i965/fs: Switch to per-subspan discard jumps.

ANY4H is more efficient than ANY8H and ANY16H because it makes sure
that whenever a whole subspan hits a discard statement it gets
disabled by the EU until the end of the program, regardless of whether
the discard condition is uniform across all channels of the SIMD8-16
thread.  OTOH ANY8H/ANY16H would cause the rest of the program to be
executed for *all* channels if only one of the channels hadn't taken
the discard branch, potentially increasing the bandwidth and ALU usage
of the program unnecessarily.

This change increases the FPS by over 3x of a simple micro-benchmark
that discards a bunch of fragments and then does a single costly
texturing operation.  I've just re-verified the FPS change on HSW and
SKL, but I expect all platforms from Gen6 up to get a similar benefit.

Note that we could potentially be more aggressive and use the NORMAL
predicate to discard individual channels, but that would need to
happen post-scheduling because the scheduler currently doesn't care to
reorder HALT instructions with respect to other instructions, and the
NORMAL predicate would cause the results of subsequent derivative
computations to become undefined -- If the scheduler didn't reorder
HALT instructions it would actually be safe to switch to NORMAL
because the behavior of derivative computations after a non-uniform
discard statement is undefined by the GLSL spec, but that would make
the optimization implemented by one of the following commits somewhat
more difficult.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>

1 files changed, 1 insertions, 3 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index b89c672..c23b558 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -1397,9 +1397,7 @@ fs_visitor::emit_discard_jump()
    fs_inst *discard_jump = bld.emit(FS_OPCODE_DISCARD_JUMP);
    discard_jump->flag_subreg = 1;
 
-   discard_jump->predicate = (dispatch_width == 8)
-                             ? BRW_PREDICATE_ALIGN1_ANY8H
-                             : BRW_PREDICATE_ALIGN1_ANY16H;
+   discard_jump->predicate = BRW_PREDICATE_ALIGN1_ANY4H;
    discard_jump->predicate_inverse = true;
 }