Occupancy (i.e., the number of waves per EU) depends, in addition to
register usage, on per-workgroup LDS usage as well as on the range of
possible workgroup sizes. Mirroring the latter, occupancy should
therefore be expressed as a range since different group sizes generally
yield different achievable occupancies.
`getOccupancyWithLocalMemSize` currently returns a scalar occupancy
based on the maximum workgroup size and LDS usage. With respect to the
workgroup size range, this scalar can be the minimum, the maximum, or
neither of the two of the range of achievable occupancies. This commit
fixes the function by making it compute and return the range of
achievable occupancies w.r.t. workgroup size and LDS usage; it also
renames it to `getOccupancyWithWorkGroupSizes` since it is the range of
workgroup sizes that produces the range of achievable occupancies.
Computing the achievable occupancy range is surprisingly involved.
Minimum/maximum workgroup sizes do not necessarily yield maximum/minimum
occupancies i.e., sometimes workgroup sizes inside the range yield the
occupancy bounds. The implementation finds these sizes in constant time;
heavy documentation explains the rationale behind the sometimes
relatively obscure calculations.
As a justifying example, consider a target with 10 waves / EU, 4 EUs/CU,
64-wide waves. Also consider a function with no LDS usage and a flat
workgroup size range of [513,1024].
- A group of 513 items requires 9 waves per group. Only 4 groups made up
of 9 waves each can fit fully on a CU at any given time, for a total of
36 waves on the CU, or 9 per EU. However, filling as much as possible
the remaining 40-36=4 wave slots without decreasing the number of groups
reveals that a larger group of 640 items yields 40 waves on the CU, or
10 per EU.
- Similarly, a group of 1024 items requires 16 waves per group. Only 2
groups made up of 16 waves each can fit fully on a CU ay any given time,
for a total of 32 waves on the CU, or 8 per EU. However, removing as
many waves as possible from the groups without being able to fit another
equal-sized group on the CU reveals that a smaller group of 896 items
yields 28 waves on the CU, or 7 per EU.
Therefore the achievable occupancy range for this function is not [8,9]
as the group size bounds directly yield, but [7,10].
Naturally this change causes a lot of test churn as instruction
scheduling is driven by achievable occupancy estimates. In most unit
tests the flat workgroup size range is the default [1,1024] which,
ignoring potential LDS limitations, would previously produce a scalar
occupancy of 8 (derived from 1024) on a lot of targets, whereas we now
consider the maximum occupancy to be 10 in such cases. Most tests are
updated automatically and checked manually for sanity. I also manually
changed some non-automatically generated assertions when necessary.
Fixes#118220.
MSG_DEALLOC_VGPRS slows down very small waveslot limited kernels. It's
been identified this message is only really needed for VGPR limited
kernels. A kernel becomes VGPR limited if a total number of VGPRs per
SIMD / number of used VGPRs is more than a number of wave slots.
This reverts commit adaff46d087799072438dd744b038e6fd50a2d78.
Drop the -O3 checks from default-attributes.hip. I don't know why they
are different on some bots but reverting this is far too disruptive.
Removing it from the codegen pipeline induces a lot of test churn
because llc is no longer optimizing out implicit arguments to kernels.
Mostly mechanical, but there are some creative test updates. I preferred
to take the changes as-is in tests where the ABI isn't relevant. In
cases where it's more relevant, or the optimize out logic was too
ingrained in the test, I pre-run the optimization. Some cases manually
add attributes to disable inputs.
Reland the original patch with additional commit containing fix for two
issues:
1. Attempting to bitcast using MVTs with no corresponding LLVM type.
getDWordFromOffset now works directly with the original vector to get
the corresponding elements given the DWordOffset.
2. Improper bit tracking in CalculateByteProvider for vector types using
certain ops. Previously, bit tracking for certain ops (e.g.
ISD::TRUNCATE) assumed operands were scalar types, which is not correct
since these ops have different semantics depending on vector / scalar.
