The implementation method refer to the maxas.
- The test environment： ubuntu18.04, cuda10, 1080ti
- The code only supports limited input matrix, not universal adaptation, only for learning. Here is the GFLOP for testing different size matrices
| N | cublas | sgemm | sgemm/cublas |
|---|---|---|---|
| 512 | 4451.6069 | 3587.3280 | 80% |
| 1024 | 7856.5241 | 6640.6945 | 84% |
| 2048 | 9409.4447 | 8769.9500 | 93% |
| 4096 | 10180.4288 | 9708.4873 | 95% |