WMMA XDL instructions are tracked as TRANs ops and the compiler should
consider them the same as TRANS in S_DELAY_ALU insertion. We use a searchable
table for the InsertDelayAlu pass to recognize these WMMA XDL instructions.
Co-authored-by: Stefan Stipanovic <Stefan.Stipanovic@amd.com>
This patch tracks the register operands of both VMEM (FLAT, MUBUF,
MTBUF) and SMEM load-store operations and inserts a S_WAIT_XCNT
instruction with sufficient wait-count before potentially redefining
them. For VMEM instructions, XNACK is returned in the same order as
they were issued and hence non-zero counter values can be inserted.
However, SMEM execution is out-of-order and so is their XNACK reception.
Thus, only zero counter value can be inserted to capture SMEM dependencies.
Move canGuaranteeTCO and mayTailCallThisCC into AMDGPUBaseInfo instead
of keeping two copies for DAG/Global ISel.
Also remove isKernelCC, which doesn't agree with isKernel and doesn't
seem very useful.
While at it, also move all the CC-related helpers into AMDGPUBaseInfo.h and
mark them constexpr.
This reapplies 067caaa and 382a085 (reverting b35f6e2) with fixes to
issues detected by the address sanitizer (MIs have to be removed from
live intervals before being removed from their parent MBB).
Original commit description below.
AMDGPU scheduler's `PreRARematStage` attempts to increase function
occupancy w.r.t. ArchVGPR usage by rematerializing trivial
ArchVGPR-defining instruction next to their single use. It first
collects all eligible trivially rematerializable instructions in the
function, then sinks them one-by-one while recomputing occupancy in all
affected regions each time to determine if and when it has managed to
increase overall occupancy. If it does, changes are committed to the
scheduler's state; otherwise modifications to the IR are reverted and
the scheduling stage gives up.
In both cases, this scheduling stage currently involves repeated queries
for up-to-date occupancy estimates and some state copying to enable
reversal of sinking decisions when occupancy is revealed not to
increase. The current implementation also does not accurately track
register pressure changes in all regions affected by sinking decisions.
This commit refactors this scheduling stage, improving RP tracking and
splitting the stage into two distinct steps to avoid repeated occupancy
queries and IR/state rollbacks.
- Analysis and collection (`canIncreaseOccupancyOrReduceSpill`). The
number of ArchVGPRs to save to reduce spilling or increase function
occupancy by 1 (when there is no spilling) is computed. Then,
instructions eligible for rematerialization are collected, stopping as
soon as enough have been identified to be able to achieve our goal
(according to slightly optimistic heuristics). If there aren't enough of
such instructions, the scheduling stage stops here.
- Rematerialization (`rematerialize`). Instructions collected in the
first step are rematerialized one-by-one. Now we are able to directly
update the scheduler's state since we have already done the occupancy
analysis and know we won't have to rollback any state. Register
pressures for impacted regions are recomputed only once, as opposed to
at every sinking decision.
In the case where the stage attempted to increase occupancy, and if both
rematerializations alone and rescheduling after were unable to improve
occupancy, then all rematerializations are rollbacked.
And related "[AMDGPU] Regenerate mfma-loop.ll test"
Introduce memory error detected by Asan #125885.
This reverts commit 382a085a95b0abeac77b150b7b644b372bd08e78.
This reverts commit 067caaafb58a156d0d77229422607782a639f5b5.
AMDGPU scheduler's `PreRARematStage` attempts to increase function
occupancy w.r.t. ArchVGPR usage by rematerializing trivial
ArchVGPR-defining instruction next to their single use. It first
collects all eligible trivially rematerializable instructions in the
function, then sinks them one-by-one while recomputing occupancy in all
affected regions each time to determine if and when it has managed to
increase overall occupancy. If it does, changes are committed to the
scheduler's state; otherwise modifications to the IR are reverted and
the scheduling stage gives up.
In both cases, this scheduling stage currently involves repeated queries
for up-to-date occupancy estimates and some state copying to enable
reversal of sinking decisions when occupancy is revealed not to
increase. The current implementation also does not accurately track
register pressure changes in all regions affected by sinking decisions.
This commit refactors this scheduling stage, improving RP tracking and
splitting the stage into two distinct steps to avoid repeated occupancy
queries and IR/state rollbacks.
- Analysis and collection (`canIncreaseOccupancyOrReduceSpill`). The
number of ArchVGPRs to save to reduce spilling or increase function
occupancy by 1 (when there is no spilling) is computed. Then,
instructions eligible for rematerialization are collected, stopping as
soon as enough have been identified to be able to achieve our goal
(according to slightly optimistic heuristics). If there aren't enough of
such instructions, the scheduling stage stops here.
- Rematerialization (`rematerialize`). Instructions collected in the
first step are rematerialized one-by-one. Now we are able to directly
update the scheduler's state since we have already done the occupancy
analysis and know we won't have to rollback any state. Register
pressures for impacted regions are recomputed only once, as opposed to
at every sinking decision.
