Stage 2a: split-radix C++20 port beside the C, bit-identical, nothing routed

[tap]

What this changes

Stage 2a of docs/audit-fft-and-code-smells.md (Part 3; spec in Part 4, provenance in Part 5). Adds tap::dsp::detail::split_radix_rdft<Sample> (include/tap/dsp/fft/split_radix.h), a statement-for-statement C++20 transliteration of the rdft path of Ooura's fftsg.c, beside the vendored C. Nothing is routed: fft.h is untouched, basic_real_fft still runs the C. The parity gate's engine_under_test alias is re-pointed at the port (the one place its header comment names), the bench builds every float scenario and the float size probe a second time against the port under a _port suffix and reports port/C without gating it, bench/README.md records the mechanism, and docs/fft-design.md records the measured numbers below from this PR's own CI logs (run and job ids on every number).

Eleven commits: the six of the original series (tables, bit reversal, complex kernel, real post-pass + transforms + gate re-point, bench, docs), a docs label correction (a9b53b4), and four review-fix commits after the two hostile reviews (d137a2e NOTICE.md provenance post-2a; 37ca717 bench: an informational port binary never reds the ratchet, size step pipefail and explicit failure on a missing .text, _port size probe under the same engine guard as icount, --merge drops nested _port keys; cb92d50 noise-floor TODO re-labelled to Stage 2b; d8eccf6 rule-2 sentence naming the intentional -Wdouble-promotion sites). Each of the first three compiles standalone; the class is complete in the fourth, which is where the parity target starts linking against it.

Why

The port is the long pole of the FFT plan: Stage 2b's routing flip (one line in fft.h), the MuTap bump with every float pin unchanged, and Stage 2c's removal of the C all rest on the port being bit-identical to the C for both precisions (Decision D10). This PR is where that identity is established and gated, with the C still in the tree to compare against.

How identity was achieved and how it is kept, stated in the header so nobody undoes it:

1. Statement fidelity. Every Ooura arithmetic statement is textually intact (no cmul helpers, lambdas, std::complex or reassociation). Verified mechanically as well as by the gate: a normalizer that strips whitespace, maps Sample→double and undoes only the documented differences finds all 26 mechanically ported kernels token-identical to the C (DIFFERENCES: 0).
2. Table semantics per precision. The C's float build is #define double float, so makewt/makect's locals and the products fed to the trig calls are float arithmetic while cos/sin/atan stay the double libm functions with promoted arguments and the results converted on assignment (and 0.5 / cos(...), 0.5 / wk1r, 0.5 * cos(...) are double arithmetic converted on assignment). The port keeps the locals typed Sample, gives every libm call an explicit static_cast<double> argument and converts to Sample exactly where the C's assignment did. Breaking this on one line (std::cos of the float product, i.e. cosf) turns the three float gates red while the three double gates stay green.
3. Tables built once in the constructor with rdft's first-call protocol; the lazy ip[0] = 0 init is gone, the transforms are const noexcept, and m_w.data() / m_ip.data() are loaded into locals once per transform so the helpers receive a, ip, w by parameter as the C functions do.
4. Index types stay the C's int (Part 4 permits widening; not widening keeps the statements identical); the single size_t→int narrowing is the constructor's.

Not ported: cdft, the DCT/DST family, the thread scaffolding.

Verification

The gate (-ffp-contract=off both sides, memcmp identity)

Green on the first CI run of the PR on every leg: linux (g++ 13.3.0), windows (MSVC 19.51.36256.0), macos (AppleClang 21.0.0, arm64), and at N ≤ 4096 against newlib's libm on cortex-m4-softfp, cortex-m4f, cortex-m33, cortex-m55 (arm-none-eabi-gcc 13.2.1, QEMU 8.2.2). Locally also clang++ 18.1.3. Every N from 4 to 65536 plus 2^20 (hosts), double and float, forward and inverse, five materials.

Informational max-ulp at default flags (CI run 35295206578 at 8afe6cf, 2026-09-18)

Platform (job)	Toolchain	Sizes	max ulp double	max ulp float
linux-ooura (105446230919)	ubuntu-24.04, GNU 13.3.0, x86-64 no -march	4 … 65536, 2^20	0	0
windows-ooura (105446230835)	windows-2025-vs2026, MSVC 19.51.36256.0	4 … 65536, 2^20	0	0
macos-vdsp (105446230934)	macos-26-arm64, AppleClang 21.0.0.21000101 (FMA ISA)	4 … 65536, 2^20	0	0
cortex-m4-softfp (105446230884)	arm-none-eabi-gcc 13.2.1 (15:13.2.rel1-2), QEMU 8.2.2, soft-float	4 … 4096	0	0
cortex-m4f (105446230771)	same, fpv4-sp-d16 (VFMA)	4 … 4096	0	0
cortex-m33 (105446230920)	same, single-precision FPU (VFMA)	4 … 4096	0	0
cortex-m55 (105446230931)	same, MVE (VFMA)	4 … 4096	0	0
local	g++ 13.3.0 and clang++ 18.1.3, x86-64 no -march	4 … 65536, 2^20	0	0

