Bibliography of Source Binaries

This appendix is the formal, SHA-256-pinned citation list for the whole reference: one row per shipped artifact the wiki derives a fact from, with its role, its absolute on-disk location, its hash, its byte size, its file class, the tool that decodes it, and the wiki Parts it grounds. It is the citation-grade companion to the working catalogue in The Corpus, Tiers & Binary Inventory: that page tiers and narrates the corpus; this page is the bibliography you cite a hash from.

Every hash and size below was re-computed directly from the file on disk this pass (sha256sum, stat -c %s, readelf -h) — none was copied from a prior report. A reader who meets any OBSERVED claim anywhere in the guide can come here, re-hash the named file on their own disk, and confirm they are looking at the same bytes. Where a figure re-derived this pass disagreed with the working inventory, an in-line CORRECTION says so and flags it for the Part-16 reconcile; no such SHA/size disagreement was found — every figure on this page matches corpus-inventory to the byte.

NOTE — what "binary-derived" means here. Everything catalogued is a shipped, redistributable artifact: an ELF, a static archive, a JSON config, a Python pickle, or a C header that the AWS Neuron firmware / toolchain actually ships. Reading, hashing, disassembling, and (for the freely-callable value/decode oracles) executing in-process these bytes is the entire method. No vendor source tree was consulted; the JSON/pickle/header artifacts are themselves shipped files, as citeable as any ELF. See Methodology for how each is driven.

How to read a row

• SHA-256 — the first 16 hex are shown in the table for density; the full 64-hex for every artifact is listed once in § Full hashes so a citation can pin the complete digest.
• Size — exact byte count from stat -c %s, [HIGH/OBSERVED].
• Class/machine — readelf -h: device firmware containers are x86-64 host wrappers (the device code is embedded inside them as ELF32-Xtensa e_machine = 0x5e = 94 blobs, not a standalone machine on the file itself); the ncore2gp config DLLs and libnrt.so are x86-64; JSON / pickle / header / .a-archive artifacts are tagged as such.
• How-accessed — the tool that turns the bytes into a fact (readelf/nm/objdump for host ELFs; the shipped xtensa-elf-objdump --xtensa-core=ncore2gp for device Xtensa; ctypes for the freely-callable value oracle; jq for JSON; pickletools only — never pickle.load — for the pickle).

All confidence tags are [HIGH/OBSERVED] for every hash and size, because each was measured against the file this pass (see The Confidence & Walls Model).

1. Device firmware containers (the Vision-Q7 images live inside these)

The GPSIMD device code that runs on the Q7 DSP is not a standalone file. It is a set of ELF32-Xtensa (e_machine = 94) images embedded as .rodata data inside host x86-64 wrapper libraries, reached through …_SO_get / …_JSON_get accessor symbols. The container file is x86-64; the firmware it vends is Xtensa. [HIGH/OBSERVED]

CORRECTION — libnrtucode_extisa.so and libncfw.so are NOT standalone files in this checkout. TASK #995's scope names …_extisa.so and libncfw.so among the firmware containers. Re-checked against disk this pass (fd --no-ignore over the whole tree): neither exists as a file here.

• The "EXTISA" content is the embedded blob set inside libnrtucode_internal.so (CAYMAN_Q7_POOL_PERF_EXTISA_<n>_SO_get, …) — 66 EXTISA accessor symbols, not a separate .so.
• NCFW (the scalar-Xtensa-LX management core) ships only as an embedded image; no standalone libncfw.so is present. The flat-container libnrtucode_extisa.so variant ships in the sibling neuronx-runtime corpus, so any fact sourced from it is CARRIED across the package boundary.

This matches corpus-inventory §3 "NOT-PRESENT" exactly; no reconcile divergence — recorded here so a citation never reaches for a file that is not on disk.

