRF cookie not set -

error: reason given failure: csrf cookie not set. in general, can occur when there genuine cross site request forgery, or when django's csrf mechanism has not been used correctly. post forms, need ensure: browser accepting cookies. view function uses requestcontext template, instead of context. in template, there {% csrf_token %} template tag inside each post form targets internal url. if not using csrfviewmiddleware, must use csrf_protect on views use csrf_token template tag, accept post data. you're seeing section of page because have debug = true in django settings file. change false, , initial error message displayed. can customize page using csrf_failure_view setting. did find, nothing helped, must foolish(( my template <form method="post">{% csrf_token %} {% field in form %} <div class="control-group"> {{ field }} </div> {% endfor %} <input type=...
Read more