Zero everywhere because the statements are identical, so each compiler fuses both sides alike. Not measured and not claimed: x86-64 with -march (FMA), where the objdump probe shows g++ fusing the two sides differently after inlining. The gate stays at -ffp-contract=off.

objdump FMA probe (-O3 -march=haswell; reference C = fftsg.c + fftsg_float.c; port = a TU instantiating both precisions)

compiler	default: C fused	default: port fused	-ffp-contract=off: C	-ffp-contract=off: port
gcc / g++ 13.3.0	372	588	0	0
clang / clang++ 18.1.3	333	221	0	0

Bench, informational (bench run 35295262684 at 8afe6cf, arm-none-eabi-gcc 13.2.1, QEMU 8.2.2; nothing gated at the port, no seeding needed, all five icount jobs green with the C at +0.00 %)

Scenario	m4-softfp	m4f	m33	m55 (C = CMSIS)	m55-ooura
rfft_f32_512 (C, seeded baseline)	1,868,441,244	98,090,666	102,248,169	52,382,331	94,561,954
rfft_f32_512_port	1,864,904,581	97,126,274	100,833,265	89,407,453	89,047,005
ratio port / C	0.9981	0.9902	0.9862	1.7068 (vs CMSIS)	0.9417
output checksums	identical	identical	identical	differ (CMSIS ≠ Ooura)	identical
rfft_f32_2048 (C, seeded baseline)	2,296,984,479	111,859,257	116,385,409	54,858,120	107,806,480
rfft_f32_2048_port	2,294,343,851	111,270,714	115,396,527	102,870,129	102,644,849
ratio port / C	0.9989	0.9947	0.9915	1.8752 (vs CMSIS)	0.9521
output checksums	identical	identical	identical	differ (CMSIS ≠ Ooura)	identical

.text of the MinSizeRel float size probe (size -A), bytes:

key	C	port	port / C
m4-softfp	51,729	53,289	1.0302
m4f	43,153	44,601	1.0336
m33	42,601	44,009	1.0331
m55 (C = CMSIS)	107,681	39,281	0.3648
m55-ooura	38,505	39,281	1.0202

Read against the ±3 % ratchet 2b applies: the port executes 0.1 – 5.8 % fewer instructions than the C on every Ooura key and is bit-identical to it at the bench's default flags (identical checksums, including on the VFMA cores); m55-ooura improves beyond 3 %, so 2b re-records that key rather than absorbing it. Against CMSIS-DSP Helium on the deployed m55 profile the port is 1.7 – 1.9× the count; m55 stays on CMSIS.

Local

g++ 13.3.0 and clang++ 18.1.3, Release, -DTAP_DSP_WERROR=ON -DTAP_DSP_BUILD_CAPI=ON -DTAP_DSP_BUILD_BENCH=ON: full builds warning-free, ctest 283/283 each (incl. the new _port smoke tests; host checksums C vs port identical: 0x662dd085b5b88325 at N = 512, 0x806d1c5dbb65c525 at N = 2048). Mutation check as described above (3 float gates red / 3 double green, restored 7/7). scripts/icount.py's informational path dry-run with a stubbed QEMU measure through compare, no-baseline, --update, --merge and ignored-baseline cases. scripts/tidy.sh full sweep (clang-tidy-18) clean; pre-commit run --all-files clean for this PR's files (it reformats the pre-existing tests/reference/frontend_vectors.h, not touched here).

Notes for the reviewer

• Documented exception. The port header's banner is the four house lines plus a // prose block carrying Ooura's notice verbatim and the derivative statement, exactly the text docs/fft-design.md and NOTICE.md settled in wave 1; SPDX LicenseRef-Ooura AND MIT.
• No contract change. No documented contract point moves; the port is bit-identical to what basic_real_fft runs today and nothing is routed at it. No consumer pin moves.
• Files deliberately not touched: fft.h (2b flips routing), README.md, tests/CMakeLists.txt (the parity targets already existed), ci.yml, third_party/**, the taphouse style files. scripts/icount.py is touched, per the bench brief's "documented informational prefix in icount.py" option: the mechanism lives where the QEMU invocation already is, so the ratio prints next to the counts.
• The m55 key's C sibling is CMSIS-DSP, so its printed port/C ratio is port-vs-CMSIS; the port-vs-Ooura number on the M55 is the m55-ooura key's.
• For the 2b flip agent: basic_real_fft can hold a split_radix_rdft<Sample> by value; its transforms are const, so forward_inplace/inverse_inplace on the wrapper can become const too. The engine's constructor allocates (two vectors) and asserts the geometry the same way fft.h does. Expect icount.py to flag m55-ooura as IMPROVED beyond tolerance at the flip (re-record, as the policy says).

