ashvardanian/ParallelReductionsBenchmark

📦 Patch

• Make: Correct Metal/OpenCL associations (b282335)
• Make: Freeze Rust toolchain to `nightly` (9de592d)

📦 Patch

• Improve: `ChillOut` to send/sync pointers (a0986f2)

📦 Minor

• Add: `oneapi::tbb::task_arena` backend (bb6602b)

📦 Patch

• Docs: Put build notes in the end (dc6e246)
• Make: Fetch CLI options (33c6b59)
• Fix: Guard OpenMP calls (2056636)
• Fix: `serial` / `unrolled` kernel naming (ff715f6)

📦 Patch

• Docs: Apple M2 Pro small results (9788a53)
• Make: Bump Fork Union with CrossBeam (860b5eb)
• Improve: Use `std::hardware_destructive_interference_size` (13e6cf9)

📦 Minor

• Add: Fork Union Rust benchmarks (9ec7f0c)

📦 Patch

• Docs: Rust results vs Rayon, Tokio, Smol.rs (9cd6268)
• Fix: Wait for Smol.rs results (48ab934)
• Fix: Cast pointers to `usize` (ed113e9)
• Improve: Reuse partial sums buffer (ca2b2c6)
• Make: Pass small size to C++ bench (1d17f93)
• Fix: Rust drafts (8e11aea)
• Make: Rust draft (b3ddf30)

📦 Patch

• Make: Reuse Taskflow executor (2daf80c)
• Make: Skip Taskflow example builds (450093e)
• Fix: Unrolled loop length (7c1b859)

📦 Patch

• Docs: Polish namespaces! (5fa1c2b)

📦 Patch

• Docs: Shrink logs column widths (7ec74f8)

📦 Minor

• Add: Taskflow backend (0514627)

📦 Patch

• Docs: Graviton Taskflow vs OpenMP vs Fork Union (c1862d9)
• Docs: OpenMP vs Fork Union (8ed7080)

📦 Patch

• Improve: Over-align partial sums to avoid false-sharing (c2b3284)
• Improve: Slicing with `scalars_per_core` (1259361)

📦 Patch

• Make: VS Code launchers for macOS (1198e49)
• Fix: Avoid deprecated 32-bit API (71de4ea)
• Fix: `openmp_gt` start offset (d19038f)
• Docs: Apple M2 Pro results (a8d8360)
• Docs: Compiling on macOS (c86a8dd)
• Make: Skip BLAS on macOS (e73f155)

📦 Patch

• Fix: OpenMP split size (d83bb02)
• Fix: Use <Accelerate> on macOS (e4f601b)
• Make: Avoid CUDA on non-Linux builds (975d013)

📦 Patch

• Docs: Single-socket Graviton4 results (c923a60)

📦 Minor

• Add: `fork_union` parallel version (c40b7f3)
• Add: Generic OpenMP pool (cce8be4)
• Add: NEON and SVE kernels (32d7d3e)

📦 Patch

• Docs: Remove namespace nesting (5e73225)
• Make: Formatting CMake (82119ea)
• Docs: Ordering header notes (a7891c7)

📦 Patch

• Docs: Mark programming language used (8680dcc)

📦 Patch

• Docs: H200 benchmarks (4b877b9)
• Fix: Missing `blasint` type (fad513b)

📦 Patch

• Fix: Measuring AVX-512 throughput (a301c39)

📦 Patch

• Docs: Need CUDA Toolkit (6ca84ed)
• Improve: Disambiguate buffer length (86a9794)
• Fix: `dataset_t` move-constructor (434cad7)
• Fix: `memset` entire dataset (e70d6a8)
• Fix: Pattern-matching version in CMake (458342d)

📦 Patch

• Docs: List notable features (c340f4c)

📦 Minor

• Add: BLAS with zero-stride (c39768e)
• Add: Latency Hiding in AVX-512 (9a5a61a)

📦 Patch

• Fix: `fmt::print` thousand separator (6250d0f)
• Fix: Detecting NUMA (8da88f6)
• Improve: Uniform benchmark naming (567d2e8)
• Improve: Scaling CUDA kernels (e57eff0)
• Make: Find `libnuma` (d0f74e4)
• Improve: Huge Pages and code style (aeabdce)
• Improve: `setzero` over `set1` intrinsics (1fbcd94)
• Fix: Handling SSE tail (2731049)
• + 3 more

📦 Minor

• Add: Metal draft (db0dc92)

📦 Patch

• Docs: File headers (e1ac216)
• Make: Switch to `.cuh` extension for CUDA (37c0581)
• Make: Drop deprecated assets (f311a61)
• Fix: Missing `fmt` include for OpenCL (0c0bc0f)
• Fix: Feature-testing `__cpp_lib_execution` (0c80b36)
• Make: Switch to `.cuh` extension for CUDA (818851f)