/*========================== begin_copyright_notice ============================

Copyright (C) 2017-2021 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

/*
 * ISA Byte Code Reader
 *
 * This library is designed to be extremely reusable and general in nature, and
 * as a result the following byte code reader code primarily uses the following
 * IR and data types:
 *
 * - common_isa_header
 * - kernel_format_t
 * - attribute_info_t
 * - CISA_opnd
 * - vector_opnd
 * - raw_opnd
 * - CISA_INST
 * - std::list<CISA_INST*>
 * - primitives (please use stdint.h types)
 *
 * which are used to load the byte code from a buffer into a simple structured
 * IR.
 *
 * Use of any other data types should be discussed by several members of the CM
 * jitter team before hand.
 *
 */

#include "Assertions.h"
#include "Attributes.hpp"
#include "Common_ISA.h"
#include "Common_ISA_framework.h"
#include "Common_ISA_util.h"
#include "G4_Kernel.hpp"
#include "IsaDisassembly.h"
#include "JitterDataStruct.h"
#include "Mem_Manager.h"
#include "common.h"
#include "visa_igc_common_header.h"

#include "VISAKernel.h"

#include <list>

// bfi can have 7 operands
#define COMMON_ISA_MAX_NUM_OPND_ARITH_LOGIC 7
#define STRING_LEN 1024

using namespace vISA;

struct RoutineContainer {
  RoutineContainer()
      : generalVarDecls(NULL), generalVarsCount(0), addressVarDecls(NULL),
        addressVarsCount(0), predicateVarDecls(NULL), predicateVarsCount(0),
        samplerVarDecls(NULL), samplerVarsCount(0), surfaceVarDecls(NULL),
        surfaceVarsCount(0), labelVarDecls(NULL), labelVarsCount(0),
        inputVarDecls(NULL), inputVarsCount(0), majorVersion(0),
        minorVersion(0) {}

  ~RoutineContainer() { stringPool.clear(); }
  RoutineContainer(const RoutineContainer&) = delete;
  RoutineContainer& operator=(const RoutineContainer&) = delete;
  VISA_GenVar **generalVarDecls;
  unsigned generalVarsCount;
  VISA_AddrVar **addressVarDecls;
  unsigned addressVarsCount;
  VISA_PredVar **predicateVarDecls;
  unsigned predicateVarsCount;
  VISA_SamplerVar **samplerVarDecls;
  unsigned samplerVarsCount;
  VISA_SurfaceVar **surfaceVarDecls;
  unsigned surfaceVarsCount;
  VISA_LabelOpnd **labelVarDecls;
  unsigned labelVarsCount;
  CISA_GEN_VAR **inputVarDecls;
  unsigned inputVarsCount;

  std::vector<std::string> stringPool;

  CISA_IR_Builder *builder = nullptr;
  VISAKernel *kernelBuilder = nullptr;
  uint8_t majorVersion;
  uint8_t minorVersion;
};

/// Assumming buf is start of the CISA byte code.
#define GET_MAJOR_VERSION(buf) (*((unsigned char *)&buf[4]))
#define GET_MINOR_VERSION(buf) (*((unsigned char *)&buf[5]))

#define READ_CISA_FIELD(dst, type, bytePos, buf)                               \
  do {                                                                         \
    dst = *((type *)&buf[bytePos]);                                            \
    bytePos += sizeof(type);                                                   \
  } while (0)

#define PEAK_CISA_FIELD(dst, type, bytePos, buf)                               \
  do {                                                                         \
    dst = *((type *)&buf[bytePos]);                                            \
  } while (0)

typedef enum {
  CISA_EMASK_M0,
  CISA_EMASK_M1,
  CISA_EMASK_M2,
  CISA_EMASK_M3,
  CISA_EMASK_M4,
  CISA_EMASK_M5,
  CISA_EMASK_M6,
  CISA_EMASK_M7,
  CISA_NO_EMASK,
  CISA_DEF_EMASK
} Common_ISA_EMask_Ctrl_3_0;

enum class FIELD_TYPE { DECL, INPUT };

// vISA 3.4+ supports 32-bit general variable IDs
// vISA 3.5+ supports 32-bit input count
template <typename T>
inline void readVarBytes(uint8_t major, uint8_t minor, T &dst,
                         uint32_t &bytePos, const char *buf,
                         FIELD_TYPE field = FIELD_TYPE::DECL) {
  vISA_ASSERT(std::is_integral<T>::value &&
                    (sizeof(T) == 2 || sizeof(T) == 4),
                "T should be short or int");
  uint32_t version = getVersionAsInt(major, minor);
  uintptr_t ptrval = reinterpret_cast<uintptr_t>(&buf[bytePos]);
  bool get4Bytes = false;
  if (field == FIELD_TYPE::DECL) {
    get4Bytes = (version >= getVersionAsInt(3, 4));
  } else if (field == FIELD_TYPE::INPUT) {
    get4Bytes = (version >= getVersionAsInt(3, 5));
  }

  if (get4Bytes) {
    dst = *(reinterpret_cast<uint32_t *>(ptrval));
    bytePos += sizeof(uint32_t);
  } else if (field == FIELD_TYPE::INPUT) {
    dst = *(reinterpret_cast<uint8_t *>(ptrval));
    bytePos += sizeof(uint8_t);
  } else {
    dst = *(reinterpret_cast<uint16_t *>(ptrval));
    bytePos += sizeof(uint16_t);
  }
}

static VISA_EMask_Ctrl transformMask(RoutineContainer &container,
                                     uint8_t maskVal) {
  VISA_EMask_Ctrl mask = vISA_EMASK_M1;
  if (container.majorVersion == 3 && container.minorVersion == 0) {
    Common_ISA_EMask_Ctrl_3_0 tMask = Common_ISA_EMask_Ctrl_3_0(maskVal);
    switch (tMask) {
    case CISA_EMASK_M0: {
      mask = vISA_EMASK_M1;
      break;
    }
    case CISA_EMASK_M1: {
      mask = vISA_EMASK_M2;
      break;
    }
    case CISA_EMASK_M2: {
      mask = vISA_EMASK_M3;
      break;
    }
    case CISA_EMASK_M3: {
      mask = vISA_EMASK_M4;
      break;
    }
    case CISA_EMASK_M4: {
      mask = vISA_EMASK_M5;
      break;
    }
    case CISA_EMASK_M5: {
      mask = vISA_EMASK_M6;
      break;
    }
    case CISA_EMASK_M6: {
      mask = vISA_EMASK_M7;
      break;
    }
    case CISA_EMASK_M7: {
      mask = vISA_EMASK_M8;
      break;
    }
    case CISA_NO_EMASK: {
      mask = vISA_EMASK_M1_NM;
      break;
    }
    case CISA_DEF_EMASK: {
      mask = vISA_EMASK_M1;
      break;
    }
    default:
      break;
    }
  } else {
    mask = VISA_EMask_Ctrl(maskVal);
  }

  return mask;
}
static void readExecSizeNG(unsigned &bytePos, const char *buf,
                           VISA_Exec_Size &size, VISA_EMask_Ctrl &mask,
                           RoutineContainer &container) {
  uint8_t execSize = 0;
  READ_CISA_FIELD(execSize, uint8_t, bytePos, buf);
  uint8_t maskVal = (execSize >> 0x4) & 0xF;

  mask = transformMask(container, maskVal);

  size = (VISA_Exec_Size)((execSize)&0xF);
}

template <typename T>
T readPrimitiveOperandNG(unsigned &bytePos, const char *buf) {
  vISA_ASSERT(buf != nullptr, "Argument Exception: argument buf  is NULL.");
  T data = 0;
  READ_CISA_FIELD(data, T, bytePos, buf);
  return data;
}

static VISA_PredOpnd *readPredicateOperandNG(unsigned &bytePos, const char *buf,
                                             RoutineContainer &container) {
  uint16_t predOpnd = 0;
  READ_CISA_FIELD(predOpnd, uint16_t, bytePos, buf);

  if (0 == predOpnd)
    return nullptr;

  VISAKernel *kernelBuilder = container.kernelBuilder;
  unsigned predID = (predOpnd & 0xfff);
  VISA_PREDICATE_CONTROL control =
      (VISA_PREDICATE_CONTROL)((predOpnd & 0x6000) >> 13);
  VISA_PREDICATE_STATE state =
      (VISA_PREDICATE_STATE)((predOpnd & 0x8000) >> 15);
  VISA_PredVar *decl = container.predicateVarDecls[predID];
  VISA_PredOpnd *opnd = nullptr;

  kernelBuilder->CreateVISAPredicateOperand(opnd, decl, state, control);

  return opnd;
}

static VISA_RawOpnd *readRawOperandNG(unsigned &bytePos, const char *buf,
                                      RoutineContainer &container) {
  vISA_ASSERT_INPUT(buf != nullptr, "Argument Exception: argument buf  is NULL.");
  uint8_t majorVersion = container.majorVersion;
  uint8_t minorVersion = container.minorVersion;

  uint32_t index = 0;
  uint16_t offset = 0;
  readVarBytes(majorVersion, minorVersion, index, bytePos, buf);
  READ_CISA_FIELD(offset, uint16_t, bytePos, buf);

  VISAKernelImpl *kernelBuilderImpl =
      ((VISAKernelImpl *)container.kernelBuilder);

  unsigned numPreDefinedVars = Get_CISA_PreDefined_Var_Count();
  VISA_GenVar *decl = NULL;
  VISA_RawOpnd *opnd = NULL;

  /**
      Null register is treated differently now. There is special operand NullReg
     created. In it field isNullReg is set to true.

      TODO:? To make things more generic need to mark decl created during
     initialization as null register, then when CreateVisaRawOperand is called
     check that decl passed in is null register decl, and mark operand is
     nullReg, also create region <0;1,0>
  */
  if (index == 0) {
    kernelBuilderImpl->CreateVISANullRawOperand(opnd, true); // dst
  } else {
    if (index >= numPreDefinedVars)
      decl = container.generalVarDecls[index];
    else
      kernelBuilderImpl->GetPredefinedVar(decl, (PreDefined_Vars)index);

    kernelBuilderImpl->CreateVISARawOperand(opnd, decl, offset);
  }

  return opnd;
}

static VISA_PredVar *readPreVarNG(unsigned &bytePos, const char *buf,
                                  RoutineContainer &container) {
  vISA_ASSERT_INPUT(buf != nullptr, "Argument Exception: argument buf  is NULL.");

  [[maybe_unused]] uint8_t tag = 0;
  READ_CISA_FIELD(tag, uint8_t, bytePos, buf);

  uint16_t index = 0;
  READ_CISA_FIELD(index, uint16_t, bytePos, buf);

  uint16_t predIndex = index & 0xfff;
  VISA_PredVar *decl = NULL;

  if (predIndex >= COMMON_ISA_NUM_PREDEFINED_PRED)
    decl = container.predicateVarDecls[predIndex];
  return decl;
}

static uint32_t readOtherOperandNG(unsigned &bytePos, const char *buf,
                                   VISA_Type visatype) {
  union {
    uint32_t other_opnd;
    struct {
      uint8_t b[4];
    };
  } v;

  unsigned bsize = CISATypeTable[visatype].typeSize;
  vISA_ASSERT_INPUT(bsize <= 4, " Unsupported other_opnd whose size > 4 bytes!");
  v.other_opnd = 0;
  for (int i = 0; i < (int)bsize; ++i) {
    v.b[i] = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  }
  return v.other_opnd;
}

static VISA_VectorOpnd *readVectorOperandNG(unsigned &bytePos, const char *buf,
                                            uint8_t &tag,
                                            RoutineContainer &container,
                                            unsigned int size, bool isDst,
                                            bool isAddressoff = false) {
  vISA_ASSERT_INPUT(buf != nullptr, "Argument Exception: argument buf  is NULL.");

  VISAKernelImpl *kernelBuilderImpl =
      ((VISAKernelImpl *)container.kernelBuilder);

  uint8_t majorVersion = container.majorVersion;
  uint8_t minorVersion = container.minorVersion;

  READ_CISA_FIELD(tag, uint8_t, bytePos, buf);
  VISA_Modifier modifier = ((VISA_Modifier)((tag >> 3) & 0x7));

  switch ((Common_ISA_Operand_Class)(tag & 0x7)) /// getOperandClass
  {
  case OPERAND_GENERAL: {
    uint32_t index = 0;
    uint8_t rowOffset = 0;
    uint8_t colOffset = 0;
    uint16_t region = 0;

    readVarBytes(majorVersion, minorVersion, index, bytePos, buf);
    READ_CISA_FIELD(rowOffset, uint8_t, bytePos, buf);
    READ_CISA_FIELD(colOffset, uint8_t, bytePos, buf);
    READ_CISA_FIELD(region, uint16_t, bytePos, buf);

    uint16_t v_stride =
        Get_Common_ISA_Region_Value((Common_ISA_Region_Val)(region & 0xF));
    uint16_t width = Get_Common_ISA_Region_Value(
        (Common_ISA_Region_Val)((region >> 4) & 0xF));
    uint16_t h_stride = Get_Common_ISA_Region_Value(
        (Common_ISA_Region_Val)((region >> 8) & 0xF));

    unsigned numPreDefinedVars = Get_CISA_PreDefined_Var_Count();

    VISA_Modifier mod = modifier;
    VISA_VectorOpnd *opnd = NULL;
    VISA_GenVar *decl = NULL;

    if (index >= numPreDefinedVars)
      decl = container.generalVarDecls[index];
    else
      kernelBuilderImpl->GetPredefinedVar(decl, (PreDefined_Vars)index);

    if (isDst)
      kernelBuilderImpl->CreateVISADstOperand(opnd, decl, h_stride, rowOffset,
                                              colOffset);
    else if (isAddressoff) {
      VISA_Type vType = decl->genVar.getType();
      G4_Type gType = GetGenTypeFromVISAType(vType);
      unsigned int offset =
          colOffset * TypeSize(gType) +
          rowOffset * kernelBuilderImpl->getKernel()->numEltPerGRF<Type_UB>();
      kernelBuilderImpl->CreateVISAAddressOfOperand(opnd, decl, offset);
    } else {
      kernelBuilderImpl->CreateVISASrcOperand(opnd, decl, mod, v_stride, width,
                                              h_stride, rowOffset, colOffset);
    }

    return opnd;
  }
  case OPERAND_ADDRESS: {
    uint16_t index = 0;
    uint8_t offset = 0;
    uint16_t width = 0;

    READ_CISA_FIELD(index, uint16_t, bytePos, buf);
    READ_CISA_FIELD(offset, uint8_t, bytePos, buf);
    READ_CISA_FIELD(width, uint8_t, bytePos, buf);

    VISA_VectorOpnd *opnd = NULL;
    VISA_AddrVar *decl = container.addressVarDecls[index];
    kernelBuilderImpl->CreateVISAAddressOperand(
        opnd, decl, offset, Get_VISA_Exec_Size((VISA_Exec_Size)width), isDst);

    return opnd;
  }
  case OPERAND_PREDICATE: {
    uint16_t index = 0;
    READ_CISA_FIELD(index, uint16_t, bytePos, buf);

    uint16_t predIndex = index & 0xfff;
    vASSERT(predIndex >= COMMON_ISA_NUM_PREDEFINED_PRED);

    VISA_PredVar *decl = container.predicateVarDecls[predIndex];
    VISA_VectorOpnd *opnd = nullptr;
    if (isDst) {
      kernelBuilderImpl->CreateVISAPredicateDstOperand(opnd, decl, size);
      return opnd;
    } else {
      kernelBuilderImpl->CreateVISAPredicateSrcOperand(opnd, decl, size);
      return opnd;
    }
  }
  case OPERAND_INDIRECT: {
    uint16_t index = 0;
    uint8_t addr_offset = 0;
    int16_t indirect_offset = 0;
    uint8_t bit_property = 0;
    uint16_t region = 0;

    READ_CISA_FIELD(index, uint16_t, bytePos, buf);
    READ_CISA_FIELD(addr_offset, uint8_t, bytePos, buf);
    READ_CISA_FIELD(indirect_offset, int16_t, bytePos, buf);
    READ_CISA_FIELD(bit_property, uint8_t, bytePos, buf);
    READ_CISA_FIELD(region, uint16_t, bytePos, buf);

    uint16_t v_stride =
        Get_Common_ISA_Region_Value((Common_ISA_Region_Val)((region)&0xF));
    uint16_t width = Get_Common_ISA_Region_Value(
        (Common_ISA_Region_Val)((region >> 4) & 0xF));
    uint16_t h_stride = Get_Common_ISA_Region_Value(
        (Common_ISA_Region_Val)((region >> 8) & 0xF));

