1xN convolutions2019-05-21T09:50:02+00:00
Viewing 0 reply threads
  • Author
    • saseptimsaseptim
      Post count: 2


      I did not notice any support for non-square convolutions. In audio processing, one could use 1×3, 1×5, etc – 1D convolutions.
      Do you have plans to support these convolutions? I guess one solution would be to use a 3×3 filter where only the first 3 are non-zero but it seems quite wasteful.

      Also, I am looking at the basic convolution implementation and trying to understand the line:
      gap8_clip(gap8_roundnorm_reg(Acc, Norm), 7)

      I understand that you are shifting Acc right by Norm and then clipping to int8 (byte). Where does “Norm” come from? I see that in the mnist example, you use 14.


      static void __attribute__ ((noinline)) KerConvNxNStrideS_Body_fps(
      signed char *__restrict__ In,
      signed char *__restrict__ Out,
      signed char *__restrict__ Filter,
      int Fw,
      int Fh,
      int W,
      int H,
      int Wo,
      int Wo_F,
      int Wo_L,
      int Ho,
      int Ho_F,
      int Ho_L,
      int Stride,
      v4s Pad,
      int Norm
      unsigned short int PadL = Pad[0], PadT = Pad[2];

      signed char *PtO = Out+Wo*Ho_F+Wo_F;
      signed char *PtC = Filter;
      for ( int h=Ho_F; h

Viewing 0 reply threads
  • You must be logged in to reply to this topic.