Archives Discussions

[arvin99]

Hi friends,

I read AMD APP OpenCL Programming Guide rev 2.7 and I am confused about the meaning of max in-flight hw threads in table 5.4 CPU and GPU Performance Characteristic.

I am understand about the meaning of Simultaneous Executing Threads in table 5.4.

It means number of threads (for CPU) that executed together and for AMD GPU, it means number of stream processor  that executed together.

In my GPU (AMD Radeon  HD 6630M),  the number of stream processor  that can be executed together is 480 ( 6 SIMD unit x 16 stream core x 5 stream processor / processing element)

So, what is the meaning of max in-flight hw threads??

How do I find the number of  max in-flight hw threads in my GPU (AMD Radeon  HD 6630M) ??

[nou · Accepted solution]

to hide memory access latencies the work items are executed in cycles. that mean workitem 0-15 one tick, then 16-31, 32-47, 48-63 and then again 0-15. also if there are enough register and local memory there can be more workgroups executing on same CU at the same time. so that max in flight threads mean how many work items can start on GPU on the same time.

[arvin99]

Thanks