NVVMToLLVMIRTranslation.cpp

Go to the documentation of this file.

//===- NVVMToLLVMIRTranslation.cpp - Translate NVVM to LLVM IR ------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements a translation between the MLIR NVVM dialect and
// LLVM IR.
//
//===----------------------------------------------------------------------===//
 
#include "mlir/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.h"
#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
#include "mlir/IR/Operation.h"
#include "mlir/Target/LLVMIR/ModuleTranslation.h"
 
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/iterator_range.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/IntrinsicsNVPTX.h"
#include "llvm/Support/FormatVariadic.h"
#include "llvm/Support/NVVMAttributes.h"
#include <cmath>
 
using namespace mlir;
using namespace mlir::LLVM;
using mlir::LLVM::detail::createIntrinsicCall;
 
#define REDUX_F32_ID_IMPL(op, abs, hasNaN)                                     \
  hasNaN ? llvm::Intrinsic::nvvm_redux_sync_f##op##abs##_NaN                   \
         : llvm::Intrinsic::nvvm_redux_sync_f##op##abs
 
#define GET_REDUX_F32_ID(op, hasAbs, hasNaN)                                   \
  hasAbs ? REDUX_F32_ID_IMPL(op, _abs, hasNaN) : REDUX_F32_ID_IMPL(op, , hasNaN)
 
static llvm::Intrinsic::ID getReduxIntrinsicId(llvm::Type *resultType,
                                               NVVM::ReductionKind kind,
                                               bool hasAbs, bool hasNaN) {
  switch (kind) {
  case NVVM::ReductionKind::ADD:
    return llvm::Intrinsic::nvvm_redux_sync_add;
  case NVVM::ReductionKind::UMAX:
    return llvm::Intrinsic::nvvm_redux_sync_umax;
  case NVVM::ReductionKind::UMIN:
    return llvm::Intrinsic::nvvm_redux_sync_umin;
  case NVVM::ReductionKind::AND:
    return llvm::Intrinsic::nvvm_redux_sync_and;
  case NVVM::ReductionKind::OR:
    return llvm::Intrinsic::nvvm_redux_sync_or;
  case NVVM::ReductionKind::XOR:
    return llvm::Intrinsic::nvvm_redux_sync_xor;
  case NVVM::ReductionKind::MAX:
    return llvm::Intrinsic::nvvm_redux_sync_max;
  case NVVM::ReductionKind::MIN:
    return llvm::Intrinsic::nvvm_redux_sync_min;
  case NVVM::ReductionKind::FMIN:
    return GET_REDUX_F32_ID(min, hasAbs, hasNaN);
  case NVVM::ReductionKind::FMAX:
    return GET_REDUX_F32_ID(max, hasAbs, hasNaN);
  }
  llvm_unreachable("unknown reduction kind");
}
 
static llvm::Intrinsic::ID getShflIntrinsicId(llvm::Type *resultType,
                                              NVVM::ShflKind kind,
                                              bool withPredicate) {
 
  if (withPredicate) {
    resultType = cast<llvm::StructType>(resultType)->getElementType(0);
    switch (kind) {
    case NVVM::ShflKind::bfly:
      return resultType->isFloatTy()
                 ? llvm::Intrinsic::nvvm_shfl_sync_bfly_f32p
                 : llvm::Intrinsic::nvvm_shfl_sync_bfly_i32p;
    case NVVM::ShflKind::up:
      return resultType->isFloatTy() ? llvm::Intrinsic::nvvm_shfl_sync_up_f32p
                                     : llvm::Intrinsic::nvvm_shfl_sync_up_i32p;
    case NVVM::ShflKind::down:
      return resultType->isFloatTy()
                 ? llvm::Intrinsic::nvvm_shfl_sync_down_f32p
                 : llvm::Intrinsic::nvvm_shfl_sync_down_i32p;
    case NVVM::ShflKind::idx:
      return resultType->isFloatTy() ? llvm::Intrinsic::nvvm_shfl_sync_idx_f32p
                                     : llvm::Intrinsic::nvvm_shfl_sync_idx_i32p;
    }
  } else {
    switch (kind) {
    case NVVM::ShflKind::bfly:
      return resultType->isFloatTy() ? llvm::Intrinsic::nvvm_shfl_sync_bfly_f32
                                     : llvm::Intrinsic::nvvm_shfl_sync_bfly_i32;
    case NVVM::ShflKind::up:
      return resultType->isFloatTy() ? llvm::Intrinsic::nvvm_shfl_sync_up_f32
                                     : llvm::Intrinsic::nvvm_shfl_sync_up_i32;
    case NVVM::ShflKind::down:
      return resultType->isFloatTy() ? llvm::Intrinsic::nvvm_shfl_sync_down_f32
                                     : llvm::Intrinsic::nvvm_shfl_sync_down_i32;
    case NVVM::ShflKind::idx:
      return resultType->isFloatTy() ? llvm::Intrinsic::nvvm_shfl_sync_idx_f32
                                     : llvm::Intrinsic::nvvm_shfl_sync_idx_i32;
    }
  }
  llvm_unreachable("unknown shuffle kind");
}
 
