v0.6.3b3

What's Changed

• [Perf] Tiles 1: _load, _store, eye by @hughperkins in #466
• [Misc] Remove dead InternalFuncStmt type_check override by @hughperkins in #471
• [Perf] Tiles 2a: add cholesky and ger by @hughperkins in #472
• [Perf] Tiles 2b: add triangular solve by @hughperkins in #474
• [Misc] Tiles 2c: Refactor: use _get_col/_set_col in tiles load/store/init by @hughperkins in #475
• [Build] Fix flaky test_clock_accuracy by @hughperkins in #436
• Fix AARCH64 emitting invalid asm in CUDA kernels. by @duburcqa in #473
• [AMDGPU] Enable HIP memory pool and surface pool-exhaustion errors. by @duburcqa in #485
• [AMDGPU] Scope hsaco tmp dir per-user to avoid collisions. by @duburcqa in #484
• [Perf] Tiles 3: Add slice syntax, qd.outer() and initial doc by @hughperkins in #477
• [AMDGPU] Fix gradient computation. by @duburcqa in #486
• Enable all backends that are supported in unit tests. by @duburcqa in #488
• Fix SPIRV ID overflow for large kernels due to autodiff. by @duburcqa in #489
• [Misc] Fix purity checker to allow accessing constants from quadrants modules by @hughperkins in #487
• [Misc] Increase tolerance for clock monotonic test by @hughperkins in #492
• [CI] Serialize api doc workflow by @hughperkins in #494
• [CI] Increase tolerance for clock test by @hughperkins in #506
• [CI] Increase clock test tolerance to 20% by @hughperkins in #509
• [Perf] Tiles 4a: Add tensor_type parametrization to tile16 tests by @hughperkins in #504
• [Perf] Tiles 4b: Migrate tiles16 tests to enable fastcache by @hughperkins in #505
• [Perf] Tiles 4c: add Tiles16x16 proxy by @hughperkins in #507
• [Perf] Tiles 4d: Consolidate slice error tests using parametrize by @hughperkins in #508
• [Perf] Tiles 4: add SharedArray slice support by @hughperkins in #482
• [Doc] Add user guide page for subgroup shuffle by @hughperkins in #512
• [Perf] Implement cross-platform shuffle_down by @hughperkins in #510
• [Perf] Add portable subgroup reduce_add and reduce_all_add by @hughperkins in #511

Full Changelog: v0.6.2...v0.6.3b3

v0.6.3b2

Pre-release v0.6.3b2

This pre-release add 16x16 register-only tiles for Cholesky, portable across all GPUs supporte by Quadrants (AMD, Vulkan, Metal, CUDA).

What's Changed

Perf

• [Perf] Tiles 1: _load, _store, eye by @hughperkins in #466
• [Perf] Tiles 2a: add cholesky and ger by @hughperkins in #472
• [Perf] Tiles 2b: add triangular solve by @hughperkins in #474
• [Perf] Tiles 2c: Refactor: use _get_col/_set_col in tiles load/store/init by @hughperkins in #475
• [Perf] Add 16x16 tiles for Cholesky factorization by @hughperkins in #449

Misc

• [Misc] Remove dead InternalFuncStmt type_check override by @hughperkins in #471

Tests

• [Tests] Fix flaky test_clock_accuracy by @hughperkins in #436

Full Changelog: v0.6.2...v0.6.3b2

v0.6.2

Release v0.6.2

This release adds support for atomics on float sharedarrays in spirv, and no longer needs CUDA Toolkit on SM90+ GPUs when using graph do while.

What's Changed

SPIRV

• [SPIRV] Feature Parity Atomics & Shared Array by @duburcqa in #432

Misc

• [Misc] Change clang format to 120 characters by @hughperkins in #463
• [Misc] CUDA graph 5 Add fatbin by @hughperkins in #464

Bug

• [Bug] Reuse VkInstance across init/reset cycles by @hughperkins in #465

Full Changelog: v0.6.1...v0.6.2

v0.6.3b1

Pre-release v0.6.3b1

This pre-release add 16x16 register-only tiles for Cholesky, portable across all GPUs supporte by Quadrants (AMD, Vulkan, Metal, CUDA).

What's Changed

• [Perf] [Perf] Add 16x16 tiles for Cholesky factorization by @hughperkins in #449

Full Changelog: v0.6.2...v0.6.3b1

v0.6.1

Release v0.6.1

This release improves search for CUDA toolkit on Windows.

What's Changed

• [Bug] Also search default CUDA toolkit install location on Windows by @hughperkins in #461

Full Changelog: v0.6.0...v0.6.1

v0.6.0

Release v0.6.0

This release fixes a bug with CUDA graph on Windows. We also rename gpu_graph to just graph, which is backwards incompatible, hence the minor version bump.

What's Changed

Misc

• [Misc] Rename gpu_graph to graph by @hughperkins in #446

SIMT

• [SIMT] Add cross-platform shuffle by @hughperkins in #447

Bug

• [Bug] Fix graph_do_while on Windows: search for cudadevrt.lib by @hughperkins in #456

Full Changelog: v0.5.2...v0.6.0

v0.6.0b13

Pre-release v0.6.0b12

This pre-release add 16x16 register-only tiles for Cholesky, portable across all GPUs supporte by Quadrants (AMD, Vulkan, Metal, CUDA).

What's Changed

• [Misc] Rename gpu_graph to graph by @hughperkins in #446
• [Misc] Add cross-platform shuffle by @hughperkins in #447
• [Perf] [Perf] Add 16x16 tiles for Cholesky factorization by @hughperkins in #449

Full Changelog: v0.5.2...v0.6.0b12

v0.6.0b12

Pre-release v0.6.0b12

This pre-release add 16x16 register-only tiles for Cholesky, portable across all GPUs supporte by Quadrants (AMD, Vulkan, Metal, CUDA).

What's Changed

• [Misc] Rename gpu_graph to graph by @hughperkins in #446
• [Misc] Add cross-platform shuffle by @hughperkins in #447
• [Perf] Add 16x16 tiles for Cholesky factorization by @hughperkins in #449

Full Changelog: v0.5.2...v0.6.0b12

v0.6.0b11

Pre-release v0.6.0b11

This pre-release add 16x16 register-only tiles for Cholesky, portable across all GPUs supporte by Quadrants (AMD, Vulkan, Metal, CUDA).

What's Changed

• [Misc] Rename gpu_graph to graph by @hughperkins in #446
• [Misc] Add cross-platform shuffle by @hughperkins in #447
• [Perf] Add 16x16 tiles for Cholesky factorization by @hughperkins in #449

Full Changelog: v0.5.2...v0.6.0b11

v0.6.0b10

Pre-release v0.6.0b10

This pre-release add 16x16 register-only tiles for Cholesky, portable across all GPUs supporte by Quadrants (AMD, Vulkan, Metal, CUDA).

What's Changed

• [Misc] Rename gpu_graph to graph by @hughperkins in #446
• [Misc] Add cross-platform shuffle by @hughperkins in #447
• [Perf] Add 16x16 tiles for Cholesky factorization by @hughperkins in #449

Full Changelog: v0.5.2...v0.6.0b10