All pull requests

Pull requests

• [nccl] Add tornado-nccl: NCCL collectives, send/recv and groups on TornadoVM device buffers

  #1180·kotselidis opened Oct 3, 2026Member·2 of 6 tasks·

  ·

• [cuda] Give each GPU its own primary context instead of sharing GPU 0's

  #1179·kotselidis opened Oct 3, 2026Member·2 of 6 tasks·

  ·

• [cuda] Abandon a CUDA-graph capture that fails part-way so later captures on the device work

  #1178·kotselidis opened Oct 3, 2026Member·2 of 6 tasks·

  ·

• [cuda] Capture CUDA graphs in thread-local mode so plans on other threads can keep working

  #1177·kotselidis opened Oct 3, 2026Member·2 of 6 tasks·

  ·

• Fix device index for tasks on identical GPUs

  7.2.0
  #1173·kotselidis opened Oct 3, 2026Member·2 of 6 tasks·

  ·

• [api] Support native arrays with more than Integer.MAX_VALUE elements

  #1163·kotselidis opened Oct 2, 2026Member·5 of 6 tasks·

  ·

• [runtime][opencl][cuda][metal] Launch reductions on a power-of-two grid with the work-group reduction after the loop

  7.3.0
  #1147·kotselidis opened Sep 29, 2026Member·5 of 6 tasks·

  ·

• [cuda] Pinned and mapped (zero-copy) host memory: XArray.allocate(n, HostMemoryType)

  #1139·mikepapadim opened Sep 28, 2026Member·

  ·
  1

• Create the OpenCL context concurrently with building the compiler backend

  #1130·kotselidis opened Sep 25, 2026Member·4 of 6 tasks·

  ·

• Create the CUDA context concurrently with building the compiler backend

  #1129·kotselidis opened Sep 25, 2026Member·2 of 6 tasks·

  ·

• Keep only the latest execution result by default, with an option to keep all

  7.3.0
  #1128·kotselidis opened Sep 25, 2026Member·2 of 6 tasks·

  ·

• [kotlin] Kotlin support for TornadoVM: API, examples, benchmarks, hybrid API, and cuTile

  kotlin
  #1123·kotselidis opened Sep 24, 2026Member·5 of 6 tasks·

  ·

• [backends] Emit Java's wrapping integer arithmetic with defined wrap-around

  #1104·mikepapadim opened Sep 18, 2026Member·

  ·
  10

• [cuda] Allow pinning NVRTC explicitly, and name the loaded toolkit in the fallback warning

  #1069·mikepapadim opened Sep 4, 2026Member·2 of 6 tasks·

  ·
  8

• [cuda] Cache the Graal compilation result per device so a new execution plan does not re-run the front end

  #1051·mikepapadim opened Aug 28, 2026Member·

  ·
  8

• feat: Add an optional native bytecode loop (experimental)

  #1041·Deepak-Sangle opened Aug 8, 2026·2 of 6 tasks·

  ·
  14

• [runtime] Issue on-demand copy-outs together and wait once instead of once per object (2.6x for 8 objects)

  #1031·mikepapadim opened Aug 6, 2026Member·

  ·
  6

• [api] TornadoExecutionPlan.withDeferredOutputs(): overlap the read-back with host work and pipeline consecutive executions (1.59x back-to-back)

  #1030·mikepapadim opened Aug 6, 2026Member·

  ·
  7

• [profiler] Harvest device timestamps after the stream drains instead of blocking on every operation (profiler-on cost -34%)

  #1024·mikepapadim opened Aug 5, 2026Member·

  ·
  6

• Make --printBytecodes readable: verbosity modes, per-execution summaries and a Graphviz export

  #1009·mikepapadim opened Jul 30, 2026Member·

  ·
  12

• CUDA stream usage: visible routing, DAG-based stream selection, plan-level tuning, examples and measurements

  CUDA
  #982·mikepapadim opened Jul 25, 2026Member·2 of 7 tasks·

  ·
  2