static llvm::Intrinsic::ID getMatchSyncIntrinsicId(Type valType,
                                                   NVVM::MatchSyncKind kind) {
  switch (kind) {
  case NVVM::MatchSyncKind::any:
    return valType.isInteger(32) ? llvm::Intrinsic::nvvm_match_any_sync_i32
                                 : llvm::Intrinsic::nvvm_match_any_sync_i64;
  case NVVM::MatchSyncKind::all:
    // match.all instruction has two variants -- one returns a single value,
    // another returns a pair {value, predicate}. We currently only implement
    // the latter as that's the variant exposed by CUDA API.
    return valType.isInteger(32) ? llvm::Intrinsic::nvvm_match_all_sync_i32p
                                 : llvm::Intrinsic::nvvm_match_all_sync_i64p;
  }
  llvm_unreachable("unsupported match sync kind");
}
 
static llvm::Intrinsic::ID getVoteSyncIntrinsicId(NVVM::VoteSyncKind kind) {
  switch (kind) {
  case NVVM::VoteSyncKind::any:
    return llvm::Intrinsic::nvvm_vote_any_sync;
  case NVVM::VoteSyncKind::all:
    return llvm::Intrinsic::nvvm_vote_all_sync;
  case NVVM::VoteSyncKind::ballot:
    return llvm::Intrinsic::nvvm_vote_ballot_sync;
  case NVVM::VoteSyncKind::uni:
    return llvm::Intrinsic::nvvm_vote_uni_sync;
  }
  llvm_unreachable("unsupported vote kind");
}
 
static llvm::Intrinsic::ID
getLdMatrixIntrinsicId(NVVM::MMALayout layout, int32_t num,
                       NVVM::LdStMatrixShapeAttr shape,
                       NVVM::LdStMatrixEltType eltType) {
  if (shape.getM() == 8 && shape.getN() == 8) {
    switch (num) {
    case 1:
      return (layout == NVVM::MMALayout::row)
                 ? llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m8n8_x1_b16
                 : llvm::Intrinsic::
                       nvvm_ldmatrix_sync_aligned_m8n8_x1_trans_b16;
    case 2:
      return (layout == NVVM::MMALayout::row)
                 ? llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m8n8_x2_b16
                 : llvm::Intrinsic::
                       nvvm_ldmatrix_sync_aligned_m8n8_x2_trans_b16;
    case 4:
      return (layout == NVVM::MMALayout::row)
                 ? llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m8n8_x4_b16
                 : llvm::Intrinsic::
                       nvvm_ldmatrix_sync_aligned_m8n8_x4_trans_b16;
    }
  } else if (shape.getM() == 8 && shape.getN() == 16) {
    if (eltType == NVVM::LdStMatrixEltType::B8X16_B6X16_P32) {
      switch (num) {
      case 1:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m8n16_x1_b8x16_b6x16_p32;
      case 2:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m8n16_x2_b8x16_b6x16_p32;
      case 4:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m8n16_x4_b8x16_b6x16_p32;
      }
    } else if (eltType == NVVM::LdStMatrixEltType::B8X16_B4X16_P64) {
      switch (num) {
      case 1:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m8n16_x1_b8x16_b4x16_p64;
      case 2:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m8n16_x2_b8x16_b4x16_p64;
      case 4:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m8n16_x4_b8x16_b4x16_p64;
      }
    } else if (eltType == NVVM::LdStMatrixEltType::S8_S4) {
      switch (num) {
      case 1:
        return llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m8n16_x1_s8_s4;
      case 2:
        return llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m8n16_x2_s8_s4;
      case 4:
        return llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m8n16_x4_s8_s4;
      }
    }
  } else if (shape.getM() == 16 && shape.getN() == 16) {
    if (eltType == NVVM::LdStMatrixEltType::B8) {
      switch (num) {
      case 1:
        return llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m16n16_x1_trans_b8;
      case 2:
        return llvm::Intrinsic::nvvm_ldmatrix_sync_aligned_m16n16_x2_trans_b8;
      }
    } else if (eltType == NVVM::LdStMatrixEltType::B8X16_B6X16_P32) {
      switch (num) {
      case 1:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m16n16_x1_trans_b8x16_b6x16_p32;
      case 2:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m16n16_x2_trans_b8x16_b6x16_p32;
      }
    } else if (eltType == NVVM::LdStMatrixEltType::B8X16_B4X16_P64) {
      switch (num) {
      case 1:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m16n16_x1_trans_b8x16_b4x16_p64;
      case 2:
        return llvm::Intrinsic::
            nvvm_ldmatrix_sync_aligned_m16n16_x2_trans_b8x16_b4x16_p64;
      }
    }
  }
  llvm_unreachable("unknown ldmatrix kind");
}
 
