Skip to content

Implement all 14 topic modules with interfaces, docs, tests, benchmarks - #5

Merged
jizhenjun merged 10 commits into
mainfrom
dev_benchmark4topics
Jun 1, 2026
Merged

Implement all 14 topic modules with interfaces, docs, tests, benchmarks#5
jizhenjun merged 10 commits into
mainfrom
dev_benchmark4topics

Conversation

@k1nsom

@k1nsom k1nsom commented May 29, 2026

Copy link
Copy Markdown
Collaborator

Backend modules (7):

  • asm_beautifier.py: RISC-V assembly beautifier with alignment/commenting
  • inst_counter.py: instruction category counter with charts/HTML reports
  • asm_peephole.py: assembly-level peephole optimizer (5 default rules)
  • const_merge.py: constant load merge (lui+addi fusion, redundancy elimination)
  • regalloc_linear.py: linear scan register allocator with spill code
  • inst_scheduler.py: list scheduler with DAG construction and latency model
  • inst_select_ext.py: extended instruction selection (sqrt/min/max/abs/fp64)

Frontend/infra modules (7):

  • dsl_extended.py: DSL enhancer with if/else and while support
  • dsl_errors.py: gcc-style error beautifier with fix suggestions
  • logger.py: colored logging system with file output and phase timing
  • cfg_builder.py: CFG builder with dominator tree and loop detection
  • ir_verifier.py: IR verifier with 7 rule categories
  • bench_runner.py: benchmark suite with 23 DSL test cases and HTML reports
  • Code standards: .pre-commit-config.yaml + CODING_STANDARDS.md

Each topic includes: module + docs/topics/ guide + tests + benchmarks
Total: 125 files, 12655 insertions, 348 tests passing, lint clean

@k1nsom @claude
Implement all 14 topic modules with interfaces, docs, tests, benchmarks
e081b6f 
Backend modules (7):
- asm_beautifier.py: RISC-V assembly beautifier with alignment/commenting
- inst_counter.py: instruction category counter with charts/HTML reports
- asm_peephole.py: assembly-level peephole optimizer (5 default rules)
- const_merge.py: constant load merge (lui+addi fusion, redundancy elimination)
- regalloc_linear.py: linear scan register allocator with spill code
- inst_scheduler.py: list scheduler with DAG construction and latency model
- inst_select_ext.py: extended instruction selection (sqrt/min/max/abs/fp64)

Frontend/infra modules (7):
- dsl_extended.py: DSL enhancer with if/else and while support
- dsl_errors.py: gcc-style error beautifier with fix suggestions
- logger.py: colored logging system with file output and phase timing
- cfg_builder.py: CFG builder with dominator tree and loop detection
- ir_verifier.py: IR verifier with 7 rule categories
- bench_runner.py: benchmark suite with 23 DSL test cases and HTML reports
- Code standards: .pre-commit-config.yaml + CODING_STANDARDS.md

Each topic includes: module + docs/topics/ guide + tests + benchmarks
Total: 125 files, 12655 insertions, 348 tests passing, lint clean

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@k1nsom
k1nsom requested a review from jizhenjun May 29, 2026 07:39
k1nsom and others added 9 commits May 29, 2026 16:12
@k1nsom @claude
Update README and project landing page
ecabe5e 
README: document all 14 new topic modules, analysis/ and utils/
directories, benchmarks/, updated test count (348), CNN pipeline,
extended DSL usage, and topic table with module paths.

ScratchV.html: update topic table to include all 14 topics with
links to both implementation guides and original topic proposals.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@k1nsom @claude
Eliminate all call pseudo-instructions, map every op to real RISC-V
b068fd0 
Replace all MachineOp.CALL placeholders with inline RV32IM instruction
sequences. Every ONNX operator now maps directly to real RISC-V:

- Conv: mv(bias) + mul(x*w) + add(acc)  (MAC)
- Gemm: mv(bias) + mul(a*w) + add(acc) (MAC)
- ReLU: max rd, rs, 0
- MaxPool: slt + bnez + li + j (branch-based max)
- Sigmoid: slt + bnez + li + j (piecewise clamp [0,1])
- Reshape: mv rd, rs
- Exp: addi + max (linear approx)
- GELU: max + mul + div (simplified)
- Softmax: mv (passthrough)
- MatMul/Dot: mul

Add 15 single-precision float MachineOp codes (FADD_S, FMUL_S, etc.)
for future RV32F support. Fix branch/jump emission to output target
labels as operands instead of comments. Use .L prefix for local labels
(GAS convention).

0 call instructions remaining. 96 pure RISC-V instructions for CNN.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@k1nsom @claude
Add library-free ONNX to RISC-V standalone pipeline
83fd10a 
Complete self-contained CNN compiler: ONNX protobuf -> RISC-V RV32IM binary.
Zero external library dependencies (Python 3.8+ stdlib only).

