Adrián Castelló
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Adrián Castelló has a total of 85 co-authors. Here are the top co-authors by number of shared publications:
- Enrique S. Quintana Orti
Universitat Politécnica de Valencia, Valencia, Spain
39 shared publications - Héctor Martínez
Universidad de Córdoba, Córdoba, Spain
22 shared publications - Francisco D. Igual
Universidad Complutense de Madrid
13 shared publications - Sandra Catalán
Universitat Jaume I
11 shared publications - Antonio José Peña
Barcelona Supercomputing Center (BSC), Barcelona, Spain
9 shared publications Rafael Mayo
9 shared publications
Latest Publications
- Open Access
- 13 April 2026
- Published by Springer Nature in Computing
- Vol. 108 (5) , 66
Sparse matrix–vector product on RISC-V processors with SIMD units
The sparse matrix–vector multiplication ( SpMV ) kernel is a key kernel in scientific and engineering applications, forming the core of many iterative solvers for linear systems and eigenvalue problems. Due to its low arithmetic intensity and irregular memory access patterns, SpMV remains memory-bound on modern architectures, making its efficient implementation particularly challenging. This paper presents vectorized SpMV routines for RISC-V processors with SIMD support, exploiting the RISC-V Vector Extension (RVV 1.0). We implement and evaluate three storage formats—CSR (Compressed Sparse Row), SELL- p (a vector-friendly variant of ELLPACK), and JDS (Jagged Diagonal Storage)—providing low-level implementations that leverage RVV intrinsics. Performance is assessed on two commercial RISC-V platforms (CanMV-K230 and BananaPi F3) with 128-bit and 256-bit vector registers, and on the EPAC research system featuring 16,384-bit vectors. Results show that the vectorized routines significantly outperform scalar baselines, achieving a variety of speed-ups depending on the format and architecture. These findings highlight the potential of open RISC-V architectures for high-performance sparse linear algebra and provide a foundation for future vector-aware sparse kernel optimizations.
- 13 April 2026
- Published by Springer Nature in Computing
- Vol. 108 (5) , 66
Open AccessSparse matrix–vector product on RISC-V processors with SIMD units
The sparse matrix–vector multiplication ( SpMV ) kernel is a key kernel in scientific and engineering applications, forming the core of many iterative solvers for linear systems and eigenvalue problems. Due to its low arithmetic intensity and irregular memory access patterns, SpMV remains memory-bound on modern architectures, making its efficient implementation particularly challenging. This paper presents vectorized SpMV routines for RISC-V processors with SIMD support, exploiting the RISC-V Vector Extension (RVV 1.0). We implement and evaluate three storage formats—CSR (Compressed Sparse Row), SELL- p (a vector-friendly variant of ELLPACK), and JDS (Jagged Diagonal Storage)—providing low-level implementations that leverage RVV intrinsics. Performance is assessed on two commercial RISC-V platforms (CanMV-K230 and BananaPi F3) with 128-bit and 256-bit vector registers, and on the EPAC research system featuring 16,384-bit vectors. Results show that the vectorized routines significantly outperform scalar baselines, achieving a variety of speed-ups depending on the format and architecture. These findings highlight the potential of open RISC-V architectures for high-performance sparse linear algebra and provide a foundation for future vector-aware sparse kernel optimizations.
- Open Access
- 26 February 2026
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 82 (4) , 206
Enhancing transformer performance and portability through auto-tuning frameworks
Transformer-based models such as BERT and GPT2 have become the foundation of many modern applications, yet their execution requires substantial computational and memory resources. To address these challenges, recent advances in compiler technology and hardware accelerators have introduced new opportunities for performance portability. In this work, we evaluate JAX and TVM as high-level frameworks that combine a NumPy-like programming model with Just-In-Time (JIT) or Ahead-of-Time (AOT) code optimization and compilation, enabling efficient execution across CPUs or GPUs, and in the case of JAX, on TPUs as well. We present systematic implementations of the core Transformer encoder and decoder blocks in JAX and TVM and compare their automatically optimized code against NumPy and CuPy baselines. Our experimental study covers heterogeneous hardware platforms (AMD CPU, NVIDIA GPUs, and Google TPUs) and multiple arithmetic precisions (FP32, BF16, INT8, and INT32). Results show that JAX and TVM deliver significant performance improvements over standard libraries, while reducing the programming effort required to adapt to different hardware. These findings demonstrate the potential of JIT- and AOT-oriented frameworks to serve as a portable and efficient solution for deploying Transformer workloads in diverse computing environments.