🤖 Generated with Claude Code

https://claude.ai/code/session_019ZPTzNxo5Fe4EtpXXKf7Sy

[tap]

Hostile review A — numerics and transliteration fidelity (head 9df52d2)

Reviewed in a detached worktree of the branch; every claim below was checked against source, a local build, a local run, or the cited CI job log. Toolchains: g++ 13.3.0, clang++ 18.1.3, clang-tidy-18, clang-format 18.1.3 (the pinned pre-commit rev), python3.

Findings

Blockers: none. Should-fix: none. Three nits, all documentation/consistency; none touches the transform, the tables, the gate or the ratchet.

1. nit — docs/fft-design.md:172. "the three FMA-capable legs (macOS arm64, M4F, M33, M55)" names four legs; the table above it has four FMA rows. Fix: "four". (Line comment.)

2. nit — bench/README.md:114 vs bench/icount/CMakeLists.txt:61 vs bench/CMakeLists.txt:49. The README says the _port binaries are always built "whatever the variable says"; the icount block skips them when TAP_DSP_BENCH_ENGINE=split_radix, while the size-probe list adds rfft_f32_512_port unconditionally (two identical port probes in that configuration, and no port icount pair). Verified by configuring the host bench with -DTAP_DSP_BENCH_ENGINE=split_radix and listing the built targets. Fix: make the two CMake lists agree (the icount guard is the sensible one; apply it to the probe) and qualify the README sentence. CI never sets the variable, so no measured number is affected. (Line comment.)

3. nit — scripts/icount.py:31 (and the same phrase in bench/README.md and the bench.yml header). "it cannot fail the run" holds for the verdict but not for the process: a _port binary that hangs, faults or prints ok=0 hits the raise SystemExit in measure() (lines 86/90) before the gated scenarios are judged, so the informational binary can still turn the job red. Probably the desired behaviour; the wording should say so, or measure() should catch it for informational binaries and print did not complete. (Line comment.)

Checked and found correct (the fix pass need not re-verify)

Transliteration fidelity (attack 1). Wrote an independent normalizer (strip comments/whitespace, drop the C's local prototypes and USE_CDFT_THREADS blocks, map Sample→double, const int*/const Sample*→bare pointer, drop static/noexcept, then tokenize) and compared all 28 routines against fftsg.c. Result: 24 routines token-identical (makeipt, bitrv2, bitrv2conj, bitrv216, bitrv216neg, bitrv208, bitrv208neg, cftfsub, cftbsub, cftf1st, cftb1st, cftrec4, cfttree, cftleaf, cftmdl1, cftmdl2, cftfx41, cftf161, cftf162, cftf081, cftf082, cftf040, cftb040, cftx020). rftfsub/rftbsub differ by exactly the one documented cast (static_cast<Sample>(0.5 - c[nc - kk]), the C's assignment narrowing); makewt/makect differ only by inserted static_cast<double>(/std::/) tokens — no operand reordered, no regrouping, no type of a local changed. The rdft glue (xi = a[0] - a[1]; a[0] += a[1]; a[1] = xi;, a[1] = 0.5 * (a[0] - a[1]); a[0] -= a[1];, the n > 4 / n == 4 conditions, the nw/nc first-call protocol with ip[0] = ip[1] = 0 on entry) matches rdft statement for statement; m_nw/m_nc equal what the C re-reads from ip[0]/ip[1] on every later call, including the N = 4 case where makect is never called and nc stays 1, and N = 8 where makewt(2) writes nothing to w and makect(2) writes c[0], c[1]. Branch coverage of the sweep: cftfsub's five size branches (n > 512, > 128, ≤ 128, == 32, < 32, == 8, == 4), bitrv2's l == 8 / else (alternating powers of two), makewt's nwh == 4 / nwh > 4, makeipt's l > 32 loop (from N = 256 up) and the deep cftrec4/cfttree path (2^20) are all reached by 4 … 65536 + 2^20.

Table semantics per precision (attack 2). Every libm call in makewt/makect/the constructor takes a double argument via explicit cast (std::atan(1.0), std::cos(static_cast<double>(...)), std::sin(...), std::sqrt(static_cast<double>(n) / 2.0)); the float products delta * static_cast<Sample>(nwh), delta * 2, delta * 6, 3 * delta * static_cast<Sample>(j) are Sample arithmetic with the C's left-to-right grouping; 0.5 / cos(...), 0.5 / wk1r, 0.5 / wk3r, 0.5 * c[0], 0.5 * cos(...), 0.5 * sin(...), 0.5 * (a[0] - a[1]), 0.5 - c[nc - kk] are double arithmetic converted on assignment, exactly where the C converts. No std:: overload resolves to a float function. Mutation reproduced: in a scratch copy, w[j] = static_cast<Sample>(std::cos(delta * static_cast<Sample>(j))) (i.e. cosf) turns exactly the three float gates red (first difference at N = 2048, 128 ulp fwd / 96 ulp inv; 2^20: 131072 / 262144 ulp) and leaves the three double gates green.

