feat: add instruction count table to openvm-prof

[876pol]

Resolves INT-5845.

Only the metric post-processor is edited. To generate the metrics.json with the instruction count information, run with the perf-metrics feature. For instance, to run the fibonacci test and generate the instruction count table, run

OUTPUT_PATH="metrics.json" GUEST_SYMBOLS_PATH="guest.syms" cargo run --release --bin fibonacci --features perf-metrics
openvm-prof --json-paths metrics.json

The generated aggregated table for fibonacci (for n=3000000) looks like:

group	opcode	frequency
fibonacci_program	ADD	27,000,065
fibonacci_program	SLTU	9,000,002
fibonacci_program	BEQ	3,000,008
fibonacci_program	BNE	3,000,005
fibonacci_program	JAL	3,000,002
fibonacci_program	STOREW	32
fibonacci_program	LOADW	21
fibonacci_program	LUI	17
fibonacci_program	JALR	13
fibonacci_program	STOREB	10
fibonacci_program	AUIPC	9
fibonacci_program	LOADBU	7
fibonacci_program	AND	4
fibonacci_program	OR	3
fibonacci_program	BLT	2
fibonacci_program	BLTU	2
fibonacci_program	HINT_BUFFER	2
fibonacci_program	SUB	1
fibonacci_program	PHANTOM	1
fibonacci_program	BGEU	1
fibonacci_program	HINT_STOREW	1

The detailed table with the breakdown by segment looks like:

group	opcode	segment	frequency
fibonacci_program	ADD	0	3,655,773
fibonacci_program	SLTU	0	1,218,584
fibonacci_program	BEQ	0	406,197
fibonacci_program	BNE	0	406,196
fibonacci_program	JAL	0	406,195
fibonacci_program	LUI	0	11
fibonacci_program	STOREW	0	11
fibonacci_program	LOADW	0	10
fibonacci_program	AUIPC	0	5
fibonacci_program	JALR	0	4
fibonacci_program	AND	0	3
fibonacci_program	OR	0	3
fibonacci_program	BLTU	0	2
fibonacci_program	HINT_BUFFER	0	2
fibonacci_program	BGEU	0	1
fibonacci_program	HINT_STOREW	0	1
fibonacci_program	PHANTOM	0	1
fibonacci_program	SUB	0	1
fibonacci_program	ADD	1	3,657,000
fibonacci_program	SLTU	1	1,219,000
fibonacci_program	BNE	1	406,334
fibonacci_program	BEQ	1	406,333
fibonacci_program	JAL	1	406,333
fibonacci_program	ADD	2	3,656,999
fibonacci_program	SLTU	2	1,219,001
fibonacci_program	JAL	2	406,334
fibonacci_program	BEQ	2	406,333
fibonacci_program	BNE	2	406,333
fibonacci_program	ADD	3	3,657,001
fibonacci_program	SLTU	3	1,218,999
fibonacci_program	BEQ	3	406,334
fibonacci_program	BNE	3	406,333
fibonacci_program	JAL	3	406,333
fibonacci_program	ADD	4	3,657,000
fibonacci_program	SLTU	4	1,219,000
fibonacci_program	BNE	4	406,334
fibonacci_program	BEQ	4	406,333
fibonacci_program	JAL	4	406,333
fibonacci_program	ADD	5	3,656,999
fibonacci_program	SLTU	5	1,219,001
fibonacci_program	JAL	5	406,334
fibonacci_program	BEQ	5	406,333
fibonacci_program	BNE	5	406,333
fibonacci_program	ADD	6	3,657,001
fibonacci_program	SLTU	6	1,218,999
fibonacci_program	BEQ	6	406,334
fibonacci_program	BNE	6	406,333
fibonacci_program	JAL	6	406,333
fibonacci_program	ADD	7	1,402,292
fibonacci_program	SLTU	7	467,418
fibonacci_program	BEQ	7	155,811
fibonacci_program	BNE	7	155,809
fibonacci_program	JAL	7	155,807
fibonacci_program	STOREW	7	21
fibonacci_program	LOADW	7	11
fibonacci_program	STOREB	7	10
fibonacci_program	JALR	7	9
fibonacci_program	LOADBU	7	7
fibonacci_program	LUI	7	6
fibonacci_program	AUIPC	7	4
fibonacci_program	BLT	7	2
fibonacci_program	AND	7	1

[jonathanpwang]

Can you provide instructions (in this PR desc for now) on how to generate the metrics.json needed by openvm-prof?
E.g., what feature flags to enable?

Provide explicit commands for one particular case to test.

After it looks good, we can update the benchmark docs

[876pol]

I've just updated the description with the instructions

[jonathanpwang]

Can you remove the dsl_ir column and run it on an example with multiple segments (just a larger fibonacci program would do it)? Then give a top level table where the instruction counts are aggregated across all segments. Then more detailed per-segment instrumentation.

[jonathanpwang]

@876pol can you fix the lint? Format on save is likely not being triggered because you need to install Rust nightly: rustup install nightly.

[github-actions]

group	app.proof_time_ms	app.cycles	app.cells_used	leaf.proof_time_ms	leaf.cycles	leaf.cells_used
verify_fibair	(+39 [+17.4%]) 263	322,610	2,058,654	-	-	-
fibonacci	(-38 [-3.6%]) 1,011	1,500,209	2,100,402	-	-	-
regex	(+150 [+6.5%]) 2,452	4,137,502	17,695,216	-	-	-
ecrecover	(-19 [-2.5%]) 733	122,859	(+2328 [+0.1%]) 2,265,100	-	-	-
pairing	(+102 [+7.1%]) 1,536	1,745,742	25,408,302	-	-	-

Commit: cb26827

Benchmark Workflow

[jonathanpwang]

LGTM!

Can you update the documentation for benchmarks to include a section on profiling detailed instruction count?

[jonathanpwang]

not sure if we need this since these are more detailed metrics