LLVM loves take advantage of the fact that vec3s in OpenCL are 16B
aligned and so it can just read/write them as vec4s. This results in a
LOT of vec4->vec3 casts on loads and stores. One solution to this
problem is to get rid of all vec3 variables.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6871>
LLVM loves take advantage of the fact that vec3s in OpenCL are 16B
aligned so it can just read/write them as vec4s. This is questionably
legal except that it uses a xyz write-mask when it does it. The result
is a LOT of vec4->vec3 casts on loads and stores. This optimization
detects this case as well as other bit-cast cases and rewrites them to
get rid of the cast.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6871>
These are based on the ones which already existed in the load/store
vectorization pass but I made some improvements while moving them. In
particular,
1. They're both faster if the bit sizes are equal
2. The check is faster if old_bit_size > new_bit_size
3. The check now fails if it would use more than NIR_MAX_VEC_COMPONENTS
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6871>
This pass attempts to optimize three broad categories of memcpy:
1. Self-copies: These we can discard out-of-hand.
2. Vector copies: It doesn't matter what the vector size is or if the
source and destination have different vector types, it's still easy
enough to emit a load/store pair.
3. Tightly packed copies: In the case where a type is tightly packed
(no padding bits), we can replace the memcpy with a copy_deref
instruction which the optimizer is far better at handling.
This has proven capable of getting rid of many of the memcpy instances
in some rather gnarly OpenCL C kernels I've been looking at, even after
coming out of LLVM's optimizer.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6871>
In 9f3c595dfc, we attempted to handle casts in opt_find_array_copies
but missed a critical case. In particular, in the case where we begin
finding a copy but then encounter a cast, we need to discard everything
which might alias that cast.
Fixes: 9f3c595dfc "nir/find_array_copies: Handle cast derefs"
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6871>
This adds primarily two passes: One is a lowering pass which turns
these conversion intrinsics into a series of ALU ops. The other is an
optimization pass which attempt to simplify the conversion whenever
possible in the hopes that we can turn it into a "normal" conversion op
which doesn't need special treatment.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6945>
Most of these were originally written by Daniel Stone in the Microsoft
ClOn12 branch, reworked by Jesse Natalie, fixed by Boris Brezillon, and
possibly touched by others along the way. Unfortunately, none of that
is in the commit history thanks to living in the CLOn12 branch.
I ported them to mesa master and further reworked things for better
cosmetics. In particular,
1. They now live in a builder helper rather than in vtn_alu.c.
2. Instead of looping inside each builder helper, we just trust NIR
vector instructions to handle vectors.
3. Lots of re-arranging of the helpers for clarity, better asserting,
and better re-use with the upcoming lowering pass.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6945>
This new intrinsic is capable of handling the full range of conversions
from OpenCL including rounding modes and possible saturation. The
intention is that we'll emit this intrinsic directly from spirv_to_nir
and then lower it to ALU ops later.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6945>
We're about to introduce conversion ops which are going to want two
different types. We may as well just split the one we have rather than
end up with three. There are a couple places where this is mildly
inconvenient but most of the time I find it to actually be nicer.
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Reviewed-by: Daniel Stone <daniels@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6945>
It turns out I had missed a case in my enumeration of why everything
currently was vec4-aligned.
Fixes a simple testcase of loading from a vec3[2] array in freedreno with
IR3_SHADER_DEBUG=nouboopt.
Initial shader-db results look devastating:
total instructions in shared programs: 8019997 -> 12829370 (59.97%)
total cat6 in shared programs: 87683 -> 145840 (66.33%)
Hopefully this will recover once we introduce the i/o vectorizer, but that
was blocked on getting the vec3 case fixed.
Reviewed-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6612>
When we come in from some other level or from the parent, we need to
ensure that the reach set is in prev_frontier but we also need to
consider the dominance frontier of our level. Otherwise, we may end up
leaving out possible blocks when computing the reach of a level.
Acked-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6750>
There are two issues here:
1. If there are any phi nodes, we'll make complete hash of them. This
isn't likely actually a problem because spirv_to_nir doesn't
generate any actual phi nodes today. However, if we start doing any
other passes before this, we may have a problem.
2. Even without phi nodes, we may still break SSA form. This can
happen if we ever have to stick a block inside a conditional to
satisfy weird CFG constraints. Doing so can cause it to no longer
look like it dominates some of its uses even though, at runtime,
it's guaranteed to be run first.
Reviewed-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6750>
This commit adds a number of new validation checks:
1. We now check that every block pointer in the IR points to a block
that actually exists in a block list that's reachable from the
nir_function_impl.
2. We assert that nir_function_impl::body is non-empty
3. We assert that the start block has no predecessors. This is
important because we tend to put run-once code there.
4. We now validate some stuff on the end block.
Reviewed-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6750>
In the case where end was a instruction-based cursor, we would mix up
our blocks and end up with block_begin pointing after the second split.
This causes a segfault as the cf_node list walk at the end of the
function never terminates properly. There's also a possibility of
mix-up if begin is an instruction-based cursor which was found by
inspection.
Fixes: fc7f2d2364 "nir/cf: add new control modification API's"
Reviewed-by: Connor Abbott <cwabbott0@gmail.com>
Acked-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6866>
This isn't strictly necessary for freedreno, since we aren't using it
yet, but I wanted to avoid any problems if we do. If we wanted to handle
this "properly", and handle matrix and array per-view variables, we'd
probably want to encode the "view stride" (number of views per user
location) and base view in the intrinsic, but for now we just don't do
any offsetting and assume the indirect offset is the view.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6514>
Taken mostly directly from the anv pass. A few anv-specific things that
I could leave in anv aren't included. Specifically on turnip we don't
need to set gl_Layer to 0, and we can handle the case where the FS reads
gl_ViewIndex, so that check is moved into anv.
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6514>
We might have several identical options to vectorize an entry with, but
only one might be vectorizable because of writes interfering.
An example of this is a pattern found in some CTS tests:
a = load(0)
b = load(4)
store(0, a)
store(4, b)
a = load(0)
b = load(4)
store(0, a)
store(4, b)
...
It might have attempted to vectorize the first load(0) with the second
load(4) without attempting the second load(4) when the first fails. This
changes vectorize_entries() to continue even if the first try_vectorize()
failed.
fossil-db (Navi):
Totals from 117 (0.09% of 137413) affected shaders:
SGPRs: 7040 -> 7088 (+0.68%)
CodeSize: 276504 -> 276308 (-0.07%); split: -0.08%, +0.01%
Instrs: 51152 -> 51111 (-0.08%); split: -0.09%, +0.01%
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5415>
