v3.6.4

CUDA v3.6.4

Diff since v3.6.3

Closed issues:

• Artifacts.toml has bad git-tree-sha1 values (#1309)

Merged pull requests:

• Fix artifact-related issues (#1310) (@maleadt)

v3.6.3

CUDA v3.6.3

Diff since v3.6.2

Closed issues:

• CUDA.@atomic deadlocks when overwriting NaN (#1299)
• Unreasonablely slow copy kernel (#1301)
• Passing a LogicalIndex(::CuArray) fails (#1304)

Merged pull requests:

• Allow sorting of tuples of numbers (#1196) (@mcabbott)
• Use === for generic atomic updates with compare-and-swap (#1300) (@guyvdbroeck)
• Update manifest (#1302) (@github-actions[bot])
• Store the array length next to its dimensions. (#1303) (@maleadt)
• Disallow calling CUDA device array intrinsics on the host. (#1305) (@maleadt)
• Support logical indexing with CPU sources. (#1306) (@maleadt)
• Activate a context when calling device!. (#1307) (@maleadt)

v3.6.2

CUDA v3.6.2

Diff since v3.6.1

Closed issues:

• Norm of complex-typed CuArray is not real (#1290)
• Calling @show on Symmetric of a CuArray triggers Scalar Indexing (#1294)
• CUSPARSE Error when solving a linear system (#1296)

Merged pull requests:

• Correctly handle missing cached_memory. (#1295) (@maleadt)
• Update manifest (#1297) (@github-actions[bot])

v3.6.1

CUDA v3.6.1

Diff since v3.6.0

Closed issues:

• reduce_block error on Complex type (#1289)
• cudnn_cnn_infer64_8 could not be laoded (#1291)
• Support to find the first k eigenvalues of a sparse matrix (#1292)

Merged pull requests:

• Bump CUDNN artifacts (#1293) (@maleadt)

v3.6.0

CUDA v3.6.0

Diff since v3.5.0

Closed issues:

• Conversion issue (#157)
• Extend new RNG to Complex numbers & normal distributions (#726)
• Fatal errors during sorting tests (#916)
• deepcopy failing (#1202)
• Kernel compilation fails when specifying shared memory array size as a tuple consisting of block dimension and kernel argument (#1205)
• ERROR: LoadError: The artifact at C:\Users\name.julia\artifacts\58bd87695e9ccdb508cb38be1ab717315ecc9152 is empty. (#1209)
• InvalidIRError when displaying a model which is on the GPU (#1212)
• CUDA.jl tries to load CUDA compat loaded via jll even though system package is installed (#1216)
• Synchronizing over blocks (#1220)
• assignment changes random seed (#1226)
• accumulate gives wrong answer when init != 0 (#1227)
• Generic dot kernel: use multiple kernels instead of atomics (#1244)
• integer division error creating CuVector of missing and nothing (#1251)
• unsupported dynamic function invocation with union type of more than 2 elements (#1252)
• three CUDA.@atomic in a row result in out-of-bounds error (#1254)
• Float16 CAS cannot use atom.cas.b16.global on sm_61 (#1258)
• cu(::SVector) gives SVector, cu(::MVector) gives CuArray (#1262)
• Get back unsafe_copyto!methods for unified<-unified and unified<->device (#1263)
• Passing and using a FFT plan in a CUDA kernel seems impossible (#1266)
• Inplace Complex FFT and Threads (#1268)
• sort returns nothing (#1270)
• Release a new version (#1276)
• __init_driver__ not called in 3.5 (#1280)
• Shared memory does not support isbits unions. (#1281)
• NVIDIA Nsight Systems and CUDA.@profile error (#1282)
• nvprof with using CUDA crashes julia (#1283)

Merged pull requests:

