The existing code was checking the whole interface variable rather
than its members, which is not what we want: we want to check
aliasing for each member in the interface variable.
Surprisingly, there are piglit tests that verify this and were
passing due to a bug in the existing code: when we were computing
the last component used by an interface variable we would use
the 'vector' path and multiply by vector_elements, which is 0 for
interface variables. This made the loop that checks for aliasing
be a no-op and not add the interface variable to the list of outputs
so then we would fail to link when we did not see a matching output
for the same input in the next stage. Since the tests expect a
linker error to happen, they would pass, but not for the right
reason.
Unfortunately, the current implementation uses ir_variable instances
to keep track of explicit locations. Since we don't have
ir_variables instances for individual interface members, we need
to have a custom struct with the data we need. This struct has
the ir_variable (which for interface members is the whole
interface variable), plus the data that we need to validate for
each aliased location, for now only the base type, which for
interface members we will take from the appropriate field inside
the interface variable.
Later patches will expand this custom struct so we can also check
other requirements for location aliasing, specifically that
we have matching interpolation and auxiliary storage, that once
again, we will take from the appropriate field members for the
interface variables.
v2:
- Use MAX_VARYING instead of MAX_VARYINGS_INCL_PATCH (Illia)
Fixes:
KHR-GL45.enhanced_layouts.varying_block_automatic_member_locations
Fixes (these were passing before but for incorrect reasons):
tests/spec/arb_enhanced_layouts/linker/block-member-locations/named-block-member-location-overlap.shader_test
tests/spec/arb_enhanced_layouts/linker/block-member-locations/named-block-member-mixed-order-overlap.shader_test
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Move the checks for explicit locations to a separate function. We
will use this in a follow-up patch to validate locations for interface
variables where we need to validate each interface member rather than
the interface variable itself.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
We were assuming that if an input has an invalid explicit location it would
fail to link because it would not find the corresponding output, however,
since we look for the matching output by indexing the explicit_locations
array with the input location, we still need to ensure that we don't index
out of bounds.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
This commit fixes two issues: First, we were returning false regardless
of whether or not the function made progress. Second, we were calling
nir_metadata_preserve far more often than needed; we only need to call
it once per impl.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
This helps valgrind when encode_type_to_blob is used.
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Not sure if this is the best place to put it, but we're going to need
this for NIR too.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Also needed in freedreno/ir3.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Dylan Baker <dylan@pnwbakers.com>
There are two issues with the current implementation. First, it relies
on the layout(local_size_*) happening in the same shader as the main
function, and secondly it doesn't work for variable group sizes.
In both cases, the simplest fix is to move the setup of these derived
values to a later time, similar to how the gl_VertexID workarounds are
done. There already exist system values defined for both of the derived
values, so we use them unconditionally, and lower them after linking is
performed.
While we're at it, we move to using gl_LocalGroupSizeARB instead of
gl_WorkGroupSize for variable group sizes.
Also the dead code elimination avoidance can be removed, since there
can be situations where gl_LocalGroupSizeARB is needed but has not been
inserted for the shader with main function. As a result, the lowering
code has to insert its own copies of the system values if needed.
Reported-by: Stephane Chevigny <stephane.chevigny@polymtl.ca>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103393
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
The pass only looks at var load/store intrinsics, not input load/store
intrinsics, so assert that we don't see the other type.
v2: Adjust comment indentation.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
v2:
- Use helper to add a new source to the texture instruction.
v3:
- Use nir_tex_instr_src_index() to simplify the patch (Jason).
Signed-off-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
We only need to add a check to validate output locations here. For
inputs with invalid locations we will fail to link when we can't
find a matching output in the same (invalid) location.
v2: compute location slots properly depending on shader stage and
variable type / direction
Fixes:
KHR-GL45.enhanced_layouts.varying_location_limit
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
We won't split varyings marked as always active because there
is no point in doing so. This means we need to mark both
sides of the interface as always active otherwise we will have
a mismatch and start removing things we shouldn't.
Reviewed-by: Eric Anholt <eric@anholt.net>
This is intended to be called before nir_lower_io() so that we
can do some linking optimisations with the results. It can also
be used with drivers that don't use nir_lower_io() at all such
as RADV.
v2: pass mode mask rather than first and last stage integer.
Reviewed-by: Eric Anholt <eric@anholt.net>
ssize_t is a GNU extension and is not available on Windows or MacOS.
Instead, we use intptr_t which should be effectively equivalent and is
part of the C standard. This should fix the Windows and Mac OS builds.
