platform/upstream/mesa.git
2 years agoradv/ci: remove unused files
Samuel Pitoiset [Mon, 7 Mar 2022 13:49:00 +0000 (14:49 +0100)]
radv/ci: remove unused files

These files are no longer used.

Fixes: cc327a0fe45 ("amd, ci: Remove unused runners.")
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Charlie Turner <cturner@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15264>

2 years agofreedreno: add a420 deqp-runner files
Ilia Mirkin [Tue, 16 Nov 2021 23:19:37 +0000 (18:19 -0500)]
freedreno: add a420 deqp-runner files

This doesn't actually get run in CI, but this helps track outstanding
issues / expectations. This is from a run on my IFC6540 with A420.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15251>

2 years agofreedreno/a4xx: expose shaders and images, as well as ES 3.1
Ilia Mirkin [Sun, 14 Nov 2021 18:06:49 +0000 (13:06 -0500)]
freedreno/a4xx: expose shaders and images, as well as ES 3.1

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15251>

2 years agofreedreno/ir3: disable conversion folding on a4xx
Ilia Mirkin [Sat, 4 Dec 2021 00:06:12 +0000 (19:06 -0500)]
freedreno/ir3: disable conversion folding on a4xx

Experiments suggest that e.g.

add.u r0.y, hr0.x, hr0.y

will result in the summed value in both the high and low words of r0.y.
This only happens with odd registers, not even ones (r0.x works fine).

Seen in the bit_count lowering (which turns out to be unnecessary, but
this is still a larger problem).

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15251>

2 years agofreedreno/ir3: no need to count bits 16b at a time for a4xx
Ilia Mirkin [Fri, 3 Dec 2021 08:04:19 +0000 (03:04 -0500)]
freedreno/ir3: no need to count bits 16b at a time for a4xx

This also works out nicely since a4xx has some sort of problem with the
16b-based lowering.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15251>

2 years agofreedreno/a4xx: improve condition for disabling early z
Ilia Mirkin [Sat, 20 Nov 2021 07:30:37 +0000 (02:30 -0500)]
freedreno/a4xx: improve condition for disabling early z

This helps some subtests in the early-z piglit test, but leaves one
occlusion-based test still failing.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15251>

2 years agofreedreno/a4xx: extend astc and tg4 workarounds to compute shaders
Ilia Mirkin [Wed, 17 Nov 2021 00:10:46 +0000 (19:10 -0500)]
freedreno/a4xx: extend astc and tg4 workarounds to compute shaders

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15251>

2 years agoRevert "lavapipe: accurately set image/ssbo access based on shader usage"
Mike Blumenkrantz [Mon, 7 Mar 2022 23:00:42 +0000 (18:00 -0500)]
Revert "lavapipe: accurately set image/ssbo access based on shader usage"

This reverts commit 821a49981ff386559f8a8fdf6bf3526b8deb2415.

still flaky

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15271>

2 years agointel/perf: Destination array calculation into function
Matt Turner [Thu, 3 Mar 2022 23:30:37 +0000 (15:30 -0800)]
intel/perf: Destination array calculation into function

Cuts 119 KiB from iris_dri.so and libvulkan_intel.so.

   text    data     bss     dec     hex filename
 917511       0       0  917511   e0007 meson-generated_.._intel_perf_metrics.c.o (before)
 796986       0       0  796986   c293a meson-generated_.._intel_perf_metrics.c.o (after)

   text    data     bss     dec     hex filename
14130948 365708  210004 14706660 e067e4 iris_dri.so (before)
14009332 365708  210004 14585044 de8cd4 iris_dri.so (after)

   text    data     bss     dec     hex filename
8124225  214264   22820 8361309  7f955d libvulkan_intel.so (before)
8002609  214264   22820 8239693  7dba4d libvulkan_intel.so (after)

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Fix mistake in description string
Matt Turner [Thu, 3 Mar 2022 07:28:18 +0000 (23:28 -0800)]
intel/perf: Fix mistake in description string

Along with fixing the grammar, this allows it to be deduplicated since
the properly worded description exists in later generations' XMLs.

Cuts 96 B from iris_dri.so and libvulkan_intel.so.

   text    data     bss     dec     hex filename
 917613       0       0  917613   e006d meson-generated_.._intel_perf_metrics.c.o (before)
 917511       0       0  917511   e0007 meson-generated_.._intel_perf_metrics.c.o (after)

   text    data     bss     dec     hex filename
14131044 365708  210004 14706756 e06844 iris_dri.so (before)
14130948 365708  210004 14706660 e067e4 iris_dri.so (after)

   text    data     bss     dec     hex filename
8124321  214264   22820 8361405  7f95bd libvulkan_intel.so (before)
8124225  214264   22820 8361309  7f955d libvulkan_intel.so (after)

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Mark intel_perf_counter_* enums as PACKED
Matt Turner [Wed, 2 Mar 2022 02:49:26 +0000 (18:49 -0800)]
intel/perf: Mark intel_perf_counter_* enums as PACKED

Reduces their sizes from 4 bytes to 1. Cuts 6 KiB from iris_dri.so and
libvulkan_intel.so.

   text    data     bss     dec     hex filename
 924401       0       0  924401   e1af1 meson-generated_.._intel_perf_metrics.c.o (before)
 917613       0       0  917613   e006d meson-generated_.._intel_perf_metrics.c.o (after)

   text    data     bss     dec     hex filename
14137732 365708  210004 14713444 e08264 iris_dri.so (before)
14131044 365708  210004 14706756 e06844 iris_dri.so (after)

   text    data     bss     dec     hex filename
8131009  214264   22820 8368093  7fafdd libvulkan_intel.so (before)
8124321  214264   22820 8361405  7f95bd libvulkan_intel.so (after)

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Store indices to strings rather than pointers
Matt Turner [Mon, 31 Jan 2022 21:16:26 +0000 (13:16 -0800)]
intel/perf: Store indices to strings rather than pointers

The compiler does a good job of deduplicating strings already, but we
can eliminate the pointers to each string by combining the strings into
a single char array and storing only an index into that array.

The longest of the char arrays is the descriptions array, which is a
little over 45 KiB, so still under MSVC's 64 KiB string literal limit
[0]. Because the string length is under 64 KiB we can use uint16_t as
the index type, which roughly doubles our savings as compared to an int.

This cuts 77 KiB from iris_dri.so (0.5%) and libvulkan_intel.so (0.9%).

   text    data     bss     dec     hex filename
 926811   25920       0  952731   e899b meson-generated_.._intel_perf_metrics.c.o (before)
 924401       0       0  924401   e1af1 meson-generated_.._intel_perf_metrics.c.o (after)

   text    data     bss     dec     hex filename
14190852 391628  210004 14792484 e1b724 iris_dri.so (before)
14137732 365708  210004 14713444 e08264 iris_dri.so (after)

   text    data     bss     dec     hex filename
8184097  240184   22820 8447101  80e47d libvulkan_intel.so (before)
8131009  214264   22820 8368093  7fafdd libvulkan_intel.so (after)

relinfo:
iris_dri.so (before): 17765 relocations, 17545 relative (98%), 452 PLT entries, 1 for local syms (0%), 0 users
iris_dri.so (after) : 15605 relocations, 15385 relative (98%), 452 PLT entries, 1 for local syms (0%), 0 users

libvulkan_intel.so (before): 10720 relocations, 6989 relative (65%), 355 PLT entries, 1 for local syms (0%), 0 users
libvulkan_intel.so (after) :  8560 relocations, 4829 relative (56%), 355 PLT entries, 1 for local syms (0%), 0 users

[0] https://docs.microsoft.com/en-us/cpp/cpp/string-and-character-literals-cpp?view=msvc-170&viewFallbackFrom=vs-2019

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Use slimmer intel_perf_query_counter_data struct
Matt Turner [Thu, 3 Mar 2022 20:24:02 +0000 (12:24 -0800)]
intel/perf: Use slimmer intel_perf_query_counter_data struct

intel_perf_query_counter contains fields for things we can't or don't
want to store in our static data (like runtime-determined max values) or
oa_read_counter function pointers which are dependent on the GPU gen and
would make deduplication very ineffective.

