Convolution Operator Optimization

Introduction

This project is about convolution operator optimization on GPU

Content

• Cuda core Implicit GEMM forward
• Cuda core Implicit GEMM backward
• CuTe Implicit GEMM

This blog provides a detailed introduction to the optimization steps.

/cuda Implementation on GPU
  /implicitgemm implicit gemm convolution implementation
  /implicitgemmbwd implicit gemm convolution backward implementation
/cudnn cuDNN test on GPU
/cute Using CuTe implement convolution

Build and run

$ cd cuda/implicitgemm
$ bash implgemm.sh

If you want to change the version of program, just change TARGET in Makefile

Verification

There is verification code in main.cu, which was annotated due to slow running.

// printf("===================start verfiy===================\n");
// direct_conv2dcpu(input, weight, output, n, c, h, w, k, r, s, u, v, p, q);
// int error = 0;
// for (int i = 0; i < n * k * outh * outw; i++)
// {
//     if (abs(output_host[i] - output[i]) > getPrecision(output[i]))
//     {
//         printf("error, postion:%d, gpuvalue:%f, cpuvalue:%f\n", i, output_host[i], output[i]);
//         error++;
//         break;
//     }
// }
// printf("================finish,error:%d=========================\n", error);

If you need to verify the result, just unannotate the above code to verify the correctness of the results.

TODO

• Triton Implicit GEMM
• Tensor core Implicit GEMM
• Winograd-based convolution