Quantized versions of Jacobi and Identity preconditioners - #2869

Open

pcarruscag wants to merge 26 commits into

Commits on Aug 11, 2026

message format
pcarruscag
committed
7b5ce77

Commits on Aug 12, 2026

Add quantized Jacobi and identity preconditioners

pcarruscag
and
claude
committed
4cb8044
Enable quantized matrix-vector product on GPU for Jacobi/identity

pcarruscag
and
claude
committed
31b6b8a
Group quantized storage into LDU<T> and overlap diagonal quantize with L/U upload

pcarruscag
and
claude
committed
e16cad7
Fold quantized diagonal build into the base preconditioner wrappers

pcarruscag
and
claude
committed
9bdd89c
Inline the diagonal quantization upload instead of a separate GPU hook

pcarruscag
and
claude
committed
76c27f2

Commits on Aug 15, 2026

Merge remote-tracking branch 'origin/develop' into pedro/q_jacobi
pcarruscag
committed
08d0406
Merge remote-tracking branch 'origin/develop' into pedro/q_jacobi
pcarruscag
committed
e6e020b
Pin the quantized L/U/D host arrays so the async H2D transfer actually overlaps

pcarruscag
and
claude
committed
c62a413
Batch rows per block and vectorize int8 loads in the SpMV kernels

pcarruscag
and
claude
committed
a1aaabf
Thread-per-row layout for the Jacobi preconditioner apply kernel

pcarruscag
and
claude
committed
a113c70
Try thread-per-block-entry layout for the Jacobi apply kernel

pcarruscag
and
claude
committed
56fe054
Revert "Try thread-per-block-entry layout for the Jacobi apply kernel"
pcarruscag
committed
feeae49
Batch rows per block in the Jacobi diagonal-inversion kernel

pcarruscag
and
claude
committed
e65bfa8
Revert "Batch rows per block in the Jacobi diagonal-inversion kernel"
pcarruscag
committed
7c308e6
Quantize the diagonal on the GPU straight from gpu.d, sharing the encoding logic

pcarruscag
and
claude
committed
40007df
Warp-align the row batching in the Jacobi diagonal-inversion kernel

pcarruscag
and
claude
committed
0bfd05b
Revert "Warp-align the row batching in the Jacobi diagonal-inversion kernel"
pcarruscag
committed
dc95fd8
Parallelize the Jacobi diagonal-inversion kernel across all block entries

pcarruscag
and
claude
committed
2492099
Parallelize the diagonal-block quantization kernel across rows

pcarruscag
and
claude
committed
1e71a1b
Overlap the quantized L/U H2D transfer with default-stream kernels

pcarruscag
and
claude
committed
ccd2564
Reuse the parallel Gauss-Jordan inversion for ILU's diagonal blocks

pcarruscag
and
claude
committed
36e98c4

Commits on Aug 16, 2026