Code

Native GPU support#65

Code

Open

MilesCranmer wants to merge 52 commits intomasterSymbolicML/DynamicExpressions.jl:masterfrom cuda-5SymbolicML/DynamicExpressions.jl:cuda-5Copy head branch name to clipboard

+603-1

Commits

Commits on Jan 14, 2024

• Initialize CUDA extension

  MilesCranmer

  committed
  9b8872c

  Copy full SHA for 9b8872c

• Fix output

  MilesCranmer

  committed
  666708e

  Copy full SHA for 666708e

• Switch to GPUArrays rather than CUDA

  MilesCranmer

  committed
  4484d18

  Copy full SHA for 4484d18

• Switch to GPUArrays

  MilesCranmer

  committed
  644c3b8

  Copy full SHA for 644c3b8

Commits on Jan 15, 2024

• Add file for converting to arrays

  MilesCranmer

  committed
  0bfd8ee

  Copy full SHA for 0bfd8ee

• Add GPU kernel for evaluation

  MilesCranmer

  committed
  5c90896

  Copy full SHA for 5c90896

• Allow submitting multiple trees at once

  MilesCranmer

  committed
  57fff36

  Copy full SHA for 57fff36

• Working GPU kernel for expression evaluation!

  MilesCranmer

  committed
  81529af

  Copy full SHA for 81529af

• Formatting

  MilesCranmer

  committed
  07dec6b

  Copy full SHA for 07dec6b

• Remove GPUArrays extension as redundant

  MilesCranmer

  committed
  c19f37f

  Copy full SHA for c19f37f

• Remove unneccessary sync call

  MilesCranmer

  committed
  ade1613

  Copy full SHA for ade1613

• Reduce number of allocations

  MilesCranmer

  committed
  53b42ce

  Copy full SHA for 53b42ce

• Greatly improve allocations

  MilesCranmer

  committed
  cdfcec2

  Copy full SHA for cdfcec2

• More memory allocation improvements

  MilesCranmer

  committed
  20a8f38

  Copy full SHA for 20a8f38

• Allow user to pass preallocated buffer

  MilesCranmer

  committed
  6bbff73

  Copy full SHA for 6bbff73

• Try to fix unbound type parameter

  MilesCranmer

  committed
  03d2d24

  Copy full SHA for 03d2d24

• Expand CUDA compat to 4

  MilesCranmer

  committed
  2e128bb

  Copy full SHA for 2e128bb

Commits on Jan 16, 2024

• Clean up formatting

  MilesCranmer

  committed
  da31b87

  Copy full SHA for da31b87

• Reduce allocations in as_array

  MilesCranmer

  committed
  4fdc025

  Copy full SHA for 4fdc025

• Fix indexing

  MilesCranmer

  committed
  c64a6cb

  Copy full SHA for c64a6cb

• Compatible CUDA kernel with empty operators

  MilesCranmer

  committed
  88ee6cf

  Copy full SHA for 88ee6cf

Commits on Jan 28, 2024

• Merge branch 'bump-alloc3' into cuda-5

  MilesCranmer

  authored
  f1f2221

  Copy full SHA for f1f2221

• Formatting

  MilesCranmer

  committed
  243bdda

  Copy full SHA for 243bdda

Commits on Feb 3, 2024

• Merge branch 'master' into cuda-5

  MilesCranmer

  committed
  6b0337c

  Copy full SHA for 6b0337c

Commits on Feb 25, 2024

• Ensure we always check root during array construction

  MilesCranmer

  committed
  b1aca68

  Copy full SHA for b1aca68

• Mechanism to unittest cuda kernel

  MilesCranmer

  committed
  9372150

  Copy full SHA for 9372150

• Merge branch 'master' into cuda-5

  MilesCranmer

  committed
  5241645

  Copy full SHA for 5241645

• Fix unbound type issues

  MilesCranmer

  committed
  5befede

  Copy full SHA for 5befede

• Start working on CPU-based unittests

  MilesCranmer

  committed
  74094e3

  Copy full SHA for 74094e3

• Fix indexing bugs in GPU kernels

  MilesCranmer

  committed
  9d7500c

  Copy full SHA for 9d7500c

• Fix unsafe sin in CUDA kernel tests

  MilesCranmer

  committed
  f6155ba

  Copy full SHA for f6155ba

• Fix MaybeCuArray definition

  MilesCranmer

  committed
  a7d7e79

  Copy full SHA for a7d7e79

• Test buffered-evaluation of CUDA kernels

  MilesCranmer

  committed
  46a6b48

  Copy full SHA for 46a6b48

• Only test CUDA on 1.9+

  MilesCranmer

  committed
  7ceacc0

  Copy full SHA for 7ceacc0

• Use vectors for vector input; tuple for tuple

  MilesCranmer

  committed
  08df9a3

  Copy full SHA for 08df9a3

• Fix unbound args in multi-tree eval

  MilesCranmer

  committed
  8b4cbf3

  Copy full SHA for 8b4cbf3

Commits on Feb 26, 2024

• Add test for no buffer recreation

  MilesCranmer

  committed
  676ad86

  Copy full SHA for 676ad86

• Start switching to KernelAbstractions.jl

  MilesCranmer

  committed
  c1e579a

  Copy full SHA for c1e579a

• Revert "Start switching to KernelAbstractions.jl"

  Show description for cb2d055
  MilesCranmer

  committed
  cb2d055

  Copy full SHA for cb2d055

Commits on Jul 17, 2024

• Merge branch 'master' into cuda-5

  MilesCranmer

  committed
  db94de2

  Copy full SHA for db94de2

• fix: as array imports

  MilesCranmer

  committed
  bfa9148

  Copy full SHA for bfa9148

• hack: try implementing cooperative group

  MilesCranmer

  committed
  9f49619

  Copy full SHA for 9f49619

Commits on Nov 29, 2024

• refactor: clean up extensions

  MilesCranmer

  committed
  1422e36

  Copy full SHA for 1422e36

Commits on Dec 15, 2024

• Merge tag 'v1.8.0' into cuda-5

  Show description for 6c01ded
  MilesCranmer

  committed
  6c01ded

  Copy full SHA for 6c01ded

• Revert "hack: try implementing cooperative group"

  Show description for 74afd72
  MilesCranmer

  committed
  74afd72

  Copy full SHA for 74afd72

• refactor: clean up map operations

  MilesCranmer

  committed
  38c589b

  Copy full SHA for 38c589b

• fix: stabilize CUDA ext

  MilesCranmer

  committed
  4cce551

  Copy full SHA for 4cce551

• test: modularize CUDA test

  MilesCranmer

  committed
  b7ecdd3

  Copy full SHA for b7ecdd3

• fix: potential type instability in as_array

  MilesCranmer

  committed
  8fe652c

  Copy full SHA for 8fe652c

Commits on Dec 16, 2024

• refactor: as array into modular functions

  MilesCranmer

  committed
  1d06bd0

  Copy full SHA for 1d06bd0

• refactor: ensure @inbounds for gpu kernel

  MilesCranmer

  committed
  74f56d4

  Copy full SHA for 74f56d4

• refactor: avoid SubArray in CUDA kernel

  MilesCranmer

  committed
  3da1d38

  Copy full SHA for 3da1d38