NEWS

---

cudacellr 0.4.0.9000

• Native CUDA PCA preparation preserves selected counts as a sparse matrix through upload, avoiding the dense host matrix and its transposed copy. PCA still uses float64 dense working storage on the GPU. CPU and torch preparation and public return objects are unchanged. Inputs smaller than 1 MiB retain the dense path to avoid sparse setup overhead.
• Invalid PCA component counts are rejected before preparing the selected input matrix. The development version requires cudaverse 0.4.2.
• Added pinned synthetic/PBMC benchmarks, numerical and performance gates, CUDA transfer traces, and a measured performance/memory guide. The validated sparse 50,000-cell case reduced peak host memory by 80.9%; PBMC by 25.1%.

cudacellr 0.4.0 (2026-10-27)

• Normalize very small positive libraries without overflowing an intermediate scale factor, and reject non-finite library totals.

• Reject component and feature counts beyond the supported integer range with an informative error before integer conversion.

• Replaces the former cudatensr, cudasparsr, and cudalearnr dependencies with the unified cudaverse package.

• Preserves all existing single-cell workflow and adapter entry points.

• Keeps cuda_provenance() as a re-export of the canonical cudaverse generic.

cudacellr 0.3.0

• Added cudacell_seurat() for a native Seurat v5 workflow using only the optional SeuratObject (>= 5.0.0) package. It writes normalized expression and feature statistics to an Assay5, PCA results to a DimReduc, exact kNN results to a Neighbor, size factors to cell metadata, and parameters and compute provenance to a deterministic tool record.
• Seurat output names and keys are checked for cross-type collisions before device selection, layer realization, or computation. Existing object state is preserved, and overwrite = TRUE replaces only explicitly named cudacellr outputs.
• Non-memory-backed Seurat layers are never realized silently. realize = TRUE explicitly opts into an in-memory sparse realization and records the original class and materialization decision.
• Added cudacell_sce() for a native, non-destructive SingleCellExperiment workflow. Normalized expression, PCA, HVG statistics, kNN relationships, and compute provenance are written to the corresponding assay, reduced dimension, rowData, colPair, and metadata locations.
• Existing assays, row and column metadata, reduced dimensions, alternative experiments, pairings, canonical size factors, labels, and user metadata are preserved. Output collisions fail before computation unless explicitly replaced with overwrite = TRUE.
• Delayed assays are never materialized silently. realize = TRUE is required to opt into an in-memory sparse realization.
• cuda_provenance() now dispatches through the shared cudaverse S3 generic and reads the ordered computation record directly from SingleCellExperiment and Seurat objects produced by cudacellr.
• cuda_cell_pca() and cudacell_workflow() now expose normalization scale_factor, log1p, and min_mean controls; the complete workflow also exposes PCA scaling through scale..

cudacellr 0.2.0

• Normalization, HVG selection, PCA, and kNN now compose into one ordered cudaverse-stage/1 provenance record, including upstream sparse materialization.
• cudacell_workflow() reports its aggregate CPU/CUDA/hybrid compute device, and its print method no longer implies that the PCA device represents the complete workflow.
• Strict CUDA requests are validated before CPU preprocessing begins, avoiding expensive work before an unavailable-device error.
• Re-exported cuda_provenance() as the common inspector for single-cell results.

cudacellr 0.1.2

• Feature and cell identifiers now survive normalization, HVG selection, PCA, kNN, and the complete workflow. HVG results retain original feature indices, including when feature names are duplicated.
• Added a concise print() method for workflow results and strict validation of the log1p flag.

cudacellr 0.1.1

• cuda_cell_neighbors() and cudacell_workflow() now expose batch_size, using the bounded-memory exact kNN implementation in cudalearnr.
• cudacell_workflow() now reuses its normalized matrix and variable-feature result for PCA instead of performing both preprocessing stages twice.