Version 6

What's Changed

• Remove trailing whitespaces and add missing newlines at EOF by @trueNAHO in #530
• Fix LD_AUDIT segfaulting in certain configurations by @vosen in #533
• Fix post-merge tests by @vosen in #535
• Update devcontainer Dockerfile, bump CUDA version to 13, split cuDNN into v8 and v9 by @vosen in #536
• Regenerate LLVM tests by @vosen in #538
• Add test for mma by @zluda-violet in #540
• Add MIOpen and connect it to zluda_dnn by @vosen in #539
• Add mma x2 test by @zluda-violet in #547
• Improve performance of instruction_mode_to_global_mode compiler pass by @vosen in #544
• Fix regression in instruction_mode_to_global_mode by @vosen in #552
• Add more missing BLAS by @vosen in #553
• Implement functionality required by katago by @vosen in #541
• Fix LLVM tests failing because of a variable order of declarations by @vosen in #557
• Fix bugs exposed by rust-cuda by @vosen in #563
• Small fix for Rust-CUDA support by @stevefan1999-personal in #560
• Build and distribute LLVM by @zluda-violet in #555
• Use different HiGHS solver (fixes crashes in some ptx modules) by @vosen in #565
• Emit llvm.zluda.mma intrinsic for MMA by @zluda-violet in #546
• Improve quality of instruction_mode_to_global_mode pass by @vosen in #567
• Implement more low precision ML instructions by @vosen in #554
• Update LLVM output by @zluda-violet in #568
• Implement fallback MMA for RDNA1 and RDNA2 by @vosen in #551
• Various LLVM fixes and improvements by @vosen in #570
• Try to pick a more appropriate ptx from fatbin by @vosen in #569
• Add a precompilation tool by @vosen in #558
• Support ld.global.v8 by @zluda-violet in #572
• Support st.global.v8 by @zluda-violet in #573
• Improve Windows loader (zluda.exe) by @vosen in #550
• Add parser support for .noreturn by @zluda-violet in #575
• Add createpolicy.fractional as nop by @zluda-violet in #576
• Add error message for PtxError::Todo to make debugging easier by @zluda-violet in #577
• Additional NVML functionality by @zluda-violet in #574
• Use LLVM for optimized i8 MMAs by @vosen in #571
• Enable ROCm7 support on Linux and Windows by @vosen in #579
• Fix bug where ignored directives are treated as invalid by @zluda-violet in #578
• Allow implicit conversion from vec to bit scalar for ld by @zluda-violet in #580
• Swap type and state space in variable printing for pass tests by @zluda-violet in #581
• [CLEANUP] Rename refactored modules, update test golden files by @zluda-violet in #582
• Increase sccache max frame length (fixes Windows builds) by @zluda-violet in #593
• Implement bmsk.clamp.b32 by @zluda-violet in #590
• Update SCCACHE_MAX_FRAME_LENGTH for post-merge builds by @vosen in #594
• Implement cuMemHostGetDevicePointer_v2 by @Knogle in #595
• Fix conflict for initial rounding and denormal mode for non-kernel functions by @zluda-violet in #596
• Implicitly convert constants from float to bit type by @zluda-violet in #583
• Allow implicit conversion from bit scalar to vec for st by @zluda-violet in #585
• Correctly zero-initialize globals by @vosen in #588
• When building ZLUDA in CI, make sure we build Linux binaries compatible with both ROCm 6 and ROCm 7 by @vosen in #589
• Add CUDA 13.1 compatibility by @vosen in #599
• Stop failing on bf16 uint_to_fp on amdgpu < gfx11 by @vosen in #601
• Update docs (add llama.cpp, zluda_precompile sections) by @vosen in #602
• Use partial parsing result in release mode by @vosen in #603
• Add sad and dp2a instructions by @vosen in #605
• Host functions for vLLM by @zluda-violet in #606
• Implement extended precision integer addition by @zluda-violet in #607
• ctlz bug fix by @zluda-violet in #610
• Fix loading CUDA modules from dark api and add a tool to verify Windows library loading by @vosen in #612
• Finish extended precision arithmetic by @zluda-violet in #609
• Enable handle from cublasCreate to be used in cublasLt calls by @zluda-violet in #587
• Support cuMemAllocPitch_v2 and cuMemcpy2D_v2 by @vosen in #616
• Add various bits and pieces required by pytorch by @vosen in #615
• Support some cublaslt settings required by COEIROINK by @vosen in #619
• Add minimal cuSPARSE by @vosen in #621
• PyTorch fixes and improvements by @vosen in #620
• Initial textures support by @vosen in #625
• Add more cuSPARSE functions by @vosen in #624
• Support vshr.u32.u32.u32.clamp.add by @vosen in #629
• Refactor emit_brev to use emit_intrinsic helper by @hemangjoshi37a in #631
• Update tests by @vosen in #632
• Fix typo: vec_acccess -> vector_read in emit_vector_read by @hemangjoshi37a in #633
• Remove redundant map_err(CompilerError::from) calls in compiler by @hemangjoshi37a in #635
• 32 bit support in the compiler by @vosen in #637
• Fix typo: compatiblity -> compatibility in xtask comment by @hemangjoshi37a in #639
• Fix typo: overriden -> overridden in zluda_redirect by @hemangjoshi37a in #641
• Implement match.any.sync, fix popc.b64 by @vosen in #642
• Fix clz.b64 and add bfind.shiftamt by @vosen in #644
• Add cuDeviceGetPCIBusId by @vosen in #645
• Minor improvements for PyTorch by @vosen in #643
• Update compiler to ROCm 7.2 and make some minor compiler fixes by @vosen in #649
• More minor compiler improvements by @vosen in #650

New Contributors

• @trueNAHO made their first contribution in #530
• @stevefan1999-personal made their first contribution in #560
• @Knogle made their first contribution in #595
• @hemangjoshi37a made their first contribution in #631

Full Changelog: v5...v6