Cuts 16 KiB from iris_dri.so and libvulkan_intel.so.

   text    data     bss     dec     hex filename
 926811   43200       0  970011   ecd1b meson-generated_.._intel_perf_metrics.c.o (before)
 926811   25920       0  952731   e899b meson-generated_.._intel_perf_metrics.c.o (after)

   text    data     bss     dec     hex filename
14190852 408908  210004 14809764 e1faa4 iris_dri.so (before)
14190852 391628  210004 14792484 e1b724 iris_dri.so (after)

   text    data     bss     dec     hex filename
8184097  257464   22820 8464381  8127fd libvulkan_intel.so (before)
8184097  240184   22820 8447101  80e47d libvulkan_intel.so (after)

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Use a function to initialize perf counters
Matt Turner [Thu, 3 Mar 2022 01:53:02 +0000 (17:53 -0800)]
intel/perf: Use a function to initialize perf counters

And specifically mark it with ATTRIBUTE_NOINLINE. Otherwise it will be
inlined and actually slightly increase code size.

Cuts 505 KiB from iris_dri.so and libvulkan_intel.so.

   text    data     bss     dec     hex filename
1538720       0       0 1538720  177aa0 meson-generated_.._intel_perf_metrics.c.o (before)
 926811   43200       0  970011   ecd1b meson-generated_.._intel_perf_metrics.c.o (after)

   text    data     bss     dec     hex filename
14751700 365708  210004 15327412 e9e0b4 iris_dri.so (before)
14190852 408908  210004 14809764 e1faa4 iris_dri.so (after)

   text    data     bss     dec     hex filename
8744913  214264   22820 8981997  890ded libvulkan_intel.so (before)
8184097  257464   22820 8464381  8127fd libvulkan_intel.so (after)

Relocations increase because the counter initializations are moved from
code (in .text) to pointers (in .text) to .rodata, which require
relocations.

relinfo:
iris_dri.so (before): 15605 relocations, 15385 relative (98%), 452 PLT entries, 1 for local syms (0%), 0 users
iris_dri.so (after) : 17765 relocations, 17545 relative (98%), 452 PLT entries, 1 for local syms (0%), 0 users

libvulkan_intel.so (before):  8560 relocations, 4829 relative (56%), 355 PLT entries, 1 for local syms (0%), 0 users
libvulkan_intel.so (after) : 10720 relocations, 6989 relative (65%), 355 PLT entries, 1 for local syms (0%), 0 users

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Deduplicate perf counters
Matt Turner [Thu, 3 Mar 2022 01:26:17 +0000 (17:26 -0800)]
intel/perf: Deduplicate perf counters

No changes in resulting code (yes, seriously!). GCC constant propagates
the static const arrays into the code, yielding bit for bit identical
results. This will however enable further cleanups.

Before this patch, we emit 11916 different initializations of
intel_perf_query_counter. With this patch we emit an array of 539 and
initialize the intel_perf_query_counters in terms of those.

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Don't print leading space from desc_units()
Matt Turner [Thu, 3 Mar 2022 21:13:38 +0000 (13:13 -0800)]
intel/perf: Don't print leading space from desc_units()

Just an annoyance I noticed when I needed to generate the description
string in two different places.

Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agointel/perf: Move some static blocks of C code out of the python script.
Emma Anholt [Mon, 7 Feb 2022 19:06:18 +0000 (11:06 -0800)]
intel/perf: Move some static blocks of C code out of the python script.

Now my editor can help me format code as I type.

Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15237>

2 years agovenus: fix properties of unsupported external fences/semaphores
Chia-I Wu [Sat, 5 Mar 2022 17:39:36 +0000 (09:39 -0800)]
venus: fix properties of unsupported external fences/semaphores

compatibleHandleTypes should be cleared.

Fixed dEQP-VK.api.external.semaphore.sync_fd.info_timeline.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15266>

2 years agoiris/ci: Mark amd_performance_monitor tests as flakes.
Ian Romanick [Thu, 3 Mar 2022 00:38:45 +0000 (16:38 -0800)]
iris/ci: Mark amd_performance_monitor tests as flakes.

On one attempt to merge !15210 failed because
spec@amd_performance_monitor@measure,UnexpectedPass.  In that same
report, the other two tests were listed under "some flakes."

Reviewed-by: Emma Anholt <emma@anholt.net>
Acked-by: Nanley Chery <nanley.g.chery@intel.com>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15220>

2 years agoturnip: Add "rast_order" debug option to force rast order access
Danylo Piliaiev [Mon, 7 Mar 2022 10:13:45 +0000 (12:13 +0200)]
turnip: Add "rast_order" debug option to force rast order access

Enables rasterization order attachment access for all pipelines,
see VK_ARM_rasterization_order_attachment_access for details.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15262>

2 years agogallium/tc: warn if an app is incompatible with cpu_storage
Pierre-Eric Pelloux-Prayer [Mon, 21 Feb 2022 19:14:02 +0000 (20:14 +0100)]
gallium/tc: warn if an app is incompatible with cpu_storage

Instead of silently ignoring unmap calls.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15074>

2 years agoradeonsi: enable tc cpu_storage by default
Pierre-Eric Pelloux-Prayer [Fri, 18 Feb 2022 09:32:05 +0000 (10:32 +0100)]
radeonsi: enable tc cpu_storage by default

Enable for all applications for all small buffers.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15074>

2 years agogallium/u_threaded: late alloc cpu_storage
Pierre-Eric Pelloux-Prayer [Fri, 18 Feb 2022 09:28:58 +0000 (10:28 +0100)]
gallium/u_threaded: late alloc cpu_storage

Instead of allocating cpu_storage in threaded_resource_init, defer the
allocation to first use (in tc_buffer_map).
This avoids needless memory allocation if tc_buffer_disable_cpu_storage is
called before tc_buffer_map.

map_buffer_alignment is stored and serves as a "can cpu_storage be used" flag.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15074>

2 years agoradeonsi: use 1 shader compilation thread if NIR_PRINT is used
Pierre-Eric Pelloux-Prayer [Thu, 17 Feb 2022 13:46:06 +0000 (14:46 +0100)]
radeonsi: use 1 shader compilation thread if NIR_PRINT is used

This avoids getting multiple shaders NIR code being printed at the
same time making the log unusable.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15074>

2 years agonir: Fix source type for fragment_fetch_amd.
Georg Lehmann [Fri, 4 Mar 2022 12:32:06 +0000 (13:32 +0100)]
nir: Fix source type for fragment_fetch_amd.

Like txf_ms, these take integers not floats.

Signed-off-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15242>

2 years agoradeonsi/tests: always add the --gpu argument
Pierre-Eric Pelloux-Prayer [Fri, 4 Mar 2022 10:49:08 +0000 (11:49 +0100)]
radeonsi/tests: always add the --gpu argument

To get the same argument in multi and single gpu situation.

Fixes: 21b01538331 ("radeonsi/tests: print PCI-id of GPU device under test")

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Mihai Preda <mhpreda@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15240>

2 years agoradeonsi: change rounding mode to round to even
Pierre-Eric Pelloux-Prayer [Fri, 4 Mar 2022 10:23:20 +0000 (11:23 +0100)]
radeonsi: change rounding mode to round to even

Use ROUND_TO_EVEN instead of TRUNCATE; this matches what pal and radv do.

This fixes the spec@ext_framebuffer_multisample@turn-on-off tests.

Cc: mesa-stable
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Mihai Preda <mhpreda@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15240>

2 years agoegl/wayland: fix crash in dri2_initialize_wayland_swrast
José Expósito [Thu, 3 Mar 2022 07:20:46 +0000 (08:20 +0100)]
egl/wayland: fix crash in dri2_initialize_wayland_swrast

When "dri2_wl_formats_init" fails in "dri2_initialize_wayland_swrast",
the "dri2_display_destroy" function is called for clean up. However, the
"dri2_egl_display" was not associated with the display in its
"DriverData" field yet.

