KERNELS are NOT queing , bug in cuda 2.0 ? cudaThreadSynchronize(); makes no difference ?

Accelerated Computing CUDA CUDA Programming and Performance
1

[b]the thread is being discussed here:LINK

Sorry for multiple posts, my bad[/b]

HI

I have a strange problem… I have the following kernels calls, which I think should be qued in ~ 3 microseconds.

gpu_ford_phia<<<dimGrid,dimBlock>>>((vec_space*)adds[0],(BIGspace*)adds[3],(Bspace*)adds[1]);

	 for(int i=0;i<12;i++)

	 {		   

	 gpu_mtranT_prod<<<dimGrid5,dimBlock5>>>((Bspace*)adds[2],(BIGspace*)adds[3],i);

	 gpu_T_prodmtran<<<dimGrid6,dimBlock6>>>((Bspace*)adds[2],(BIGspace*)adds[3],i);

	  gpu_mT_prod<<<dimGrid4,dimBlock4>>>((Bspace*)adds[2],(BIGspace*)adds[3],i);   

	 gpu_ford_phic<<<dimGrid3,dimBlock3>>>((vec_space*)adds[0],(BIGspace*)adds[3],(Bspace*)adds[2],(Bspace*)adds[1],i);

	 }

AND they are qued in 4e-3 seconds in cuda and nvcc version 2.0 which is fine.

Now if I add this kernel call

gpu_R<<<dimGridR,dimBlockR>>>((Bspace*)adds[2], upd*(N/Block_sizeR), (sol_space*)adds[4] ); // the GPU_R kernel does reduction sort of operations on matrices.

after the for loop of previous kernels the queuing takes 0.1 seconds ??

Hence, I am loosing all the speed up here as I want to do some CPU calculations before GPU finishes.

I then inserted cudaThreadSynchronize(); and saw that it takes the same time 0.1 seconds. External Image

hence the program actually waits for the last kernel to finsih before transferring the control to the cpu, even though am not using any synchronization. Why will this happen ?

I dont know why this is happening :wacko: , please anyone got any ideas ?

here is the pastebin link to the gpu_r kernel if anyone thinks its due to something in the kernel (which I doubt it should be)…

thanks all