• Addition over CuSparseMatrix (#1195) (@yuehhua)
• [CUSOLVER] Add ordering functions (#1198) (@amontoison)
• Correctly handle multi-GPU instances with NVML. (#1199) (@maleadt)
• CI improvements. (#1200) (@maleadt)
• fix FFT workarea typo leading to memory corruption (#1204) (@marius311)
• Update manifest (#1206) (@github-actions[bot])
• Minor improvements for library wrappers (#1207) (@maleadt)
• Various small improvements (#1210) (@maleadt)
• Extend CuDeviceArray ctors for mixed-int indices. (#1211) (@maleadt)
• Deprecate non-blocking sync, and always call the synchronization API. (#1213) (@maleadt)
• Generic CUSPARSE: use the index arguments. (#1214) (@maleadt)
• Add bitonic sort implementation (#1217) (@xaellison)
• Update manifest (#1218) (@github-actions[bot])
• Reverted deepcopy, added test (#1221) (@birkmichael)
• Use broadcast instead of copies to initialize mapreduce buffers. (#1223) (@maleadt)
• Remove some unneeded Base module prefixes. (#1224) (@maleadt)
• Update manifest (#1225) (@github-actions[bot])
• Cherry-picked improvements (#1228) (@maleadt)
• Update introduction.jl (#1232) (@aramirezreyes)
• Update manifest (#1233) (@github-actions[bot])
• Fix SpMV for CUDA 11.5 (#1234) (@amontoison)
• Add support for randn and randexp. (#1236) (@maleadt)
• Avoid double-initializing partial accumulate results. (#1237) (@maleadt)
• Fix cuTENSOR contractions not working for FP16 inputs (#1238) (@thomasfaingnaert)
• Bump CUTENSOR and fix on CUDA 11.5 (#1239) (@maleadt)
• Support dot product on GPU between CuArrays with inconsistent eltypes (#1240) (@findmyway)
• Update manifest (#1241) (@github-actions[bot])
• Optimize CUTENSOR contraction. (#1243) (@maleadt)
• Don't use nondeterministic atomics in dot when requested. (#1245) (@maleadt)
• Remove CUBLAS decomposition tests without pivoting. (#1246) (@maleadt)
• Update manifest (#1247) (@github-actions[bot])
• wrap CUBLAS spmv and spr (#1248) (@bjarthur)
• CompatHelper: bump compat for "SpecialFunctions" to "2" (#1249) (@github-actions[bot])
• Update manifest (#1250) (@github-actions[bot])
• Store array offset as elements to fix all-singleton case. (#1255) (@maleadt)
• Update CUDA to 11.5 Update 1. (#1256) (@maleadt)
• Use Base functionality for iteration Union type components. (#1257) (@maleadt)
• Bump CI to Julia 1.7. (#1260) (@maleadt)
• Update manifest (#1261) (@github-actions[bot])
• Use CUDA APIs for unoptimized copies. (#1265) (@maleadt)
• Bump CUDNN to 8.3.1, enable CUDA 11.5 by default. (#1267) (@maleadt)
• Adding stream update for inplace complex FFT (#1269) (@ovanvincq)
• Fix sort! return type. (#1272) (@maleadt)
• Add const keyword to type aliases declarations. (#1273) (@eliascarv)
• Update manifest (#1274) (@github-actions[bot])
• Avoid eager expansion of CUDA_compat artifact string. (#1275) (@maleadt)
• Allow copies between unified arrays in different contexts. (#1277) (@maleadt)
• fix zeros and ones for user defined types (#1278) (@GiggleLiu)
• Make CUDNN depend on CUBLAS. (#1279) (@maleadt)
• Update manifest (#1286) (@github-actions[bot])
• Restore call to init_driver. (#1287) (@maleadt)
• Improvements for isbits union shared memory (#1288) (@maleadt)

v3.5.0

CUDA v3.5.0

Diff since v3.4.2

Closed issues:

• Illegal memory access on 3.3 (#975)
• Forward compatibility (#1071)
• ambiguous sparse constructor (#1088)
• Map reduce with float 16 (#1124)
• Allow invalid GPU pointers not allowed in unsafe_wrap (#1125)
• Scalar Indexing error in the Introduction docs (#1127)
• stackoverflow when printing a custom subtype of AbstractCuSparseMatrix (#1128)
• missing rand methods (#1138)
• Error mapreducing over a 0 dimensional array (#1141)
• seed! is not thread safe (#1158)
• Simplify Int32-based indices (#1160)
• Concatenating a scalar to a CuArray gives an Array (#1162)
• Calling byte_perm with Int32 values inserts sign checks (#1165)
• sum! does not compile for large arrays (#1169)
• Same random sequence on GPU and CPU? (#1170)
• Specifying eltype and buffer type when adapting to CuArray? (#1171)
• Inefficient lop3.lut instructions generated (#1172)
• Writing temporary PTX files can fail (#1173)
• Switching devices doesn't switch the REPL's output task (#1175)
• GC is not working for CuSparseMatrixCSR (#1178)
• sparse*dense operations shouldn't drop sparseness (#1188)
• Raises illegal memory access error randomly (#1189)

Merged pull requests:

