Improve GPU defaults, caching, and pointwise scheduling#4668

Closed

Rolaand-Jayz wants to merge 13 commits intoROCm:developfrom

Rolaand-Jayz:opt/migraphx-performance-pr

Commits on Mar 13, 2026

Auto-tune default GPU stream count
Rolaand-Jayz
committed
Select NHWC and MLIR attention defaults by arch
Rolaand-Jayz
committed
Skip redundant host bounce on GPU fallback return
Rolaand-Jayz
committed
Fix MLIR-off GPU linkage in local build
Rolaand-Jayz
committed
Cache GPU feature probes by device
Rolaand-Jayz
committed
Cache repeated HIP code compilations
Rolaand-Jayz
committed
Cache repeated MiOpen convolution solutions
Rolaand-Jayz
committed
Avoid scheduling undersized GPU stream partitions
Rolaand-Jayz
committed
Tune GPU pointwise launch bounds by wavefront
Rolaand-Jayz
committed
Stabilize quantized GPU NHWC compilation
Rolaand-Jayz
committed
Support PReLU in MLIR pointwise fusion
Rolaand-Jayz
committed
Use MLIR tuning DB without CFG gate
Rolaand-Jayz
committed

Commits on Mar 14, 2026

fix: address review feedback for PR 4668
Rolaand-Jayz
committed