AMDGPU.jl

AMDGPU v1.0.5

Diff since v1.0.4

Merged pull requests:

• Avoid constructing MulAddMuls (#623) (@dkarrasch)
• Use AcceleratedKernels for sorting (#688) (@pxl-th)
• Fix FFT execution on separate Julia task (#689) (@jipolanco)

Closed issues:

• (2nd) AMDGPU does not load any more without error when installation not functional (#687)

AMDGPU v1.0.4

Diff since v1.0.3

Merged pull requests:

• Generate librocsolver.jl with Clang.jl (#617) (@amontoison)
• [rocSPARSE] Update the interface for sparse triangular solves (#682) (@amontoison)
• Fix import on non-functional machines (#686) (@pxl-th)

Closed issues:

• Unable to find correct HIP path (#671)
• AMDGPU does not load any more without error when installation not functional (#685)

AMDGPU v1.0.3

Diff since v1.0.2

Merged pull requests:

• Fix workspace size calculation for fwd conv (#678) (@pxl-th)

AMDGPU v1.0.2

Diff since v1.0.1

Merged pull requests:

• Add test for hardware atomic (#665) (@pxl-th)
• copy in constructor (#669) (@jariji)
• Prevent crash when libraries are unavailable during init (#673) (@AntonReinhard)
• Remove artifacts support & fix rocSPARSE bug (#674) (@pxl-th)
• Bump deps (#675) (@pxl-th)
• Fix strided matmul & minor cleanup (#677) (@pxl-th)

Closed issues:

• Failures when using ROCM builds that have particular type of debug info in them (both in JLL-mixed-mode and in system-ROCM mode), e.g. on Arch Linux (#620)
• Multiplying two views (#676)

AMDGPU v1.0.1

Diff since v1.0.0

Merged pull requests:

• Fix feature concatenation (#664) (@pxl-th)

Closed issues:

• Running AMDGPU for MultiGPU using an Array of Pointers (#662)

AMDGPU v1.0.0

Diff since v0.9.6

Merged pull requests:

• Bump LLVM to 8 (#649) (@pxl-th)
• Run TLS tests on Navi 3 again (#650) (@pxl-th)
• Re-enable rocSPARSE & rocSOLVER tests (#652) (@pxl-th)
• Add JULIA_AMDGPU_LAUNCH_BLOCKING option (#655) (@pxl-th)
• Move error checks to ccalls (#656) (@pxl-th)
• Fix context switching & minor optimizations (#658) (@pxl-th)
• Add LLVM IR attributes (target cpu and target features) (#659) (@pxl-th)
• Add ormqr test (#660) (@pxl-th)
• Enable hardware atomics & link device libraries early (#661) (@pxl-th)

Closed issues:

• Allow preserving HSA executable (#97)
• Add workaround for atomics on floating point values (#151)
• CI against system ROCM 5 (#238)
• rocSPARSE tests hang Navi 3 (#597)
• Multithreading code hangs (#606)
• Failing tests with MI250X and Rocm 6.1.1 (#648)

AMDGPU v0.9.6

Diff since v0.9.5

Merged pull requests:

• Properly allocate workspace for conv algo search (#643) (@pxl-th)
• Implement Base.dataids(A::ROCArray) (#644) (@ahojukka5)
• Do full GC if pressure is high (#645) (@pxl-th)
• Cache FFT handles & more aggressive maybe_collect (#646) (@pxl-th)
• Update LA.dot wrappers (#647) (@pxl-th)

AMDGPU v0.9.5

Diff since v0.9.4

Merged pull requests:

• Print kernel info that requires hostcall (#642) (@pxl-th)

AMDGPU v0.9.4

Diff since v0.9.3

Merged pull requests:

• Add ctor with buffer type (#641) (@pxl-th)

AMDGPU v0.9.3

Diff since v0.9.2

Merged pull requests:

• Update rocFFT (#640) (@pxl-th)

AMDGPU v0.9.2

Diff since v0.9.1

Merged pull requests:

• Bump LLVM to 7 (#639) (@pxl-th)

AMDGPU v0.9.1

Diff since v0.9.0

Merged pull requests:

• Remove NNlib from tests (#636) (@pxl-th)
• Bring back batchnorm (#637) (@pxl-th)

AMDGPU v0.9.0

Diff since v0.8.11

Merged pull requests:

• Bump julia-actions/setup-julia from 1 to 2 (#612) (@dependabot[bot])
• Update wrappers.jl (#616) (@amontoison)
• Fix blas tests (#627) (@pxl-th)
• Rework hard memory limit & add maybe collect mechanism (#628) (@pxl-th)
• Fix softmax tests (#629) (@pxl-th)
• Remove MIOpen softmax & batchnorm (#630) (@pxl-th)
• Cleanup (#631) (@pxl-th)
• Keep track of stream usage for arrays (#633) (@pxl-th)
• Allow disabling eager GC (#634) (@pxl-th)
• Add 1.11 support (#635) (@pxl-th)

Closed issues:

• @atomic is slow within AMDGPU.jl (#569)
• Rework hard memory limit with maxSize option (#611)
• i get error when i call AMDGPU.rocBLAS.gemm_strided_batched! (#615)
• error: Opaque pointers are only supported in -opaque-pointers mode (#624)
• Adopt maybe collect garbage collection scheme similar to CUDA.jl (#625)
• GPUCompiler.CodeCache not defined (#632)

AMDGPU v0.8.11

Diff since v0.8.10

Merged pull requests:

• Default to hardware floating-point atomics. (#604) (@pxl-th)
• Mark ccalls as gc safe (#605) (@pxl-th)

AMDGPU v0.8.10

Diff since v0.8.9

Merged pull requests:

• Run GPU-less CI on Buildkite (#599) (@pxl-th)
• Update profiling doc (#600) (@luraess)
• Add more BLAS tests (#601) (@pxl-th)
• Update BLAS & add more BLAS tests (#602) (@pxl-th)

AMDGPU v0.8.9

Diff since v0.8.8

Merged pull requests:

• Update docs (#593) (@pxl-th)
• Use default flag during host malloc (#594) (@pxl-th)
• Cleanup & bump deps (#595) (@pxl-th)
• Disable rocSPARSE tests (#596) (@pxl-th)
• Cleanup GitHub CI (#598) (@pxl-th)

Closed issues:

• ERROR: HIPError(code hipErrorOutOfMemory, out of memory) (#591)

AMDGPU v0.8.8

Diff since v0.8.7

Merged pull requests:

• Add warp-level primitives (#592) (@pxl-th)

AMDGPU v0.8.7

Diff since v0.8.6

Merged pull requests:

• Bump actions/cache from 3 to 4 (#586) (@dependabot[bot])
• Bump codecov/codecov-action from 3 to 4 (#587) (@dependabot[bot])
• Use adjoint instead of tranpose when returning V in SVD (#588) (@wbernoudy)
• Allow any AMD GPU in CI (#589) (@pxl-th)
• Add support for logical indexing (#590) (@pxl-th)

AMDGPU v0.8.6

Diff since v0.8.5

Merged pull requests:

• Expose alpha and beta parameters in LinearAlgebra.mul! (#585) (@wbernoudy)

AMDGPU v0.8.5

Diff since v0.8.4

Merged pull requests:

• Fix kernel launch inference (#582) (@pxl-th)
• Expose rocSOLVER eigensolver and SVD methods (#584) (@wbernoudy)