On AlphaTensor’s new matrix multiplication algorithms

For BC6H encoding in Oodle Texture, we needed a sensible error metric to use in the encoder core. BC6H is HDR which is more challenging than LDR data since we need to handle vast differences in magnitude. BC6H internally essentially treats the float16 bits as 16-bit integers (which is a semi-logarithmic mapping) and works with […]

GPUs support UNORM formats that represent a number inside [0,1] as an 8-bit unsigned integer. In exact arithmetic, the conversion to a floating-point number is straightforward: take the integer and divide it by 255. 8-bit integers are for sure machine numbers (exactly represented) in float32 and so is 255, so if you’re willing to do […]

That’s right, it’s another texture compression blog post! I’ll keep it short. By “solid-color block”, I mean a 4×4 block of pixels that all have the same color. ASTC has a dedicated encoding for these (“void-extent blocks”), BC7 does not. Therefore we have an 8-bit RGBA input color and want to figure out how to […]

The x86 instruction set has a somewhat peculiar set of SIMD integer multiply operations, and Intel’s particular implementation of several of these operations in their headline core designs has certain idiosyncrasies that have been there for literally over 25 years at this point. I don’t actually have any inside information, but it’s fun to...

This one originally came up for me in Oodle Texture’s BC7 decoder. In the BC7 format, each pixel within a 4×4 block can choose from a limited set of between 4 to 16 colors (ignoring some caveats like the dual-index modes that don’t matter here) and consequently between 2 and 4 bits per pixel are […]