/// Return the intrinsic ID associated with stmatrix for the given paramters.
static llvm::Intrinsic::ID
getStMatrixIntrinsicId(NVVM::MMALayout layout, int32_t num,
                       NVVM::LdStMatrixShapeAttr shape,
                       NVVM::LdStMatrixEltType eltType) {
  if (shape.getM() == 8 && shape.getN() == 8) {
    switch (num) {
    case 1:
      return (layout == NVVM::MMALayout::row)
                 ? llvm::Intrinsic::nvvm_stmatrix_sync_aligned_m8n8_x1_b16
                 : llvm::Intrinsic::
                       nvvm_stmatrix_sync_aligned_m8n8_x1_trans_b16;
    case 2:
      return (layout == NVVM::MMALayout::row)
                 ? llvm::Intrinsic::nvvm_stmatrix_sync_aligned_m8n8_x2_b16
                 : llvm::Intrinsic::
                       nvvm_stmatrix_sync_aligned_m8n8_x2_trans_b16;
    case 4:
      return (layout == NVVM::MMALayout::row)
                 ? llvm::Intrinsic::nvvm_stmatrix_sync_aligned_m8n8_x4_b16
                 : llvm::Intrinsic::
                       nvvm_stmatrix_sync_aligned_m8n8_x4_trans_b16;
    }
  } else if (shape.getM() == 16 && shape.getN() == 8) {
    switch (num) {
    case 1:
      return llvm::Intrinsic::nvvm_stmatrix_sync_aligned_m16n8_x1_trans_b8;
    case 2:
      return llvm::Intrinsic::nvvm_stmatrix_sync_aligned_m16n8_x2_trans_b8;
    case 4:
      return llvm::Intrinsic::nvvm_stmatrix_sync_aligned_m16n8_x4_trans_b8;
    }
  }
  llvm_unreachable("unknown stmatrix kind");
}
 
static unsigned getUnidirectionalFenceProxyID(NVVM::ProxyKind fromProxy,
                                              NVVM::ProxyKind toProxy,
                                              NVVM::MemScopeKind scope,
                                              bool isRelease) {
  if (fromProxy == NVVM::ProxyKind::GENERIC &&
      toProxy == NVVM::ProxyKind::TENSORMAP) {
    switch (scope) {
    case NVVM::MemScopeKind::CTA: {
      if (isRelease)
        return llvm::Intrinsic::nvvm_fence_proxy_tensormap_generic_release_cta;
      return llvm::Intrinsic::nvvm_fence_proxy_tensormap_generic_acquire_cta;
    }
    case NVVM::MemScopeKind::CLUSTER: {
      if (isRelease)
        return llvm::Intrinsic::
            nvvm_fence_proxy_tensormap_generic_release_cluster;
      return llvm::Intrinsic::
          nvvm_fence_proxy_tensormap_generic_acquire_cluster;
    }
    case NVVM::MemScopeKind::GPU: {
      if (isRelease)
        return llvm::Intrinsic::nvvm_fence_proxy_tensormap_generic_release_gpu;
      return llvm::Intrinsic::nvvm_fence_proxy_tensormap_generic_acquire_gpu;
    }
    case NVVM::MemScopeKind::SYS: {
      if (isRelease)
        return llvm::Intrinsic::nvvm_fence_proxy_tensormap_generic_release_sys;
      return llvm::Intrinsic::nvvm_fence_proxy_tensormap_generic_acquire_sys;
    }
    }
    llvm_unreachable("Unknown scope for uni-directional fence.proxy operation");
  }
  llvm_unreachable("Unsupported proxy kinds");
}
 
static unsigned getMembarIntrinsicID(NVVM::MemScopeKind scope) {
  switch (scope) {
  case NVVM::MemScopeKind::CTA:
    return llvm::Intrinsic::nvvm_membar_cta;
  case NVVM::MemScopeKind::CLUSTER:
    return llvm::Intrinsic::nvvm_fence_sc_cluster;
  case NVVM::MemScopeKind::GPU:
    return llvm::Intrinsic::nvvm_membar_gl;
  case NVVM::MemScopeKind::SYS:
    return llvm::Intrinsic::nvvm_membar_sys;
  }
  llvm_unreachable("Unknown scope for memory barrier");
}
 
#define TCGEN05LD(SHAPE, NUM) llvm::Intrinsic::nvvm_tcgen05_ld_##SHAPE##_##NUM
 
static llvm::Intrinsic::ID
getTcgen05LdIntrinsicID(mlir::NVVM::Tcgen05LdStShape shape, uint32_t num) {
  llvm::Intrinsic::ID Shape16x64b[] = {
      TCGEN05LD(16x64b, x1),  TCGEN05LD(16x64b, x2),   TCGEN05LD(16x64b, x4),
      TCGEN05LD(16x64b, x8),  TCGEN05LD(16x64b, x16),  TCGEN05LD(16x64b, x32),
      TCGEN05LD(16x64b, x64), TCGEN05LD(16x64b, x128),
  };
 
