https://github.com/Themaister/parallel-rdp



These fossils contain very large and complex shaders. The small_*.foz
files use 8/16-bit arithmetic.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>

Konstantin Seurer authored 1 year ago and Samuel Pitoiset committed 1 year ago

Aligns RADV stage names to the common ones, including combined stages.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

Konstantin Seurer authored 1 year ago and Samuel Pitoiset committed 1 year ago

This patch adds handling for all possible stage names returned by
_mesa_shader_stage_to_string.

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>

This adds a property to indicate that a statistic change makes all
other statistics uncomparable. This is quite useful on Intel where
shaders are compiled in different variants and a SIMD8 compute shader
cannot be compared to a SIMD16 variant. All that should be visible is
the change of SIMD width.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>

reported by pycodestyle

also change the text

The per-app table is better and the biggest changes are printed
with shader file names.

It only printed shaders that also had regressions.

This reverts commit 47ded5b3.

Mistake on my part, `V3D_DEBUG=shaderdb` is actually for manually
dumping stats from random apps in the shader-db format so that they can
be read using shader-db's `report.py`.

Having `V3D_DEBUG=shaderdb` in shader-db makes it dump the information
twice, causing issues.

Signed-off-by: Eric Engestrom <eric@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>

This fixes ./report-fossil.py --help, which previously just croaked
with an error about not being able to detect the driver, because no .csv
files were specified on the command line.  Which isn't a great error
message when you're...trying to figure out the command line.

While at it, we also drop the requirement that filenames must end with
.csv.  The code to check for that was basically to try and distinguish
filenames from other arguments...which we don't need to do now that we
properly do argument parsing beforehand.  They still have to be .csv
files, of course, but mandating extensions is a bit unusual.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>

Almost all the time ;)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

We don't want to compare a FS16 to a FS8.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

Signed-off-by: Eric Engestrom <eric@engestrom.ch>
Acked-by: David Heidelberg <david.heidelberg@collabora.com>

Ian Romanick authored 3 years ago and Matt Turner committed 2 years ago

Discovered when an experiment resulted in a driver that only gave OpenGL
3.0 without GL_ARB_uniform_buffer_object.

All changes made with this script:

for ext in GL_ARB_shader_image_load_store \
           GL_ARB_shader_storage_buffer_object \
	   GL_ARB_shading_language_420pack \
	   GL_ARB_uniform_buffer_object \
	   GL_ARB_explicit_attrib_location
do
    egrep -lr "extension[[:space:]]+${ext}.*enable" . |\
        while read s; do
            # If the shader already has the requirement, skip
            if grep -A10 '[[]require]' $s | grep -q ^$ext ; then
                continue
            fi

            # If the shader has ifdefs for the extension, skip
            if egrep -q "#[[:space:]]*if.*${ext}" $s; then
                continue
            fi

            sed --in-place -e "s/\(^GLSL >.*$\)/\1\n${ext}/" $s
        done
done

These are OpenCL kernels compiled through clvk-0.0.1-r20 on ChromeOS, to
be used by Google Meet.

Filip Gawin authored 2 years ago and Lionel Landwerlin committed 2 years ago

Acked-by: Alejandro Piñeiro <apinheiro@igalia.com>

M Henning authored 2 years ago and Mel Henning committed 2 years ago

Obtained by clicking a bunch of buttons with the default cube open.

Acked-by: Alejandro Piñeiro <apinheiro@igalia.com>

'p' is an obsolete option so it needs to be removed from print_usage.
And there is no reason to print out "Unknown option:" error string since
it is actually done by getopt (e.g. "./run: invalid option -- 'p'").

Signed-off-by: Dongwon Kim <dongwon.kim@intel.com>

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>

Eric Engestrom authored 2 years ago and Eric Engestrom committed 2 years ago

Signed-off-by: Eric Engestrom <eric@igalia.com>

Eric Engestrom authored 2 years ago and Eric Engestrom committed 2 years ago

Signed-off-by: Eric Engestrom <eric@igalia.com>

Acked-by: Iago Toral Quiroga <itoral@igalia.com>

By default Sascha Willems demos uses for their demos using a graphics
pipeline the D32_SFLOAT_S8_UINT format. v3dv doesn't support it, but
the demos properly detect that, and use a supported format in any
case.

But as the fossils were recorded with a driver supporting that format,
it tries to use it, and we got a renderpass not supported warning, and
no shader compilation.

As fossils here are mostly used for the shading compiling statistics,
changing the format shouldn't affect to other drivers. Tested with anv
and there is no changes.

As we are here we also add some extra fossils.

Acked-by: Dylan Baker <dylan@pnwbakers.com>
Acked-by: Emma Anholt <emma@anholt.net>

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Acked-by: Matt Turner <mattst88@gmail.com>

Obtained by clicking around in many of the example projects from
https://github.com/godotengine/godot-demo-projects

 commit 2e40f67b1b
in godot v3.4.4

Acked-by: Emma Anholt <emma@anholt.net>

Captured from gtk4-widget-factory on gtk 4.6.0

Reviewed-by: Emma Anholt <emma@anholt.net>

This fixes a bug where the scheduler mode for every shader was detected
as "scheduled" instead of the actual mode name.

Code above this block was detecting "scheduled with mode X" and parsing
that as field = word 0 = "scheduled" and val = word 3 = "X".  Then this
code came along and did the default parsing of "N things" as value =
word 0 and field = word 1, and overrode that.  We should just put it
in the else case so each case happens independently.

Looks like this has been broken from the day it was commited. I
beleive this likely happened when Ken changed his original
patchset to stop looking for the extension and instead look for
this string.

Reviewed-by: Emma Anholt <emma@anholt.net>

MESA_GLSL_CACHE_DISABLE envvar was renamed to MESA_SHADER_CACHE_DISABLE.

v2:
 - Keep old envvar too (Marcin).

Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>

imirkin wants the current format for shader-db for nouveau, but I
really want the current report.py script to work because it helps me
dig into which shaders are the problem.

i915 doesn't have loops at all, so just skip the check.

The performance of the binary really doesn't matter, it's Mesa that
does.  Found when my build was SIGILLing on my test system.

radv-report-fossil is more generic than anv-report-fossil, and has seen
more development work, so make it the "generic" report-fossil.py. Add
support for turnip by adding statistics and executable names generated
by turnip. Also fold in the driver detection logic from report-fossil
and call out to anv-report-fossil if the driver is detected to be anv.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>