v1.3.6

AMDGPU v1.3.6

Diff since v1.3.5

Merged pull requests:

• Fix return value of find_roc_path when ROCM_PATH is set (#798) (@giordano)

Closed issues:

• ROCM_PATH detected is actually ROCM_PATH/lib (#797)

v1.3.5

AMDGPU v1.3.5

Diff since v1.3.4

Merged pull requests:

• Remove overload for 1.12-rc1 and higher (#784) (@pxl-th)
• Fix global hostcall detection (#787) (@pxl-th)
• Refactor initialization logic to allow for automatic detection in Fedora (#788) (@gbaraldi)
• Fix printf on 1.11+ (#789) (@pxl-th)
• follow up #788, improve docs (#791) (@simeonschaub)
• Remove unsupported assertions stuff (#792) (@pxl-th)
• actually run batchnorm tests (#793) (@simeonschaub)

Closed issues:

• Force matching artifact versions and LLVM version (#309)
• Device printing is broken on Julia 1.11 (#551)
• Trying to import AMDGPU fails with (an lvm?) error (#579)
• reduce(f, A) fails to compile if Julia is started with --check-bounds=no (#581)
• Tests not passing on MI300A (#779)
• incorrect first value for simple kernel on 1.12 (#780)
• @rocprintf produces memory access fault in 1.12 (#781)
• to_power_type does not exist any more in Base (#782)

v1.3.4

AMDGPU v1.3.4

Diff since v1.3.3

Bug fixes

• sync_workgroup was lacking memory barrier semantics. They have been updated to match the expectation coming from HIP.

Merged pull requests:

• Update indexing.jl (#771) (@amontoison)
• Add memory barrier semantics to sync_workgroup (#783) (@vchuravy)

Closed issues:

• collectinvokes! error on julia 1.12.0-beta3 (#768)

v1.2.8

AMDGPU v1.2.8

Diff since v1.2.7

Bug fixes

• sync_workgroup was lacking memory barrier semantics. They have been updated to match the expectation coming from HIP.

This release has been identified as a backport.
Automated changelogs for backports tend to be wildly incorrect.
Therefore, the list of issues and pull requests is hidden.

v1.3.3

AMDGPU v1.3.3

Diff since v1.3.2

Merged pull requests:

• Export more types (#764) (@amontoison)
• HIP fix for unsafe_copy3d (#769) (@luraess)
• Include AcceleratedKernels-0.4 (#772) (@anicusan)
• Add code bloc to issue template (#773) (@luraess)
• Improve docs (#774) (@luraess)

Closed issues:

• unsafe_copy3d! requires 2^4 alignment (#330)
• ROCclr segfault when running Julia with threads (#770)

v1.3.2

AMDGPU v1.3.2

Diff since v1.3.1

Merged pull requests:

• Fixup image alignment (#760) (@luraess)
• Fix fast min/max on 1.12 (#761) (@pxl-th)
• Add Performance Tips to docs (#762) (@pxl-th)

Closed issues:

• Compilation issues with Flux.softmax() and Julia v1.12.0-beta2 (#756)

v1.3.1

AMDGPU v1.3.1

Diff since v1.3.0

Merged pull requests:

• Fix doc build CI (#757) (@luraess)
• Improve docs (#758) (@pxl-th)
• Improve docs (#759) (@pxl-th)

v1.3.0

AMDGPU v1.3.0

What's new:

• Julia 1.12 support.
• MI300X support.
• Use newer artifact device libraries.
• Update rocBLAS, rocFFT, rocSOLVER, rocSPARSE wrappers.

Diff since v1.2.7

Merged pull requests:

• Regenerate wrappers for ROCm v6.3.3 (#747) (@amontoison)
• Fix errors with istriu and istril (#750) (@amontoison)
• Initial 1.12 enablement (#751) (@pxl-th)
• More 1.12 enablement (#752) (@pxl-th)
• Test doc build on CI machine (#753) (@luraess)

Closed issues:

• MI300X (gfx942) support for broadcast operations (#621)
• multiplication of small matrices errors due to scalar indexing (#712)
• AMDGPU works on normal terminal but crashes on VS Code (#749)

v1.2.7

AMDGPU v1.2.7

Diff since v1.2.6

Merged pull requests:

• Bump deps (#748) (@pxl-th)

v1.2.6

AMDGPU v1.2.6

Diff since v1.2.5

Merged pull requests:

• Update Project.toml (#744) (@miniskar)
• Restrict GPUCompiler < 1.3 (#745) (@pxl-th)