retroSoC/docs/mini-high-performance-traceability.md

Mini High-Performance Traceability

Status vocabulary

• implemented: RTL and software-visible behavior exist and directed tests pass.
• partial: a bounded subset is implemented and the missing boundary is stated.
• verified: the required simulation and formal evidence pass.
• physical-pending: digital behavior exists but synthesis or later evidence is required.
• planned: the implementation is not complete.

Requirement matrix

Requirement	Implementation owner	Required evidence	Status
Independent HP I-cache and D-cache paths	HP subsystem and data plane	Concurrent different-target AXI test	implemented
Target-aware multi-outstanding AXI64	Data crossbar	Multi-ID, ordering, backpressure, and formal tests	implemented: SRAM/SDRAM use 4R/2W credits; serial targets use 1R/1W; full formal proof pending
LP access to every memory window through the data plane	LP data gateway	no-PLL boot and memory-window smoke	implemented
Bounded QoS and recovery priority	Data crossbar	starvation and emergency-promotion test	implemented and directed-tested at a reduced four-cycle bound; production bound is 256 cycles and full liveness proof remains pending
Initiator ACL and first-fault attribution	Generated data policy, crossbar, SYSCTRL, and Fabric Monitor	target/execute/cache/range denial and injection tests	implemented: fixed policy is topology-generated and valid/ready fault reporting retains every accepted event across the HP/PCLK CDC
Stable memory functional clocks	Memory integration	HP DFS while refresh/protocol engines continue	implemented digitally: SRAM is native HP AXI64; SDRAM and all serial targets cross directly from HP to stable memory as AXI64 before local downsizing
Reset epoch and stale-response rejection	Async bridges	unilateral reset and clock-stop matrix	partial: coordinated warm flush and epoch implemented
HP normal and forced hot reset	AON lifecycle controller and Resource Controller	cache request/ACK, drain, timeout, flush, and recovery tests	implemented digitally: Zicbom and bounded ACK window exist; range policy and Linux service remain software work
Clock monitor and programmable timeout	AON clock/reset subsystem	8x8 DFS and clock-loss tests	implemented
Operational EXT-H ACL, timeout, and data path	Extension subsystem	transfer, denial, hang, and quiesce tests	implemented
Ownership-aware IRQ handoff	Resource Controller and HP PLIC	idle handoff, owner lock, LP/HP exclusion, and fault tests	implemented for DMA, USB2, SDIO0/1, SPI-SD, EXT-H, and JPEG; downstream per-engine reset ACK remains pending
JPEG 1080p60 at 72 MHz	JPEG codec and private AXI4 DMA	complete 1080p encode/decode cycles, randomized AXI stalls, synthesis/STA/power	partial: functional direct/ring codec and cycle benchmark exist; current encode core projects to about 24 fps, and the required multi-lane pipeline remains planned
Bounded target timeout and isolation	Target guards	pre-/post-accept timeout, burst completion, and late-response isolation	implemented: synthetic SLVERR and fail-closed isolation; physical fault injection pending
Root-visible fabric observability	Fabric Monitor	counters, snapshot, high-water, promotion, timeout, isolation, flush, sticky fault, and RTL/C parity tests	implemented; histogram/interrupt/trace streaming and silicon correlation remain pending
QPI/OPI pad exclusion and inactive-window error	Memory pad manager	simulation and formal mutual-exclusion proof	implemented
Technology clock cells and physical closure	Technology/physical flow	synthesis, MMMC, SDF, CDC/RDC, and PVT	physical-pending

The architecture report under /nfs/share/home/miaoyuchi/plan is the design input. This matrix and executable configuration are the repository status sources of truth.

Historical RV32 verification evidence

The results below predate the RV64 migration and must not be used as RV64 qualification. Current generation selects rv64imafdc_zicbom_max; see HP RV64 validation for migration-specific evidence.

• The complete Python suite passes 308 tests, including directed AXI64 multi-ID/ACL/QoS, target isolation, native AXI64 SRAM, Fabric Monitor, async warm-flush, clock monitor, HP lifecycle, Resource Controller, EXT-H DMA, topology, clock/reset inventory, and physical macro mapping.
• Generated PRODUCT Vexii manifests select rv32imafdc_zicbom_max. IHP130 PRODUCT Icarus assembly reports SIM_TEST_PASS at 3.532422647 ms. Its PR ci_smoke Verilator run with SVA, fast flash, and the 32 KiB all-SRAM layout passes ARCHINFO, SRAM, Fabric Monitor HAL/APB, extensions, RNG, CLINT, timers, GPIO, SDRAM, crypto, and USB2, then reports SIM_TEST_PASS code=0 after 4,540,766 cycles.
• The historical ICS55 no-PLL/no-SRAM PRODUCT result predates the locked OpenECOS SRAM integration. A fresh ICS55 PRODUCT run with the committed 32 KiB OpenECOS SRAM profile is required before claiming current timing or simulation evidence; ignored local commercial SRAM variants are no longer part of the supported flow.
• Strict IHP130 PRODUCT Verilator lint/elaboration completes successfully. Repository format, full RTL style audit/readiness, clock/reset inventory, embedded-C format/policy/host, Ruff, YAML, Actions, regression dry-runs, and Pytest gates pass.
• The 12-target rib2apb topology passes its depth-20 SymbiYosys/Bitwuzla proof and cover. Product SYSCTRL and PLL/RCU retain their bounded proof and cover targets. A complete AXI64 liveness proof remains pending.

The strict IHP130 warning-baseline comparison remains an observation: the baseline predates the fixed HP/product topology and does not contain many generated VexiiRiscv, newly enabled peripheral, and PDK signatures. New warning signatures introduced directly by the data-plane, lifecycle, extension, and CDC integration were removed; the baseline was intentionally not regenerated or hand-edited in this change.

Synthesis, netlist simulation, STA, CDC/RDC signoff, gate-level power intent, and PVT characterization are intentionally outside this delivery, as required by the implementation scope.