The following cast in "dri2_display_destroy":

  struct dri2_egl_display *dri2_dpy = dri2_egl_display(disp);

Expands to:

  _EGL_DRIVER_TYPECAST(drvname ## _display, _EGLDisplay, obj->DriverData)

Crashing.

Signed-off-by: José Expósito <jose.exposito89@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13972>

2 years agoegl/wayland: fix crash in dri2_initialize_wayland_drm
José Expósito [Sun, 28 Nov 2021 16:57:25 +0000 (17:57 +0100)]
egl/wayland: fix crash in dri2_initialize_wayland_drm

When "dri2_wl_formats_init" fails in "dri2_initialize_wayland_drm", the
"dri2_display_destroy" function is called for clean up. However, the
"dri2_egl_display" was not associated with the display in its
"DriverData" field yet.

The following cast in "dri2_display_destroy":

  struct dri2_egl_display *dri2_dpy = dri2_egl_display(disp);

Expands to:

  _EGL_DRIVER_TYPECAST(drvname ## _display, _EGLDisplay, obj->DriverData)

Crashing.

Addresses-Coverity-ID: 1494541 ("Resource leak")
Signed-off-by: José Expósito <jose.exposito89@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13972>

2 years agozink: hide descriptor debug behind #ifdef
Mike Blumenkrantz [Fri, 4 Mar 2022 16:06:30 +0000 (11:06 -0500)]
zink: hide descriptor debug behind #ifdef

I've gotten a feel for this, but it's annoying to always see it spamming away

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: invalidate non-punted recycled descriptor sets that are not valid
Mike Blumenkrantz [Thu, 3 Mar 2022 17:25:41 +0000 (12:25 -0500)]
zink: invalidate non-punted recycled descriptor sets that are not valid

these sets may contain refs from the descriptors which need to be removed
to avoid invalid memory access if the ref is leaked

cc: mesa-stable

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: stop leaking descriptor sets
Mike Blumenkrantz [Thu, 3 Mar 2022 17:16:33 +0000 (12:16 -0500)]
zink: stop leaking descriptor sets

when migrating a recycled set here, the set was previously invalid and
in the recycled table, meaning it can be reused directly so long as
it's first invalidated

the previous code would instead pop a different set off the allocation array,
leaking this one

cc: mesa-stable

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: mark fbfetch push sets as non-cached
Mike Blumenkrantz [Thu, 3 Mar 2022 14:23:03 +0000 (09:23 -0500)]
zink: mark fbfetch push sets as non-cached

these can't be cached, so ensure the value isn't uninitialized

Test case 'KHR-GL46.blend_equation_advanced.blend_all.GL_HARDLIGHT_KHR_all_qualifier'..
==1193311== Conditional jump or move depends on uninitialised value(s)
==1193311==    at 0x634EF05: update_push_ubo_descriptors (zink_descriptors.c:1230)
==1193311==    by 0x634FCC5: zink_descriptors_update (zink_descriptors.c:1412)
==1193311==    by 0x63A5EA1: void zink_draw<(zink_multidraw)1, (zink_dynamic_state)3, true, false>(pipe_context*, pipe_draw_info const*, unsigned int, pipe_draw_indirect_info const*, pipe_draw_start_count_bias const*, unsigned int, pipe_vertex_state*, unsigned int) (zink_draw.cpp:788)
==1193311==    by 0x635A0AF: void zink_draw_vbo<(zink_multidraw)1, (zink_dynamic_state)3, true>(pipe_context*, pipe_draw_info const*, unsigned int, pipe_draw_indirect_info const*, pipe_draw_start_count_bias const*, unsigned int) (zink_draw.cpp:907)
==1193311==    by 0x6174397: tc_call_draw_single (u_threaded_context.c:3150)
==1193311==    by 0x616C403: tc_batch_execute (u_threaded_context.c:211)
==1193311==    by 0x616CA44: _tc_sync (u_threaded_context.c:362)
==1193311==    by 0x61725E9: tc_texture_map (u_threaded_context.c:2274)
==1193311==    by 0x5A9DAE9: pipe_texture_map_3d (u_inlines.h:572)
==1193311==    by 0x5A9EB80: st_ReadPixels (st_cb_readpixels.c:530)
==1193311==    by 0x5A0647A: read_pixels (readpix.c:1178)
==1193311==    by 0x5A0647A: _mesa_ReadnPixelsARB (readpix.c:1195)
==1193311==    by 0x5A06517: _mesa_ReadPixels (readpix.c:1210)

cc: mesa-stable

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: fix descriptor cache pointer array allocation
Mike Blumenkrantz [Wed, 2 Mar 2022 20:49:29 +0000 (15:49 -0500)]
zink: fix descriptor cache pointer array allocation

this got mixed up during some refactor and started indexing based on the
number of bindings instead of the number of descriptors, which means
that array descriptor bindings would have overlapping array memory

cc: mesa-stable

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: wait on program cache fences before destroying programs
Mike Blumenkrantz [Wed, 2 Mar 2022 13:48:52 +0000 (08:48 -0500)]
zink: wait on program cache fences before destroying programs

if these still have outstanding cache jobs, deleting the object now
will cause a crash

maybe fixes some cts flakiness?

cc: mesa-stable

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: use a fence for pipeline cache update jobs
Mike Blumenkrantz [Wed, 2 Mar 2022 20:50:50 +0000 (15:50 -0500)]
zink: use a fence for pipeline cache update jobs

otherwise there's nothing to wait on

cc: mesa-stable

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: add function for refcounting zink_program structs
Mike Blumenkrantz [Wed, 2 Mar 2022 13:48:28 +0000 (08:48 -0500)]
zink: add function for refcounting zink_program structs

step one in maybe cleaning up and deduplicating some program code

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agozink: always update shader variants when rebinding a gfx program
Mike Blumenkrantz [Tue, 1 Mar 2022 17:06:13 +0000 (12:06 -0500)]
zink: always update shader variants when rebinding a gfx program

the variant data may have changed in the meanwhile, so do a pass to
ensure that the correct variant is used

cc: mesa-stable

fixes caselist:
dEQP-GLES31.functional.shaders.sample_variables.sample_mask.discard_half_per_two_samples.multisample_texture_4
dEQP-GLES31.functional.shaders.sample_variables.sample_mask.discard_half_per_two_samples.singlesample_rbo

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15226>

2 years agolavapipe: accurately set image/ssbo access based on shader usage
Mike Blumenkrantz [Thu, 3 Mar 2022 18:37:51 +0000 (13:37 -0500)]
lavapipe: accurately set image/ssbo access based on shader usage

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15233>

2 years agolavapipe: scan shaders for image/ssbo access and generate per-stage masks
Mike Blumenkrantz [Thu, 3 Mar 2022 18:37:13 +0000 (13:37 -0500)]
lavapipe: scan shaders for image/ssbo access and generate per-stage masks

