Otherwise we may end up trying to coalesce in a case such as
ssa_1 = fadd r1, r2
r3.x = fneg(r2);
r3 = vec4(ssa_1, ssa_1.y, ...)
and that would cause us to move the writes to r3 from the vec to the
fadd which would re-order them with respect to the write from the fneg.
In order to solve this, we just don't coalesce if the destination of the
vec is not SSA. We could try to get clever and still coalesce if there
are no writes to the destination of the vec between the vec and the ALU
source. However, since registers only come from phi webs and indirects,
the chances of having a vec with a register destination that is actually
coalescable into its source is very slim.
Shader-db results on Haswell:
total instructions in shared programs: 13657906 -> 13659101 (<.01%)
instructions in affected programs: 149291 -> 150486 (0.80%)
helped: 0
HURT: 592
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105440
Fixes: 2458ea95c5 "nir/lower_vec_to_movs: Coalesce movs on-the-fly when possible"
Reported-by: Vadym Shovkoplias <vadym.shovkoplias@globallogic.com>
Tested-by: Vadym Shovkoplias <vadym.shovkoplias@globallogic.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Add helpers to get the number of src/dest components for an intrinsic,
and update spots that were open-coding this logic to use the helpers
instead.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Because nir_instr_remove is an inline wrapper around nir_instr_remove_v,
the compiler should be able to tell that the return value is unused and
not emit the extra code in most cases.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
ARB_gl_spirv adds the ability to use SPIR-V binaries, and a new
method, glSpecializeShader. Here we add a new function to do the
validation for this function:
From OpenGL 4.6 spec, section 7.2.1"
"Shader Specialization", error table:
INVALID_VALUE is generated if <pEntryPoint> does not name a valid
entry point for <shader>.
INVALID_VALUE is generated if any element of <pConstantIndex>
refers to a specialization constant that does not exist in the
shader module contained in <shader>.""
v2: rebase update (spirv_to_nir options added, changes on the warning
logging, and others)
v3: include passing options on common initialization, doesn't call
setjmp on common_initialization
v4: (after Jason comments):
* Rename common_initialization to vtn_builder_create
* Move validation method and their helpers to own source file.
* Create own handle_constant_decoration_cb instead of reuse existing one
v5: put vtn_build_create refactoring to their own patch (Jason)
v6: update after vtn_builder_create method renamed, add explanatory
comment, tweak existing comment and commit message (Timothy)
Future changes will add generated files used only from
src/compiler/glsl. These can't be built from Makefile.nir.am, and we
can't move all the rules from Makefile.nir.am to Makefile.spirv.am (and
it would be silly anyway).
v2: Do it for meson too.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Engestrom <eric.engestrom@imgtec.com> (the meson bits)
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com> (the automake bits)
Previously bitset.h would include u_math.h to get bitscan.h. u_math.h
lives in src/gallium/auxiliary/util while both bitset.h and bitscan.h
live in src/util. Having the one file directly include another file
that lives in the same directory makes much more sense.
As a side-effect, several files need to directly include standard header
files that were previously indirectly included.
v2: Fix build break in src/amd/common/ac_nir_to_llvm.c.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
I have no idea why but having dest_components == -1 was causing a memory
leak somewhere. Without this, you can't get through a full shader-db
run without running out of memory.
Reviewed-by: Rob Clark <robdclark@gmail.com>
When an if nesting inside anouther if is optimised away we can
end up with a loop terminator and following block that looks like
this:
if ssa_596 {
block block_5:
/* preds: block_4 */
vec1 32 ssa_601 = load_const (0xffffffff /* -nan */)
break
/* succs: block_8 */
} else {
block block_6:
/* preds: block_4 */
/* succs: block_7 */
}
block block_7:
/* preds: block_6 */
vec1 32 ssa_602 = phi block_6: ssa_552
vec1 32 ssa_603 = phi block_6: ssa_553
vec1 32 ssa_604 = iadd ssa_551, ssa_66
The problem is the phis. Loop unrolling expects the last block in
the loop to be empty once we splice the instructions in the last
block into the continue branch. The problem is we cant move phis
so here we lower the phis to regs when preparing the loop for
unrolling. As it could be possible to have multiple additional
blocks/ifs following the terminator we just convert all phis at
the top level of the loop body for simplicity.
We also add some comments to loop_prepare_for_unroll() while we
are here.
