Skip to content

0.17.0

Latest

Choose a tag to compare

@bettinaheim bettinaheim released this 16 Nov 14:28

What's Changed

This release comes with a range of exciting new features.

We now distribute a new EDEM package to facilitate building DEMs dynamically for QEC at scale. We have also created a separate cudaq-core PyPI package that contains the runtime dependencies for the cudaq-devel package that can be found as artifact in this release. These packages contain the necessary build and runtime dependencies respectively to build and distribute CUDA-Q compiler extensions and integrations.

We have completely overhauled our compiler configurability, improved performance across the board, and include a new custabilizer backend for faster stabilizer simulation. Take a look at the features and enhancements listed below.

Last but not least, the transport and dispatch APIs for CUDA-Q Realtime received a major update to make integration easier and more consistent. To facilitate exploring highly customized scenarios, we now provide the means to plug in a custom dispatcher tailored to specific applications.

Features and Enhancements 🎉

Breaking Changes 🛠

Performance⏱️

  • Speeding up cudaq.logical compilation for Gidney–Ekera by @sacpis in #5681
  • Speed up quantum allocation combining by @taalexander in #5318
  • Skipping gridsynth for exact Clifford+T angles by @sacpis in #5420
  • Add runtime support for native negated controls by @khalatepradnya in #5501
  • Speeding up cudaq-logical physical estimation by @sacpis in #5513
  • Speeding up Gidney-Ekera estimation by @sacpis in #5512
  • Adding move operations to ExecutionResult by @sacpis in #5550
  • Accelerate dynamics by adding fabric-qualified memory allocations for MNNVL and reusing allocations by @cuMikeBrown in #5545
  • Reducing Clifford overhead in GridSynth approximations by @sacpis in #5434
  • Making dead store removal linear in the number of stores by @sacpis in #5582
  • Making indirect-to-direct-calls linear in module size by @sacpis in #5593
  • Fixing quadratic erase list dedup in LiftArrayAllocPatterns by @sacpis in #5617
  • Fixing quadratic worklist scan in SROA by @sacpis in #5615
  • Fixing quadratic live scope lookup in VariableCoalesce by @sacpis in #5618
  • Reusing symbol tables in GetConcreteMatrix by @sacpis in #5623
  • Making combine-quantum-alloc lookup constant time by @sacpis in #5626
  • Making factor-quantum-alloc concat offsets linear by @sacpis in #5648
  • Speeding up device-code-loader dependency printing by @sacpis in #5644
  • Making delay measurements traversal linear by @sacpis in #5665
  • Avoiding quadratic unwrap lookup in RegToMem by @sacpis in #5667
  • Making QIR array release constant time by @sacpis in #5669
  • Using a min-heap for recycled qudit IDs by @sacpis in #5692
  • Computing state-prep rotation angles with a fast Walsh-Hadamard transform by @sacpis in #5694
  • Using Eigen matrix exponential in complex_matrix::exponential by @sacpis in #5697
  • Caching GridSynth words in qlx-synthesize-rotations by @sacpis in #5716
  • Finding patch adjacency with one traversal per patch by @sacpis in #5734
  • Fuse constant one-qubit chains for Clifford+T synthesis by @taalexander in #5575
  • Move SV sampling results instead of copying them. by @cuMikeBrown in #5766
  • Optimizations to python kernel argument processing by @cuMikeBrown in #5765
  • Adding structural fingerprint mode as an alternative to hashing a merged print of a program by @cuMikeBrown in #5788
  • Update the IB Stack in our docker containers for improved performance by @cuMikeBrown in #5009
  • Preserve native negated gate controls through full QIR by @khalatepradnya in #5516

