AMDGPU.jl Files

AMDGPU v2.9.0

Diff since v2.8.0

Merged pull requests:

• Support dims in sort, sort!, sortperm and sortperm! (#1033) (@luraess)
• Override sind/cosd to avoid the malloc hostcall (#1041) (@luraess)
• use device libs from AMDGPU_LLVM_Backend_jll (2) (#1044) (@simeonschaub)
• rocBLAS: route 16-bit mul! through gemm_ex (Float16 3→190 TFLOPS on MI300A) (#1049) (@gbaraldi)
• Transfer buffer ownership at conversion time instead of a per-launch locked phase (#1050) (@gbaraldi)
• Probe rocSPARSE with Libdl only (#1051) (@luraess)
• Fix MIOpen in CSCS CI and log if spurious async FFT failure (#1052) (@luraess)
• Run the AMDGPU attributor after optimization (kernel descriptors match hipcc) (#1056) (@gbaraldi)
• Fix off-by-one in indexing !range metadata (#1057) (@luraess)
• Read workgroup/grid dimensions from hidden kernargs instead of the dispatch packet (2x on short kernels) (#1058) (@gbaraldi)
• Fix typo (#1059) (@luraess)
• Support launch bounds (@roc maxthreads) and set them for static KernelAbstractions workgroups (#1060) (@gbaraldi)
• Add register-tiled GEMM and 3D diffusion stencil benchmarks (#1061) (@gbaraldi)
• Makes artifact ROCm survive a machine that already has a ROCm in its environment (#1062) (@luraess)
• Claim the bundle's vendor libraries, and drop the rocSPARSE version probe (#1063) (@luraess)
• Implement LinearAlgebra's storage-specific mul! methods (#1064) (@maleadt)
• Handle hipTENSOR 2.4: fix the version decode, reject the layouts it miscomputes (#1065) (@luraess)
• Skip Int128 in the GPUArrays test suite on LLVM 18 (#1066) (@maleadt)
• Exclude Int128 from testing on linalg/core (#1067) (@luraess)
• Guard empty-array launch in out-of-place reverse (#1069) (@shreyas-omkar)
• plan handle cache fix (#1070) (@gregordecristoforo)
• Adapt to the library-based AMDGPU_LLVM_Backend_jll 23 and LLVMDowngrader_jll 0.10 (#1071) (@maleadt)
• Adapt to LLVMDowngrader_jll 0.11 and GPUCompiler 2.8 (#1072) (@maleadt)
• Avoid boxing compiler jobs during kernel resolution (#1073) (@maleadt)
• Add Julia package details in versioninfo (#1074) (@christiangnrd)
• Define KernelState before the device code that references it (#1075) (@gbaraldi)
• Stdlib compat and remove unnecessary(?) Pkg dep (#1082) (@christiangnrd)
• Remove dead code get_artifact_library (#1086) (@luraess)
• Workaround precompile hang on Windows (#1087) (@luraess)
• Skip Int128 flaky test (#1088) (@luraess)
• rocFFT: capture plan context at creation time, not finalization (#1092) (@gregordecristoforo)
• [CI] Require 1.13 to pass (#1093) (@christiangnrd)
• Add Enzyme rules for the ROCArray constructors (#1095) (@michel2323)
• library_state: key handles on the creation context, and call the destructor that was passed (#1096) (@luraess)
• Remove quirks that never apply (#1097) (@maleadt)
• Pass the exception type separately to @gputhrow (#1098) (@maleadt)
• Stop excluding device code from coverage (#1099) (@maleadt)
• Add LinearAlgebra and Printf compat entries (#1102) (@christiangnrd)
• Skip broken MIOpen XDLOPS conv solvers on gfx90a too, and move the workaround into the test (#1103) (@luraess)
• Keep stream and rocFFT plan destructors out of task-local state (#1104) (@luraess)
• Skip prepare_state in GC finalizers (#1106) (@luraess)
• Switch back after device!(f, dev) on a fresh task (#1107) (@luraess)
• Fix hipTENSOR detection on ROCm < 7.2 and harden versioninfo() (#1108) (@gregordecristoforo)
• Julia 1.14 compatibility; CI step selection (#1109) (@maleadt)
• Run explicit frees in the owner's context (#1110) (@luraess)
• Bring back default_device! and document that device! sets the default (#1111) (@luraess)
• Fix ballot on wave64 GPUs and with GPUCompiler 2.9 (#1112) (@luraess)
• Document CI step selection (#1114) (@luraess)
• Launch args no tuple limits (#1115) (@sheevy)
• Make unsafe_wrap of host memory safe, and support wrapping Arrays (#1116) (@maleadt)
• Implement KernelInterface, and support KernelAbstractions 0.10 (#1117) (@maleadt)
• Use GPUToolbox's cooperative_wait for synchronization (#1118) (@maleadt)
• Work around RDNA2/gfx1036 bug with AMD_OPT_FLUSH preference (#1119) (@maleadt)
• Recycle the streams of finished tasks (#1120) (@maleadt)
• bump LLVMDowngrader to 0.12 (#1121) (@simeonschaub)
• docs: fix some typos (#1126) (@jeis4wpi)
• Un-break Int128 axpby! test (fixed by AMDGPU_LLVM_Backend_jll 23.1.1+3) (#1127) (@luraess)
• Enable hostcall on Windows without HSA dependency (#1128) (@ehgus)
• docs: fix typos (#1129) (@ehgus)
• Add regression test for short-circuit select miscompile (#1113) (#1131) (@stillyslalom)
• Upgrade to LLVM.jl 10 (#1132) (@maleadt)
• Upgrade to UnsafeAtomics 0.4 (#1135) (@maleadt)
• Test Int128 on Julia 1.10/1.11 again (#1139) (@luraess)
• Report kernel lookups to the compile hook instead of recompiling (#1140) (@maleadt)
• Drop the Atomix.modify! override, now that Atomix uses the device scope (#1141) (@maleadt)
• Let atomics assume coarse-grained, local memory unless they use the system scope (#1143) (@maleadt)
• Publish benchmarks when Buildkite reports them (#1144) (@maleadt)
• Use GPUArrays 12's reductions, sorting, scans and findall (#1145) (@maleadt)
• Reduce CI pressure on pull requests (#1146) (@maleadt)
• Adapt to GPUCompiler 2.13's atomics handling (#1147) (@maleadt)
• Make device min/max propagate NaNs and order signed zeros (#1148) (@maleadt)
• Allow CEnum 1.0 (#1156) (@maleadt)
• Pin nolebase readabilities plugin to ~2.18.2 to fix docs build (#1157) (@luraess)

Closed issues:

• Int128 regression in Julia 1.12 (#1002)
• sort!(x, dims=2) doesn't work (#1030)
• Float16 mul! falls through to GPUArrays' generic kernel (~65x slower than rocBLAS gemm_ex) (#1048)
• rocFFT: plan handle cache (IDLE_HANDLES) never evicts distinct-shape plans, leaking GPU memory (#1053)
• reverse of an empty ROCArray throws hipErrorInvalidValue (#1068)
• BUG: using AMDGPU hangs indefinitely during precompilation on Windows on main branch (#1083)
• ENHANCEMENT: Hostcall on Windows without HSA Runtime (#1089)
• library_state leaks library handles when a task switches device (#1094)
• Miscompile in AMDGPU_LLVM_Backend_jll 22.1.8 (AMDGPU.jl 2.7.1–2.8.0): (c1 || c2) ? a : b gives b in every lane (#1113)
• Broken module (addrspace(1) jl_libjulia_internal_handle) compiling a Union{Missing,Bool} broadcast on CI (#1149)