Accessing tile pixel values, and skipping denoise step for tile

Hello! I am working on implementing a small optimization (but huge for my use-case) centered around denoising. I want to check whether a tile is completely transparent, and if so skip the denoising for that tile. I’ve been through session.cpp and I implemented something that I think is on the right path. However, the pipeline is somewhat opaque to me; I’m not sure that I am trying to implement this in the right place.

Basically, I have this function:

bool FullyTransparent(ccl::Tile* tile){
      float* in_combined;
      in_combined = tile->buffers[0];

      BufferParams *params = &tile->buffers[0].params;
      int pass_stride = params->get_passes_size();
      int size = params->width * params->height;
      int offset = 0;
      int str = 0;
      float *in = tile->buffers[0] + offset;

        for (int i = 0; i < size; i++, in += pass_stride, in_combined += pass_stride, in_combined += 4) {
           if(in_combined[3] > 0.01f)
              return false;
        return true;

Down in session::AcquireTile, where we normally have rtile.task = RenderTile::DENOISE; I added a little check:

  if (tile->state == Tile::DENOISE) {
     if (!FullyTransparent(tile)) {
      rtile.task = RenderTile::DENOISE;
     else {
          tile->state = tile->DENOISED;

I’m pretty sure that FullyTransparent() is on the right track (I know that it doesn’t work quite right), but I’m not at all sure if I am on the right track about where to actually put the check.

I’m not looking for somebody to write the code for me, just point me in the right direction please!

1 Like

I would add this check in the device code that does the denoising, for example CPUDevice::denoise_openimagedenoise_buffer. For two reasons:

  • You don’t need to change the overall rendering logic, I don’t think changing the task or state in the tile will be sufficient.
  • It ensures your pixels have been copied to CPU memory, which your code seems to assume. If you are using a GPU denoiser, you’d need to either copy them to the CPU or do the test on the GPU.

Ah, I understand. I hadn’t considered that the rendering pass data might not be immediately CPU accessible. I figured that implementing it in the renderer pipeline would make it more generically useful, but as it happens I AM only interested in the OpenImage denoiser.

It looks like I can probably just pop the check it into denoise_openimagedenoise. Thanks!

Hehey, I got it! With almost no understanding of what I was doing! I think that’s a skill :stuck_out_tongue:

I’d never submit a pull requested for this code, but just adding this to the start of the denoising function cut my render time by a third (will probably be 2 thirds when I modify my project to take advantage of it):

  int offset = 0;
  int str = 0;

  BufferParams *params = &rtile.buffers[0].params;
  params->get_offset_stride(str, offset);
  int size = params->width * params->height;

  float *in = rtile.buffers[0] + offset;
  float *in_combined = rtile.buffers[0];

  bool AllTransparent = true;
  for (int i = 0; i < size;
       i++, in += pass_stride, in_combined += pass_stride) {
    if (in_combined[3] > 0.01f) {
      AllTransparent = false;
  if (AllTransparent) {

I requested this exact feature some time ago on RGC! Really hope you manage to make this and have it implemented, congrats!

It’s a little bit tricky, because 99% of the stuff I do with Blender these days is work-related. I would need a sign-off from on high to work on anything for submission upstream. Also, this needs to be toggle-able in the settings so there is a bit more to do.
But you know, if you want a custom build the fix is right there :wink: