HLSL doesn't support bitcasting a 64bit integer to a double. DXIL
doesn't have generic pack/unpack instructions, so we lower those to
integer bitwise ops. As a result, NIR generic double pack/unpack would
require our backend to emit a bitcast to get a double, but we want
to match HLSL semantics and emit MakeDouble/SplitDouble.
Adding a dedicated opcode for double pack/unpack allows us to add a
pass to emit that instead, which lets our backend emit the right
instruction to pack and unpack doubles.
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10063>
GLSL and SPIR-V GLSL.std.450 don't have any requirements for fsign(NaN),
and both only require that FSign(-0.0) == 0.0. OpenCL, on the other
hand, requires sign(-0.0) be exactly -0.0. It also requires that
sign(NaN) be exactly 0.0.
In practice, this change is difficult to test. Our GLSL frontend
already constant folds sign(NaN) to 0.0 before even getting to NIR. As
far as I can tell, glslang does the same. I don't have a good way to
run an OpenCL SPIR-V test. Maybe SPIR-V GLSL.std.450 assembly?
No shader-db or fossil-db changes on any Intel platform.
Acked-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6358>
This commit only documents the current behavior, even if that behavior
is not the behavior preferred by the relevant specs.
In SPIR-V, there are two flavors of the sign instruction, and each lives
in an extended instruction set. The GLSL.std.450 FSign instruction is
defined as:
Result is 1.0 if x > 0, 0.0 if x = 0, or -1.0 if x < 0.
This also matches the GLSL 4.60 definition.
However, the OpenCL.ExtendedInstructionSet.100 sign instruction is
defined as:
Returns 1.0 if x > 0, -0.0 if x = -0.0, +0.0 if x = +0.0, or -1.0 if
x < 0. Returns 0.0 if x is a NaN.
There are two differences. Each treats -0.0 differently, and each also
treats NaN differently. Specifically, GLSL.std.450 FSign does not
define any specific behavior for NaN.
There has been some discussion in Khronos about the NaN behavior of
GLSL.std.450 FSign. As part of that discussion, I did some research
into how we treat NaN for nir_op_fsign, and this commit just captures
some of those notes.
v2: Document the expected behavior of nir_op_fsign more thoroughly.
Suggested by Rhys. Note that the current implementation of constant
folding does not produce the expected result for NaN. Suggested by
Caio.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com> [v1]
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6358>
This will create code that is easier to combine into MADs/FMA when the
last component is 1.0.
nir_opt_algebraic_late has an optimization to do something similar but it
only works for inexact code, if the multiplication-by-1 optimization is
done before it and if the backend enables fuse_ffma.
fossil-db (Navi):
Totals from 85583 (74.64% of 114665) affected shaders:
SGPRs: 4556060 -> 4558596 (+0.06%); split: -0.07%, +0.12%
VGPRs: 3315060 -> 3312984 (-0.06%); split: -0.23%, +0.17%
SpillSGPRs: 13552 -> 13553 (+0.01%)
CodeSize: 184962756 -> 184431388 (-0.29%); split: -0.32%, +0.03%
MaxWaves: 1208693 -> 1209361 (+0.06%); split: +0.17%, -0.11%
Instrs: 35678819 -> 35361617 (-0.89%); split: -0.91%, +0.02%
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5631>
All these instructions replicate the result of a N-component dot-product
to a vec4. Naming them fdot_replicatedN gives the impression that are
some sort of abstract dot-product that replicates the result to a vecN.
They also deviate from fdph_replicated... which nobody would reasonably
consider naming fdot_replicatedh.
Naming these opcodes fdotN_replicated more closely matches what they
are, and it matches the pattern of fdph_replicated.
I believe that the only reason these opcodes were named this way was
because it simplified the implementation of the binop_reduce function in
nir_opcodes.py. I made some fairly simple changes to that function, and
I think the end result is ok.
The bulk of the changes come from the sed rename:
sed --in-place -e 's/fdot_replicated\([234]\)/fdot\1_replicated/g' \
$(grep -r 'fdot_replicated[234]' src/)
v2: Use a named parameter to binop_reduce instead of using
isinstance(name, str). Suggested by Jason.
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5725>
These exist to convert between different types of boolean values. In
particular, we want to use these for uniform and shared memory
operations where we need to convert to a reasonably sized boolean but we
don't care what its format is so we don't want to make the back-end
insert an actual i2b/b2i. In the case of uniforms, Mesa can tweak the
format of the uniform boolean to whatever the driver wants. In the case
of shared, every value in a shared variable comes from the shader so
it's already in the right boolean format.
The new boolean conversion opcodes get replaced with mov in
lower_bool_to_int/float32 so the back-end will hopefully never see them.
However, while we're in the middle of optimizing our NIR, they let us
have sensible load_uniform/ubo intrinsics and also have the bit size
conversion.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4338>
This opcode is the same as the f2f16 opcode except that it comes with
a promise that it is safe to optimise it out if the result is
immediately converted back to a 32-bit float again. Normally this
would be a lossy conversion and so it would be visible to the
application, but if the conversion is generated as part of the mediump
lowering process then this removal doesn’t matter. The opcode is
eventually replaced with a regular f2f16 in the late optimisations so
the backends don’t need to handle it.
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3822>
uctz isn't added because it will implemented in the GLSL path and the
SPIR-V path using other pre-existing instructions.
v2: Avoid signed integer overflow for uabs_isub(0, INT_MIN). Noticed by
Caio.
v3: Alternate fix for signed integer overflow for abs_sub(0, INT_MIN).
I tried the previous methon in a small test program with -ftrapv, and it
failed.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com> [v1]
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/767>
This introduces new vec8 and vec16 instructions (which are the only
instructions taking more than 4 sources), in order to construct 8 and 16
component vectors.
In order to avoid fixing up the non-autogenerated nir_build_alu() sites
and making them pass 16 src args for the benefit of the two instructions
that take more than 4 srcs (ie vec8 and vec16), nir_build_alu() is has
nir_build_alu_tail() split out and re-used by nir_build_alu2() (which is
used for the > 4 src args case).
v2 (Karol Herbst):
use nir_build_alu2 for vec8 and vec16
use python's array multiplication syntax
add nir_op_vec helper
simplify nir_vec
nir_build_alu_tail -> nir_builder_alu_instr_finish_and_insert
use nir_build_alu for opcodes with <= 4 sources
v3 (Karol Herbst):
fix nir_serialize
v4 (Dave Airlie):
fix serialization of glsl_type
handle vec8/16 in lowering of bools
v5 (Karol Herbst):
fix load store vectorizer
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Adds binop_reduce_all_sizes which generates both 1-bit and 32-bit
versions of the reduce operation. This reduces the code duplication a
bit and will make it easier to later add 16-bit versions as well.
Reviewed-by: Rob Clark <robdclark@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Adds binop_compare_all_sizes which generates both 1-bit and 32-bit
versions of the comparison operation. This reduces the code
duplication a bit and will make it easier to later add 16-bit versions
as well.
Reviewed-by: Rob Clark <robdclark@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Used for address/offset calculation (ie. array derefs), where we can
potentially use less than 32b for the multiply of array idx by element
size. For backends that support `imul24`, this gives a lowering pass
an easy way to find multiplies that potentially can be converted to
`imul24`.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Some hardware can do 24b multiply in a single instruction, but not 32b.
However in most cases 24b is sufficient for address/offset calculation.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Jason Ekstrand