A small unpretentious software rasterizer written in Odin, just for fun.

NOTE: The pipeline only supports 4D floating-point vectors. This includes buffers, varyings, uniforms, and framebuffers.

• SIMD rasterization
• Vertex and fragment shaders
• Shader uniforms
• Dynamic attributes/varyings
• Perspective correct
• Texture sampling
• Depth function
• Blend function
• Depth mask
• Clipping
• Configurable face culling
• MRT

Performance isn't "wow", but for a triangle in a 400×225 box on my Ryzen 3600, here are the 10 second averages:

With shader indirection:

Clear:   101.81 µs (9.8%)
Render:   92.15 µs (8.9%)
Blit:    269.04 µs (26.0%)
Present: 572.09 µs (55.3%)

Without shader indirection:

Clear:   114.07 µs (11.3%)
Render:   58.12 µs (5.7%)
Blit:    269.47 µs (26.6%)
Present: 569.83 µs (56.3%)

Render is the part that rasterizes the triangle (~90k pixels).

• With indirection: ~92 µs
• Without indirection: ~58 µs

Which roughly gives (ballpark, please don’t quote me in court):

• Total cycles: 0.000058s * 3.6GHz ~= 208,800 cycles
• Per SIMD iteration: 208'800 / 11'250 ~= 18.6 cycles
• Per pixel: 18.6 / 8 ~= 2.3 cycles

The most expensive step is Blit, because the rendering is done in an internal framebuffer where each component is stored as vectors of 8 floats (SoA), and then needs to be converted to an RGBA8 buffer.

Compilation flags used:

-o:speed -microarch:native

So these tests are rough and I'm not aiming for maximum performance. This is mainly for fun, and if it allows cool offline renders that's more than enough :D