    VISA_Modifier mod = modifier;
    VISA_VectorOpnd *opnd = NULL;
    VISA_AddrVar *decl = container.addressVarDecls[index];

    kernelBuilderImpl->CreateVISAIndirectGeneralOperand(
        opnd, decl, mod, addr_offset, indirect_offset, v_stride, width,
        h_stride, (VISA_Type)(bit_property & 0xF), isDst);

    return opnd;
  }
  case OPERAND_IMMEDIATE: {
    uint8_t type = 0;
    READ_CISA_FIELD(type, uint8_t, bytePos, buf);
    VISA_Type immedType = (VISA_Type)(type & 0xF);

    VISA_VectorOpnd *opnd = NULL;

    if (immedType == ISA_TYPE_DF) {
      double val = 0;
      READ_CISA_FIELD(val, double, bytePos, buf);
      kernelBuilderImpl->CreateVISAImmediate(opnd, &val, immedType);
    } else if (immedType == ISA_TYPE_Q || immedType == ISA_TYPE_UQ) {
      uint64_t val = 0;
      READ_CISA_FIELD(val, uint64_t, bytePos, buf);
      kernelBuilderImpl->CreateVISAImmediate(opnd, &val, immedType);
    } else /// Immediate operands are at least 4 bytes.
    {
      unsigned val = 0;
      READ_CISA_FIELD(val, unsigned, bytePos, buf);
      kernelBuilderImpl->CreateVISAImmediate(opnd, &val, immedType);
    }

    return opnd;
  }
  case OPERAND_STATE: {
    uint8_t opnd_class = 0;
    uint16_t index = 0;
    uint8_t offset = 0;

    READ_CISA_FIELD(opnd_class, uint8_t, bytePos, buf);
    READ_CISA_FIELD(index, uint16_t, bytePos, buf);
    READ_CISA_FIELD(offset, uint8_t, bytePos, buf);

    VISA_VectorOpnd *opnd = NULL;

    switch ((Common_ISA_State_Opnd_Class)opnd_class) {
    case STATE_OPND_SURFACE: {
      if (isAddressoff) {
        VISA_SurfaceVar *decl = container.surfaceVarDecls[index];
        unsigned int offsetB = offset * TypeSize(Type_UW);
        kernelBuilderImpl->CreateVISAAddressOfOperand(opnd, decl, offsetB);
      } else {
        VISA_SurfaceVar *decl = container.surfaceVarDecls[index];
        kernelBuilderImpl->CreateVISAStateOperand(opnd, decl, (uint8_t)size,
                                                  offset, isDst);
      }
      break;
    }
    case STATE_OPND_SAMPLER: {
      VISA_SamplerVar *decl = container.samplerVarDecls[index];
      if (isAddressoff) {
        unsigned int offsetB = offset * TypeSize(Type_UW);
        kernelBuilderImpl->CreateVISAAddressOfOperandGeneric(opnd, decl,
                                                             offsetB);
      } else {
        kernelBuilderImpl->CreateVISAStateOperand(opnd, decl, (uint8_t)size,
                                                  offset, isDst);
      }
      break;
    }
    default: {
      vISA_ASSERT_UNREACHABLE("Invalid state operand class: only surface and "
                          "sampler are supported.");
      break;
    }
    }

    return opnd;
  }
  default:
    vISA_ASSERT_UNREACHABLE("Operand class not recognized");
    return NULL;
  }
}

static VISA_VectorOpnd *readVectorOperandNG(unsigned &bytePos, const char *buf,
                                            RoutineContainer &container,
                                            unsigned int size) {
  uint8_t tag = 0;
  bool isDst = false;
  return readVectorOperandNG(bytePos, buf, tag, container, size, isDst);
}

static VISA_VectorOpnd *readVectorOperandNG(unsigned &bytePos, const char *buf,
                                            RoutineContainer &container,
                                            bool isDst) {
  uint8_t tag = 0;
  return readVectorOperandNG(bytePos, buf, tag, container, 1, isDst);
}

static VISA_VectorOpnd *
readVectorOperandNGAddressOf(unsigned &bytePos, const char *buf,
                             RoutineContainer &container) {
  uint8_t tag = 0;
  bool isDst = false;
  bool isAddressOff = true;
  return readVectorOperandNG(bytePos, buf, tag, container, 1, isDst,
                             isAddressOff);
}

static void readInstructionCommonNG(unsigned &bytePos, const char *buf,
                                    ISA_Opcode opcode,
                                    RoutineContainer &container) {
  VISA_EMask_Ctrl emask = vISA_EMASK_M1;
  VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;

  VISA_INST_Desc *inst_desc = &CISA_INST_table[opcode];
  unsigned opnd_count = inst_desc->opnd_num;
  unsigned opnd_skip = 0;

  for (unsigned i = 0; i < 2; i++) {
    if ((opnd_count - opnd_skip) > 0 &&
        (inst_desc->opnd_desc[i].opnd_type == OPND_EXECSIZE ||
         inst_desc->opnd_desc[i].opnd_type == OPND_PRED)) {
      opnd_skip++;
    }
  }

  VISAKernel *kernelBuilder = container.kernelBuilder;

  switch (ISA_Inst_Table[opcode].type) {
  case ISA_Inst_Mov:
  case ISA_Inst_Arith:
  case ISA_Inst_Logic:
  case ISA_Inst_Address:
  case ISA_Inst_Compare: {
    VISA_VectorOpnd *opnds[COMMON_ISA_MAX_NUM_OPND_ARITH_LOGIC];
    vISA_ASSERT_INPUT(
        opnd_count <= COMMON_ISA_MAX_NUM_OPND_ARITH_LOGIC,
        "Insturction operand count exceeds maximum supported operands.");
    memset(opnds, 0,
           sizeof(VISA_VectorOpnd *) * COMMON_ISA_MAX_NUM_OPND_ARITH_LOGIC);

    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = hasPredicate(opcode)
                              ? readPredicateOperandNG(bytePos, buf, container)
                              : NULL;

    uint8_t opSpec = 0;
    if (ISA_FMINMAX == opcode || ISA_CMP == opcode) {
      opSpec =
          readPrimitiveOperandNG<uint8_t>(bytePos, buf); /// rel_Op or opext
      opnd_skip++;
    }
    uint8_t bfn_func_ctrl = 0;

    uint32_t exSize = Get_VISA_Exec_Size(esize);
    uint8_t tag = 0;
    VISA_PredVar *dstDcl = NULL;
    bool cmpHasDst = false;
    for (unsigned i = 0; i < opnd_count - opnd_skip; i++) {
      bool isDst = i == 0;

      if ((OPND_DST_GEN & inst_desc->opnd_desc[i + opnd_skip].opnd_type) != 0) {
        isDst = true;
      }

      if (isDst) {
        if (ISA_Inst_Table[opcode].type == ISA_Inst_Compare) {
          opnds[i] = NULL;
          PEAK_CISA_FIELD(tag, uint8_t, bytePos, buf);

          if ((tag & 0x7) == OPERAND_GENERAL) {
            opnds[i] = readVectorOperandNG(bytePos, buf, tag, container,
                                           Get_VISA_Exec_Size(esize), true);
            cmpHasDst = true;

          } else {
            dstDcl = readPreVarNG(bytePos, buf, container);
          }
        } else
          opnds[i] =
              readVectorOperandNG(bytePos, buf, tag, container, exSize, isDst);
      } else if (ISA_Inst_Table[opcode].type == ISA_Inst_Address && i == 1) {
        // for first source of address add instruction.
        opnds[i] = readVectorOperandNGAddressOf(bytePos, buf, container);
      } else if (i == 4 && opcode == ISA_BFN) {
        // read bfn booleanFuncCtrl from the last opnd
        int opnd_ix = i + opnd_skip;
        vISA_ASSERT_INPUT(inst_desc->opnd_desc[opnd_ix].opnd_type == OPND_OTHER,
               "BFN: FuncCtrl opnd_desc's type should be OPND_OTHER!");
        VISA_Type visatype = (VISA_Type)inst_desc->opnd_desc[opnd_ix].data_type;
        bfn_func_ctrl = readOtherOperandNG(bytePos, buf, visatype);
      } else {
        opnds[i] = readVectorOperandNG(bytePos, buf, container, exSize);
      }
    }

    opnd_count -= opnd_skip;

    // skip booleanFuncCtrl, which is the last opnd
    if (opcode == ISA_BFN)
      --opnd_count;

    bool saturate = (((VISA_Modifier)((tag >> 3) & 0x7)) == MODIFIER_SAT);
    VISA_VectorOpnd *dst = opnds[0];
    VISA_VectorOpnd *src0 = opnds[1];
    VISA_VectorOpnd *src1 = opnd_count > 2 ? opnds[2] : NULL;
    VISA_VectorOpnd *src2 = opnd_count > 3 ? opnds[3] : NULL;
    VISA_VectorOpnd *src3 = opnd_count > 4 ? opnds[4] : NULL;

    switch (ISA_Inst_Table[opcode].type) {
    case ISA_Inst_Mov:
      if (opcode == ISA_FMINMAX)
        kernelBuilder->AppendVISAMinMaxInst((CISA_MIN_MAX_SUB_OPCODE)opSpec,
                                            saturate, emask, esize, dst, src0,
                                            src1);
      else
        kernelBuilder->AppendVISADataMovementInst(opcode, pred, saturate, emask,
                                                  esize, dst, src0, src1);
      break;
    case ISA_Inst_Arith:
      if (opcode == ISA_ADDC || opcode == ISA_SUBB) {
        kernelBuilder->AppendVISATwoDstArithmeticInst(
            opcode, pred, emask, esize, dst, src0, src1, src2);
      } else {
        kernelBuilder->AppendVISAArithmeticInst(opcode, pred, saturate, emask,
                                                esize, dst, src0, src1, src2);
      }
      break;
    case ISA_Inst_Logic:
      if (opcode == ISA_BFN)
        kernelBuilder->AppendVISABfnInst(bfn_func_ctrl, pred, saturate, emask,
                                         esize, dst, src0, src1, src2);
      else
        kernelBuilder->AppendVISALogicOrShiftInst(
            opcode, pred, saturate, emask, esize, dst, src0, src1, src2, src3);
      break;
    case ISA_Inst_Address:
      kernelBuilder->AppendVISAAddrAddInst(emask, esize, dst, src0, src1);
      break;
    case ISA_Inst_Compare:
      if (dstDcl)
        kernelBuilder->AppendVISAComparisonInst(
            (VISA_Cond_Mod)(opSpec & 0x7), emask, esize, dstDcl, src0, src1);
      else if (cmpHasDst)
        kernelBuilder->AppendVISAComparisonInst((VISA_Cond_Mod)(opSpec & 0x7),
                                                emask, esize, dst, src0, src1);
      break;
    default:
      break;
    }

    break;
  }
  case ISA_Inst_SIMD_Flow: {
    vISA_ASSERT_INPUT(opcode == ISA_GOTO, "expect goto instruction");
    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = hasPredicate(opcode)
                              ? readPredicateOperandNG(bytePos, buf, container)
                              : nullptr;
    VISA_LabelOpnd *label =
        opcode == ISA_GOTO
            ? container
                  .labelVarDecls[readPrimitiveOperandNG<uint16_t>(bytePos, buf)]
            : nullptr;
    kernelBuilder->AppendVISACFGotoInst(pred, emask, esize, label);
    break;
  }
  case ISA_Inst_Sync: {
    if (opcode == ISA_WAIT) {
      VISA_VectorOpnd *mask = NULL;
      if (getVersionAsInt(container.majorVersion, container.minorVersion) >=
          getVersionAsInt(3, 1)) {
        // additional vector operand
        mask = readVectorOperandNG(bytePos, buf, container, false);
      } else {
        // set mask to 0
        uint16_t value = 0;
        kernelBuilder->CreateVISAImmediate(mask, &value, ISA_TYPE_UW);
      }
      kernelBuilder->AppendVISAWaitInst(mask);
    } else if (opcode == ISA_SBARRIER) {
      uint32_t mode = readOtherOperandNG(bytePos, buf, ISA_TYPE_UB);
      kernelBuilder->AppendVISASplitBarrierInst(mode != 0);
    } else if (opcode == ISA_NBARRIER) {
      // Still support reading visa binary ?
      uint32_t mode = readOtherOperandNG(bytePos, buf, ISA_TYPE_UB);
      auto barrierId = readVectorOperandNG(bytePos, buf, container, false);
      VISA_VectorOpnd *barrierType =
          readVectorOperandNG(bytePos, buf, container, false);
      VISA_VectorOpnd *numProds =
          readVectorOperandNG(bytePos, buf, container, false);
      VISA_VectorOpnd *numCons =
          readVectorOperandNG(bytePos, buf, container, false);
      bool isWait = (mode == 0);
      if (isWait) {
        kernelBuilder->AppendVISANamedBarrierWait(barrierId);
      } else {
        const auto &vo = barrierType->_opnd.v_opnd;
        if ((vo.tag & 0x7) == OPERAND_IMMEDIATE &&
            vo.opnd_val.const_opnd._val.lval == 0 && numProds == numCons) {
          kernelBuilder->AppendVISANamedBarrierSignal(barrierId, numProds);
        } else {
          kernelBuilder->AppendVISANamedBarrierSignal(barrierId, barrierType,
                                                      numProds, numCons);
        }
      }
    } else {
      bool hasMask = (opcode == ISA_FENCE);
      uint8_t mask =
          hasMask ? readPrimitiveOperandNG<uint8_t>(bytePos, buf) : 0;
      kernelBuilder->AppendVISASyncInst(opcode, mask);
    }
    break;
  }
  default: {
    vISA_ASSERT_UNREACHABLE("Invalid common instruction type.");
  }
  }
}

/// Read a byte which encodes the atomic opcode and a flag indicating whether
/// this is a 16bit atomic operation.
std::tuple<VISAAtomicOps, unsigned short>
getAtomicOpAndBitwidth(unsigned &bytePos, const char *buf) {
  // bits 0-4 atomic op and bit 5-6 encode the bitwidth
  uint8_t data = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  VISAAtomicOps op = static_cast<VISAAtomicOps>(data & 0x1F);
  unsigned short bitwidth;
  if (data >> 5 == 1)
    bitwidth = 16;
  else if (data >> 6 == 1)
    bitwidth = 64;
  else
    bitwidth = 32;
  return std::tie(op, bitwidth);
}

static void readInstructionDataportNG(unsigned &bytePos, const char *buf,
                                      ISA_Opcode opcode,
                                      RoutineContainer &container) {
  VISAKernel *kernelBuilder = container.kernelBuilder;
  VISAKernelImpl *kernelBuilderImpl = ((VISAKernelImpl *)kernelBuilder);

  switch (opcode) {
  case ISA_MEDIA_ST:
  case ISA_MEDIA_LD: {
    uint8_t modifier = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t plane = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t width = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t height = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_opnd *xoffset = readVectorOperandNG(bytePos, buf, container, false);
    VISA_opnd *yoffset = readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *msg = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);

    kernelBuilder->AppendVISASurfAccessMediaLoadStoreInst(
        opcode, (MEDIA_LD_mod)modifier, surfaceHnd, width, height,
        (VISA_VectorOpnd *)xoffset, (VISA_VectorOpnd *)yoffset, msg,
        (CISA_PLANE_ID)plane);
    break;
  }
  case ISA_OWORD_ST:
  case ISA_OWORD_LD:
  case ISA_OWORD_LD_UNALIGNED: {
    uint8_t size = readPrimitiveOperandNG<uint8_t>(bytePos, buf) & 0x7;
    if (ISA_OWORD_ST != opcode) {
      readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    } // modifier
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *offset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *msg = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilder->AppendVISASurfAccessOwordLoadStoreInst(
        opcode, vISA_EMASK_M1, surfaceHnd, (VISA_Oword_Num)size,
        (VISA_VectorOpnd *)offset, msg);

    break;
  }
  case ISA_GATHER:
  case ISA_SCATTER: {
    uint8_t elt_size = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    if (ISA_GATHER == opcode) {
      readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    } // modifier
    uint8_t num_elts = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_VectorOpnd *globalOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *elementOffset = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *msg = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;

