It's just v_fma with fixed DPP8 and builtin s_waitcnt_expcnt, so it can mostly
be handled as a pure VALU instruction.
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/21023>
When a shader has a comparison with the subgroup invocation id,
we can use a constant instead, saving a VALU instruction.
When the constant can't be represented as a 64-bit literal,
use the s_bfm_b64 instruction to generate it instead, which
is still a win.
Fossil DB stats on GFX11:
Totals from 300 (0.22% of 134913) affected shaders:
CodeSize: 2223052 -> 2214336 (-0.39%); split: -0.43%, +0.04%
Instrs: 430216 -> 429882 (-0.08%); split: -0.14%, +0.06%
Latency: 5881180 -> 5878181 (-0.05%); split: -0.05%, +0.00%
InvThroughput: 731846 -> 729293 (-0.35%)
Copies: 31662 -> 31847 (+0.58%); split: -0.03%, +0.61%
Branches: 8241 -> 8100 (-1.71%)
PreVGPRs: 15788 -> 15786 (-0.01%)
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20843>
It looks like the LLVM assembler promotes true 16-bit instructions to VOP3
in this case.
No fossil-db changes.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20251>
fossil-db (gfx1100):
Totals from 112 (0.08% of 134574) affected shaders:
Scratch: 1513472 -> 1455360 (-3.84%)
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20534>
p_cvt_f16_f32_rtne will be lowered to v_cvt_f16_f32 and we already know that
preserves the high bits.
I tested the others on GFX1036.
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20574>
Apparently the TLS constructor doesn't work well if RADV
is instantiated multiple times and/or used by a program with
already existing threads.
Fixes: a128d444cb ('aco: use monotonic_buffer_resource for instructions')
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19219>
GFX11 has a lot of complicated data dependency hazards.
For debugging GFX10+ data dependency hazards. This creates an excessive
amount of s_waitcnt_depctr.
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/18273>
Add GFX11 support and use wait_imm::unset_counter. Looping in the waitcnt
pass was probably broken on GFX11 because of this.
fossil-db (gfx1100):
Totals from 899 (0.56% of 161689) affected shaders:
Instrs: 1319368 -> 1319179 (-0.01%)
CodeSize: 7124640 -> 7123884 (-0.01%)
Latency: 26554304 -> 26404606 (-0.56%)
InvThroughput: 9032485 -> 8978773 (-0.59%); split: -0.59%, +0.00%
No navi10 fossil-db changes.
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/17710>
As monotonic_buffer_resource is not thread-safe,
we use a thread_local instance which gets allocated once.
This change reduces the compile time spent in ACO by
approximately 10%.
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18112>
We would assemble an instruction writing vcc instead.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Fixes: 5ffc73896f ("aco/assembler: Fix v_cmpx with SDWA.")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18077>
The old name is no longer accurate.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Georg Lehmann <dadschoorse@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18077>
Instead of SDWA v_mov_b32/v_xor_b32, we can use a combination of
v_add_u16/v_sub_u16 (add/sub swap, similar to xor swap) and v_perm_b32
with a literal.
I don't know yet if GFX11 adds any new instructions which makes this
easier, but this approach should have full functionality.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16595>
This should be kept to only things aco uses, and expanded when
radeonsi support is added. Things should be removed if lowered in NIR.
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16342>
