FlagFFT v0.2.0

FlagFFT v0.2.0

Part of FlagOS 2.2.

Changes since v0.1.0 (FlagOS 2.1).

9 pull requests and 384 additional commits.

Source: v0.2.0, commit 46338c973526.

New Features

• Add PPU backend support and improve codegen for libtriton_jit (#16) by @Artlesbol
• feat(npu): Ascend NPU backend for 1D/2D/3D FFT + launch-limit, codegen-import and kernel-cache fixes (#21) by @Artlesbol
• feat: add contiguous row-major rank-2 C2C/Z2Z 2D FFT via RTRT decomposition (c6ce5ab63794; commit)
• feat: add 2D FFT benchmark cases to all suites (8d9154e02837; commit)
• feat: implement 2D R2C/C2R/D2Z/Z2D FFT support (ac9b600d4fc5; commit)
• feat: add Rader algorithm for prime-size FFT acceleration (193ffb6395b5; commit)
• feat: add TestParams struct, CLI arg parsing, and JSON listener to ctest (d245300a21fa; commit)
• feat: add tools/consts.py and tools/run_tests.py unified test runner (37aecdf4c981; commit)
• feat: add CI workflow with unified test runner (c5fe12c72d39; commit)
• feat: align run_tests.py CLI and output format with upstream FlagGems (7352bd1b0c2d; commit)
• feat: add decomposition tuner for 2p20 (a72bfe1a0e9f; commit)
• feat: 支持 3D FFT（RTRT 分解 + 3D 轴置换 kernel） (0f879c2e610d; commit)
• feat: tune batched complex FFT algorithms and decompositions (873bfedf5b49; commit)
• feat: accept 36 FFT operators with NumPy validation and runtime plans (b01be40b38cc; commit)
• feat: add native MACA backend and portable Triton leaf exchange (6c23cd2e3074; commit)
• feat: align MACA planning with warp64 and split Bluestein execution (e21eae4b7ac6; commit)
• Add queried hardware profiles and isolated FP64 capability diagnostics (14ed0a580b76; commit)
• Add paired correctness-gated end-to-end hardware policy experiments (570208c495a3; commit)
• feat: enable the portable register-tile 3D transpose for MACA (0cf86bf9bfa9; commit)
• feat(maca): env-gated experiment knobs and a variant A/B harness (0e36c81deaba; commit)
• feat: 添加 FlagFFT 测试环境搭建文档 (45eb2a8a40b0; commit)
• feat(maca): tune bounded power-of-two prime convolution alternatives (ce5d27dc422f; commit)
• Add HCU backend adaptor for BW1000 (a296f1f6a862; commit)
• feat(maca): scope 2d codegen policy through axis compilation (225e4881368a; commit)
• feat(maca): add scoped opt-in tail policy and variant-safe codegen caches (f50973b75502; commit)
• feat(maca): make tail policy API-aware for real transforms (b35e93467bdc; commit)
• feat(maca): extend tail strategy to measured real prime pairs (7835e1fba765; commit)
• feat(maca): generalize small real direct-dft strategy bounds (d616489724f8; commit)
• feat(maca): enable the tail candidate policy by default for rank-1 batch-1 (6952a9e108bf; commit)
• Add native multiprocess JIT cache regression test (536f64c92f5d; commit)

Performance

• perf: optimize 2^20 FFT with Triton TLE (d6ee407f013b; commit)
• perf: reuse shared buffer for 2p20 leaves (3064a3494d6e; commit)
• perf: add distributed radix-32 leaf for 2p20 (c0a1b798ef3b; commit)
• perf: use thread-local radix-32 for 2p20 (586f695c887c; commit)
• perf: generalize thread-local mixed-radix 1d path (b9ea0d7e9ce2; commit)
• perf: extend mixed-radix 1d leaves (c36fc9b77e1c; commit)
• perf: repack large mixed-radix four-step leaves (00e2af4a24d7; commit)
• perf: distribute low-lane mixed-radix stages (feb6939efe34; commit)
• perf: generalize mixed-radix stage parallelism (ac4219ee29b9; commit)
• perf: penalize shared-heavy FFT leaves (b15ea1b253d7; commit)
• perf: model generated mixed-radix leaves (c49a70fe9f9f; commit)
• perf: extend thread-local FFT to real transforms (4af1fc44391b; commit)
• perf: improve large-prime Bluestein plan selection (aef798d5a2b5; commit)
• perf: remove legacy 16384 plan special case (8395b551e266; commit)
• perf: generalize fused Bluestein for 1D primes (fdf2bbb04eb6; commit)
• perf: add strided RC and register-transpose paths for 2D FFT (7ee19a7bd900; commit)
• perf: pack tiny leaves and real-transform row kernels (3a21d08115bd; commit)
• perf: keep RTRT for large rectangular 2D shapes (8d08f16a833d; commit)
• perf: scale Bluestein batch chunk by convolution size; revert leaf pack cap (7c6d82673006; commit)
• perf: vectorize 3D transpose with 8-byte ld/st.global.v2 (c02a626bf5e4; commit)
• perf: raise four-step col inner pack 2->4 for small-batch throughput (f2cdcf7b9eba; commit)
• perf: replay 2D RTRT/RC kernel sequences via CUDA Graph (dce53456dab1; commit)
• perf(cuda): improve small batch-1 z2z leaves on A100 (7c2684658a1e; commit)
• perf(cuda): generalize small-batch leaf optimization across FFT APIs (5796dd89c3f4; commit)
• perf(cuda): generalize z2z FourStep with f64 thread-local path (04e718891746; commit)
• perf(cuda): vectorize generic FourStep complex I/O and twiddles (fe8c5b23138b; commit)
• Optimize A100 FP64 real and Bluestein FFT paths ([...

FlagOS 2.1 — FlagFFT v0.1.0

Release v0.1.0

Initial Release

Other

• Add fp64/Z2Z support, R2C FFT, and real transform execution (#10) by @Artlesbol
• Add Docker environment files (#11) by @spring-wcb

Contributors

Thanks to all contributors who made this release possible:

• @Artlesbol
• @spring-wcb

---

Generated by Release Notes Generator