- 26 February 2026
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 82 (4) , 206
Open AccessEnhancing transformer performance and portability through auto-tuning frameworks
Transformer-based models such as BERT and GPT2 have become the foundation of many modern applications, yet their execution requires substantial computational and memory resources. To address these challenges, recent advances in compiler technology and hardware accelerators have introduced new opportunities for performance portability. In this work, we evaluate JAX and TVM as high-level frameworks that combine a NumPy-like programming model with Just-In-Time (JIT) or Ahead-of-Time (AOT) code optimization and compilation, enabling efficient execution across CPUs or GPUs, and in the case of JAX, on TPUs as well. We present systematic implementations of the core Transformer encoder and decoder blocks in JAX and TVM and compare their automatically optimized code against NumPy and CuPy baselines. Our experimental study covers heterogeneous hardware platforms (AMD CPU, NVIDIA GPUs, and Google TPUs) and multiple arithmetic precisions (FP32, BF16, INT8, and INT32). Results show that JAX and TVM deliver significant performance improvements over standard libraries, while reducing the programming effort required to adapt to different hardware. These findings demonstrate the potential of JIT- and AOT-oriented frameworks to serve as a portable and efficient solution for deploying Transformer workloads in diverse computing environments.
- Open Access
- 12 February 2026
- Published by Elsevier in SSRN Electronic Journal
Inference Acceleration for Graph Convolutional Networks on RISC-V Systems
Graph Convolutional Networks (GCNs) are a widely used class of Graph Neural Networks for learning on graph-structured data. Despite their shallow architectures, inference with this class of models remains computationally challenging due to the scale and irregularity of real-world graphs, particularly on low-power multicore CPUs. In this paper, we investigate the optimization of GCN inference on RISC-V processor architectures. We evaluate three popular deep learning frameworks, namely DGL, PyTorch Geometric, and PyTorch, using standard citation benchmarks (Cora, PubMed, and CiteSeer) on a Banana Pi BPI-F3 board with an 8-core RISC-V processor. Based on the initial experimental analysis, we develop optimized kernels for the dense and sparse matrix multiplications, showing that these components dominate GCN inference time on low-power CPUs. Finally, we exploit mixed-precision arithmetic in GEMM to reduce memory bandwidth pressure and improve inference throughput. Overall, this work demonstrates the performance benefits of targeted kernel optimizations for efficient GCN inference on low-power multicore processors.
- 12 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessInference Acceleration for Graph Convolutional Networks on RISC-V Systems
Graph Convolutional Networks (GCNs) are a widely used class of Graph Neural Networks for learning on graph-structured data. Despite their shallow architectures, inference with this class of models remains computationally challenging due to the scale and irregularity of real-world graphs, particularly on low-power multicore CPUs. In this paper, we investigate the optimization of GCN inference on RISC-V processor architectures. We evaluate three popular deep learning frameworks, namely DGL, PyTorch Geometric, and PyTorch, using standard citation benchmarks (Cora, PubMed, and CiteSeer) on a Banana Pi BPI-F3 board with an 8-core RISC-V processor. Based on the initial experimental analysis, we develop optimized kernels for the dense and sparse matrix multiplications, showing that these components dominate GCN inference time on low-power CPUs. Finally, we exploit mixed-precision arithmetic in GEMM to reduce memory bandwidth pressure and improve inference throughput. Overall, this work demonstrates the performance benefits of targeted kernel optimizations for efficient GCN inference on low-power multicore processors.
- Open Access
- 25 January 2026
- Published by Association for Computing Machinery (ACM)
- p. 239-246
Migration of Ginkgo's Jacobi-Preconditioned CG Solver to Vector RISC-V
- 25 January 2026
- Published by Association for Computing Machinery (ACM)
- p. 239-246
Open AccessMigration of Ginkgo's Jacobi-Preconditioned CG Solver to Vector RISC-V
- 24 November 2025
- Published by Springer Nature
- p. 431-443
Generation of Mixed-Precision Kernels for Quantized Transformer Encoders with Exo
- 24 November 2025
- Published by Springer Nature
- p. 431-443
Generation of Mixed-Precision Kernels for Quantized Transformer Encoders with Exo
- 24 November 2025
- Published by Springer Nature
- p. 534-547
Evaluation of RVV-Enabled COTS Platforms with Matrix Multiplication and Exo
- 24 November 2025
- Published by Springer Nature
- p. 534-547
Evaluation of RVV-Enabled COTS Platforms with Matrix Multiplication and Exo
- Open Access
- 9 November 2025
- Published by Elsevier in Future Generation Computer Systems
- Vol. 177
The cambrian explosion of mixed-precision matrix multiplication for quantized deep learning inference
- 9 November 2025
- Published by Elsevier in Future Generation Computer Systems
- Vol. 177
Open AccessThe cambrian explosion of mixed-precision matrix multiplication for quantized deep learning inference
- Open Access
- 29 October 2025
- Published by Springer Nature in Research Square
Enhancing Transformer Performance and Portability through Auto-tuning Frameworks
Transformer-based models such as BERT and GPT2 have become the foundation of many modern applications, yet their execution requires substantial computational and memory resources. To address these challenges, recent advances in compiler technology and hardware accelerators have introduced new opportunities for performance portability. In this work, we evaluate JAX and TVM as high-level frameworks that combine a NumPy-like programming model with Just-In-Time (JIT) or Ahead-of-Time (AOT) code optimization and compilation, enabling efficient execution across CPUs or GPUs, and, in the case of JAX, on TPUs as well. We present systematic implementations of the core Transformer encoder and decoder blocks in JAX and TVM and compare their automatically optimized code against NumPy and CuPy baselines. Our experimental study covers heterogeneous hardware platforms (AMD CPU, NVIDIA GPUs, and Google TPUs) and multiple arithmetic precisions (FP32, INT8, and INT32). Results show that JAX and TVM deliver significant performance improvements over standard libraries, while reducing the programming effort required to adapt to different hardware. These findings demonstrate the potential of JIT- and AOT-oriented frameworks to serve as a portable and efficient solution for deploying Transformer workloads in diverse computing environments.