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15233>

2 years agolavapipe: heap-allocate rendering_state struct
Mike Blumenkrantz [Fri, 4 Mar 2022 18:17:06 +0000 (13:17 -0500)]
lavapipe: heap-allocate rendering_state struct

this thing is like 28k now, which is just way too big to have on the stack

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15233>

2 years agogallivm: avoid division by zero when computing cube face
Mike Blumenkrantz [Fri, 4 Mar 2022 15:59:12 +0000 (10:59 -0500)]
gallivm: avoid division by zero when computing cube face

this is illegal and produces NaNs which blow up the sample instr

cc: mesa-stable

fixes (llvmpipe and zink):
KHR-GL45.incomplete_texture_access.sampler
dEQP-GLES31.functional.program_uniform.by_pointer.render.array_in_struct.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_pointer.render.array_in_struct.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_pointer.render.array_in_struct.sampler2D_samplerCube_vertex
dEQP-GLES31.functional.program_uniform.by_pointer.render.basic.samplerCube_both
dEQP-GLES31.functional.program_uniform.by_pointer.render.basic.samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_pointer.render.basic.samplerCube_vertex
dEQP-GLES31.functional.program_uniform.by_pointer.render.basic_struct.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_pointer.render.basic_struct.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_pointer.render.nested_structs_arrays.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_pointer.render.nested_structs_arrays.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_pointer.render.nested_structs_arrays.sampler2D_samplerCube_vertex
dEQP-GLES31.functional.program_uniform.by_pointer.render.struct_in_array.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_pointer.render.struct_in_array.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_pointer.render.struct_in_array.sampler2D_samplerCube_vertex
dEQP-GLES31.functional.program_uniform.by_value.render.array_in_struct.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_value.render.array_in_struct.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_value.render.array_in_struct.sampler2D_samplerCube_vertex
dEQP-GLES31.functional.program_uniform.by_value.render.basic.samplerCube_both
dEQP-GLES31.functional.program_uniform.by_value.render.basic.samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_value.render.basic.samplerCube_vertex
dEQP-GLES31.functional.program_uniform.by_value.render.basic_struct.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_value.render.basic_struct.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_value.render.nested_structs_arrays.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_value.render.nested_structs_arrays.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_value.render.nested_structs_arrays.sampler2D_samplerCube_vertex
dEQP-GLES31.functional.program_uniform.by_value.render.struct_in_array.sampler2D_samplerCube_both
dEQP-GLES31.functional.program_uniform.by_value.render.struct_in_array.sampler2D_samplerCube_fragment
dEQP-GLES31.functional.program_uniform.by_value.render.struct_in_array.sampler2D_samplerCube_vertex

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15246>

2 years agogallivm: fix debug prints for halfs
Mike Blumenkrantz [Fri, 4 Mar 2022 15:42:59 +0000 (10:42 -0500)]
gallivm: fix debug prints for halfs

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15246>

2 years agopan/bi: Don't assign slots for the blend second source
Icecream95 [Mon, 21 Feb 2022 03:54:50 +0000 (16:54 +1300)]
pan/bi: Don't assign slots for the blend second source

Another instruction might write to the second source, and then an
INSTR_INVALID_ENC fault will be raised because the tuple will write to
and read from the register at the same time.

Fixes: 795638767d1 ("pan/bi: Use fused dual source blending")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopan/bi: Skip psuedo sources in ISA.xml
Icecream95 [Mon, 21 Feb 2022 03:50:56 +0000 (16:50 +1300)]
pan/bi: Skip psuedo sources in ISA.xml

The second staging register source for the +BLEND instruction should
not be packed nor disassembled, so skip it when include_pseudo is not
set.

Fixes: 795638767d1 ("pan/bi: Use fused dual source blending")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopanfrost: Fix ubo_mask calculation
Icecream95 [Fri, 31 Dec 2021 13:15:21 +0000 (02:15 +1300)]
panfrost: Fix ubo_mask calculation

BITSET_MASK returns ~0 when given an input of zero, when we need it to
return 0 instead.

Fixes shaders with only sysvals but no UBOs when push constants are
disabled.

This breaks when 31 or 32 UBOs are used, but PAN_MAX_CONST_BUFFERS is
currently set to 16.

Fixes: c246af0dd80 ("panfrost: Only upload UBOs when needed")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopanfrost: Improve comment for emit_fragment_job
Icecream95 [Sun, 20 Feb 2022 09:45:02 +0000 (22:45 +1300)]
panfrost: Improve comment for emit_fragment_job

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopan/bi: Add documentation for bifrost_nir_lower_store_component
Icecream95 [Wed, 23 Feb 2022 10:16:35 +0000 (23:16 +1300)]
pan/bi: Add documentation for bifrost_nir_lower_store_component

Taken from the commit that introduced the function,
95458c40330 ("pan/bi: Lower stores with component != 0").

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopan/bi: Make disassembler build reproducibly
Icecream95 [Thu, 27 Jan 2022 04:46:54 +0000 (17:46 +1300)]
pan/bi: Make disassembler build reproducibly

Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopanfrost: Re-emit descriptors after resource shadowing
Icecream95 [Tue, 15 Feb 2022 07:34:57 +0000 (20:34 +1300)]
panfrost: Re-emit descriptors after resource shadowing

This could be made slightly more efficient by only setting the dirty
state that is needed, but eventually you reach a point where it's
cheaper to re-emit everything than work out what can or can't be kept.

Fixes rendering issues in Duckstation.

Fixes: cd2c1ef9da6 ("panfrost: Dirty track textures/samplers")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopanfrost: Set dirty state in set_shader_buffers
Icecream95 [Tue, 15 Feb 2022 07:24:01 +0000 (20:24 +1300)]
panfrost: Set dirty state in set_shader_buffers

Otherwise the pointer (which is uploaded as a sysval) won't be updated
when a new SSBO is bound.

Fixes: c34b760b9f9 ("panfrost: Dirty track constant buffers")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopan/bi: Check dependencies of both destinations of instructions
Icecream95 [Mon, 14 Feb 2022 03:26:59 +0000 (16:26 +1300)]
pan/bi: Check dependencies of both destinations of instructions

TEXC can have two destinations; the value for neither of them can be
used in the same bundle, so extend the code to check for this to
iterate over both destinations.

Fixes artefacts in the game "LIMBO".

Fixes: a303076c1ab ("pan/bi: Add bi_instr_schedulable predicate")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopan/bi: Add interference between destinations
Icecream95 [Mon, 14 Feb 2022 03:20:47 +0000 (16:20 +1300)]
pan/bi: Add interference between destinations

Trying to write to overlapping register ranges from a single
instruction is undefined behaviour, so add interference between the
nodes to avoid this.

Hit in a dual-texture instruction in LIMBO.

Fixes: 9146bafbb42 ("pan/bi: Add dual texture fusing pass")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopanfrost: Disable point size upper limit clamping
Icecream95 [Tue, 18 Jan 2022 03:02:12 +0000 (16:02 +1300)]
panfrost: Disable point size upper limit clamping

The hardware already clamps this, there is no need to do it in the
shader.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopanfrost: Update point size limits to match hardware behaviour
Icecream95 [Tue, 18 Jan 2022 02:48:17 +0000 (15:48 +1300)]
panfrost: Update point size limits to match hardware behaviour

Found while reverse-engineering the tiler heap format.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopanfrost: Set PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION
Icecream95 [Tue, 18 Jan 2022 02:16:49 +0000 (15:16 +1300)]
panfrost: Set PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION

Fixes arb-provoking-vertex-render Piglit test.

Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agopan/mdg: Use util_logbase2 instead of C99 log2
Icecream95 [Sat, 19 Jun 2021 03:01:15 +0000 (15:01 +1200)]
pan/mdg: Use util_logbase2 instead of C99 log2

log2 operates on double, we only need the integer util/ function.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15250>

2 years agoa4xx: add emission of compute state, and compute dispatch
Ilia Mirkin [Sun, 14 Nov 2021 18:01:57 +0000 (13:01 -0500)]
a4xx: add emission of compute state, and compute dispatch

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14794>

2 years agoa4xx: add logic to emit image/ssbo state
Ilia Mirkin [Sun, 14 Nov 2021 17:59:40 +0000 (12:59 -0500)]
a4xx: add logic to emit image/ssbo state

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14794>

2 years agofreedreno/ir3: support a4xx compute differences
Ilia Mirkin [Sun, 14 Nov 2021 18:05:07 +0000 (13:05 -0500)]
freedreno/ir3: support a4xx compute differences

Mainly the workgroup id comes injected via consts by the hardware (or
CP), and we must make room for it, otherwise the driver won't know where
to put it.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14794>

2 years agofreedreno/ir3: support a4xx in load/store buffer/image emission
Ilia Mirkin [Sun, 14 Nov 2021 18:04:26 +0000 (13:04 -0500)]
freedreno/ir3: support a4xx in load/store buffer/image emission