  llvm::Intrinsic::ID Shape16x128b[] = {
      TCGEN05LD(16x128b, x1),  TCGEN05LD(16x128b, x2),  TCGEN05LD(16x128b, x4),
      TCGEN05LD(16x128b, x8),  TCGEN05LD(16x128b, x16), TCGEN05LD(16x128b, x32),
      TCGEN05LD(16x128b, x64),
  };
 
  llvm::Intrinsic::ID Shape16x256b[] = {
      TCGEN05LD(16x256b, x1), TCGEN05LD(16x256b, x2),  TCGEN05LD(16x256b, x4),
      TCGEN05LD(16x256b, x8), TCGEN05LD(16x256b, x16), TCGEN05LD(16x256b, x32),
  };
 
  llvm::Intrinsic::ID Shape16x32bx2[] = {
      TCGEN05LD(16x32bx2, x1),  TCGEN05LD(16x32bx2, x2),
      TCGEN05LD(16x32bx2, x4),  TCGEN05LD(16x32bx2, x8),
      TCGEN05LD(16x32bx2, x16), TCGEN05LD(16x32bx2, x32),
      TCGEN05LD(16x32bx2, x64), TCGEN05LD(16x32bx2, x128),
  };
 
  llvm::Intrinsic::ID Shape32x32b[] = {
      TCGEN05LD(32x32b, x1),  TCGEN05LD(32x32b, x2),   TCGEN05LD(32x32b, x4),
      TCGEN05LD(32x32b, x8),  TCGEN05LD(32x32b, x16),  TCGEN05LD(32x32b, x32),
      TCGEN05LD(32x32b, x64), TCGEN05LD(32x32b, x128),
  };
 
  // `num` contains the length of vector and log2 of `num` returns the index
  // into the shape array
  unsigned Idx = std::log2(num);
 
  switch (shape) {
  case NVVM::Tcgen05LdStShape::SHAPE_16X64B:
    return Shape16x64b[Idx];
  case NVVM::Tcgen05LdStShape::SHAPE_16X128B:
    return Shape16x128b[Idx - 1];
  case NVVM::Tcgen05LdStShape::SHAPE_16X256B:
    return Shape16x256b[Idx - 2];
  case NVVM::Tcgen05LdStShape::SHAPE_32X32B:
    return Shape32x32b[Idx];
  case NVVM::Tcgen05LdStShape::SHAPE_16X32BX2:
    return Shape16x32bx2[Idx];
  }
  llvm_unreachable("unhandled tcgen05.ld lowering");
}
 
#define TCGEN05ST(SHAPE, NUM) llvm::Intrinsic::nvvm_tcgen05_st_##SHAPE##_##NUM
 
static llvm::Intrinsic::ID
getTcgen05StIntrinsicID(mlir::NVVM::Tcgen05LdStShape shape, uint32_t num) {
  llvm::Intrinsic::ID Shape16x64b[] = {
      TCGEN05ST(16x64b, x1),  TCGEN05ST(16x64b, x2),   TCGEN05ST(16x64b, x4),
      TCGEN05ST(16x64b, x8),  TCGEN05ST(16x64b, x16),  TCGEN05ST(16x64b, x32),
      TCGEN05ST(16x64b, x64), TCGEN05ST(16x64b, x128),
  };
 
  llvm::Intrinsic::ID Shape16x128b[] = {
      TCGEN05ST(16x128b, x1),  TCGEN05ST(16x128b, x2),  TCGEN05ST(16x128b, x4),
      TCGEN05ST(16x128b, x8),  TCGEN05ST(16x128b, x16), TCGEN05ST(16x128b, x32),
      TCGEN05ST(16x128b, x64),
  };
 
  llvm::Intrinsic::ID Shape16x256b[] = {
      TCGEN05ST(16x256b, x1), TCGEN05ST(16x256b, x2),  TCGEN05ST(16x256b, x4),
      TCGEN05ST(16x256b, x8), TCGEN05ST(16x256b, x16), TCGEN05ST(16x256b, x32),
  };
 
  llvm::Intrinsic::ID Shape16x32bx2[] = {
      TCGEN05ST(16x32bx2, x1),  TCGEN05ST(16x32bx2, x2),
      TCGEN05ST(16x32bx2, x4),  TCGEN05ST(16x32bx2, x8),
      TCGEN05ST(16x32bx2, x16), TCGEN05ST(16x32bx2, x32),
      TCGEN05ST(16x32bx2, x64), TCGEN05ST(16x32bx2, x128),
  };
 
  llvm::Intrinsic::ID Shape32x32b[] = {
      TCGEN05ST(32x32b, x1),  TCGEN05ST(32x32b, x2),   TCGEN05ST(32x32b, x4),
      TCGEN05ST(32x32b, x8),  TCGEN05ST(32x32b, x16),  TCGEN05ST(32x32b, x32),
      TCGEN05ST(32x32b, x64), TCGEN05ST(32x32b, x128),
  };
 