Constness, aliasing, UB (attack 3). Every transform helper takes const int* ip / const Sample* w, so nothing can write a table after construction (compiler-enforced); makewt/makect are reached only from the constructor. Built the gate under ASan + UBSan (-fno-sanitize-recover=all, Debug so the constructor assert is live): 6/6 pass over 4 … 65536 + 2^20, both precisions, both directions — no out-of-bounds read at N = 4 / 8 (ip = 4 ints, w = 2 / 4 samples), no signed overflow or shift UB in bitrv2/makeipt at 2^20 (largest ip write is index 511 of 727). -Wall -Wextra -Wpedantic -Wconversion -Wshadow -Werror full builds (capi + bench on) warning-free on both g++ 13.3 and clang++ 18.1; ctest 283/283 on each.

Parity gate integrity (attack 4). expect_identical_all_sizes runs every size × every material for the given direction; the four gates cover {double, float} × {forward, inverse}; the 2^20 test covers both directions per material; the inverse is fed the reference forward's output on both sides. sweep_sizes() clamps at TAP_DSP_PARITY_MAX_N with a static_assert that it is a power of two ≥ 4 (no vacuous pass); the 2^20 test is compiled out, not skipped, below the cap (bare-metal main counts SKIP as a failure). tap_dsp_fft_parity gets -ffp-contract=off on its own TU (which is where the header-only port is compiled) and links tap_dsp_fft_reference_nocontract; tap_dsp_fft_parity_default_flags links tap_dsp_fft_reference_default with no flag on either side. Neither parity target links tap::dsp, so no backend define reaches them.

Bench and ratchet (attack 5). split_radix_bench_adapter::forward/inverse/copy are line-for-line basic_real_fft::forward/inverse/copy (if (input != output) copy loop; const Sample scale = Sample(2) / static_cast<Sample>(m_size); output[i] *= scale;); member init order matches declaration order. Host checksums C vs port identical on both compilers: 0x662dd085b5b88325 (512), 0x806d1c5dbb65c525 (2048). icount.py driven with a stubbed measure through compare / port-key-in-baseline / no-baseline / --update --record / --merge: _port never enters failures, never reaches baselines[target], is filed under top-level informational in the record file, and --merge skips that key — so the seeding path on a push to main cannot pick it up. The m55 sibling is CMSIS (documented three times: bench.yml, README, design note), and the port-vs-Ooura M55 figure is taken from m55-ooura.

Docs and numbers (attack 6). Pulled all twelve job logs. Every number in the PR body and docs/fft-design.md matches: the ten _port counts, the ten C counts (+0.00 % vs baselines.json), the ten port/C ratios, the five checksum verdicts (identical × 4 keys, DIFFER on m55), the ten .text figures and five size ratios, and the max over all N and materials = 0 ulp lines for double and float on all seven CI legs with the stated compiler ids (GNU 13.3.0, MSVC 19.51.36256.0, AppleClang 21.0.0.21000101, GNU 13.2.1 × 4). The derived figures (+2.0 … +3.4 % .text, 0.1 … 5.8 % fewer instructions, 0.4 % m55-vs-m55-ooura port delta, 1.7 – 1.9× vs CMSIS) recompute correctly. objdump FMA probe reproduced exactly (-O3 -march=haswell, vfn?m(add|sub)): gcc 372 / 588, clang 333 / 221, both 0 / 0 at -ffp-contract=off. The header banner is the design note's block verbatim plus the one modifications line, SPDX LicenseRef-Ooura AND MIT, and is consistent with NOTICE.md, LICENSES/LicenseRef-Ooura.txt and the Part 5 / Part 13 corrections (derivative-work statement, modification grant, readme path fixed by D6). The final commit 9df52d2 touches only docs/fft-design.md, so the 8afe6cf attribution of the measurements is right.

Style gates (attack 7). clang-format-18 --dry-run --Werror and pre-commit run --files clean on split_radix.h, test_fft_parity_ooura.cpp, bench_common.h; scripts/tidy.sh tests/test_fft_parity_ooura.cpp (which pulls the header through HeaderFilterRegex) clean, also with --warnings-as-errors=readability-*. (Note, pre-existing and not this PR's: the tidy compile database is configured without TAP_DSP_BUILD_BENCH, so bench/** is never tidied in CI or by tidy.sh.)

---

Generated by Claude Code

[tap]

Hostile review B (process, build, CI, downstream) of head a9b53b4. Numerics and transliteration fidelity are the other reviewer's. Every finding below was reproduced on a detached worktree of the branch or read from the named CI log; nothing is from memory. No blockers.

