[S4] Quantize, and emit the error bound into the generated table - #50
Conversation
Fill each declared node from on-knot samples, apply the stored scale with round-to-nearest (ties away from zero), and measure deviation of the quantized table from every supplied sample. The maximum is a durable i32-value-LSB const fragment for later emission; operator stats stay on the CLI. Missing nodes, ambiguous nodes, non-invertible scale, and i32 overflow are closed BakeError variants. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: eece7a42df
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Unrounded host bilinear can understate deviation at u16 off-knot samples because each runtime interpolation step rounds. The bound now takes the max of that host reconstruction and the integer X-then-Y path whenever both sample coordinates are exact u16 values. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 47bacba037
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Interpolate the i32 grid with weighted f64 mixes so extreme dequantized endpoints cannot overflow v1 - v0 to NaN. Reject NX*NY above 1_048_576 with a closed GridTooLarge error before allocating the cell matrix. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1e6bce7830
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Validated axes are strictly increasing, so fill and segment location can binary-search instead of scanning. A 65_536-by-2 grid with one sample per node no longer does billions of linear comparisons. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 452aea5bef
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Adding 0.5 then flooring maps the f64 immediately below 0.5 to 1. Split into trunc and fraction so values just under a half stay on the nearer integer, while exact halves still round away from zero. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
1 similar comment
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c0473f7502
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
mul_add can round a residual that is slightly above an integer onto that integer, so a plain ceil understated the Codex 1+eps / next_down(1) case. next_up on a nonzero magnitude makes that case emit n+1; exact-zero tables stay 0, and a residual just below an integer does not jump an extra LSB. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
Coupled-edit for the new host quantize API: the firmware README and crate docs name that ph-surfaces-bake quantizes a caller-stated grid and emits MAX_ERR_LSB as deviation from supplied samples, without dumping baker types into the runtime rustdoc. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7c5fadb0da
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Three f64 lerps can round a reconstruction upward so the residual lands just under an integer while the exact bilinear of the i32 grid is just over it. For exact u16 coordinates, MAX_ERR_LSB now uses num/den of that grid; fractional coordinates keep an outward lerp envelope. Exact-zero tables stay 0. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
A u16-only rational plus an 8-ULP lerp envelope still treats host float as the oracle for fractional coordinates. Decode IEEE bit-patterns as dyadics, reconstruct the i32 grid as a ratio at every sample, and ceil that residual. Drop the one-ULP ceil padding. Co-authored-by: Cursor <cursoragent@cursor.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 0e4be2195b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Unreduced add/mul overflowed i128 on ordinary 0.1 off-knot coordinates because shared 2^55 denominators were multiplied again. Cancel gcd first. A dyadic whose exponent is below -126 keeps m/2^126 so a finite 1e-300 sample still emits a bound instead of NonFiniteDeviation. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
…y finite. Unreduced i128 n/d overflows on 0.1-class interpolants and treats 1e-300 as non-finite. Keep powers of two in the exponent, and tell implementing agents in AGENTS.md not to substitute host f64 or ULP envelopes. Co-authored-by: Cursor <cursoragent@cursor.com>
|
@codex review
|
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e73087ae8a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Aligning 1e-300 with a unit reconstruction needs a ~1049-bit shift, so add returns the dominant term when checked_shl fails. A finite ceil that does not fit i32 is BakeError::BoundOverflow, not NonFiniteDeviation. Raise the baker implementation cap to 1,600. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 76a5a89441
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Returning only the dominant operand when exponent alignment exceeds 256 bits understates ceil for 1 - (-1e-300). Opposite-signed reconstructs keep the dominant term; same-sign exact integers round the bound up. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review Re-review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 719c615454
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
A ceil-oriented shortcut inside lerp turned 1+1e-300 into 2 and understated the residual of sample 2 against rows [1, 2]. Carry the unaligned addend through reconstruction; ceil of the two-term residual must not understate, and a finite 3e9+tiny still maps to BoundOverflow. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
The baker may take reviewed std crates, so MAX_ERR_LSB can ceil an exact finite residual without a firmware-shaped 256-bit add and without treating overflow as non-finite. Co-authored-by: Cursor <cursoragent@cursor.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f648b0bf9a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The coupled baker-dep trail missed a direct crate, and squaring 1e-200 underflows the operator statistic without changing MAX_ERR_LSB. Co-authored-by: Cursor <cursoragent@cursor.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 614e38e185
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Shifting both BigInt limbs by max(bits)-1023 zeros 1/2^1023, so RMS of MIN_POSITIVE/2 printed 0. Cap the shift by the smaller limb so that subnormal ratio stays representable. This is not MAX_ERR_LSB. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 532a95791c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
1/2^1074 cannot be a finite f64 numerator over a 2^1074 denominator, so to_f64 returned infinity and a valid bake became NonFiniteDeviation. Keep the smallest subnormal as the operator statistic, not the bound. Co-authored-by: Steven Giacomelli <swgiacomelli@users.noreply.github.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 07c0575922
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| _ if n_shift > d_shift => return f64::INFINITY, | ||
| _ => return 0.0, | ||
| }; | ||
| let x = (nf / df) * 2.0_f64.powi(exp); |
There was a problem hiding this comment.
