v1.0.2

Release v1.0.2

This release reverts 'fix ndarrays on data oriented' which caused a regression in Genesis.

What's Changed

• [Type] Add unpacked form for qd.vector for indexed register access by @hughperkins in #718
• [DataOriented] Revert "[DataOriented] Fix ndarrays on data oriented (#704)" by @hughperkins in #719

Full Changelog: v1.0.1...v1.0.2

v1.0.1

Release v1.0.1

This release adds axes= to ndrange, bitonic sort, and adds 32x32 Cholesky tiles.

What's Changed

Perf

• [Perf] Move register-tile Cholesky optimizations from Genesis back into quadrants by @hughperkins in #714
• [Perf] Add bitonic sort to subgroup ops by @hughperkins in #713

DataOriented

• [DataOriented] Fix ndarrays on data oriented by @hughperkins in #704

Lang

• [Lang] Add axes= to ndrange by @hughperkins in https://github.com/Genesis-Embodied-

CI

• [CI] Add slow marker and remove un-necessary tests by @hughperkins in #711
• [CI] Upgrade PR change report from composer 2 to composer 2.5 by @hughperkins in #716
  AI/quadrants/pull/710

Doc

• [Doc] Fix spelling of 'Ying' to 'Yin' in README by @zhouxian in #715

Test

• [Test] Drop taichi xdist fork, use stock pytest-xdist by @hughperkins in #556

Full Changelog: v1.0.0...v1.0.1

v1.0.1b2

Pre-release v1.0.1b2

Changes:

Data-oriented

• [DataOriented] Fastcache, perf, pruning by @hughperkins in #705

Full Changelog: v1.0.0...v1.0.1b2

v1.0.1b1

Pre-release v1.0.1b1

Changes:

Data-oriented

• [DataOriented] Fastcache, perf, pruning by @hughperkins in #705

Full Changelog: v1.0.0...v1.0.1b1

v1.0.0

Release v1.0.0

This release adds new device-level ops for QIPC, and volatile_load.

What's Changed

GPU

• [GPU] New device-level ops for QIPC by @hughperkins in #693

Cleaning

• [Cleaning] PrefixSumExecutor: drop unused GRID_SZ local by @hughperkins in #701
• [Cleaning] sync(): fix unsupported-arch error message by @hughperkins in #700

Atomics

• [Atomics] add qd.volatile_load primitive (closes #648) by @hughperkins in #702

AutoDiff

• [AutoDiff] Reject recycled identity_key in AdStackCache::register_adstack_sizing_info by @hughperkins in #708

Vulkan

• [Vulkan] Declare GroupNonUniform SPIR-V caps and enable shaderSubgroupExtendedTypes by @hughperkins in #707

Full Changelog: v0.8.0...v1.0.0

v0.8.1b2

Pre-release v0.8.1b2

This pre-release is to test a new faster more streamlined data_oriented class on Genesis.

What's Changed

GPU

• [GPU] New device-level ops for QIPC by @hughperkins in #693

Cleaning

• [Cleaning] PrefixSumExecutor: drop unused GRID_SZ local by @hughperkins in #701
• [Cleaning] sync(): fix unsupported-arch error message by @hughperkins in #700

Atomics

• [Atomics] add qd.volatile_load primitive (closes #648) by @hughperkins in #702

AutoDiff

• [AutoDiff] Reject recycled identity_key in AdStackCache::register_adstack_sizing_info by @hughperkins in #708

DataOriented

• [DataOriented] Fastcache, perf, pruning by @hughperkins in #705

Full Changelog: v0.8.0...v0.8.1b2

v0.8.1b1

Pre-release v0.8.1b1

This pre-release is to test a new faster more streamlined data_oriented class on Genesis.