Findings

1. should-fix — NOTICE.md (canonical provenance statement) still calls the port "planned". NOTICE.md:39 "What DspTap relies on going forward: the planned C++20 port of rdft (include/tap/dsp/fft/split_radix.h …)" and :31 "What DspTap ships today: one source file of the package, fftsg.c, plus its readme.txt". On merge DspTap ships the derivative header, so the file that README.md and docs/fft-design.md both defer to ("This file is the canonical statement") is wrong on the one fact this PR changes. docs/fft-design.md's provenance section was updated in this PR to say the port landed; NOTICE.md was not, and it is not a Part 12 contention file. Fix: two sentences ("landed at Stage 2a, #28, beside the C and routed nowhere until 2b; the decision on asking the author is still recorded here at 2c"). README.md:436-440 ("planned C++20 port", "Nothing from either has shipped yet") and :449 say the same thing; README is 2b's file, so that is a hand-off to the 2b README rewrite, not a change here.

2. should-fix — .github/workflows/bench.yml:168-169: the size step pipes size through tee under bash -e {0} with no pipefail. Inline comment. A failing size on the port probe yields ratio port/C = 0.0000 and a passing step; the C line has the same pre-existing hole. The Count step in the same file sets set -o pipefail; Part 13 says every tee step must.

3. nit — -Wdouble-promotion. Six lines in the float instantiation (split_radix.h:187, 245, 246, 290, 2531, 2554) fail under -Wdouble-promotion -Werror on g++ 13.3.0 and clang++ 18.1.3. Not in any Tap or MuTap flag set today; inline comment explains why it is still worth a thought for M4 soft-float consumers, and why it is the numerics reviewer's call.

4. nit — bench/CMakeLists.txt:49 has no NOT TAP_DSP_BENCH_ENGINE STREQUAL "split_radix" guard on the _port size probe while bench/icount/CMakeLists.txt:61 has one; with the non-default engine the port is compared against itself. Inline.

5. nit — bench.yml:140 comment "Release (-O2)" is contradicted by this PR's own last commit (a9b53b4: Release is CMake's GNU default -O3 -DNDEBUG, no toolchain file overrides it). The commit message says a workflow file was out of that commit's scope, but the PR touches bench.yml, so the one-word fix belongs here rather than in a later PR.

6. nit — docs/fft-design.md:48 still carries TODO(stage 2a) / "pinned at Stage 2a" for the noise-floor row; this is the Stage 2a PR and it does not do it. Inline; re-label to the owner.

7. nit — PR body says "Six commits"; there are seven (a9b53b4, the docs label fix, was added after the body was written). The commit list in the body is otherwise accurate.

8. note — scripts/icount.py:108 --merge filters only the top-level informational key, not a _port key nested under a real target; no pipeline path produces one. Inline, no change requested.

