§ 4.2.4. Generated Workloads¶

Benchmark reference · Fixed programs · Campaigns

Catalogue¶

The native catalogue contains these 15 generated families in addition to 22 fixed programs. A family is not one timing result: parameters, target, policy, and measurement window define the actual case.

Workload ID / CLI name	Purpose	Generator
`dependent`	Dependent integer recurrence	Legacy loop
`alternating`	Alternating branch loop	Legacy loop
`nested`	Nested counted loops	Legacy loop
`independent`	Independent ALU chains	Parameterized kernel
`dependency`	Dependency distance	Parameterized kernel
`load-use`	Immediate load-use pairs	Parameterized kernel
`load-scheduled`	Scheduled load-use pairs	Parameterized kernel
`branch-pattern`	Controlled branch outcomes	Parameterized kernel
`branch-alias`	Predictor table aliasing	Parameterized kernel
`calls`	Nested direct and indirect calls	Parameterized kernel
`stream`	Streaming copy and checksum	Parameterized kernel
`pointer-chase`	Dependent pointer chasing	Parameterized kernel
`sort`	Unsigned insertion sort	Parameterized kernel
`crc32`	Bitwise CRC32 kernel	Parameterized kernel
`scan`	Byte scan and classification	Parameterized kernel

Shared setup¶

The three legacy loops accept an iteration count and use their own registers; they do not implement the parameterized body's automatic training/window markers. The other 12 kernels share this setup:

Property	Contract
Iterations	1..1,000,000 measured iterations; optional training plus measured total also at most 1,000,000
Size	1..256 data elements, default 16; does not resize fixed-shape ALU/call/branch bodies
Seed	Nonzero 32-bit integer, default 1; xorshift32 makes generated input reproducible
Memory	Separate zero-filled instruction/data images; default 256 KiB each
Data addresses	Input/nodes 0x1000, output 0x2000, result word 0x3000
Stack	Calls kernel initializes sp=0x4000; other kernels are not ABI applications
Registers	Reset to zero; x28 is the outer counter and x10 is the result accumulator
Termination	Store the checksum at 0x3000, then EBREAK
Arithmetic	Unsigned 32-bit wraparound unless stated otherwise

--distance may be nondefault only for dependency; --branch-spacing only for branch-alias; --pattern only for branch-pattern. Shape parameters recorded in a case are not necessarily used by every family.

Behavioral description and checking¶

Each page describes the generated instruction sequence using compact pseudocode, not a replacement simulator implementation. The actual generator emits RV32I machine words. The runner checks those words on the selected core against the instruction oracle; parameterized kernels additionally provide host-computed expected checksums and, where applicable, expected output bytes.

In the page formulas, N = measured iterations + training iterations, and n is the data size. Whole-program final state includes training. A body-window counter covers only the selected measured region; its result need not be a checksum of that region alone.

Use --measure-body --training-iterations 32 for a parameterized body measurement after 32 real in-program training iterations. Without a body window, setup, loop control, final checksum reduction/store, and termination all belong to the whole-program measurement. See execution for exact retirement/window accounting.

Running¶

build/rv32i_experiment --core advanced --workload scan \
  --iterations 2 --size 16 --seed 1 --fidelity behavioral \
  --period 4000 --allocation-limit-mib 512 --max-cycles 4096

Here 4000 is an explicit conservative scheduling period for a small functional run, not a measured best clock or performance claim. Use a clock-certified campaign for cross-core timing comparisons. These headless generated workloads do not currently have individual embedded visualizer scenarios.

Source and tests¶

src/analysis/RV32IBenchmarks.cpp implements the 12 kernels and independent expected results. src/analysis/RV32IWorkloads.cpp owns the three legacy loops and shared preparation. Workload validation documents host-result, CLI, campaign, and corner-case tests.