Some nir lowerings might need to know if txs is supported by
the backend.
Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8898>
On Intel platforms before Gen6, there is no min or max instruction.
Instead, a comparison instruction (*more on this below) and a SEL
instruction are used. Per other IEEE rules, the regular comparison
instruction, CMP, will always return false if either source is NaN. A
sequence like
cmp.l.f0.0(16) null<1>F g30<8,8,1>F g22<8,8,1>F
(+f0.0) sel(16) g8<1>F g30<8,8,1>F g22<8,8,1>F
will generate the wrong result for min if g22 is NaN. The CMP will
return false, and the SEL will pick g22.
To account for this, the hardware has a special comparison instruction
CMPN. This instruction behaves just like CMP, except if the second
source is NaN, it will return true. The intention is to use it for min
and max. This sequence will always generate the correct result:
cmpn.l.f0.0(16) null<1>F g30<8,8,1>F g22<8,8,1>F
(+f0.0) sel(16) g8<1>F g30<8,8,1>F g22<8,8,1>F
The problem is... for whatever reason, we don't emit CMPN. There was
even a comment in lower_minmax that calls out this very issue! The bug
is actually older than the "Fixes" below even implies. That's just when
the comment was added. That we know of, we never observed a failure
until #4254.
If src1 is known to be a number, either because it's not float or it's
an immediate number, use CMP. This allows cmod propagation to still do
its thing. Without this slight optimization, about 8,300 shaders from
shader-db are hurt on Iron Lake.
Fixes the following piglit tests (from piglit!475):
tests/spec/glsl-1.20/execution/fs-nan-builtin-max.shader_test
tests/spec/glsl-1.20/execution/fs-nan-builtin-min.shader_test
tests/spec/glsl-1.20/execution/vs-nan-builtin-max.shader_test
tests/spec/glsl-1.20/execution/vs-nan-builtin-min.shader_test
Closes: #4254
Fixes: 2f2c00c727 ("i965: Lower min/max after optimization on Gen4/5.")
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Iron Lake and GM45 had similar results. (Iron Lake shown)
total instructions in shared programs: 8115134 -> 8115135 (<.01%)
instructions in affected programs: 229 -> 230 (0.44%)
helped: 0
HURT: 1
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9027>
Since the CMPN builder was never used, there was no reason to make its
interface usable. :)
Fixes: 2f2c00c727 ("i965: Lower min/max after optimization on Gen4/5.")
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9027>
These checks were originally assertions elsewhere either in the existing
code or later in this MR.
Suggested-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9027>
On Gen7, we have to split shuffles into two MOVs for 64-bit types so we
can't handle source modifiers. On Gen12.5, we have to use integer types
all the time so we can't use them there either. Fixing that will be a
different commit but it interacts with this one.
Fixes: 90c9f29518 "i965/fs: Add support for nir_intrinsic_shuffle"
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9068>
Ensure that all resources are properly released by
properly parenting them to a memory context and releasing
the context during test teardown.
Signed-off-by: Rohan Garg <rohan.garg@collabora.com>
Reviewed-by: Tapani Pälli <tapani.palli@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8162>
This halves the number of calls to get_register_pressure_benefit
and decreases shader-db CPU time by ~1.5%.
Signed-off-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8741>
I meant to do this years ago when I first added SHADER_OPCODE_SEND. At
the time, the only use for the extended descriptor was bindless handles
which were always one thing and never non-constant. However, it doesn't
actually require any extra instructions because we have to OR in ex_mlen
anyway.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8748>
It is currently a bitset on top of a uint64_t but there are already
more than 64 values. Change to use BITSET to cover all the
SYSTEM_VALUE_MAX bits.
Cc: mesa-stable
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Karol Herbst <kherbst@redhat.com>
Acked-by: Jesse Natalie <jenatali@microsoft.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8585>
We were using nir_tex_instr::dest_type to a glsl_type, then passing it
to emit_texture(), only to just check the number of components. Just
pass the number of components directly. This lets us delete
brw_glsl_base_type_for_nir_type, which was asserting with
nir_texop_all_samples_equal because it didn't handle bool32.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7989>
This adds an emit_scan_step helper which gives us a place to do
something a bit more interesting than emitting a single op.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7329>
There are two problems this commit solves: First, is that the 64x64 MUL
lowering generates a Q MOV which, because of how late it runs in the
compile pipeline, it never gets removed. Second, it generates 32x32
MULs and we have to run it a second time to lower those.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7329>
We could probably support some strides if we tried hard enough but the
whole point of this opcode is to accelerate things with crazy Align16 or
crazy regions. It's ok if we have to emit an extra MOV to get a packed
source.