What's Changed

GPU

• [GPU] New device-level ops for QIPC by @hughperkins in #693

Cleaning

• [Cleaning] PrefixSumExecutor: drop unused GRID_SZ local by @hughperkins in #701
• [Cleaning] sync(): fix unsupported-arch error message by @hughperkins in #700

Atomics

• [Atomics] add qd.volatile_load primitive (closes #648) by @hughperkins in #702

AutoDiff

• [AutoDiff] Reject recycled identity_key in AdStackCache::register_adstack_sizing_info by @hughperkins in #708

DataOriented

• [DataOriented] Fastcache, perf, pruning by @hughperkins in #705

Full Changelog: v0.8.0...v0.8.1b1

v0.8.0

Release v0.8.0

This release brings many cross-GPU SIMT primitives, at both subgroup and block level. Note that subgroup reductions no longer take log2_size parameter, which is a breaking change, hence the minor version bump. In addition, AMD always uses wave64 going forward, to simplify testing.

What's Changed

GPU

• [GPU] Cross-GPU for grid ops by @hughperkins in #670
• [GPU] New bit ops for QIPC by @hughperkins in #679
• [GPU] Subgroup ops cross-gpu by @hughperkins in #665
• [BREAKING][GPU] New QIPC ops for subgroups by @hughperkins in #676
• [GPU] New QIPC ops for block by @hughperkins in #684

Math

• [Math] Make bitop operations portable cross-gpu by @hughperkins in #662
• [Math] New QIPC ops for single-threaded linalg by @hughperkins in #683

AMDGPU

• [AMDGPU] Always use wave64, on both RDNA and CDNA by @hughperkins in #687
• [AMDGPU] Use syncscope("agent") for atomix xor to avoid CAS livelock by @hughperkins in #672

Graph

• [Graph] Rename CUDA Graph to Graph in docs by @hughperkins in #691
• [Graph] HIP graph runtime support for @qd.kernel(graph=True) by @hughperkins in #692

Metal

• [Metal] Fix FIFO-queue ordering when sharing command queue. by @duburcqa in #694

Atomics

• [Atomics] New QIPC ops for atomics by @hughperkins in #690

Structs

• [Structs] Pass dataclass sub-structs into qd.func by @hughperkins in #698

CI

• [CI] Add per-file timing report to Mac Metal test job by @hughperkins in #695
• [CI] Enable kernel disk cache during tests by @hughperkins in #696

Full Changelog: v0.7.8...v0.8.0

v0.7.8

Release v0.7.8

This release contains further autodiff optimizations; and generalizes block and atomic operations across all GPU architectures.

What's Changed

Perf

• [Perf] Adstack max-reducer: launch cache + zero-copy result map; content-stable registry_id by @duburcqa in #671
• [Perf] CPU LLVM adstack-cache: skip per-launch bump-writes + ndarray_shapes capture on forward-only handles by @duburcqa in #685

SPIR-V

• [SPIR-V] dispatch_max_reducers: register each task with the real kernel name by @duburcqa in #675

AutoDiff

• [AutoDiff] Debug-mode field/grad/dual: dtype, layout, and access-time invariants by @duburcqa in #677

Docs

• [Docs] Add user-guide page for qd.algorithms.* device-wide algorithms by @hughperkins in #642
• [Docs] Doc for existing atomics: switch support table to per-backend columns by @hughperkins in #657

GPU

• [GPU] Cross gpu atomics by @hughperkins in #666
• [GPU] Make block operations portable cross-gpu by @hughperkins in #664

Full Changelog: v0.7.6...v0.7.8

v0.7.7

v0.7.7

This release mainly targets autodiff. It fixes SPIR-V backends (Metal, Vulkan), significantly improves runtime speed (up to 30%), and add full support of debug mode.

What's Changed

AutoDiff

• [Perf] Adstack max-reducer: launch cache + zero-copy result map; content-stable registry_id by @duburcqa in #671
• [SPIR-V] dispatch_max_reducers: register each task with the real kernel name by @duburcqa in #675
• [AutoDiff] Debug-mode field/grad/dual: dtype, layout, and access-time invariants by @duburcqa in #677

Full Changelog: v0.7.6...v0.7.7