- 29 October 2025
- Published by Springer Nature in Research Square
Open AccessEnhancing Transformer Performance and Portability through Auto-tuning Frameworks
Transformer-based models such as BERT and GPT2 have become the foundation of many modern applications, yet their execution requires substantial computational and memory resources. To address these challenges, recent advances in compiler technology and hardware accelerators have introduced new opportunities for performance portability. In this work, we evaluate JAX and TVM as high-level frameworks that combine a NumPy-like programming model with Just-In-Time (JIT) or Ahead-of-Time (AOT) code optimization and compilation, enabling efficient execution across CPUs or GPUs, and, in the case of JAX, on TPUs as well. We present systematic implementations of the core Transformer encoder and decoder blocks in JAX and TVM and compare their automatically optimized code against NumPy and CuPy baselines. Our experimental study covers heterogeneous hardware platforms (AMD CPU, NVIDIA GPUs, and Google TPUs) and multiple arithmetic precisions (FP32, INT8, and INT32). Results show that JAX and TVM deliver significant performance improvements over standard libraries, while reducing the programming effort required to adapt to different hardware. These findings demonstrate the potential of JIT- and AOT-oriented frameworks to serve as a portable and efficient solution for deploying Transformer workloads in diverse computing environments.
- Open Access
- 13 July 2025
- Published by SAGE Publications in The International Journal of High Performance Computing Applications
- Vol. 39 (6) , 803-821
Characterization of quantized inference with transformer encoders on low power CPUs
Transformers, particularly large language models (LLMs), are revolutionizing applications in natural language processing and computer vision but at a high cost in memory, energy, and computational resources. Quantization has emerged as an effective compression method to alleviate these demands, reducing the bitwidth of model data and arithmetic precision to enable efficient inference on resource-constrained devices. This paper focuses on optimizing inference with transformer encoders on low-power general-purpose CPUs, as those often found in edge devices. Our key contributions include exposing the critical role of linear layers within transformer encoders on CPUs with a limited number of cores; developing mixed integer precision matrix multiplication on ARM and RISC-V CPUs; and evaluating performance impact and energy savings of quantized inference. In summary, this work highlights the advantages of applying quantization to transformer encoders on current single-core and multi-core low power CPUs, offering insights for efficient LLM deployment on edge platforms.
- 13 July 2025
- Published by SAGE Publications in The International Journal of High Performance Computing Applications
- Vol. 39 (6) , 803-821
Open AccessCharacterization of quantized inference with transformer encoders on low power CPUs
Transformers, particularly large language models (LLMs), are revolutionizing applications in natural language processing and computer vision but at a high cost in memory, energy, and computational resources. Quantization has emerged as an effective compression method to alleviate these demands, reducing the bitwidth of model data and arithmetic precision to enable efficient inference on resource-constrained devices. This paper focuses on optimizing inference with transformer encoders on low-power general-purpose CPUs, as those often found in edge devices. Our key contributions include exposing the critical role of linear layers within transformer encoders on CPUs with a limited number of cores; developing mixed integer precision matrix multiplication on ARM and RISC-V CPUs; and evaluating performance impact and energy savings of quantized inference. In summary, this work highlights the advantages of applying quantization to transformer encoders on current single-core and multi-core low power CPUs, offering insights for efficient LLM deployment on edge platforms.
- 10 June 2025
- Published by Springer Nature
- p. 288-292
One-to-Many Communication Primitives in Dragonfly Networks
- 10 June 2025
- Published by Springer Nature
- p. 288-292
One-to-Many Communication Primitives in Dragonfly Networks