Fixes: 8b4a5e641b "intel/fs: Add support for subgroup quad operations"
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7329>
Previously, we were returning 2 whenever the source was a Q type. As
far as I can tell, the only reason why this hasn't blown up before is
that it was only ever used for VGRFs until the SWSB pass landed which
uses it for everything. This wasn't a problem because Q types generally
aren't a thing on TGL. However, they are for a small handful of
instructions.
Cc: mesa-stable@lists.freedesktop.org
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7329>
We currently don't use push constants with the COMPUTE_WALKER command.
Make all uniforms to be pull constants.
The local group id previously was a push constant, but is now
available in R0.2[7:0].
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8342>
If the instruction being coalesced would be vectorized but the target
doesn't support vectorizing that op, skip coalescing.
Reuse the callbacks from alu_to_scalar to describe which ops should not
be vectorized.
Signed-off-by: Erico Nunes <nunes.erico@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6506>
Instead of making it a fragment-specific thing based on uses_kill, track
whether or not we need one in fs_visitor and emit HALT_TARGET at the end
of emit_nir_code() if needed.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5071>
This means the pass has to walk all the instructions but it was doing
that in a bunch of cases anyway when it didn't have a HALT_TARGET.
However, removing HALT_TARGET frees up the scheduler a bit because
HALT_TARGET is considered a scheduling barrier. The shader-db results
are kind-of a wash but we're about to add HALT_TARGET unconditionally so
we want to be able to get rid of it.
Shader-db results on Ice Lake:
total instructions in shared programs: 19935623 -> 19935623 (0.00%)
instructions in affected programs: 0 -> 0
helped: 0
HURT: 0
total cycles in shared programs: 976758472 -> 976766135 (<.01%)
cycles in affected programs: 11097707 -> 11105370 (0.07%)
helped: 1750
HURT: 875
helped stats (abs) min: 1 max: 866 x̄: 26.39 x̃: 4
helped stats (rel) min: <.01% max: 39.24% x̄: 1.25% x̃: 0.46%
HURT stats (abs) min: 1 max: 1678 x̄: 61.54 x̃: 10
HURT stats (rel) min: <.01% max: 65.69% x̄: 1.86% x̃: 0.42%
95% mean confidence interval for cycles value: -2.48 8.32
95% mean confidence interval for cycles %-change: -0.40% -0.03%
Inconclusive result (value mean confidence interval includes 0).
LOST: 62
GAINED: 46
All of the lost/gained programs are SIMD32 fragment shaders.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5071>
We're about to start using it to implement nir_jump_halt which has
nothing inherently to do with fragment shaders or discards. May as well
name it for the HW instruction it generates.
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5071>
The code works but is a bit fragile if we ever add a case that has a
less strict requirement (a smaller gen) than the case above. To avoid
having to reason about this, refactor code to use a variable to
indicate whether the SFID is supported or not.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Marcin Ślusarz <marcin.slusarz@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7742>
They're not really "push" anymore but that's because there is no such
thing as push constants in bindless shaders on Intel. They should be
fast enough, though. There is some room for debate here as to whether
we want to do the pull in NIR or push it into the back-end. The
advantage of doing it in the back-end is that it'd be easier to use
MOV_INDIRECT for indirect push constant access rather than falling back
to a dataport message.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7356>
For triangle geometry, the hit attributes are always two floats which
contain the barycentric coordinates of the hit. For procedural
geometry, they're an arbitrary blob of data passed from the intersection
shader to the hit shaders. In our implementation, we stash that data
right after the HW RayQuery in the ray stack.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7356>
Unlike graphics and compute pipelines, Vulkan ray-tracing pipelines do
not have a single entrypoint. Instead, the raygen shader is specified
as a one-element shader binding table in the vkCmdTraceRay call. This
means that raygen shaders have to be bindless shaders just like any
other ray tracing shader. To launch them, we have a tiny compute shader
that acts as a trampoline and sets up the hotzone and uses btd_spawn to
fire off the raygen shader.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7356>
Most of the work for this is done for us by spirv_to_nir which gives us
a load_global from a memory address based on the shader_record_ptr
system values.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7356>