  // `num` contains the length of vector and log2 of `num` returns the index
  // into the shape array
  unsigned Idx = std::log2(num);
 
  switch (shape) {
  case NVVM::Tcgen05LdStShape::SHAPE_16X64B:
    return Shape16x64b[Idx];
  case NVVM::Tcgen05LdStShape::SHAPE_16X128B:
    return Shape16x128b[Idx - 1];
  case NVVM::Tcgen05LdStShape::SHAPE_16X256B:
    return Shape16x256b[Idx - 2];
  case NVVM::Tcgen05LdStShape::SHAPE_32X32B:
    return Shape32x32b[Idx];
  case NVVM::Tcgen05LdStShape::SHAPE_16X32BX2:
    return Shape16x32bx2[Idx];
  }
  llvm_unreachable("unhandled tcgen05.st lowering");
}
 
static llvm::Intrinsic::ID getFenceSyncRestrictID(NVVM::MemOrderKind order) {
  return order == NVVM::MemOrderKind::ACQUIRE
             ? llvm::Intrinsic::
                   nvvm_fence_acquire_sync_restrict_space_cluster_scope_cluster
             : llvm::Intrinsic::
                   nvvm_fence_release_sync_restrict_space_cta_scope_cluster;
}
 
static llvm::Intrinsic::ID
getFenceProxyID(NVVM::ProxyKind kind, std::optional<NVVM::SharedSpace> space) {
  switch (kind) {
  case NVVM::ProxyKind::alias:
    return llvm::Intrinsic::nvvm_fence_proxy_alias;
  case NVVM::ProxyKind::async:
    return llvm::Intrinsic::nvvm_fence_proxy_async;
  case NVVM::ProxyKind::async_global:
    return llvm::Intrinsic::nvvm_fence_proxy_async_global;
  case NVVM::ProxyKind::async_shared:
    return *space == NVVM::SharedSpace::shared_cta
               ? llvm::Intrinsic::nvvm_fence_proxy_async_shared_cta
               : llvm::Intrinsic::nvvm_fence_proxy_async_shared_cluster;
  default:
    llvm_unreachable("unsupported proxy kind");
  }
}
 
static llvm::Intrinsic::ID
getFenceProxySyncRestrictID(NVVM::MemOrderKind order) {
  return order == NVVM::MemOrderKind::ACQUIRE
             ? llvm::Intrinsic::
                   nvvm_fence_proxy_async_generic_acquire_sync_restrict_space_cluster_scope_cluster
             : llvm::Intrinsic::
                   nvvm_fence_proxy_async_generic_release_sync_restrict_space_cta_scope_cluster;
}
 
static llvm::RoundingMode
getLLVMRoundingModeForFPArith(NVVM::FPRoundingMode rndMode) {
  switch (rndMode) {
  case NVVM::FPRoundingMode::RN:
    return llvm::RoundingMode::NearestTiesToEven;
  case NVVM::FPRoundingMode::RM:
    return llvm::RoundingMode::TowardNegative;
  case NVVM::FPRoundingMode::RP:
    return llvm::RoundingMode::TowardPositive;
  case NVVM::FPRoundingMode::RZ:
    return llvm::RoundingMode::TowardZero;
  default:
    // default rounding mode is RN
    assert(rndMode == NVVM::FPRoundingMode::NONE &&
           "unsupported rounding mode for nvvm fp arithmetic");
    return llvm::RoundingMode::NearestTiesToEven;
  }
}
 
// Calls an LLVM intrinsic on the given operands. For f32/f64 vector types,
// the intrinsic is called per-element and the results are packed back into a
// vector. If retType is non-null, it is forwarded as the return-type
// overload to `createIntrinsicCall`.
static llvm::Value *
createScalarizedIntrinsicCall(llvm::IRBuilderBase &builder,
                              llvm::Intrinsic::ID IID, llvm::Type *opTypeLLVM,
                              ArrayRef<llvm::Value *> operands,
                              llvm::Type *retType) {
  if (opTypeLLVM->isVectorTy() && (opTypeLLVM->getScalarType()->isFloatTy() ||
                                   opTypeLLVM->getScalarType()->isDoubleTy())) {
    llvm::Value *result = llvm::PoisonValue::get(
        llvm::FixedVectorType::get(opTypeLLVM->getScalarType(), 2));
    for (int64_t i = 0; i < 2; ++i) {
      llvm::SmallVector<llvm::Value *> scalarArgs;
      for (llvm::Value *op : operands)
        scalarArgs.push_back(
            op->getType()->isVectorTy()
                ? builder.CreateExtractElement(op, builder.getInt32(i))
                : op);
      llvm::Value *res = createIntrinsicCall(builder, IID, retType, scalarArgs);
      result = builder.CreateInsertElement(result, res, builder.getInt32(i));
    }
    return result;
  }
 
  return createIntrinsicCall(builder, IID, retType, operands);
}
 
// Lowers a binary FP arithmetic op to the intrinsic selected from `ids`, which
// is indexed by [ftz][sat].
static void lowerFloatBinaryOpToLLVMIR(const llvm::Intrinsic::ID (&ids)[2][2],
                                       llvm::Value *argLHS, llvm::Value *argRHS,
                                       Value res, NVVM::FPRoundingMode rndMode,
                                       NVVM::SaturationMode satMode, bool isFTZ,
                                       LLVM::ModuleTranslation &mt,
                                       llvm::IRBuilderBase &builder) {
  llvm::Type *opTypeLLVM = argLHS->getType();
  bool isSat = satMode != NVVM::SaturationMode::NONE;
 
  llvm::Intrinsic::ID id = ids[isFTZ][isSat];
  llvm::Value *rnd = builder.getInt32(
      static_cast<int>(getLLVMRoundingModeForFPArith(rndMode)));
 