Verified correct

• Scope. Nine files; none of fft.h, README.md, CLAUDE.md, ci.yml, tests/CMakeLists.txt, test_fft.cpp, test_fft_oracle.cpp, test_fft_rt.cpp, third_party/**, or the style configs. The only edit to tests/test_fft_parity_ooura.cpp is the engine_under_test alias plus its comment; the fft.h include there is still load-bearing (it supplies the rdft/rdft_f prototypes ooura_ref calls). Nothing routed, nothing of 2b/2c.
• Header-only, standalone. A TU including only split_radix.h, and one instantiating both precisions, compile clean under -std=c++20 -Wall -Wextra -Wconversion -Wsign-conversion -Wshadow -pedantic -Werror on g++ 13.3.0 and clang++ 18.1.3, also with -fno-exceptions -fno-rtti -DNDEBUG. Five standard includes, no macro defined, no statics or globals (every helper is a private static member), clang-tidy-18 with the repo .clang-tidy clean over the header, clang-format --dry-run --Werror clean on the three C++ files, each of the first three commits' header compiles on its own as the body claims. The constructor's assert is the same idiom as fft.h:340; Stage 6 owns TAP_EXPECTS.
• Compile time. Both precisions at -O2: g++ 1.72 s, clang++ 1.10 s (this host; fftsg.c as C++ 0.94 / 0.73 s, the two C files as C 1.67 s gcc). Same order as Part 4's 1.3 s; D8 stands.
• Test registration. tap_dsp_fft_parity links tap_dsp_fft_reference_nocontract + tap_dsp_warnings only (tests/CMakeLists.txt:221-223), not tap::dsp; gtest names unchanged, so ci.yml's -L parity -R default_flags -V step still matches (ran on linux, job 105446230919). QEMU legs at 9df52d2 (run 35295707940): all four report TAP_DSP_TESTS_COMPLETE rc=0 with selected=223/4/1 skipped=0; the parity binaries ran whole at N ≤ 4096 (the two 2^20 tests are compiled out under the cap, as documented); M55 tap_dsp_tests .text 1,077,672 + heap 2,087,896 fits AN547.
• CMake/CI. bench/icount second loop goes through the same function, so the smoke test stays host-only and the _port binaries build on every leg (the QEMU jobs need them); the _port size probe builds on Generic because bench.yml builds it. TAP_DSP_BENCH_ENGINE default unchanged and only visible under TAP_DSP_BUILD_BENCH.
• icount.py. Exercised compare / no-baseline / --update / --record / --merge with a stubbed QEMU: _port never affects the exit code, a baseline naming a _port key prints IGNORED and --update drops it, --record files it under informational, --merge skips that key; the seed-only-on-main logic is untouched and baselines-<key> artifacts come from --update, which never writes _port. The informational lines appear in the job log (m33 105446394276, m4-softfp 105446394297, m55 105446393983) and are appended to the step summary via tee → icount.txt; the size line is echoed and appended too.
• Provenance. The header's notice block is token-identical to LICENSES/LicenseRef-Ooura.txt and to third_party/ooura/readme.txt:141-145; NOTICE.md's quoted grant matches; the derivative-work sentence is present; no "unmodified", "public domain" or "permissive" wording was added; docs/fft-design.md's banner block matches the header except the documented extra "Modifications (…) September 2026" sentence.
• Numbers traced to logs. m33 counts 100,833,265 / 115,396,527 and .text 42,601 / 44,009 (105446394276); m4-softfp 1,864,904,581 / 2,294,343,851 (105446394297); m55 .text 107,681 / 39,281 (105446393983); ulp 0 / 0 on linux (105446230919), macos (105446230934), cortex-m55 (105446230931); runner images ubuntu-24.04 / macos-26-arm64 / windows-2025-vs2026, compilers GNU 13.3.0 / AppleClang 21.0.0.21000101 / MSVC 19.51.36256.0 / arm-none-eabi-gcc 13.2.1 (15:13.2.rel1-2), QEMU 8.2.2 (1:8.2.2+ds-0ubuntu1.18) — all as the design note states. bench/README.md's tables agree with bench/baselines.json (unchanged) and the new _port semantics. The .text deltas (+776 … +1,560 B, +2.0 … +3.4 %) and the "0.1 – 5.8 % fewer instructions" arithmetic check out.
• Downstream (MuTap). Scratch copies of MuTap configured with -DMUTAP_WERROR=ON against DspTap origin/main and against this tree: CMakeCache.txt, target list and configure output identical; mutap_tests builds warning-free with -Werror against this tree (its suite was 123/197 green and still running at post time; nothing in it can see the port). Nothing in MuTap includes the new header; MuTap's CI gcc -c submodules/dsptap/third_party/ooura/fftsg*.c path is untouched; no new cache variable, option, CMake minimum or default. The "no pin bump needed" claim holds.
• Hygiene. All seven commits carry both trailers; PR footer present; no other model identifiers in the diff; PR template sections filled; each commit message matches its diff. Local: full build -DTAP_DSP_WERROR=ON -DTAP_DSP_BUILD_BENCH=ON -DTAP_DSP_BUILD_CAPI=ON, ctest 283/283. Head a9b53b4: all 24 check runs green.

---

Generated by Claude Code

[tap]

Fix pass for reviews A (#pullrequestreview-5243359055) and B (#pullrequestreview-5243370615)

Head a9b53b4 → d8eccf6, four commits, seven files; the port header change is comment-only (git diff --stat: +7 lines in the banner; parity label 7/7 re-run on g++ 13.3.0 and clang++ 18.1.3). Every finding was re-verified against source before it was changed.

finding	fix	commit
B1 should-fix — NOTICE.md calls the port "planned", describes the C as everything shipped	provenance bullet states the post-2a truth: the port exists at include/tap/dsp/fft/split_radix.h beside the vendored C, bit-identical under the parity gate, routed nowhere until 2b, the C leaves at 2c; "SPDX plan" → "SPDX". No licensing fact moved (grant quotation, derivative-work and modification-grant sentences, readme.txt stays, LicenseRef-Ooura AND MIT all as wave 1 settled them)	d137a2e
B2 should-fix — size step pipes size through tee without pipefail	set -eo pipefail at the top of the step; each of $text / $text_port checked to be a non-empty digit string, else ::error::<key>: C probe .text unavailable … / … port .text unavailable … and exit 1. Both holes (C line included) closed; no more 0.0000 ratio or skipped ceiling comparison	37ca717
A2 / B4 nit — _port size probe unguarded while the icount pairs are guarded; README says "whatever the variable says"	bench/CMakeLists.txt applies the icount guard (NOT TAP_DSP_BENCH_ENGINE STREQUAL "split_radix"); README states the actual rule. Verified by configuring both ways: default builds both _port icount pairs and the _port probe; split_radix builds none of the three	37ca717
A3 nit — "cannot fail the run" overstated; B note 8 — --merge only filters the top-level key	catch-and-print: measure() raises MeasurementError; a gated binary still aborts the run, a _port binary that times out / faults / prints ok=0 prints <key>: informational binary failed: <reason> in the informational block and the gated verdict runs. --merge drops _port keys nested under a real target with a DROPPED line. Docstring, bench/README.md and the bench.yml header say exactly what holds	37ca717
B5 nit — bench.yml comment "Release (-O2)"	"CMake's Release default for GCC, -O3 -DNDEBUG; no toolchain file overrides it"	37ca717
B6 nit — docs/fft-design.md noise-floor row TODO(stage 2a)	both cells and the rationale paragraph re-labelled TODO(stage 2b) with the reason (2a's scope was the port and its gate; test_fft.cpp untouched; 2b re-points the battery at the port, so a number pinned there is measured on what ships; 3b's widening is over the fixed-point profiles)	cb92d50
B7 nit — PR body "Six commits"	body now says eleven and names the five added since	(PR body)
B3 nit — -Wdouble-promotion on lines 187, 245, 246, 290, 2531, 2554	documented, not cast: reproduced on both compilers (exactly those six, nothing else); rule 2 in the banner now names the sites, states they are the rule-2 contract (the C's float build promoted in the same places), that no cast is added (rule 1 keeps the statements textually intact; both reviews' normalizers and the cosf mutation are evidence attached to the current text), and that no Tap flag set enables the warning	d8eccf6
A1 nit — "three FMA-capable legs"	already fixed at a9b53b4 (review A was written against 9df52d2)	—

