AlphaGitHub
Source/Packages

Accelerator.SM86.Operands

packages/hardware/architectures/nvidia-sm86/src/Accelerator/SM86/Operands.alpha

699 lines54 declarations58.4 KiBSHA-256 ac21d61cced1
Download exact fileView at revisionbcc11a4e7647d31628

def · lines 648–699

sm86OpSummaryOfBody

Full file
648def sm86OpSummaryOfBody =
649  (lambda unrestricted body : (family SM86InstructionBody) .
650    (eliminate SM86InstructionBody
651      (lambda unrestricted current : (family SM86InstructionBody) . (family SM86OpSummary))
652      body
653      (branch SM86MoveConstant destination bank offset control . (sm86OpSummaryMake sm86NoRegisters (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
654      (branch SM86SpecialToRegister destination source control . (sm86OpSummaryMake sm86NoRegisters (sm86One destination sm86NoRegisters) sm86NoRegisters 1 0 1 control))
655      (branch SM86MoveImmediate destination immediate control . (sm86OpSummaryMake sm86NoRegisters (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
656      (branch SM86IntegerMultiplyAddConstant destination left bank offset addend control . (sm86OpSummaryMake (sm86One left (sm86One addend sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
657      (branch SM86IntegerMultiplyAddImmediate destination left immediate addend control . (sm86OpSummaryMake (sm86One left (sm86One addend sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
658      (branch SM86IntegerMultiplyAddWideConstant destination left right bank offset control . (sm86OpSummaryMake (sm86One left (sm86Pair right sm86NoRegisters)) (sm86Pair destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
659      (branch SM86IntegerAddThreeImmediate destination left immediate control . (sm86OpSummaryMake (sm86One left sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
660      (branch SM86IntegerAddThreeRegister destination left right control . (sm86OpSummaryMake (sm86One left (sm86One right sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
661      (branch SM86ShiftRightImmediate destination source amount control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
662      (branch SM86LogicThreeInputTruthTable destination left right truthTable control . (sm86OpSummaryMake (sm86One left (sm86One right sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
663      (branch SM86IntegerToFloat destination source control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
664      (branch SM86FloatAdd destination left right control . (sm86OpSummaryMake (sm86One left (sm86One right sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
665      (branch SM86FloatMultiply destination left right control . (sm86OpSummaryMake (sm86One left (sm86One right sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
666      -- Double arithmetic has variable latency, including late source reads.
667      -- Mesa NAK ir.rs, has_fixed_latency, revision b1fe47e94409d81924c8663f58fac97e66ab4527:
668      -- https://chromium.googlesource.com/external/gitlab.freedesktop.org/mesa/mesa/+/b1fe47e94409d81924c8663f58fac97e66ab4527/src/nouveau/compiler/nak/ir.rs
669      -- Both halves participate; waiting on the low word alone is insufficient.
670      (branch SM86Float64Binary operation destination left right control . (sm86OpSummaryMake (sm86Pair (sm86RegisterPairBase left) (sm86Pair (sm86RegisterPairBase right) sm86NoRegisters)) (sm86Pair (sm86RegisterPairBase destination) sm86NoRegisters) sm86NoRegisters 1 0 1 control))
671      (branch SM86FloatConvert conversion control . (sm86OpSummaryMake (sm86ConversionReads conversion) (sm86ConversionWrites conversion) sm86NoRegisters 1 0 1 control))
672      (branch SM86FloatFusedMultiplyAdd destination left right addend control . (sm86OpSummaryMake (sm86One left (sm86One right (sm86One addend sm86NoRegisters))) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
673      (branch SM86MultiFunctionUnitApproximation destination source operation control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 1 0 1 control))
674      (branch SM86FloatMinimumOrMaximum destination left right mode control . (sm86OpSummaryMake (sm86One left (sm86One right sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
675      (branch SM86FloatNegate destination source control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
676      (branch SM86FloatPairToPackedHalfPair destination high low control . (sm86OpSummaryMake (sm86One high (sm86One low sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
