We would like to eliminate not just entire dead instructions, but also
dead components, which increases scheduler flexibility (since some
vector instructions can become scalar after eliminating dead
components). This also will allow better RA in the future.
Results are meh.
total instructions in shared programs: 3453 -> 3451 (-0.06%)
instructions in affected programs: 60 -> 58 (-3.33%)
helped: 2
HURT: 0
total bundles in shared programs: 1826 -> 1824 (-0.11%)
bundles in affected programs: 33 -> 31 (-6.06%)
helped: 2
HURT: 0
total quadwords in shared programs: 3144 -> 3144 (0.00%)
quadwords in affected programs: 0 -> 0
helped: 0
HURT: 0
total registers in shared programs: 321 -> 321 (0.00%)
registers in affected programs: 45 -> 45 (0.00%)
helped: 11
HURT: 11
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 16.67% max: 50.00% x̄: 39.70% x̃: 50.00%
HURT stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
HURT stats (rel) min: 100.00% max: 100.00% x̄: 100.00% x̃: 100.00%
95% mean confidence interval for registers value: -0.45 0.45
95% mean confidence interval for registers %-change: -1.87% 62.18%
Inconclusive result (value mean confidence interval includes 0).
total threads in shared programs: 445 -> 447 (0.45%)
threads in affected programs: 2 -> 4 (100.00%)
helped: 1
HURT: 0
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This allows for vec16 dependencies in the scheduler, not that we have
any yet (thankfully).
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Now that we have notion of byte masks, liveness tracking can be updated
to reflect this extra granularity without loss of correctness.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Read component masks don't have a particular type associated, since the
type of the ALU operation may not match the type of the operands in
question. So let's generate byte masks instead, and update the rest of
the compiler to use byte masks when analyzing reads.
Preparation for mixed types.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
There are essentially two formats of masks in play beginning with this
commit: masks per-channel and masks per-byte. The former make sense
within a given fixed-size instruction; the latter are
typesize-independent. It turns out you need the latter to meaningfully
manipulate instructions containing multiple sizes (which is quite
possible with ALU operations).
Similarly, we have mir_srcsize. We calculate the size of the source by
analyzing the size of the instruction itself and stepping down if there
is a half-modifier.
Finally, we have mir_round_bytemask_down, for when we want to take a
byte mask and "round it down" to a given component size, so that we can
use it as a component mask.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
This will allow us to encode properties about the load/store ops like we
do for ALU ops. We include now properties about whether we have a store,
and if there are special cases on the load/store op. We also tag each
instruction by its natural size... this is probably not totally right,
but it's a start.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
The trick is realizing even with a destination override, the masks are encoded in the same mode as the
instruction itself, rather than stepping down. The override means that
the smaller type is used, but the mask is parsed as if it were the
higher type. Overriding down is down by printed by blinding doing this. Overriding up can be thought of as printing in the upper size, but shifting the alphabet to use the upper half, i.e. shifting xyzw to become abcd.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Add some comments explaining what's going on in a more natural flow in
order to solve the actual bug.
Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Fixes: 2d914ebe81 ("pan/midgard: Fix memory corruption in register spilling")
This allows a write to proceed to an uninitialized part of a buffer
even when the GPU is using the previously-initialized portions.
Such a situation can be triggered with the following API usage example:
glBufferSubData(..., offset, size, data1);
glDrawArrays(...);
// append new vertex data
glBufferSubData(..., offset+size, size, data2);
glDrawArrays(...);
Same is done for freedreno, nouveau and radeon.
Signed-off-by: Christian Gmeiner <christian.gmeiner@gmail.com>
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>
This is the layout used in the GL API, and maps directly to PIPE
formats with no endianness trickery. As with the LA change, this
fixes big-endian fetching from texbos. Also cleans up some endian
shenanigans in shader images.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Now that Mesa is also using an array format for LA, nothing was using
these. (And, clearly, no HW driver had exposed them).
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The array format is what the GL API wants (fixing texbos on
big-endian), and matches directly to gallium's corresponding array
format. The only driver exposing A8L8 was radeon/r200 in big-endian,
where the HW's underlying format was trying to read as array and we
needed to flip things around to make our packed format come out right
(note that while the radeon format tables had both AL and LA,
ChooseTextureFormat would only pick one of them based on endianness).
v2: Don't make r200/radeon use endian swaps.
v3: Rebase on dropping the r200 _be/_le format table removal patch
v4: reword commit message to explain why we can drop both formats
from radeon.
