CHANGES IN VERSION 0.99.1 (2023 Nov)

• the R global option variable 'HIBAG_GPU_INIT_TRAIN_PREC' can be set before loading the HIBAG.gpu package via e.g., options(HIBAG_GPU_INIT_TRAIN_PREC="half"). It should be NULL (unset), 'auto', 'half', 'mixed', 'single' or 'double'. It can be used without calling hlaGPU_Init(,train_prec="") to reset the training precision.
• fix the GPU memory leaks

CHANGES IN VERSION 0.99.0 (2021 Oct)

• remove the dependency of the OpenCL R package
• reimplement the HIBAG GPU algorithm for speed-up
• new implementation using half and mixed precisions
• a new function hlaAttrBagging_MultiGPU() to leverage multiple GPU devices

• add KIR information
• optimize the GPU kernel by avoiding unnecessary single-precision calculation
• support the Windows platform

Performance

The ratios of running times for training HIBAG models:

using HIBAG v1.14.0 and HIBAG.gpu v0.9.0

CPU (1 core), the default installation from Bioconductor supporting SIMD SSE2 instructions, using Intel(R) Xeon(R) CPU E5-2630L @2.40GHz

CPU (1 core, POPCNT), optimization with Intel/AMD POPCNT instruction, using Intel(R) Xeon(R) CPU E5-2630L @2.40GHz

This work was made possible, in part, through HPC time donated by Microway, Inc. We gratefully acknowledge Microway for providing access to their GPU-accelerated compute cluster (http://www.microway.com/gpu-test-drive/).