Hand-off to Stage 2b (not changed here)

README.md:436-440 ("planned C++20 port", "Nothing from either has shipped yet") and :449 say what NOTICE.md used to say. README is the Stage 2b README rewrite's file (audit Part 3) and another agent's this wave, so it is left as is; the 2b agent should bring those sentences to the same post-2a wording NOTICE.md now carries (port landed at 2a, bit-identical, routed at 2b).

Local verification of d8eccf6

• g++ 13.3.0 and clang++ 18.1.3, Release, -DTAP_DSP_WERROR=ON -DTAP_DSP_BUILD_CAPI=ON -DTAP_DSP_BUILD_BENCH=ON: zero warnings, ctest 283/283 each; ctest -L parity 7/7 each.
• -DTAP_DSP_BENCH_ENGINE=split_radix vs default: target lists as in the A2/B4 row.
• python3 -m py_compile scripts/icount.py; pre-commit run (clang-format 18.1.3) clean; scripts/tidy.sh tests/test_fft_parity_ooura.cpp (clang-tidy-18, pulls the header) clean.

scripts/icount.py driven with a stubbed measure()

Fake build dir with the four tap_dsp_icount_rfft_f32_{512,2048}{,_port} binaries, counts = the real m33 numbers, measure() replaced by a stub that can raise MeasurementError per key. <dir> = the scratch directory.

$ icount.py --target m33 --build-dir <dir>/build-m33 --plugin <plugin>   # compare
rfft_f32_2048: 116385409 insns vs baseline 116385409 (+0.00%) ok
rfft_f32_512: 102248169 insns vs baseline 102248169 (+0.00%) ok
--- informational: '_port' scenarios (counted, never gated, never baselined; Stage 2a until 2c) ---
rfft_f32_2048_port: 115396527 insns [engine=split_radix backend=split_radix]; rfft_f32_2048: 116385409 insns [engine=reference_c backend=ooura]; ratio rfft_f32_2048_port/rfft_f32_2048 = 0.9915; output checksums identical
rfft_f32_512_port: 100833265 insns [engine=split_radix backend=split_radix]; rfft_f32_512: 102248169 insns [engine=reference_c backend=ooura]; ratio rfft_f32_512_port/rfft_f32_512 = 0.9862; output checksums identical
-> exit 0
   baselines file now: {"m33": {"rfft_f32_2048": 116385409, "rfft_f32_512": 102248169}}

$ icount.py --target m33 --build-dir <dir>/build-m33 --plugin <plugin>   # no-baseline
rfft_f32_2048: 116385409 insns (NO BASELINE — commit this value)
rfft_f32_512: 102248169 insns (NO BASELINE — commit this value)
--- informational: '_port' scenarios (counted, never gated, never baselined; Stage 2a until 2c) ---
rfft_f32_2048_port: 115396527 insns [engine=split_radix backend=split_radix]; rfft_f32_2048: 116385409 insns [engine=reference_c backend=ooura]; ratio rfft_f32_2048_port/rfft_f32_2048 = 0.9915; output checksums identical
rfft_f32_512_port: 100833265 insns [engine=split_radix backend=split_radix]; rfft_f32_512: 102248169 insns [engine=reference_c backend=ooura]; ratio rfft_f32_512_port/rfft_f32_512 = 0.9862; output checksums identical
-> exit 1
   baselines file now: {"m33": {}}

