Several optimization methods of half-precision general matrix multiplication (HGEMM) using tensor core with WMMA API and MMA PTX instruction. The calculation expression is as follows, where the precision of matrix A (M * K), B (K * N) and C (M * N) is FP16. Through exploring various matrix tiling and optimization methods, the current performance between 256 to 16384 dimensions is not less than 95% of the performance of cublas, and in many scenarios, it exceeds the performance of cublas.
C (M * N) = A (M * K) * B (K * N)
- Tiling: 256 * 128 for block tiling size and 64 * 64 for warp tiling size
- Coalescing Access: using wide instruction access to global memory
- Data Reuse: using shared memory to reuse data of matrix A and B
- Async Copy: using asynchronous copy operation with non-blocking instruction
- Bank Conflict: using padding method for WMMA API and permuted method for MMA PTX instruction to eliminate bank conflict
- L2 Cache: using swizzle access mode to increase L2 cache hit ratio
- Register Reuse: calculating as "Right Left Right Left" for the internal tile of warp
- Pg2s: double-buffer algorithm using prefetching global memory to shared memory
- Ps2r: double-buffer algorithm using prefetching shared memory to register
- Stage: multi-buffer algorithm using prefetching global memory to shared memory
- OS: Linux
- Cmake Version: >= 3.12
- GCC Version: >= 4.8
- CUDA Version: >= 11.0
- Gflags: install on ubuntu as follows
sudo apt-get install libgflags-dev
git clone https://github.com/Bruce-Lee-LY/cuda_hgemm.git
cd cuda_hgemm
./build.sh -a 80 -t Release -b OFF
./build.sh -a 80 -t Debug -b OFF
cd cuda_hgemm
./build.sh -a 86 -t Release -b OFF
./build.sh -a 86 -t Debug -b OFF
./run_sample.sh
Process the data in the log and plot it as a line chart.
cd tools/performance
./performance.sh
- CUDA Version: 11.3
The best performance that can be achieved.
Performance achieved by current optimization methods.
- CUDA Version: 11.3
The best performance that can be achieved.