Convert to display string (e.g., "AhKh")

Parse a two-character card string (e.g., "As") to Card value

Horizontal sum of __m256, used by reductions.

Set a card in a bitmask

Mixed-magnitude payload that hits zero, negative, near-zero, and large. Use this for elementwise kernels (no reduction) where catastrophic cancellatio

Extract suit from card

Parse a board string like "AsKd7c" into array of Cards

Extract rank from card

Build a label→pct map from a solver result so two configs can be compared even if action ordering shifts.

Inline card construction

Get hand category string from rank

Compare two global_strategy maps. Every action present in either side must match within `tol` percentage points, and the action sets must agree.

Build dead card mask from board

Sparse valid mask: ~70% ones, 30% zeros (matches the typical density of the matchup_valid table after dead-card filtering).

Returns the grid label for a combo (e.g., "AA", "AKs", "AKo")

EV row for showdown: roughly equal split of {-1, 0, 1} when valid, 0 elsewhere. The category threshold is >0.5 / <-0.5, so this must include true win/

Check if this combo conflicts with a card mask (dead cards)

Parse TexasSolver range string, e.g. "AA:1.0,AKs:0.5,A4o:1.0" Returns a map of grid_label -> weight

Bytes required to hold all per-runout matchup tables. runout_tables × nc² × 2 (EV+valid) × sizeof(float) `runout_tables` is the number of distinct b

Convert Card to two-character string (e.g., Card(50) -> "As")

v1.8.2 A2: helper to compute the pre-thresholded category byte from an (ev, valid) pair, mirroring precompute_matchups()'s bucketing rules: valid == 0

============================================================================ DCFR discount-factor computation (shared by CPU + GPU backends) =========

* @brief Compute the canonical suit mapping for a given board. * * Algorithm: * 1. Analyze the board's suit pattern. * 2. For each starting co

--------------------------------------------------------------------------- Scenario runner. Two regimes: hot — 1 table, K calls. Same data reused →

Format internal chip amount as BB for display. 1 BB = 10 chips (same * convention as MATCHUPS / derivePotStack). Trims trailing .0 so whole * valu

Rough JSON response size estimator. Strategy tree dominates: each emitted node carries ~5 lists × 169 grid labels × ~30 bytes/entry.

Find the path to the deepsolver_core binary. In development, looks for it in the core/build/Release directory. In production, looks for the sidecar bi

Check if rank_bits form a straight, return straight "height" (0-9) or -1