The embedded-image counts were re-counted this pass by scanning \x7fELF magic and reading e_machine: 16 ELF32/e_machine=94 images in libnrtucode_internal.so, 12 in libnrtucode.so. [HIGH/OBSERVED] per-library; the cross-library aggregate ≈29 is [HIGH/CARRIED] (spans both libs — the dedup overlap is why it is not a simple 16+12).

2. Host runtime — libnrt.so (sibling neuronx-runtime corpus)

CORRECTION — libnrt.so is file-absent from the gpsimd checkout; it lives in the sibling neuronx-runtime tree. corpus-inventory §3 and methodology §1 both mark the host NeuronCore runtime as NOT-PRESENT in this gpsimd checkout. It is present, and hashed here, in the sibling neuronx-runtime package — so host-runtime facts are OBSERVED against that corpus and CARRIED into gpsimd pages. The DWARF (.debug_*) corroboration leg referenced in methodology §2(c) reads from this very file.

The package build-id is 0b044f4ce (from the extraction directory name); the host runtime version is 2.31.24.0. [HIGH/OBSERVED]

3. ncore2gp ISS config DLLs (decode / value / timing oracles)

The nine host x86-64 DLLs under TOOLS/ncore2gp/config/ are the executable model of the Vision-Q7 core — callable in-process, not merely readable. All ship not stripped (full .symtab, no DWARF .debug_*). The -ref- twins are differential cross-check references. [HIGH/OBSERVED]

GOTCHA — the .data VMA ≠ file-offset for these DLLs (delta 0x200000). When xxd/objdump a .data-resident struct in any ncore2gp config DLL, subtract 0x200000 from the VMA to get the file offset (.text/.rodata are VMA==offset; .data is not). The delta is binary-specific — 0x200000 here, 0x3000 for the nrtucode pair — read the real Addr/Off columns with readelf -SW before addressing a .data struct. See Methodology §5.

QUIRK — libnrtucode_internal.so has zero C++ RTTI; this DLL set retains a full .symtab. The device firmware is -fno-rtti (_ZTV/_ZTI/_ZTS count = 0), so its class hierarchy is not recoverable from RTTI scaffold. The config DLLs are the opposite: named-symbol-rich, which is exactly why nm + ctypes reach OBSERVED-by-execution on the value lane.

4. customop-lib arch-ISA headers + instruction_mapping.json (4 generations)

The cleartext per-generation specification. Four unified generations ship as CUSTOMOP/c10/include/neuron_<gen>_arch_isa/ (aws_neuron_isa_* headers); Tonga (v1) is the pre-unified outlier under arch-headers/tonga/ + arch-isa/, not one of the four. [HIGH/OBSERVED]

The monotone climb 89 → 99 → 108 → 114 is the OBSERVED per-generation opcode growth; each instruction_mapping.json has top-level keys {struct2opcode, struct2pseudo_opcode} with struct2pseudo_opcode = 2 in every gen. The header .h counts (100/111/120/126) and JSON sizes (13 619 / 14 482 / 15 483 / 16 158) match corpus-inventory §5 to the byte. [HIGH/OBSERVED]

The firmware-side core-kind enum that ties these names to the loader is read from libnrtucode_internal.so strings: NRTUCODE_CORE_{SUNDA,CAYMAN,MARIANA,MARIANA_PLUS,MAVERICK}_NX_POOL — 5 NX-POOL core kinds. [HIGH/OBSERVED]

5. CSR JSON + RTL address maps + the Maverick address-map pickle

GOTCHA — al_address_map_db.pkl is Maverick, not Cayman; and never deserialize it. The pickle lives under the maverick/ customop arch-headers (…/arch-headers/maverick/ext/), verified this pass — not in the cayman-arch-regs tarball. Its first two bytes are 80 04 (pickle PROTO 4), confirming a valid pickle; inspect it with pickletools.genops / stat / sha256sum only. A pickle.load of an untrusted 207 MB blob executes arbitrary __reduce__ opcodes — it is treated as opaque bytes here. [HIGH/OBSERVED] (path provenance + magic).

