Mesa (master): intel/ir/gen12+: Work around FS performance regressions due to SIMD32 discard divergence.

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 4d73988f6fef39e9263ec0bb49cd5efff68393bc
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=4d73988f6fef39e9263ec0bb49cd5efff68393bc

Author: Francisco Jerez <currojerez at riseup.net>
Date:   Sun May 31 14:56:40 2020 -0700

intel/ir/gen12+: Work around FS performance regressions due to SIMD32 discard divergence.

This avoids some performance regressions on Gen12 platforms caused by
SIMD32 fragment shaders reported in titles like Dota2, TF2, Xonotic,
and GFXBench5 Car Chase and Aztec Ruins.

The most obvious pattern in the regressing shaders I identified among
these workloads is that they all had non-uniform discard statements,
which are handled rather optimistically by the current IR analysis
pass: No penalty is currently applied to the SIMD32 variant of the
shader in the form of differing branching weights like we do for other
control flow instructions in order to account for the greater
likelihood of divergence of a SIMD32 shader.

Simply changing that by giving the same treatment to discard
statements as we give to other branching instructions seemed to hurt
more than it helped on platforms earlier than Gen12, since it reversed
most of the improvement obtained from SIMD32 fragment shaders in
Manhattan for no measurable benefit in other workloads (Manhattan has
a handful of shaders with statically non-uniform discard statements
which actually perform better in SIMD32 mode due to their approximate
dynamic uniformity).  For that reason this change is applied to Gen12+
platforms only.

I've been running a number of tests trying to understand the
difference in behavior between Gen12 and earlier platforms, and most
of the evidence I've gathered seems to point at EU fusion being the
culprit: Unlike previous generations, on Gen12 EUs are arranged in
pairs which execute instructions in lockstep, giving an effective warp
size of 64 threads in SIMD32 mode, which seems to increase the
likelihood for control flow divergence in some of the affected shaders
significantly.

Fixes: 188a3659aea6dec9acf1 "intel/ir: Import shader performance analysis pass."
Reported-by: Caleb Callaway <caleb.callaway at intel.com>
Reviewed-by: Matt Turner <mattst88 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5910>

---

 src/intel/compiler/brw_ir_performance.cpp | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/src/intel/compiler/brw_ir_performance.cpp b/src/intel/compiler/brw_ir_performance.cpp
index 8d02db932ba..3c39594d121 100644
--- a/src/intel/compiler/brw_ir_performance.cpp
+++ b/src/intel/compiler/brw_ir_performance.cpp
@@ -1522,9 +1522,19 @@ namespace {
        *       difference is the worst-case scenario branch_weight used for
        *       SIMD32 which accounts for the possibility of a dynamically
        *       uniform branch becoming divergent in SIMD32.
+       *
+       *       Note that we provide slightly more pessimistic weights on
+       *       Gen12+ for SIMD32, since the effective warp size on that
+       *       platform is 2x the SIMD width due to EU fusion, which increases
+       *       the likelihood of divergent control flow in comparison to
+       *       previous generations, giving narrower SIMD modes a performance
+       *       advantage in several test-cases with non-uniform discard jumps.
        */
       const float branch_weight = (dispatch_width > 16 ? 1.0 : 0.5);
+      const float discard_weight = (dispatch_width > 16 || s->devinfo->gen < 12 ?
+                                    1.0 : 0.5);
       const float loop_weight = 10;
+      unsigned discard_count = 0;
       unsigned elapsed = 0;
       state st;
 
@@ -1538,6 +1548,8 @@ namespace {
 
             if (inst->opcode == BRW_OPCODE_ENDIF)
                st.weight /= branch_weight;
+            else if (inst->opcode == FS_OPCODE_PLACEHOLDER_HALT && discard_count)
+               st.weight /= discard_weight;
 
             elapsed += (st.unit_ready[unit_fe] - clock0) * st.weight;
 
@@ -1547,6 +1559,8 @@ namespace {
                st.weight *= loop_weight;
             else if (inst->opcode == BRW_OPCODE_WHILE)
                st.weight /= loop_weight;
+            else if (inst->opcode == FS_OPCODE_DISCARD_JUMP && !discard_count++)
+               st.weight *= discard_weight;
          }
 
          p.block_latency[block->num] = elapsed - elapsed0;