Features:
- Manual protobuf wire-format parser (no 'onnx' package needed)
- Q16.16 fixed-point weight conversion
- Memory planner for intermediate tensors
- Inline RISC-V codegen for all 15 CNN operators:
  Conv2D (6-deep nested loops), MaxPool (5-deep nested loops),
  Gemm/FC (3-deep nested loops), ReLU (branch-free, 3 instructions),
  Sigmoid (piecewise linear), Reshape (element-wise copy)
- Direct RISC-V machine code encoding (no assembler needed)
- Generated binary has zero runtime library calls (bare-metal ready)
- Position-independent code via AUIPC data addressing

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@k1nsom @claude
Add RISC-V performance benchmark engine (emulator + analytical estima…
ef957b7 
…tor)

Library-free RV32IM emulator with inline performance counters:
- Dynamic instruction mix (ALU/memory/branch/jump/upper percentages)
- Memory access stats (loads, stores, L/S ratio)
- Compute-to-memory ratio (CNN workload classification)
- Branch behaviour (taken/not-taken rate)
- Top-10 hottest PC addresses (sampled every 1024 insns)
- Per-operator (per-label) instruction counts
- Progress logging every 10M instructions
- Host execution time and simulated MIPS

Analytical instruction estimator (instant, no emulation):
- Estimates total dynamic instructions from CNN dimensions
- Per-layer breakdown with percentages
- Estimated hardware execution time @50/100MHz

Usage:
  python onnx_to_riscv_standalone.py model.onnx --benchmark --max-instr 10000000
  python onnx_to_riscv_standalone.py model.onnx --estimate

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@k1nsom @claude
Add CI-visualizable benchmark reports (HTML/JSON/GitHub summary)
43eca5c 
New bench_report.py generates three output formats:
- HTML: self-contained page with CSS bar charts, C/M gauge,
  per-layer breakdown, summary cards (zero JS dependencies)
- JSON: machine-parseable for CI dashboards and metrics tracking
- GitHub Actions job summary: markdown table for CI step summary

CI integration:
- New 'cnn-benchmark' job in ci.yml: compiles cnn.onnx to RISC-V,
  generates all reports, uploads HTML artifact (30-day retention),
  writes GitHub step summary
- Makefile: 'make bench-cnn' (estimate + reports) and
  'make bench-cnn-emu' (emulation + reports)
- .gitignore: exclude benchmark_reports/

CLI: --report flag on standalone pipeline generates all formats.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@k1nsom @claude
Add CNN pipeline benchmark reports, PPT assets, and misc updates
cb4c646 
- bench_report.py: HTML/JSON/GitHub summary report generator
- benchmark.py: RV32IM emulator with inline performance counters
- onnx_to_riscv_standalone.py: --benchmark, --estimate, --report flags
- CI: cnn-benchmark job with HTML artifact upload
- Makefile: bench-cnn, bench-cnn-emu targets
- riscv_encoder.py: minor updates
- PPT and gen_ppt scripts for project presentation

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@k1nsom
add cnn.onnx
e66a842
@k1nsom @claude
Add per-instruction cycle-accurate microarchitecture model
3abb273 
Four configurable RISC-V microarchitecture profiles:
  single:  CPI=1.00  (baseline, all ops 1 cycle)
  fast:    CPI=1.06  (MUL=1, DIV=4, LW=1, branch-taken=2)
  basic:   CPI=1.55  (MUL=4, DIV=34, LW=2, branch-taken=3)
  slow:    CPI=5.02  (MUL=32, DIV=34, LW=5, branch-taken=3)

Features:
- MicroArch class with per-category cycle costs (ALU_R, ALU_I,
  SHIFT, MUL, DIV, LOAD, STORE, BRANCH taken/not, JUMP, JALR)
- Emulator hot-loop tracks cycles per instruction inline
- PerfCounters tracks total_cycles, cat_cycles[], CPI
- Cycle distribution in benchmark report (bar chart)
- Analytical estimator projects cycles for all 4 profiles
- --uarch CLI flag (single|fast|basic|slow, default: basic)

CNN model estimates (basic profile):
  CPI=1.55, 12.1B cycles, 242s @50MHz, 121s @100MHz
  (vs CPI=1.0 single-cycle: 7.8B cycles, 156s @50MHz)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@k1nsom @claude
Reorganize CI: simplify to test + benchmark jobs
9dd3644 
CI restructured from 6 jobs to 2:
  test:      all 14 topic module tests (pytest tests/, single python3.12)
  benchmark: model perf tests (ONNX pipeline + 23 DSL cases + CNN RISC-V)

Removed:
  - Python version matrix (3.9/3.10/3.11/3.12) → single python3.12
  - lint job (flake8/mypy — kept in Makefile for dev use)
  - coverage job
  - smoke job (DSL examples — covered by DSL bench cases)

Makefile simplified to match:
  make test       → topic tests
  make bench      → model benchmarks
  make bench-cnn  → CNN RISC-V compilation + estimation
  make lint       → dev-only linting
  make clean      → cleanup including benchmark_reports/

Also add examples/cnn_model.dsl (ONNX→DSL conversion output).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@jizhenjun
jizhenjun merged commit 6088b3d into main Jun 1, 2026
0 of 2 checks passed
k1nsom added a commit that referenced this pull request Jun 4, 2026
@k1nsom @claude
fix: review 全部10个问题修复
5bcefa2 
#1+#2 dashboard数据管线: _run()从--json-output路径读取,generate_dashboard_html接收参数
#4 CI CNN benchmark: 模型不存在时自动生成最小CNN onnx,不再静默跳过
#5 tinyfive CI兼容: --scratchv-asm/--llvm-asm CLI参数,文件不存在容错
#6 estimator参数化: estimate_cnn_model接受model_spec dict
#8 LLVM TinyFive内核: 用真实LLVM O3汇编提取的指令序列替代手写
#9 import统一: llvm_cache_compare使用绝对路径导入cache_model

未修(需更大重构):
#3 run_spike_bench field() bug (需dataclass重构,影响力低)
#10 CPU仿真器重复代码 (需提取共享模块)
#7 缓存分析模型 (需trace-driven,实现成本高)

348 tests passed

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@github-actions github-actions Bot mentioned this pull request Sep 13, 2026
Merged
1 task
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@jizhenjun jizhenjun jizhenjun approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@k1nsom @jizhenjun