v1.1gpu

What's Changed

• final_optim: solver stall break (C4), a larger late time step (D1), Cornelius without per-element allocations (A12) by @jhputschke in #22

Full Changelog: v1.0gpu...v1.1gpu

v1.0gpu

What's Changed

• Freeze-out search: alternating step buffers, e scan and cell gather (A10, A11) by @jhputschke in #21

Full Changelog: v1.0gpu_beta...v1.0gpu

v1.0gpu_beta

What's Changed

• Less host work per run: string-source check, diagnostic switch, store packing, GPU edge test, optional frame store by @jhputschke in #19
• MusicCheckpoint: a run can start from another run's state (MUSIC_2 from MUSIC_1's, A8) by @jhputschke in #20

Full Changelog: v1.0...v1.0gpu_beta

v1.0gpu_alpha

What's Changed

• fix(cuda): repair CUDA backend on the XSCAPE Fields path by @jhputschke in #1
• test(cuda): add MUSIC_CUDA_FORCE_COHERENT to validate the GB10 path by @jhputschke in #2
• Xscape cuda 3 last bits of GPU tuning by @jhputschke in #3
• Fix rk1-boundary swap gate; close 100x GPU-vs-CPU precision gap by @jhputschke in #4
• Xscape cuda cpu part improvements by @jhputschke in #5
• Add gpu tuning extra metal as standard by @jhputschke in #6
• freeze_out_surface = 0: stop without building the freeze-out surface by @jhputschke in #9
• Source term on the GPU path: skip empty fills, bin strings by transverse reach by @jhputschke in #10
• StringFind4: fail loudly on a parameter file without EndOfData by @jhputschke in #11
• Freeze-out surface: parallel over eta slices again, deterministic; store P by @jhputschke in #12
• GPU: vacuum cells at rest, non-finite W/Pi guard with counter (grid freeze in dilute regions) by @jhputschke in #13
• HydroinfoMUSIC::clean_hydro_event: release the store's memory by @jhputschke in #15
• docs/MUSIC_CPU_vs_GPU.md; README: validated against the CPU build, large-scale campaign still missing by @jhputschke in #16
• docs/: GPU port documents moved there, every reference updated; test_metal_input removed by @jhputschke in #17
• MUSIC_CUDA_SYNC selects the host wait; string source with less work per pair by @jhputschke in #18

Full Changelog: v3.1...v1.0

v0.99gpu

What's Changed

• freeze_out_surface = 0: stop without building the freeze-out surface by @jhputschke in #9
• Source term on the GPU path: skip empty fills, bin strings by transverse reach by @jhputschke in #10
• StringFind4: fail loudly on a parameter file without EndOfData by @jhputschke in #11
• Freeze-out surface: parallel over eta slices again, deterministic; store P by @jhputschke in #12
• GPU: vacuum cells at rest, non-finite W/Pi guard with counter (grid freeze in dilute regions) by @jhputschke in #13

Full Changelog: v0.96gpu...v0.99gpu

v0.97gpu

What's Changed

• freeze_out_surface = 0: stop without building the freeze-out surface by @jhputschke in #9
• Source term on the GPU path: skip empty fills, bin strings by transverse reach by @jhputschke in #10
• StringFind4: fail loudly on a parameter file without EndOfData by @jhputschke in #11
• Freeze-out surface: parallel over eta slices again, deterministic; store P by @jhputschke in #12

Full Changelog: v0.96gpu...v0.97gpu

v0.96gpu

Clean ups and minor improvements.

Full Changelog: v0.9.5gpu...v0.96gpu

v0.9.5gpu

This release is the first version of a GPU accelerated MUSIC for XSCAPE. The main_gou branch for MUSIC standalone is not optimized at the same level as the XSCAPE branch (they are different and need seperate integration/tuning). The branch XSCAPEv1 is verified with an OO test event using XSCAPE 2.1 with 3d Glauber. Due to fp32 on GPUs they are not byte identical by differences are < 1e-3 - 1e-4. This release includes further significant speed-up improvements, but before production needs a larger scale with hadronic observables validation.

l2_max_errors_vs_tau.pdf

XSCAPE GPU Metal/CUDA implementaton

Based and validated against main_gpu release v0.9gpu

First gpu Metal/Cuda acceleration of MUSIC

First gpu Metal/Cuda acceleration of MUSIC