Skip to content

tacuda v0.9.2

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Nov 11:39
· 1 commit to main since this release

0.9.2 — 2026-11-20

  • All six host Hilbert indicators now use the GPU's Hilbert core and the same
    build-generated phase weights. This fixes a nearly 25-degree phase difference
    and a trend-mode change on flat input when glibc selected different FMA paths.
  • The batch dispatcher processes 118 more legacy indicator families across rows.
    It also runs eligible short histories for the remaining 29 legacy families in
    one launch. Set TACUDA_BATCH_FUSED=0 to select the per-series legacy path.
    TA-Lib 0.8.3 formulas now size scratch per formula and distribute long batches
    across CUDA blocks.
  • STDDEV uses centered window moments. SUM and STDDEV recover after a missing
    sample leaves the window. Leading NaNs and row-local warm-up now agree across
    affected host, device and batch indicators, including MIDPOINT.
  • Python rejects invalid CUDA stream handles with ValueError before calling
    the driver. Use apo_ma or ppo_ma to select a moving-average type.
  • Added regression coverage for Hilbert FMA dispatch, batch row isolation,
    short-series dispatch, NaN handling and Python stream validation.
    73347c7738243979567603663b45d03efa6ccb627427ae8e15a65e81b6de7c34 libtacuda.so
    9a3840c8949e4bc1e7412c296fc4e7dd3ec8a42433b1dd8c34ccc98366b34e4b tacuda.dll