introduction

A simple high performance CUDA GEMM, Block Sparse GEMM and Non-uniform Quantized GEMM implementation.

C = alpha * A * B + beta * C

algorithm

located in src/cuda/

• MatrixMulCUDA
  • one element of C is assigned one thread
  • global memory coalesce of B
• MatrixMulCUDA1
  • texture load
• MatrixMulCUDA2
  • one 4 * 4 grid of C is assigned one thread
• MatrixMulCUDA3
  • vectorized A B load
• MatrixMulCUDA4
  • vectorized C store
• MatrixMulCUDA5
  • block sparse version
• MatrixMulCUDA6
  • vectorized A B load coalesce
• MatrixMulCUDA7
  • warp shuffle to enable C store coalesce
• MatrixMulCUDAQuantize8bit
  • 8 bit non-uniform quantized matmul

experiments

located in benchmark/

• benchmark_dense
  • Compare My Gemm with Cublas
• benchmark_sparse
  • Compare My block sparse Gemm with Cusparse
• benchmark_quantization_8bit
  • Compare My Gemm with Cublas
• benchmark_quantization
  • Compare My Gemm with My quantized non-uniform 8 bit Gemm

TODO

• (MatrixMulCUDA7) write back to C matrix, warp shuffle to enable global memory coalesce
• (MatrixMulCUDA8) double buffering

run

make benchmark_[experiment name]
bash scripts/benchmark_[experiment name].sh

Note

• sparsity约为1%的时候, cusparse的性能可以超越cublas
• 合理分配寄存器 尽可能让参数在编译器确定节省计算资源和寄存器数目