The scissor test is implemented as another per-quad operation in
the JIT code. The four scissor box params are passed via the
lp_jit_context. In the JIT code we compare the quad's x/y coords
against the clip bounds and create a new in/out mask that's AND'd
with the main quad mask.
Note: we should also do scissor testing in the triangle setup code
to improve efficiency. That's not done yet.
When the incoming c0,c1,c2 values are equal to INT_MIN it means that
all pixels are inside the triangle. Thus we can skip the detailed
pixel inside/outside triangle tests. Use the new lp_build_if()/endif()
functions to generate the branching code.
The code is disabled ATM however because it's actually a little slower
than the original code. A little more tuning may fix that though...
Was used only as a reference, since texture sampling is now code generated.
Already axed in the lp-binning branch too.
This fixes the llvmpipe build after recent sampling changes.
The setup tiling engine is now plugged directly into the draw module
as a rendering backend.
Removed a couple of layering violations such that the setup code no
longer reaches out into the surrounding llvmpipe state or context.
That is:
- check for no op
- update/flush draw module
- update bound state and mark it as dirty
In particular flushing the draw module is important since it may contain
unflushed primitives which would otherwise be draw with wrong state.
The test to determine which of the pixels in a 2x2 quad is now done in
the fragment shader rather than in the calling C code. This is a little
faster but there's a few more things to do.
Note that the step[] array elements are in a different order now. Rather
than being in row-major order for the 4x4 grid, they're in "quad-major"
order. The setup of the step arrays is a little more complicated now.
So is the course/intermediate tile test code, but some lookup tables
help with that.
Next steps:
- early-cull 2x2 quads which are totally outside the triangle.
- skip the in/out test for fully contained quads
- make the in/out comparison code tighter/faster.
This matches the convention used by the recursive rasterizer.
Also fixed assorted typos, comments, etc.
Now tri-z.c, gears.c, etc look basically right but there's still some
cracks in triangle rasterization.
This was redundant as drivers can just keep track of whether they are
inside a begin/end query pair. We want to add more query types later
and also support nested queries, none of which map well onto a flag like
this. No driver appeared to be using the flag.
New control flow helper functions which keep track of all variables
and generate the correct Phi functions.
This re-enables skipping the fs execution of quads masked out by
the rasterizer, early z testing, and kill opcode.
This yields a performance improvement of around 20%.
Finally a substantial performance improvement: framerates of apps using
texturing tripled, and furthermore, enabling/disabling texturing only
affects around 15% of the framerate, which means the bottleneck is now
somewhere else.
Generated texture sampling code is not complete though -- we always
sample from the base level -- so final figures will be different.
Special attention is given to the interpolation of side by side quads.
Multiplications are made only for the first quad. Interpolation of
inputs for posterior quads are done exclusively with additions, and
perspective divide if necessary.
Brian Paul
José Fonseca
Keith Whitwell
Michal Krol
Vinson Lee
Zack Rusin