arXiv subjects

Sonia Rani Gupta

Publications and source records attributed to Sonia Rani Gupta.

3 recordsLinked to original sources

arXiv2026

FlashAttention for Scalable Vector Architectures

Inference with transformer models on CPUs is increasingly important, especially for Small Language Models (SLMs), where vector architectures are emerging as a promising execution substrate. The attention module is a major bottleneck due to high memory bandwidth requirements; FlashAttention mitigates this by fusing operations to improve data locality and reduce intermediate memory traffic. In this paper, we present FlashAttention-V, a blocked FlashAttention for scalable vector architectures that adapts efficiently from short to very long vectors by exploiting parallelism across attention heads, inter-head packing to enable efficient utilization of vector lengths beyond the head dimension, and improving vector register utilization and memory access locality. We integrate FlashAttention-V into ggml within llama.cpp and evaluate it on TinyLlama, Llama 3.2, Qwen2.5, and Pythia-410M using gem5 and a Banana Pi BPI-F3. On the Banana Pi BPI-F3, we confirm that loop reordering and loop unrolling across attention heads are effective optimization principles, scaling performance gains with larger models and most pronounced with short contexts and during decoding. Simulation-based analysis shows that FlashAttention-V achieves 22x-42x speedup over scalar FlashAttention at 512-bit VL in prefill, with an additional 2x-2.5x gain scaling to 64 lanes and 4096-bit VL. During decode, FlashAttention-V achieves 8x-11x speedup using 512-bit vector lengths over scalar FlashAttention, with performance showing diminishing sensitivity to vector width and lane count due to single-token, memory-bound execution. We further identify structural bottlenecks in Q8_0 quantized linear layers that limit arithmetic amortization under long-vector execution, consistent across RVV and Arm SVE, indicating that current quantization formats pose a fundamental challenge to long-vector scalability.

cs.LG↗

arXiv2023

Challenges and Opportunities in the Co-design of Convolutions and RISC-V Vector Processors

The RISC-V "V" extension introduces vector processing to the RISC-V architecture. Unlike most SIMD extensions, it supports long vectors which can result in significant improvement of multiple applications. In this paper, we present our ongoing research to implement and optimize a vectorized Winograd algorithm used in convolutional layers on RISC-V Vector(RISC-VV) processors. Our study identifies effective techniques for optimizing the kernels of Winograd on RISC-VV using intrinsic instructions, and showcases how certain instructions offer better performance. Our co-design findings suggest that the Winograd algorithm benefits from vector lengths up to 2048 bits and cache sizes up to 64MB. We use our experience with Winograd to highlight potential enhancements for the standard that would simplify code generation and aid low-level programming. Finally, we share our experience from experimenting with forks of gem5 for RISC-VV and stress the importance of a mature software ecosystem, to facilitate design space exploration and architectural optimization.

cs.DC↗

arXiv2022

Accelerating CNN inference on long vector architectures via co-design

CPU-based inference can be an alternative to off-chip accelerators, and vector architectures are a promising option due to their efficiency. However, the large design space of convolutional algorithms and hardware implementations makes it challenging to select the best options. This paper presents ongoing research into co-designing vector architectures for CPU-based CNN inference, focusing on the im2col+GEMM and Winograd kernels. Using the Gem5 simulator, we examine the impact of various hardware microarchitectural features on RISC-V Vector and ARM-SVE ISAs. We also study the impact of several BLIS-like algorithmic optimizations on im2col+GEMM. Our co-design study shows that longer vector lengths and larger caches can improve performance by 5x with our optimized CNN kernels, compared to a vector length of 512-bit and 1MB of L2 cache. For Winograd, we present a novel approach of inter-tile parallelization that exploits longer vector lengths and offers high memory reuse, resulting in up to 2.4x performance improvement for non-strided convolutional layers with 3x3 kernel size. Our study also shows that Winograd requires smaller cache sizes compared to im2col+GEMM.

cs.DC↗