Fixes: 51daccb289 "nir: add a loop unrolling pass"
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105670
Apparently it is not happy about things like: .foo = {}
So skip over initializers for empty lists.
Fixes: 76dfed8ae2
Reported-by: Roland Scheidegger <sroland@vmware.com>
Signed-off-by: Rob Clark <robdclark@gmail.com>
I threatened to do this a long time ago.. I probably *should* have done
it a long time ago when there where many fewer intrinsics. But the
system of macro/#include magic for dealing with intrinsics is a bit
annoying, and python has the nice property of optional fxn params,
making it possible to define new intrinsics while ignoring parameters
that are not applicable (and naming optional params). And not having to
specify various array lengths explicitly is nice too.
I think the end result makes it easier to add new intrinsics.
v2: couple small fixes found with a test program to compare the old and
new tables
v3: misc comments, don't rely on capture=true for meson.build, get rid
of system_values table to avoid return value of intrinsic() and
*mostly* remove side-effects, add autotools build support
v4: scons build
Signed-off-by: Rob Clark <robdclark@gmail.com>
Acked-by: Dylan Baker <dylan@pnwbakers.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Not used in GL but 8 and 16 component vectors exist in OpenCL.
Signed-off-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Generated with
git grep -l nir_intrinsic_image | xargs \
sed -i 's/nir_intrinsic_image/nir_intrinsic_image_var/g'
and some manual fixing in nir_intrinsics.h
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Shader-db runtime change avarage of five runs:
Before 125,77 seconds (+/- 0,09%)
After 124,48 seconds (+/- 0,07%)
Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>
Reviewed-by: Eric Anholt <eric at anholt.net>
Make a simple worklist by basically just wrapping u_vector.
This is intended used in nir_opt_dce to reduce the number of calls
to ralloc, as we are currenlty spamming ralloc quite bad. It should
also give better cache locality and much lower memory usage.
Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>
Reviewed-by: Eric Anholt <eric at anholt.net>
Generalize the code for remove dead loops to also remove dead if
nodes. The conditions are the same in both cases, if the node (and
it's children) don't have side-effects AND the nodes after it don't
use the values produced by the node.
The only difference is when evaluating side effects: loops consider
only return jumps as a side-effect -- they can stop execution of nodes
after it; 'if' nodes outside loops should consider all kinds of
jumps (return, break, continue) since all of them can cause execution
of nodes after it to be skipped.
After this patch, empty ifs (those which both then and else blocks are
empty) will be removed by nir_opt_dead_cf.
It caused no change to shader-db, in part because the removal of empty
ifs is currently covered by nir_opt_peephole_select.
v2: Improve the identification of cases where break/continue can cause
side-effects. (Jason)
v3: Move code comment changes to a different patch. (Jason)
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
i965 and gallium handle the atomic buffer index differently. It was
just by luck that the single piglit test for this was passing.
For gallium we use the atomic binding so that we match the handling
in st_bind_atomics().
On radeonsi this fixes the CTS test:
KHR-GL43.shader_storage_buffer_object.advanced-write-fragment
It also fixes tressfx hair rendering in Tomb Raider.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This will only ever be used by gallium drivers so it probably doesn't
belong in the nir toolkit. Also we want to pass it some non NIR
things in the following patch.
To avoid regressions we wrap the lowering calls that have been moved
to st_glsl_to_nir with a quick hack so that they are only called for
radeonsi, we will replace the hack with a check for uniform packing
in a following patch.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This pass moves load UBO operations just before their first use,
loosely based on nir_opt_move_comparisons.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
OpenCL kernels also have int8/uint8.
v2: remove changes in nir_search as Jason posted a patch for that
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Rob Clark <robdclark@gmail.com>
Signed-off-by: Karol Herbst <kherbst@redhat.com>
This is based heavily on 97f10934ed, "ac/nir: Add vote_ieq/vote_feq
lowering pass." from Bas Nieuwenhuizen. This version is a bit more
general since it's in common code. It also properly handles NaN due to
not flipping the comparison for floats.
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The SPIR-V extension wants us to be able to do an AllEqual on any vector
or scalar type. This has two implications:
1) We need to be able to handle vectors so we switch the vote_eq
intrinsics to be vectorized intrinsics.
2) We need to handle floats which have different behavior with respect
to +-0, NaN, etc. than the integer variant so we need two variants.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Someone can make the lowering optional later if they want something
different for their hardware.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Samuel Iglesias Gonsálvez <siglesias@igalia.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Jason Ekstrand