freedreno/ir3: scheduler improvements

For instructions that increase the # of live values, apply a threshold
to avoid scheduling them too early.  And factor the net change of # of
live values that would result from scheduling an instruction, to
prioritize instructions that reduce number of live values as the number
of live values increases.

For manhattan:

  total instructions in shared programs: 27869 -> 28413 (1.95%)
  instructions in affected programs: 26756 -> 27300 (2.03%)
  helped: 102
  HURT: 87

  total full in shared programs: 1903 -> 1719 (-9.67%)
  full in affected programs: 1390 -> 1206 (-13.24%)
  helped: 124
  HURT: 9

The reduction in register usage nets ~20% gain in manhattan.  (So
getting mediump support should be a huge win for gles gfxbench.)

Also significantly helps some of the more complex shadertoy shaders,
like IQ's Piano (32 to 18 regs, doubles fps).

The effect is less pronounced on smaller shaders.

Signed-off-by: Rob Clark <[email protected]>
Reviewed-by: Eric Anholt <[email protected]>

1 files changed, 5 insertions, 0 deletions

diff --git a/src/freedreno/ir3/ir3.h b/src/freedreno/ir3/ir3.h
index f3c25ea2792..ccd102b8e44 100644
--- a/src/freedreno/ir3/ir3.h
+++ b/src/freedreno/ir3/ir3.h
@@ -292,6 +292,11 @@ struct ir3_instruction {
 	};
 
 	/* used for per-pass extra instruction data.
+	 *
+	 * TODO we should remove the per-pass data like this and 'use_count'
+	 * and do something similar to what RA does w/ ir3_ra_instr_data..
+	 * ie. use the ir3_count_instructions pass, and then use instr->ip
+	 * to index into a table of pass-private data.
 	 */
 	void *data;