CalculateByteProvider / CalculateSrcByte now exit on vector types,
handling which is a TODO.
This allows working with e.g. v8i8 / v16i8 sources.
It is generally useful, but is primarily beneficial when allowing e.g. v8i8s to be passed to branches directly through registers. As such, this is the first in a series of patches to enable that work. However, it effects https://reviews.llvm.org/D155995, so it has been implemented on top of that.
Differential Revision: https://reviews.llvm.org/D159036
Change-Id: Idfcb57dacd0c32cab040fe4dd4ac2ec762750664
PR #66334 tried to renumber slot indexes before register allocation, but
the numbering was still affected by list entries for instructions which
had been erased. Fix this to make the register allocator's live range
length heuristics even less dependent on the history of how instructions
have been added to and removed from SlotIndexes's maps.
RegAllocGreedy uses SlotIndexes::getApproxInstrDistance to approximate
the length of a live range for its heuristics. Renumbering all slot
indexes with the default instruction distance ensures that this estimate
will be as accurate as possible, and will not depend on the history of
how instructions have been added to and removed from SlotIndexes's maps.
This also means that enabling -early-live-intervals, which runs the
SlotIndexes analysis earlier, will not cause large amounts of churn due
to different register allocator decisions.
Documentation for TargetLowering::getShiftAmountTy says that LegalTypes
should generally be true during type legalization, so this patch does
that.
On AMDGPU the effect is that we use i32 (a sane type) instead of i64
(pointer sized type) for more shift amounts, which in turn allows more
formation of rotates and funnel shifts pre-legalization.
Differential Revision: https://reviews.llvm.org/D154960
Occupancy is expressed as waves per SIMD. This means that we need to
take into account the number of SIMDs per "CU" or, to be more precise,
the number of SIMDs over which a workgroup may be distributed.
getOccupancyWithLocalMemSize was wrong because it didn't take SIMDs
into account at all.
At the same time, we need to take into account that WGP mode offers
access to a larger total amount of LDS, since this can affect how
non-power-of-two LDS allocations are rounded. To make this work
consistently, we distinguish between (available) local memory size and
addressable local memory size (which is always limited by 64kB on
gfx10+, even with WGP mode).
This change results in a massive amount of test churn. A lot of it is
caused by the fact that the default work group size is 1024, which means
that (due to rounding effects) the default occupancy on older hardware
is 8 instead of 10, which affects scheduling via register pressure
estimates. I've adjusted most tests by just running the UTC tools, but
in some cases I manually changed the work group size to 32 or 64 to make
sure that work group size chunkiness has no effect.
Differential Revision: https://reviews.llvm.org/D139468
If we can not prove that f16 operands of a buildvector are canonicalized, then we can not lower into a V_PACK. In this scenario, we would previously lower into some combination of and(sdwa), shr, or. This patch allows for matching into V_PERM instead.
Change-Id: Ifa4a74fdb81ef44f22ba490c7fdf81ec8aebc945
Add GFX11 test coverage to a bunch of tests where it was easy to do so,
mostly because the checks are autogenerated and/or GFX11 can share the
same checks as GFX10.
Differential Revision: https://reviews.llvm.org/D129295
Fold immediates regardless of how many uses they have. This is expected
to increase overall code size, but decrease register usage.
Differential Revision: https://reviews.llvm.org/D114644
Previously SIFoldOperands::foldInstOperand would only fold a
non-inlinable immediate into a single user, so as not to increase code
size by adding the same 32-bit literal operand to many instructions.
This patch removes that restriction, so that a non-inlinable immediate
will be folded into any number of users. The rationale is:
- It reduces the number of registers used for holding constant values,
which might increase occupancy. (On the other hand, many of these
registers are SGPRs which no longer affect occupancy on GFX10+.)
- It reduces ALU stalls between the instruction that loads a constant
into a register, and the instruction that uses it.
