D3D12: Crash when rendering a non-square texture with mipmaps

[BlueCube3310]

Tested versions

Reproducible in 4.3-stable, 4.4-dev1

System information

Windows 10.0.19045 - d3d12 (Forward+) - dedicated NVIDIA GeForce RTX 4060 Ti (NVIDIA; 31.0.15.5244) - AMD Ryzen 9 5900X 12-Core Processor (24 Threads)

Issue description

The engine crashes when trying to create a non-square texture with mipmaps on the GPU under D3D12. The crash always occurs, even if both the texture's width and height are powers of 2.

Steps to reproduce

1. Open the MRP.
2. Click on aa.png to preview it in the inspector.
3. The crash should occur.

Minimal reproduction project (MRP)

d3d12-mip-crash.zip

[alvinhochun]

Here it is crashing with 0xc0000094 (STATUS_INTEGER_DIVIDE_BY_ZERO) at

godot/drivers/d3d12/rendering_device_driver_d3d12.cpp

Line 4890 in 61598c5

uint32_t row_pitch = image_size / (p_regions[i].texture_region_size.y * computed_d) * block_h;

p_regions[i].texture_region_size.y is 0:

p_regions._ptr[9]
(const RenderingDeviceDriver::BufferTextureCopyRegion) {
  buffer_offset = 0
  texture_subresources = {
    aspect = (value = 1)
    mipmap = 9
    base_layer = 0
    layer_count = 1
  }
  texture_offset = {
     = {
       = (x = 0, y = 0, z = 0)
      coord = ([0] = 0, [1] = 0, [2] = 0)
    }
  }
  texture_region_size = {
     = {
       = (x = 1, y = 0, z = 1)
      coord = ([0] = 1, [1] = 0, [2] = 1)
    }
  }
}

Just a quick guess, it might be coming from

godot/servers/rendering/rendering_device.cpp

Lines 1632 to 1647 in 61598c5

	for (uint32_t i = 0; i < tex->mipmaps; i++) {
	RDD::BufferTextureCopyRegion copy_region;
	copy_region.buffer_offset = mip_layouts[i].offset;
	copy_region.texture_subresources.aspect = tex->read_aspect_flags;
	copy_region.texture_subresources.mipmap = i;
	copy_region.texture_subresources.base_layer = p_layer;
	copy_region.texture_subresources.layer_count = 1;
	copy_region.texture_region_size.x = w;
	copy_region.texture_region_size.y = h;
	copy_region.texture_region_size.z = d;
	command_buffer_texture_copy_regions_vector.push_back(copy_region);
	w = (w >> 1);
	h = (h >> 1);
	d = MAX(1u, d >> 1);
	}

Might be a regression from #94069 (CC @RandomShaper)