Repository navigation
tacuda v0.9.1
0.9.1 — 2026-11-16
TA-Lib 0.8.3 compatibility: 237 functions, typed outputs and numerical corrections.
This patch release adds APIs without changing existing signatures. Existing
RSI/IMI flat-series values and the other corrections below intentionally change
numerical results; pin 0.9.0 when reproducing its historical output.
- Pin the Python/native correctness oracle to 0.8.3 and its complete
237-function manifest. Add the 76 functions introduced since 0.7.1,
reusing NVI/PVI and keeping distinct compatible ASI/kurtosis variants. - Extend MA types through ALMA (15), add selectable-MA device/batch variants,
and correct recursive composite seeding where the older implementation
only approximated the reference. - Generate int32 and mixed output APIs and bindings, including four-output
functions. Existing C signatures and existing float output types remain. - Align flat RSI/IMI with neutral 50, newest-tie index behavior, WILLR range
clipping, NATR's scale-invariant zero guard, and leading-NaN NVI/PVI
alignment with the new reference. Restore the formerly excluded NATR
zero-close dataset now that upstream's write bug is fixed. - New reference-backed functions use a CUDA batch baseline in double with
stream-owned scratch. Performance of long sequential histories remains a
separate optimization target; no universal speedup is claimed. MAVP and
MACDEXT also use this path for exact seeding/NaN semantics; their long-row
throughput and memory use can regress from 0.9.0. Ordinary MACD retains
its optimized path.
CUDA Graph compatibility:
- Scratch-backed device and batch capture paths check graph memory-node
support before allocating. A device without memory pools now receives
CT_STATUS_UNSUPPORTEDwith a capability diagnostic, including mixed-
precision wrappers. Ordinary fallback and allocation-free fused graphs
remain available; public signatures and ABI symbols are unchanged. - Capture tests isolate their own failed captures/runtime errors and validate
unsupported paths separately from kernel-only graph replay. Linux CTest
covers both behaviors with memory-pool support disabled by an interposer.
Indicator correctness corrections from the 2026-11-15 audit:
- STOCH retains its first valid K/D pair at the minimum sufficient input
length, including batch fallback. The lookback guard uses 64-bit arithmetic. - CORREL uses rebased local sliding sums instead of differences of long raw
prefixes. This fixes low-variance accuracy at large price offsets and makes
eager/graph results reproducible. Non-finite suffix behavior is preserved. - PVO computes full SMA-seeded volume EMAs and starts at
slowPeriod-1.
The old truncated-window approximation and its extra warm-up bar are removed. - NVI/PVI keep price ratios and cumulative products in double and round only
public outputs to float32. Fixed-tree scans eliminate scheduling-dependent
rounding. BBANDS (including periods above the fused limit) and BETA also use
fixed scan trees for reproducible eager execution and CUDA Graph replay. - GPU Hilbert DC phase uses build-host sin/cos coefficients and separate
multiply/add operations, correcting flat/candle HT_DCPHASE, HT_SINE and
HT_TRENDMODE differences. Coefficients are compiled into constant memory;
first-use CUDA Graph capture needs no host initialization or transfer. - Independent batch/device mathematical corpora, dispatch boundaries, long
numerical/repeatability checks and six native regressions now run in CTest.
These corrections change affected results even with TACUDA_BATCH_FUSED=0.
See docs/TA_LIB_COMPATIBILITY.md for numerical contracts and validation.
30738684ebb3dee36fe9e09fbf1e9ad3b4b04acb1ab80b79b6af53c7c65a33b5 libtacuda.so
074bd33d1550d9e189b0abc4a63caca26c7c8d653c13f244730fc52c794a37cd tacuda.dll