Now that we generate built-in functions inline, there's no need to link
against the built-in shader, and no built-in prototypes to consider.
This lets us delete a bunch of code.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by; Ian Romanick <ian.d.romanick@intel.com>
This is now handled directly by ast_function.cpp.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by; Ian Romanick <ian.d.romanick@intel.com>
In the past, we imported the prototypes of built-in functions, generated
calls to those, and waited until link time to resolve the calls and
import the actual code for the built-in functions.
This severely limited our compile-time optimization opportunities: even
trivial functions like dot() were represented as function calls. We
also had no way of reasoning about those calls; they could have been
1,000 line functions with side-effects for all we knew.
Practically all built-in functions are trivial translations to
ir_expression opcodes, so it makes sense to just generate those inline.
Since we eventually inline all functions anyway, we may as well just do
it for all built-in functions.
There's only one snag: built-in functions that refer to built-in global
variables need those remapped to the variables in the shader being
compiled, rather than the ones in the built-in shader. Currently,
ftransform() is the only function matching those criteria, so it seemed
easier to just make it a special case.
On Skylake:
total instructions in shared programs: 12023491 -> 12024010 (0.00%)
instructions in affected programs: 77595 -> 78114 (0.67%)
helped: 97
HURT: 309
total cycles in shared programs: 137239044 -> 137295498 (0.04%)
cycles in affected programs: 16714026 -> 16770480 (0.34%)
helped: 4663
HURT: 4923
while these statistics are in the wrong direction, the number of
hurt programs is small (309 / 41282 = 0.75%), and I don't think
anything can be done about it. A change like this significantly
alters the order in which optimizations are performed.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by; Ian Romanick <ian.d.romanick@intel.com>
We want to check prior to optimization - otherwise we might fail to
detect cases where barrier() is in control flow which is always taken
(and therefore gets optimized away).
We don't currently loop unroll if there are function calls inside;
otherwise we might have a problem detecting barrier() in loops that
get unrolled as well.
Tapani's switch handling code adds a loop around switch statements, so
even with the mess of if ladders, we'll properly reject it.
Enforcing these rules at compile time makes more sense more sense than
link time. Doing it at ast-to-hir time (rather than as an IR pass)
allows us to emit an error message with proper line numbers.
(Otherwise, I would have preferred the IR pass...)
Fixes spec/arb_tessellation_shader/compiler/barrier-switch-always.tesc.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by; Ian Romanick <ian.d.romanick@intel.com>
It makes more sense to have this here where we store the other values
from xfb qualifiers. The struct it was previously part of is now only
used to store values that come from the api.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
The expectation is that drivers will set this based on
OES_geometry_shader and ARB_viewport_array support. This is a separate
enable on the same reasoning as for OES_texture_cube_map_array.
Signed-off-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Anuj Phogat <anuj.phogat@gmail.com>
OpAtomicLoad/Store should have pointer to images just like the rest of the
atomic operators. These couple of lines were poorly copied from the
ssbo/shared_vars cases (the only ones currently tests by the CTS).
Fixes 2afb950161 ("spirv/nir: Add support for OpAtomicLoad/Store")
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
VC4 was running into a major performance regression from enabling control
flow in the glmark2 conditionals test, because of short if statements
containing an ffract.
This pass seems like it was was trying to ensure that we only flattened
IFs that should be entirely a win by guaranteeing that there would be
fewer bcsels than there were MOVs otherwise. However, if the number of
ALU ops is small, we can avoid the overhead of branching (which itself
costs cycles) and still get a win, even if it means moving real
instructions out of the THEN/ELSE blocks.
For now, just turn on aggressive flattening on vc4. i965 will need some
tuning to avoid regressions. It does looks like this may be useful to
replace freedreno code.
Improves glmark2 -b conditionals:fragment-steps=5:vertex-steps=0 from 47
fps to 95 fps on vc4.
vc4 shader-db:
total instructions in shared programs: 101282 -> 99543 (-1.72%)
instructions in affected programs: 17365 -> 15626 (-10.01%)
total uniforms in shared programs: 31295 -> 31172 (-0.39%)
uniforms in affected programs: 3580 -> 3457 (-3.44%)
total estimated cycles in shared programs: 225182 -> 223746 (-0.64%)
estimated cycles in affected programs: 26085 -> 24649 (-5.51%)
v2: Update shader-db output.
