Function operator ()

devices/gpu/gpu_input_process.h:82–176

    }

    oidn_device_inline void operator ()(const oidn_private WorkGroupItem<2>& it) const
    {
      const int hDst = it.getGlobalID<0>();
      const int wDst = it.getGlobalID<1>();

      const int h = hDst - tile.hDstBegin;
      const int w = wDst - tile.wDstBegin;

      // Gather and process the input channel values
      float values[dstPaddedC] = {}; // = 0

      if (h >= 0 && h < tile.H && w >= 0 && w < tile.W)
      {
        const int hSrc = h + tile.hSrcBegin;
        const int wSrc = w + tile.wSrcBegin;

        const vec3f inputValue = getInput(hSrc, wSrc);
        values[0] = inputValue.x;
        values[1] = inputValue.y;
        values[2] = inputValue.z;

        if (dstPaddedC >= 6 && albedo.ptr)
        {
          const vec3f albedoValue = getAlbedo(hSrc, wSrc);
          values[3] = albedoValue.x;
          values[4] = albedoValue.y;
          values[5] = albedoValue.z;

          if (dstPaddedC >= 9 && normal.ptr)
          {
            const vec3f normalValue = getNormal(hSrc, wSrc);
            values[6] = normalValue.x;
            values[7] = normalValue.y;
            values[8] = normalValue.z;
          }
        }
      }

    #if defined(OIDN_COMPILE_SYCL) || defined(OIDN_COMPILE_CUDA) || defined(OIDN_COMPILE_HIP)
      // Transpose the values in the subgroup into coalesced blocks and store them to memory (fast)
      // All work-items in the subgroup are assumed to be in the same row
      const int subgroupLocalID = it.getSubgroupLocalID();
      const int wDstBegin = it.subgroupBroadcast(wDst, 0);
      GlobalPtr<DstT> dstPtr = &dst(0, hDst, wDstBegin);

      #if defined(OIDN_COMPILE_SYCL)
      // The subgroup size is assumed to be equal to the channel count
      constexpr int subgroupSize = dstPaddedC;

      #pragma unroll
      for (int i = 0; i < subgroupSize; ++i)
      {
        float dstBlock = 0;

        #pragma unroll
        for (int c = 0; c < min(dstPaddedC, 9); ++c) // only up to 9 non-zero channels
        {
          const auto value = it.subgroupBroadcast(values[c], i);