$ icount.py --target m33 --build-dir <dir>/build-m33 --plugin <plugin>   # port-key-in-baseline
rfft_f32_2048: 116385409 insns vs baseline 116385409 (+0.00%) ok
rfft_f32_512: 102248169 insns vs baseline 102248169 (+0.00%) ok
rfft_f32_512_port: baseline 100833265 IGNORED (informational scenario; remove it from bench/baselines.json)
--- informational: '_port' scenarios (counted, never gated, never baselined; Stage 2a until 2c) ---
rfft_f32_2048_port: 115396527 insns [engine=split_radix backend=split_radix]; rfft_f32_2048: 116385409 insns [engine=reference_c backend=ooura]; ratio rfft_f32_2048_port/rfft_f32_2048 = 0.9915; output checksums identical
rfft_f32_512_port: 100833265 insns [engine=split_radix backend=split_radix]; rfft_f32_512: 102248169 insns [engine=reference_c backend=ooura]; ratio rfft_f32_512_port/rfft_f32_512 = 0.9862; output checksums identical
-> exit 0
   baselines file now: {"m33": {"rfft_f32_2048": 116385409, "rfft_f32_512": 102248169, "rfft_f32_512_port": 100833265}}

$ icount.py --target m33 --build-dir <dir>/build-m33 --plugin <plugin> --update --record <dir>/measured-m33.json   # update-record
rfft_f32_2048: 116385409 insns (NO BASELINE — commit this value)
rfft_f32_512: 102248169 insns (NO BASELINE — commit this value)
--- informational: '_port' scenarios (counted, never gated, never baselined; Stage 2a until 2c) ---
rfft_f32_2048_port: 115396527 insns [engine=split_radix backend=split_radix]; rfft_f32_2048: 116385409 insns [engine=reference_c backend=ooura]; ratio rfft_f32_2048_port/rfft_f32_2048 = 0.9915; output checksums identical
rfft_f32_512_port: 100833265 insns [engine=split_radix backend=split_radix]; rfft_f32_512: 102248169 insns [engine=reference_c backend=ooura]; ratio rfft_f32_512_port/rfft_f32_512 = 0.9862; output checksums identical
recorded <dir>/measured-m33.json
updated <dir>/update-record.json
-> exit 0
   baselines file now: {"m33": {"rfft_f32_2048": 116385409, "rfft_f32_512": 102248169}}
   measured-m33.json: {"informational": {"m33": {"rfft_f32_2048_port": 115396527, "rfft_f32_512_port": 100833265}}, "m33": {"rfft_f32_2048": 116385409, "rfft_f32_512": 102248169}}

$ icount.py --target m33 --build-dir <dir>/build-m33 --plugin <plugin>   # port-timeout-informational (new path)
rfft_f32_2048: 116385409 insns vs baseline 116385409 (+0.00%) ok
rfft_f32_512: 102248169 insns vs baseline 102248169 (+0.00%) ok
--- informational: '_port' scenarios (counted, never gated, never baselined; Stage 2a until 2c) ---
rfft_f32_512_port: informational binary failed: <dir>/build-m33/bench/icount/tap_dsp_icount_rfft_f32_512_port: timed out after 1200 s under QEMU
rfft_f32_2048_port: 115396527 insns [engine=split_radix backend=split_radix]; rfft_f32_2048: 116385409 insns [engine=reference_c backend=ooura]; ratio rfft_f32_2048_port/rfft_f32_2048 = 0.9915; output checksums identical
-> exit 0

$ icount.py --target m33 --build-dir <dir>/build-m33 --plugin <plugin>   # port-ok0-informational (new path)
rfft_f32_2048: 116385409 insns vs baseline 116385409 (+0.00%) ok
rfft_f32_512: 102248169 insns vs baseline 102248169 (+0.00%) ok
--- informational: '_port' scenarios (counted, never gated, never baselined; Stage 2a until 2c) ---
rfft_f32_2048_port: informational binary failed: <dir>/build-m33/bench/icount/tap_dsp_icount_rfft_f32_2048_port: workload did not complete cleanly
rfft_f32_512_port: 100833265 insns [engine=split_radix backend=split_radix]; rfft_f32_512: 102248169 insns [engine=reference_c backend=ooura]; ratio rfft_f32_512_port/rfft_f32_512 = 0.9862; output checksums identical
-> exit 0

$ icount.py --target m33 --build-dir <dir>/build-m33 --plugin <plugin>   # gated-binary-fails (unchanged: aborts)
rfft_f32_2048: 116385409 insns vs baseline 116385409 (+0.00%) ok
-> exit SystemExit(<dir>/build-m33/bench/icount/tap_dsp_icount_rfft_f32_512: workload did not complete cleanly)

$ icount.py --merge <dir>/measured-m33.json <dir>/hand-edited-m4f.json   # merge; the second file smuggles a _port key under "m4f"
m33: 2 scenario(s) from <dir>/measured-m33.json
m4f: rfft_f32_512_port from <dir>/hand-edited-m4f.json DROPPED (informational scenario; never a gate entry)
m4f: 1 scenario(s) from <dir>/hand-edited-m4f.json
merged into <dir>/merge.json
-> exit 0
   baselines file now: {"m33": {"rfft_f32_2048": 116385409, "rfft_f32_512": 102248169}, "m4f": {"rfft_f32_512": 98090666}}

CI on d8eccf6 is being watched; not merging.

---

Generated by Claude Code