I have one kernel on multiple devices at the same time. Currently i enqueue one kernel and than a read-command to check the return value of the kernel. The kernel currently only contains the md5-algorithm.
My Problem is that every read causes a gab in the timeline on the GPU. The Cpu doesn't have these gabs.
I can avoid these gabs with enqueing multiple kernels but in this case i had to sync the devices and i don't know how. I want to interrupt the kernels on other devices, if the solution is found on one device.
I am using a Radeon HD 5850 with Catalyst 11.6 on Win7 x64 with an Core i7 920 as CPU.
A screenshot of the profileroutput can be found at: here
Does anybody know how to sync within multiple devices or why the GPU has these gabs?