    /// TODO: Conversions like these make using vISA builder cumbersome.
    switch (num_elts & 0x3) {
    case 0:
      esize = EXEC_SIZE_8;
      break;
    case 1:
      esize = EXEC_SIZE_16;
      break;
    case 2:
      esize = EXEC_SIZE_1;
      break;
    default:
      vISA_ASSERT_UNREACHABLE("Invalid Number of Elements for Gather/Scatter.");
    }

    emask = transformMask(container, num_elts >> 4);

    kernelBuilderImpl->AppendVISASurfAccessGatherScatterInst(
        opcode, emask, (GATHER_SCATTER_ELEMENT_SIZE)(elt_size & 0x3), esize,
        surfaceHnd, globalOffset, elementOffset, msg);
    break;
  }
  case ISA_GATHER4_TYPED:
  case ISA_SCATTER4_TYPED: {
    if (getVersionAsInt(container.majorVersion, container.minorVersion) >=
        getVersionAsInt(3, 2)) {
      VISA_EMask_Ctrl emask = vISA_EMASK_M1;
      VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
      readExecSizeNG(bytePos, buf, esize, emask, container);

      VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
      unsigned ch_mask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
      uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

      VISA_RawOpnd *uOffset = readRawOperandNG(bytePos, buf, container);
      VISA_RawOpnd *vOffset = readRawOperandNG(bytePos, buf, container);
      VISA_RawOpnd *rOffset = readRawOperandNG(bytePos, buf, container);
      VISA_RawOpnd *lod = readRawOperandNG(bytePos, buf, container);
      VISA_RawOpnd *msg = readRawOperandNG(bytePos, buf, container);

      VISA_StateOpndHandle *surfaceHnd = NULL;
      kernelBuilderImpl->CreateVISAStateOperandHandle(
          surfaceHnd, container.surfaceVarDecls[surface]);
      kernelBuilderImpl->AppendVISASurfAccessGather4Scatter4TypedInst(
          opcode, pred, ChannelMask::createAPIFromBinary(opcode, ch_mask),
          emask, esize, surfaceHnd, uOffset, vOffset, rOffset, lod, msg);
    } else {
      uint8_t ch_mask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

      VISA_EMask_Ctrl emask = vISA_EMASK_M1;
      VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
      readExecSizeNG(bytePos, buf, esize, emask, container);

      vISA_ASSERT_INPUT(esize == 0, "Unsupported number of elements for "
                               "ISA_SCATTER4_TYPED/ISA_GATHER4_TYPED.");
      esize = EXEC_SIZE_8;

      uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

      VISA_RawOpnd *uOffset = readRawOperandNG(bytePos, buf, container);
      VISA_RawOpnd *vOffset = readRawOperandNG(bytePos, buf, container);
      VISA_RawOpnd *rOffset = readRawOperandNG(bytePos, buf, container);
      VISA_RawOpnd *lod = NULL;
      kernelBuilderImpl->CreateVISANullRawOperand(lod, false);
      VISA_RawOpnd *msg = readRawOperandNG(bytePos, buf, container);

      VISA_StateOpndHandle *surfaceHnd = NULL;
      kernelBuilderImpl->CreateVISAStateOperandHandle(
          surfaceHnd, container.surfaceVarDecls[surface]);
      ch_mask = ~ch_mask;
      kernelBuilderImpl->AppendVISASurfAccessGather4Scatter4TypedInst(
          opcode, NULL, ChannelMask::createAPIFromBinary(opcode, ch_mask),
          emask, esize, surfaceHnd, uOffset, vOffset, rOffset, lod, msg);
    }
    break;
  }
  case ISA_3D_RT_WRITE: {
    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

    uint16_t mode = readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *r1HeaderOpnd = readRawOperandNG(bytePos, buf, container);

    vISA_RT_CONTROLS cntrls;
    cntrls.s0aPresent = (mode & (0x1 << 3)) ? true : false;
    cntrls.oMPresent = (mode & (0x1 << 4)) ? true : false;
    cntrls.zPresent = (mode & (0x1 << 5)) ? true : false;
    cntrls.isStencil = (mode & (0x1 << 6)) ? true : false;
    cntrls.isLastWrite = (mode & (0x1 << 7)) ? true : false;
    bool CPSEnable = (mode & (0x1 << 8)) ? true : false;
    cntrls.isPerSample = (mode & (0x1 << 9)) ? true : false;
    cntrls.isCoarseMode = (mode & (0x1 << 10)) ? true : false;
    cntrls.isSampleIndex = (mode & (0x1 << 11)) ? true : false;
    cntrls.RTIndexPresent = (mode & (0x1 << 2)) ? true : false;
    cntrls.isNullRT = (mode & (0x1 << 12)) ? true : false;
    cntrls.isHeaderMaskfromCe0 = 0;

    VISA_VectorOpnd *sampleIndex =
        cntrls.isSampleIndex
            ? readVectorOperandNG(bytePos, buf, container, false)
            : NULL;
    VISA_VectorOpnd *cpsCounter =
        CPSEnable ? readVectorOperandNG(bytePos, buf, container, false) : NULL;
    VISA_VectorOpnd *rti =
        cntrls.RTIndexPresent
            ? readVectorOperandNG(bytePos, buf, container, false)
            : NULL;
    VISA_RawOpnd *s0a =
        cntrls.s0aPresent ? readRawOperandNG(bytePos, buf, container) : NULL;
    VISA_RawOpnd *oM =
        cntrls.oMPresent ? readRawOperandNG(bytePos, buf, container) : NULL;
    VISA_RawOpnd *R = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *G = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *B = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *A = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *Z =
        cntrls.zPresent ? readRawOperandNG(bytePos, buf, container) : NULL;
    VISA_RawOpnd *S =
        cntrls.isStencil ? readRawOperandNG(bytePos, buf, container) : NULL;

    std::vector<VISA_RawOpnd *> rawOpndVector;
    if (s0a)
      rawOpndVector.push_back(s0a);
    if (oM)
      rawOpndVector.push_back(oM);
    if (R)
      rawOpndVector.push_back(R);
    if (G)
      rawOpndVector.push_back(G);
    if (B)
      rawOpndVector.push_back(B);
    if (A)
      rawOpndVector.push_back(A);
    if (Z)
      rawOpndVector.push_back(Z);
    if (S)
      rawOpndVector.push_back(S);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->AppendVISA3dRTWriteCPS(
        pred, emask, esize, rti, cntrls, surfaceHnd, r1HeaderOpnd, sampleIndex,
        cpsCounter, (uint8_t)rawOpndVector.size(), rawOpndVector.data());
    break;
  }
  case ISA_GATHER4_SCALED:
  case ISA_SCATTER4_SCALED: {
    VISA_EMask_Ctrl eMask = vISA_EMASK_M1;
    VISA_Exec_Size exSize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, exSize, eMask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    unsigned channelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    // scale is ignored and must be zero
    (void)readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *globalOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *offsets = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dstOrSrc = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->AppendVISASurfAccessGather4Scatter4ScaledInst(
        opcode, pred, eMask, exSize,
        ChannelMask::createAPIFromBinary(opcode, channelMask), surfaceHnd,
        globalOffset, offsets, dstOrSrc);
    break;
  }
  case ISA_GATHER_SCALED:
  case ISA_SCATTER_SCALED: {
    VISA_EMask_Ctrl eMask = vISA_EMASK_M1;
    VISA_Exec_Size exSize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, exSize, eMask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    // block size is ignored (MBZ)
    (void)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_SVM_Block_Num numBlocks =
        VISA_SVM_Block_Num(readPrimitiveOperandNG<uint8_t>(bytePos, buf));
    // scale is ignored (MBZ)
    (void)readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *globalOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *offsets = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dstOrSrc = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->AppendVISASurfAccessScatterScaledInst(
        opcode, pred, eMask, exSize, numBlocks, surfaceHnd, globalOffset,
        offsets, dstOrSrc);
    break;
  }
  case ISA_DWORD_ATOMIC: {
    VISAAtomicOps subOpc;
    unsigned short bitwidth;
    std::tie(subOpc, bitwidth) = getAtomicOpAndBitwidth(bytePos, buf);

    VISA_EMask_Ctrl eMask = vISA_EMASK_M1;
    VISA_Exec_Size exSize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, exSize, eMask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    unsigned surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *offsets = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src0 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src1 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->AppendVISASurfAccessDwordAtomicInst(
        pred, subOpc, bitwidth == 16, eMask, exSize, surfaceHnd, offsets, src0,
        src1, dst);
    break;
  }
  case ISA_3D_TYPED_ATOMIC: {
    VISAAtomicOps subOpc;
    unsigned short bitwidth;
    std::tie(subOpc, bitwidth) = getAtomicOpAndBitwidth(bytePos, buf);

    VISA_EMask_Ctrl eMask = vISA_EMASK_M1;
    VISA_Exec_Size exSize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, exSize, eMask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    unsigned surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *u = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *v = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *r = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *lod = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src0 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src1 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->AppendVISA3dTypedAtomic(subOpc, bitwidth == 16, pred,
                                               eMask, exSize, surfaceHnd, u, v,
                                               r, lod, src0, src1, dst);
    break;
  }
  case ISA_QW_GATHER:
  case ISA_QW_SCATTER: {
    VISA_EMask_Ctrl eMask = vISA_EMASK_M1;
    VISA_Exec_Size exSize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, exSize, eMask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    VISA_SVM_Block_Num numBlocks =
        VISA_SVM_Block_Num(readPrimitiveOperandNG<uint8_t>(bytePos, buf));
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *offsets = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dstOrSrc = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = nullptr;
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);

    if (opcode == ISA_QW_GATHER) {
      kernelBuilderImpl->AppendVISAQwordGatherInst(
          pred, eMask, exSize, numBlocks, surfaceHnd, offsets, dstOrSrc);
    } else {
      kernelBuilderImpl->AppendVISAQwordScatterInst(
          pred, eMask, exSize, numBlocks, surfaceHnd, offsets, dstOrSrc);
    }
    break;
  }
  default: {
    vISA_ASSERT_UNREACHABLE("Unimplemented or Illegal DataPort Opcode.");
  }
  }
}

static void readInstructionControlFlow(unsigned &bytePos, const char *buf,
                                       ISA_Opcode opcode,
                                       RoutineContainer &container) {
  VISAKernel *kernelBuilder = container.kernelBuilder;

  VISA_EMask_Ctrl emask = vISA_EMASK_M1;
  VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;

  switch (opcode) {
  case ISA_SUBROUTINE:
  case ISA_LABEL: {
    uint16_t labelId = readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    VISA_LabelOpnd *label = container.labelVarDecls[labelId];
    kernelBuilder->AppendVISACFLabelInst(label);
    return;
  }
  case ISA_JMP:
  case ISA_RET:
  case ISA_CALL:
  case ISA_FRET:
  case ISA_FCALL: {
    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = hasPredicate(opcode)
                              ? readPredicateOperandNG(bytePos, buf, container)
                              : NULL;

    uint16_t labelId =
        (opcode == ISA_JMP || opcode == ISA_CALL || opcode == ISA_FCALL)
            ? readPrimitiveOperandNG<uint16_t>(bytePos, buf)
            : 0;

    if (opcode == ISA_FCALL) {
      uint8_t argSize = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
      uint8_t retSize = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
      kernelBuilder->AppendVISACFFunctionCallInst(
          pred, emask, esize, container.stringPool[labelId], argSize, retSize);
      return;
    }

    switch (opcode) {
    case ISA_JMP:
      kernelBuilder->AppendVISACFJmpInst(pred,
                                         container.labelVarDecls[labelId]);
      return;
    case ISA_CALL:
      kernelBuilder->AppendVISACFCallInst(pred, emask, esize,
                                          container.labelVarDecls[labelId]);
      return;

    case ISA_RET:
      kernelBuilder->AppendVISACFRetInst(pred, emask, esize);
      return;
    case ISA_FRET:
      kernelBuilder->AppendVISACFFunctionRetInst(pred, emask, esize);
      return;

    default:
      vISA_ASSERT_UNREACHABLE("Unimplemented or Illegal Control Flow Opcode.");
      return;
    }

    return;
  }
  case ISA_IFCALL: {
    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = hasPredicate(opcode)
                              ? readPredicateOperandNG(bytePos, buf, container)
                              : nullptr;

    bool isUniform = false;
    VISA_VectorOpnd *funcAddr =
        readVectorOperandNG(bytePos, buf, container, false);
    uint8_t argSize = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t retSize = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    kernelBuilder->AppendVISACFIndirectFuncCallInst(
        pred, emask, esize, isUniform, funcAddr, argSize, retSize);
    return;
  }
  case ISA_FADDR: {
    uint16_t sym_name_idx = readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    VISA_VectorOpnd *dst = readVectorOperandNG(bytePos, buf, container, true);
    kernelBuilder->AppendVISACFSymbolInst(container.stringPool[sym_name_idx],
                                          dst);
    return;
  }
  case ISA_SWITCHJMP: {
    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    uint8_t numLabels = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    vISA_ASSERT_INPUT(0 < numLabels && numLabels < 33,
                 "Number of labels in SWITCHJMP must be between 1 and 32.");

    VISA_VectorOpnd *index =
        readVectorOperandNG(bytePos, buf, container, false);

    VISA_LabelOpnd *labels[32]; /// 32 is max
    for (unsigned i = 0; i < numLabels; i++)
      labels[i] =
          container
              .labelVarDecls[readPrimitiveOperandNG<uint16_t>(bytePos, buf)];

    kernelBuilder->AppendVISACFSwitchJMPInst(index, numLabels, labels);
    break;
  }
  default:
    vISA_ASSERT_UNREACHABLE("Unimplemented or Illegal Control Flow Opcode.");
  }
}

static void readInstructionMisc(unsigned &bytePos, const char *buf,
                                ISA_Opcode opcode,
                                RoutineContainer &container) {
  VISAKernel *kernelBuilder = container.kernelBuilder;

  switch (opcode) {
  case ISA_FILE: {
    uint32_t versionInt =
        getVersionAsInt(container.majorVersion, container.minorVersion);
    bool is3Dot4Plus = versionInt >= getVersionAsInt(3, 4);
    uint32_t filenameIndex =
        is3Dot4Plus ? readPrimitiveOperandNG<uint32_t>(bytePos, buf)
                    : readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    const char *filename = container.stringPool[filenameIndex].c_str();
    kernelBuilder->AppendVISAMiscFileInst((char *)filename);
    break;
  }
  case ISA_LOC: {
    unsigned lineNumber = readPrimitiveOperandNG<unsigned>(bytePos, buf);
    kernelBuilder->AppendVISAMiscLOC(lineNumber);
    break;
  }
  case ISA_RAW_SEND: {
    uint8_t modifier = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

    unsigned exMsgDesc = readPrimitiveOperandNG<unsigned>(bytePos, buf);
    uint8_t numSrc = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t numDst = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_VectorOpnd *desc = readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *src = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    kernelBuilder->AppendVISAMiscRawSend(pred, emask, esize, modifier,
                                         exMsgDesc, numSrc, numDst, desc, src,
                                         dst);
    break;
  }
  case ISA_RAW_SENDS: {
    uint8_t modifier = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    bool hasEOT = modifier & 0x2;

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

    uint8_t numSrc0 = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t numSrc1 = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t numDst = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t ffid = 0;
    if (getVersionAsInt(container.majorVersion, container.minorVersion) >
        getVersionAsInt(3, 5)) {
      ffid = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    }
    VISA_VectorOpnd *exMsgDesc =
        readVectorOperandNG(bytePos, buf, container, false);

    VISA_VectorOpnd *desc = readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *src0 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src1 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    kernelBuilder->AppendVISAMiscRawSends(pred, emask, esize, modifier, ffid,
                                          exMsgDesc, numSrc0, numSrc1, numDst,
                                          desc, src0, src1, dst, hasEOT);
    break;
  }
  case ISA_VME_FBR: {
    VISA_RawOpnd *UNIInput = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *FBRInput = readRawOperandNG(bytePos, buf, container);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *FBRMbMode =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *FBRSubMbShape =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *FBRSubPredMode =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *output = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilder->AppendVISAMiscVME_FBR(surfaceHnd, UNIInput, FBRInput,
                                         FBRMbMode, FBRSubMbShape,
                                         FBRSubPredMode, output);
    break;
  }
  case ISA_VME_IME: {
    uint8_t streamMode = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t searchCtrl = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_RawOpnd *UNIInput = readRawOperandNG(bytePos, buf, container);