Fixes: 3af1c82989
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103253
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
Tested-by: Vinson Lee <vlee@freedesktop.org>
Since blob.h moved up to src/compiler the test should include that
instead of src/compiler/glsl
fixes: 0e3bd56c6e ("compiler: Move blob up a level")
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This looks like a copy+paste error. They don't actually write into that
variable as would be implied by putting the return there.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: mesa-stable@lists.freedesktop.org
We didn't fold correctly in the case of 0x1 because we never let the
loop counter hit 0. Switching it to bit >= 0 solves this problem.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Cc: mesa-stable@lists.freedesktop.org
There are certain advantages to using uint8_t internally such as
well-defined arithmetic on all platforms. However, interfaces that
work in terms of raw data should use a void* type.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
These helpers not only call blob_reserve_bytes but also make sure that
the blob is properly aligned as if blob_write_* were called.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Despite the name, it could only be used if you immediately wrote to the
pointer. Noboby was using it outside of one test, so clearly this
behavior wasn't that useful. Instead, make it return an offset into the
data buffer so that the result isn't invalidated if you later write to
the blob. In conjunction with blob_overwrite_bytes(), this will be
useful for leaving a placeholder and then filling it in later, which
we'll need to do for handling phi nodes when serializing NIR.
v2 (Jason Ekstrand):
- Detect overflow in the offset + to_write computation
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
These can be used to easily count up the number of bytes that will be
required by "writing" it into the NULL blob.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
There's no reason why that tiny bit of memory needs to be on the heap.
We always put blob_reader on the stack, so why not do the same with the
writable blob.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
We're going to want to use the blob for Vulkan pipeline caching so it
makes sense to have it in libcompiler not libglsl.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Otherwise we could have a failure followed by a smaller write that
succeeds and get a corrupted blob. If we ever OOM, we should stop.
v2 (Jason Ekstrand):
- Initialize the new boolean member in create_blob
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Cc: mesa-stable@lists.freedesktop.org
Otherwise, if you have a large read fail and then try to do a small
read, the small read may succeed even though it's at the wrong offset.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Cc: mesa-stable@lists.freedesktop.org
For TGSI-based drivers, st_glsl_to_tgsi records this information.
For NIR-based drivers, nir_shader_gather_info() will do so.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
I'd like to put this sort of metadata in the shader_info structure,
rather than adding more things to gl_program.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
With the ssao demo from Vulkan demos:
radv/rx480: 440->440fps
anv/haswell: 24->34 fps
The demo does a 0->32 loop across a ubo with 32 members.
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
I've been doing this inside of vc4, but vc5 wants it as well and it may be
useful for other drivers (Intel has a related path for pre-gen6 with MRT,
and freedreno had a TGSI path for it at one point).
This required defining a common enum for the standard comparison
functions, but other lowering passes are likely to also want that enum.
v2: Add to meson.build as well.
Acked-by: Rob Clark <robdclark@gmail.com>
Unlike uniforms, the limit on shared memory size is not called out
explicitly in the list of things that cause linker errors, but presumably
that's just an oversight in the spec.
Fixes dEQP-GLES31.functional.debug.negative_coverage.{callbacks,get_error,log}.compute.exceed_shared_memory_size_limit
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
c7affbf687 enabled GLSLOptimizeConservatively on some
drivers. The idea was to speed up compile times by running
the GLSL IR passes only once each time do_common_optimization()
is called. However loop unrolling can create a big mess and
with large loops can actually case compile times to increase
significantly due to a bunch of redundant if statements being
propagated to other IRs.
Here we make sure to clean things up before moving on.
There was no measureable difference in shader-db compile times,
but it makes compile times of some piglit tests go from a couple
of seconds to basically instant.
The shader-db results seemed positive also:
Totals:
SGPRS: 2829456 -> 2828376 (-0.04 %)
VGPRS: 1720793 -> 1721457 (0.04 %)
Spilled SGPRs: 7707 -> 7707 (0.00 %)
Spilled VGPRs: 33 -> 33 (0.00 %)
Private memory VGPRs: 3140 -> 2060 (-34.39 %)
Scratch size: 3308 -> 2180 (-34.10 %) dwords per thread
Code Size: 79441464 -> 79214616 (-0.29 %) bytes
LDS: 436 -> 436 (0.00 %) blocks
Max Waves: 558670 -> 558571 (-0.02 %)
Wait states: 0 -> 0 (0.00 %)
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
The old code assumed that loop terminators will always be at
the start of the loop, resulting in otherwise unrollable
loops not being unrolled at all. For example the current
code would unroll:
int j = 0;
do {
if (j > 5)
break;
... do stuff ...
j++;
} while (j < 4);
But would fail to unroll the following as no iteration limit was
calculated because it failed to find the terminator:
int j = 0;
do {
... do stuff ...
j++;
} while (j < 4);
Also we would fail to unroll the following as we ended up
calculating the iteration limit as 6 rather than 4. The unroll
code then assumed we had 3 terminators rather the 2 as it
wasn't able to determine that "if (j > 5)" was redundant.
int j = 0;
do {
if (j > 5)
break;
... do stuff ...
if (bool(i))
break;
j++;
} while (j < 4);
This patch changes this pass to be more like the NIR unrolling pass.
With this change we handle loop terminators correctly and also
handle cases where the terminators have instructions in their
branches other than a break.
V2:
- fixed regression where loops with a break in else were never
unrolled in v1.
- fixed confusing/wrong naming of bools in complex unrolling.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
do-while loops can increment the starting value before the
condition is checked. e.g.
do {
ndx++;
} while (ndx < 3);
This commit changes the code to detect this and reduces the
iteration count by 1 if found.
V2: fix terminator spelling
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Elie Tournier <elie.tournier@collabora.com>
Tested-by: Dieter Nützel <Dieter@nuetzel-hh.de>
