v1.15.1

GPUCompiler v1.15.1

Diff since v1.15.0

Merged pull requests:

• Metal: extend address-space inference to integer pointer arguments (#819) (@maleadt)

v1.15.0

GPUCompiler v1.15.0

Diff since v1.14.0

Merged pull requests:

• Metal: include the debug level in the runtime slug (#817) (@maleadt)
• Add debug level to the compiler options (#818) (@maleadt)

v1.14.0

GPUCompiler v1.14.0

Diff since v1.13.3

Merged pull requests:

• Gate VectorCombinePass on can_vectorize(job) (#814) (@michel2323)
• Metal: extend address-space inference across call arguments (#816) (@maleadt)

Closed issues:

• VectorCombinePass in buildScalarOptimizerPipeline produces invalid SPIR-V on OpenCLBackend (#813)
• emit_trap removed (#815)

v1.13.3

GPUCompiler v1.13.3

Diff since v1.13.2

Merged pull requests:

• Metal: turn device-exception traps into returns to avoid GPU hangs (#810) (@maleadt)
• PTX: Fix regression in kernel metadata on LLVM 20. (#811) (@maleadt)
• Unify unreachable control flow lowering across back-ends (#812) (@maleadt)

v1.13.2

GPUCompiler v1.13.2

Diff since v1.13.1

Merged pull requests:

• Improve the pipeline with callbacks (#785) (@maleadt)
• Clean-up some FileCheck invocations. (#806) (@maleadt)
• PTX: add PTXRSqrtFastPass to fold afn 1/sqrt(x) to nvvm.rsqrt.approx (#807) (@maleadt)
• Make gpu_* runtime stubs CPU-AOT-safe via weak linkage. (#808) (@maleadt)
• Strip back-end runtime stub bodies before InternalizePass. (#809) (@maleadt)

Closed issues:

• PTX: investigate if custom NVVM Reflect pass is still needed (#547)

v1.13.1

GPUCompiler v1.13.1

Diff since v1.13.0

v1.13.0

GPUCompiler v1.13.0

Diff since v1.12.0

Merged pull requests:

• Add PTXFDivFastPass to lower fdiv fast to NVPTX approximate division (#800) (@vchuravy)
• Register deferred_codegen in GlobalJD on Julia 1.14+. (#802) (@maleadt)
• Fixes for nightly (#803) (@maleadt)
• Add a pass to apply fastmath attributes. (#804) (@maleadt)

v1.12.0

GPUCompiler v1.12.0

Diff since v1.11.1

Merged pull requests:

• Bump julia-actions/setup-julia from 2 to 3 (#786) (@dependabot[bot])
• PTX: Add support for selecting the GPU feature set. (#798) (@maleadt)

v1.11.1

GPUCompiler v1.11.1

Diff since v1.11.0

Merged pull requests:

• Revert "Fix always_inline via inlining policy override" (#797) (@maleadt)

Closed issues:

• Pass registration broken on recent release? (#789)

v1.11.0

GPUCompiler v1.11.0

Diff since v1.10.0

Merged pull requests:

• Turn instcombine into a flag (and use instsimplify otherwise) (#791) (@wsmoses)
• Adjust to upcoming sparam_vals representation change (#792) (@Keno)
• Consider assertions for runtime module (#793) (@vchuravy)
• Fix always_inline via inlining policy override (#795) (@maleadt)
• Adapt to LLVM.jl change. (#796) (@maleadt)

Closed issues:

• always_inline is broken on 1.11 (#527)