- The above benefits are expected to outweigh any increase in code size.
Differential Revision: https://reviews.llvm.org/D114643
If we're using shift pairs to mask, then relax the one use limit if the shift amounts are equal - we'll only be generating a single AND node.
AArch64 has a couple of regressions due to this, so I've enforced the existing one use limit inside a AArch64TargetLowering::shouldFoldConstantShiftPairToMask callback.
Part of the work to fix the regressions in D77804
Differential Revision: https://reviews.llvm.org/D125607
Pulled out of D77804 as its going to be easier to address the regressions individually.
This patch allows SimplifyDemandedBits to call SimplifyMultipleUseDemandedBits in cases where the source operand has other uses, enabling us to peek through the shifted value if we don't demand all the bits/elts.
The lost RISCV gorc2 fold shouldn't be a problem - instcombine would have already destroyed that pattern - see https://github.com/llvm/llvm-project/issues/50553
Differential Revision: https://reviews.llvm.org/D124839
As suggested by @foad on D124839
If we're extracting a vector element that originally came from a scalar_to_vector, then avoid the bitcasting of a vector type and perform the shift masking on the (any-extended) scalar source directly, making use of the fact that the upper elements of a scalar_to_vector are all undef.
Differential Revision: https://reviews.llvm.org/D125173
Using a BufferSize of one for memory ProcResources will result in better
ILP since it more accurately models the dependencies between memory ops
and their consumers on an in-order processor. After this change, the
scheduler will treat the data edges from loads as blocking so that
stalls are guaranteed when waiting for data to be retreaved from memory.
Since we don't actually track waitcnt here, this should do a better job
at modeling their behavior.
Practically, this means that the scheduler will trigger the 'STALL'
heuristic more often.
This type of change needs to be evaluated experimentally. Preliminary
results are positive.
Fixes: SWDEV-282962
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D114777
Change VOP_PAT_GEN to default to not generating an instruction selection
pattern for the VOP2 (e32) form of an instruction, only for the VOP3
(e64) form. This allows SIFoldOperands maximum freedom to fold copies
into the operands of an instruction, before SIShrinkInstructions tries
to shrink it back to the smaller encoding.
This affects the following VOP2 instructions:
v_min_i32
v_max_i32
v_min_u32
v_max_u32
v_and_b32
v_or_b32
v_xor_b32
v_lshr_b32
v_ashr_i32
v_lshl_b32
A further cleanup could simplify or remove VOP_PAT_GEN, since its
optional second argument is never used.
Differential Revision: https://reviews.llvm.org/D114252
Use GCNHazardRecognizer in postra sched.
Updated tests for the new schedules.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D109536
Change-Id: Ia86ba2ae168f12fb34b4d8efdab491f84d936cde
This is to allow 64 bit constant rematerialization. If a constant
is split into two separate moves initializing sub0 and sub1 like
now RA cannot rematerizalize a 64 bit register.
This gives 10-20% uplift in a set of huge apps heavily using double
precession math.
Fixes: SWDEV-292645
Differential Revision: https://reviews.llvm.org/D104874
Add SReg_224, VReg_224, AReg_224, etc.
Link 224-bit types with v7i32/v7f32.
Link existing 192-bit types to newly added v3i64/v3f64/v6i32/v6f32.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D104622
Support for XNACK and SRAMECC is not static on some GPUs. We must be able
to differentiate between different scenarios for these dynamic subtarget
features.
The possible settings are:
- Unsupported: The GPU has no support for XNACK/SRAMECC.
- Any: Preference is unspecified. Use conservative settings that can run anywhere.
- Off: Request support for XNACK/SRAMECC Off
- On: Request support for XNACK/SRAMECC On
GCNSubtarget will track the four options based on the following criteria. If
the subtarget does not support XNACK/SRAMECC we say the setting is
"Unsupported". If no subtarget features for XNACK/SRAMECC are requested we
must support "Any" mode. If the subtarget features XNACK/SRAMECC exist in the
feature string when initializing the subtarget, the settings are "On/Off".