  // For f64 vectors, and f32 vectors with saturation, we need to scalarize
  // the intrinsic call.
  llvm::Type *scalarTypeLLVM = opTypeLLVM->getScalarType();
  if (opTypeLLVM->isVectorTy() && (scalarTypeLLVM->isDoubleTy() ||
                                   (isSat && scalarTypeLLVM->isFloatTy()))) {
    mt.mapValue(res, createScalarizedIntrinsicCall(builder, id, opTypeLLVM,
                                                   {argLHS, argRHS, rnd},
                                                   scalarTypeLLVM));
    return;
  }
 
  mt.mapValue(
      res, createIntrinsicCall(builder, id, opTypeLLVM, {argLHS, argRHS, rnd}));
}
 
void NVVM::AddFOp::lowerAddFToLLVMIR(llvm::Value *argLHS, llvm::Value *argRHS,
                                     Value res, NVVM::FPRoundingMode rndMode,
                                     NVVM::SaturationMode satMode, bool isFTZ,
                                     LLVM::ModuleTranslation &mt,
                                     llvm::IRBuilderBase &builder) {
  static constexpr llvm::Intrinsic::ID addIDs[2][2] = {
      {llvm::Intrinsic::nvvm_fadd, llvm::Intrinsic::nvvm_fadd_sat},
      {llvm::Intrinsic::nvvm_fadd_ftz, llvm::Intrinsic::nvvm_fadd_ftz_sat}};
  lowerFloatBinaryOpToLLVMIR(addIDs, argLHS, argRHS, res, rndMode, satMode,
                             isFTZ, mt, builder);
}
 
void NVVM::MulFOp::lowerMulFToLLVMIR(llvm::Value *argLHS, llvm::Value *argRHS,
                                     Value res, NVVM::FPRoundingMode rndMode,
                                     NVVM::SaturationMode satMode, bool isFTZ,
                                     LLVM::ModuleTranslation &mt,
                                     llvm::IRBuilderBase &builder) {
  static constexpr llvm::Intrinsic::ID mulIDs[2][2] = {
      {llvm::Intrinsic::nvvm_fmul, llvm::Intrinsic::nvvm_fmul_sat},
      {llvm::Intrinsic::nvvm_fmul_ftz, llvm::Intrinsic::nvvm_fmul_ftz_sat}};
  lowerFloatBinaryOpToLLVMIR(mulIDs, argLHS, argRHS, res, rndMode, satMode,
                             isFTZ, mt, builder);
}
 
void NVVM::FmaOp::lowerFmaToLLVMIR(Operation &op, LLVM::ModuleTranslation &mt,
                                   llvm::IRBuilderBase &builder) {
  auto thisOp = cast<NVVM::FmaOp>(op);
  mlir::NVVM::FPRoundingMode rndMode = thisOp.getRnd();
  bool isFTZ = thisOp.getFtz();
  bool isRelu = thisOp.getRelu();
  bool isSat = thisOp.getSat() == NVVM::SaturationMode::SAT;
  bool isOOB = thisOp.getOob();
 
  mlir::Type opType = thisOp.getRes().getType();
  llvm::Type *opTypeLLVM = mt.convertType(opType);
 
  llvm::Value *argA = mt.lookupValue(thisOp.getA());
  llvm::Value *argB = mt.lookupValue(thisOp.getB());
  llvm::Value *argC = mt.lookupValue(thisOp.getC());
 
  // The oob variants only support RN and take no rounding-mode operand.
  if (isOOB) {
    llvm::Intrinsic::ID id = isRelu ? llvm::Intrinsic::nvvm_fma_rn_oob_relu
                                    : llvm::Intrinsic::nvvm_fma_rn_oob;
    mt.mapValue(thisOp.getRes(), createIntrinsicCall(builder, id, opTypeLLVM,
                                                     {argA, argB, argC}));
    return;
  }
 
  static constexpr llvm::Intrinsic::ID fmaIDs[2][3] = {
      {llvm::Intrinsic::nvvm_fma, llvm::Intrinsic::nvvm_fma_sat,
       llvm::Intrinsic::nvvm_fma_relu},
      {llvm::Intrinsic::nvvm_fma_ftz, llvm::Intrinsic::nvvm_fma_ftz_sat,
       llvm::Intrinsic::nvvm_fma_ftz_relu}};
 
  llvm::Intrinsic::ID id = fmaIDs[isFTZ][isSat ? 1 : (isRelu ? 2 : 0)];
  llvm::Value *rnd = builder.getInt32(
      static_cast<int>(getLLVMRoundingModeForFPArith(rndMode)));
 
