We can only use linear for 2D images, not even 2D arrays. Even for 2D
images, we only want to use linear if:
* We are required to use linear due to window system requirements.
* The texture is streaming.
Otherwise, we want to use tiled textures. (Or better, compressed, but we
don't support that yet.)
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14903>
nir_alu_instr_is_comparison needs to consider all comparison opcodes regardless
of size. Otherwise, they will be missed by nir_opt_move/sink.
Without this change, lowering booleans to integers regresses register
pressure (and spills/fills) significantly in certain shaders on Panfrost,
like android/com.miHoYo.GenshinImpact/1420.shader_test.
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15073>
The Vulkan 1.3 spec says:
"The implementation-dependent limit bufferImageGranularity specifies
a page-like granularity at which linear and non-linear resources
must be placed in adjacent memory locations to avoid aliasing. Two
resources which do not satisfy this granularity requirement are said
to alias. bufferImageGranularity is specified in bytes, and must be
a power of two. Implementations which do not impose a granularity
restriction may report a bufferImageGranularity value of one.
Note: Despite its name, bufferImageGranularity is really a
granularity between "linear" and "non-linear" resources."
We set this limit to 64 bytes (a cacheline) at the dawn of time, without
any real rationale attached. There shouldn't be any restrictions here.
Our tile sizes are typically 4K, and tiled resource addresses are
aligned to the tile size, and the extent is also a multiple of the tile
sized. So if a linear resource occurs before a tiled one, there will
naturally be some space due to the alignment of the tiled resource's
starting address. If a linear resource occurs after a tiled one, the
tiled resource's ending address is already 4K aligned, which is already
guaranteeing that they won't share a cacheline.
So I think it should be fine to reduce this to 1. The other Vulkan
driver for our hardware seems to advertise 1 here as well.
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15066>
Instead, we only recompute liveness and we add new nodes and
interferences to the graph manually (we also need to patch
register classes in some cases).
To assist in this process, we also add an ip counter to our
instructions that we also recompute after each spill, which we use
to identify registers that cross thrsw boundries introduced with
TMU spills and fills and adjust their register classes accordingly
(removing their capacity to use accumulators).
This significantly reduces the CPU cost of spills. Using
shaders/closed/gputest/piano/7.shader_test as reference:
Compile time up to the first successful compile strategy in main is
~24s and with this change it is ~11s. With this speed up, we can now
try all 2-thread compile strategies (including the fallback scheduler)
in only ~15s.
A full shader-db run results in:
Total CPU time (seconds): 9904.67 -> 9087.98 (-8.25%)
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15041>
We may be pipelining TMU writes and reads, in which case we can
see both TMUWT and LDTMU at the end of a TMU sequence, so we should
not assume that a TMUWT always terminates a sequence.
Also, we had a bug where we were using inst instead of scan_inst
to check if we find another TMUWT after the curent instruction.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15041>
Instead of whether they are allowed to spill or not. This is more flexible.
Also, while we are not currently enabling spilling on any 4-thread strategies,
should we do that in the future, always prefer a 4-thread compile.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15041>
Until now we would only allow spilling as a last resort in the
last 2 strategies, however, it is possible that in some cases
earlier strategies may produce less spills if we allowed spilling
on them.
Likewise, the fallback scheduler can sometimes produce less spills
than 2 threads with optimizations disabled.
With this change, we start allowing all our 2-thread strategies to
spill, and instead of choosing the first strategy that is successful,
we choose the one that doesn't spill or the one with the least amount
of spilling.
It should be noted that this may incur in a significant increase
of compile times. We will address this in a follow-up patch.
Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15041>
Commit 38800b38 changed nir_opcodes.py, but that doesn't seem to have
triggered nir_opt_algebraic.py. The change in 75ef5991 depends on
opt_algebraic lowering 16-bit versions of slt, but if opt_algebraic is
not rebuilt, this may not happen. This resulted in some people seeing
assertion failures in, for example,
dEQP-VK.spirv_assembly.instruction.compute.float16.arithmetic_3.step,
due to the backend seeing nir_op_slt that it didn't know how to handle.
v2: Add nir_opcodes.py to nir_algebraic_py so that all the per-driver
algebraic passes pick up the dependency too. Rename it to
nir_algebraic_depends. Suggested by Emma.
Closes: #6047
Fixes: d1992255bb ("meson: Add build Intel "anv" vulkan driver")
Reviewed-by: Emma Anholt <emma@anholt.net>
Acked-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15050>
memcpy is divided into chunks that are vec4 sized max. The problem
here happens with a structure of 24 bytes :
struct {
float3 a;
float3 b;
}
If you memcpy that struct, the lowering will emit 2 load/store, one of
sized 8, next one sized 16. But both end up located at offset 0, so we
effectively drop 2 floats.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: a3177cca99 ("nir: Add a lowering pass to lower memcpy")
Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15049>
Alyssa Rosenzweig