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14794>

2 years agofreedreno/perfetto+fdperf: Set SYSPROF param
Rob Clark [Thu, 3 Mar 2022 23:38:22 +0000 (15:38 -0800)]
freedreno/perfetto+fdperf: Set SYSPROF param

No need to check error return and deal with older kernels.  Older
kernels won't have this param but their default behavior allows for
systemwide perfcntr collection.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15236>

2 years agofreedreno/drm: Add SYSPROF param
Rob Clark [Thu, 3 Mar 2022 23:34:36 +0000 (15:34 -0800)]
freedreno/drm: Add SYSPROF param

Add new param for putting kernel in system-profiling mode and add
corresponding fd_pipe_set_param() mechanism.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15236>

2 years agofreedreno: Update uapi header
Rob Clark [Thu, 3 Mar 2022 23:17:13 +0000 (15:17 -0800)]
freedreno: Update uapi header

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15236>

2 years agoegl+libsync: Add helper to complain about invalid fence fd's
Rob Clark [Fri, 25 Feb 2022 17:08:34 +0000 (09:08 -0800)]
egl+libsync: Add helper to complain about invalid fence fd's

Debugging fd lifetime issues can be hard.  Add a helper for debug builds
to print out an error if an fd is not a fence fd, and sprinkle it around

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15094>

2 years agoandroid: Push in-fence-fd down to driver
Rob Clark [Sat, 19 Feb 2022 16:38:26 +0000 (08:38 -0800)]
android: Push in-fence-fd down to driver

Rather than immediately stall on the CPU in SwapBuffers() if the
in-fence for the dequeued buffer is not yet signaled, push it down
to the driver.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6048
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15094>

2 years agogallium/dri: Extend image extension to support in-fence
Rob Clark [Sat, 19 Feb 2022 16:36:43 +0000 (08:36 -0800)]
gallium/dri: Extend image extension to support in-fence

Extend dri so that an in-fence-fd can be plumbed through to driver.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15094>

2 years agoradv/ci: update list of expected failures
Samuel Pitoiset [Thu, 3 Mar 2022 19:26:54 +0000 (20:26 +0100)]
radv/ci: update list of expected failures

Add dEQP-VK.glsl.builtin.precision_double.determinant.compute.mat3
which fails on all generations.

It looks like CTS should relax tolerance slightly.

Co-authored-by: Charlie Turner <cturner@igalia.com>
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Martin Roukala <martin.roukala@mupuf.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15234>

2 years agoradv/ci: skip dEQP-VK.renderpass2.depth_stencil_resolve.*_samplemask
Samuel Pitoiset [Fri, 4 Mar 2022 15:20:25 +0000 (16:20 +0100)]
radv/ci: skip dEQP-VK.renderpass2.depth_stencil_resolve.*_samplemask

They randomly hang on Navi10 and randomly fail on Sienna Cichlid.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Martin Roukala <martin.roukala@mupuf.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15234>

2 years agov3d: rebind sampler view if resource changed the BO
Juan A. Suarez Romero [Thu, 3 Mar 2022 11:09:32 +0000 (12:09 +0100)]
v3d: rebind sampler view if resource changed the BO

When discarding the whole resource to create a new one, if this resource
is used by a sampler view, a rebind must be done to use the new
resource.

But this must be done when setting the sampler views, because we don't
have access to those samplers before.

v2:
 - Pack shader state on setting sampler views (Iago)
 - Use a serial ID to know when to rebind sampler views (Juan)

v3:
 - Move check to caller (Iago)
 - Keep rebind sampler view on BO change (Iago)

v4:
 - Rename "serial_bo" to "serial_id" (Iago)
 - Add comments (Iago)

Fixes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6027
Signed-off-by: Juan A. Suarez Romero <jasuarez@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15171>

2 years agopanfrost: Push twice as many uniforms
Alyssa Rosenzweig [Thu, 3 Mar 2022 22:39:02 +0000 (17:39 -0500)]
panfrost: Push twice as many uniforms

The limit for Bifrost is twice as high as previously thought -- the limit is 64
*slots* of FAU, not 64 words. Each slot is 2 words. We can push twice as much,
saving a considerable number of cycles in some cases.

total instructions in shared programs: 2454260 -> 2431502 (-0.93%)
instructions in affected programs: 845176 -> 822418 (-2.69%)
helped: 3376
HURT: 304
helped stats (abs) min: 1.0 max: 60.0 x̄: 7.92 x̃: 6
helped stats (rel) min: 0.13% max: 45.45% x̄: 4.60% x̃: 4.11%
HURT stats (abs)   min: 1.0 max: 60.0 x̄: 13.06 x̃: 8
HURT stats (rel)   min: 0.16% max: 35.09% x̄: 7.58% x̃: 6.52%
95% mean confidence interval for instructions value: -6.50 -5.87
95% mean confidence interval for instructions %-change: -3.75% -3.43%
Instructions are helped.

total tuples in shared programs: 1963383 -> 1951560 (-0.60%)
tuples in affected programs: 638622 -> 626799 (-1.85%)
helped: 2959
HURT: 573
helped stats (abs) min: 1.0 max: 54.0 x̄: 5.61 x̃: 4
helped stats (rel) min: 0.15% max: 28.57% x̄: 3.61% x̃: 3.12%
HURT stats (abs)   min: 1.0 max: 50.0 x̄: 8.35 x̃: 6
HURT stats (rel)   min: 0.25% max: 27.34% x̄: 6.24% x̃: 4.92%
95% mean confidence interval for tuples value: -3.61 -3.08
95% mean confidence interval for tuples %-change: -2.18% -1.85%
Tuples are helped.

total clauses in shared programs: 387817 -> 365111 (-5.85%)
clauses in affected programs: 135527 -> 112821 (-16.75%)
helped: 3489
HURT: 25
helped stats (abs) min: 1.0 max: 43.0 x̄: 6.52 x̃: 5
helped stats (rel) min: 0.82% max: 58.33% x̄: 17.48% x̃: 15.87%
HURT stats (abs)   min: 1.0 max: 3.0 x̄: 1.56 x̃: 1
HURT stats (rel)   min: 2.94% max: 11.11% x̄: 6.87% x̃: 6.67%
95% mean confidence interval for clauses value: -6.67 -6.26
95% mean confidence interval for clauses %-change: -17.65% -16.96%
Clauses are helped.

total cycles in shared programs: 201842.21 -> 168754.04 (-16.39%)
cycles in affected programs: 84035.50 -> 50947.33 (-39.37%)
helped: 3547
HURT: 136
helped stats (abs) min: 0.041665999999999315 max: 54.0 x̄: 9.33 x̃: 8
helped stats (rel) min: 0.17% max: 80.77% x̄: 36.10% x̃: 36.84%
HURT stats (abs)   min: 0.041665999999999315 max: 1.0 x̄: 0.12 x̃: 0
HURT stats (rel)   min: 0.18% max: 12.24% x̄: 1.18% x̃: 0.61%
95% mean confidence interval for cycles value: -9.26 -8.71
95% mean confidence interval for cycles %-change: -35.34% -34.11%
Cycles are helped.

total arith in shared programs: 74918.46 -> 75022.62 (0.14%)
arith in affected programs: 22471.04 -> 22575.21 (0.46%)
helped: 1571
HURT: 1492
helped stats (abs) min: 0.041665999999999315 max: 1.125 x̄: 0.17 x̃: 0
helped stats (rel) min: 0.17% max: 40.00% x̄: 2.50% x̃: 1.96%
HURT stats (abs)   min: 0.041665999999999315 max: 2.375 x̄: 0.25 x̃: 0
HURT stats (rel)   min: 0.16% max: 100.00% x̄: 5.35% x̃: 2.37%
95% mean confidence interval for arith value: 0.02 0.05
95% mean confidence interval for arith %-change: 1.08% 1.56%
Arith are HURT.

total ldst in shared programs: 174812 -> 137889 (-21.12%)
ldst in affected programs: 81319 -> 44396 (-45.41%)
helped: 3722
HURT: 0
helped stats (abs) min: 1.0 max: 62.0 x̄: 9.92 x̃: 8
helped stats (rel) min: 1.82% max: 100.00% x̄: 47.18% x̃: 43.75%
95% mean confidence interval for ldst value: -10.20 -9.64
95% mean confidence interval for ldst %-change: -47.97% -46.39%
Ldst are helped.

total quadwords in shared programs: 1757124 -> 1714130 (-2.45%)
quadwords in affected programs: 584065 -> 541071 (-7.36%)
helped: 3474
HURT: 173
helped stats (abs) min: 1.0 max: 90.0 x̄: 12.66 x̃: 9
helped stats (rel) min: 0.26% max: 34.18% x̄: 8.78% x̃: 8.33%
HURT stats (abs)   min: 1.0 max: 26.0 x̄: 5.76 x̃: 4
HURT stats (rel)   min: 0.45% max: 20.66% x̄: 4.48% x̃: 2.63%
95% mean confidence interval for quadwords value: -12.21 -11.37
95% mean confidence interval for quadwords %-change: -8.36% -7.95%
Quadwords are helped.

total threads in shared programs: 52898 -> 53142 (0.46%)
threads in affected programs: 262 -> 506 (93.13%)
helped: 250
HURT: 6
helped stats (abs) min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
helped stats (rel) min: 100.00% max: 100.00% x̄: 100.00% x̃: 100.00%
HURT stats (abs)   min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 50.00% max: 50.00% x̄: 50.00% x̃: 50.00%
95% mean confidence interval for threads value: 0.92 0.99
95% mean confidence interval for threads %-change: 93.69% 99.28%
Threads are helped.

total spills in shared programs: 161 -> 107 (-33.54%)
spills in affected programs: 54 -> 0
helped: 27
HURT: 0

total fills in shared programs: 1386 -> 796 (-42.57%)
fills in affected programs: 590 -> 0
helped: 27
HURT: 0

Fixes: d4dccea0ba3 ("panfrost: Add UBO push data structure")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15239>

2 years agopan/bi: Run CSE after lowering FAU
Alyssa Rosenzweig [Fri, 4 Mar 2022 02:04:28 +0000 (21:04 -0500)]
pan/bi: Run CSE after lowering FAU

Lowering FAU can add moves from uniforms. If a uniform is moved out to a
register mulitple times in a basic block, these moves can be CSE'd, saving
instructions at the cost of register pressure.

854 shaders in my shader-db are helped on cycle count (average 2.94% reduction
in cycles). Only 9 shaders have hurt thread count, and there is no change in
spills or fills. Overall, this seems to be a win.

Prevents instruction count regressions from the next commit.

total instructions in shared programs: 2454423 -> 2444690 (-0.40%)
instructions in affected programs: 386274 -> 376541 (-2.52%)
helped: 2105
HURT: 0
helped stats (abs) min: 1.0 max: 116.0 x̄: 4.62 x̃: 2
helped stats (rel) min: 0.04% max: 27.27% x̄: 3.64% x̃: 1.92%
95% mean confidence interval for instructions value: -4.91 -4.33
95% mean confidence interval for instructions %-change: -3.83% -3.45%
Instructions are helped.

total tuples in shared programs: 1963534 -> 1957106 (-0.33%)
tuples in affected programs: 233562 -> 227134 (-2.75%)
helped: 1491
HURT: 117
helped stats (abs) min: 1.0 max: 63.0 x̄: 4.44 x̃: 2
helped stats (rel) min: 0.04% max: 24.53% x̄: 4.39% x̃: 2.59%
HURT stats (abs)   min: 1.0 max: 5.0 x̄: 1.61 x̃: 1
HURT stats (rel)   min: 0.18% max: 8.33% x̄: 1.44% x̃: 1.05%
95% mean confidence interval for tuples value: -4.28 -3.71
95% mean confidence interval for tuples %-change: -4.20% -3.73%
Tuples are helped.

total clauses in shared programs: 387848 -> 387079 (-0.20%)
clauses in affected programs: 13718 -> 12949 (-5.61%)
helped: 583
HURT: 60
helped stats (abs) min: 1.0 max: 16.0 x̄: 1.42 x̃: 1
helped stats (rel) min: 1.11% max: 25.00% x̄: 8.28% x̃: 6.67%
HURT stats (abs)   min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 0.86% max: 20.00% x̄: 4.58% x̃: 4.00%
95% mean confidence interval for clauses value: -1.29 -1.10
95% mean confidence interval for clauses %-change: -7.57% -6.58%
Clauses are helped.

total cycles in shared programs: 201866.21 -> 201682.92 (-0.09%)
cycles in affected programs: 6241.79 -> 6058.50 (-2.94%)
helped: 952
HURT: 98
helped stats (abs) min: 0.04166399999999726 max: 2.625 x̄: 0.20 x̃: 0
helped stats (rel) min: 0.12% max: 26.00% x̄: 4.05% x̃: 2.38%
HURT stats (abs)   min: 0.041665999999999315 max: 0.16666700000000034 x̄: 0.07 x̃: 0
HURT stats (rel)   min: 0.18% max: 8.70% x̄: 1.60% x̃: 1.43%
95% mean confidence interval for cycles value: -0.19 -0.16
95% mean confidence interval for cycles %-change: -3.80% -3.24%
Cycles are helped.

total arith in shared programs: 74924.00 -> 74660.12 (-0.35%)
arith in affected programs: 9303.67 -> 9039.79 (-2.84%)
helped: 1513
HURT: 118
helped stats (abs) min: 0.04166399999999726 max: 2.625 x̄: 0.18 x̃: 0
helped stats (rel) min: 0.07% max: 33.33% x̄: 4.68% x̃: 2.67%
HURT stats (abs)   min: 0.041665999999999315 max: 0.16666800000000137 x̄: 0.07 x̃: 0
HURT stats (rel)   min: 0.18% max: 8.70% x̄: 1.55% x̃: 1.37%
95% mean confidence interval for arith value: -0.17 -0.15
95% mean confidence interval for arith %-change: -4.48% -3.98%
Arith are helped.

total quadwords in shared programs: 1757254 -> 1751978 (-0.30%)
quadwords in affected programs: 197399 -> 192123 (-2.67%)
helped: 1464
HURT: 110
helped stats (abs) min: 1.0 max: 51.0 x̄: 3.73 x̃: 2
helped stats (rel) min: 0.04% max: 21.95% x̄: 4.16% x̃: 2.52%
HURT stats (abs)   min: 1.0 max: 7.0 x̄: 1.71 x̃: 1
HURT stats (rel)   min: 0.21% max: 13.04% x̄: 1.65% x̃: 0.93%
95% mean confidence interval for quadwords value: -3.58 -3.13
95% mean confidence interval for quadwords %-change: -3.97% -3.53%
Quadwords are helped.

total threads in shared programs: 52899 -> 52890 (-0.02%)
threads in affected programs: 18 -> 9 (-50.00%)
helped: 0
HURT: 9
HURT stats (abs)   min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
HURT stats (rel)   min: 50.00% max: 50.00% x̄: 50.00% x̃: 50.00%
95% mean confidence interval for threads value: -1.00 -1.00
95% mean confidence interval for threads %-change: -50.00% -50.00%
Threads are HURT.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15239>

2 years agofrontends/va: ignore incoming frame_num from VA picture parameters
Henry Goffin [Wed, 29 Dec 2021 09:20:30 +0000 (09:20 +0000)]
frontends/va: ignore incoming frame_num from VA picture parameters

The Gallium pipe video "frame_num" variable is internally used as a
counter of elapsed reference frames since the last IDR. The incoming
frame_num field from VA picture parameters is not equivalent; the VA
value may wrap to zero prematurely, as it is a 16-bit struct field with
a documented max value of 2^(log2_max_frame_num_minus4 + 4)-1.