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
We already support all of the decorations that require this capability.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Dave Airlie <airlied@redhat.com>
This was something that I wrote in the early days of the spirv_to_nir code
but deleted once we had a real driver. However, in the absence of a
shader_runner equivalent, it's extremely useful for debugging the
spirv_to_nir code so let's bring it back.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
This extension is a combination of AMD_vertex_shader_viewport_index and
AMD_vertex_shader_layer, making it rather trivial to implement.
For gallium I *think* this needs a new cap because of the addition of
support in tessellation evaluation shaders, and since I don't have any
hardware to test it on, I've left that for someone else to wire up.
Signed-off-by: Dylan Baker <dylanx.c.baker@intel.com>
Reviewed-by: Ilia Mirkin <imirkin@alum.mit.edu>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
This fixes 8 fs-interpolateat* piglit crashes on radeonsi, because it can't
handle non-input operands in interpolateAt*.
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
This adds support for the input attachments subpass type
to the SPIRV->NIR pass.
v1.1: drop handling from vtn_handle_texture
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
SPIR-V/Vulkan have a special image type for input attachments
called the subpass type. It has different characteristics than
other images types.
The main one being it can only be an input image to fragment
shaders and loads from it are relative to the frag coord.
This adds support for it to the GLSL types. Unfortunately
we've run out of space in the sampler dim in types, so we
need to use another bit.
v2: Fixup subpass input name (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Eero Tamminen <eero.t.tamminen@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
This is mandatory.
Cc: mesa-stable@lists.freedesktop.org
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Numeric 2 is actually GLSL_SAMPLER_DIM_3D, which I don't think is what
was intended.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
I want to re-use this in a different pass, so move to nir.h
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
This clears the last bits of the usecases of the hash table
located in mesa/program, allowing us to remove it.
V2: Rebase on top of changes to Makefile.sources
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
And change the include in glcpp.h accordingly.
V2: Whitespace fix
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Also do some minor whitespace cleanups
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
We are getting the util hash table through the include in
program/hash_table.h for the moment until we migrate the
string_to_uint_map to a separate file.
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
The "locals" hash table is used as a set, so use a set to
avoid confusion and also spare some minor memory.
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
V2: Fix incorrect ordering on hash table insert
V3: null check value returned by _mesa_hash_table_search()
(Timothy Arceri)
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
V2: Rebase to the adaption of new hashing functions
V3: move previous_label declaration to where it is used
(Timothy Arceri)
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
Here we make the prog_hash_table functionally equivalent to
the one in util by wrapping the remaing functions that differ.
We also move the functions to the header so we can remove the c
file.
This enables us to do a step-by-step replacement of the table.
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
This should prevent us from rebuilding the world.
Signed-off-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Timothy Arceri <timothy.arceri@collabora.com>
This change makes copy propagation pass faster. Complete link time
spent in test case attached to bug 94477 goes down to ~400 secs from
over 500 secs on my HSW machine. Does not fix the actual issue but
brings down the total. No regressions seen in CI.
v2: do not leak hash_table structure
Signed-off-by: Tapani Pälli <tapani.palli@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Unlike the current CSE pass, global value numbering is capable of detecting
common values even if one does not dominate the other. For instance, in
you have
if (...) {
ssa_1 = ssa_0 + 7;
/* use ssa_1 */
} else {
ssa_2 = ssa_0 + 7;
/* use ssa_2 */
}
Global value numbering doesn't care about dominance relationships so it
figures out that ssa_1 and ssa_2 are the same and converts this to
if (...) {
ssa_1 = ssa_0 + 7;
/* use ssa_1 */
} else {
/* use ssa_1 */
}
Obviously, we just broke SSA form which is bad. Global code motion,
however, will repair this for us by turning this into
ssa_1 = ssa_0 + 7;
if (...) {
/* use ssa_1 */
} else {
/* use ssa_1 */
}
This intended to eventually mostly replace CSE. However, conventional CSE
may still be useful because it's less of a scorched-earth approach and
doesn't require GCM. This makes it a bit more appropriate for use as a
clean-up in a late optimization run.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Previously, we dind't apply variable decorations to the members of a split
structure variable. This doesn't quite work, unfortunately, because things
such as the "flat" qualifier may get applied to an entire structure instead
of propagated to the members. This fixes 9 of the new CTS tests in the
dEQP-VK.glsl.linkage.varying.struct.* group.
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Cc: "12.0" <mesa-stable@lists.freedesktop.org>
