LLVM and the proprietary compiler seem to do this
Fixes: b01847bd9 ("aco/gfx10: Fix mitigation of VMEMtoScalarWriteHazard.")
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 93c8ebfa ('aco: Initial commit of independent AMD compiler')
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes:
Wolfenstein:Youngblood (w/o shader_ballot)
dEQP-VK.descriptor_indexing.combined_image_sampler_in_loop_with_lod
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
I don't think the bug applies for global/scratch instructions and
load_barycentric_at_sample selection expects this feature to work.
Fixes various dEQP-VK.pipeline.multisample_interpolation.* tests on GFX10.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-By: Timur Kristóf <timur.kristof@gmail.com>
It shouldn't matter, but the 1 was leftover from when it was handled
together with workgroup_size and num_work_groups.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
The former was always true and hence dead code. We will want to
explicitly declare the ring offset register with ACO, but we also want
to declare the scratch offset too, and we can't try to disable it since
ACO also supports spilling and the determination of whether spilling has
to happen occurs well after setting up registers. So replace
supports_spill with something that will actually be used for ACO.
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Due to how LLVM works we have to make some of the FS inputs become
vectors, and therefore have to split them early so that they don't take
up extra register pressure due to how RA currently works.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Fixes: d1b9deee ('aco: improve waitcnt insertion around loops')
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Do this by repeating processing of loops until no progress is made.
Totals from affected shaders:
SGPRS: 162576 -> 162576 (0.00 %)
VGPRS: 145228 -> 145228 (0.00 %)
Spilled SGPRs: 668 -> 668 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 15778640 -> 15771336 (-0.05 %) bytes
LDS: 146 -> 146 (0.00 %) blocks
Max Waves: 6087 -> 6087 (0.00 %)
v2: use block_kind_loop_header/block_kind_loop_exit to repeat at the end
of loops instead of at each continue
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
The multiplication reduction is larger than it could be, but it should be
easier to implement this way.
No failures with dEQP-VK.subgroups.*int64* except those caused by LLVM
being used for other stages.
v2: don't call setFixed() for v_add carry-out, since setHint sets physReg
v3: add and use emit_vadd32() helper
v4: use num_opcodes instead of last_opcode
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev> (v3)
Should make 64-bit integer reductions easier to implement.
v4: use num_opcodes instead of last_opcode
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev> (v3)
This actually supports more of the extension than the LLVM backend but we
can't enable it because ACO doesn't work with all stages yet.
With more of it enabled, some CTS tests fail because our 64-bit sqrt
is very imprecise. I can't find any precision requirements for it
anywhere, so I'm thinking it might be a CTS issue.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Fixes: 93c8ebfa ('aco: Initial commit of independent AMD compiler')
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
No pipeline-db changes
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Fixes: 93c8ebfa ('aco: Initial commit of independent AMD compiler')
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
This should remove most of the excess code size that was
introduced by making all booleans per-lane.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Previously, instruction selection had two kinds of booleans:
1. divergent which was per-lane and stored in s2 (VCC size)
2. uniform which was stored in s1
Additionally, uniform booleans were made per-lane when they resulted
from operations which were supported only by the VALU.
To decide which type was used, we relied on the destination size,
which was not reliable due to the per-lane uniform bools, but it
mostly works on wave64.
However, in wave32 mode (where VCC is also s1) this approach
makes it impossible keep track of which boolean is uniform and
which is divergent.
This commit makes all booleans per-lane.
The resulting excess code size will be taken care of by the optimizer.
v2 (by Daniel Schürmann):
- Better names for some functions
- Use s_andn2_b64 with exec for nir_op_inot
- Simplify code due to using s_and_b64 in bool_to_scalar_condition
v3 (by Timur Kristóf):
- Fix several subgroups regressions
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
ACO's optimizer would try to propagate 64-bit constants, but
does so in such a way that wouldn't work due to how the 64-bit
constants are handled in the IR.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
This patch tries to give instructions with the same execution
mask also the same pass_flags and enables VN for SALU instructions
using exec as Operand.
This patch also adds back VN for VOPC instructions and removes VN for phis.
v2 (by Timur Kristóf):
- Fix some regressions.
v3 (by Daniel Schürmann):
- Fix additional issues
Reviewed-By: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 93c8ebfa ('aco: Initial commit of independent AMD compiler')
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 93c8ebfa ('aco: Initial commit of independent AMD compiler')
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Fixes: 93c8ebfa ('aco: Initial commit of independent AMD compiler')
Seems we can use DPP's row_mask field to get an effect similar to
modifying exec.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
It happens that some games try to access a vertex buffer without
a valid format. This case was incorrectly handled by
ac_get_tbuffer_format which made ACO emit an invalid instruction.
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Cc: 19.3 <mesa-stable@lists.freedesktop.org>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
In particular, increase the cost of 64-bit integer division.
Fixes huge shaders with dEQP-VK.spirv_assembly.type.scalar.i64.mod_geom
, with ACO used for GS this creates shaders requiring a branch with
>32767 dword offset.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
The workaround got accidentally moved to the wrong place
Fixes: 08d510010b aco: increase accuracy of SGPR limits
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
