This enables the cache on vertex and fragment shaders only.
v2:
* Use MAYBE_UNUSED. (Matt)
[jordan.l.justen@intel.com: reword subject]
[jordan.l.justen@intel.com: *_cached_program => brw_disk_cache_*_program]
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This uses the Mesa disk_cache support to write out the final linked
binary for vertex and fragment shader programs.
This is based off the initial implementation done by Carl Worth. It
has been significantly reworked, first by Tim Arceri, and then by
Jordan Justen.
v2:
* Squash 'i965: add image param shader cache support'
* Squash 'i965: add shader cache support for pull param pointers'
* Sustantially simplified by a rework on top of Jason's 2975e4c56a.
* Rename load_program_data to read_program_data. (Jason)
v3:
* Simplify and align program read/write. (Jason)
v4:
* Don't save prog_data size since we know it from the stage. (Ken)
* Don't save program size, since prog_data includes the size. (Ken)
* Remove `assert` that potentially could be triggered by disk
corruption of the cache entries. (Ken)
* Fix compute shader scratch allocation. (Ken)
* Remove special case mapping for non-LLC. (Ken)
* Remove SET_UPLOAD_PARAMS macro
[jordan.l.justen@intel.com: *_cached_program => brw_disk_cache_*_program]
[jordan.l.justen@intel.com: brw_shader_cache.c => brw_disk_cache.c]
[jordan.l.justen@intel.com: don't map to write program when LLC is present]
[jordan.l.justen@intel.com: set program_written_to_cache on read from cache]
[jordan.l.justen@intel.com: only try cache when status is linking_skipped]
[jordan.l.justen@intel.com: all v2-v4 changes noted above]
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, thread_count was sent in from the stage after some stage
specific calculations. Those stage specific calculations were moved
into brw_alloc_stage_scratch, which will allow the shader cache to
also use the same calculations.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The caller can now use brw_stage_prog_data::program_size which is set
by the brw_compile_* functions.
Cc: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This will be used by the on disk shader cache.
v2:
* Set in brw_compile_* rather than brw_codegen_*. (Jason)
Signed-off-by: Timothy Arceri <timothy.arceri@collabora.com>
[jordan.l.justen@intel.com: Only add to brw_stage_prog_data]
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
When a program is restored from the shader cache, prog->nir will be
NULL, but prog->info will be restored.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
If the shader cache is enabled, after linking the program, we
serialize the program to nir. This will be saved out by the glsl
shader cache support.
Later, if the same program is found in the cache, we can use the nir
for a fallback in the unlikely case that the gen binary program is not
found in the cache.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
These fields can be used to optionally save off a driver blob with the
program metadata. For example, serialized nir, or tgsi.
v3:
* Rename serialized_nir* to driver_cache_blob*. (Tim)
* Free memory. (Jason)
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
v2 (Jason Ekstrand):
- Various whitespace cleanups
- Add helpers for reading/writing objects
- Rework derefs
- [de]serialize nir_shader::num_*
- Fix uses of blob_reserve_bytes
- Use a bitfield struct for packing tex_instr data
v3:
- Zero nir_variable struct on deserialization. (Jordan)
- Allow nir_serialize.h to be included in C++. (Jordan)
- Handle NULL info.name. (Jason)
- Set info.name to NULL when name is NULL. (Jordan)
Acked-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
if the driver sets the cap, then use the value it gives us.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Some hw (evergreen) has a limit on how many combined (images/buffers/mrts)
a fragment shader can access.
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Signed-off-by: Dave Airlie <airlied@redhat.com>
It is possible that the optimizer ends up in an infinite loop in
post_scheduler::schedule_alu(), because post_scheduler::prepare_alu_group()
does not find a proper scheduling. This can be deducted from
pending.count() being larger than zero and not getting smaller.
This patch works around this problem by signalling this failure so that the
optimizers bails out and the un-optimized shader is used.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103142
Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Gert Wollny <gw.fossdev@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
The shared si_create_shader_selector() code already offsets the mask.
Fixes the following piglit tests:
arb_cull_distance/clip-cull-3.shader_test
arb_cull_distance/clip-cull-4.shader_test
Fixes: 29d7bdd179 (radeonsi: scan NIR shaders to obtain required info)
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Previously the values were calculated by just shifting ~0 by the
invocation ID. This would end up including bits that are higher than
gl_SubGroupSizeARB. The corresponding CTS test effectively requires that
these high bits be zero so it was failing. There is a Piglit test as
well but this appears to checking the wrong values so it passes.
For the two greater-than bitmasks, this patch adds an extra mask with
(~0>>(64-gl_SubGroupSizeARB)) to force these bits to zero.