This change improves "infinite GOP" single-client live streaming, where
it is reasonable for the server to desire an endless series of P-frames
without IDR. Without this change, it is difficult/impossible for an
application to encode a P- or B-frame after the VA frame_num field wraps
around to zero, depending on the backend encoder implementation.

This change has no effect on existing applications that always signal an
IDR frame and reset the VA frame_num to zero before it wraps around. For
example, the FFmpeg vaapi encoder ignores the VA documentation and sends
an un-wrapped VA frame_num, which results in identical computation of
the internal frame_num (as long as each GOP is less than 65536 frames).

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/5768

Reviewed-by: Thong Thai <thong.thai@amd.com>
patch revision 3: correctly avoid incrementing frame_num when the encoded
frame is not a reference, per h264 spec and ffmpeg behavior

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14332>

2 years agoaco: rework removal of jumps over branches
Rhys Perry [Wed, 2 Mar 2022 13:32:42 +0000 (13:32 +0000)]
aco: rework removal of jumps over branches

Only allow this in situations where we know it's safe. In particular, this
stops removal of unconditional branches like with
block_kind_continue_or_break.

Fixes dEQP-VK.graphicsfuzz.fragcoord-control-flow hang.

fossil-db (Sienna Cichlid):
Totals from 34 (0.02% of 162293) affected shaders:
Instrs: 84115 -> 84178 (+0.07%); split: -0.00%, +0.08%
CodeSize: 463372 -> 463624 (+0.05%); split: -0.00%, +0.06%
Latency: 3467316 -> 3467652 (+0.01%)
InvThroughput: 3085493 -> 3085578 (+0.00%)
Branches: 3221 -> 3284 (+1.96%); split: -0.03%, +1.99%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: f030b75b7d2 ("aco: relax condition to remove branches in case of few instructions")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15214>

2 years agoac/nir: implement nir_op_pack_{uint,sint}_2x16
Samuel Pitoiset [Thu, 3 Mar 2022 12:08:17 +0000 (13:08 +0100)]
ac/nir: implement nir_op_pack_{uint,sint}_2x16

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15231>

2 years agoaco: implement nir_op_pack_{uint,sint}_2x16
Samuel Pitoiset [Thu, 3 Mar 2022 07:53:06 +0000 (08:53 +0100)]
aco: implement nir_op_pack_{uint,sint}_2x16

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15231>

2 years agonir: introduce nir_pack_{sint,uint}_2x16 instructions
Samuel Pitoiset [Thu, 3 Mar 2022 07:37:29 +0000 (08:37 +0100)]
nir: introduce nir_pack_{sint,uint}_2x16 instructions

These instructions have AMD hardware equivalent and they will be used
to lower fragment shader outputs in NIR.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15231>

2 years agoiris: Mark a dirty update when vs_needs_sgvs_element value changed
Xiaohui Gu [Mon, 25 Oct 2021 05:58:03 +0000 (13:58 +0800)]
iris: Mark a dirty update when vs_needs_sgvs_element value changed

Add vs_needs_sgvs_element value check when updating vertex
element dirty state in iris_update_compiled_vs to solve
render error of Android game "Genshin Impact".

Signed-off-by: Xiaohui Gu <xiaohui.gu@intel.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15142>

2 years agovenus: add VK_EXT_image_robustness support
Yiwei Zhang [Tue, 1 Mar 2022 19:11:29 +0000 (19:11 +0000)]
venus: add VK_EXT_image_robustness support

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15205>

2 years agovenus: add VK_EXT_provoking_vertex support
Yiwei Zhang [Tue, 1 Mar 2022 18:50:40 +0000 (18:50 +0000)]
venus: add VK_EXT_provoking_vertex support

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15205>

2 years agovenus: add VK_EXT_line_rasterization support
Yiwei Zhang [Tue, 1 Mar 2022 18:40:31 +0000 (18:40 +0000)]
venus: add VK_EXT_line_rasterization support

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15205>

2 years agovenus: update to latest venus protocol
Yiwei Zhang [Tue, 1 Mar 2022 18:30:36 +0000 (18:30 +0000)]
venus: update to latest venus protocol

Added the below extension support:
- VK_EXT_line_rasterization
- VK_EXT_provoking_vertex

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15205>

2 years agovenus: group extensions promoted to 1.3
Yiwei Zhang [Thu, 3 Mar 2022 22:28:03 +0000 (22:28 +0000)]
venus: group extensions promoted to 1.3

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15205>

2 years agovenus: clean up physical device features and properties
Yiwei Zhang [Thu, 3 Mar 2022 22:23:23 +0000 (22:23 +0000)]
venus: clean up physical device features and properties

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15205>

2 years agonir/opt_shrink_vectors: update docstring
Daniel Schürmann [Thu, 19 Aug 2021 08:46:09 +0000 (10:46 +0200)]
nir/opt_shrink_vectors: update docstring

in order to reflect the various recent improvements.

Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12468>

2 years agonir/opt_shrink_vectors: remove duplicate components from vecN
Daniel Schürmann [Tue, 17 Aug 2021 12:01:36 +0000 (14:01 +0200)]
nir/opt_shrink_vectors: remove duplicate components from vecN

vecN instructions which are only used by other ALU
will now get duplicate channels removed.

i915g:
total instructions in shared programs: 396309 -> 396294 (<.01%)
instructions in affected programs: 186 -> 171 (-8.06%)

r300:
total instructions in shared programs: 1165059 -> 1164354 (-0.06%)
instructions in affected programs: 35884 -> 35179 (-1.96%)
total temps in shared programs: 165497 -> 165326 (-0.10%)
temps in affected programs: 2990 -> 2819 (-5.72%)

softpipe:
total instructions in shared programs: 2860028 -> 2859084 (-0.03%)
instructions in affected programs: 55539 -> 54595 (-1.70%)
total temps in shared programs: 516939 -> 516546 (-0.08%)
temps in affected programs: 6623 -> 6230 (-5.93%)

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12468>

2 years agonir/opt_shrink_vectors: shrink load_const properly
Daniel Schürmann [Tue, 17 Aug 2021 11:23:22 +0000 (13:23 +0200)]
nir/opt_shrink_vectors: shrink load_const properly

This patch enables removal of arbitrary channels in
load_const instructions, if they are either unused or
duplicates of other channels and only used by ALU.

Totals from 692 (0.51% of 134913) affected shaders: (GFX10.3)
VGPRs: 21832 -> 21544 (-1.32%)
CodeSize: 1322016 -> 1313080 (-0.68%); split: -0.68%, +0.01%
Instrs: 243635 -> 242231 (-0.58%); split: -0.58%, +0.00%
Latency: 1856138 -> 1857237 (+0.06%); split: -0.09%, +0.15%
InvThroughput: 424298 -> 421671 (-0.62%); split: -0.62%, +0.01%
VClause: 4580 -> 4583 (+0.07%); split: -0.02%, +0.09%
SClause: 14336 -> 14354 (+0.13%); split: -0.04%, +0.17%
Copies: 8897 -> 8859 (-0.43%); split: -0.45%, +0.02%
PreSGPRs: 20439 -> 20437 (-0.01%)
PreVGPRs: 16011 -> 15907 (-0.65%); split: -0.97%, +0.32%

i915g:
total instructions in shared programs: 396471 -> 396309 (-0.04%)
instructions in affected programs: 6408 -> 6246 (-2.53%)
total const in shared programs: 56458 -> 56422 (-0.06%)
const in affected programs: 407 -> 371 (-8.85%)
LOST:   shaders/closed/steam/trine-2/fp-3.shader_test FS

r300:
total instructions in shared programs: 1164421 -> 1165059 (0.05%)
instructions in affected programs: 143981 -> 144619 (0.44%)
total temps in shared programs: 165488 -> 165497 (<.01%)
temps in affected programs: 318 -> 327 (2.83%)
total consts in shared programs: 922140 -> 921952 (-0.02%)
consts in affected programs: 12438 -> 12250 (-1.51%)

softpipe:
total instructions in shared programs: 2859978 -> 2860028 (<.01%)
instructions in affected programs: 183355 -> 183405 (0.03%)
total temps in shared programs: 517071 -> 516939 (-0.03%)
temps in affected programs: 1416 -> 1284 (-9.32%)
total imm in shared programs: 103601 -> 102767 (-0.81%)
imm in affected programs: 3928 -> 3094 (-21.23%)

Acked-by: Emma Anholt <emma@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12468>

2 years agocrocus: change the line width workaround for gfx4/5
Dave Airlie [Thu, 3 Mar 2022 05:37:25 +0000 (15:37 +1000)]
crocus: change the line width workaround for gfx4/5

This fixes piglit line-flat-clip-color and the hud fps counter.

