Title:
SERIALIZING DIVERGENT ACCESSES USING PEELING
Document Type and Number:
WIPO Patent Application WO/2020/186630
Kind Code:
A1
Abstract:
Divergent memory accesses to a storage device are serialized using a per-warp address register. Serializing a divergent memory access includes storing, in per-thread general purpose registers (GPRs), address values indicating locations of thread data within one or more storage devices associated with a plurality of threads, each of the plurality of threads included in a warp. A peeling operation is iteratively performed until all threads corresponding to address values stored in the per-thread GPRs are processed. The peeling operation includes selecting an address value to be processed from among the address values stored in the per-thread GPR, and storing that address value in the per-warp address register. The address value to be processed is associated with an unprocessed thread. Threads associated with address values different from the address value to be processed are deactivated, threads associated with address values matching the address value to be processed are processed.
More Like This:
| JPH0312742 | CENTRAL PROCESSING UNIT |
| JP2015184979 | ARITHMETIC UNIT AND INSTRUCTION FETCH METHOD |
| JP4662744 | Use of SIMD instructions in Montgomery multiplication |
Inventors:
CHEN LIN (US)
HONG ZHOU (US)
HONG ZHOU (US)
Application Number:
PCT/CN2019/090160
Publication Date:
September 24, 2020
Filing Date:
June 05, 2019
Export Citation:
Assignee:
HUAWEI TECH CO LTD (CN)
International Classes:
G06F9/38
Foreign References:
| US20150277921A1 | 2015-10-01 | |||
| US20130346727A1 | 2013-12-26 | |||
| US20090182984A1 | 2009-07-16 | |||
| US20050125802A1 | 2005-06-09 |
Download PDF:
Previous Patent: CHILD SEAT RING, INTELLIGENT TOILET COVER AND INTELLIGENT TOILET
Next Patent: COMPUTE SHADER WARPS WITHOUT RAMP UP
Next Patent: COMPUTE SHADER WARPS WITHOUT RAMP UP