imx6 opencl zero copy

cancel
Showing results for 
Show  only  | Search instead for 
Did you mean: 

imx6 opencl zero copy

Jump to solution
2,669 Views
liujkian
Contributor I

Hi, i have a question about the opencl map(clEnqueueMapBuffer ) mechanism. if a opencl host program allocs a block of memory  by itself (such as malloc) not clCreateBuffer,then it calls clEnqueueMapBuffer to map the memory to gpu.can gpu use the mapped memory directly (or DMA )without any data copy (or DMA transfer )to vivante opencl runtime lib or to gpu global memory. Any Tips?thanks

Tags (3)
1 Solution
1,423 Views
andre_silva
NXP Employee
NXP Employee

I just checked and cl_mem must be used to allocate memory clEnqueueMapBuffer, there isn´t another way to do fill cpu buffers to be used in gpu, so, zero copies in this case is not possible.

regards,

Andre

View solution in original post

0 Kudos
5 Replies
1,424 Views
andre_silva
NXP Employee
NXP Employee

I just checked and cl_mem must be used to allocate memory clEnqueueMapBuffer, there isn´t another way to do fill cpu buffers to be used in gpu, so, zero copies in this case is not possible.

regards,

Andre

0 Kudos
1,158 Views
korabelnikov
Contributor III

Hi,

Is any updates or tips on it? How to make host<->opencl memory access faster?

 

 

data transfers take a half of computation time on imx6qp

0 Kudos
1,423 Views
arthassin
Contributor II

I've tried this on imx6. Seems like the buffer is still copied. No any performance improvement was achieved.

0 Kudos
1,423 Views
andre_silva
NXP Employee
NXP Employee

I´m checking this information and will be back soon with the answer.

0 Kudos
1,423 Views
liujkian
Contributor I

anyone can help me?thanks

0 Kudos