Optimization of loops with deep learning functions

Hello,

I am trying to optimize the following code, which performs the deep learning convolutions on input arrays:

       parfor k=1:500 %number of images
        for j =1:8 %number of channels of the relevant filters
            dldf_O_dlconv3_temp_1(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,1,k),[8 8 1]),'SSC'),0,'Padding',padding);%1st filter
            dldf_O_dlconv3_temp_2(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,2,k),[8 8 1]),'SSC'),0,'Padding',padding);%2nd filter
            dldf_O_dlconv3_temp_3(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,3,k),[8 8 1]),'SSC'),0,'Padding',padding);%3rd filter
            dldf_O_dlconv3_temp_4(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,4,k),[8 8 1]),'SSC'),0,'Padding',padding);%4th filter
            dldf_O_dlconv3_temp_5(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,5,k),[8 8 1]),'SSC'),0,'Padding',padding);%5th filter
            dldf_O_dlconv3_temp_6(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,6,k),[8 8 1]),'SSC'),0,'Padding',padding);%6th filter
            dldf_O_dlconv3_temp_7(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,7,k),[8 8 1]),'SSC'),0,'Padding',padding);%7th filter
            dldf_O_dlconv3_temp_8(:,:,j,k)=dlconv(dlarray(reshape(O_maxpool2(:,:,j,k),[8 8 1]), 'SSC'),dlarray(reshape(DLDO_O_dlconv3(:,:,8,k),[8 8 1]),'SSC'),0,'Padding',padding);%8th filter
        end
       end

As you can see, I am already using the parfor loop. I tried using GPU arrays with O_maxpool2 and DLDO_O_dlconv3, but instead of speeding anything up I think it became a bit slower if anything. My GPU device details are as follows:

         CUDADevice with properties:

                      Name: 'GeForce RTX 2080 Ti'
                     Index: 1
         ComputeCapability: '7.5'
            SupportsDouble: 1
             DriverVersion: 11.2000
            ToolkitVersion: 11
        MaxThreadsPerBlock: 1024
          MaxShmemPerBlock: 49152
        MaxThreadBlockSize: [1024 1024 64]
               MaxGridSize: [2.1475e+09 65535 65535]
                 SIMDWidth: 32
               TotalMemory: 1.1811e+10
           AvailableMemory: 9.4411e+09
       MultiprocessorCount: 68
              ClockRateKHz: 1545000
               ComputeMode: 'Default'
      GPUOverlapsTransfers: 1
    KernelExecutionTimeout: 1
          CanMapHostMemory: 1
           DeviceSupported: 1
           DeviceAvailable: 1
            DeviceSelected: 1

Please let me know if there is anything else I could do to speed this code, and also why using gpuArrays have not sped it up.

Much thanks.

Answer 1

[Joss Knight]

dlconv is designed to work in batch with multiple input channels, multiple filters, and multiple input observations in a single call. Read the documentation for dlconv.

[Radians]

Thanks alot...Even though it was not straight forward what I was trying to do(that's why I had to use forloops), but your comment made me think about how I could utilize the built-in functionality of dlconv to remove the loops, and after reshaping my data a bit I was able to do it. Thanks again.