  // For f64 vector fma, and f32 vector fma with saturation, we need to
  // scalarize the intrinsic call.
  llvm::Type *scalarTypeLLVM = opTypeLLVM->getScalarType();
  if (opTypeLLVM->isVectorTy() && (scalarTypeLLVM->isDoubleTy() ||
                                   (isSat && scalarTypeLLVM->isFloatTy()))) {
    mt.mapValue(thisOp.getRes(), createScalarizedIntrinsicCall(
                                     builder, id, opTypeLLVM,
                                     {argA, argB, argC, rnd}, scalarTypeLLVM));
    return;
  }
 
  mt.mapValue(thisOp.getRes(), createIntrinsicCall(builder, id, opTypeLLVM,
                                                   {argA, argB, argC, rnd}));
}
 
// Returns the LLVM type the sp{de}compress intrinsics use for metadata: i32 for
// a single register (including vector<1xi32>) and a vector of i32 otherwise.
static llvm::Type *getSPMetadataType(Type type, llvm::IRBuilderBase &builder) {
  auto vecTy = dyn_cast<VectorType>(type);
  if (!vecTy || vecTy.getNumElements() == 1)
    return builder.getInt32Ty();
  return llvm::FixedVectorType::get(builder.getInt32Ty(),
                                    vecTy.getNumElements());
}
 
void NVVM::SparseCompressOp::lowerToLLVMIR(Operation &op,
                                           LLVM::ModuleTranslation &mt,
                                           llvm::IRBuilderBase &builder) {
  auto thisOp = cast<NVVM::SparseCompressOp>(op);
  Value metadata = thisOp.getMetadata();
  Value compressedData = thisOp.getCompressedData();
  llvm::Type *resultTy =
      llvm::StructType::get(getSPMetadataType(metadata.getType(), builder),
                            mt.convertType(compressedData.getType()));
 
  llvm::Value *result = createIntrinsicCall(
      builder, llvm::Intrinsic::nvvm_spcompress, resultTy,
      {mt.lookupValue(thisOp.getData()), mt.lookupValue(thisOp.getSpDesc()),
       builder.getInt32(thisOp.getIdxSize()),
       // spcompress only supports the 2:4 sparsity factor.
       builder.getInt32(4)});
 
  mt.mapValue(metadata) =
      builder.CreateBitCast(builder.CreateExtractValue(result, 0),
                            mt.convertType(metadata.getType()));
  mt.mapValue(compressedData) = builder.CreateExtractValue(result, 1);
}
 
void NVVM::SparseDecompressOp::lowerToLLVMIR(Operation &op,
                                             LLVM::ModuleTranslation &mt,
                                             llvm::IRBuilderBase &builder) {
  auto thisOp = cast<NVVM::SparseDecompressOp>(op);
  Value metadata = thisOp.getMetadata();
  mt.mapValue(thisOp.getData()) = createIntrinsicCall(
      builder, llvm::Intrinsic::nvvm_spdecompress,
      mt.convertType(thisOp.getData().getType()),
      {builder.CreateBitCast(mt.lookupValue(metadata),
                             getSPMetadataType(metadata.getType(), builder)),
       mt.lookupValue(thisOp.getCompressedData()),
       builder.getInt32(thisOp.getIdxSize()),
       builder.getInt32(thisOp.getNumTgt())});
}
 
namespace {
/// Implementation of the dialect interface that converts operations belonging
/// to the NVVM dialect to LLVM IR.
class NVVMDialectLLVMIRTranslationInterface
    : public LLVMTranslationDialectInterface {
public:
  using LLVMTranslationDialectInterface::LLVMTranslationDialectInterface;
 
  /// Translates the given operation to LLVM IR using the provided IR builder
  /// and saving the state in `moduleTranslation`.
  LogicalResult
  convertOperation(Operation *op, llvm::IRBuilderBase &builder,
                   LLVM::ModuleTranslation &moduleTranslation) const final {
    // All NVVM ops are instruction-level and require an active insertion point.
    // A null insert block means the op is misplaced (e.g., at module scope),
    // which would otherwise cause a null dereference in createIntrinsicCall.
    if (!builder.GetInsertBlock())
      return op->emitOpError(
          "cannot be translated to LLVM IR without an active insertion "
          "point; make sure the op is inside a function");
    Operation &opInst = *op;
#include "mlir/Dialect/LLVMIR/NVVMConversions.inc"
 