    VISA_RawOpnd *IMEInput = readRawOperandNG(bytePos, buf, container);

    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *ref0 = readRawOperandNG(bytePos, buf, container); //, 2);
    VISA_RawOpnd *ref1 = readRawOperandNG(bytePos, buf, container); //, 2);
    VISA_RawOpnd *costCenter =
        readRawOperandNG(bytePos, buf, container); //, 4);
    VISA_RawOpnd *output = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilder->AppendVISAMiscVME_IME(surfaceHnd, streamMode, searchCtrl,
                                         UNIInput, IMEInput, ref0, ref1,
                                         costCenter, output);
    break;
  }
  case ISA_VME_SIC: {
    VISA_RawOpnd *UNIInput = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *SICInput =
        readRawOperandNG(bytePos, buf, container); //, 128); //SICInput
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *output = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);

    kernelBuilder->AppendVISAMiscVME_SIC(surfaceHnd, UNIInput, SICInput,
                                         output);
    break;
  }
  case ISA_VME_IDM: {
    VISA_RawOpnd *UNIInput =
        readRawOperandNG(bytePos, buf, container); //, 4*32);
    VISA_RawOpnd *IDMInput =
        readRawOperandNG(bytePos, buf, container); //, 32); //DMInput
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *output = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);

    kernelBuilder->AppendVISAMiscVME_IDM(surfaceHnd, UNIInput, IDMInput,
                                         output);
    break;
  }
  case ISA_3D_URB_WRITE: {
    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

    uint8_t numOut = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_RawOpnd *channelMask = readRawOperandNG(bytePos, buf, container);
    uint16_t globalOffset = readPrimitiveOperandNG<uint16_t>(bytePos, buf);

    VISA_RawOpnd *urbHandle = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *perSlotOffset = readRawOperandNG(bytePos, buf, container);

    VISA_RawOpnd *vertexData = readRawOperandNG(bytePos, buf, container);

    kernelBuilder->AppendVISA3dURBWrite(pred, emask, esize, numOut, channelMask,
                                        globalOffset, urbHandle, perSlotOffset,
                                        vertexData);
    break;
  }
  case ISA_DPAS:
  case ISA_DPASW: {
    GenPrecision A = GenPrecision::INVALID, W = GenPrecision::INVALID;
    uint8_t D = 0;
    uint8_t C = 0;

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src0 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src1 = readRawOperandNG(bytePos, buf, container);

    uint8_t tag = 0;
    VISA_VectorOpnd *src2 = readVectorOperandNG(
        bytePos, buf, tag, container, Get_VISA_Exec_Size(esize), false);

    VISA_INST_Desc *inst_desc = &CISA_INST_table[opcode];

    // sanity check
    vISA_ASSERT_INPUT(inst_desc->opnd_desc[5].opnd_type == OPND_OTHER,
           "opnd_desc's type at index = 5 should be OPND_OTHER!");

    VISA_Type visatype = (VISA_Type)inst_desc->opnd_desc[5].data_type;
    uint32_t dpasOtherOpnd = readOtherOperandNG(bytePos, buf, visatype);
    UI32ToDpasInfo(dpasOtherOpnd, A, W, D, C);

    kernelBuilder->AppendVISADpasInst(opcode, emask, esize, dst, src0, src1,
                                      src2, A, W, D, C);
    break;
  }
  case ISA_LIFETIME: {
    VISA_VectorOpnd *opnd = NULL;
    Common_ISA_Operand_Class opndClass;
    VISAVarLifetime lifetime;

    unsigned char properties = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint32_t versionInt =
        getVersionAsInt(container.majorVersion, container.minorVersion);
    uint32_t varId = versionInt >= getVersionAsInt(3, 4)
                         ? readPrimitiveOperandNG<uint32_t>(bytePos, buf)
                         : readPrimitiveOperandNG<uint16_t>(bytePos, buf);

    opndClass = (Common_ISA_Operand_Class)(properties >> 4);
    lifetime = (VISAVarLifetime)(properties & 0x1);

    if (opndClass == OPERAND_GENERAL) {
      VISA_GenVar *decl;

      decl = container.generalVarDecls[varId];

      if (lifetime == LIFETIME_START) {
        kernelBuilder->CreateVISADstOperand(opnd, decl, 1, 0, 0);
      } else {
        kernelBuilder->CreateVISASrcOperand(opnd, decl, MODIFIER_NONE, 0, 1, 0,
                                            0, 0);
      }
    } else if (opndClass == OPERAND_ADDRESS) {
      VISA_AddrVar *decl;
      decl = container.addressVarDecls[varId];

      if (lifetime == LIFETIME_START) {
        kernelBuilder->CreateVISAAddressDstOperand(opnd, decl, 0);
      } else {
        kernelBuilder->CreateVISAAddressSrcOperand(opnd, decl, 0, 1);
      }
    } else if (opndClass == OPERAND_PREDICATE) {
      VISA_PredVar *decl;

      decl = container.predicateVarDecls[varId];
      VISA_PredOpnd *predOpnd;
      kernelBuilder->CreateVISAPredicateOperand(
          predOpnd, decl, PredState_NO_INVERSE, PRED_CTRL_NON);
      opnd = (VISA_VectorOpnd *)predOpnd;
    } else {
      vISA_ASSERT_UNREACHABLE("Unimplemented or Illegal Operand Class");
    }

    kernelBuilder->AppendVISALifetime(lifetime, opnd);
    break;
  }
  default: {
    vISA_ASSERT_UNREACHABLE("Unimplemented or Illegal Misc Opcode.");
  }
  }
}

static void readInstructionSVM(unsigned &bytePos, const char *buf,
                               ISA_Opcode opcode, RoutineContainer &container) {
  VISA_EMask_Ctrl emask = vISA_EMASK_M1;
  VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;

  VISAKernel *kernelBuilder = container.kernelBuilder;

  SVMSubOpcode subOpcode =
      (SVMSubOpcode)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  switch (subOpcode) {
  case SVM_BLOCK_ST:
  case SVM_BLOCK_LD: {
    uint8_t numOWords = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    bool unaligned = (numOWords & 8) != 0;
    numOWords &= 7;
    VISA_VectorOpnd *address =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    if (subOpcode == SVM_BLOCK_LD) {
      kernelBuilder->AppendVISASvmBlockLoadInst(VISA_Oword_Num(numOWords),
                                                unaligned, address, dst);
    } else {
      kernelBuilder->AppendVISASvmBlockStoreInst(VISA_Oword_Num(numOWords),
                                                 unaligned, address, dst);
    }
    break;
  }
  case SVM_GATHER:
  case SVM_SCATTER: {
    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

    VISA_SVM_Block_Type blockSize = VISA_SVM_Block_Type(
        readPrimitiveOperandNG<uint8_t>(bytePos, buf) & 0x3);
    VISA_SVM_Block_Num numBlocks =
        VISA_SVM_Block_Num(readPrimitiveOperandNG<uint8_t>(bytePos, buf) & 0x3);
    VISA_RawOpnd *addresses = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    if (subOpcode == SVM_GATHER) {
      kernelBuilder->AppendVISASvmGatherInst(pred, emask, esize, blockSize,
                                             numBlocks, addresses, dst);
    } else {
      kernelBuilder->AppendVISASvmScatterInst(pred, emask, esize, blockSize,
                                              numBlocks, addresses, dst);
    }
    break;
  }
  case SVM_ATOMIC: {
    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

    VISAAtomicOps op;
    unsigned short bitwidth;
    std::tie(op, bitwidth) = getAtomicOpAndBitwidth(bytePos, buf);

    VISA_RawOpnd *addresses = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src0 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *src1 = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    kernelBuilder->AppendVISASvmAtomicInst(pred, emask, esize, op, bitwidth,
                                           addresses, src0, src1, dst);
    break;
  }
  case SVM_GATHER4SCALED: {
    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    unsigned channelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    // scale is ignored and MBZ
    (void)readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    VISA_VectorOpnd *address =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *offsets = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
    kernelBuilder->AppendVISASvmGather4ScaledInst(
        pred, emask, esize,
        ChannelMask::createAPIFromBinary(ISA_SVM, channelMask), address,
        offsets, dst);
    break;
  }
  case SVM_SCATTER4SCALED: {
    readExecSizeNG(bytePos, buf, esize, emask, container);
    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    unsigned channelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    // scale is ignored and MBZ
    (void)readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    VISA_VectorOpnd *address =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *offsets = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
    kernelBuilder->AppendVISASvmScatter4ScaledInst(
        pred, emask, esize,
        ChannelMask::createAPIFromBinary(ISA_SVM, channelMask), address,
        offsets, dst);
    break;
  }
  default:
    vISA_ASSERT_UNREACHABLE("Unimplemented or Illegal SVM Sub Opcode.");
  }
}

template <class T>
static VISA3DSamplerOp readSubOpcodeByteNG(unsigned &bytePos, const char *buf) {
  T val = 0;
  READ_CISA_FIELD(val, T, bytePos, buf);
  return VISA3DSamplerOp::extractSamplerOp<T>(val);
}

static VISA_VectorOpnd *readAoffimmi(uint32_t &bytePos, const char *buf,
                                     RoutineContainer &container) {
  VISAKernel *kernelBuilder = container.kernelBuilder;
  uint32_t versionInt =
      getVersionAsInt(container.majorVersion, container.minorVersion);
  bool is3Dot4Plus = versionInt >= getVersionAsInt(3, 4);
  VISA_VectorOpnd *aoffimmi = nullptr;
  if (is3Dot4Plus) {
    aoffimmi = readVectorOperandNG(bytePos, buf, container, false);
  } else {
    uint16_t aoffimmiVal = readPrimitiveOperandNG<uint16_t>(bytePos, buf);
    kernelBuilder->CreateVISAImmediate(aoffimmi, &aoffimmiVal, ISA_TYPE_UW);
  }
  return aoffimmi;
}

static void readInstructionSampler(unsigned &bytePos, const char *buf,
                                   ISA_Opcode opcode,
                                   RoutineContainer &container) {
  VISAKernel *kernelBuilder = container.kernelBuilder;
  VISAKernelImpl *kernelBuilderImpl = ((VISAKernelImpl *)kernelBuilder);

  switch (opcode) {
  case ISA_AVS: {
    uint8_t chanelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISAChannelMask channel =
        ChannelMask::createAPIFromBinary(ISA_AVS, chanelMask);
    uint8_t sampler = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_VectorOpnd *uOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *vOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *deltaU =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *deltaV =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *u2d = readVectorOperandNG(bytePos, buf, container, false);

    VISA_VectorOpnd *groupID =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *verticalBlockNumber =
        readVectorOperandNG(bytePos, buf, container, false);
    uint8_t cntrl = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *v2d = readVectorOperandNG(bytePos, buf, container, false);
    uint8_t execMode = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *IEFBypass =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    VISA_StateOpndHandle *samplerHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilder->CreateVISAStateOperandHandle(
        samplerHnd, container.samplerVarDecls[sampler]);

    kernelBuilder->AppendVISAMEAVS(
        surfaceHnd, samplerHnd, channel, uOffset, vOffset, deltaU, deltaV, u2d,
        v2d, groupID, verticalBlockNumber, (OutputFormatControl)cntrl,
        (AVSExecMode)execMode, IEFBypass, dst);
    break;
  }
  case ISA_LOAD:
  case ISA_SAMPLE: {
    uint8_t mode = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t sampler = (ISA_SAMPLE == opcode)
                          ? readPrimitiveOperandNG<uint8_t>(bytePos, buf)
                          : 0;
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_RawOpnd *uOffset = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *vOffset = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *roffset = readRawOperandNG(bytePos, buf, container);
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    uint8_t channel = mode & 0xF;
    bool isSimd16 = 0x0 != ((mode >> 4) & 0x3);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    VISA_StateOpndHandle *samplerHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);

    if (opcode == ISA_SAMPLE) {
      kernelBuilderImpl->CreateVISAStateOperandHandle(
          samplerHnd, container.samplerVarDecls[sampler]);
      kernelBuilderImpl->AppendVISASISample(
          vISA_EMASK_M1, surfaceHnd, samplerHnd,
          ChannelMask::createAPIFromBinary(opcode, channel), isSimd16, uOffset,
          vOffset, roffset, dst);
    } else
      kernelBuilderImpl->AppendVISASILoad(
          surfaceHnd, ChannelMask::createAPIFromBinary(opcode, channel),
          isSimd16, uOffset, vOffset, roffset, dst);
    break;
  }
  case ISA_SAMPLE_UNORM: {
    uint8_t channelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t sampler = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    CHANNEL_OUTPUT_FORMAT channelOutput =
        (CHANNEL_OUTPUT_FORMAT)ChannelMask::getChannelOutputFormat(channelMask);

    VISA_VectorOpnd *uOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *vOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *deltaU =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *deltaV =
        readVectorOperandNG(bytePos, buf, container, false);

    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    VISA_StateOpndHandle *samplerHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        samplerHnd, container.samplerVarDecls[sampler]);
    kernelBuilderImpl->AppendVISASISampleUnorm(
        surfaceHnd, samplerHnd,
        ChannelMask::createAPIFromBinary(opcode, channelMask), uOffset, vOffset,
        deltaU, deltaV, dst, channelOutput);
    break;
  }
  case ISA_3D_SAMPLE: {
    // 0x6D <op> <pixel_null_mask> <cps_enable> <exec_size> <pred>
    // <channels> <aoffimmi> <sampler> <surface> <dst> <numParams> <params>
    // <paired_surface>
       // The vISA version that widens the opcode field is 4
       // the below check is necessary to pick the correct
       // width during decoding based on input binary version
    const int WIDE_OPCODE_ISA_VER = 4;
    auto op = (container.majorVersion >= WIDE_OPCODE_ISA_VER)
                  ? readSubOpcodeByteNG<uint16_t>(bytePos, buf)
                  : readSubOpcodeByteNG<uint8_t>(bytePos, buf);

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    uint8_t channelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *aoffimmi = readAoffimmi(bytePos, buf, container);
    uint8_t sampler = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
    bool isPairedSurfaceAvailable =
        getVersionAsInt(container.majorVersion, container.minorVersion) >=
        getVersionAsInt(3, 8);
    VISA_RawOpnd *pairedSurface =
        isPairedSurfaceAvailable ? readRawOperandNG(bytePos, buf, container)
                                 : nullptr;
    if (pairedSurface == nullptr) {
      kernelBuilder->CreateVISANullRawOperand(pairedSurface, false /*isDst*/);
    }

    uint8_t numParams = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    vISA_ASSERT_INPUT(numParams < 16,
                 "number of parameters for 3D_Sample should be < 16");

    VISA_RawOpnd *params[16];
    for (int i = 0; i < numParams; ++i) {
      params[i] = readRawOperandNG(bytePos, buf, container);
    }
    VISA_StateOpndHandle *surfaceHnd = NULL;
    VISA_StateOpndHandle *samplerHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        samplerHnd, container.samplerVarDecls[sampler]);
    kernelBuilderImpl->AppendVISA3dSampler(
        op.opcode, op.pixelNullMask, op.cpsEnable, !op.nonUniformSampler, pred,
        emask, esize, ChannelMask::createAPIFromBinary(opcode, channelMask),
        aoffimmi, samplerHnd, 0, surfaceHnd, 0, pairedSurface, dst, numParams,
        params);
    break;
  }
  case ISA_3D_LOAD: {
    // 0x6E <op> <pixel_null_mask> <exec_size> <pred> <channels>
    // <aoffimmi> <surface> <dst> <numParams> <params>
    // <paired_surface>
       // same as 3D_SAMPLE, except that sampler is missing.

