经管之家送您一份
应届毕业生专属福利!
求职就业群
感谢您参与论坛问题回答
经管之家送您两个论坛币!
+2 论坛币
Unlimited Vector Extension (UVE) is a novel ISA extension that incorporates data streaming and scalablevectorization. To do so, it implements a data streaming engine that allows detaching the main memoryaccess from computation, releasing some pressure from the main pipeline of the processor. This isdone by statically encoding the memory access pattern of streamable loops through a set of specialdescription instructions, to offload the corresponding address sequence generation to the streamingengine. This not only allows the computational code to be simplified by removing memory addressinginstructions,but also simplifies the vectorization of the loop, in turn increasing the throughput of theprocessing cores. However a new extension that implements target specific instructions, it still lackscompiler support to produce target code. Accordingly, in this thesis aims at taking the first steps in thedevelopment of a compiler for UVE by instantiating a new subtarget from RISC-V’s LLVM backend andcreating and encoding the extension’s instructions. To provide an initial integration with LLVM IR, a setof intrinsics is also proposed that match the instructions. Due to the intrinsic incompatibility betweenthe streaming paradigm from UVE and LLVM IR SSA form, a new approach to overcome this issue isintroduced based on pseudo-instructions . The new backend and LLVM IR intrinsics were evaluated witha set of benchmarks that highlight the main introduced features
扫码加我 拉你入群
请注明:姓名-公司-职位
以便审核进群资格,未注明则拒绝
|