6. Toolchain — the native Xtensa device disassembler + ISS cores

QUIRK — xtensa-elf-objdump has no default core. Invoked bare it errors "there is no Xtensa core registered as the default"; you must export XTENSA_CORE=ncore2gp (the only core in the bundled registry). It is the Vision config — the wrong config for the scalar Xtensa-LX NCFW management core, which must be decoded with the manual LX op0 ∈ {e,f} ⇒ 3-byte length rule (resync at retw.n). Using ncore2gp on NCFW bytes produces the spurious "~26–28% FLIX" artifact. See Toolchain Inventory & Versions §B.1.

Plain host GNU binutils (objdump/nm/readelf/sha256sum, 2.46-3.fc44) decode every x86-64 host ELF and .a archive above; the device xtensa-elf-objdump is reserved for e_machine=94 device code. [HIGH/OBSERVED]

7. Sibling cross-check corpora (T2 — corroboration only)

Three independently-shipped AWS Neuron corpora live in the same repo tree, used only to corroborate a gpsimd-internal fact from a separate build — never as the sole source of one. [HIGH/OBSERVED] (presence re-checked this pass.)

The libnrt.so host runtime (§2) is itself a sibling-corpus artifact (neuronx-runtime), catalogued separately because it is the host-runtime primary source, not merely a cross-check.

8. Lawful-interop posture

Every artifact catalogued here is an already-shipped, redistributable AWS Neuron package component, and every result derived from it is the product of static analysis — reading, hashing, disassembling, and (for the freely-callable value/decode oracles) executing in-process — of those shipped bytes. No vendor source tree was consulted; no design document was read; no proprietary firmware was decrypted or circumvented. The work is authored under the interoperability research exemption of 17 U.S.C. §1201(f) (and the analogous EU software-directive provisions): the sole purpose is to document the interfaces and behaviour necessary to build an independently interoperable Vision-Q7-compatible GPSIMD engine, toolchain, runtime, or simulator.

Two license walls are reported, not crossed: the cycle/timing oracle (libcas-core.so, libsimxtcore.so) gates instruction retirement on a FlexNet check (AUTH::check_iss_licenses), and no license circumvention is performed or documented — the wall is recorded as a wall (see The Confidence & Walls Model). The freely-callable value lane (libfiss-base.so) requires no license and is the only oracle this reference executes. The 207 MB Maverick pickle is treated as opaque bytes (pickletools only — never deserialized), so no arbitrary-code path is reached. [HIGH/OBSERVED]

Full SHA-256 digests

The complete 64-hex digest for every hashed artifact (all [HIGH/OBSERVED], computed this pass):

Path convention. CUSTOMOP/ = neuronx-gpsimd/extracted/aws-neuronx-gpsimd-customop-lib_0.21.2.0_amd64/opt/aws/neuron/gpsimd/custom_op/; TOOLS/ = neuronx-gpsimd/extracted/nested/gpsimd_tools_tgz/tools/. extracted/ is git-ignored — locate files there with fd --no-ignore or an absolute path. Package versions (themselves OBSERVED, from the extraction directory names): customop-lib 0.21.2.0; tools 0.21.0.0-bc9b5fad5; runtime 2.31.24.0-0b044f4ce.

Cross-references

• The Corpus, Tiers & Binary Inventory — the tiered, role-narrated working catalogue this bibliography pins to (the reference for every figure here).
• Methodology — How This Was Reverse-Engineered — how each artifact above is driven (the value-oracle ctypes drive, the device objdump, the offsetof compile-verify, the four verification gotchas).
• Toolchain Inventory & Versions — the exact xtensa-elf-objdump / ncore2gp / clang-10 / FlexLM version pinning for §6.
• The Confidence & Walls Model — the [HIGH/OBSERVED] tag system and the file-absent / license walls this page records.
• Codename ↔ Generation Cross-Walk — the five-codename axis the §4 header sets and §1 core-kind enum are keyed on.