GPU Pro 7

224 V 3D Engine Design

output compressed blocks from the pixel shader. The render target should be

16 times smaller than the source texture, so one texel corresponds to one BC6H

block (16 source texels). The final step is to copy the results from the temporary

render target to the destination BC6H texture using the

CopyResource function.

To achieve optimal performance, the algorithm requires a native float to half

conversion instructions, which are available in Shader Model 5. Additionally, it

is preferable to avoid integer operations altogether (especially 32-bit divides and

multiplies), as they are very costly on modern GPUs. For example, the pop-

ular ...