    // The vISA version that widens the opcode field is 4
    // the below check is necessary to pick the correct
    // width during decoding based on input binary version
    const int WIDE_OPCODE_ISA_VER = 4;
    auto op = (container.majorVersion >= WIDE_OPCODE_ISA_VER)
                  ? readSubOpcodeByteNG<uint16_t>(bytePos, buf)
                  : readSubOpcodeByteNG<uint8_t>(bytePos, buf);

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
    uint8_t channelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *aoffimmi = readAoffimmi(bytePos, buf, container);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
    bool isPairedSurfaceAvailable =
        getVersionAsInt(container.majorVersion, container.minorVersion) >=
        getVersionAsInt(3, 8);
    VISA_RawOpnd *pairedSurface =
        isPairedSurfaceAvailable ? readRawOperandNG(bytePos, buf, container)
                                 : nullptr;
    if (pairedSurface == nullptr) {
      kernelBuilder->CreateVISANullRawOperand(pairedSurface, false /*isDst*/);
    }

    uint8_t numParams = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    vISA_ASSERT_INPUT(numParams < 16,
                 "number of parameters for 3D_Load should be < 16");

    VISA_RawOpnd *params[16];
    for (int i = 0; i < numParams; ++i) {
      params[i] = readRawOperandNG(bytePos, buf, container);
    }
    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->AppendVISA3dLoad(
        op.opcode, op.pixelNullMask, pred, emask, esize,
        ChannelMask::createAPIFromBinary(opcode, channelMask), aoffimmi,
        surfaceHnd, 0,
        pairedSurface,
        dst, numParams, params);
    break;
  }
  case ISA_3D_GATHER4: {
    // The vISA version that widens the opcode field is 4
    // the below check is necessary to pick the correct
    // width during decoding based on input binary version
    const int WIDE_OPCODE_ISA_VER = 4;
    auto op = (container.majorVersion >= WIDE_OPCODE_ISA_VER)
                  ? readSubOpcodeByteNG<uint16_t>(bytePos, buf)
                  : readSubOpcodeByteNG<uint8_t>(bytePos, buf);

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);

    VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

    uint8_t channel = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    VISA_VectorOpnd *aoffimmi = readAoffimmi(bytePos, buf, container);
    uint8_t sampler = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
    bool isPairedSurfaceAvailable =
        getVersionAsInt(container.majorVersion, container.minorVersion) >=
        getVersionAsInt(3, 8);
    VISA_RawOpnd *pairedSurface =
        isPairedSurfaceAvailable ? readRawOperandNG(bytePos, buf, container)
                                 : nullptr;
    if (pairedSurface == nullptr) {
      kernelBuilder->CreateVISANullRawOperand(pairedSurface, false /*isDst*/);
    }

    uint8_t numParams = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    vISA_ASSERT_INPUT(numParams < 8,
                 "number of parameters for 3D_Gather4 should be < 8");

    VISA_RawOpnd *params[16];
    for (int i = 0; i < numParams; ++i) {
      params[i] = readRawOperandNG(bytePos, buf, container);
    }
    VISA_StateOpndHandle *surfaceHnd = NULL;
    VISA_StateOpndHandle *samplerHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilderImpl->CreateVISAStateOperandHandle(
        samplerHnd, container.samplerVarDecls[sampler]);
    kernelBuilder->AppendVISA3dGather4(op.opcode, op.pixelNullMask, pred, emask,
                                       esize, (VISASourceSingleChannel)channel,
                                       aoffimmi, samplerHnd, 0, surfaceHnd, 0,
                                       pairedSurface,
                                       dst, numParams, params);
    break;
  }
  case ISA_3D_INFO: {
    VISASampler3DSubOpCode subOpcode =
        (VISASampler3DSubOpCode)readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_EMask_Ctrl emask = vISA_EMASK_M1;
    VISA_Exec_Size esize = EXEC_SIZE_ILLEGAL;
    readExecSizeNG(bytePos, buf, esize, emask, container);
    uint8_t channelMask = 0xF;
    channelMask = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_RawOpnd *lod = subOpcode == VISA_3D_RESINFO
                            ? readRawOperandNG(bytePos, buf, container)
                            : NULL;
    VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);
    kernelBuilder->AppendVISA3dInfo(
        subOpcode, emask, esize,
        ChannelMask::createAPIFromBinary(opcode, channelMask), surfaceHnd, 0, lod,
        dst);

    break;
  }
  case ISA_VA: {
    /// subOpcode
    ISA_VA_Sub_Opcode subOpcode =
        (ISA_VA_Sub_Opcode)readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    VISA_StateOpndHandle *surfaceHnd = NULL;
    VISA_StateOpndHandle *samplerHnd = NULL;
    uint8_t sampler = 0;

    switch (subOpcode) {
    case MINMAXFILTER_FOPCODE:
    case Convolve_FOPCODE:
    case Dilate_FOPCODE:
    case ERODE_FOPCODE: {
      sampler = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
      kernelBuilderImpl->CreateVISAStateOperandHandle(
          samplerHnd, container.samplerVarDecls[sampler]);
      break;
    }
    default:
      break; // Prevent gcc warning
    }

    uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    kernelBuilder->CreateVISAStateOperandHandle(
        surfaceHnd, container.surfaceVarDecls[surface]);

    VISA_VectorOpnd *uOffset =
        readVectorOperandNG(bytePos, buf, container, false);
    VISA_VectorOpnd *vOffset =
        readVectorOperandNG(bytePos, buf, container, false);

    switch (subOpcode) {
    case MINMAX_FOPCODE: {
      /// mmf mode
      VISA_VectorOpnd *mmfMode =
          readVectorOperandNG(bytePos, buf, container, false);

      /// dst
      VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
      kernelBuilder->AppendVISAVAMinMax(surfaceHnd, uOffset, vOffset, mmfMode,
                                        dst);
      break;
    }
    case MINMAXFILTER_FOPCODE: {

      /// cntrl
      OutputFormatControl cntrl =
          (OutputFormatControl)readPrimitiveOperandNG<uint8_t>(bytePos, buf);

      /// execMode
      MMFExecMode execMode =
          (MMFExecMode)readPrimitiveOperandNG<uint8_t>(bytePos, buf);

      /// mmf mode
      VISA_VectorOpnd *mmfMode =
          readVectorOperandNG(bytePos, buf, container, false);

      /// dst
      VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
      kernelBuilder->AppendVISAVAMinMaxFilter(samplerHnd, surfaceHnd, uOffset,
                                              vOffset, cntrl, execMode, mmfMode,
                                              dst);
      break;
    }
    case BoolCentroid_FOPCODE:
    case Centroid_FOPCODE: {
      /// v size
      VISA_VectorOpnd *vSize =
          readVectorOperandNG(bytePos, buf, container, false);

      if (subOpcode == BoolCentroid_FOPCODE) {
        /// h size
        VISA_VectorOpnd *hSize =
            readVectorOperandNG(bytePos, buf, container, false);
        VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
        kernelBuilder->AppendVISAVABooleanCentroid(surfaceHnd, uOffset, vOffset,
                                                   vSize, hSize, dst);
      } else {
        VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
        kernelBuilder->AppendVISAVACentroid(surfaceHnd, uOffset, vOffset, vSize,
                                            dst);
      }

      break;
    }
    case Convolve_FOPCODE: {
      /// size for convolve, execMode for erode/dilate
      uint8_t properties = readPrimitiveOperandNG<uint8_t>(bytePos, buf);

      bool isBigKernel = (bool)((properties >> 4) & 0x1);
      CONVExecMode execMode = (CONVExecMode)(properties & 0x3);

      /// dst
      VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
      kernelBuilder->AppendVISAVAConvolve(samplerHnd, surfaceHnd, uOffset,
                                          vOffset, execMode, isBigKernel, dst);
      break;
    }
    case Dilate_FOPCODE:
    case ERODE_FOPCODE: {
      /// size for convolve, execMode for erode/dilate
      EDExecMode execMode =
          (EDExecMode)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
      EDMode mode = (subOpcode == Dilate_FOPCODE) ? VA_DILATE : VA_ERODE;
      /// dst
      VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
      kernelBuilder->AppendVISAVAErodeDilate(mode, samplerHnd, surfaceHnd,
                                             uOffset, vOffset, execMode, dst);
      break;
    }
    default:
      break; // Prevent gcc warning
    }

    break;
  }
  case ISA_VA_SKL_PLUS: {
    VISA_INST_Desc *instDesc = &CISA_INST_table[opcode];

    /// subOpcode
    ISA_VA_Sub_Opcode subOpcode =
        (ISA_VA_Sub_Opcode)readPrimitiveOperandNG<uint8_t>(bytePos, buf);

    if ((subOpcode < VA_OP_CODE_1D_CONVOLVE_VERTICAL) ||
        (subOpcode >= VA_OP_CODE_UNDEFINED)) {
      vISA_ASSERT_INPUT(false, "Invalid VA sub-opcode");
      return;
    }

#define MAX_NUM_VOPNDS 10
    VISA_VectorOpnd *vOpnds[MAX_NUM_VOPNDS];
    uint8_t numVSrcs = 0;
    memset(vOpnds, 0, sizeof(VISA_VectorOpnd *) * MAX_NUM_VOPNDS);

#define MAX_NUM_MOPNDS 10
    unsigned int miscOpnds[MAX_NUM_MOPNDS];
    uint8_t numMiscOpnds = 0;
    memset(miscOpnds, 0, sizeof(uint32_t) * MAX_NUM_MOPNDS);

    VISA_RawOpnd *dst = NULL;

#define MAX_NUM_RSRCS 5
    VISA_RawOpnd *rawSrcs[MAX_NUM_RSRCS];
    uint8_t numRawSrcs = 0;
    memset(rawSrcs, 0, sizeof(VISA_RawOpnd *) * MAX_NUM_RSRCS);

#define MAX_NUM_SOPNDS 4
    VISA_StateOpndHandle *stateOpnds[MAX_NUM_SOPNDS];
    uint8_t numStateOpnds = 0;
    memset(stateOpnds, 0, sizeof(VISA_StateOpndHandle *) * MAX_NUM_SOPNDS);

    int numTotalOperands = instDesc->getSubInstDesc(subOpcode).opnd_num;

    for (int i = 0; i < numTotalOperands; i++) {
      const OpndDesc *opndDesc =
          &(instDesc->getSubInstDesc(subOpcode).opnd_desc[i]);

      if (opndDesc->opnd_type == OPND_SAMPLE) {
        uint8_t sampler = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
        kernelBuilderImpl->CreateVISAStateOperandHandle(
            stateOpnds[numStateOpnds++], container.samplerVarDecls[sampler]);
      } else if (opndDesc->opnd_type == OPND_SURFACE) {
        uint8_t surface = readPrimitiveOperandNG<uint8_t>(bytePos, buf);
        kernelBuilderImpl->CreateVISAStateOperandHandle(
            stateOpnds[numStateOpnds++], container.surfaceVarDecls[surface]);
      } else if ((opndDesc->opnd_type & OPND_SRC_GEN) == OPND_SRC_GEN) {
        vOpnds[numVSrcs++] =
            readVectorOperandNG(bytePos, buf, container, false);
      } else if (opndDesc->opnd_type == OPND_RAW_SRC) {
        rawSrcs[numRawSrcs++] = readRawOperandNG(bytePos, buf, container);
      } else if (opndDesc->opnd_type == OPND_RAW_DST) {
        dst = readRawOperandNG(bytePos, buf, container);
      } else if (opndDesc->opnd_type == OPND_OTHER) {
        // in theory this is not necessary since all of them will be UB
        // but to demonstrate usage model
        if (opndDesc->data_type == ISA_TYPE_UB) {
          miscOpnds[numMiscOpnds++] =
              readPrimitiveOperandNG<uint8_t>(bytePos, buf);
        } else if (opndDesc->data_type == ISA_TYPE_UW) {
          miscOpnds[numMiscOpnds++] =
              readPrimitiveOperandNG<uint16_t>(bytePos, buf);
        } else if (opndDesc->data_type == ISA_TYPE_UD) {
          miscOpnds[numMiscOpnds++] =
              readPrimitiveOperandNG<unsigned int>(bytePos, buf);
        } else {
          vISA_ASSERT_INPUT(false, "Invalid misc opnd data type");
          return;
        }
      } else {
        vISA_ASSERT_INPUT(false, "Invalid opnd type");
        return;
      }
    }
    numVSrcs = 0;
    numRawSrcs = 0;
    numMiscOpnds = 0;
    numStateOpnds = 0;

    if (subOpcode >= 8 && subOpcode <= 15 && subOpcode != 14)
      vASSERT(dst);

    switch (subOpcode) {
    case VA_OP_CODE_FLOOD_FILL:
      kernelBuilderImpl->AppendVISAVAFloodFill(
          miscOpnds[0] != 0, rawSrcs[0], vOpnds[0], vOpnds[1], vOpnds[2], dst);
      break;
    case VA_OP_CODE_1D_CONVOLVE_VERTICAL:
      kernelBuilderImpl->AppendVISAVAConvolve1D(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (CONVExecMode)miscOpnds[0], VA_V_DIRECTION, dst);
      break;
    case VA_OP_CODE_1D_CONVOLVE_HORIZONTAL:
      kernelBuilderImpl->AppendVISAVAConvolve1D(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (CONVExecMode)miscOpnds[0], VA_H_DIRECTION, dst);
      break;
    case VA_OP_CODE_1PIXEL_CONVOLVE:
      kernelBuilderImpl->AppendVISAVAConvolve1Pixel(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (CONV1PixelExecMode)miscOpnds[0], rawSrcs[0], dst);
      break;
    case VA_OP_CODE_LBP_CORRELATION:
      kernelBuilderImpl->AppendVISAVALBPCorrelation(stateOpnds[0], vOpnds[0],
                                                    vOpnds[1], vOpnds[2], dst);
      break;
    case VA_OP_CODE_LBP_CREATION:
      kernelBuilderImpl->AppendVISAVALBPCreation(
          stateOpnds[0], vOpnds[0], vOpnds[1], (LBPCreationMode)miscOpnds[0],
          dst);
      break;
    case VA_OP_CODE_CORRELATION_SEARCH:
      kernelBuilderImpl->AppendVISAVACorrelationSearch(
          stateOpnds[0], vOpnds[0], vOpnds[1], vOpnds[2], vOpnds[3], vOpnds[4],
          vOpnds[5], vOpnds[6], vOpnds[7], dst);
      break;
    case ISA_HDC_ERODE:
      kernelBuilderImpl->AppendVISAVAHDCErodeDilate(
          VA_ERODE, stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          stateOpnds[2], vOpnds[2], vOpnds[3]);
      break;
    case ISA_HDC_DILATE:
      kernelBuilderImpl->AppendVISAVAHDCErodeDilate(
          VA_DILATE, stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          stateOpnds[2], vOpnds[2], vOpnds[3]);
      break;
    case ISA_HDC_LBPCORRELATION:
      kernelBuilderImpl->AppendVISAVAHDCLBPCorrelation(
          stateOpnds[0], vOpnds[0], vOpnds[1], vOpnds[2], stateOpnds[1],
          vOpnds[3], vOpnds[4]);
      break;
    case ISA_HDC_LBPCREATION:
      kernelBuilderImpl->AppendVISAVAHDCLBPCreation(
          stateOpnds[0], vOpnds[0], vOpnds[1], (LBPCreationMode)miscOpnds[0],
          stateOpnds[1], vOpnds[2], vOpnds[3]);
      break;
    case ISA_HDC_MMF:
      kernelBuilderImpl->AppendVISAVAHDCMinMaxFilter(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (HDCReturnFormat)miscOpnds[0], (MMFEnableMode)miscOpnds[1],
          stateOpnds[2], vOpnds[2], vOpnds[3]);
      break;
    case ISA_HDC_1PIXELCONV:
      kernelBuilderImpl->AppendVISAVAHDCConvolve1Pixel(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (HDCReturnFormat)miscOpnds[0], rawSrcs[0], stateOpnds[2], vOpnds[2],
          vOpnds[3]);
      break;
    case ISA_HDC_CONV:
      kernelBuilderImpl->AppendVISAVAHDCConvolve(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (HDCReturnFormat)(miscOpnds[0] & 0xF),
          (CONVHDCRegionSize)(miscOpnds[0] >> 4), stateOpnds[2], vOpnds[2],
          vOpnds[3]);
      break;
    case ISA_HDC_1DCONV_H:
      kernelBuilderImpl->AppendVISAVAHDCConvolve1D(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (HDCReturnFormat)miscOpnds[0], VA_H_DIRECTION, stateOpnds[2],
          vOpnds[2], vOpnds[3]);
      break;
    case ISA_HDC_1DCONV_V:
      kernelBuilderImpl->AppendVISAVAHDCConvolve1D(
          stateOpnds[0], stateOpnds[1], vOpnds[0], vOpnds[1],
          (HDCReturnFormat)miscOpnds[0], VA_V_DIRECTION, stateOpnds[2],
          vOpnds[2], vOpnds[3]);
      break;
    default:
      vISA_ASSERT_UNREACHABLE("Invalid VA sub-opcode");
    }

    break;
  }
  default: {
    vISA_ASSERT_UNREACHABLE("Unimplemented or Illegal Sampler Opcode.");
  }
  }
}

static LSC_CACHE_OPTS readLscCacheOpts(unsigned &bytePos, const char *buf) {
  LSC_CACHE_OPTS caching{};
  caching.l1 = (LSC_CACHE_OPT)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  caching.l3 = (LSC_CACHE_OPT)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  return caching;
}

static LSC_ADDR readLscAddr(unsigned &bytePos, const char *buf) {
  LSC_ADDR addrInfo{};
  addrInfo.type = (LSC_ADDR_TYPE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  addrInfo.immScale = readPrimitiveOperandNG<uint16_t>(bytePos, buf);
  addrInfo.immOffset = readPrimitiveOperandNG<int32_t>(bytePos, buf);
  addrInfo.size = (LSC_ADDR_SIZE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  return addrInfo;
}

static LSC_DATA_SHAPE readLscDataShape(unsigned &bytePos, const char *buf,
                                       bool useChMask) {
  LSC_DATA_SHAPE dataInfo{};
  dataInfo.size = (LSC_DATA_SIZE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  dataInfo.order =
      (LSC_DATA_ORDER)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  // even though both chmask and elems are a sum type, the binary holds both
  auto dataElems =
      (LSC_DATA_ELEMS)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  int chMask = (int)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  if (useChMask) {
    dataInfo.chmask = chMask;
  } else {
    dataInfo.elems = dataElems;
  }
  return dataInfo;
}

static void readInstructionLscUntyped(LSC_OP subOpcode, unsigned &bytePos,
                                      const char *buf,
                                      RoutineContainer &container) {
  const LscOpInfo opInfo = LscOpInfoGet(subOpcode);