CUDA-Q Realtime

  • Adding unified "kernel" support for UDP by @Renaud-K in #5337
  • Isolating RoCE transport dependencies into their own functions and out of the unified kernel dispatcher by @Renaud-K in #5484
  • Moving the unified cpu_roce implementation behind the cpu_dataplane by @Renaud-K in #5535
  • Returning value through TX for the unified kernel by @Renaud-K in #5580
  • Pass (rx_slot, tx_slot, slot_size) to the realtime DEVICE_CALL handlers by @cketcham2333 in #5602
  • Add LTO optimization for realtime apps by @Renaud-K in #5583
  • Introducing cudaq_dispatcher_set_bridge by @Renaud-K in #5610
  • Updating realtime to Hololink 2.7 and DOCA 3.4 by @Renaud-K in #5670
  • Remove the cooperative dispatch kernel mode from cudaq-realtime by @cketcham2333 in #5673
  • Run DEVICE_CALL handlers on every thread of the dispatch block by @cketcham2333 in #5675
  • Remove host-side graph launch from the host dispatch path by @cketcham2333 in #5736
  • Introduce the target agnostic UNIFIED kernel architecture with nvJitLink by @Renaud-K in #5722
  • Improve the performance of the unified dispatcher by @Renaud-K in #5782
  • Loading nvjitlink dynamically by @Renaud-K in #5797
  • Adding support for a custom GPU dispatcher by @Renaud-K in #5778
  • Add cudaq_bridge_get_ring_buffer and deprecate cudaq_bridge_get_transport_context by @cketcham2333 in #5840

Documentation Updates ✏️

Other Changes

Bug Fixes 🐛

  • Fix C++ lowering of cudaq::range and add better tests. by @schweitzpgi in #5355
  • Prevent infinite recursion in repair-linear-type pass by @schweitzpgi in #5351
  • Better support for references to kernel decorators in Python bridge by @atgeller in #5372
  • Update canonicalization of cc.if by @atgeller in #5303
  • Harden ApplyOpSpecialization to callable values by @atgeller in #5395
  • Harden quake.phase lowering and canonicalization by @cabreraam in #5322
  • Supporting comma expressions by @sacpis in #5407
  • Fix redundant sub-steps in dynamics integrators by @1tnguyen in #5446
  • Fix empty control-array packing by @khalatepradnya in #5458
  • Unify Quake gate definitions and fix negated-control expansion by @khalatepradnya in #5449
  • Fix bit order for qBraid integration by @ryanhill1 in #5462
  • Fix transposed Gray-code angle matrix in from_state for 3+ qubits by @ikkoham in #5486
  • Fix batched super-operators without a dissipator by @udsy19 in #5357
  • Fix PhasedRx adjoints and negated custom controls by @khalatepradnya in #5511
  • Respect negated controls in phase folding by @khalatepradnya in #5506
  • Fix the qubit vector that apply_noise receives for a qvector by @khalatepradnya in #5528
  • Allow atom_filling=None in RydbergHamiltonian by @udsy19 in #5527
  • Fix integrator_options leaking across BaseIntegrator instances by @udsy19 in #5345
  • Fixing multi-controlled gates on cudaq.logical targets by @sacpis in #5566
  • Fixing cudaq.logical import_cudaq for kernels that allocate qubits by @sacpis in #5603
  • Fix bugs in C++ template function kernels by @schweitzpgi in #5606
  • Reset cudaq.run's noise model when the launch fails by @khalatepradnya in #5634
  • Fix bug in loop normalization, step expression rewrite by @schweitzpgi in #5645
  • Lower switch instructions for the adaptive QIR profile by @mitchdz in #5663
  • Fix loop analysis on a loop condition without a defining op by @khalatepradnya in #5676
  • Honor if filters in gate-applying list comprehensions by @huaweil-nv in #5739
  • Binding the CUDA device before launching on MQPU workers by @sacpis in #5757
  • Fixing stack growth from captured lists in repeated kernel calls by @sacpis in #5781
  • Fix crash when applies share a constant-propagated callable by @iarjunganesh in #5493
  • Apply a run noise model to the requested QPU by @khalatepradnya in #5802
  • Prevent cudaq.run from leaking result storage by @khalatepradnya in #5786
  • Fixing out-of-range qubit index check in QIR Array indexing by @sacpis in #5805
  • Keeping CUDA-Q's LLVM symbols out of the global scope by @sacpis in #5804
  • Loading nvq++ compiled kernel libraries without exposing LLVM globally by @sacpis in #5811
  • Fix general Kraus noise channels on fp32 backends by @huaweil-nv in #5584
  • Fix analytical estimation of single-attempt retries by @huaweil-nv in #5830
  • Fix sporadic race condition in state resize during trajectory noisy statevector simulation by @1tnguyen in #5868
  • Fixing QPU queue lifetime during target reset and endpoint reuse by @sacpis in #5876
  • Fixing race in matrix operator registration by @sacpis in #5879

New Contributors

Full Changelog: 0.16.0...0.17.0
Release created by workflow 43856736126.