Mesa (master): tgsi_exec: Fix NaN behavior of min and max

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 939bf7a4198cecff57510b7fe5d38ef9b5dd22be
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=939bf7a4198cecff57510b7fe5d38ef9b5dd22be

Author: Ian Romanick <ian.d.romanick at intel.com>
Date:   Thu Apr 22 17:46:39 2021 -0700

tgsi_exec: Fix NaN behavior of min and max

Modern shader APIs, like DX10 and GLSL 1.30, want min() and max() to
"cleanse" NaN.  If one source is NaN, the other value should be chosen.
If both sources are NaN, the result may be either.

There are many cases where TGSI is generate from NIR, and many
optimizations in NIR expect this behavior.  Not meeting these
expectations can lead to unexpected results.

Reviewed-by: Eric Anholt <eric at anholt.net>
Fixes: ffe58739da9 ("Softpipe: import TGSI tree. Not hooked-up yet.")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10419>

---

 src/gallium/auxiliary/tgsi/tgsi_exec.c             | 32 +++++++++++-----------
 src/gallium/drivers/softpipe/ci/softpipe-quick.txt |  4 ---
 2 files changed, 16 insertions(+), 20 deletions(-)

diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c b/src/gallium/auxiliary/tgsi/tgsi_exec.c
index b4c6f60f888..fbfb04ef360 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_exec.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c
@@ -263,20 +263,20 @@ static void
 micro_dmax(union tgsi_double_channel *dst,
            const union tgsi_double_channel *src)
 {
-   dst->d[0] = src[0].d[0] > src[1].d[0] ? src[0].d[0] : src[1].d[0];
-   dst->d[1] = src[0].d[1] > src[1].d[1] ? src[0].d[1] : src[1].d[1];
-   dst->d[2] = src[0].d[2] > src[1].d[2] ? src[0].d[2] : src[1].d[2];
-   dst->d[3] = src[0].d[3] > src[1].d[3] ? src[0].d[3] : src[1].d[3];
+   dst->d[0] = src[0].d[0] > src[1].d[0] || isnan(src[1].d[0]) ? src[0].d[0] : src[1].d[0];
+   dst->d[1] = src[0].d[1] > src[1].d[1] || isnan(src[1].d[1]) ? src[0].d[1] : src[1].d[1];
+   dst->d[2] = src[0].d[2] > src[1].d[2] || isnan(src[1].d[2]) ? src[0].d[2] : src[1].d[2];
+   dst->d[3] = src[0].d[3] > src[1].d[3] || isnan(src[1].d[3]) ? src[0].d[3] : src[1].d[3];
 }
 
 static void
 micro_dmin(union tgsi_double_channel *dst,
            const union tgsi_double_channel *src)
 {
-   dst->d[0] = src[0].d[0] < src[1].d[0] ? src[0].d[0] : src[1].d[0];
-   dst->d[1] = src[0].d[1] < src[1].d[1] ? src[0].d[1] : src[1].d[1];
-   dst->d[2] = src[0].d[2] < src[1].d[2] ? src[0].d[2] : src[1].d[2];
-   dst->d[3] = src[0].d[3] < src[1].d[3] ? src[0].d[3] : src[1].d[3];
+   dst->d[0] = src[0].d[0] < src[1].d[0] || isnan(src[1].d[0]) ? src[0].d[0] : src[1].d[0];
+   dst->d[1] = src[0].d[1] < src[1].d[1] || isnan(src[1].d[1]) ? src[0].d[1] : src[1].d[1];
+   dst->d[2] = src[0].d[2] < src[1].d[2] || isnan(src[1].d[2]) ? src[0].d[2] : src[1].d[2];
+   dst->d[3] = src[0].d[3] < src[1].d[3] || isnan(src[1].d[3]) ? src[0].d[3] : src[1].d[3];
 }
 
 static void
@@ -1357,10 +1357,10 @@ micro_max(union tgsi_exec_channel *dst,
           const union tgsi_exec_channel *src0,
           const union tgsi_exec_channel *src1)
 {
-   dst->f[0] = src0->f[0] > src1->f[0] ? src0->f[0] : src1->f[0];
-   dst->f[1] = src0->f[1] > src1->f[1] ? src0->f[1] : src1->f[1];
-   dst->f[2] = src0->f[2] > src1->f[2] ? src0->f[2] : src1->f[2];
-   dst->f[3] = src0->f[3] > src1->f[3] ? src0->f[3] : src1->f[3];
+   dst->f[0] = src0->f[0] > src1->f[0] || isnan(src1->f[0]) ? src0->f[0] : src1->f[0];
+   dst->f[1] = src0->f[1] > src1->f[1] || isnan(src1->f[1]) ? src0->f[1] : src1->f[1];
+   dst->f[2] = src0->f[2] > src1->f[2] || isnan(src1->f[2]) ? src0->f[2] : src1->f[2];
+   dst->f[3] = src0->f[3] > src1->f[3] || isnan(src1->f[3]) ? src0->f[3] : src1->f[3];
 }
 
 static void
@@ -1368,10 +1368,10 @@ micro_min(union tgsi_exec_channel *dst,
           const union tgsi_exec_channel *src0,
           const union tgsi_exec_channel *src1)
 {
-   dst->f[0] = src0->f[0] < src1->f[0] ? src0->f[0] : src1->f[0];
-   dst->f[1] = src0->f[1] < src1->f[1] ? src0->f[1] : src1->f[1];
-   dst->f[2] = src0->f[2] < src1->f[2] ? src0->f[2] : src1->f[2];
-   dst->f[3] = src0->f[3] < src1->f[3] ? src0->f[3] : src1->f[3];
+   dst->f[0] = src0->f[0] < src1->f[0] || isnan(src1->f[0]) ? src0->f[0] : src1->f[0];
+   dst->f[1] = src0->f[1] < src1->f[1] || isnan(src1->f[1]) ? src0->f[1] : src1->f[1];
+   dst->f[2] = src0->f[2] < src1->f[2] || isnan(src1->f[2]) ? src0->f[2] : src1->f[2];
+   dst->f[3] = src0->f[3] < src1->f[3] || isnan(src1->f[3]) ? src0->f[3] : src1->f[3];
 }
 
 static void
diff --git a/src/gallium/drivers/softpipe/ci/softpipe-quick.txt b/src/gallium/drivers/softpipe/ci/softpipe-quick.txt
index 94a71929e0d..d77786acd8d 100644
--- a/src/gallium/drivers/softpipe/ci/softpipe-quick.txt
+++ b/src/gallium/drivers/softpipe/ci/softpipe-quick.txt
@@ -2684,10 +2684,6 @@ spec/glsl-1.10/execution/fs-dfdx-accuracy: warn
 spec/glsl-1.10/execution/fs-dfdy-accuracy: warn
 spec/glsl-1.10/preprocessor/extension-defined-test: skip
 spec/glsl-1.10/preprocessor/extension-if-1: skip
-spec/glsl-1.20/execution/fs-nan-builtin-max: fail
-spec/glsl-1.20/execution/fs-nan-builtin-min: fail
-spec/glsl-1.20/execution/vs-nan-builtin-max: fail
-spec/glsl-1.20/execution/vs-nan-builtin-min: fail
 spec/glsl-1.30/execution/tex-miplevel-selection texturegrad 1d: fail
 spec/glsl-1.30/execution/tex-miplevel-selection texturegrad 1darray: fail
 spec/glsl-1.30/execution/tex-miplevel-selection texturegrad 1darrayshadow: fail