Use tensor core to calculate back-to-back HGEMM (half-precision general matrix multiplication) with MMA PTX instruction. The calculation expression is as follows, where the precision of matrix A (M * K), B (K * N), C (N * L) and D (M * L) is FP16. Tested in special scenarios, the performance can exceed cublas performance by 20% to 120%.

D (M * L) = A (M * K) * B (K * N) * C (N * L)
HGEMM0: P (M * N) = A (M * K) * B (K * N)
HGEMM1: D (M * L) = P (M * N) * C (N * L)

• Simt Naive: naive algorithm using cuda core
• Mma Naive Reg: naive algorithm using tensor core and transfer the intermediate results through shared memeory
• Mma Naive Shm: naive algorithm using tensor core and transfer the intermediate results through register
• Mma Async Reg: async algorithm using tensor core and transfer the intermediate results through shared memeory
• Mma Async Shm: async algorithm using tensor core and transfer the intermediate results through register

• 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_back2back_hgemm.git

cd cuda_back2back_hgemm
./build.sh -a 80 -t Release -b OFF
./build.sh -a 80 -t Debug -b OFF

cd cuda_back2back_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: 11.8

L = 8
N = 16
K = 512

Performance achieved by current optimization methods.

Currently only supports N=16 back-to-back HGEMM.

• N = 32, 64, 128 etc.