The defaults are updated to be conservatively correct, meaning if no setting
for XNACK or SRAMECC is explicitly requested, defaults will be used which
generate code that can be run anywhere. This corresponds to the "Any" setting.
Differential Revision: https://reviews.llvm.org/D85882
Treat a non-atomic volatile load and store as a relaxed atomic at
system scope for the address spaces accessed. This will ensure all
relevant caches will be bypassed.
A volatile atomic is not changed and still only bypasses caches upto
the level specified by the SyncScope operand.
Differential Revision: https://reviews.llvm.org/D94214
- Add an internal option `-amdgpu-use-aa-in-codegen` to enable or
disable this feature. By Default, it's enabled.
Differential Revision: https://reviews.llvm.org/D89320
Change waitcnt insertion to check the memory operand tokens to see if
flat memory operations access VMEM in the same way it does to check if
accessing LDS. This avoids adding waitcnt for counters for address
spaces that are not accessed.
In addition, only generate the pessimistic waitcnt 0 if a flat memory
operation appears to access both VMEM and LDS.
This benefits flat memory operations that explicitly specify the
address space as GLOBAL or LOCAL.
Differential Revision: https://reviews.llvm.org/D89618
This tends to increase code size but more importantly it reduces vgpr
usage, and could avoid costly readfirstlanes if the result needs to be
in an sgpr.
Differential Revision: https://reviews.llvm.org/D88245
The previous implementation was incorrect, and based off incorrect
instruction definitions. Unfortunately we can't match natural
addressing in a lot of cases due to the shift/scale applied in
getelementptrs. This relies on reducing the 64-bit shift to 32-bits.
tryLatency compares two sched candidates. For the top zone it prefers
the one with lesser depth, but only if that depth is greater than the
total latency of the instructions we've already scheduled -- otherwise
its latency would be hidden and there would be no stall.
Unfortunately it only tests the depth of one of the candidates. This can
lead to situations where the TopDepthReduce heuristic does not kick in,
but a lower priority heuristic chooses the other candidate, whose depth
*is* greater than the already scheduled latency, which causes a stall.
The fix is to apply the heuristic if the depth of *either* candidate is
greater than the already scheduled latency.
All this also applies to the BotHeightReduce heuristic in the bottom
zone.
Differential Revision: https://reviews.llvm.org/D72392
This patch allows ISD::FSHR(i32) patterns to lower to ALIGNBIT instructions.
This improves test coverage of ISD::FSHR matching - x86 has both FSHL/FSHR instructions and we prefer FSHL by default.
Differential Revision: https://reviews.llvm.org/D76070
Summary:
pickNodeBidirectional tried to compare the best top candidate and the
best bottom candidate by examining TopCand.Reason and BotCand.Reason.
This is unsound because, after calling pickNodeFromQueue, Cand.Reason
does not reflect the most important reason why Cand was chosen. Rather
it reflects the most recent reason why it beat some other potential
candidate, which could have been for some low priority tie breaker
reason.
I have seen this cause problems where TopCand is a good candidate, but
because TopCand.Reason is ORDER (which is very low priority) it is
repeatedly ignored in favour of a mediocre BotCand. This is not how
bidirectional scheduling is supposed to work.
To fix this I changed the code to always compare TopCand and BotCand
directly, like the generic implementation of pickNodeBidirectional does.
This removes some uncommented AMDGPU-specific logic; if this logic turns
out to be important then perhaps it could be moved into an override of
tryCandidate instead.
Graphics shader benchmarking on gfx10 shows a lot more positive than
negative effects from this change.
Reviewers: arsenm, tstellar, rampitec, kzhuravl, vpykhtin, dstuttard, tpr, atrick, MatzeB
Subscribers: jvesely, wdng, nhaehnle, yaxunl, t-tye, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68338
