The Fetch Size != (threads*fetch instructions*size of fetch), is this intentional?
For example, in DCT there are 24 Fetch instructions of float type each from global memory reported by the profiler but the Fetch Size is 65KB.
24*32*4096*4096 = 12884901888 bits
65536*1024*8 = 536870912
Hi ryta,
This is a known issue and we are working towards it.
Can you please help us by providing the following details about your system:
CPU,GPU, SDK version, driver version, profiler version
So the Fetch Size is reporting incorrect values right? The correct value for teh above example is the "24*32*4096*4096=1288490188" right?
Please let me know.
System:
MSI 790FX-GD70
AMD Phenom x4 II 965
AMD 5870 Radeon
4GB DDR3 1600
Catalyst 10.11
Profiler 1.4 (2.0 is also broken, too much to use at the moment)
SDK 2.2
Originally posted by: himanshu.gautam Hi ryta,
This is a known issue and we are working towards it.
Can you please help us by providing the following details about your system:
CPU,GPU, SDK version, driver version, profiler version
Also, even then for the above example the Fetch Busy is 22.44 and the execution time is 26.45, so the actual time the fetch units are busy should be 26.45*.2244 = 5.93538ms
However, if I calculate the time using bits fethced, I get (24*32*4096*4096)/(256*4800*4) = 10.48576ms
Can you tell me what I might be missing here? The Cache hit value is 0 btw.
Thanks Ryta.
I have forwarded this to AMD profiler team.
Originally posted by: himanshu.gautam Thanks Ryta.
I have forwarded this to AMD profiler team.
Could you please help me with the above questions?
This issue is being looked into by the developer teams. I will let you know of the progress. Thanks for reporting this issue.