Matrix Multiplication (MMul) Benchmarks

This repository contains a number of serial and parallel benchmarks for matrix multiplication in C++. Matrix multiplication is a wonderful first operation to try your hand at optimizing for the following reasons:

• It is a very common operation in popular fields (e.g., ML)
• The optimizations are fairly easy to understand (they primarily deal with simple access patterns)
• The optimizations are composable (they work better together!)
• It is fairly easy to parallelize

And many more!

The benchmarks in this repository were written using Google Benchmark. For simplicity, all benchmarks assume square matrix of dimension N x N, where N is 384, 768, and 1152.

Benchmarks

The following section breaks down the benchmarks contained in each subdirectory.

Baseline

• serial_mmul_bench
  • Baseline serial mmul implementation (using the classical triply-nested for loop)
• parallel_mmul_bench
  • Baseline parallel mmul implementation (splits rows of output matrix across threads)

Blocked

• blocked_mmul_bench
  • A serial mmul implementation which processes a block of elements at a time to exploit locality in the B matrix
• blocked_aligned_mmul_bench
  • Same as blocked_mmul_bench but using 64-byte aligned allocations to prevent blocks from spanning cache lines
• parallel_blocked_mmul_bench
  • A parallel blocked mmul implementation (splits rows of output matrix across threads)

Blocked Column

• blocked_column_aligned_mmul_bench
  • A serial mmul implementation which processes a block of elements at a time, but traverses output blocks of elements in column-major order to exploit locality in the columns of B between blocks of output elements
• parallel_blocked_column_mmul_bench
  • A parallel blocked column implementation (splits columns between threads) where work is statically mapped

Blocked Column Multi Output

• blocked_column_multi_output_aligned_mmul_bench
  • A serial mmul implementation which processes a tile of output elements at a time, exploiting locality from each tile of B across output tile elements
• parallel_blocked_column_multi_output_mmul_bench
  • A parallel blocked column multi output implementation (splits columns between threads) where work is statically mapped

GPU Implementations (CUDA)

• baseline_cuda_mmul
  • A naive mmul implementation for NVIDIA GPUs written in CUDA
• shmem_cuda_mmul
  • A cache-tiled mmul implementation for NVIDIA GPUs using shared memory

Contact Information

• Email: [email protected]
• YouTube: Link
• Twitter: Link