stzGpu

(no comment in the source: the name reads as a sentence) initno extension

(no comment in the source: the name reads as a sentence) is availableno extension

(no comment in the source: the name reads as a sentence) device nameno extension

(no comment in the source: the name reads as a sentence) adapter countno extension

Upload a vector; it lives on the device until Free() (or eviction under VRAM pressure -- bounded, counted, never silent).no extension

One-shot convenience: upload the bound vectors, run the kernel, read the result back, free everything temporary. Data in, DATA out.QCSXTZZZIBWU

Multi-input kernels over ALREADY-RESIDENT buffers: no transfer at all. paBindings binds vector names to stzGpuBuffer objects (plus scalars): oG.ApplyOnQ(k, [ :A = oBuf1, :B = oBuf2, :alpha = 2 ])no extension

Measure where the GPU actually starts winning ON THIS MACHINE, store the crossover in the calibration store, and persist it (the faces auto-load it in later sessions). Warm-min discipline throughout -- G0's clock inversion means sustained numbers flatter the GPU. Returns [ nThreshold, [ [nProblemSize, nCpuMs, nGpuMs], ... ] ].no extension

The same, on a caller-chosen ladder of corpus sizes (d fixed at 64: a representative embedding width; problem size = n*d).no extension

pCand may be an stzKernelMaker on the SAME declarations, or raw WGSL on the same binding contract (how a hand-written variant -- or a cheat -- gets checked). Options: [ :reps = 7, :band = 0.000001 ].no extension

(no comment in the source: the name reads as a sentence) verify withno extension

The roofline alone: could a dispatch touching nBytes finish in nMs on this device? Pure judgement over the MEASURED floors -- exposed so a guard can hand it an impossible number and watch it refuse.no extension

The measured floors themselves: [ :floorgbs, :submitfloorms ]no extension

(no comment in the source: the name reads as a sentence) load calibrationno extension

one rung: warm-min per-query ms through the REAL seam, both routes Measured 2026-09-09 on this laptop's RTX 3050: after ~10 s idle, every submit pays a fixed ~2 ms and the copy floor reads 13 GB/s instead of 85 -- and hundreds of tiny queries never lift it, while ~150 ms of heavy work does. Wake() is that work, ADAPTIVE: it copies until a copy runs at full speed, settles, and stops -- the budget is a cap. Every timing this face takes calls it first; a caller measuring on its own should too. Returns the copies it took (an awake device answers in ~20). THE CAUSE AND THE SYMPTOM. A shader that fails validation raises two errors: the parser's, with the offending name and line, and then the pipeline's, which only says the module is invalid. FirstError() is the parser's; LastError() is whatever came last. A compile clears both first.no extension

(no comment in the source: the name reads as a sentence) last errorno extension

(no comment in the source: the name reads as a sentence) clear errorsno extension

(no comment in the source: the name reads as a sentence) wakeno extension

Measured 2026-09-09 on the RTX 3050 laptop at 384 dims: the GPU loses at 1,000 (0.39x) and 4,000 (0.98x) and wins at 16,000 (1.42x); the multicore top-k switches on at n*d >= 8M and the window may close again above it -- which is why the rung above that gate is measured too, and why the store is SHAPED: a class that wins carries its win, and the flat line stays beyond the ladder where a class lost.no extension

(no comment in the source: the name reads as a sentence) calibrate knn resident withno extension

This is the ONE-SHOT doorway: upload, run, download. A plane holding its signal on the device uses the buffer form (StzEngineGpuOpConvolveReal over buffer ids) and pays no marshalling -- the sound desk's route.no extension

Report: [ :m, :n, :d, :hiddenn, :clocks, :refgpums, :refwallms, :variants = [ [name, verified, gpums, wallms, ratio, ratiowall], ... ], :winner = name, :ratio = generic/winner on the GPU clock, :stored = bool ]no extension

(no comment in the source: the name reads as a sentence) foundry pairdist withno extension

THE ATTENTION FOUNDRY (GK2c). The backbone's fused attention -- one workgroup per (head, query row) -- against three reduction-form variants at workgroup widths 64, 128 and 256, under the checker on the neural DLL's device; the winner recorded for the (tokens, width, head_dim) class under the op name "attention", persisted, and pushed into the neural DLL's table so the next forward dispatches it.no extension

(no comment in the source: the name reads as a sentence) foundry attention withno extension

THE MATMUL FOUNDRY (GK2b's matmul leg). C(m x n) = A(m x k) B(k x n): tile16 (the generic), tile8, reg2, reg4 under the checker at the asked shape and a hidden one; the winner recorded for the (m, n, k) class under the op name "matmul" and persisted. The backbone compiles the table's winner for its own shapes with the bias fused.no extension

(no comment in the source: the name reads as a sentence) foundry matmul withno extension

THE k-NN FOUNDRY (GS6e). The UMAP k-NN kernel lives in the stats DLL on its own device, so its enumeration runs there; the verdict is recorded in that DLL's variant table AND persisted here under the op name "umap_knn", beside pairdist's rows, so the next session finds it. The stats DLL reads the persisted rows back through StzUmapKnnVariantsSync() at its first fit.no extension

(no comment in the source: the name reads as a sentence) foundry umap knn withno extension

The kill line, per adapter: a grid of single-query shapes; TRUE if any class found a variant worth the margin.no extension

Report: [ :cells = [[d, n, cpuMs, gpuMs, ratio], ...], :flat = the flat line stored, :flatwhy = "crossover" | "beyond-ladder", :hidden = [ [d, n, cpuMs, gpuMs, predicted, measured, agree], ... ], :adapter = name ]no extension

(no comment in the source: the name reads as a sentence) calibrate shaped withno extension