You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit d9df40d
Browse filesBrowse the repository at this point in the historyBrowse files
tinytpu: column-summed drain for op-46 + matching tests
Controller's do_drain branch for the generic-VPU epilogue now reads
array.getResults (column-summed GEMM output, one row per dispatch)
instead of array.getMatrix (per-PE rank-1 state). This unlocks the
fused op-46 path for standard NN kernels — linear, attention, FFN,
RoPE all need 'GEMM-then-elementwise', which is exactly what
getResults produces.
Controller.bsv:
- applyCuratedOpRow / applyIpairRotateRow: 1-row helpers replacing
the matrix-shaped ones. The case statement still elaborates once
(no G0024 warnings).
- Drain branch: reduced = array.getResults; src2_row = src2Reg[0]
(caller materializes one VMEM slot per output row, with the
coefficients placed at row 0); result_row written to outm[0],
outm[1..rows-1] zero-filled. The SXU writeback still reads the
full tile so existing read paths keep working.
tests/test_tinytpu_backend.py (TestMxuVpuEpilogue): updated expected
outputs to row 0 = op(act, src2_row0), lanes COLS..15 = 0. Helper
_expected_row keeps the assertion form short.
tests/test_e2e_rope_fusion.py: test_rope_runtime_matches_numpy_reference
now uses a non-trivial 4x4 activation tile and asserts all 4 dispatch
outputs match IPAIR_ROTATE(act_row_r, interleaved_cs_row_r) per row
(meaningful lanes 0..3, zeros after). Identity weights so the GEMM
result equals the input activation row; the fusion path itself is
under test, not the systolic-array correctness.
doc/tinyspec_coverage.md: added IPAIR_ROTATE and
DISPATCH_MXU_VPU_EPILOGUE to the supported op lists.
tinygrad submodule bump: matching lowerer + INTERLEAVE_CS materializer
changes (per-(row, tile) src2 slots).
Verification: 1231/1231 tests pass (full suite minus pre-existing
test_benchmark.py::test_benchmark_runs_add_kernel which is a
sim-side cycle-counter issue, not introduced by this work).
0 commit comments