Workarounds

24 bit depth support

Not all vulkan platforms support 24 bit depth textures. They prefer to use 32F depth textures. Blender still rely on using 24 bit depth textures, but should be phased out after we fully migrate to Vulkan.

RGB Texture support

GPU_RGB16F isn't support by almost all platforms. This texture format is in used by color management, external render engines and the real-time compositor. If a texture is created this format it is automatically converted to GPU_RGBA16F.

GPU_RGB32F isn't support by almost all platforms. This texture format is in used by color management, external render engines and the real-time compositor. If a texture is created this format it is automatically converted to GPU_RGBA32F.

Data formats without benefits

Vulkan doesn't support texture formats that don't have any benefit performance wise. For example using a 32 bit integer and using it as a float requires conversion, but doesn't have any bandwidth improvements. Therefor in the core of vulkan it isn't allowed to use these configurations. As workaround we convert the 32 bit interger to float when uploading the data to the GPU.

Some platforms don't allow to using 3 component byte data in a vertex buffer. Internally those formats are 32 bits in the vertex buffer. The workaround is to use 4 components and fill the last component with the value 255.

Depth Range

The default OpenGL depth range is between -1 and 1. Before entering the fragment stage the depth value is normalized. Vulkan has a depth range of 0 to 1. Depth values between -1 and 0 are clipped away. To workaround this issue the gl_Position.z is retargetted to 0 to 1.

The main difference is that precision is lost as the order of clipping and normalization is swapped. The same solution is done in the Metal backend.

Framebuffer flipping

OpenGL uses a different coordinate system then Vulkan (and Metal) We need to flip the Y coordinates to convert to the same coordinate system. For Vulkan we flip the coordinate system when blitting/drawing the editors to the swap chain image.

Note

Metal uses a different approach where all Y coordiates are flipped by the vertex stage.