TornadoVM 5.0.0

[stratika]

09/07/26

Improvements

• [hybrid] Hybrid API: native library tasks with cuDNN provider #888: [hybrid] Hybrid API: native library tasks with cuDNN provider
• [hybrid] Hybrid API: native library tasks with cuFFT provider #887: [hybrid] Hybrid API: native library tasks with cuFFT provider
• [JDK25][PTX] Tensor Core MMA intrinsics for FP16 and INT8 #869: [JDK25][PTX] Tensor Core MMA intrinsics for FP16 and INT8
• [Metal] Add hardware matrix units (simdgroup_float8x8) for fast gemms to M-silicon #850: [Metal] Add hardware matrix units (`simdgroup_float8x8`) for fast gemms to M-silicon
• [PTX][CUDA] Tensor Core MMA intrinsics for FP16 and INT8 #867: [PTX][CUDA] Tensor Core MMA intrinsics for FP16 and INT8
• [PTX] Add atan2 and asin math intrinsics and software atan implementation #854: [PTX] Add atan2 and asin math intrinsics and software atan implementation
• [jdk25] Add swizzled local memory accessors for FP16 and INT8 on the PTX backend #843: [jdk25] Add swizzled local memory accessors for FP16 and INT8 on the PTX backend
• Add swizzled local memory accessors for FP16 and INT8 on the PTX backend #841: Add swizzled local memory accessors for FP16 and INT8 on the PTX backend

Compatibility

• [ci] Automate multi-platform SDK release pipeline (+ Ray-Tracer/Metal CI and launcher exit-code fix) #857: [ci] Automate multi-platform SDK release pipeline (+ Ray-Tracer/Metal CI and launcher exit-code fix)

Bug Fixes

• Fix for stride-16 swizzled FP16 shared-memory load and store that caused segfault in tests #903: Fix for stride-16 swizzled FP16 shared-memory load and store that caused segfault in tests
• [hotfix] Multiple fixes after enabling CUDA backend into develop #874: [hotfix] Multiple fixes after enabling CUDA backend into develop
• Fix batch processing with a remainder chunk under device buffer reuse #886: Fix batch processing with a remainder chunk under device buffer reuse
• Fix repeated execution of batched plans #883: Fix repeated execution of batched plans
• Fix repeated execution of batched plans #882: Fix repeated execution of batched plans
• [hotfix] Resolving unit-tests failure in CUDA backend #877: [hotfix] Resolving unit-tests failure in CUDA backend
• [fix] Handle ObjectStamp compatibility for vector types in OpenCL, Metal and SPIRV backends #856: [fix] Handle ObjectStamp compatibility for vector types in OpenCL, Metal and SPIRV backends
• [fix] NPE when calling HalfFloat.getFloat32() on a value read from a local HalfFloat[] #844: [fix] NPE when calling HalfFloat.getFloat32() on a value read from a local HalfFloat[]
• [fix] Honour compile-time local work size in SPIR-V Level Zero launcher #838: [fix] Honour compile-time local work size in SPIR-V Level Zero launcher

