Compositor: implement bicubic interpolation for Stabilize2DNode #122288

Habib Gahbiche · 2024-05-26T19:44:12+02:00

Habib Gahbiche commented

2024-05-26 19:44:12 +02:00

This patch also fixes a crash when image input of Stabilize2DNode is unconnected and interpolation is set to bicubic.

Differences to GPU compositor:

~1px difference is observed due to different rounding of domain size vs. canvas size. This difference is constant for all image sizes.
If image input is unconnected but other inputs are, CPU compositor doesn't consider the operation to be constant, whereas GPU compositor still outputs a constant result.

Todos:

update tests

This patch also fixes a crash when image input of `Stabilize2DNode` is unconnected and interpolation is set to bicubic. Differences to GPU compositor: - ~1px difference is observed due to different rounding of domain size vs. canvas size. This difference is constant for all image sizes. - If image input is unconnected but other inputs are, CPU compositor doesn't consider the operation to be constant, whereas GPU compositor still outputs a constant result. Todos: - [x] update tests

Habib Gahbiche added 3 commits 2024-05-26 19:44:22 +02:00

Fix bicubic interpolation 0b08c715ca

Fix build error 1988b0763b

Fix: unconnected Stabilize2dNode crashes with bicubic interp

buildbot/vexp-code-patch-lint Build done.

Details

buildbot/vexp-code-patch-linux-x86_64 Build done.

Details

buildbot/vexp-code-patch-darwin-x86_64 Build done.

Details

buildbot/vexp-code-patch-windows-amd64 Build done.

Details

buildbot/vexp-code-patch-darwin-arm64 Build done.

Details

buildbot/vexp-code-patch-coordinator Build done.

Details

051029a1b7

Habib Gahbiche requested review from Omar Emara 2024-05-26 19:49:24 +02:00

Omar Emara requested changes 2024-05-27 09:48:05 +02:00

Dismissed

Omar Emara left a comment

Didn't a similar change introduce bad results according to a past bug report? (I can't find it right now)
Do you remember that report?

Didn't a similar change introduce bad results according to a past bug report? (I can't find it right now) Do you remember that report?

source/blender/compositor/intern/COM_MemoryBuffer.h

						
				@ -291,2 +291,2 @@

				    math::interpolate_cubic_bspline_fl(

				        buffer_, out, this->get_width(), this->get_height(), num_channels_, x, y);

				    if (is_a_single_elem_) {

				      memcpy(out, buffer_, get_elem_bytes_len());

Omar Emara commented

2024-05-27 09:46:11 +02:00

This probably won't be optimized away by the compiler, so I would rather hard code it using a switch case or something.

Habib Gahbiche commented

2024-05-27 10:22:33 +02:00

This is the same way read_elem_bilinear() is implemented. I'll see if there is a better way.

This is the same way `read_elem_bilinear()` is implemented. I'll see if there is a better way.

Habib Gahbiche commented

2024-05-30 14:04:16 +02:00

From a quick test, the following code slows down Stabilize 2D Node from 220ms to 280ms (disconnected image input) and translate node (image input disconnected, X input connected to image) from 1ms to 20ms on a 8000x4000 image:

if (is_a_single_elem_) {
  switch (datatype_) {
    case DataType::Value:
      out[0] = buffer_[0];
      break;
    case DataType::Float2:
      copy_v2_v2(out, buffer_);
      break;
    case DataType::Vector:
      copy_v3_v3(out, buffer_);
      break;
    case DataType::Color:
    default:
      copy_v4_v4(out, buffer_);
  }
  return;
}

So hardcoding doesn't seem faster. Or did you have something else in mind?

Also, I think these are rather exotic cases, in most cases the code inside if (is_a_single_elem_) will either get called once (unconnected node) or not at all (node connected to valid input). So I suggest to keep it simple and use memcpy()

From a quick test, the following code slows down Stabilize 2D Node from 220ms to 280ms (disconnected image input) and translate node (image input disconnected, X input connected to image) from 1ms to 20ms on a 8000x4000 image: ```cpp if (is_a_single_elem_) { switch (datatype_) { case DataType::Value: out[0] = buffer_[0]; break; case DataType::Float2: copy_v2_v2(out, buffer_); break; case DataType::Vector: copy_v3_v3(out, buffer_); break; case DataType::Color: default: copy_v4_v4(out, buffer_); } return; } ``` So hardcoding doesn't seem faster. Or did you have something else in mind? Also, I think these are rather exotic cases, in most cases the code inside `if (is_a_single_elem_)` will either get called once (unconnected node) or not at all (node connected to valid input). So I suggest to keep it simple and use `memcpy()`

Omar Emara commented

2024-05-30 14:56:34 +02:00

That doesn't make sense, see: https://quick-bench.com/q/cgUMr8uXSU4PxLbwjiGohYnzImg.

But since this is copied from similar code, we can look into it later.

That doesn't make sense, see: https://quick-bench.com/q/cgUMr8uXSU4PxLbwjiGohYnzImg. But since this is copied from similar code, we can look into it later.

Habib Gahbiche commented

2024-05-30 15:58:12 +02:00

My tests only include the case where datatype_ == Color and the value to be copied is always the same. For number_of_elements = 4 results look different: https://quick-bench.com/q/FV5Ai_bHNldJ6aw3kLZKEBpWWPU

My tests only include the case where `datatype_ == Color` and the value to be copied is always the same. For `number_of_elements = 4` results look different: https://quick-bench.com/q/FV5Ai_bHNldJ6aw3kLZKEBpWWPU

Omar Emara commented

2024-05-30 16:03:15 +02:00

@zazizizou Well the benchmark is not very useful when you set number_of_elements to 4 in the source, because the compiler will detect that and remove the memcpy call entirely. While this is not known at compile time in real world cases.

Habib Gahbiche commented

2024-05-30 16:55:44 +02:00

Ah yes true. It's still not completely random though and that might play a role.

OmarEmaraDev marked this conversation as resolved

Omar Emara commented

2024-05-27 09:48:35 +02:00

@blender-bot build

Habib Gahbiche commented

2024-05-27 10:22:51 +02:00

Yes, a previous patch for translate node introduced this bug: #121436.
The difference now is the argument get_relative_x(x) is given to the function interpolate_cubic_bspline_fl() instead of just x. So this is the proper fix for the bug above.

Yes, a previous patch for translate node introduced this bug: https://projects.blender.org/blender/blender/issues/121436. The difference now is the argument `get_relative_x(x)` is given to the function `interpolate_cubic_bspline_fl()` instead of just `x`. So this is the proper fix for the bug above.

Omar Emara approved these changes 2024-05-30 14:57:31 +02:00

Habib Gahbiche referenced this issue from a commit

2024-05-30 17:09:33 +02:00

Compositor: update CPU distort tests after bicubic sampling fix.

Habib Gahbiche added 2 commits 2024-05-30 17:11:01 +02:00

Merge remote-tracking branch 'origin/main' into fix-bicubic-interp 43f396bb57

Update tests

buildbot/vexp-code-patch-lint Build done.

Details

buildbot/vexp-code-patch-linux-x86_64 Build done.

Details

buildbot/vexp-code-patch-windows-amd64 Build done.