  VISA_EMask_Ctrl execMask = vISA_EMASK_M1;
  VISA_Exec_Size execSize = EXEC_SIZE_ILLEGAL;

  readExecSizeNG(bytePos, buf, execSize, execMask, container);
  VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

  auto lscSfid = (LSC_SFID)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  LSC_CACHE_OPTS caching = readLscCacheOpts(bytePos, buf);
  LSC_ADDR addrInfo = readLscAddr(bytePos, buf);
  LSC_DATA_SHAPE dataInfo = readLscDataShape(
      bytePos, buf, subOpcode == LSC_LOAD_QUAD || subOpcode == LSC_STORE_QUAD);

  VISA_VectorOpnd *surface =
      readVectorOperandNG(bytePos, buf, container, false);
  auto ssIdx = readPrimitiveOperandNG<uint32_t>(bytePos, buf);
  VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
  VISA_RawOpnd *src0 = readRawOperandNG(bytePos, buf, container);
  VISA_VectorOpnd *src0Pitch = nullptr;
  if (opInfo.isStrided()) {
    src0Pitch = readVectorOperandNG(bytePos, buf, container, false);
  }
  VISA_RawOpnd *src1 = readRawOperandNG(bytePos, buf, container);
  VISA_RawOpnd *src2 = nullptr;
  if (!opInfo.isStrided()) {
    src2 = readRawOperandNG(bytePos, buf, container);
  }

  if (opInfo.isStrided()) {
    container.kernelBuilder->AppendVISALscUntypedStridedInst(
        subOpcode, lscSfid, pred, execSize, execMask, caching, addrInfo,
        dataInfo, surface, ssIdx, dst, src0, src0Pitch, src1);
  } else {
    container.kernelBuilder->AppendVISALscUntypedInst(
        subOpcode, lscSfid, pred, execSize, execMask, caching, false, addrInfo,
        dataInfo, surface, ssIdx, dst, src0, src1, src2);
  }
}

static void readInstructionLscUntypedBlock2D(LSC_OP subOpcode,
                                             unsigned &bytePos, const char *buf,
                                             RoutineContainer &container) {
  VISA_EMask_Ctrl execMask = vISA_EMASK_M1;
  VISA_Exec_Size execSize = EXEC_SIZE_ILLEGAL;

  readExecSizeNG(bytePos, buf, execSize, execMask, container);
  VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
  //
  auto sfid = (LSC_SFID)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  //
  LSC_CACHE_OPTS caching = readLscCacheOpts(bytePos, buf);
  //
  LSC_DATA_SHAPE_BLOCK2D dataShape2D{};
  dataShape2D.size =
      (LSC_DATA_SIZE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  dataShape2D.order =
      (LSC_DATA_ORDER)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  dataShape2D.blocks = (int)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  dataShape2D.width = (int)readPrimitiveOperandNG<uint16_t>(bytePos, buf);
  dataShape2D.height = (int)readPrimitiveOperandNG<uint16_t>(bytePos, buf);
  dataShape2D.vnni = readPrimitiveOperandNG<uint8_t>(bytePos, buf) != 0;
  //
  VISA_RawOpnd *dstData = readRawOperandNG(bytePos, buf, container);
  VISA_VectorOpnd *src0Addrs[LSC_BLOCK2D_ADDR_PARAMS]{};
  for (size_t i = 0; i < LSC_BLOCK2D_ADDR_PARAMS; i++) {
    src0Addrs[i] = readVectorOperandNG(bytePos, buf, container, false);
  }
  VISA_RawOpnd *src1Data = readRawOperandNG(bytePos, buf, container);
  //
  container.kernelBuilder->AppendVISALscUntypedBlock2DInst(
      subOpcode, sfid, pred, execSize, execMask, caching, dataShape2D, dstData,
      src0Addrs, 0, 0, src1Data);
}

static void readInstructionLscTyped(LSC_OP subOpcode, unsigned &bytePos,
                                    const char *buf,
                                    RoutineContainer &container) {
  VISA_EMask_Ctrl execMask = vISA_EMASK_M1;
  VISA_Exec_Size execSize = EXEC_SIZE_ILLEGAL;

  readExecSizeNG(bytePos, buf, execSize, execMask, container);
  VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

  // LSC_SFID lscSfid = ... always TGM
  LSC_CACHE_OPTS caching = readLscCacheOpts(bytePos, buf);
  auto addrModel = (LSC_ADDR_TYPE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  auto addrSize = (LSC_ADDR_SIZE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  //
  auto dataSize = (LSC_DATA_SIZE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  auto dataOrder =
      (LSC_DATA_ORDER)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  auto dataElems =
      (LSC_DATA_ELEMS)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  auto chMask = (int)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  LSC_DATA_SHAPE dataInfo{};
  dataInfo.size = dataSize;
  dataInfo.order = dataOrder;
  if (subOpcode == LSC_LOAD_QUAD || subOpcode == LSC_STORE_QUAD) {
    dataInfo.chmask = chMask;
  } else {
    dataInfo.elems = dataElems;
  }

  VISA_VectorOpnd *surface =
      readVectorOperandNG(bytePos, buf, container, false);
  auto ssIdx = readPrimitiveOperandNG<uint32_t>(bytePos, buf);
  VISA_RawOpnd *dstData = readRawOperandNG(bytePos, buf, container);
  VISA_RawOpnd *src0AddrsU = readRawOperandNG(bytePos, buf, container);
  auto uOff = readPrimitiveOperandNG<int32_t>(bytePos, buf);
  VISA_RawOpnd *src0AddrsV = readRawOperandNG(bytePos, buf, container);
  auto vOff = readPrimitiveOperandNG<int32_t>(bytePos, buf);
  VISA_RawOpnd *src0AddrsR = readRawOperandNG(bytePos, buf, container);
  auto rOff = readPrimitiveOperandNG<int32_t>(bytePos, buf);
  VISA_RawOpnd *src0AddrsLOD = readRawOperandNG(bytePos, buf, container);
  VISA_RawOpnd *src1Data = readRawOperandNG(bytePos, buf, container);
  VISA_RawOpnd *src2Data = readRawOperandNG(bytePos, buf, container);

  container.kernelBuilder->AppendVISALscTypedInst(
      subOpcode, pred, execSize, execMask, caching, addrModel, addrSize,
      dataInfo,
      surface, ssIdx,
      dstData,
      src0AddrsU, uOff,
      src0AddrsV, vOff,
      src0AddrsR, rOff,
      src0AddrsLOD, src1Data, src2Data);
}
static void readInstructionLscFence(unsigned &bytePos, const char *buf,
                                    RoutineContainer &container) {
  VISA_EMask_Ctrl execMask = vISA_EMASK_M1;
  VISA_Exec_Size execSize = EXEC_SIZE_1;

  readExecSizeNG(bytePos, buf, execSize, execMask, container);
  (void)readPredicateOperandNG(bytePos, buf, container);
  LSC_SFID sfid = (LSC_SFID)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  LSC_FENCE_OP fenceOp =
      (LSC_FENCE_OP)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  LSC_SCOPE scope = (LSC_SCOPE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);

  container.kernelBuilder->AppendVISALscFence(sfid, fenceOp, scope);
}

static void readInstructionLscTypedBlock2D(LSC_OP subOpcode, unsigned &bytePos,
                                           const char *buf,
                                           RoutineContainer &container) {
  VISA_EMask_Ctrl execMask = vISA_EMASK_M1;
  VISA_Exec_Size execSize = EXEC_SIZE_ILLEGAL;
  readExecSizeNG(bytePos, buf, execSize, execMask, container);
  VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);
  (void)pred;

  LSC_CACHE_OPTS caching = readLscCacheOpts(bytePos, buf);

  auto addrModel = (LSC_ADDR_TYPE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);

  LSC_DATA_SHAPE_TYPED_BLOCK2D dataShape2D{};
  dataShape2D.width = (int)readPrimitiveOperandNG<uint16_t>(bytePos, buf);
  dataShape2D.height = (int)readPrimitiveOperandNG<uint16_t>(bytePos, buf);

  VISA_VectorOpnd *surface =
      readVectorOperandNG(bytePos, buf, container, false);
  unsigned surfaceIndex = readPrimitiveOperandNG<uint32_t>(bytePos, buf);

  VISA_RawOpnd *dstData = readRawOperandNG(bytePos, buf, container);

  VISA_VectorOpnd *xOffset =
      readVectorOperandNG(bytePos, buf, container, false);
  VISA_VectorOpnd *yOffset =
      readVectorOperandNG(bytePos, buf, container, false);

  VISA_RawOpnd *src1Data = readRawOperandNG(bytePos, buf, container);

  container.kernelBuilder->AppendVISALscTypedBlock2DInst(
      subOpcode, caching, addrModel, dataShape2D,
      surface, surfaceIndex,
      dstData,
      xOffset, yOffset, 0, 0, src1Data);
}

static void
readInstructionLscUntypedAppendCounterAtomic(LSC_OP subOpcode,
                                             unsigned &bytePos, const char *buf,
                                             RoutineContainer &container) {

  VISA_EMask_Ctrl execMask = vISA_EMASK_M1;
  VISA_Exec_Size execSize = EXEC_SIZE_ILLEGAL;

  readExecSizeNG(bytePos, buf, execSize, execMask, container);
  VISA_PredOpnd *pred = readPredicateOperandNG(bytePos, buf, container);

  auto lscSfid = (LSC_SFID)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  (void)lscSfid;

  LSC_CACHE_OPTS caching = readLscCacheOpts(bytePos, buf);
  auto addrModel = (LSC_ADDR_TYPE)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  LSC_DATA_SHAPE dataInfo = readLscDataShape(bytePos, buf, false);

  VISA_VectorOpnd *surface =
      readVectorOperandNG(bytePos, buf, container, false);
  auto surfaceIndex = readPrimitiveOperandNG<uint32_t>(bytePos, buf);
  VISA_RawOpnd *dst = readRawOperandNG(bytePos, buf, container);
  VISA_RawOpnd *srcData = readRawOperandNG(bytePos, buf, container);

  container.kernelBuilder->AppendVISALscUntypedAppendCounterAtomicInst(
      subOpcode, pred, execSize, execMask, caching, addrModel, dataInfo,
      surface, surfaceIndex, dst, srcData);
}

static void readInstructionLSC(unsigned &bytePos, const char *buf,
                               ISA_Opcode opcode, RoutineContainer &container) {
  if (opcode == ISA_Opcode::ISA_LSC_FENCE) {
    readInstructionLscFence(bytePos, buf, container);
  } else if (opcode == ISA_Opcode::ISA_LSC_UNTYPED) {
    auto subOpcode = (LSC_OP)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    if (subOpcode == LSC_LOAD_BLOCK2D || subOpcode == LSC_STORE_BLOCK2D) {
      readInstructionLscUntypedBlock2D(subOpcode, bytePos, buf, container);
    } else if (subOpcode == LSC_APNDCTR_ATOMIC_ADD ||
               subOpcode == LSC_APNDCTR_ATOMIC_SUB) {
      readInstructionLscUntypedAppendCounterAtomic(subOpcode, bytePos, buf,
                                                   container);
    } else {
      readInstructionLscUntyped(subOpcode, bytePos, buf, container);
    }
  } else if (opcode == ISA_Opcode::ISA_LSC_TYPED) {
    auto subOpcode = (LSC_OP)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
    if (subOpcode == LSC_LOAD_BLOCK2D || subOpcode == LSC_STORE_BLOCK2D) {
      readInstructionLscTypedBlock2D(subOpcode, bytePos, buf, container);
    } else {
      readInstructionLscTyped(subOpcode, bytePos, buf, container);
    }
  } else {
    vISA_ASSERT_INPUT(false, "invalid LSC op");
  }
}

void readInstructionNG(unsigned &bytePos, const char *buf,
                       RoutineContainer &container, unsigned instID) {
  ISA_Opcode opcode = (ISA_Opcode)readPrimitiveOperandNG<uint8_t>(bytePos, buf);
  vISA_ASSERT_INPUT(opcode < ISA_NUM_OPCODE,
               "Illegal or unimplemented CISA opcode.");

