| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Throughput seems to be 4x higher.
|
|
|
|
|
|
Not all successfully yet though..
|
|
|
|
|
|
With the meaning that all values are assumed to be positive.
|
|
|
|
So that we have a comparable number for different formats.
|
|
|
|
|
|
Otherwise we get -O2 optimization flag, preventing proper debugging.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ubytes.
|
|
|
|
As it will be shared with more modules.
|
|
|
|
LLVM during optimization reorders the rdtsc instructions next to each
other, which makes it pretty useless.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Straightforward representation of floating-point/fixed-point/integer,
normalized/scaled, signed/unsigned SIMD vector types.
|
|
|
|
|
|
|
|
The code
|
|
|
|
Because this is interpolated (ie. early) depth, we can build in an
assumption about the quads emitted by triangle setup, ie that they
are actually linear spans. Interpolate z over those spans in z16
format to save on math & conversion.
|
|
Disable blend code when no color buffer
|
|
|
