LOCAL_STORAGE.zw is workgroup local memory, whereas LOCAL_STORAGE.xy is
thread local memory. Likewise PC_SP.zw is the stack pointer, which is
initialized to (LOCAL_STORAGE.zw + offset) but is modifiable by the
shader. Panfrost doesn't modify the s tack pointer, and the register
allocation logic assumes a zero offset, so let's always spill to thread
local memory = LOCAL_STORAGE.xy, as was intended by Italo's cleanup.
This is visible on any shader that spills. Compute shaders aren't
advertised yet, so WLS will be null, causing a fault like the following
(reproduced on Mali T860 with the glyphy trace):
[15634.148873] panfrost ff9a0000.gpu: Unhandled Page fault in AS0 at VA 0x0000000000000000
Reason: TODO
raw fault status: 0x70003C2
decoded fault status: SLAVE FAULT
exception type 0xC2: TRANSLATION_FAULT_LEVEL2
access type 0x3: WRITE
source id 0x700
[15634.658170] panfrost ff9a0000.gpu: gpu sched timeout, js=0, config=0x3300, status=0x8, head=0x31d4540,
tail=0x31d4540, sched_job=00000000e8101b2e
Fixes: 6a12ea02fe ("pan/mdg: properly encode/decode ldst instructions")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10866>
The compiler ABI specifies push uniforms at a 4-byte granularity (like
Bifrost), but Midgard require a 16-byte granularity. As such if the
number of pushed words is not a multiple of 4, there is a buffer overrun
at shader load time. Ordinarily this is inaccessible so the garbage is
ignored.
However, there was a great deal of confusion around the `uniform_cutoff`
variable. In some cases (such a full glmark2 run on a 64-bit processor),
the push uniforms would be at the end of a BO and the overrun would
cause a page fault.
Remove uniform_cutoff entirely and work with count directly to avoid
faulting, and round the count up to be defensive.
Closes: #4289
Fixes: ed810eb0a0 ("panfrost: Don't truncate uniform_count")
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Tested-by: Robin Murphy <robin.murphy@arm.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/9109>
While at it, rework the code to avoid copies between intermediate
structures: the pan_shader_info is passed to the compiler context so
the compiler can fill shader information directly.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8963>
So we can re-use the panfrost_sysvals definition outside of the
compiler without dragging the sysval_to_id hash table.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Acked-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8963>
LD_UNIFORM supports constant shifts and biases, just like LD, so take
advantage of that. Will avoid a regression in code quality from lowering
uniforms to UBOs.
No shader-db changes.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8973>
We want to add support for more memory types, so replace the is_shared
bool with an integer that is directly stored to load_store.arg_1.
The new memory type values are off by 0x40, as that bit now comes from
the index type.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8264>
Some MSAA+fmt combination require writeout to be split. Right now, it
only impacts blend shaders since we only support MSAA 4x, and the only
formats that could exceed the 128bit/pixel limit in MSAA 4x are
not supported by the fixed-function blend unit. We thus rely on the
blend shader to split things properly. Things will change once we add
MSAA 8x/16x to the mix, since even the blendable formats will exceed
the 128b/pixel limit in that case.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7984>
When constants are used in the blend equation we simply recompile the
shader.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7066>
No need to put it on the context, we can keep it local in mir_squeeze
and drop when we're done.
15.77KB leaked over 85 calls from:
0xffffaed3bfc3
in ??
_mesa_hash_table_rehash
at ../src/util/hash_table.c:368
in /home/alyssa/rockchip_dri.so
hash_table_insert
at ../src/util/hash_table.c:403
in /home/alyssa/rockchip_dri.so
find_or_allocate_temp
at ../src/panfrost/midgard/mir_squeeze.c:48
in /home/alyssa/rockchip_dri.so
find_or_allocate_temp
at ../src/panfrost/midgard/mir_squeeze.c:35
in /home/alyssa/rockchip_dri.so
mir_squeeze_index
at ../src/panfrost/midgard/mir_squeeze.c:76
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Cc: mesa-stable
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6373>
The alpha test is lowered by the frontend so we can drop that unused
code path, and stencil state can be computed at CSO create time to
reduce draw-time complexity.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Reviewed-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6195>
This commit removes the `ins->alu` field from midgard_instruction,
simplifying the code by just recreating midgard_vector_alu later when we
have to emit it.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5933>
This commit makes `ins->op` the correct field to use with load_store
instructions.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5933>
In an effort to simplify MIR by not prepacking instructions, this commit
removes references to `ins->alu.outmod` so that we can later remove the
`ins->alu` field from midgard_instruction.
Every place that was using `ins->alu.outmod` was changed to now use the
generic `ins->outmod` field instead.
We then reconstruct the outmod field right before emission.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5933>
In an effort to simplify MIR by not prepacking instructions, this commit
removes references to `ins->alu.reg_mode` so that we can later remove
the `ins->alu` field from midgard_instruction.
Every place that was using reg_mode was changed to now use the generic
`ins->src_type` field instead.
We then reconstruct the reg_mode field right before emission.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5933>
In an effort to simplify MIR by not prepacking instructions, this commit
removes references to `ins->alu.op` so that we can later remove the
`ins->alu` field from midgard_instruction.
Every place that was using ins->op was changed to now use the generic
`ins->op` field instead.
We then reconstruct the `alu.op` field right before emission.
This new field is generic and can contain opcodes for ALU, texture or
load/store instructions. It should be used in conjunction with
`ins->type`, just like the current prepacked `op` field.
Signed-off-by: Italo Nicola <italonicola@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5933>
We write to r2, which is preseved through to the blend shader, from
where it is read. We won't worry about MRT to keep things simple.
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5620>
Instead of requiring an explicit unoptimized move, we can implicitly
colour the blend input intrinsic to r0, where it will be preloaded; this
is a simple task for RA, and does not conflict with anything. If there
are multiple duplicate loads, the latter ones can just be simple moves
which will be copypropped.
We don't need to include a explicit synthetic load, since (scanning
backwards) the read will cause the input to become live at the right
time and the lack of an explicit write will keep it live from the
beginning of the shader. So no need to make it more complicated than it
needs to be.
Saves a cycle in blend shaders.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5449>
We don't make great use of it (due to the scheduler not being aware
yet), but we can pack for it regardless and maybe pick up some win.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5286>
We can pass it all off to emit time, and let the types in the IR do the
heavylifting in the meantime, which is a lot easier to get right.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5151>
Rather than recompute liveness every block, compute it just once for the
whole shader, which ends up more efficient.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5123>
Rather than O(N) each call, we can precompute the whole set - also O(N)
- and then subsequent checks are O(1).
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5123>
The test and rewrite were both accidentally O(N) to the shader size when
they should be O(1), so overall this takes the pass from O(N^2) to O(N).
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5123>