  // cout << "Opcode: " << ISA_Inst_Table[opcode].str << endl;

  switch (ISA_Inst_Table[opcode].type) {
  case ISA_Inst_Mov:
  case ISA_Inst_Sync:
  case ISA_Inst_Arith:
  case ISA_Inst_Logic:
  case ISA_Inst_Address:
  case ISA_Inst_Compare:
  case ISA_Inst_SIMD_Flow:
    readInstructionCommonNG(bytePos, buf, (ISA_Opcode)opcode, container);
    break;
  case ISA_Inst_Data_Port:
    readInstructionDataportNG(bytePos, buf, (ISA_Opcode)opcode, container);
    break;
  case ISA_Inst_Flow:
    readInstructionControlFlow(bytePos, buf, (ISA_Opcode)opcode, container);
    break;
  case ISA_Inst_Misc:
    readInstructionMisc(bytePos, buf, (ISA_Opcode)opcode, container);
    break;
  case ISA_Inst_SVM:
    readInstructionSVM(bytePos, buf, (ISA_Opcode)opcode, container);
    break;
  case ISA_Inst_Sampler:
    readInstructionSampler(bytePos, buf, (ISA_Opcode)opcode, container);
    break;
  case ISA_Inst_LSC:
    readInstructionLSC(bytePos, buf, (ISA_Opcode)opcode, container);
    break;
  default: {
    std::stringstream sstr;
    sstr << "Illegal or unimplemented ISA opcode " << ISA_Inst_Table[opcode].str
         << " (" << (unsigned)opcode << ")"
         << " at byte position " << bytePos - 1 << "(0x" << std::hex
         << bytePos - 1 << ")"
         << ".";
    vISA_ASSERT_UNREACHABLE(sstr.str());
  }
  }
}

static void readAttributesNG(uint8_t major, uint8_t minor, unsigned &bytePos,
                             const char *buf, kernel_format_t &header,
                             attribute_info_t *attributes, int numAttributes,
                             vISA::Mem_Manager &mem) {
  vISA_ASSERT_INPUT(buf != nullptr, "Argument Exception: argument buf    is NULL.");

  for (int i = 0; i < numAttributes; i++) {
    vISA_ASSERT_INPUT(attributes,
                "Argument Exception: argument 'attributes' is NULL");

    readVarBytes(major, minor, attributes[i].nameIndex, bytePos, buf);
    READ_CISA_FIELD(attributes[i].size, uint8_t, bytePos, buf);

    const char *attrName = header.strings[attributes[i].nameIndex];
    char *valueBuffer =
        (char *)mem.alloc(sizeof(char) * (attributes[i].size + 1));
    memcpy_s(valueBuffer, attributes[i].size, &buf[bytePos],
             attributes[i].size);
    bytePos += attributes[i].size;
    vISA::Attributes::ID attrID = vISA::Attributes::getAttributeID(attrName);
    if (vISA::Attributes::isInt32(attrID) || vISA::Attributes::isBool(attrID)) {
      attributes[i].isInt = true;
      switch (attributes[i].size) {
      case 0:
        // No value attribute, treat it as boolean internally.
        attributes[i].value.intVal = 1;
        break;
      case 1:
        attributes[i].value.intVal = *valueBuffer;
        break;
      case 2:
        attributes[i].value.intVal = *((short *)valueBuffer);
        break;
      case 4:
        attributes[i].value.intVal = *((int *)valueBuffer);
        break;
      default:
        vISA_ASSERT_UNREACHABLE("Unsupported attribute size.");
        break;
      }
    } else if (vISA::Attributes::isCStr(attrID)) {
      valueBuffer[attributes[i].size] = '\0';
      attributes[i].isInt =
          false; // by default assume attributes have string value
      attributes[i].value.stringVal = valueBuffer;
    } else {
      std::string errMsg(attrName);
      vISA_ASSERT_INPUT(false, "%s: unsupported attribute!", errMsg.c_str());
    }
  }
}

static std::string getDeclLabelString(const char *prefix, uint32_t index,
                                      kernel_format_t &header,
                                      VISA_Label_Kind kind) {
  if (index) {
    if (kind == LABEL_FC)
      return header.strings[index];
    std::stringstream sstr;
    sstr << header.strings[index] << "_" << index;
    return sstr.str();
  }
  return prefix;
}

static void addAllAttributesNG(kernel_format_t &Header, VISAKernelImpl *KBImpl,
                               CISA_GEN_VAR *GenVar, const uint32_t AttrsCount,
                               attribute_info_t *Attrs) {
  if (AttrsCount > 0) {
    KBImpl->resizeAttribute(GenVar, AttrsCount);

    for (unsigned ai = 0; ai < AttrsCount; ai++) {
      attribute_info_t *attribute = &Attrs[ai];
      KBImpl->AddAttributeToVarGeneric(
          GenVar, Header.strings[attribute->nameIndex], attribute->size,
          attribute->isInt ? (const char *)&attribute->value.intVal
                           : attribute->value.stringVal);
    }
  }
}

static void readRoutineNG(unsigned &bytePos, const char *buf,
                          vISA::Mem_Manager &mem, RoutineContainer &container) {
  kernel_format_t header;
  uint8_t majorVersion = container.majorVersion;
  uint8_t minorVersion = container.minorVersion;

  VISAKernelImpl *kernelBuilderImpl =
      ((VISAKernelImpl *)container.kernelBuilder);
  bool isKernel = kernelBuilderImpl->getIsKernel();

  unsigned kernelStart = bytePos;

  readVarBytes(majorVersion, minorVersion, header.string_count, bytePos, buf);
  header.strings =
      (const char **)mem.alloc(header.string_count * sizeof(char *));
  container.stringPool.resize(header.string_count);
  for (unsigned i = 0; i < header.string_count; i++) {
    char *str = (char *)mem.alloc(STRING_LEN);
    unsigned j = 0;
    while (buf[bytePos] != '\0' && j < STRING_LEN) {
      str[j++] = buf[bytePos++];
    }
    vISA_ASSERT_INPUT(j < STRING_LEN, "string exceeds the maximum length allowed");
    str[j] = '\0';
    bytePos++;
    header.strings[i] = str;
    container.stringPool[i] = str;
  }
  readVarBytes(majorVersion, minorVersion, header.name_index, bytePos, buf);

  /// read general variables
  unsigned numPreDefinedVars = Get_CISA_PreDefined_Var_Count();
  readVarBytes(majorVersion, minorVersion, header.variable_count, bytePos, buf);
  header.variables = (var_info_t *)mem.alloc(
      sizeof(var_info_t) * (header.variable_count + numPreDefinedVars));
  container.generalVarDecls = (VISA_GenVar **)mem.alloc(
      sizeof(VISA_GenVar *) * (header.variable_count + numPreDefinedVars));
  container.generalVarsCount = (header.variable_count + numPreDefinedVars);

  for (unsigned i = numPreDefinedVars;
       i < header.variable_count + numPreDefinedVars; i++) {
    unsigned declID = i;
    readVarBytes(majorVersion, minorVersion,
                 header.variables[declID].name_index, bytePos, buf);
    READ_CISA_FIELD(header.variables[declID].bit_properties, uint8_t, bytePos,
                    buf);
    READ_CISA_FIELD(header.variables[declID].num_elements, uint16_t, bytePos,
                    buf);
    readVarBytes(majorVersion, minorVersion,
                 header.variables[declID].alias_index, bytePos, buf);
    READ_CISA_FIELD(header.variables[declID].alias_offset, uint16_t, bytePos,
                    buf);

    READ_CISA_FIELD(header.variables[declID].alias_scope_specifier, uint8_t,
                    bytePos, buf);

    READ_CISA_FIELD(header.variables[declID].attribute_count, uint8_t, bytePos,
                    buf);

    header.variables[declID].attributes = (attribute_info_t *)mem.alloc(
        sizeof(attribute_info_t) * header.variables[declID].attribute_count);
    readAttributesNG(majorVersion, minorVersion, bytePos, buf, header,
                     header.variables[declID].attributes,
                     header.variables[declID].attribute_count, mem);
    header.variables[declID].dcl = NULL;

    /// VISA Builder Call
    var_info_t *var = &header.variables[declID];
    VISA_GenVar *decl = NULL;
    VISA_Type varType = (VISA_Type)((var->bit_properties) & 0xF);
    VISA_Align varAlign = (VISA_Align)((var->bit_properties >> 4) & 0xF);
    [[maybe_unused]] uint8_t aliasScopeSpecifier =
        header.variables[declID].alias_scope_specifier;
    [[maybe_unused]] int status = VISA_SUCCESS;

    vISA_ASSERT(aliasScopeSpecifier == 0,
           "file scope variables are no longer supported");

    {
      VISA_GenVar *parentDecl = NULL;
      uint16_t aliasOffset = 0;
      uint32_t aliasIndex = header.variables[declID].alias_index;
      if (aliasIndex > 0) {
        if (aliasIndex < numPreDefinedVars) {
          status = kernelBuilderImpl->GetPredefinedVar(
              parentDecl, (PreDefined_Vars)aliasIndex);
          vISA_ASSERT(status == VISA_SUCCESS,
                      "Invalid index for pre-defined variables");
        } else {
          parentDecl = container.generalVarDecls[aliasIndex];
        }
        aliasOffset = header.variables[declID].alias_offset;
      }

      status = kernelBuilderImpl->CreateVISAGenVar(
          decl, header.strings[var->name_index], var->num_elements, varType,
          varAlign, parentDecl, aliasOffset);
      vISA_ASSERT(VISA_SUCCESS == status,
                  "Failed to add VISA general variable.");
    }

    addAllAttributesNG(header, kernelBuilderImpl, decl, var->attribute_count,
                       var->attributes);
    container.generalVarDecls[i] = decl;
  }

  /// read address variables
  READ_CISA_FIELD(header.address_count, uint16_t, bytePos, buf);
  header.addresses =
      (addr_info_t *)mem.alloc(sizeof(addr_info_t) * header.address_count);
  container.addressVarDecls = (VISA_AddrVar **)mem.alloc(
      sizeof(VISA_AddrVar *) * (header.address_count));
  container.addressVarsCount = (header.address_count);
  for (unsigned i = 0; i < header.address_count; i++) {
    unsigned declID = i;
    readVarBytes(majorVersion, minorVersion,
                 header.addresses[declID].name_index, bytePos, buf);
    READ_CISA_FIELD(header.addresses[declID].num_elements, uint16_t, bytePos,
                    buf);
    READ_CISA_FIELD(header.addresses[declID].attribute_count, uint8_t, bytePos,
                    buf);
    header.addresses[declID].attributes = (attribute_info_t *)mem.alloc(
        sizeof(attribute_info_t) * header.addresses[declID].attribute_count);
    readAttributesNG(majorVersion, minorVersion, bytePos, buf, header,
                     header.addresses[declID].attributes,
                     header.addresses[declID].attribute_count, mem);
    header.addresses[declID].dcl = NULL;

    /// VISA Builder Call
    addr_info_t *var = &header.addresses[declID];
    VISA_AddrVar *decl = NULL;
    [[maybe_unused]] int status = kernelBuilderImpl->CreateVISAAddrVar(
        decl, header.strings[var->name_index], var->num_elements);
    vISA_ASSERT(VISA_SUCCESS == status, "Failed to add VISA address variable.");

    addAllAttributesNG(header, kernelBuilderImpl, decl, var->attribute_count,
                       var->attributes);
    container.addressVarDecls[i] = decl;
  }

  // read predicate variables
  READ_CISA_FIELD(header.predicate_count, uint16_t, bytePos, buf);
  header.predicates = (pred_info_t *)mem.alloc(
      sizeof(pred_info_t) *
      (header.predicate_count + COMMON_ISA_NUM_PREDEFINED_PRED));
  container.predicateVarDecls = (VISA_PredVar **)mem.alloc(
      sizeof(VISA_PredVar *) *
      (header.predicate_count + COMMON_ISA_NUM_PREDEFINED_PRED));
  container.predicateVarsCount =
      (header.predicate_count + COMMON_ISA_NUM_PREDEFINED_PRED);
  for (unsigned i = COMMON_ISA_NUM_PREDEFINED_PRED;
       i < (unsigned)(header.predicate_count + COMMON_ISA_NUM_PREDEFINED_PRED);
       i++) {
    unsigned declID = i;
    readVarBytes(majorVersion, minorVersion,
                 header.predicates[declID].name_index, bytePos, buf);
    READ_CISA_FIELD(header.predicates[declID].num_elements, uint16_t, bytePos,
                    buf);
    READ_CISA_FIELD(header.predicates[declID].attribute_count, uint8_t, bytePos,
                    buf);
    header.predicates[declID].attributes = (attribute_info_t *)mem.alloc(
        sizeof(attribute_info_t) * header.predicates[declID].attribute_count);
    readAttributesNG(majorVersion, minorVersion, bytePos, buf, header,
                     header.predicates[declID].attributes,
                     header.predicates[declID].attribute_count, mem);
    header.predicates[declID].dcl = NULL;

    /// VISA Builder Call
    pred_info_t *var = &header.predicates[declID];
    VISA_PredVar *decl = NULL;
    [[maybe_unused]] int status = kernelBuilderImpl->CreateVISAPredVar(
        decl, header.strings[var->name_index], var->num_elements);
    vISA_ASSERT(VISA_SUCCESS == status,
                "Failed to add VISA predicate vairable.");

    addAllAttributesNG(header, kernelBuilderImpl, decl, var->attribute_count,
                       var->attributes);
    container.predicateVarDecls[i] = decl;
  }

  // read label variables
  READ_CISA_FIELD(header.label_count, uint16_t, bytePos, buf);
  header.labels =
      (label_info_t *)mem.alloc(sizeof(label_info_t) * header.label_count);
  container.labelVarDecls = (VISA_LabelOpnd **)mem.alloc(
      sizeof(VISA_LabelOpnd *) * (header.label_count));
  container.labelVarsCount = header.label_count;
  for (unsigned i = 0; i < header.label_count; i++) {
    readVarBytes(majorVersion, minorVersion, header.labels[i].name_index,
                 bytePos, buf);
    READ_CISA_FIELD(header.labels[i].kind, uint8_t, bytePos, buf);
    READ_CISA_FIELD(header.labels[i].attribute_count, uint8_t, bytePos, buf);
    header.labels[i].attributes = (attribute_info_t *)mem.alloc(
        sizeof(attribute_info_t) * header.labels[i].attribute_count);
    readAttributesNG(majorVersion, minorVersion, bytePos, buf, header,
                     header.labels[i].attributes,
                     header.labels[i].attribute_count, mem);

    /// VISA Builder Call
    unsigned declID = i;
    label_info_t *var = &header.labels[declID];
    VISA_LabelOpnd *decl = NULL;
    [[maybe_unused]] int status = kernelBuilderImpl->CreateVISALabelVar(
        decl,
        getDeclLabelString("L", var->name_index, header,
                           VISA_Label_Kind(var->kind))
            .c_str(),
        VISA_Label_Kind(var->kind));
    vISA_ASSERT(VISA_SUCCESS == status, "Failed to add VISA label variable.");

    for (unsigned ai = 0; ai < var->attribute_count; ai++) {
      /// TODO: How to Add label decls and attributes correctly.
      /// kernelBuilderImpl->AddAttributeToVar(decl,
      /// header.strings[attribute->nameIndex], attribute->size,
      /// attribute->value.stringVal);
      vISA_ASSERT(false, "Currently the builder API does not support label "
                         "attributes. Please file a bug.");
    }

    container.labelVarDecls[i] = decl;
  }

  // read sampler variables
  READ_CISA_FIELD(header.sampler_count, uint8_t, bytePos, buf);
  // up to 31 pre-defined samplers are allowed
  vISA_ASSERT(header.sampler_count < COMMON_ISA_MAX_NUM_SAMPLERS,
               "number of vISA samplers exceeds the max");
  header.samplers = (state_info_t *)mem.alloc(sizeof(state_info_t) *
                                              COMMON_ISA_MAX_NUM_SAMPLERS);
  container.samplerVarDecls = (VISA_SamplerVar **)mem.alloc(
      sizeof(VISA_SamplerVar *) * COMMON_ISA_MAX_NUM_SAMPLERS);
  container.samplerVarsCount = header.sampler_count;
  for (unsigned i = 0; i < header.sampler_count; i++) {
    readVarBytes(majorVersion, minorVersion, header.samplers[i].name_index,
                 bytePos, buf);
    READ_CISA_FIELD(header.samplers[i].num_elements, uint16_t, bytePos, buf);
    READ_CISA_FIELD(header.samplers[i].attribute_count, uint8_t, bytePos, buf);
    header.samplers[i].attributes = (attribute_info_t *)mem.alloc(
        sizeof(attribute_info_t) * header.samplers[i].attribute_count);
    readAttributesNG(majorVersion, minorVersion, bytePos, buf, header,
                     header.samplers[i].attributes,
                     header.samplers[i].attribute_count, mem);

    /// VISA Builder Call
    unsigned declID = i;
    state_info_t *var = &header.samplers[declID];
    VISA_SamplerVar *decl = NULL;
    [[maybe_unused]] int status = kernelBuilderImpl->CreateVISASamplerVar(
        decl, header.strings[var->name_index], var->num_elements);
    vISA_ASSERT(VISA_SUCCESS == status, "Failed to add VISA sampler variable.");

    addAllAttributesNG(header, kernelBuilderImpl, decl, var->attribute_count,
                       var->attributes);
    container.samplerVarDecls[i] = decl;
  }
  // bindless sampler field
  constexpr int BINDLESS_SAMPLER_ID = 31;
  kernelBuilderImpl->GetBindlessSampler(
      container.samplerVarDecls[BINDLESS_SAMPLER_ID]);

  // read surface variables
  READ_CISA_FIELD(header.surface_count, uint8_t, bytePos, buf);
  unsigned num_pred_surf = Get_CISA_PreDefined_Surf_Count();
  header.surface_count += (uint8_t)num_pred_surf;
  header.surface_attrs = (bool *)mem.alloc(sizeof(bool) * header.surface_count);
  memset(header.surface_attrs, 0, sizeof(bool) * header.surface_count);
  header.surfaces =
      (state_info_t *)mem.alloc(sizeof(state_info_t) * header.surface_count);
  container.surfaceVarDecls = (VISA_SurfaceVar **)mem.alloc(
      sizeof(VISA_SurfaceVar *) * (header.surface_count));
  container.surfaceVarsCount = header.surface_count;

