The algebraic pass was exhibiting O(n^2) behavior in
dEQP-GLES2.functional.uniform_api.random.3 and
dEQP-GLES31.functional.ubo.random.all_per_block_buffers.13 (along with
other code-generated tests, and likely real-world loop-unroll cases).
In the process of using fmul(b2f(x), b2f(x)) -> b2f(iand(x, y)) to
transform:
result = b2f(a == b);
result *= b2f(c == d);
...
result *= b2f(z == w);
->
temp = (a == b)
temp = temp && (c == d)
...
temp = temp && (z == w)
result = b2f(temp);
nir_opt_algebraic, proceeding bottom-to-top, would match and convert
the top-most fmul(b2f(), b2f()) case each time, leaving the new b2f to
be matched by the next fmul down on the next time algebraic got run by
the optimization loop.
Back in 2016 in 7be8d07732 ("nir: Do opt_algebraic in reverse
order."), Matt changed algebraic to go bottom-to-top so that we would
match the biggest patterns first. This helped his cases, but I
believe introduced this failure mode. Instead of reverting that, now
that we've got the automaton, we can update the automaton's state
recursively and just re-process any instructions whose state has
changed (indicating that they might match new things). There's a
small chance that the state will hash to the same value and miss out
on this round of algebraic, but this seems to be good enough to fix
dEQP.
Effects with NIR_VALIDATE=0 (improvement is better with validation enabled):
Intel shader-db runtime -0.954712% +/- 0.333844% (n=44/46, obvious throttling
outliers removed)
dEQP-GLES2.functional.uniform_api.random.3 runtime
-65.3512% +/- 4.22369% (n=21, was 1.4s)
dEQP-GLES31.functional.ubo.random.all_per_block_buffers.13 runtime
-68.8066% +/- 6.49523% (was 4.8s)
v2: Use two worklists, suggested by @cwabbott, to cut out a bunch of
tricky code. Runtime of uniform_api.random.3 down -0.790299% +/-
0.244213% compred to v1.
v3: Re-add the nir_instr_remove() that I accidentally dropped in v2,
fixing infinite loops.
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
My motivation was to clarify the changes in the following commit, but
incidentally, it reduces runtime of
dEQP-GLES2.functional.uniform_api.random.3 (an algebraic-heavy
testcase) by -5.39524% +/- 2.21179% (n=15)
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
In order to have nir_opt_algebraic be able to do further algebraic
work on the output of a replacement, we need to maintain the
automaton's state.
Reviewed-by: Eric Anholt <eric@anholt.net>
If not all bits are cleared, then BLT needs to be given the current clear
value and not the new one.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Fixes:
Wolfenstein:Youngblood (w/o shader_ballot)
dEQP-VK.descriptor_indexing.combined_image_sampler_in_loop_with_lod
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
I don't think the bug applies for global/scratch instructions and
load_barycentric_at_sample selection expects this feature to work.
Fixes various dEQP-VK.pipeline.multisample_interpolation.* tests on GFX10.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-By: Timur Kristóf <timur.kristof@gmail.com>
Commit 0899bf55 made some deqp-gles3 tests related to RGB8 PBOs fail
on R600 because it exposed PIPE_FORMAT_R8G8B8_UNORM and R600 doesn't
propely handle this. Disabling this format also for buffers fixes the
issue.
