This introduces two new lowering passes. One to lower VS to explicit
outputs using STLW and one to lower GS to load input using LDLW and
implement the GS specific functionality.
Signed-off-by: Kristian H. Kristensen <hoegsberg@google.com>
This allows us to make sure clipdist is emitted as a scalar array rather
than two vec4s. This matches SPIR-V semantics, and will be useful for
Zink.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
We want this for tessellation eventually, but we can turn it on now.
Shader-db results:
total instructions in shared programs: 8612905 -> 8611387 (-0.02%)
instructions in affected programs: 164952 -> 163434 (-0.92%)
total dwords in shared programs: 11952000 -> 11950560 (-0.01%)
dwords in affected programs: 68096 -> 66656 (-2.11%)
total full in shared programs: 315019 -> 315009 (<.01%)
full in affected programs: 1642 -> 1632 (-0.61%)
total constlen in shared programs: 2463654 -> 2463654 (0.00%)
constlen in affected programs: 0 -> 0
total (ss) in shared programs: 152379 -> 152409 (0.02%)
(ss) in affected programs: 1503 -> 1533 (2.00%)
total (sy) in shared programs: 96473 -> 96525 (0.05%)
(sy) in affected programs: 654 -> 706 (7.95%)
total max_sun in shared programs: 1172454 -> 1172472 (<.01%)
max_sun in affected programs: 104 -> 122 (17.31%)
Signed-off-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Set of opcodes doesn't have enough flexibility in certain cases. E.g.
Utgard PP has vector conditional select operation, but condition is always
scalar. Lowering all the vector selects to scalar increases instruction
number, so we need a way to filter only those ops that can't be handled
in hardware.
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>
This better matches all the other atomic intrinsics such as those for
SSBOs and shared variables where the sign is part of the intrinsic
opcode. Both generators (GLSL and SPIR-V) know the sign from the type
of the image variable or handle. In SPIR-V, signed min/max are separate
opcodes from unsigned.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
This is mostly the same as nir_move_load_const() but can also move
undef instructions, comparisons and some intrinsics (being careful with
loops).
v2: actually delete nir_move_load_const.c
v3: fix nir_opt_sink() usage in freedreno
v3: update Makefile.sources
v4: replace get_move_def with nir_can_move_instr and nir_instr_ssa_def
v4: handle if uses
v4: fix handling of nested loops
v5: re-write adjust_block_for_loops
v5: re-write setting of use_block for if uses
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Co-authored-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Eric Anholt <eric@anholt.net>
ir3_nir_analyze_ubo_ranges() has already told us how much of cb0 we
need to upload (all of it, since it will lower indirect UBO 0 accesses
from load_ubo back to indirection on the constant buffer).
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Rob Clark <robdclark@gmail.com>
That is: the five least significant bits provide the values of
'bits' and 'offset' which is the case for all hardware currently
supported by NIR and using the bfm/bfe instructions.
This patch also changes the lowering of bitfield_insert/extract
using shifts to not use bfm and removes the flag 'lower_bfm'.
Tested-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
total constlen in shared programs: 2485933 -> 2462236 (-0.95%)
Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
We only ever return the shader we were passed in (but internally
modified).
Reviewed-by: Rob Clark <robdclark@gmail.com>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
We originally had a single lower_fmod option. In commit 2ab2d2e5, Sam
split 32 and 64-bit lowering into separate flags, with the rationale
that some drivers might want different options there. This left 16-bit
unhandled, so Iago added a lower_fmod16 option in commit ca31df6f.
Now that lower_fmod64 is gone (in favor of nir_lower_doubles and
nir_lower_dmod), we re-combine lower_fmod16 and lower_fmod32 into a
single lower_fmod flag again. I'm not aware of any hardware which
need lowering for one bitsize and not the other.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
This can be used by both etnaviv and freedreno/a2xx as they are both vec4
architectures with some instructions being scalar-only.
Signed-off-by: Jonathan Marek <jonathan@marek.ca>
Reviewed-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
For a6xx, we construct/emit a single VS const state used for both
binning pass and draw pass. So far we were mostly getting lucky that
there were not (obvious) mismatches between the const_state (like
different lowered immediates) between the binning and draw pass
VS ir3_shader_variant.
And I guess this situation will come up more as GS and tess is added
into the equation.
Since really everything about the const state is not specific to the
variant, move this. The main exception is lowered immediates, but these
are the last to appear in the layout, and it doesn't hurt for each new
shader variant to just append any immed's it lowers to the end of the
immediate state.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Next patch moves const_state to ir3_shader, before the compile context
is created. So move the code around in prep to call it earlier.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Combine the offsets of differenet parts of the constant space with (what
was formerly known as) ir3_driver_const_layout. Bunch of churn, but no
functional change.
