Attachment 143048 Details for Bug 109216 – intel/compiler: Try nir_opt_if once before nir_opt_peephole_select

[patch] intel/compiler: Try nir_opt_if once before nir_opt_peephole_select

0001-intel-compiler-Try-nir_opt_if-once-before-nir_opt_pe.patch (text/plain), 2.82 KB, created by Ian Romanick on 2019-01-09 23:52:21 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Ian Romanick

Created: 2019-01-09 23:52:21 UTC

Size: 2.82 KB

patch

obsolete

>From e54d29e881d2f5bce62d58b55d9e417c945558ae Mon Sep 17 00:00:00 2001
>From: Ian Romanick <ian.d.romanick@intel.com>
>Date: Wed, 9 Jan 2019 14:16:35 -0800
>Subject: [PATCH] intel/compiler: Try nir_opt_if once before
> nir_opt_peephole_select
>MIME-Version: 1.0
>Content-Type: text/plain; charset=UTF-8
>Content-Transfer-Encoding: 8bit
>
>The four helped sharders are two pair of SIMD8/SIMD16 fragment shaders
>in Bioshock Infinite.  In all cases the shaders are helped because some
>comparisons are moved closer to the if-statements that use the
>condition, and the extra comparison to regenerat the flag is no longer
>necessary.
>
>This should also fix the loop-unrolling regression in every SPIR-V
>shader generated by glslang.
>
>All Gen7+ platforms had similar results. (Skylake shown)
>total instructions in shared programs: 15047460 -> 15047452 (<.01%)
>instructions in affected programs: 820 -> 812 (-0.98%)
>helped: 4
>HURT: 0
>helped stats (abs) min: 2 max: 2 xÌ: 2.00 xÌ: 2
>helped stats (rel) min: 0.98% max: 0.98% xÌ: 0.98% xÌ: 0.98%
>95% mean confidence interval for instructions value: -2.00 -2.00
>95% mean confidence interval for instructions %-change: -0.98% -0.98%
>Instructions are helped.
>
>total cycles in shared programs: 369988163 -> 369988143 (<.01%)
>cycles in affected programs: 7970 -> 7950 (-0.25%)
>helped: 4
>HURT: 0
>helped stats (abs) min: 2 max: 8 xÌ: 5.00 xÌ: 5
>helped stats (rel) min: 0.11% max: 0.36% xÌ: 0.24% xÌ: 0.24%
>95% mean confidence interval for cycles value: -10.51 0.51
>95% mean confidence interval for cycles %-change: -0.46% -0.01%
>Inconclusive result (value mean confidence interval includes 0).
>
>Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=109216
>Fixes: 8fb8ebfbb05 ("intel/compiler: More peephole select")
>---
> src/intel/compiler/brw_nir.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
>diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
>index 77c6c4e0e5a..6ba59932b0f 100644
>--- a/src/intel/compiler/brw_nir.c
>+++ b/src/intel/compiler/brw_nir.c
>@@ -569,6 +569,16 @@ brw_nir_optimize(nir_shader *nir, const struct brw_compiler *compiler,
>       OPT(nir_opt_dce);
>       OPT(nir_opt_cse);
> 
>+      /* Try nir_opt_if once before doing nir_opt_peephole_select.  glslang
>+       * generates for-loops with a pattern that nir_opt_peephole_select will
>+       * convert to a nir_op_bcsel.  However, this prevents nir_opt_if
>+       * (opt_peel_loop_initial_if specifically) from moving the code to the
>+       * end of the loop and eliminating the if-statement).  This prevents the
>+       * loop analysis code from detecting the induction variable, and the
>+       * loop is therefore not unrolled.
>+       */
>+      OPT(nir_opt_if);
>+
>       /* Passing 0 to the peephole select pass causes it to convert
>        * if-statements that contain only move instructions in the branches
>        * regardless of the count.
>-- 
>2.14.5
>

Actions: View | Splinter Review

Attachments on bug 109216: 142959 | 143043 | 143048

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.