Skip to content

v2.11.0

Latest

Choose a tag to compare

@github-actions github-actions released this 01 Nov 21:17
de02505

AMDGPU v2.11.0

Diff since v2.10.0

Merged pull requests:

  • Record more benchmark data, and publish it to juliagpu.org/perf (#1193) (@maleadt)
  • Comment on pull requests from the JuliaGPU benchmark history (#1194) (@maleadt)
  • Redirect the benchmark page to juliagpu.org/perf (#1195) (@maleadt)
  • Cache the launch limits queried by KernelAbstractions (#1196) (@maleadt)
  • Reduce first-launch and first-free compile time (#1197) (@maleadt)
  • Accept callable objects as kernels (#1198) (@maleadt)
  • List only deployed folders in the docs version picker [only docs] (#1199) (@luraess)
  • Bump benchmark-action, to flag benchmarks that start allocating (#1200) (@maleadt)
  • Only collect garbage before benchmark samples that left some on the GPU (#1201) (@maleadt)
  • Define the KernelAbstractions methods in AMDGPU rather than an extension (#1202) (@maleadt)
  • Build the documentation with Julia 1.13 (#1203) (@maleadt)
  • Avoid boxed captured variables (#1204) (@maleadt)
  • Define device-only functions as overlays, so sysimages load (#1205) (@maleadt)
  • Precompile displaying arrays, and link kernels without a closure (#1206) (@maleadt)
  • Create memory pools with an export handle type, making ROCArray memory IPC-capable (GPU-aware MPI) (#1208) (@luraess)
  • Add AMDGPU.enable_synchronization! to opt out of the implicit stream synchronization of an array (#1209) (@luraess)

Closed issues:

  • Create the device memory pool with an exportable handle type (GPU-aware MPI / IPC on ROCm >= 7.1) (#1207)