Signed-off-by: Rob Clark <robdclark@chromium.org>
I tried to be very careful while updating all the various drivers, but I
don't have any of that hardware for testing. :(
i965 is the only platform that sets always_precise = true, and it is
only set true for fragment shaders. Gen4 and Gen5 both set lower_flrp32
only for vertex shaders. For fragment shaders, nir_op_flrp is lowered
during code generation as a(1-c)+bc. On all other platforms 64-bit
nir_op_flrp and on Gen11 32-bit nir_op_flrp are lowered using the old
nir_opt_algebraic method.
No changes on any other Intel platforms.
v2: Add panfrost changes.
Iron Lake and GM45 had similar results. (Iron Lake shown)
total cycles in shared programs: 188647754 -> 188647748 (<.01%)
cycles in affected programs: 5096 -> 5090 (-0.12%)
helped: 3
HURT: 0
helped stats (abs) min: 2 max: 2 x̄: 2.00 x̃: 2
helped stats (rel) min: 0.12% max: 0.12% x̄: 0.12% x̃: 0.12%
Reviewed-by: Matt Turner <mattst88@gmail.com>
Driver which do not support native integers should use a lowering
pass to go from integers to floats.
Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Lower load_output to txf_ms_fb and add support for the new texture fetch
instruction.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Otherwise we zero out the state again, but all the UBO loads that we
could lower are already lowered. End result is that we didn't emit the
uniforms for lowered UBO access in any case where multiple shader
variants are used.
Fixes: 893425a607 freedreno/ir3: Push UBOs to constant file
Fixes: 3c8779af32 freedreno/ir3: Enable PIPE_CAP_PACKED_UNIFORMS
Signed-off-by: Rob Clark <robdclark@chromium.org>
We have a pass to lower global registers to locals and many drivers
dutifully call it. However, no one ever creates a global register ever
so it's all dead code. It's time we bury it.
Acked-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
When I implemented opt_if_loop_last_continue() I had restricted
this pass from moving other if-statements inside the branch opposite
the continue. At the time it was causing a bunch of spilling in
shader-db for i965.
However Samuel Pitoiset noticed that making this pass more aggressive
significantly improved the performance of Doom on RADV. Below are
the statistics he gathered.
28717 shaders in 14931 tests
Totals:
SGPRS: 1267317 -> 1267549 (0.02 %)
VGPRS: 896876 -> 895920 (-0.11 %)
Spilled SGPRs: 24701 -> 26367 (6.74 %)
Code Size: 48379452 -> 48507880 (0.27 %) bytes
Max Waves: 241159 -> 241190 (0.01 %)
Totals from affected shaders:
SGPRS: 23584 -> 23816 (0.98 %)
VGPRS: 25908 -> 24952 (-3.69 %)
Spilled SGPRs: 503 -> 2169 (331.21 %)
Code Size: 2471392 -> 2599820 (5.20 %) bytes
Max Waves: 586 -> 617 (5.29 %)
The codesize increases is related to Wolfenstein II it seems largely
due to an increase in phis rather than the existing jumps.
This gives +10% FPS with Doom on my Vega56.
Rhys Perry also benchmarked Doom on his VEGA64:
Before: 72.53 FPS
After: 80.77 FPS
v2: disable pass on non-AMD drivers
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Acked-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Add support for load_barycentric_pixel, load_interpolated_input, and
friends. For now, this retains support for old-style inputs, which can
probably be dropped with some ttn work.
Prep work for sample-shading support.
Signed-off-by: Rob Clark <robdclark@gmail.com>
This commit turns on the gallium cap and adds a pass to lower the
load_ubo intrinsics for block 0 back to load_uniform intrinsics and
adjust the backend where the cap switches units from vec4s to dwords.
As we stop using ir3_glsl_type_size() for uniform layout, this also
corrects an issue where we would allocate a vec4 slot for samplers in
uniforms, fixing:
dEQP-GLES3.functional.shaders.struct.uniform.sampler_array_fragment
dEQP-GLES3.functional.shaders.struct.uniform.sampler_array_vertex
dEQP-GLES3.functional.shaders.struct.uniform.sampler_nested_fragment
dEQP-GLES2.functional.shaders.struct.uniform.sampler_nested_vertex
dEQP-GLES2.functional.shaders.struct.uniform.sampler_nested_fragment
Signed-off-by: Kristian H. Kristensen <hoegsberg@chromium.org>
Reviewed-by: Rob Clark <robdclark@gmail.com>
For some things that show up when we expose higher glsl
TODO check blob traces to see if we have instructions for some of this?