In the case where the stage attempted to increase occupancy, and if both
rematerializations alone and rescheduling after were unable to improve
occupancy, then all rematerializations are rollbacked.
In dynamic VGPR mode, we can allocate up to 8 blocks of either 16 or 32
VGPRs (based on a chip-wide setting which we can model with a Subtarget
feature). Update some of the subtarget helpers to reflect this.
In particular:
- getVGPRAllocGranule is set to the block size
- getAddresableNumVGPR will limit itself to 8 * size of a block
We also try to be more careful about how many VGPR blocks we allocate.
Therefore, when deciding if we should revert scheduling after a given
stage, we check that we haven't increased the number of VGPR blocks that
need to be allocated.
---------
Co-authored-by: Jannik Silvanus <jannik.silvanus@amd.com>
From GFX10 onwards it is possible to employ benevolent scheduling of
waves. This patch unconditionally enables, for the `amdhsa` OS, the bit
which controls that capability, as it is beneficial for algorithms that
rely on more complex concurrent coordination and it is generally
performance neutral otherwise.
Enable gisel selection for uaddsat and usubsat in true16 flow
This patch includes:
1. Added VGPR_16_Lo128/VGPR_16 to register bank and update register info
for recognizing 16bit regclass id and bit width
2. uaddsat/usubsat test update
Previous patch is merged
https://github.com/llvm/llvm-project/pull/114500 and it hit a buildbot
failure and thus reverted
It seems the AMDGPU::OpName::OPERAND_LAST is removed at the meantime
when previous patch is merged and that's causing the compile error.
Fixed and reopen it here
This reverts commit f7a5f067885b7f6cc4a000c8392adf6b777a9108.
Fails to build with:
llvm/lib/Target/AMDGPU/AMDGPUMCInstLower.cpp:126:37: error: no member named 'OPERAND_LAST' in 'llvm::AMDGPU::OpName'
126 | uint16_t OpName = AMDGPU::OpName::OPERAND_LAST;
Implement new pseudos with the suffix _t16 for FLAT_LOAD which have
VGPR_16 as the load dst. Lower the pseudos to the existing real
instructions with VGPR_32 src or dst (which makes them consistent with
the hardware encoding). This patch reduces VGPR usage by making hi
halves of VGPRs available for other values.
There are more 8/16 bits ld/st instructions to be supported in the
up-coming patches
If a function has `amdgpu-flat-work-group-size`, honor it in `initialize` by
taking its value directly; otherwise, it uses the default range as a starting
point. We will no longer manipulate the known range, which can cause issues
because the known range is a "throttle" to the assumed range such that the
assumed range can't get widened properly in `updateImpl` if the known range is
not set properly for whatever reasons. Another benefit of not touching the known
range is, if we indicate pessimistic state, it also invalidates the AA such that
`manifest` will not be called. Since we honor the attribute, we don't want and
will not add any half-baked attribute added to a function.
Presently, compiler selectivelly adds nop when opsel != 0 i.e. only when
partially writing to high bytes.
Experiments in SWDEV-499733 and SWDEV-501347 suggest that we need nop
for above cases irrespective of opsel values.
Note: We might need to add few others into the same table.
gfx950 SP changes doc says:
No 4 clk forwarding on opcodes that convert from
F32/F16->F8 or F32/F16->F4. Must insert a NOP or
instruction writing some other destination VREG
after a conversion to F4/F8 since it writes either
low/high half or bytes.
Co-authored-by: Pravin Jagtap <Pravin.Jagtap@amd.com>
Co-authored-by: Jeffrey Byrnes <Jeffrey.Byrnes@amd.com>
The encoding of v_dot2c_f32_bf16 opcode is same as v_mac_f32 in gfx90a,
both from gfx9 series. This required a new decoderNameSpace GFX950_DOT.
Co-authored-by: Sirish Pande <Sirish.Pande@amd.com>
These use a new VOP3PX encoding for the v_mfma_scale_* instructions,
which bundles the pre-scale v_mfma_ld_scale_b32. None of the modifiers
are supported yet (op_sel, neg or clamp).
I'm not sure the intrinsic should really expose op_sel (or any of the
others). If I'm reading the documentation correctly, we should be able
to just have the raw scale operands and auto-match op_sel to byte
extract patterns.
The op_sel syntax also seems extra horrible in this usage, especially with the
usual assumed op_sel_hi=-1 behavior.
Existing Fake16 versions of these instructions do not support op_sel on
the _e32 encoding, which leaves a hole in the disassembler support.
Implement the true16 version of the instructions in the MC layer.
0 does not make sense as a value for this to be, much less the default.
Also stop emitting each individual field if it is the default, rather than
if any element was the default. Also fix the name of the test since it didn't
exactly match the real attribute name.
Change the names of the TableGen features to match the names used by
AMDGPUSubtarget. "Addressable" refers to the amount that can be accessed
by a single workgroup. Add some explanatory comments. NFC.
