--------------------------------------------------------------------------- hsum256 — copied from cpu_kernels_avx2.cpp so the micro-bench matches the

Get a unique index in [0, 1325] for this combo Formula: index = c2*(c2-1)/2 + c1 (where c1 < c2)

Parse "2s" into {rank:0, suit:'s'}. Returns nulls if invalid.

Range-clamped payload for inputs feeding into reductions. No 1e6 spikes (which would invalidate dot-product comparisons due to FP non-associativity ra

* @brief All precomputed inputs a backend needs to run DCFR. * * All pointers are borrowed: the Solver orchestrator owns the underlying data * and

--------------------------------------------------------------------------- Baseline kernel — bit-identical to the production showdown_oop_full body.

--------------------------------------------------------------------------- Signed-valid zero-rake kernel. Exact only when tie payoff is zero and lose

Returns total frequency of "aggressive" actions (Bet_*, Raise*, All-in) * for ranking. Folds/Checks/Calls don't count. Used by `sortAggregates`.

Name of the active backend (for logging / UI indicator)

* One spot to solve. Filename is the cache key; mirrors §6.4 "Phase 1 spot ID * scheme" of PRESOLVE_BUNDLE_PLAN.md.

v1.2.2: build the CLI argv vector for `deepsolver_core` from a SolverRequest. Extracted from `try_run_solver` so the estimate-only path (`run_estimate

Bytes for the CPU CFR state. Counts regrets + strategy_sum + current_strategy (three arrays). action_values is a transient stack-allocated vector in C

Bytes for the GPU CFR state. Lives in VRAM. Includes: - regrets, strategy_sum, current_strategy, action_values: each N×A×nc floats - reach_scratch_oop

v1.7.0: extra heap held by the levelized CPU backend on top of the reference state. LevelizedCpuBackend allocates three flat [N × nc] float buffers —

Bytes for the strategy-tree EV cache. The current implementation stores `<node_id, vector<float>>` for every visited inner node × 2 perspectives (OOP-

Map our MATCHUPS positions to chart hero_position labels. They differ * for one role: our "MP" is the same seat as the dataset's "HJ" in * 6-hande

Combine all non-fold actions in a chart into a single in-range string. * Mirrors the helper in GtoChartBrowser but available standalone here.

Check if the user has PRO solver access

Per-combo best-response value at root for the given player (0 = OOP, 1 = IP). Empty on backends that don't support GPU postsolve.

Returns CUDA compute capability of device 0 as `major*10 + minor` (e.g. 89 for Ada, 61 for Pascal, 120 for Blackwell). Returns 0 when no CUDA device i

Why was the CUDA backend skipped? Returns an empty string when the GPU IS usable (caller should not surface anything in that case). Otherwise returns

Decide whether the planned solve fits the budget. Order of severity (returned first wins): GPU > matchup > host total > strategy tree > JSON. We retu

* Pull a top-level `export const NAME ... = [ ... ];` block out of a TS file * and `eval` the array body as JS. Works because object/array literals a

Find the best chart in `scenarios` matching (folder, position, scenario, * effective_bb). Tries the scenario candidates in order; first folder+posit

Locate the bundled `gto_output/` directory. Search order — covers every layout we care about: 1. `<exe>/resources/gto_output/` Tauri 2 production wit

Mirrors backend should_force_allin: if committing this bet leaves SPR * below threshold after opponent calls, the backend collapses the sizing * i

Format an EV value for cell display. Signed; keep 1 decimal below 100 so * close mixes stay comparable (the old >=10 branch rounded 24.3 -> "24", *

Build a human-readable diagnostic line. Shape matches the format the CLI's error JSON expects so the message can be embedded directly: "Matchup preco

* Get bet sizes for the current street.

Generate heatmap color for a single action frequency

Parse board string into individual card strings

Parse simple range text: supports "AA,AKs,KK" and "AA:1.0,AKs:0.5"

Count how many bits are set in the lower 13 bits

Base64url (no padding) of N random bytes. Length of output is ~ceil(4N/3).

Run the solver engine with the given request, with automatic GPU→CPU fallback. Behavior: - If `request.backend` is "cpu" (or unset in a CPU-only cont

For a (potType, side) combination, return scenario_type names ranked * by how well they fit. The first match found in the bundled charts wins. * M

Phase 2 (10-point plan): callers can hand in the canonical-combo count and a memory budget so the runout-enumeration gate is byte-based instead of an