• CI fixes (#950) (@maleadt)
• implement sparse (#1093) (@CarloLucibello)
• Use the kernel state object to pass the exception flag location. (#1110) (@maleadt)
• Update manifest (#1123) (@github-actions[bot])
• Improve show methods in sparse GPU arrays. (#1129) (@maleadt)
• Use warp intrinsics for a wider range of reductions. (#1130) (@maleadt)
• Support wrapping a host buffer with a CuArray (#1131) (@maleadt)
• support transpose CSC to CUDA CSR (#1132) (@Roger-luo)
• Small improvements to discovery of local toolkits. (#1134) (@maleadt)
• Rework device and context getters. (#1135) (@maleadt)
• Avoid memory operations during graph capture. (#1137) (@maleadt)
• Streamline the random number interface. (#1146) (@maleadt)
• Native device synchronization (#1147) (@maleadt)
• support interpret(reshape) (#1149) (@Roger-luo)
• add a gitignore (#1150) (@Roger-luo)
• Fix normalize on complex number (#1151) (@maleadt)
• Addition and multiplication over cuarray and cusparse (#1152) (@maleadt)
• Preserve Int32 hardware indices (#1153) (@maleadt)
• remove mutable to make device sparse type bitstype (#1154) (@Roger-luo)
• Update manifest (#1155) (@github-actions[bot])
• CompatHelper: bump compat for "BFloat16s" to "0.2" (#1156) (@github-actions[bot])
• Perform actual synchronization API calls when we need the memory (#1157) (@maleadt)
• Binary dependency changes (#1159) (@maleadt)
• Bump dependencies. (#1161) (@maleadt)
• Generalize Sparse Array Indices Type in Struct Def (#1163) (@Roger-luo)
• Use unchecked type conversions for byte_perm arguments (#1166) (@eschnett)
• Fix performance regressions (#1167) (@maleadt)
• Fix big mapreduce kernel for inputs without neutral element. (#1174) (@maleadt)
• Switch contexts before performing memory operations on arrays (#1176) (@maleadt)
• Improvements to stream-ordered memory management (#1177) (@maleadt)
• Update manifest (#1180) (@github-actions[bot])
• Consistently use chars instead of raw enums in CUSPARSE/CUSOLVER functions. (#1181) (@maleadt)
• Implement forward compatibility (#1182) (@maleadt)
• Bump GPUCompiler for 1.8 compat. (#1183) (@maleadt)
• Bump GPUArrays. (#1186) (@maleadt)
• Update documentation (#1187) (@maleadt)

v3.4.2

CUDA v3.4.2

Diff since v3.4.1

Closed issues:

• Broadcasting a datatype does not work (#261)
• CUDA error: invalid argument during Zygote/Flux gradient computation (#1107)
• EXCEPTION_ACCESS_VIOLATION when using shared memory allocations. (#1116)

Merged pull requests:

• add symmetric support for mul (#217) (@Roger-luo)
• adds a device array type for CuSparseMatrixCSR to support using it in kernel functions (#1106) (@Roger-luo)
• Update manifest (#1108) (@github-actions[bot])
• Specialize Ref{<:Type} for GPU compatibility. (#1109) (@maleadt)
• Use the documented version of the enable_finalizers API. (#1111) (@maleadt)
• Don't embed the method table in the AST. (#1112) (@maleadt)
• Remove the hacky unique'ing of shmem GVs. (#1114) (@maleadt)
• Introduce a macro for marking multiple functions as device-only. (#1117) (@maleadt)
• Simplify library loading. (#1121) (@maleadt)
• Backports for 3.4.2 (#1122) (@maleadt)

v3.4.1

CUDA v3.4.1

Diff since v3.4.0

Closed issues:

• cudnnFindConvolutionAlgorithmWorkspaceSize uses removed function cached_memory (#1101)

Merged pull requests:

• Update manifest (#1102) (@github-actions[bot])
• Release hotfixes (#1103) (@maleadt)
• Reverse CI for NNlibCUDA.jl (#1104) (@maleadt)

v3.4.0

CUDA v3.4.0

Diff since v3.3.6

Merged pull requests:

• Update GPUArrays and GPUCompiler. (#1100) (@maleadt)

v3.3.6

CUDA v3.3.6

Diff since v3.3.5

Closed issues:

• LinearAlgebra.mul! with scalar arguments triggers scalar iteration (#790)
• Kernel fails if input is struct with function (#1094)
• cusparse: sparse matrix - matrix multiplication broken with transpose operation (#1095)

Merged pull requests:

• lib cusparse: fix #1095 (broken sparse matrix-matrix multiplication with transpose operation) (#1096) (@frapac)
• Only export the atomic macro on 1.6. (#1097) (@maleadt)
• Support more inplace atomic operations. (#1098) (@maleadt)
• Backports for 3.3.6 (#1099) (@maleadt)