v0.4.9

AMDGPU v0.4.9

Diff since v0.4.8

Closed issues:

• State of queues and streams (#337)
• rocBLAS: Remove old hand-wrapped code (#384)
• HSA memory fault upon switching from default device on multi-GPU node (#385)
• Test fail locally with AssertionError: AMDGPU.Runtime.LOGGING_STATIC_ENABLED (#399)

Merged pull requests:

• Switch to task-focused synchronization model (#374) (@jpsamaroo)
• Use broadcast instead of copies to initialize mapreduce buffers. (#390) (@maleadt)
• tests: Skip logging tests if disabled (#391) (@jpsamaroo)
• Add blas wrappers for triangular matrix mul / div (#392) (@pxl-th)
• Simplify signal pooling (#393) (@pxl-th)
• Adapt to GPUCompiler 0.18 (#394) (@pxl-th)
• Reduce memory usage (#395) (@pxl-th)
• Add support for KernelAbstraction 0.9 (#398) (@vchuravy)
• Update to GPUCompiler 0.19 & LLVM 5 (#407) (@pxl-th)
• Fix compiler timespan logging (#408) (@pxl-th)
• rocBLAS: define highlevel dot, gemm, axpy functions for FP16 (#409) (@pxl-th)
• Add KernelAbstractions.jl unsafe_free! (#410) (@pxl-th)

v0.4.8

AMDGPU v0.4.8

Diff since v0.4.7

Merged pull requests:

• ROCSignal: Pool signals in ctor (#369) (@jpsamaroo)
• Reduce allocations (#376) (@pxl-th)
• Report and exit on memory fault (#379) (@jpsamaroo)
• versioninfo: Indicate if using JLLs or System (#381) (@jpsamaroo)
• ROCSignal: Disable IPC by default (#383) (@jpsamaroo)

v0.4.7

AMDGPU v0.4.7

Diff since v0.4.6

Closed issues:

• Segfault on ROCm 5.0 (#239)

Merged pull requests:

• Add more MIOpen functionality (#370) (@pxl-th)
• Store HSA.Signal instead of Ref{HSA.Signal} in ROCSignal (#375) (@pxl-th)

v0.4.6

AMDGPU v0.4.6

Diff since v0.4.5

Closed issues:

• Implement occupancy API (#271)
• getinfo should determine the Ref output container automatically (#273)

Merged pull requests:

• Add timespan logging via TimespanLogging.jl (#263) (@jpsamaroo)
• Add occupancy API and groupsize tuning (#326) (@jpsamaroo)
• Reduce signal wait allocations (#361) (@jpsamaroo)
• Add more intrinsics, enable always_inline (#362) (@jpsamaroo)
• Simplify math intrinsics (#363) (@pxl-th)
• Implement unified getinfo interface (#364) (@jpsamaroo)
• Assorted fixes (#365) (@jpsamaroo)
• Add memory allocation limiters (#366) (@jpsamaroo)
• Specify return types for getinfo calls (#368) (@pxl-th)

v0.4.5

AMDGPU v0.4.5

Diff since v0.4.4

Closed issues:

• Mem.alloc: Allow using hipMalloc to service allocations (#286)
• rocBLAS GEMM ignores @view (#319)
• sincospi intrinsic is broken (#334)
• #jps/dev segfaults on MI250x (#340)
• method ambiguity in rand! (#343)
• Add function or macro for AMDGPU.jl equivalent to CUDA.CuDynamicSharedArray (and CUDA.CuStaticSharedArray) (#347)
• Free KernelState in finalizer (#352)
• --check-bounds=no is broken on Julia 1.9.0-beta3 (#354)

Merged pull requests:

• Fix GEMM (regular & batched) and support batched GEMM for 3D array (#318) (@pxl-th)
• Add MIOpen (#320) (@pxl-th)
• Add support for 2D * 3D batched GEMM (#321) (@pxl-th)
• Support NNlib batched gemm format (#322) (@pxl-th)
• Add pointer() method for ROCArray and some library tests (#323) (@torrance)
• Fix double unsafe_free calls (#324) (@jpsamaroo)
• Mem: Allow using hipMalloc/hipFree for allocations (#325) (@jpsamaroo)
• Cast to Ptr before checking NULL pointer (#328) (@torrance)
• Resize! support (#333) (@matinraayai)
• Add sincos/sincospi/frexp/ldexp intrinsics (#336) (@jpsamaroo)
• Add local memory allocation helpers (#348) (@jpsamaroo)
• Add GPUCompiler 0.17 to compat (#349) (@jpsamaroo)
• Preserve UInt32 in indexing intrinsics (#351) (@pxl-th)
• Fix unsafe_free! not actually freeing (#353) (@jpsamaroo)
• Don't sync on default HIP stream every time (#356) (@pxl-th)
• Make alignment generated (#358) (@pxl-th)
• tests: Properly unwrap Distributed exceptions (#359) (@jpsamaroo)

v0.4.4

AMDGPU v0.4.4

Diff since v0.4.3

Closed issues:

