Pyqrack

While we've tried various "hybridizations" of quantum binary decision diagrams (QBDD) with state vector simulation and other methods in the past, they as of yet have not given any compelling practical advantages in general use cases, as best the Qrack team can assess.

Trying a slightly different approach with the new QBdtHybrid, this might remain the case, like for the quantum Fourier transform or random circuit sampling. However, holistic integration with the overall "layer stack" seems significantly improved, with this approach, and special-case advantages might have improved. (We still need to collect additional benchmark data, to test this.)

File SHA-1 sums:
cda0a18756f8a003fc1c224d0bc105943d7b9fb1 pyqrack-1.24.0-py3-none-any.whl
c3ffd1fd5772c9c2389fd2b499c79a81cc968118 pyqrack-1.24.0-py3-none-linux_aarch64.whl
44fa68ca664235476285f34bd5e8cb7ff33a1dec pyqrack-1.24.0-py3-none-linux_armv7l.whl
ef0c039595287f15ea8b81919a44ccdd1a7eb558 pyqrack-1.24.0-py3-none-macosx_10_4_x86_64.whl
562d7ebd24fc81274241e0c0a8d90a6b6cf21f4a pyqrack-1.24.0-py3-none-manylinux2014_x86_64.whl
2f5bb69632e03596f2eac79990868f3979d52ef5 pyqrack-1.24.0-py3-none-manylinux_2_35_x86_64.whl
31809acfa00b87ca914d9f3ae209c5db774feb97 pyqrack-1.24.0-py3-none-win32.whl
7991aa8136263fa22b1be660347c3d88a1def546 pyqrack-1.24.0-py3-none-win_amd64.whl
c171fcaf9e1311ec454c2a4bafedb5feba9f2851 pyqrack-1.24.0.tar.gz

This patch release debugs `QUnitMulti` for heterogeneous OpenCL devices. **Note that this is the first ever release with CMake `-DUINTPOW=5` option, hence you will need to recompile your OpenCL kernels, if you use "pre-compilation."** `-DUINTPOW=5` limits `QPager` "page" width to 31 qubits or less, but this is literally larger than the maximum allocation segment of an A100 GPU, and this has no impact on multi-GPU scaling, hence there should be no new limitations so long as environment variable `QRACK_MAX_PAGE_QB` is set to `31` or less, (and most systems, including the A100, will typically want to set this variable even _lower._)

With this release, interoperability between NVIDIA GPU and Intel HD integrated graphics can offer significant speed advantages on quantum volume circuits, for example. For a gaming laptop with two OpenCL devices, an Intel HD (device number `0`) and an RTX 3080 (device number `1`), here is an example of how the environment variable might be set:

export QRACK_OCL_DEFAULT_DEVICE=1
export QRACK_MAX_PAGE_QB=27
export QRACK_QUNITMULTI_DEVICES=1,0
export QRACK_QUNITMULTI_DEVICES_MAX_QB=-1,15
export QRACK_QPAGER_DEVICES=8.1,24.0
export QRACK_QPAGER_DEVICES_HOST_POINTER=8.0,24.1
export QRACK_MAX_ALLOC_MB=24576,16108
export QRACK_MAX_PAGING_QB=32
export QRACK_MAX_CPU_QB=32
export QRACK_QTENSORNETWORK_THRESHOLD_QB=32
These commented variables might also be tuned for performance and approximation
export QRACK_PSTRIDEPOW=11
export QRACK_QUNIT_SEPARABILITY_THRESHOLD=0.00001
export QRACK_QBDT_SEPARABILITY_THRESHOLD=0.00001
export QRACK_NONCLIFFORD_ROUNDING_THRESHOLD=1


File SHA-1 sums:
5a003eee9f12f66adb8d6c7be8a8752f08f074fc pyqrack-1.23.1-py3-none-any.whl
26099f2c626d412c14785095596a2aa173fabaca pyqrack-1.23.1-py3-none-linux_aarch64.whl
53abb05db9fa7270b6fd86380191a7817aebbb5d pyqrack-1.23.1-py3-none-linux_armv7l.whl
cdcc0a05163f704ae0bdff99c8820c460d5b7666 pyqrack-1.23.1-py3-none-macosx_10_4_x86_64.whl
b9f4468a749cf5bca2889dcbc54cb8a115fd72d1 pyqrack-1.23.1-py3-none-manylinux2014_x86_64.whl
1f94239f491a8c6bac53b0d98bf46bf41426168a pyqrack-1.23.1-py3-none-manylinux_2_35_x86_64.whl
00316ea4521800890e33be8a5853f93d10b12424 pyqrack-1.23.1-py3-none-win32.whl
68f2d661593aa65b9aacc7ac4d54be37bc73b4e5 pyqrack-1.23.1-py3-none-win_amd64.whl
fc60129a0c20be898469076a89e146ef919c3553 pyqrack-1.23.1.tar.gz

This feature release adds the `QRACK_QUNITMULTI_DEVICES_MAX_QB` environment variable. Per device, as listed and ordered in `QRACK_QUNITMULTI_DEVICES`, this new variable allows the user to specify a maximum subsystem qubit width, above which no larger subsystem will be allocated on the particular device. This is particularly useful for combining OpenCL CPUs with OpenCL GPUs and other accelerators, since CPUs can give a significant advantage but are best suited for lower parallelism widths than GPUs.