Fixes: KHR-GL45.shader_ballot_tests.ShaderBallotBitmasks
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102680#c3
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Neil Roberts <nroberts@igalia.com>
Only use CCS_E to render to a texture that is CCS_E-compatible with the
original texture's miptree (linear) format. This prevents render
operations from writing data that can't be decoded with the original
miptree format.
On Gen10, with the new CCS_E-enabled formats handled, this enables the
driver to pass the arb_texture_view-rendering-formats piglit test.
v2. Add a TODO for texturing. (Jason)
Cc: <mesa-stable@lists.freedesktop.org>
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
CannonLake additionally supports R11G11B10_FLOAT and four 10-10-10-2
formats with CCS_E. None of these formats fit within the current
blorp_copy framework so disable them until support is added.
Signed-off-by: Nanley Chery <nanley.g.chery@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
We might want to add a symbol check for the glvnd variant though.
Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
This is done in autotools, and is an oversight in the meson build.
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Tested-by: Eric Engestrom <eric.engestrom@imgtec.com>
It's supposed to be linked with pthread-stubs (if the platform needs
pthread-stubs). Pthread stubs support isn't (yet) implemented in the
meson build, so add a TODO.
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
This allows a user to not care whether they're setting a tristate or a
boolean option, which is a nice user facing feature, and something I've
personally run into.
Suggested-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
If we don't want to use these deps, there's no good reason to search
for them in the first place. This should shave a bit of time for the
initial build.
Signed-off-by: Erik Faye-Lund <kusmabite@gmail.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
The u_format_other.c users sqrtf, which on some systems require
a math-library. So let's make sure we link with it.
Signed-off-by: Erik Faye-Lund <kusmabite@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Fixes regression in:
dEQP-VK.api.object_management.alloc_callback_fail.graphics_pipeline
Fixes: 1e84e53712 "radv: add cache items to in memory cache when reading from disk"
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This patch modifies the ARB_indirect_parameters logic in
brw_draw_prims, so that our implementation isn't affected if
another application attempts to use predicates. Previously we
were using a predicate with a DELTAS_EQUAL comparison operation
and relying on the MI_PREDICATE_DATA register being 0. Our code
to initialize MI_PREDICATE_DATA to 0 was incorrect, so we were
accidentally using whatever value was written there. Because the
kernel does not initialize the MI_PREDICATE_DATA register on
hardware context creation, we might inherit the value from whatever
context was last running on the GPU (likely another process).
The Haswell command parser also does not currently allow us to write
the MI_PREDICATE_DATA register. Rather than fixing this and requiring
an updated kernel, we switch to a different approach which uses a
SRCS_EQUAL predicate that makes no assumptions about the states of any
of the predicate registers.
Fixes Piglit's spec/arb_indirect_parameters/tf-count-arrays test.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103085
Signed-off-by: Plamena Manolova <plamena.manolova@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Due to a gaffe on my part, we were re-emitting all binding table entries
on every single draw call. The push_constant_packets atom listens to
BRW_NEW_DRAW_CALL, but skips emitting 3DSTATE_CONSTANT_XS for each stage
unless stage_state->push_constants_dirty is true. However, it flagged
BRW_NEW_SURFACES unconditionally at the end, by mistake.
Instead, it should only flag it if we actually emit 3DSTATE_CONSTANT_XS
for a stage. We can move it a few lines up, inside the loop - the early
continues will skip over it if push constants aren't dirty for a stage.
With INTEL_NO_HW=1 set, improves performance of GFXBench5 gl_driver_2
on Apollolake at 1280x720 by 1.01122% +/- 0.470723% (n=35).
Reviewed-by: Rafael Antognolli <rafael.antognolli@intel.com>
Groups containing fields smaller than a DWord were not being decoded
correctly. For example:
<group count="32" start="32" size="4">
<field name="Vertex Element Enables" start="0" end="3" type="uint"/>
</group>
gen_field_iterator_next would properly walk over each element of the
array, incrementing group_iter, and calling iter_group_offset_bits()
to advance to the proper DWord. However, the code to print the actual
values only considered iter->field->start/end, which are 0 and 3 in the
above example. So it would always fetch bits 3:0 of the current DWord
when printing values, instead of advancing to each element of the array,
printing bits 0-3, 4-7, 8-11, and so on.
To fix this, we add new iter->start/end tracking, which properly
advances for each instance of a group's field.
Caught by Matt Turner while working on 3DSTATE_VF_COMPONENT_PACKING,
with a patch to convert it to use an array of bitfields (the example
above).
This also fixes the decoding of 3DSTATE_SBE's "Attribute Active
Component Format" fields.
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Fixes fbo-blending-formats on RGB8 and 565. We will still need to demote
blending to shader code in the MRT case to fix it in general, but that can
be added when we start doing 32F blending (which also needs to be done in
the shader).