• Repetetive AMDGPU.ones calls crash runtime (#299)
• Add AMDGPU.jl equivalent to CUDA.CuDynamicSharedArray (and CUDA.CuStaticSharedArray) (#304)
• Segfault with basic kernel from AMDGPU.jl doc on LUMI (#308)
• ROC kernel faulting upon having AMDGPU and CUDA loaded (#312)
• AMDGPU.rand failing to create a ROCArray (#315)

Merged pull requests:

• Remove waiter and error monitor threads (#306) (@pxl-th)
• Update bindeps search path (#307) (@luraess)
• Prioritise ENV var to use or not artifacts (#310) (@luraess)
• Add dynamic local memory support (#311) (@jpsamaroo)
• random: Load definitions without rocRAND (#316) (@jpsamaroo)

v0.4.3

AMDGPU v0.4.3

Diff since v0.4.2

Closed issues:

• Queue selection test fail (#274)

Merged pull requests:

• Add device quirks from CUDA.jl, enhance at-rocprintf (#269) (@jpsamaroo)
• Use an optimized norm function for ROCBLASArray (#282) (@amontoison)
• Add rocBLAS_jll and rocSPARSE_jll deps (#284) (@jpsamaroo)
• active_kernels: Use WeakKeyDict (#285) (@jpsamaroo)
• CI: Add gfx90a to more jobs (#289) (@jpsamaroo)
• build: Remove build step, run at toplevel (#290) (@jpsamaroo)
• Mapreducedim support for AnyROCArray (#291) (@matinraayai)
• Parallelize tests (#293) (@jpsamaroo)
• Fix precompilation (#294) (@pxl-th)
• Do not rethrow EOF (#296) (@pxl-th)
• Use correct queue for kernels (#297) (@pxl-th)
• Implement kernel hashing system (#302) (@jpsamaroo)

v0.4.2

AMDGPU v0.4.2

Diff since v0.4.1

Closed issues:

• build failure on Julia 1.8.1 (#278)

Merged pull requests:

• Mem: Retry failing allocations (#251) (@jpsamaroo)
• Add device-to-device unsafe_copy3d test (#260) (@luraess)
• Fix allocation retry mechanism, add slow allocation fallback (#262) (@jpsamaroo)
• Run wavefront tests with detected wavefrontsize (#264) (@torrance)
• During HostCall, ensure device has finished using buffers before freeing (#266) (@torrance)
• Expand fft tests (#267) (@torrance)
• Remove code that duplicates AbstractFFTs; add tests for casting (#268) (@torrance)
• Don't embed the method table in the AST (#276) (@jpsamaroo)
• deps: Don't access is_available unless using succeeds (#279) (@jpsamaroo)
• device: Add ROCDevice() ctor (#280) (@jpsamaroo)

v0.4.1

AMDGPU v0.4.1

Diff since v0.4.0

Closed issues:

• Add option to disable automatic mark/wait of specific arrays (#126)
• Limit multi-dimensional groupsize properly (#150)
• Optimize kernarg allocations in kernel construction (#247)
• Add priority kwarg to ROCQueue ctor (#256)

Merged pull requests:

• Add BackToCPU struct to reduce 'view' allocations (#246) (@pxl-th)
• LB GPUCompiler to 0.16.2 (#248) (@jpsamaroo)
• Optimize kernel setup and launch (#249) (@jpsamaroo)
• launch: Fix groupsize dimension check (#250) (@jpsamaroo)
• device: Add device_id method (#253) (@jpsamaroo)
• Re-export indexing intrinsics (#254) (@jpsamaroo)
• CI: Switch GHA to 1.7 release (#257) (@jpsamaroo)
• queue: Allow setting priority from ctor (#258) (@jpsamaroo)
• math: Make signbit return Bool (#259) (@jpsamaroo)

v0.4.0

AMDGPU v0.4.0

Diff since v0.3.7

Closed issues:

• Use atomics/locking in hostcall (#4)
• Bump Setfield compat (#242)

Merged pull requests:

• Remove launch export (#232) (@matinraayai)
• Remove indirection layer, use modules (#240) (@jpsamaroo)
• Update Setfield compat (#243) (@luraess)