Co-designing ab initio electronic structure methods on a RISC-V vector architecture

Area of Science:

• Computational physics and chemistry
• High-Performance Computing (HPC)
• Emerging computer architectures

Background:

• Electronic structure calculations are vital for materials and molecular science but computationally intensive.
• The eigenvalue problem is a common bottleneck in these scientific applications.
• High-Performance Computing (HPC) platforms are essential for tackling these demanding calculations.

Purpose of the Study:

• To investigate the feasibility of porting ab initio electronic structure codes to RISC-V Vector Processing Units (VPUs).
• To evaluate different vectorization strategies for optimizing performance on a RISC-V prototype.
• To provide feedback for hardware and software co-design for future exascale systems.

Main Methods:

• Developed a software tester based on a mini-app from the ELPA eigensolver library.
• Tested a user-space emulator (Vehave) and a RISC-V vector architecture on an FPGA.
• Implemented and compared various vectorization techniques, including autovectorization, loop fusion, and intrinsics.

Main Results:

• Observed a progressive reduction in vectorised instructions, executed instructions, and computing cycles with increasing vectorization complexity.
• Demonstrated the potential for substantial speed-up in eigenvalue problem computations.
• Validated the performance of different vectorization strategies on the RISC-V prototype.

Conclusions:

• Porting electronic structure codes to RISC-V VPUs is a viable approach for accelerating scientific computations.
• The findings support the advancement of computational materials and molecular science on EU-developed RISC-V technologies.
• This work offers valuable insights for hardware designers, engineers, and compiler developers involved in RISC-V co-design efforts.