677      (branch SM86FloatPairToPackedBFloat16Pair destination high low control . (sm86OpSummaryMake (sm86One high (sm86One low sm86NoRegisters)) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
678      (branch SM86HalfToFloat destination source selector control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
679      (branch SM86BFloat16ToFloat destination source selector control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 0 sm86FixedLatencyCycles 1 control))
680      (branch SM86TensorCoreHalfMatrixMultiplyAccumulate16x8x16Float32 destination a b accumulator control . (sm86OpSummaryMake (sm86Quad a (sm86Pair b (sm86Quad accumulator sm86NoRegisters))) (sm86Quad destination sm86NoRegisters) sm86NoRegisters 0 (naturalSelect (sm86ControlWritesBarrier control) 0 sm86TensorLatencyCycles) (naturalSelect (sm86ControlWritesBarrier control) 1 sm86TensorIssueStall) control))
681      (branch SM86TensorCoreBFloat16MatrixMultiplyAccumulate16x8x16Float32 destination a b accumulator control . (sm86OpSummaryMake (sm86Quad a (sm86Pair b (sm86Quad accumulator sm86NoRegisters))) (sm86Quad destination sm86NoRegisters) sm86NoRegisters 0 (naturalSelect (sm86ControlWritesBarrier control) 0 sm86TensorLatencyCycles) (naturalSelect (sm86ControlWritesBarrier control) 1 sm86TensorIssueStall) control))
682      (branch SM86LoadGlobal destination address offset control . (sm86OpSummaryMake (sm86Pair address sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 1 0 1 control))
683      (branch SM86LoadGlobalWide destination address offset control . (sm86OpSummaryMake (sm86Pair address sm86NoRegisters) (sm86Quad destination sm86NoRegisters) sm86NoRegisters 1 0 1 control))
684      (branch SM86WarpShuffle destination source lane segment mode control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 1 0 1 control))
685      (branch SM86LoadShared destination address offset control . (sm86OpSummaryMake (sm86One address sm86NoRegisters) (sm86One destination sm86NoRegisters) sm86NoRegisters 1 0 1 control))
686      (branch SM86LoadSharedMatrix destination address offset count transpose control . (sm86OpSummaryMake (sm86One address sm86NoRegisters) (sm86RegisterRun destination (sm86SharedMatrixRegisters count) sm86NoRegisters) sm86NoRegisters 1 0 1 control))
687      (branch SM86StoreShared address value offset control . (sm86OpSummaryMake (sm86One address (sm86One value sm86NoRegisters)) sm86NoRegisters sm86NoRegisters 0 0 1 control))
688      (branch SM86LoadGlobalToShared address offset source sourceOffset control . (sm86OpSummaryMake (sm86One address (sm86Pair source sm86NoRegisters)) sm86NoRegisters sm86NoRegisters 0 0 1 control))
689      (branch SM86CommitAsyncGroup control . (sm86OpSummaryMake sm86NoRegisters sm86NoRegisters sm86NoRegisters 0 0 1 control))
690      (branch SM86WaitAsyncGroups count control . (sm86OpSummaryMake sm86NoRegisters sm86NoRegisters sm86NoRegisters 0 0 sm86WaitAsyncGroupsStall control))
691      (branch SM86BarrierSynchronize control . (sm86OpSummaryMake sm86NoRegisters sm86NoRegisters sm86NoRegisters 0 0 sm86BarrierSynchronizeStall control))
692      (branch SM86BarrierSynchronizeCount barrier threads control . (sm86OpSummaryMake sm86NoRegisters sm86NoRegisters sm86NoRegisters 0 0 sm86BarrierSynchronizeStall control))
693      (branch SM86StoreGlobal address value offset control . (sm86OpSummaryMake (sm86Pair address (sm86One value sm86NoRegisters)) sm86NoRegisters sm86NoRegisters 0 0 1 control))
694      (branch SM86StoreGlobalWide address value offset control . (sm86OpSummaryMake (sm86Pair address (sm86Quad value sm86NoRegisters)) sm86NoRegisters sm86NoRegisters 0 0 1 control))
695      (branch SM86StoreGlobal64 address value offset control . (sm86OpSummaryMake (sm86Pair address (sm86Pair value sm86NoRegisters)) sm86NoRegisters sm86NoRegisters 0 0 1 control))
696      (branch SM86ReduceGlobalAddFloat32 address value offset control . (sm86OpSummaryMake (sm86Pair address (sm86One value sm86NoRegisters)) sm86NoRegisters sm86NoRegisters 0 0 1 control))
697      (branch SM86PredicateGreaterThanImmediate predicate source immediate control . (sm86OpSummaryMake (sm86One source sm86NoRegisters) sm86NoRegisters (sm86PredicateKeys predicate) 0 sm86FixedLatencyCycles 1 control))
698      (branch SM86Branch offset descriptor control . (sm86OpSummaryMake sm86NoRegisters sm86NoRegisters sm86NoRegisters 0 0 1 control))
699      (branch SM86Exit control . (sm86OpSummaryMake sm86NoRegisters sm86NoRegisters sm86NoRegisters 0 0 1 control))))

The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.