ChASE Library Release Notes

Overview

This release represents significant advancements in the ChASE (Chebyshev Accelerated Subspace Extraction) library, with major enhancements to pseudo-Hermitian matrix support to compute interior eigenpairs with eigenvalues around 0, distributed-memory GPU capabilities, QR factorization algorithms, and user interfaces.

🎯 Major Features

1. Pseudo-Hermitian Matrix Support

• Extended ChASE to compute lower excited states of full Bethe-Salpeter Equation (BSE) matrices
• Implemented positive-negative symmetry handling in ChASE-CPU, ChASE-GPU, pChASE-CPU and pChASE-GPU
• Improved locking mechanism for negative components with K-conjugate operation
• Added damping factor for lower block initialization to improve numerical stability
• Paper on arxiv: Chebyshev Accelerated Subspace Eigensolver for Pseudo-hermitian Hamiltonians

2. Distributed-Memory GPU Support (pChASE-GPU)

• GPU-resident Lanczos implementation with fused kernels
• Optional warm-up procedures in constructor for NCCL initialization

3. Advanced Householder QR Factorization

• CUDA-aware-MPI and NCCL Householder QR: Two-level padding/cleaning strategy with full-path diagnostics
  • Level-1: Panel pre-clean (coarse)
  • Level-2: Column split-and-pad (fine)
  • Full-height GEMM updates with NCCL allreduce
• Distributed CPU Householder QR:
  • cpu_distributed_houseQR_panel_factor_block_cyclic_1d
  • cpu_distributed_blocked_houseQR_formQ_block_cyclic_1d

4. Enhanced Fortran/C Interfaces

• Added second initializer for Fortran interfaces (internal buffer management)
• Users no longer need to provide external buffers for eigenvectors/eigenvalues
• Configuration setter functions added to Fortran/C interfaces
• Support for both block-block and block-cyclic distributions simultaneously

📦 Build & Configuration

CMake Improvements

• Minimal drivers are split into chase_driver_cpu (links chase_cpu, and cublaspp/cusolverpp when CUDA is enabled so shared headers link cleanly) and chase_driver_gpu (links chase_gpu when CUDA is enabled); the old single target chase_driver is removed.
• Made ScaLAPACK optional
• Added pkg-config support
• Improved CMake configuration with chase_config.h.in
• Default CUDA architectures: 80;90 (Ampere and Hopper)

Configuration Options

• Environment variable support for Householder parameters
• CHASE_PH_LANCZOS_DIAG: Optional damping factor info print (default: OFF)

🔍 Logging & Diagnostics

• Centralized ChASE logger introduced
• Replaced direct stdout with configurable logging
• Log levels: Trace, Debug, Info, Warning, Error

🧪 Testing

• Added Fortran interface tests
• Full runs of all builds in unit tests

🔗 API Changes

New Unified Configuration Setters

All precision/matrix-type independent:

chase_set_tol(tol)
chase_set_deg(deg)
chase_set_max_deg(max_deg)
chase_set_max_iter(max_iter)
chase_set_lanczos_iter(lanczos_iter)
chase_set_num_lanczos(num_lanczos)
chase_set_approx(flag)
chase_set_opt(flag)
chase_set_cholqr(flag)
chase_enable_sym_check(flag)
chase_set_cluster_aware_degrees(flag)

📌 Migration Notes

For Users Upgrading

1. ScaLAPACK is now optional: Distributed Householder QR is the default
2. Logger API: Direct printf/cout calls replaced with GetLogger().Log()
3. Configuration: Fortran: new unified setters work across all precision/matrix types
4. Fortran buffers: Optional external buffers - library can manage internally

👥 Contributors

This release includes contributions from Edoardo Di Napoli, Clément Richefort and Xinzhe Wu, with major efforts in:

• Pseudo-Hermitian extensions
• Distributed GPU implementations
• NCCL-based linear algebra
• Interface improvements