Other Changes

• Fix for stride-16 swizzled FP16 shared-memory load and store that caused segfault in tests #904: Fix for stride-16 swizzled FP16 shared-memory load and store that caused segfault in tests
• Fix CUDA-backend build on develop (cuFFT broken merge + cudnn SDPA) #893: Fix CUDA-backend build on develop (cuFFT broken merge + cudnn SDPA)
• Feat/update logo #901: Feat/update logo
• [JDK25] Support for CUDA streams for overlapping data transfers with TaskGraph execution #897: [JDK25] Support for CUDA streams for overlapping data transfers with TaskGraph execution
• [JDK21] Support for CUDA streams for overlapping data transfers with TaskGraph execution #800: [JDK21] Support for CUDA streams for overlapping data transfers with TaskGraph execution
• Feat/update logo #902: Feat/update logo
• Revise README to include latest TornadoVM updates #852: Revise README to include latest TornadoVM updates
• Backport Cuda fixes to jdk25 #900: Backport Cuda fixes to jdk25
• Fix batch processing with a remainder chunk under device buffer reuse #892: Fix batch processing with a remainder chunk under device buffer reuse
• Backport Readme to jdk25 #896: Backport Readme to jdk25
• Backport Hybrid API to jdk25 #898: Backport Hybrid API to jdk25
• [hybrid] Hybrid API: native library tasks with cuBLAS/cuBLASLt provider #879: [hybrid] Hybrid API: native library tasks with cuBLAS/cuBLASLt provider
• [cuda] Add CUDA C backend via NVRTC to complement the PTX backend #861: [cuda] Add CUDA C backend via NVRTC to complement the PTX backend
• [OpenCL] Fix pointer-to-ulong conversions rejected by strict OpenCL compilers (Intel iGPUs) #885: [OpenCL] Fix pointer-to-ulong conversions rejected by strict OpenCL compilers (Intel iGPUs)
• [hotfix][PTX] Align compileTask exception handling with other backends to preserve original exception on bailout-disabled path #891: [hotfix][PTX] Align compileTask exception handling with other backends to preserve original exception on bailout-disabled path
• [hotfix][PTX] Align compileTask exception handling with other backends to preserve original exception on bailout-disabled path #890: [hotfix][PTX] Align compileTask exception handling with other backends to preserve original exception on bailout-disabled path
• Port the CUDA backend (CUDA C / NVRTC) from develop to jdk25 #881: Port the CUDA backend (CUDA C / NVRTC) from develop to jdk25
• [OpenCL] Fix pointer-to-ulong conversions rejected by strict OpenCL compilers (Intel iGPUs) #884: [OpenCL] Fix pointer-to-ulong conversions rejected by strict OpenCL compilers (Intel iGPUs)
• Replace heuristic brace-placement with structured control-flow recovery (OpenCL + Metal) #848: Replace heuristic brace-placement with structured control-flow recovery (OpenCL + Metal)
• [jdk25] Replace CFG block visitors with structured control-flow recovery (OpenCL + Metal) #849: [jdk25] Replace CFG block visitors with structured control-flow recovery (OpenCL + Metal)
• Fix nested persisted-object list on task-graph reuse and empty consume #870: Fix nested persisted-object list on task-graph reuse and empty consume
• Backport Persist obj to jdk25 #880: Backport Persist obj to jdk25
• Metal mma tiled jdk25 #876: Metal mma tiled jdk25
• Update level-zero library version to v1.18.1 #866: Update level-zero library version to v1.18.1
• Update level-zero library version to v1.18.1 #864: Update level-zero library version to v1.18.1
• Release 4.0.1-jdk25 #834: Release 4.0.1-jdk25

How to use TornadoVM SDK

To use the TornadoVM SDK, ensure that the environment variable JAVA_HOME points to a valid JDK 21 installation. After that, select the appropriate sdk distribution for your target architecture and the accelerator backends you intend to use.

Linux Systems

wget https://github.com/beehive-lab/TornadoVM/releases/download/v5.0.0-jdk21/tornadovm-5.0.0-jdk21-opencl-linux-amd64.zip
unzip tornadovm-5.0.0-jdk21-opencl-linux-amd64.zip
export TORNADOVM_HOME="$(pwd)/tornadovm-5.0.0-jdk21-opencl"
export PATH=$TORNADOVM_HOME/bin:$PATH
tornado --devices
tornado --version

macOS Systems

wget https://github.com/beehive-lab/TornadoVM/releases/download/v5.0.0-jdk21/tornadovm-5.0.0-jdk21-opencl-mac-aarch64.zip
unzip tornadovm-5.0.0-jdk21-opencl-mac-aarch64.zip
export TORNADOVM_HOME="$(pwd)/tornadovm-5.0.0-jdk21-opencl"
export PATH=$TORNADOVM_HOME/bin:$PATH
tornado --devices
tornado --version

Windows (10+) Systems

curl -L -o tornadovm-5.0.0-jdk25-opencl-windows-amd64.zip https://github.com/beehive-lab/TornadoVM/releases/download/v5.0.0-jdk25/tornadovm-5.0.0-jdk25-opencl-windows-amd64.zip
tar -xf tornadovm-5.0.0-jdk25-opencl-windows-amd64.zip
set TORNADOVM_HOME=%cd%\tornadovm-5.0.0-jdk25-opencl
set PATH=%TORNADOVM_HOME%\bin;%PATH%
tornado --devices
tornado --version