Method submitKernels

devices/cpu/cpu_conv.cpp:68–101

  }

  void CPUConv::submitKernels(const Ref<CancellationToken>& ct)
  {
    if (!src || !dst || !weight || !bias)
      throw std::logic_error("convolution argument not set");

    ispc::CPUConvKernel kernel;
    kernel.src    = *src;
    kernel.weight = *weight;
    kernel.bias   = *bias;
    kernel.dst    = *dst;
    kernel.relu   = activation == Activation::ReLU;

    engine->submitFunc([=]
    {
      const int OH = kernel.dst.H; // output height
      const int OW = kernel.dst.W; // output width
      const size_t N = size_t(OCBB) * OH * OWC; // total number of work items

      parallel_for(N, [&](size_t i)
      {
        const size_t j = i / OCBB;
        const int ocbb = int(i % OCBB); // output channel block block index
        const int oh   = int(j % OH);   // output height index
        const int owc  = int(j / OH);   // output width chunk index

        constexpr int PW = 1; // KW = 3
        const int owr = OWC * (blockOW - PW - 1);
        const int owBegin = owc   > 0   ? (owc     * OW + owr) / (OWC*blockOW) * blockOW + PW : 0;
        const int owEnd   = owc+1 < OWC ? ((owc+1) * OW + owr) / (OWC*blockOW) * blockOW + PW : OW;

        ispc::CPUConvKernel_run_f32(&kernel, blockOCB, ocbb * blockOCB, oh, owBegin, owEnd);
      });
    }, ct);
  }

OIDN_NAMESPACE_END