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

Copyright (C) 2017-2021 Intel Corporation

SPDX-License-Identifier: MIT

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

extern __constant int __hasHWLocalThreadID;


#define MAX_DIM 2
#define BuiltinVector(BuiltinName) \
    (size_t3)(BuiltinName(0),      \
              BuiltinName(1),      \
              BuiltinName(2))

#ifdef NO_ASSUME_SUPPORT
    #define BuiltinAssumeGE0(s)
    #define BuiltinVectorAssumeGE0(v)
#else
    #define BuiltinAssumeGE0(s) __builtin_assume( s >= 0 )
    #define BuiltinVectorAssumeGE0(v) \
        __builtin_assume( (v.x) >= 0 ); \
        __builtin_assume( (v.y) >= 0 ); \
        __builtin_assume( (v.z) >= 0 );
#endif

// Helper functions prefixed with '__intel'

uint __intel_WorkgroupSize()
{
    uint totalWorkGroupSize =
        __builtin_IB_get_local_size(0) *
        __builtin_IB_get_local_size(1) *
        __builtin_IB_get_local_size(2);

    BuiltinAssumeGE0(totalWorkGroupSize);
    return totalWorkGroupSize;
}

size_t __intel_EnqueuedWorkgroupSize()
{
    size_t totalWorkGroupSize =
        __builtin_IB_get_enqueued_local_size(0) *
        __builtin_IB_get_enqueued_local_size(1) *
        __builtin_IB_get_enqueued_local_size(2);

    BuiltinAssumeGE0(totalWorkGroupSize);
    return totalWorkGroupSize;
}

size_t3 OVERLOADABLE __intel_WorkgroupId()
{
    size_t3 v = BuiltinVector(__builtin_IB_get_group_id);
    BuiltinVectorAssumeGE0(v);
    return v;
}

uint OVERLOADABLE __intel_LocalInvocationId(uint dim)
{
    uint v = 0;
    if (dim == 0) {
        v = __builtin_IB_get_local_id_x();
    }
    else if (dim == 1) {
        v = __builtin_IB_get_local_id_y();
    }
    else if (dim == 2) {
        v = __builtin_IB_get_local_id_z();
    }

    // local id is a 16 bit number in curbe.
#ifndef NO_ASSUME_SUPPORT
    __builtin_assume(v >= 0);
    __builtin_assume(v <= 0xffff);
#endif
    return v;
}

size_t3 OVERLOADABLE __intel_LocalInvocationId()
{
    size_t3 v = (size_t3)(__builtin_IB_get_local_id_x(),
                     __builtin_IB_get_local_id_y(),
                     __builtin_IB_get_local_id_z());

    BuiltinVectorAssumeGE0(v);
#ifndef NO_ASSUME_SUPPORT
    __builtin_assume(v.x <= 0xffff);
    __builtin_assume(v.y <= 0xffff);
    __builtin_assume(v.z <= 0xffff);
#endif

    return v;
}

size_t OVERLOADABLE __intel_GlobalInvocationId(uint dim)
{
    if (dim > MAX_DIM)
        return 0;

    size_t v =
        __builtin_IB_get_group_id(dim) * __builtin_IB_get_enqueued_local_size(dim) +
        __intel_LocalInvocationId(dim) + __builtin_IB_get_global_offset(dim);

#ifndef NO_ASSUME_SUPPORT
     BuiltinAssumeGE0(v);
     // We want to show, that the value is positive.
     // On LLVM level, where the signedness of the type is lost, the only way to prove
     // that the value is positive is to check the sign bit.
     __builtin_assume((v & 0x8000000000000000ULL) == 0);
#endif

    return v;
}

size_t3 OVERLOADABLE __intel_GlobalInvocationId()
{
    return BuiltinVector(__intel_GlobalInvocationId);
}

uint __intel_LocalInvocationIndex()
{
#if 0
    // This doesn't work right now due to a bug in the runtime.
    // If/when they fix their bug we can experiment if spending the
    // register(s) for get_local_linear_id() is better than spending
    // the math to compute the linear local ID.
    return __builtin_IB_get_local_linear_id();
#else
    uint llid;
    llid = (uint)__intel_LocalInvocationId(2);
    llid *= (uint)__builtin_IB_get_local_size(1);
    llid += (uint)__intel_LocalInvocationId(1);
    llid *= (uint)__builtin_IB_get_local_size(0);
    llid += (uint)__intel_LocalInvocationId(0);

    BuiltinAssumeGE0(llid);
    return llid;
#endif
}

////////////////////////

#if !defined(__USE_KHRONOS_SPIRV_TRANSLATOR__)

size_t3 __builtin_spirv_BuiltInNumWorkgroups()
{
    size_t3 v = BuiltinVector(__builtin_IB_get_num_groups);
    BuiltinVectorAssumeGE0(v);
    return v;
}

size_t3 __builtin_spirv_BuiltInWorkgroupSize()
{
   size_t3 v = BuiltinVector(__builtin_IB_get_local_size);
   BuiltinVectorAssumeGE0(v);
   return v;
}

size_t3 __builtin_spirv_BuiltInWorkgroupId()
{
    return __intel_WorkgroupId();
}

size_t3 __builtin_spirv_BuiltInLocalInvocationId()
{
    return __intel_LocalInvocationId();
}

size_t3 __builtin_spirv_BuiltInGlobalInvocationId()
{
    return __intel_GlobalInvocationId();
}

size_t3 __builtin_spirv_BuiltInGlobalSize()
{
    size_t3 v = BuiltinVector(__builtin_IB_get_global_size);
    BuiltinVectorAssumeGE0(v);
    return v;
}

size_t3 __builtin_spirv_BuiltInEnqueuedWorkgroupSize()
{
    size_t3 v = BuiltinVector(__builtin_IB_get_enqueued_local_size);
    BuiltinVectorAssumeGE0(v);
    return v;
}

size_t3 __builtin_spirv_BuiltInGlobalOffset()
{
    return BuiltinVector(__builtin_IB_get_global_offset);
}

