ParallelReductionsBenchmark

ParallelReductionsBenchmark copied to clipboard

Add support for partial builds

1

On some architectures, we won't have access to CUDA. On others, OpenCL might be an issue. Furthermore, the compiler flags for non-GCC & non-NVCC builds have to be properly configured....

ashvardanian

Using Mat-Mul Instructions, like Arm SME and Intel AMX

5

Array reductions can be represented as a two-stage pipeline built on top of matrix-vector multiplications, where the vector is made of all ones. Let's say our hardware supports fast 16...

ashvardanian

It's hard to control the affinity of `std::thread` in conjunction with NUMA node-assignment, so we need to add a lower-level thread-pool implementation that directly uses the POSIX API.

ashvardanian

The project currently provides multiple backends for x86: SSE, AVX, AVX-512; but none for Arm. NEON and SVE backends should be added.

ashvardanian

Add Python CUDA reduction benchmark with cuda.cccl

2

Introduces reduce_bench.py, a Python script to benchmark parallel reductions on NVIDIA GPUs using the cuda.cccl library, and updates the README with usage instructions and example output. This allows users to...

AnshSinghSonkhia

Add Python benchmarks for the new CUDA DSL/JIT

2

Now that [CCCL v3](https://github.com/NVIDIA/cccl/releases/tag/v3.0.0) can be used for [efficient parallel reductions in Python](https://developer.nvidia.com/blog/delivering-the-missing-building-blocks-for-nvidia-cuda-kernel-fusion-in-python) it would be great to create an additional benchmark file - `reduce_bench.py` with Python-ic JIT-ed kernels for...

ashvardanian

Thanks for sharing your benchmarks and [your thoughts on your blog](https://ashvardanian.com/posts/beyond-openmp-in-cpp-rust/). Another framework you may want to consider in Rust is [Paralight](https://docs.rs/paralight), which offers the best of both worlds: a...

gendx