search

jiazhihao / TASO

The Tensor Algebra SuperOptimizer for Deep Learning
Apache License 2.0
693 stars 90 forks source link

copy weights incorrectly in merge_gconv_kernel.cn #79

Closed jiahuiyang closed 3 years ago

jiahuiyang commented 3 years ago

Hi, All

Based on my understanding, the weights should be copied to upper or bottom half in new gconv kernel interleavingly. Howerver, I found it is a sequential process in the optiztimized network. So I change void merge_gconv_kernel() in merge_gconv_kernel.cu as following

  void merge_gconv_kernel(DATATYPE* dst_ptr,
                          const DATATYPE* src_ptr,
                          int volume,
                          int c_in_h_w,
                          int c_out,
                          int count)
  {
    assert(c_out % count == 0);
    CUDA_KERNEL_LOOP(i, volume)
    {
      int mod = i % c_in_h_w;//kernel内位置
      int div = i / c_in_h_w;//所属kernel

      // int dst_i = div * c_in_h_w * count + div / (c_out / count) * c_in_h_w + mod;
      int dst_i = div* c_in_h_w *count + div % count * c_in_h_w + mod;

       dst_ptr[dst_i] = src_ptr[i];
    }
  }

Hoping it helps.