intel/fs: Add a NONE scheduling mode

While our LIFO scheduling mode attempts to optimize for register
pressure, it's often hard for a scheduling algorithm to do better than
the instruction order provided by the shader author.  Shader authors
often do perfectly reasonable things like using texture results
immediately after fetching them or constructing texture coordinates
immediately before the texture op.  When we throw all the instruction
ordering information away, we loose any help the author may have given
us.  By attempting NONE before we fall back to the worst case LIFO mode.

And, yes, I tried this with NONE both before and after LIFO and doing
NONE before LIFO is substantially better, according to shader-db.

    total instructions in shared programs: 19673152 -> 19665202 (-0.04%)
    instructions in affected programs: 33669 -> 25719 (-23.61%)
    helped: 20
    HURT: 0
    helped stats (abs) min: 15 max: 4609 x̄: 397.50 x̃: 107
    helped stats (rel) min: 2.33% max: 67.50% x̄: 14.60% x̃: 9.12%
    95% mean confidence interval for instructions value: -867.61 72.61
    95% mean confidence interval for instructions %-change: -21.74% -7.46%
    Inconclusive result (value mean confidence interval includes 0).

    total cycles in shared programs: 935562500 -> 935020920 (-0.06%)
    cycles in affected programs: 18620349 -> 18078769 (-2.91%)
    helped: 104
    HURT: 48
    helped stats (abs) min: 88 max: 60986 x̄: 8031.48 x̃: 3680
    helped stats (rel) min: 0.61% max: 51.44% x̄: 14.95% x̃: 8.87%
    HURT stats (abs)   min: 10 max: 54724 x̄: 6118.62 x̃: 1530
    HURT stats (rel)   min: 0.13% max: 46.45% x̄: 10.28% x̃: 6.46%
    95% mean confidence interval for cycles value: -5724.34 -1401.71
    95% mean confidence interval for cycles %-change: -9.86% -4.10%
    Cycles are helped.

    total spills in shared programs: 12158 -> 10327 (-15.06%)
    spills in affected programs: 1831 -> 0
    helped: 20
    HURT: 0

    total fills in shared programs: 14749 -> 12635 (-14.33%)
    fills in affected programs: 2114 -> 0
    helped: 20
    HURT: 0

    LOST:   8
    GAINED: 649

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13734>

diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index 32adfea..a3c2eb4 100644 (file)
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -8652,12 +8652,14 @@ fs_visitor::allocate_registers(bool allow_spilling)
    static const enum instruction_scheduler_mode pre_modes[] = {
       SCHEDULE_PRE,
       SCHEDULE_PRE_NON_LIFO,
+      SCHEDULE_NONE,
       SCHEDULE_PRE_LIFO,
    };
 
    static const char *scheduler_mode_name[] = {
       "top-down",
       "non-lifo",
+      "none",
       "lifo"
    };
 
@@ -8697,7 +8699,8 @@ fs_visitor::allocate_registers(bool allow_spilling)
          invalidate_analysis(DEPENDENCY_INSTRUCTIONS);
       }
 
-      schedule_instructions(pre_modes[i]);
+      if (pre_modes[i] != SCHEDULE_NONE)
+         schedule_instructions(pre_modes[i]);
       this->shader_stats.scheduler_mode = scheduler_mode_name[i];
 
       if (0) {

diff --git a/src/intel/compiler/brw_shader.h b/src/intel/compiler/brw_shader.h
index d2eb75d..6afe9c7 100644 (file)
--- a/src/intel/compiler/brw_shader.h
+++ b/src/intel/compiler/brw_shader.h
@@ -38,6 +38,7 @@ enum instruction_scheduler_mode {
    SCHEDULE_PRE_NON_LIFO,
    SCHEDULE_PRE_LIFO,
    SCHEDULE_POST,
+   SCHEDULE_NONE,
 };
 
 #define UBO_START ((1 << 16) - 4)