  /// Populate the predefined surfaces.
  for (unsigned i = 0; i < num_pred_surf; i++) {
    VISA_SurfaceVar *surfaceHnd = NULL;
    kernelBuilderImpl->GetPredefinedSurface(surfaceHnd, (PreDefined_Surface)i);
    container.surfaceVarDecls[i] = surfaceHnd;
  }

  /// Populate the rest of the surfaces.
  for (unsigned i = num_pred_surf; i < header.surface_count; i++) {
    readVarBytes(majorVersion, minorVersion, header.surfaces[i].name_index,
                 bytePos, buf);
    READ_CISA_FIELD(header.surfaces[i].num_elements, uint16_t, bytePos, buf);
    READ_CISA_FIELD(header.surfaces[i].attribute_count, uint8_t, bytePos, buf);
    header.surfaces[i].attributes = (attribute_info_t *)mem.alloc(
        sizeof(attribute_info_t) * header.surfaces[i].attribute_count);
    readAttributesNG(majorVersion, minorVersion, bytePos, buf, header,
                     header.surfaces[i].attributes,
                     header.surfaces[i].attribute_count, mem);

    /// VISA Builder Call
    unsigned declID = i;
    state_info_t *var = &header.surfaces[declID];
    VISA_SurfaceVar *decl = NULL;
    [[maybe_unused]] int status = kernelBuilderImpl->CreateVISASurfaceVar(
        decl, header.strings[var->name_index], var->num_elements);
    vISA_ASSERT(VISA_SUCCESS == status, "Failed to add VISA surface variable.");

    addAllAttributesNG(header, kernelBuilderImpl, decl, var->attribute_count,
                       var->attributes);

    for (unsigned ai = 0; ai < var->attribute_count; ai++) {
      attribute_info_t *attribute = &var->attributes[ai];

      /// TODO: Does this code even make sense anymore???
      const char *attrName = header.strings[attribute->nameIndex];
      if (Attributes::isAttribute(Attributes::ATTR_SurfaceUsage, attrName)) {
        header.surface_attrs[i] = (attribute->value.intVal == 2);
        break;
      }
    }

    container.surfaceVarDecls[i] = decl;
  }

  [[maybe_unused]] int vmeCount = 0;
  READ_CISA_FIELD(vmeCount, uint8_t, bytePos, buf);
  vISA_ASSERT(vmeCount == 0, "VME variable is no longer supported");
  header.vme_count = 0;

  // read input variables
  if (isKernel) {
    readVarBytes(container.majorVersion, container.minorVersion,
                 header.input_count, bytePos, buf, FIELD_TYPE::INPUT);

    header.inputs =
        (input_info_t *)mem.alloc(sizeof(input_info_t) * header.input_count);
    container.inputVarDecls = (CISA_GEN_VAR **)mem.alloc(
        sizeof(CISA_GEN_VAR *) * (header.input_count));
    container.inputVarsCount = header.input_count;
    for (unsigned i = 0; i < header.input_count; i++) {
      READ_CISA_FIELD(header.inputs[i].kind, uint8_t, bytePos, buf);
      readVarBytes(majorVersion, minorVersion, header.inputs[i].index, bytePos,
                   buf);
      READ_CISA_FIELD(header.inputs[i].offset, int16_t, bytePos, buf);
      READ_CISA_FIELD(header.inputs[i].size, uint16_t, bytePos, buf);

      unsigned declID = i;
      input_info_t *var = &header.inputs[declID];
      CISA_GEN_VAR *decl = NULL;

      switch (var->getInputClass()) {
      case INPUT_GENERAL:
        decl = container.generalVarDecls[var->index];
        break;
      case INPUT_SAMPLER:
        decl = container.samplerVarDecls[var->index];
        break;
      case INPUT_SURFACE:
        decl = container.surfaceVarDecls[var->index];
        break;
      default:
        vISA_ASSERT_UNREACHABLE("Incorrect input variable type.");
      }

      [[maybe_unused]] int status = kernelBuilderImpl->CreateVISAInputVar(
          decl, var->offset, var->size, var->getImplicitKind());
      vISA_ASSERT(VISA_SUCCESS == status, "Failed to add VISA input variable.");

      container.inputVarDecls[i] = decl;
    }
  }

  READ_CISA_FIELD(header.size, unsigned, bytePos, buf);
  READ_CISA_FIELD(header.entry, unsigned, bytePos, buf);

  if (!isKernel) {
    READ_CISA_FIELD(header.input_size, uint8_t, bytePos, buf);
    READ_CISA_FIELD(header.return_value_size, uint8_t, bytePos, buf);

    // Store size of arg/ret registers for stack call functions
    kernelBuilderImpl->setInputSize(header.input_size);
    kernelBuilderImpl->setReturnSize(header.return_value_size);
  }

  /// read kernel attributes
  READ_CISA_FIELD(header.attribute_count, uint16_t, bytePos, buf);
  header.attributes = (attribute_info_t *)mem.alloc(sizeof(attribute_info_t) *
                                                    header.attribute_count);
  readAttributesNG(majorVersion, minorVersion, bytePos, buf, header,
                   header.attributes, header.attribute_count, mem);

  for (unsigned ai = 0; ai < header.attribute_count; ai++) {
    attribute_info_t *attribute = &header.attributes[ai];
    /// TODO: This parameter ordering is inconsistent.
    kernelBuilderImpl->AddKernelAttribute(
        header.strings[attribute->nameIndex], attribute->size,
        attribute->isInt ? (char *)&attribute->value.intVal
                         : attribute->value.stringVal);
  }
  if (!kernelBuilderImpl->getKernelAttributes()->isKernelAttrSet(
          vISA::Attributes::ATTR_Target)) {
    VISATarget target = kernelBuilderImpl->getOptions()->getTarget();
    kernelBuilderImpl->AddKernelAttribute("Target", sizeof(target), &target);
  }

  unsigned kernelEntry = kernelStart + header.entry;
  unsigned kernelEnd = kernelEntry + header.size;

  bytePos = kernelEntry;

  for (unsigned i = 0; bytePos < kernelEnd; i++) {
    readInstructionNG(bytePos, buf, container, i);
  }
}

#define READ_FIELD_FROM_BUF(dst, type)                                         \
  dst = *((type *)&buf[byte_pos]);                                             \
  byte_pos += sizeof(type);

static int processCommonISAHeader(common_isa_header &cisaHdr,
                                  unsigned &byte_pos, const void *cisaBuffer,
                                  vISA::Mem_Manager *mem) {
  const char *buf = (const char *)cisaBuffer;
  READ_FIELD_FROM_BUF(cisaHdr.magic_number, uint32_t);
  READ_FIELD_FROM_BUF(cisaHdr.major_version, uint8_t);
  READ_FIELD_FROM_BUF(cisaHdr.minor_version, uint8_t);
  READ_FIELD_FROM_BUF(cisaHdr.num_kernels, uint16_t);

  vISA_ASSERT(cisaHdr.major_version >= 3,
              "only vISA version 3.0 and above are supported");

  if (cisaHdr.num_kernels) {
    cisaHdr.kernels = (kernel_info_t *)mem->alloc(sizeof(kernel_info_t) *
                                                  cisaHdr.num_kernels);
    vASSERT(cisaHdr.kernels != nullptr);
  } else {
    cisaHdr.kernels = NULL;
  }

  for (int i = 0; i < cisaHdr.num_kernels; i++) {
    if (cisaHdr.major_version == 3 && cisaHdr.minor_version < 7) {
      READ_FIELD_FROM_BUF(cisaHdr.kernels[i].name_len, uint8_t);
    } else {
      READ_FIELD_FROM_BUF(cisaHdr.kernels[i].name_len, uint16_t);
    }
    cisaHdr.kernels[i].name =
        (char *)mem->alloc(cisaHdr.kernels[i].name_len + 1);
    memcpy_s(cisaHdr.kernels[i].name,
             cisaHdr.kernels[i].name_len * sizeof(uint8_t), &buf[byte_pos],
             cisaHdr.kernels[i].name_len * sizeof(uint8_t));
    cisaHdr.kernels[i].name[cisaHdr.kernels[i].name_len] = '\0';
    byte_pos += cisaHdr.kernels[i].name_len;
    READ_FIELD_FROM_BUF(cisaHdr.kernels[i].offset, uint32_t);
    READ_FIELD_FROM_BUF(cisaHdr.kernels[i].size, uint32_t);
    READ_FIELD_FROM_BUF(cisaHdr.kernels[i].input_offset, uint32_t);

    READ_FIELD_FROM_BUF(cisaHdr.kernels[i].variable_reloc_symtab.num_syms,
                        uint16_t);
    vISA_ASSERT(cisaHdr.kernels[i].variable_reloc_symtab.num_syms == 0,
                "relocation symbols not allowed");
    cisaHdr.kernels[i].variable_reloc_symtab.reloc_syms = nullptr;

    READ_FIELD_FROM_BUF(cisaHdr.kernels[i].function_reloc_symtab.num_syms,
                        uint16_t);

    vISA_ASSERT(cisaHdr.kernels[i].function_reloc_symtab.num_syms == 0,
                "relocation symbols not allowed");
    cisaHdr.kernels[i].function_reloc_symtab.reloc_syms = nullptr;
    READ_FIELD_FROM_BUF(cisaHdr.kernels[i].num_gen_binaries, uint8_t);
    cisaHdr.kernels[i].gen_binaries = (gen_binary_info *)mem->alloc(
        cisaHdr.kernels[i].num_gen_binaries * sizeof(gen_binary_info));
    for (int j = 0; j < cisaHdr.kernels[i].num_gen_binaries; j++) {
      READ_FIELD_FROM_BUF(cisaHdr.kernels[i].gen_binaries[j].platform, uint8_t);
      READ_FIELD_FROM_BUF(cisaHdr.kernels[i].gen_binaries[j].binary_offset,
                          uint32_t);
      READ_FIELD_FROM_BUF(cisaHdr.kernels[i].gen_binaries[j].binary_size,
                          uint32_t);
    }

    cisaHdr.kernels[i].cisa_binary_buffer = NULL;
    cisaHdr.kernels[i].genx_binary_buffer = NULL;
  }

  READ_FIELD_FROM_BUF(cisaHdr.num_filescope_variables, uint16_t);
  vISA_ASSERT(cisaHdr.num_filescope_variables == 0,
              "file scope variables are no longer supported");

  READ_FIELD_FROM_BUF(cisaHdr.num_functions, uint16_t);

  if (cisaHdr.num_functions) {
    cisaHdr.functions = (function_info_t *)mem->alloc(sizeof(function_info_t) *
                                                      cisaHdr.num_functions);
    vASSERT(cisaHdr.functions != nullptr);
  } else {
    cisaHdr.functions = NULL;
  }

  for (int i = 0; i < cisaHdr.num_functions; i++) {
    // field is deprecated
    READ_FIELD_FROM_BUF(cisaHdr.functions[i].linkage, uint8_t);

    if (cisaHdr.major_version == 3 && cisaHdr.minor_version < 7) {
      READ_FIELD_FROM_BUF(cisaHdr.functions[i].name_len, uint8_t);
    } else {
      READ_FIELD_FROM_BUF(cisaHdr.functions[i].name_len, uint16_t);
    }
    cisaHdr.functions[i].name =
        (char *)mem->alloc(cisaHdr.functions[i].name_len + 1);
    memcpy_s(cisaHdr.functions[i].name,
             cisaHdr.functions[i].name_len * sizeof(uint8_t), &buf[byte_pos],
             cisaHdr.functions[i].name_len * sizeof(uint8_t));
    cisaHdr.functions[i].name[cisaHdr.functions[i].name_len] = '\0';
    byte_pos += cisaHdr.functions[i].name_len;
    READ_FIELD_FROM_BUF(cisaHdr.functions[i].offset, uint32_t);
    READ_FIELD_FROM_BUF(cisaHdr.functions[i].size, uint32_t);

    READ_FIELD_FROM_BUF(cisaHdr.functions[i].variable_reloc_symtab.num_syms,
                        uint16_t);
    vISA_ASSERT(cisaHdr.functions[i].variable_reloc_symtab.num_syms == 0,
                "variable relocation not supported");
    cisaHdr.functions[i].variable_reloc_symtab.reloc_syms = nullptr;

    READ_FIELD_FROM_BUF(cisaHdr.functions[i].function_reloc_symtab.num_syms,
                        uint16_t);
    vISA_ASSERT(cisaHdr.functions[i].function_reloc_symtab.num_syms == 0,
                "function relocation not supported");
    cisaHdr.functions[i].function_reloc_symtab.reloc_syms = nullptr;

    cisaHdr.functions[i].cisa_binary_buffer = NULL;
    cisaHdr.functions[i].genx_binary_buffer = NULL;
  }

  return 0;
}

//
// buf -- vISA binary to be processed.  For offline compile it's always the
// entire vISA object.
//     For JIT mode it's the entire isa file for 3.0, the kernel isa only
//     for 2.x
// builder -- the vISA builder
// kernels -- IR for the vISA kernel
//      if kernelName is specified, return that kernel only in kernels[0]
//      otherwise, all kernels in the isa are processed and returned in kernel
// kernelName -- name of the kernel to be processed.  If null, all kernels will
// be built majorVerion/minorVersion -- version of the vISA binary returns true
// if IR build succeeds, false otherwise
//
bool readIsaBinaryNG(const char *buf, CISA_IR_Builder *builder,
                     std::vector<VISAKernel *> &kernels, const char *kernelName,
                     unsigned int majorVersion, unsigned int minorVersion) {
  vISA_ASSERT(buf != nullptr, "Argument Exception: argument buf  is NULL.");

  unsigned bytePos = 0;
  vISA::Mem_Manager binaryReaderMem(4096);
  common_isa_header isaHeader;
  isaHeader.num_functions = 0;

  processCommonISAHeader(isaHeader, bytePos, buf, &binaryReaderMem);

  // we have to set the CISA builder version to the binary version,
  // or some instructions that behave differently based on vISA version (e.g.,
  // unaligned oword read) would not work correctly
  builder->CISA_IR_setVersion(isaHeader.major_version, isaHeader.minor_version);

  if (kernelName) {
    int kernelIndex = -1;
    for (unsigned i = 0; i < isaHeader.num_kernels; i++) {
      if (!strcmp(isaHeader.kernels[i].name, kernelName)) {
        kernelIndex = i;
        break;
      }
    }

    if (kernelIndex == -1) {
      return false;
    }

    bytePos = isaHeader.kernels[kernelIndex].offset;

    RoutineContainer container;
    container.builder = builder;
    container.kernelBuilder = NULL;
    container.majorVersion = isaHeader.major_version;
    container.minorVersion = isaHeader.minor_version;

    builder->AddKernel(container.kernelBuilder,
                       isaHeader.kernels[kernelIndex].name);
    kernels.push_back(container.kernelBuilder);

    readRoutineNG(bytePos, buf, binaryReaderMem, container);

    for (unsigned int i = 0; i < isaHeader.num_functions; i++) {
      bytePos = isaHeader.functions[i].offset;

      VISAFunction *funcPtr = NULL;
      builder->AddFunction(funcPtr, isaHeader.functions[i].name);

      container.kernelBuilder = (VISAKernel *)funcPtr;
      kernels.push_back(container.kernelBuilder);

      readRoutineNG(bytePos, buf, binaryReaderMem, container);
    }
  } else {
    for (unsigned int k = 0; k < isaHeader.num_kernels; k++) {
      bytePos = isaHeader.kernels[k].offset;

      RoutineContainer container;
      container.builder = builder;
      container.kernelBuilder = NULL;
      container.majorVersion = isaHeader.major_version;
      container.minorVersion = isaHeader.minor_version;

      builder->AddKernel(container.kernelBuilder, isaHeader.kernels[k].name);
      kernels.push_back(container.kernelBuilder);

      readRoutineNG(bytePos, buf, binaryReaderMem, container);
    }

    for (unsigned int i = 0; i < isaHeader.num_functions; i++) {
      RoutineContainer container;

      container.builder = builder;
      container.majorVersion = isaHeader.major_version;
      container.minorVersion = isaHeader.minor_version;

      bytePos = isaHeader.functions[i].offset;

      VISAFunction *funcPtr = NULL;
      builder->AddFunction(funcPtr, isaHeader.functions[i].name);

      container.kernelBuilder = (VISAKernel *)funcPtr;
      kernels.push_back(container.kernelBuilder);

      readRoutineNG(bytePos, buf, binaryReaderMem, container);
    }
  }

  return true;
}