Skip to content

README.md

Latest commit

 

History

History
72 lines (50 loc) · 2.46 KB

README.md

File metadata and controls

72 lines (50 loc) · 2.46 KB

bmpu_verify

bmpu_verify is the fastest correctness-oriented app in the repository for validating BMPU mixed-precision behavior.

What It Checks

  • generated test vectors match the declared benchmark cases
  • BMPU execution completes successfully
  • packed output can be unpacked back into column-major form
  • unpacked output matches the reference result tensor

Current Coverage

  • 18 directed cases
  • precisions: binary, INT2, INT4
  • group shapes: gm x gn = 2x2, 4x1, 1x2, 1x4, 8x1, 1x8
  • tail coverage for non-full tiles and larger ntile edge cases

The gm=8 / gn=8 cases are legal even when the last tile group is smaller than 8. Execution clips the tail group to the remaining tile count, so these cases explicitly exercise the clipped-group path.

Additional current coverage includes small and tail-oriented binary cases such as:

  • mtile=8, ntile=128 minimal and tail cases
  • mtile=16, ntile=64 minimal and tail cases

Data Layout Notes

  • result_torch is stored as column-major reference data.
  • result_hp is not a plain M*N matrix in this app. It is a packed BMPU store buffer laid out tile-by-tile and block-by-block.
  • For non-full edge tiles, result_hp must be sized by packed tile capacity: ceil(M/mtile) * ceil(N/ntile) * ceil(mtile/8) * ceil(ntile/16) * 8 * 16 int16 elements.

If result_hp is only allocated as M*N, tail-tile cases can overwrite following symbols in the generated dataset and cause false mismatches or hangs.

Why It Matters

Use this app before large benchmark campaigns when you changed:

  • BMPU RTL
  • low-precision packing logic
  • benchmark data generation
  • shared mixed-matmul helpers

This app should remain the quickest full-flow signal that a maintained BMPU-related change is still functionally safe.

Local Structure

  • main.c: executes the validation loop and compares results
  • kernel/: generated tensors and benchmark case metadata
  • script/gen_data.py: generator for the validation dataset

Debug Controls

  • app-side compare/debug prints are controlled by BMPU_VERIFY_DEBUG
  • hardware-side debug prints are controlled from src/hardware/rtl/extended/include/bitfly_debug.svh
  • keep both disabled by default for full regressions

Typical Flow

source /data2/zzx/data/miniconda3/etc/profile.d/conda.sh
conda activate bitfly
./scripts/sync_src_to_ara.sh --no-patch
make -C ara/apps bin/bmpu_verify -j8
make -C ara/hardware simv app=bmpu_verify

A healthy run ends with ALL CASES PASSED.