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
The array format is what the GL API wants (and we made a mistake in
the format returned for texbos on big-endian!), and it's exactly what
the gallium-side PIPE_FORMAT_L16A16 is. The only downside is that
dri_util tries to fall back to sampling RG16 using LA16, which doesn't
have a match for big-endian any more. No HW drivers supported A16L16
anyway.
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
The first arg to OUT_BATCH_RELOC is ignored, we actually wanted these
in the third arg. They're always 0 so far, so it didn't matter.
v2: Reword commit message that I don't end up using the tile bits, but
keep the commit as a cleanup anyway.
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
No matter what, we deref the texFormat from the table, except for a
mistake in cpp=4 where we pulled a 0 out of the table either way.
v2: Rebase on dropping r200 table deduplication patch.
Reviewed-by: Marek Olšák <marek.olsak@amd.com> (v1)
PP stack size should be set to maximum PP stack size, not to stack size of
last shader.
Fixes: 27e7603c34 ("lima: fix ppir spill stack allocation")
Tested-by: Icenowy Zheng <icenowy@aosc.io>
Reviewed-by: Erico Nunes <nunes.erico@gmail.com>
Signed-off-by: Vasily Khoruzhick <anarsoul@gmail.com>
Similiar to iadd, we can fold an added constant value from an imad24_ir3
into the load_uniform's constant offset. This avoids some cases where
the addition of imad24_ir3 could otherwise be a regression in instr
count.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
We can't encode immed sources for cat3 (mad) instructions, but we can
use const in first or third src. We handled this case already, but we
weren't considering that we could lower immed to const.
For manhattan:
total instructions in shared programs: 35202 -> 34718 (-1.37%)
instructions in affected programs: 14931 -> 14447 (-3.24%)
helped: 90
HURT: 0
total full in shared programs: 2451 -> 2359 (-3.75%)
full in affected programs: 653 -> 561 (-14.09%)
helped: 69
HURT: 2
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Lower amul to either imul or imul24, depending on whether 24b is enough
bits to calculate an offset within the thing being dereferenced.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Used for address/offset calculation (ie. array derefs), where we can
potentially use less than 32b for the multiply of array idx by element
size. For backends that support `imul24`, this gives a lowering pass
an easy way to find multiplies that potentially can be converted to
`imul24`.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Some hardware can do 24b multiply in a single instruction, but not 32b.
However in most cases 24b is sufficient for address/offset calculation.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
ir3 compiler has a signed integer multiply-add instruction (MAD_S24)
that is used for different offset calculations in the backend.
Since we intend to move some of these calculations to NIR, we need
a new ALU op that can directly represent it.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Acked-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
Otherwise, if the base type is (for example) uint32, we would
incorrectly think that PoT optimizations could not apply.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Jason Ekstsrand <jason@jleksrand.net>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Reviewed-by: Eduardo Lima Mitev <elima@igalia.com>
The pass should run once at the end of shader compilation, for a4xx
onwards. It iterates texture sampling instructions and mark those
eligibile for pre-dispatch by changing the tex op from 'tex' to
'tex_prefetch'. An instruction is eligibile if:
* The coordinate is a vector where all its components come from a
shader input.
* The order of the components match exactly that of the input (no
swizzles).
* The instruction is in the 'main' function, and in the outer
most-block.
The first two restrictions were arrived to empirically, so more
testing could tighten or loosen it.
The 3rd restriction is there to allow moving the instructions
eligible for pre-dispatch to the beginning of the shader, so
that we don't block the registers holding the result for too
long.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
It seems that pre-fs texture fetch only works if ij_pix ends up in r0.x.
I've tried unknown zero bits, to no avail, and blob also seems to force
r0.x when this feature is used.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Useful to see in disassembly listing texture fetches that were moved to
pre-dispatch.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
If the only use of varyings is a pre-shader texture-fetch, we still need
to issue a bary.f with the end-input flag, otherwise we'll block further
VS invocations, as the hw will think varying storage is still busy.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
It is possible that the result of a pre-fs texture fetch is an output
(or partially an output) of the FS. Sine the meta:tex_prefetch
instructions are dropped before the assembler, we need to account for
this when we fixup the register footprint.
Signed-off-by: Rob Clark <robdclark@chromium.org>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