Fixes: 6b7a68b7c21e ("crocus: add missing line smooth bits.")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15229>

2 years agovenus: abort when stuck
Chia-I Wu [Mon, 28 Feb 2022 22:28:36 +0000 (14:28 -0800)]
venus: abort when stuck

This gives

  MESA-VIRTIO: debug: stuck in ring seqno wait with iter at 4096
  MESA-VIRTIO: debug: stuck in ring seqno wait with iter at 8192
  MESA-VIRTIO: debug: stuck in ring seqno wait with iter at 12288
  MESA-VIRTIO: debug: stuck in ring seqno wait with iter at 16384
  MESA-VIRTIO: debug: aborting
  Aborted

which should be more friendly than printing the messages forever.

On my i7-7820HQ, this aborts after roughly 4+8+16+32=60 seconds

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15200>

2 years agoaco/ra: don't immediately assign a register for p_branch
Daniel Schürmann [Wed, 2 Mar 2022 23:52:06 +0000 (00:52 +0100)]
aco/ra: don't immediately assign a register for p_branch

These get now assigned after handling phis.

Totals from 564 (0.42% of 134913) affected shaders: (GFX10.3)
CodeSize: 5519744 -> 5515308 (-0.08%)
Instrs: 1063045 -> 1061936 (-0.10%)
Latency: 11880452 -> 11875904 (-0.04%)
InvThroughput: 2259933 -> 2259581 (-0.02%); split: -0.02%, +0.00%
Copies: 86908 -> 85799 (-1.28%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13432>

2 years agoaco/tests: add test for branch definition RA
Rhys Perry [Tue, 27 Jul 2021 10:24:53 +0000 (11:24 +0100)]
aco/tests: add test for branch definition RA

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13432>

2 years agoaco: fix branch definition validation
Rhys Perry [Fri, 10 Sep 2021 17:20:03 +0000 (18:20 +0100)]
aco: fix branch definition validation

Like how they have to be register allocated differently, branch
definitions at merge block predecessors need to be validated differently.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13432>

2 years agoaco: add validate_instr_defs()
Rhys Perry [Wed, 20 Oct 2021 08:11:51 +0000 (09:11 +0100)]
aco: add validate_instr_defs()

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13432>

2 years agoaco/ra: fix register allocation of branch definitions
Rhys Perry [Wed, 28 Jul 2021 15:30:57 +0000 (16:30 +0100)]
aco/ra: fix register allocation of branch definitions

fossil-db (Sienna Cichlid):
Totals from 704 (0.52% of 134913) affected shaders:
CodeSize: 7177288 -> 7182072 (+0.07%); split: -0.00%, +0.07%
Instrs: 1371781 -> 1372977 (+0.09%); split: -0.00%, +0.09%
Latency: 17993572 -> 18001344 (+0.04%); split: -0.00%, +0.04%
InvThroughput: 4198996 -> 4199569 (+0.01%); split: -0.00%, +0.01%
Copies: 122456 -> 123516 (+0.87%); split: -0.01%, +0.88%
Branches: 43815 -> 43818 (+0.01%); split: -0.02%, +0.03%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13432>

2 years agoaco/ra: add get_reg_phi() helper
Rhys Perry [Wed, 28 Jul 2021 15:20:07 +0000 (16:20 +0100)]
aco/ra: add get_reg_phi() helper

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13432>

2 years agoaco: remove vcc hint from branch definitions
Rhys Perry [Wed, 28 Jul 2021 14:11:03 +0000 (15:11 +0100)]
aco: remove vcc hint from branch definitions

This doesn't seem to have much benefit anymore.

fossil-db (Sienna Cichlid):
Totals from 198 (0.15% of 134913) affected shaders:
CodeSize: 2610536 -> 2610872 (+0.01%); split: -0.01%, +0.02%
Instrs: 479001 -> 479085 (+0.02%); split: -0.01%, +0.03%
Latency: 7310684 -> 7300735 (-0.14%); split: -0.16%, +0.02%
InvThroughput: 2439084 -> 2437446 (-0.07%); split: -0.07%, +0.00%
SClause: 14760 -> 14722 (-0.26%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13432>

2 years agor300: schedule TEX instructions before OUT instructions
Pavel Ondračka [Thu, 3 Mar 2022 11:59:00 +0000 (12:59 +0100)]
r300: schedule TEX instructions before OUT instructions

NIR-to-TGSI produces partial output writes contrary to the old paths
that always wrote the full outputs. Therefore if there is now a partial
output write ready to be scheduled and nothing else besides a tex
is ready, we would schedule the output write first. This was not a
problem before as usually at last some component of the full output write
depended on the tex result.

This is not optimal from the performance point of view and resulted in
~20% slowdown in the Unigine demos. The docs say:

The first OUTPUT instruction will reserve space in the output register
fifo. This space is limited, therefore issuing an OUTPUT earlier than
necessary may cause threads to stall earlier than necessary. You
should not set an ALU instruction as type OUTPUT unless it is actually
writing to an output register, or it is the last instruction of
the program.

Fix it by explicitly prefering a TEX before OUT and restore the
performance: 9.66 -> 12.12 fps (as compared to 11.83 with the old
glsl-to-TGSI path) in Unigine Sanctuary. No change in Lightsmark or
GLmark.

This is also a win from the intructions point of view as we are usually
able to schedule the partial output writes in a single pair at the end.

total instructions in shared programs: 106009 -> 105891 (-0.11%)
instructions in affected programs: 10153 -> 10035 (-1.16%)
helped: 118
HURT: 0

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/5840
Signed-off-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15165>

2 years agor300: remove some dead logic in tex pair scheduling
Pavel Ondračka [Thu, 3 Mar 2022 11:24:37 +0000 (12:24 +0100)]
r300: remove some dead logic in tex pair scheduling

The max_score == -1 condition is already before so this
will never trigger. Its unclear what was the intention anyway. Now we
emit either:
- if we have accumulated enough tex intructions for a full block
- if we have nothing else to emit
- or if we can emit all remaining tex instructions already.

Signed-off-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15165>

2 years agoVenus: Add `vn_physical_device_{features, properties}` for better organization
Igor Torrente [Mon, 28 Feb 2022 11:59:41 +0000 (08:59 -0300)]
Venus: Add `vn_physical_device_{features, properties}` for better organization

New extensions properties/feature are being put in the `vn_physical_device`
which is not ideal from an organization point of view.

Here the `vn_physical_device_{features,properties}` are two new struct to
help the `vn_physical_device` organzation.

Signed-off-by: Igor Torrente <igor.torrente@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15170>

2 years agofreedreno/a4xx: fix integer tg4
Ilia Mirkin [Mon, 29 Nov 2021 06:44:32 +0000 (01:44 -0500)]
freedreno/a4xx: fix integer tg4

Something is slightly off in the integer values returned. It passes many
tests without the fixup, but the dEQP-GLES31 tests complain. The blob
ends up doing 3x gathers, and selects between them based on getinfo
results. Since we already have a per-sampler key with some spare bits,
just stick the bit-size info in there. And we can derive signedness from
the associated type info.

Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14670>