#else // defined(__USE_KHRONOS_SPIRV_TRANSLATOR__)

size_t OVERLOADABLE __spirv_BuiltInNumWorkgroups(int dimindx)
{
    size_t v = __builtin_IB_get_num_groups(dimindx);
    BuiltinAssumeGE0(v);
    return v;
}

size_t OVERLOADABLE __spirv_BuiltInWorkgroupSize(int dimindx)
{
    size_t v = __builtin_IB_get_local_size(dimindx);
    BuiltinAssumeGE0(v);
    return v;
}

size_t OVERLOADABLE __spirv_BuiltInWorkgroupId(int dimindx)
{
    size_t v = __builtin_IB_get_group_id(dimindx);
    BuiltinAssumeGE0(v);
    return v;
}

size_t OVERLOADABLE __spirv_BuiltInLocalInvocationId(int dimindx)
{
    return __intel_LocalInvocationId(dimindx);
}

size_t OVERLOADABLE __spirv_BuiltInGlobalInvocationId(int dimindx)
{
    return __intel_GlobalInvocationId(dimindx);
}

size_t OVERLOADABLE __spirv_BuiltInGlobalSize(int dimindx)
{
    size_t v = __builtin_IB_get_global_size(dimindx);
    BuiltinAssumeGE0(v);
    return v;
}

size_t OVERLOADABLE __spirv_BuiltInEnqueuedWorkgroupSize(int dimindx)
{
    size_t v = __builtin_IB_get_enqueued_local_size(dimindx);
    BuiltinAssumeGE0(v);
    return v;
}

size_t OVERLOADABLE __spirv_BuiltInGlobalOffset(int dimindx)
{
    return __builtin_IB_get_global_offset(dimindx);
}

#endif

size_t SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInGlobalLinearId, , )()
{
  uint dim = SPIRV_BUILTIN_NO_OP(BuiltInWorkDim, , )();
  size_t result = 0;

  switch (dim) {
    default:
    case 1:
      result = __intel_GlobalInvocationId(0) - __builtin_IB_get_global_offset(0);
      break;
    case 2:
      result = (__intel_GlobalInvocationId(1) - __builtin_IB_get_global_offset(1))*
                __builtin_IB_get_global_size(0) + (__intel_GlobalInvocationId(0) - __builtin_IB_get_global_offset(0));
      break;
    case 3:
      result = ((__intel_GlobalInvocationId(2) - __builtin_IB_get_global_offset(2)) *
                __builtin_IB_get_global_size(1) * __builtin_IB_get_global_size(0)) +
               ((__intel_GlobalInvocationId(1) - __builtin_IB_get_global_offset(1)) * __builtin_IB_get_global_size(0)) +
               (__intel_GlobalInvocationId(0) - __builtin_IB_get_global_offset(0));
      break;
  }

  BuiltinAssumeGE0(result);
  return result;
}

size_t SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInLocalInvocationIndex, , )()
{
    return __intel_LocalInvocationIndex();
}

uint SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInWorkDim, , )()
{
    uint dim = __builtin_IB_get_work_dim();

#ifndef NO_ASSUME_SUPPORT
    __builtin_assume(dim >= 0);
    __builtin_assume(dim <= 3);
#endif

    return dim;
}

uint SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )()
{
    uint v = __builtin_IB_get_simd_size();

#ifndef NO_ASSUME_SUPPORT
    __builtin_assume(v >= 8);
    __builtin_assume(v <= 32);
#endif

    return v;
}

uint SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInSubgroupId, , )()
{
    if(__hasHWLocalThreadID)
    {
        return __builtin_IB_get_local_thread_id();
    }

    uint v = (uint)SPIRV_BUILTIN_NO_OP(BuiltInLocalInvocationIndex, , )() / SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )();

    BuiltinAssumeGE0(v);
    return v;
}

uint SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInNumSubgroups, , )()
{
    uint totalWorkGroupSize = __intel_WorkgroupSize() + SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )() - 1;
    return totalWorkGroupSize / SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )();
}

uint SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInSubgroupSize, , )()
{
    uint    remainder =
                __intel_WorkgroupSize() & ( SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )() - 1 );
    bool    fullSubGroup =
                ( remainder == 0 ) ||
                ( SPIRV_BUILTIN_NO_OP(BuiltInSubgroupId, , )() < SPIRV_BUILTIN_NO_OP(BuiltInNumSubgroups, , )() - 1 );

    return fullSubGroup ? SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )() : remainder;
}

uint SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInNumEnqueuedSubgroups, , )()
{
    uint totalEnqueuedWorkGroupSize = __intel_EnqueuedWorkgroupSize() + SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )() - 1;
    return totalEnqueuedWorkGroupSize / SPIRV_BUILTIN_NO_OP(BuiltInSubgroupMaxSize, , )();
}

uint SPIRV_OVERLOADABLE SPIRV_BUILTIN_NO_OP(BuiltInSubgroupLocalInvocationId, , )()
{
    uint simd_id = __builtin_IB_get_simd_id();

#ifndef NO_ASSUME_SUPPORT
    __builtin_assume(simd_id >= 0);
    __builtin_assume(simd_id < 32);
#endif

    return simd_id;
}