In addition, disabling also the related RGB8 integer formats for buffers
fixes some deqp-gles3 tests:
dEQP-GLES3.functional.texture.specification.teximage2d_pbo.rgb8ui_cube
dEQP-GLES3.functional.texture.specification.texsubimage2d_pbo.rgb8i_2d
dEQP-GLES3.functional.texture.specification.texsubimage2d_pbo.rgb8i_cube
dEQP-GLES3.functional.texture.specification.texsubimage2d_pbo.rgb8ui_2d
dEQP-GLES3.functional.texture.specification.texsubimage2d_pbo.rgb8ui_cube
dEQP-GLES3.functional.texture.specification.teximage3d_pbo.rgb8i_2d_array
dEQP-GLES3.functional.texture.specification.teximage3d_pbo.rgb8i_3d
dEQP-GLES3.functional.texture.specification.teximage3d_pbo.rgb8ui_2d_array
dEQP-GLES3.functional.texture.specification.teximage3d_pbo.rgb8ui_3d
dEQP-GLES3.functional.texture.specification.texsubimage3d_pbo.rgb8i_2d_array
dEQP-GLES3.functional.texture.specification.texsubimage3d_pbo.rgb8i_3d
dEQP-GLES3.functional.texture.specification.texsubimage3d_pbo.rgb8ui_2d_array
dEQP-GLES3.functional.texture.specification.texsubimage3d_pbo.rgb8ui_3d
Fixes: 0899bf55
st/mesa: Map MESA_FORMAT_RGB_UNORM8 <-> PIPE_FORMAT_R8G8B8_UNORM
Closes#2118
Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
../src/amd/llvm/ac_llvm_build.c: In function ‘ac_build_canonicalize’:
../src/amd/llvm/ac_llvm_build.c:4567:9: warning: ‘intr’ may be used uninitialized in this function [-Wmaybe-uninitialized]
4567 | return ac_build_intrinsic(ctx, intr, type, params, 1,
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
4568 | AC_FUNC_ATTR_READNONE);
| ~~~~~~~~~~~~~~~~~~~~~~
../src/amd/llvm/ac_llvm_build.c:4567:9: warning: ‘type’ may be used uninitialized in this function [-Wmaybe-uninitialized]
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
In the case of promoted extensions we can end up with an entrypoint that
we support being an alias of an entrypoint we do not support. For
instance, if an extension gets promoted from EXT to KHR, the EXT entry-
points may be aliases of the KHR ones. We want to leave everything as
EXT until we get around to advertising the KHR so that we don't break
things when we update the XML and headers.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
The current code can only handle enum aliases if the original enum is
declared first followed by the alias as we walk the XML in a linear
fashion. This commit allows us to handle aliases where the alias
declaration comes before the thing it's aliasing.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
We may have replaced the backing storage for a texture buffer while it
was unbound, at which point iris_rebind_buffer would not have caught it
and updated it. We need to ensure that the current resource's address
matches the one our SURFACE_STATE points at. If not, update addresses
and re-upload the SURFACE_STATE.
Shader images and buffers do not suffer from this problem because we
re-stream the surface state on every set call, since there isn't a
created CSO object for those with a saved SURFACE_STATE. Constant
buffers are also currently re-streamed (we pitch the SURFACE_STATE
on every set_constant_buffer call). Surfaces would need this
treatment (as they're created CSOs) except that we never swap out
their backing storage today (we only do it for buffers), so it's OK
for now.
Fixes misrendering in Unreal 4 demos (Elemental, Matinee Fight Scene).
Huge thanks to Andrii Simiklit for tracking down the problem - it was
quite difficult to find! Also fixes Andrii's new Piglit test for the
bug, 'arb_texture_buffer_object-re-init'.
Closes: https://gitlab.freedesktop.org/mesa/mesa/issues/1365
When replacing the backing storage for texture buffers, image buffers,
and so on, we may need to update the "Surface Base Address" field in
any corresponding SURFACE_STATE. This is easier to accomplish if we
have a copy on the CPU - we can just compare the current field, update
it, and re-upload.
This patch adds a CPU-side copy to the new iris_surface_state wrapper
struct, and reworks allocation and upload to fill things out on the
CPU copy first, then upload that to the GPU when finished.
This will be necessary to fix iris_invalidate_resource bugs shortly.
Technically, we never replace the backing storage for pipe_surfaces
(render targets), so we don't need to make this change there. However,
it's nice to have surfaces, sampler views, and image views handled
similarly. Plus, if we ever wanted to swap out backing storage for
busy textures, we'd need this infrastructure.
v2: Properly free memory (caught by Andrii Simiklit)
Today, we only have a state reference to the GPU buffer containing our
uploaded SURFACE_STATEs. However, we're going to want a CPU-side copy
soon. Making a wrapper struct means we can talk about both together,
and also put both in the field called "surface_state".
We can just compare the VERTEX_BUFFER_STATE address field to the
current BO's address. When calling rebind, we've already updated
the resource to the new buffer, but the state will have the old
address.
Mutating fields of global resources is generally not safe, and the only
reason we were doing it was to avoid passing an extra parameter to
the fill_surface_state helper.
This takes a noticable amount of time in piglit and some tests don't
need it.
Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
This is similar to a scheduler I've written for vc4 and i965, but this
time written at the NIR level so that hopefully it's reusable. A notable
new feature it has is Goodman/Hsu's heuristic of "once we've started
processing the uses of a value, prioritize processing the rest of their
uses", which should help avoid the heuristic otherwise making such
systematically bad choices around getting texture results consumed.