File SHA-1 sums:
ab71d9cd6fa0837bb918223decffe56c165fecef pyqrack-1.23.0-py3-none-any.whl
079d1b332d267943e584796ee2daf9a951923082 pyqrack-1.23.0-py3-none-linux_aarch64.whl
270077e6196a9409377523e8fb507288f8bc2ff8 pyqrack-1.23.0-py3-none-linux_armv7l.whl
06937245b6c6e1e8980731df29adc698be2393bd pyqrack-1.23.0-py3-none-macosx_10_4_x86_64.whl
09e9ea682b07ea6e1b8bbbbae215d7becc8d30a1 pyqrack-1.23.0-py3-none-manylinux2014_x86_64.whl
ce62cf47b6c7dcaa8e21ab28338e3fc1772e6dc3 pyqrack-1.23.0-py3-none-manylinux_2_35_x86_64.whl
0a0efeab0167a9879460ad431953bfde79c291bb pyqrack-1.23.0-py3-none-win32.whl
d07f28a6abe312325fd331af0b9843c22c0b6392 pyqrack-1.23.0-py3-none-win_amd64.whl
a40990c51eae76d66a6c84fbff792d92271fc444 pyqrack-1.23.0.tar.gz

QrackCircuit` now exposes a `bool` constructor options that, if `true`, only "simplifies" Clifford-with-Clifford and non-Clifford-with-non-Clifford gates, but not Clifford-with-non-Clifford. `QTensorNetwork` automatically leverages this option when `QStabilizerHybrid` is in use, as in default "layer stack" case. (Some circuits still might complete more efficiently without `QStabilizerHybrid`.)

File SHA-1 sums:
ac001ecca1ee0d0382feb2e1328d70c84de37c99 pyqrack-1.22.0-py3-none-any.whl
058c12cf0cc9906d5dd599bf6e8cfbb8df5df119 pyqrack-1.22.0-py3-none-linux_aarch64.whl
8cee1857997247fe1a9b61ab973cb1b05c7cce6d pyqrack-1.22.0-py3-none-linux_armv7l.whl
b961a023eec554266a750d863005fe6c655bc012 pyqrack-1.22.0-py3-none-macosx_10_4_x86_64.whl
64175fd6cdc60d9ac9c0027c2ebb45d9ab66c0de pyqrack-1.22.0-py3-none-manylinux2014_x86_64.whl
3e5decfdd7273b405f228110dbc2736813be8a38 pyqrack-1.22.0-py3-none-manylinux_2_35_x86_64.whl
2920d120edd764cc1f616d7309086d047184ad7e pyqrack-1.22.0-py3-none-win32.whl
847a29ed0da0131edde51e0b899addb0bc8b5485 pyqrack-1.22.0-py3-none-win_amd64.whl
17bbea4729be8d02bc8f9bc7b40765b68d44f932 pyqrack-1.22.0.tar.gz

For CPU-based and OpenCL-based back ends, a minor improvement has been made in the algorithm of MAll(), such that single-shot simulation might see a small average-case performance benefit.

Also, Qrack's intention was always that internal "epsilon" for n bit precision was 2^-(n+1); these values were accidentally set much lower. (The change has been tested with the "normalization" features of Qrack, which are basically all this change should affect, and it works.)

File SHA-1 sums:
a9ae91e05a5f3e2c70f303c22a9b411faae4da7d pyqrack-1.21.4-py3-none-any.whl
1ad84872c22fdb50a64ca5b18d594d4dbd3571e6 pyqrack-1.21.4-py3-none-linux_aarch64.whl
1ff3b9f12f9dfa27cc0ac36e6037cc63d0b767d2 pyqrack-1.21.4-py3-none-linux_armv7l.whl
4d655ec89c0800ff9c3c55e694c2661da5077ab5 pyqrack-1.21.4-py3-none-macosx_10_4_x86_64.whl
24615f1e4de600611a686022b4c4dae7da8f70b3 pyqrack-1.21.4-py3-none-manylinux2014_x86_64.whl
062a3b1bd9c4a0ef115694b61d4ad64cfc2bb820 pyqrack-1.21.4-py3-none-manylinux_2_35_x86_64.whl
bd9e38a1bd560685739f06c7770d0a8945e62014 pyqrack-1.21.4-py3-none-win32.whl
d7245e3f196332e4de55939abe909aad78f51301 pyqrack-1.21.4-py3-none-win_amd64.whl
417809d2006bb1277d4590560ecd49fec11da6a2 pyqrack-1.21.4.tar.gz

QTensorNetwork` has been changed from asynchronous (for gate dispatch without output) to synchronous (with threading for `QCircuit`, but with usual `QInterface` threading). This should fix segmentation faults with no major performance drawbacks.

File SHA-1 sums:
fa33202dcf53f5b4fe72ce1b876569b04350aa86 pyqrack-1.21.3-py3-none-any.whl
632944d0b9e3f5465737de15677b045256ab2d8d pyqrack-1.21.3-py3-none-linux_aarch64.whl
b40cc1ed817c3fcc83e89a5bd8a9298ba698ed4a pyqrack-1.21.3-py3-none-linux_armv7l.whl
281fc6605bfbacb723e8c6474ae7bf81df0314a8 pyqrack-1.21.3-py3-none-macosx_10_4_x86_64.whl
3ba35cd891dc4897dc8dd2ff345f869179cf5a4e pyqrack-1.21.3-py3-none-manylinux2014_x86_64.whl
261d74cbfce7731401a63d0c3aa6b47eca2ff0e1 pyqrack-1.21.3-py3-none-manylinux_2_35_x86_64.whl
d015035915d6b2fff2ed45f55eb8248e3bd56cc3 pyqrack-1.21.3-py3-none-win32.whl
f1d7065bbb8d88e6fc8da0bc4dc5aa1d9715cee8 pyqrack-1.21.3-py3-none-win_amd64.whl
7cfc12082485f48c6e8017c3650f0ae9b5f8a2be pyqrack-1.21.3.tar.gz