fp: trace GPU vs CPU row-reduce dispatch (fp::rr) - #16

Draft

JoeyBF wants to merge 17 commits into

hpcfrom

worktree-hpc-build

Commits on Jul 20, 2026

Revert "nassau: remove the ParallelGuard priority-inversion retry mechanism"
JoeyBF
committed
83819ef

Commits on Jul 22, 2026

nassau: park retries instead of busy-spinning on ParallelGuard
claude
committed
5beaf22
nassau: make the parallel-section guard per-thread
claude
committed
80d614f
nassau: test that the parallel-section guard is thread-local
claude
committed
540b31f
Merge remote-tracking branch 'origin/claude/nassau-relaxed-dependency-graph-d69k0s' into hpc
JoeyBF
committed
2b54cf9
milnor_gpu: drop the global device mutex; make the resident store thread-local

JoeyBF
and
claude
committed
dbc645e

Commits on Jul 23, 2026

milnor_gpu: bound GPU-path memory (row blocks, launch permits, shared resident master)

JoeyBF
and
claude
committed
f1cc8fa
milnor_gpu: byte-weighted launch budget + prefix-doubling master uploads

JoeyBF
and
claude
committed
a0022b7
nassau: offload the d_s image build (signature_matrix) to the GPU multiply path

JoeyBF
and
claude
committed
a04dd91
milnor_gpu: zero the out_h accumulator on-device instead of uploading a host zero buffer

JoeyBF
and
claude
committed
bc253f6

Commits on Jul 24, 2026

milnor_gpu: move the resident-master device upload off the handle lock

JoeyBF
and
claude
committed
c433556
milnor_gpu: fill the term-data upload buffers directly, killing the per-product alloc storm

JoeyBF
and
claude
committed
164b129
milnor_gpu: raise GPU_PAIR_CHUNK to ~2^32 so giant multiplies stop over-splitting

JoeyBF
and
claude
committed
b02dcb3
milnor_gpu: make the Milnor basis resident on the GPU, upload term indices

JoeyBF
and
claude
committed
e6b85fd

Commits on Jul 25, 2026

milnor_gpu: safe multi-stream via in-place-grown resident globals

JoeyBF
and
claude
committed
4efee67
fp: trace GPU vs CPU row-reduce dispatch (fp::rr)

JoeyBF
and
claude
committed
0c1712a

Commits on Jul 26, 2026

fp,algebra: serialize cubecl multiply vs cooperative GPU row-reduce

JoeyBF
and
claude
committed
f22f8a7