4 ms·
I don't think CUB is doing decoupled look-back, the reference you want is: https://research.nvidia.com/publication/single-pass-parallel-prefix-scan-decoupled-lo
by raphlinus 5y ago
I don't think CUB is doing decoupled look-back, the reference you want is: https://research.nvidia.com/publication/single-pass-parallel-prefix-scan-decoupled-look-back https://research.nvidia.com/publication/single-pass-parallel...
It doesn't use the word "spin" but repeated polling (step 4 in the algorithm presented in section 4.1, particularly when the flag is X) is basically the same.
- dragontamer 5y ago3rd paragraph: > In this report, we describe the decoupled-lookback method of single-pass parallel prefix scan and its implementation within the open-source CUB library of GPU parallel primitives The CUB-library also states: https://nvlabs.github.io/cub/structcub_1_1_device_scan.html https://nvlabs.github.io/cub/structcub_1_1_device_scan.html >> As of CUB 1.0.1 (2013), CUB's device-wide scan APIs have implemented our "decoupled look-back" algorithm for performing global prefix scan with only a single pass through the input data, as described in our 2016 technical report [1] Where [1] is a footnote pointing at the exact paper you just linked. ----------- > It doesn't use the word "spin" but repeated polling (step 4 in the algorithm presented in section 4.1, particularly when the flag is X) is basically the same. That certainly sounds spinlock-ish. At least that gives me what to look for in the code.
- raphlinus 5y agoAh, good then, I didn't know that, thanks for the cite. I haven't studied the CUB code on it carefully.