    return failure();
  }
 
  /// Attaches module-level metadata for functions marked as kernels
  /// and managed annotations for global variables.
  LogicalResult
  amendOperation(Operation *op, ArrayRef<llvm::Instruction *> instructions,
                 NamedAttribute attribute,
                 LLVM::ModuleTranslation &moduleTranslation) const final {
    if (auto globalOp = dyn_cast<LLVM::GlobalOp>(op)) {
      if (attribute.getName() == NVVM::NVVMDialect::getManagedAttrName()) {
        auto *gv = cast<llvm::GlobalVariable>(
            moduleTranslation.lookupGlobal(globalOp));
        llvm::Module *m = gv->getParent();
        llvm::LLVMContext &ctx = m->getContext();
        llvm::NamedMDNode *md = m->getOrInsertNamedMetadata("nvvm.annotations");
        md->addOperand(llvm::MDNode::get(
            ctx, {llvm::ConstantAsMetadata::get(gv),
                  llvm::MDString::get(ctx, "managed"),
                  llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
                      llvm::Type::getInt32Ty(ctx), 1))}));
      }
      return success();
    }
 
    auto func = dyn_cast<LLVM::LLVMFuncOp>(op);
    if (!func)
      return failure();
    llvm::Function *llvmFunc = moduleTranslation.lookupFunction(func.getName());
 
    if (attribute.getName() == NVVM::NVVMDialect::getMaxntidAttrName()) {
      if (!isa<DenseI32ArrayAttr>(attribute.getValue()))
        return failure();
      auto values = cast<DenseI32ArrayAttr>(attribute.getValue());
      const std::string attr = llvm::formatv(
          "{0:$[,]}", llvm::make_range(values.asArrayRef().begin(),
                                       values.asArrayRef().end()));
      llvmFunc->addFnAttr(llvm::NVVMAttr::MaxNTID, attr);
    } else if (attribute.getName() == NVVM::NVVMDialect::getReqntidAttrName()) {
      if (!isa<DenseI32ArrayAttr>(attribute.getValue()))
        return failure();
      auto values = cast<DenseI32ArrayAttr>(attribute.getValue());
      const std::string attr = llvm::formatv(
          "{0:$[,]}", llvm::make_range(values.asArrayRef().begin(),
                                       values.asArrayRef().end()));
      llvmFunc->addFnAttr(llvm::NVVMAttr::ReqNTID, attr);
    } else if (attribute.getName() ==
               NVVM::NVVMDialect::getClusterDimAttrName()) {
      if (!isa<DenseI32ArrayAttr>(attribute.getValue()))
        return failure();
      auto values = cast<DenseI32ArrayAttr>(attribute.getValue());
      const std::string attr = llvm::formatv(
          "{0:$[,]}", llvm::make_range(values.asArrayRef().begin(),
                                       values.asArrayRef().end()));
      llvmFunc->addFnAttr(llvm::NVVMAttr::ClusterDim, attr);
    } else if (attribute.getName() ==
               NVVM::NVVMDialect::getClusterMaxBlocksAttrName()) {
      auto value = dyn_cast<IntegerAttr>(attribute.getValue());
      llvmFunc->addFnAttr(llvm::NVVMAttr::MaxClusterRank,
                          llvm::utostr(value.getInt()));
    } else if (attribute.getName() ==
               NVVM::NVVMDialect::getMinctasmAttrName()) {
      auto value = dyn_cast<IntegerAttr>(attribute.getValue());
      llvmFunc->addFnAttr(llvm::NVVMAttr::MinCTASm,
                          llvm::utostr(value.getInt()));
    } else if (attribute.getName() == NVVM::NVVMDialect::getMaxnregAttrName()) {
      auto value = dyn_cast<IntegerAttr>(attribute.getValue());
      llvmFunc->addFnAttr(llvm::NVVMAttr::MaxNReg,
                          llvm::utostr(value.getInt()));
    } else if (attribute.getName() ==
               NVVM::NVVMDialect::getKernelFuncAttrName()) {
      llvmFunc->setCallingConv(llvm::CallingConv::PTX_Kernel);
    } else if (attribute.getName() ==
               NVVM::NVVMDialect::getBlocksAreClustersAttrName()) {
      llvmFunc->addFnAttr(llvm::NVVMAttr::BlocksAreClusters);
    }
 
    return success();
  }
 
  LogicalResult
  convertParameterAttr(LLVMFuncOp funcOp, int argIdx, NamedAttribute attribute,
                       LLVM::ModuleTranslation &moduleTranslation) const final {
 
    llvm::LLVMContext &llvmContext = moduleTranslation.getLLVMContext();
    llvm::Function *llvmFunc =
        moduleTranslation.lookupFunction(funcOp.getName());
 
    if (attribute.getName() == NVVM::NVVMDialect::getGridConstantAttrName()) {
      llvmFunc->addParamAttr(
          argIdx,
          llvm::Attribute::get(llvmContext, llvm::NVVMAttr::GridConstant));
    }
    return success();
  }
};
} // namespace
 
void mlir::registerNVVMDialectTranslation(DialectRegistry &registry) {
  registry.insert<NVVM::NVVMDialect>();
  registry.addExtension(+[](MLIRContext *ctx, NVVM::NVVMDialect *dialect) {
    dialect->addInterfaces<NVVMDialectLLVMIRTranslationInterface>();
  });
}
 
void mlir::registerNVVMDialectTranslation(MLIRContext &context) {
  DialectRegistry registry;
  registerNVVMDialectTranslation(registry);
  context.appendDialectRegistry(registry);
}