I guess we don't but worth a check..
Signed-off-by: Rob Clark <robdclark@gmail.com>
These intrinsics have the offset in dwords already computed in the last
source, so the change here is basically using that instead of emitting
the ir3_SHR to divide the byte-offset by 4.
The improvement in shader stats is significant, of up to ~15% in
instruction count in some cases. Tested only on a5xx.
shader-db is unfortunately not very useful here because shaders that use
SSBO require GLSL versions that are not supported by freedreno yet.
For examples, most Khronos CTS tests under 'dEQP-GLES31.functional.ssbo.*'
are helped.
A random case:
dEQP-GLES31.functional.ssbo.layout.2_level_array.packed.row_major_mat3x2
with current master:
; CL prog 14/1: 1252 instructions, 0 half, 48 full
; 8 const, 8 constlen
; 61 (ss), 43 (sy)
with the SSBO dword-offset moved to NIR:
; CL prog 14/1: 1053 instructions, 0 half, 45 full
; 7 const, 7 constlen
; 34 (ss), 73 (sy)
The SHR previously emitted for every single SSBO instruction disappears
in most cases, and the dword-offset ends up embedded in the STGB
instruction as immediate in many cases as well.
There are also a few of those tests that are currently failing on register
allocation, that start to pass as a result of reducing the pressure. At least
these, probably more:
dEQP-GLES31.functional.ssbo.layout.random.unsized_arrays.24
dEQP-GLES31.functional.ssbo.layout.random.arrays_of_arrays.6
dEQP-GLES31.functional.ssbo.layout.random.arrays_of_arrays.17
dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays.14
dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.5
dEQP-GLES31.functional.ssbo.layout.random.nested_structs_arrays_instance_arrays.7
No regressions observed with relevant CTS and piglit tests.
Reviewed-by: Rob Clark <robdclark@gmail.com>
Which also requires uadd_carry lowering
Until recently this was lowered in glsl ir so it went unnoticed that we
weren't lowering it.
Fixes: 1d8994a63b glsl: [u/i]mulExtended optimization for GLSL
Signed-off-by: Rob Clark <robdclark@gmail.com>
I think this will save an instruction and hopefully not increase any other
costs (possibly the immediate -1 and 1?), but I haven't actually tested.
Reviewed-by: Kristian H. Kristensen <hoegsberg@chromium.org>
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
the naming is a bit confusing no matter how you look at it. Within SPIR-V
"global" memory is memory accessible from all threads. glsl "global" memory
normally refers to shader thread private memory declared at global scope. As
we already use "shared" for memory shared across all thrads of a work group
the solution where everybody could be happy with is to rename "global" to
"private" and use "global" later for memory usually stored within system
accessible memory (be it VRAM or system RAM if keeping SVM in mind).
glsl "local" memory is memory only accessible within a function, while SPIR-V
"local" memory is memory accessible within the same workgroup.
v2: rename local to function as well
v3: rename vtn_variable_mode_local as well
Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
On some GPUs, especially older Intel GPUs, some math instructions are
very expensive. On those architectures, don't reduce flow control to a
csel if one of the branches contains one of these expensive math
instructions.
This prevents a bunch of cycle count regressions on pre-Gen6 platforms
with a later patch (intel/compiler: More peephole select for pre-Gen6).
v2: Remove stray #if block. Noticed by Thomas.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Thomas Helland <thomashelland90@gmail.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
That flow control may be trying to avoid invalid loads. On at least
some platforms, those loads can also be expensive.
No shader-db changes on any Intel platform (even with the later patch
"intel/compiler: More peephole select").
v2: Add a 'indirect_load_ok' flag to nir_opt_peephole_select. Suggested
by Rob. See also the big comment in src/intel/compiler/brw_nir.c.
v3: Use nir_deref_instr_has_indirect instead of deref_has_indirect (from
nir_lower_io_arrays_to_elements.c).
v4: Fix inverted condition in brw_nir.c. Noticed by Lionel.
Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Move (most of) the ir3 compiler to src/freedreno/ir3 so that it can be
re-used by some future vulkan driver. The parts that are gallium
specific have been refactored out and remain in the gallium driver.
Getting the move done now so that it can happen before further
refactoring to support a6xx specific instructions.
NOTE also removes ir3_cmdline compiler tool from autotools build since
that was easier than fixing it and I normally use meson build. Waiting
patiently for the day that we can remove *everything* from the autotools
build.
Signed-off-by: Rob Clark <robdclark@gmail.com>