Results for v3d:
total instructions in shared programs: 6497588 -> 6518242 (0.32%)
total threads in shared programs: 154000 -> 152828 (-0.76%)
total uniforms in shared programs: 2119629 -> 2068681 (-2.40%)
total spills in shared programs: 4984 -> 472 (-90.53%)
total fills in shared programs: 6418 -> 1546 (-75.91%)
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com> (v1)
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com> (v2)
v2: Use the DAG datastructure, fold in the scheduling-for-parallelism
patch, include SSA defs in live values so we can switch to bottom-up
if we want.
v3: Squash in improvements from Alejandro Piñeiro for getting V3D to
successfully register allocate on GLES3.1 dEQP. Make sure that
discards don't move after store_output. Comment spelling fix.
At the same time, update etna_clear_blit_pack_rgba to work with integer
formats.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
At the same time, this change allows using BLT for 8bpp formats
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Use the extended format if an such a format was passed.
v1 -> v2:
- set FORMAT_MASK bit when using ext PE format as suggested
by Wladimir J. van der Laan
Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>
Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Wladimir J. van der Laan <laanwj@gmail.com>
Reviewed-by: Jonathan Marek <jonathan@marek.ca>
No functional changes as the driver still uses the depth+stencil
pipeline.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
This refactoring will help for creating more decompress pipelines.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
No functional changes, but it will be used to decompress
separate depth/stencil aspects.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
No functional changes because the aspect mask is still not used
during image transitions but it will be needed for the separate
depth/stencil aspects logic.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
v7: run nir_opt_algebraic
v9: rework the callback function
v9: update alignment on all loads/stores, even if they're not vectorized
v10: add tests for 64-bit offsets
v10: add tests for signed offsets
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v9)
This pass combines intersecting, adjacent and identical loads/stores into
potentially larger ones and will be used by ACO to greatly reduce the
number of memory operations.
v2: handle nir_deref_type_ptr_as_array
v3: assume explicitly laid out types for derefs
v4: create less deref casts
v4: fix shared boolean vectorization
v4: fix copy+paste error in resources_different
v4: fix extract_subvector() to pass
nir_load_store_vectorize_test.ssbo_load_intersecting_32_32_64
v4: rebase
v5: subtract from deref/offset instead of scheduling offset calculations
v5: various non-functional changes/cleanups
v5: require less metadata and preserve more
v5: rebase
v6: cleanup and improve dependency handling
v6: emit less deref casts
v6: pass undef to components not set in the write_mask for new stores
v7: fix 8-bit extract_vector() with 64-bit input
v7: cleanup creation of store write data
v7: update align correctly for when the bit size of load/store increases
v7: rename extract_vector to extract_component and update comment
v8: prevent combining of row-major matrix column acceses
v9: rework process_block() to be able to vectorize more
v9: rework the callback function
v9: update alignment on all loads/stores, even if they're not vectorized
v9: remove entry::store_value, since it will not be updated if it's was
from a vectorized load
v9: fix bug in subtract_deref(), causing artifacts in Dishonored 2
v9: handle nir_intrinsic_scoped_memory_barrier
v10: use nir_ssa_scalar
v10: handle non-32-bit offsets
v10: use signed offsets for comparison
v10: improve create_entry_key_from_offset()
v10: support load_shared/store_shared
v10: remove strip_deref_casts()
v10: don't ever pass NULL to memcmp
v10: remove recursion in gcd()
v10: fix outdated comment
v11: use the new nir_extract_bits()
v12: remove use of nir_src_as_const_value in resources_different
v13: make entry key hash function deterministic
v13: simplify mask_sign_extend()
v14: add comment in hash_entry_key() about hashing pointers
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com> (v9)
Otherwise, nir_intrinsic_align() will assert when called on the intrinsics
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
These will be useful in the upcoming load/store vectorizer.
v11: rebase
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
It shouldn't matter, but the 1 was leftover from when it was handled
together with workgroup_size and num_work_groups.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
The former was always true and hence dead code. We will want to
explicitly declare the ring offset register with ACO, but we also want
to declare the scratch offset too, and we can't try to disable it since
ACO also supports spilling and the determination of whether spilling has
to happen occurs well after setting up registers. So replace
supports_spill with something that will actually be used for ACO.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Due to how LLVM works we have to make some of the FS inputs become
vectors, and therefore have to split them early so that they don't take
up extra register pressure due to how RA currently works.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
ac_shader_args will be similar to ac_shader_abi, except for being free
from LLVM-specific concepts and therefore capable of being shared
between LLVM and ACO. This will help us accomplish a few different
things:
- Decouple setting up SGPR and VGPR arguments from translating to LLVM,
so that we can reference these arguments in NIR lowering passes, which
will let us lower e.g. descriptor sets in NIR.
- Stop using radv-specific structures for things like determining the
chip generation in ACO.
In the end, we should replace ac_shader_abi with this structure +
driver-specific lowering passes.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>