Hello, this is my first topic and i hope that you understand my doubt.
In the page 34 of the best practices guide v3.0, it is said that the
listing 3.12 has ‘wasted bandwidth’ for the compute capabilities 1.2 and higher.
I believe that this is an error because in my opinion this case is resolved
with only one transaction of 64bytes. There’s not wasted bandwidth, which bandwidth?
e.g: shift = 13, griddim=(16384,1,1) and blockdim=(64,1,1)
first warp, first halfwarp
thread 0 accesses to pos 13 and take from 13 to 44, and in this segment lies the rest
of the threads of this halfwarp. Then, this segment degrades to 64 bytes et voilà ,
anything byte is wasted.
Each float is 4 bytes, so threads 0 to 31 would access bytes 52 to 176. 1st half warp fetches 128 bytes, second does 64+32 bytes. If shift were 0 then it would only be 128 bytes.
Each float is 4 bytes, so threads 0 to 31 would access bytes 52 to 176. 1st half warp fetches 128 bytes, second does 64+32 bytes. If shift were 0 then it would only be 128 bytes.
Hello! Thank your for your response. I’ve been investigating and this is my conclusion:
It’s almost perfect:
For the first halfwarp, it takes a 128bytes segment.
For the second halfwarp, it takes the same 128 segment but degradated first to 64 and second to 32bytes segment. Then it takes the following 128 bytes that are degradated to a 64 bytes segment.
If we consider that are 4 byte words, the first halfwarp wastes 16 words( 64 bytes) and the second halfwarp wastes 3 words(12 bytes)
Hello! Thank your for your response. I’ve been investigating and this is my conclusion:
It’s almost perfect:
For the first halfwarp, it takes a 128bytes segment.
For the second halfwarp, it takes the same 128 segment but degradated first to 64 and second to 32bytes segment. Then it takes the following 128 bytes that are degradated to a 64 bytes segment.
If we consider that are 4 byte words, the first halfwarp wastes 16 words( 64 bytes) and the second halfwarp wastes 3 words(12 bytes)