I started seeing
ACO ERROR:
In file ../src/amd/compiler/aco_validate.cpp:98
Operand and Definition types do not match: s1: %44 = p_parallelcopy %158
test_basic: ../src/amd/compiler/aco_interface.cpp:85: void validate(aco::Program*):
Assertion `is_valid' failed.
since commit 52ee4cf229 ("nir/builder: Teach nir_pack_bits and
nir_unpack_bits about 32_4x8").
Fixes: e0d232c2fc ("aco: implement nir_op_pack_32_4x8"). I
Suggested-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27972>
If otherwise no helper lanes are required by the shader, then demote
behaves like discard and immediately terminates the invocations.
With maximal reconvergence, however, we need to ensure that helper lanes
are not terminated unless the entire quad was demoted.
In order to fix this, generally enable helper lanes in this unlikely
corner case and avoid a major refactor.
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27277>
The upside is that this removes 600 lines of code. The downside is
that if instance divisors are used, we will compile the VS on demand.
Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/27120>
Because we use unaligned dispatches, 1D launches only use 8 threads per
wave. Converting to 2D and fixing up launch IDs in the prolog
significantly increases occupancy.
Gives ~30% uplift in Ghostwire Tokyo.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26105>
This currently has no effects because the store_output instructions
are removed earlier (in ac_nir_lower_ps). Though, this will be needed
for exporting MRTZ from PS epilogs for alpha to coverage on RDNA3.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26413>
Otherwise, we can transition to exact before p_jump_to_epilog, then
transition to WQM again and then back to exact:
p_jump_to_epilog //transitions to exact
p_logical_end //transitions to wqm
p_end_wqm //transitions to exact
We rely on ssa elimination to clean most of this up.
fossil-db (navi21):
Totals from 1 (0.00% of 79330) affected shaders:
Instrs: 111 -> 110 (-0.90%)
CodeSize: 572 -> 568 (-0.70%)
Copies: 16 -> 15 (-6.25%)
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25440>
The main change is to use struct radv_vs_prolog_key directly instead of
the compressed representation to simplify an upcoming rework in prolog /
epilog caching. In doing so the state struct pointer was replaced with
an inline struct.
Care was also taken to pre-mask all the states with the active attribute
mask and other masks when it makes sense; this ensures that we don't
accidentally use information not hashed into the key during compilation.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/26023>
Because a LDS load is 2 separate loads on gfx10+ with wave64, a different wave
can write LDS in between and cause a non uniform result. Use v_readfirst_lane
instead of p_as_uniform because it cannot be copy propagated.
Fixes a OpenCL CTS test with zink+rusticl.
Totals from 136 (0.17% of 78196) affected shaders:
MaxWaves: 3236 -> 3244 (+0.25%)
Instrs: 130069 -> 131221 (+0.89%)
CodeSize: 698048 -> 703436 (+0.77%)
VGPRs: 5464 -> 5440 (-0.44%)
SpillSGPRs: 94 -> 96 (+2.13%)
Latency: 5361017 -> 5363781 (+0.05%); split: -0.00%, +0.05%
InvThroughput: 883010 -> 884100 (+0.12%)
SClause: 3822 -> 3821 (-0.03%); split: -0.05%, +0.03%
Copies: 14220 -> 14314 (+0.66%); split: -0.01%, +0.68%
Branches: 4549 -> 4551 (+0.04%)
PreSGPRs: 4934 -> 4940 (+0.12%)
PreVGPRs: 4666 -> 4655 (-0.24%)
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25973>
According to RadeonSI, this is required for preemption, user queues,
and we only have to wait for VS after streamout which should be more
performant.
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25284>
PS with epilog does not need to fix_exports. And radeonsi use
p_end_with_regs so does not have jump instruction at last.
radeonsi may also have exec restore instruction, so may break
before reach to p_end_with_regs.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24973>
radeonsi need to compact color export for ps epilog while radv does not.
radv will fill empty color slot, so won't affected by this change.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24973>
