Speed up the matmul op, use the gemm replace the batch gemm (!22926) · 合并请求 · PaddlePaddle / Paddle

Speed up the matmul op, use the gemm replace the batch gemm !22926

Created by: wawltor

matmul性能优化 1.性能对比1 X = [1024, 10, 12] Y = [12, 10]

未优化前 matmul gpu_time 0.591ms

优化后 matmul gpu_time 0.326ms

速度提升 1.8x

2.性能对比2 X = [128, 50, 1000] Y = [1000, 1000]

未优化前 matmul gpu_time 20.9ms

优化后 matmul gpu_time 15.6ms

速度提升 1.34x

3.性能对比3 X = [512, 50, 1000] Y = [1000, 1000]

未优化前 matmul gpu_time 61.3ms

优化后 matmul gpu_time 43.2ms

速度提升 1.41x