When both ratio limbs require truncation and their shift difference is -1075, powi(exp) underflows to zero before the normalized nf / df factor can raise the result into the representable subnormal range. Fresh evidence beyond the earlier scale-1 fixture is a zero-quantizing 2×2 table whose node and off-knot values are all f64::from_bits(0x0c70_0000_0000_0001) with scale f64::from_bits(0x3040_0000_0000_0001): every exact residual is slightly above half of 2^-1074, so its f64 value and RMS are the smallest subnormal, but this expression returns zero and publishes rms_lsb = 0. Convert the normalized significand and exponent together so intermediate exponentiation cannot underflow.
Useful? React with 👍 / 👎.
2 participants
Closes #40.
Stacked S4 on the merged S3 ingest (
epic/36-surfaces-bake). S4 applies the stored scale and measures deviation; it does not invent a fitter or a report-only baker.The baker now:
(x_knot, y_knot)from samples whose X and Y equal that knot (f64equality withf64::fromof theu16). Missing nodes and two samples claiming the same node with different values are closedBakeErrors. Off-knot samples do not fill a node. Knot lookup is binary search on the strictly increasing lists.round(value * scale)toi32, nearest, exact half-way away from zero. The host helper splits trunc and fraction so thef64immediately below0.5stays0(adding0.5then flooring would emit1). Overflow outsidei32is a closedBakeError. Zero / non-invertible scale is rejected at quantize. S3's non-finite sample/scale gate is unchanged.MAX_ERR_LSBisceilof an exact rational residual incrates/surfaces-bake/src/bound.rs(IEEEf64bit-patterns as dyadics; bilinear as an exact ratio of thei32grid on the host vianum-rationalBigRational).ceilapplies only to the finished residual. Exactu16coordinates also include the runtime-rounded integer path (evaluate_u16). Hostf64lerp is not the bound. A finite residual whose ceil does not fit ini32isBakeError::BoundOverflow;NonFiniteDeviationis for a true non-finite residual. Exact-zero tables stay 0. Reviewed host crates arenum-bigint,num-rational, andnum-traits; they stay off the runtime graph. RMS is an operator statistic:Ratio::to_f64uses a 53-bit window plus a binary exponent sof64::from_bits(1)(2^-1074) stays finite, notNonFiniteDeviation.NX * NYaboveMAX_GRID_CELLS(1_048_576) with closedBakeError::GridTooLargebefore allocating the cell matrix.pub const MAX_ERR_LSB: i32 = …;viaemit_max_err_lsb. Full--emit-rustis not in this PR.Runtime README / crate-doc status points at baker quantization and
MAX_ERR_LSB. Baker types stay out of the firmware rustdoc.Wording is deviation between the supplied samples and the table built from them. No device, vendor, sensor, calibration, accuracy, timing, flash, or WCET claim.
Bound proof
A test builds
ph_surfaces::BilinearSurfacefrom the quantized i32 grid. Exact-u16samples useevaluate.[0, 2], grid[[0, 1], [1, 2]], sample(1, 1, 1): host reconstruct1, runtimeevaluate2. Bound is1.[0, 3], grid[[4, 13], [75, 94]], sample(1, 1, 32.77777777777778). Exact bilinear is286/9; threef64lerps make the residual< 1; exact residual emits bound2.[0, 1], grid[[1, 2], [3, 7]], sample(0.1, 0.1)stays finite.1. Tiny sample(0.5, 0.5, 1e-300)against a unit table also emits bound1.[0, 1], rows[0, -1], extra sample(1e-300, 0.5, 1). Exact reconstruction is-1e-300; residual is slightly greater than1; bound is2.[0, 1], both rows[1, 2], extra sample(1e-300, 0.5, 2). Exact reconstruction is1 + 1e-300; residual ceiling is1.3_000_000_000isBakeError::BoundOverflow, notNonFiniteDeviation. The same overflow with a same-sign tiny on rows[0, -1]is alsoBoundOverflow.n, notn+1. Exact-zero tables still emit0.1e-200is1e-200 / sqrt(5), not0. RMS off64::MIN_POSITIVE / 2is that value oversqrt(5), not0. A 2×2 off64::from_bits(1)plus an off-knot of the same value quantizes to zeros, bound1, RMS2^-1074, notNonFiniteDeviation.Not in this PR
--emit-rusttable emission,cargo xtask generate, goldens,docs/generating-tables.md(S5–S7 / [S5] Emit static tables and gate checked-in generated source against drift #41–[S7] Document the host/target split and the boundary rules #43)[dependencies]on the firmware crate, clap,genfeatureVerification
cargo xtask cion07c0575(full profile, not release): all runnable checks PASS; baker line budget 1409 / 1600. SKIP: secret scan (gitleaks not installed). A skip is not a pass. This is not release evidence.PR base is
epic/36-surfaces-bake(S3 ingest is merged there), notmain.