xref: /libCEED/backends/cuda-gen/ceed-cuda-gen-operator-build.cpp (revision fd364f385fcf8d7fc7b4a70eff195ce997d4a2a4)

// Copyright (c) 2017-2018, Lawrence Livermore National Security, LLC.
// Produced at the Lawrence Livermore National Laboratory. LLNL-CODE-734707.
// All Rights reserved. See files LICENSE and NOTICE for details.
//
// This file is part of CEED, a collection of benchmarks, miniapps, software
// libraries and APIs for efficient high-order finite element and spectral
// element discretizations for exascale applications. For more information and
// source code availability see http://github.com/ceed.
//
// The CEED research is supported by the Exascale Computing Project 17-SC-20-SC,
// a collaborative effort of two U.S. Department of Energy organizations (Office
// of Science and the National Nuclear Security Administration) responsible for
// the planning and preparation of a capable exascale ecosystem, including
// software, applications, hardware, advanced system engineering and early
// testbed platforms, in support of the nation's exascale computing imperative.
#include <ceed-backend.h>
#include "ceed-cuda-gen.h"
#include <iostream>
#include <sstream>
#include "../cuda/ceed-cuda.h"
#include "../cuda-reg/ceed-cuda-reg.h"
#include "../cuda-shared/ceed-cuda-shared.h"

static const char *atomicAdd = QUOTE(
//------------------------------------------------------------------------------
// Atomic add, for older CUDA
//------------------------------------------------------------------------------
__device__ double atomicAdd(double *address, double val) {
  unsigned long long int *address_as_ull = (unsigned long long int *)address;
  unsigned long long int old = *address_as_ull, assumed;
  do {
    assumed = old;
    old =
      atomicCAS(address_as_ull, assumed,
                __double_as_longlong(val +
                                     __longlong_as_double(assumed)));
    // Note: uses integer comparison to avoid hang in case of NaN
    // (since NaN != NaN)
  } while (assumed != old);
  return __longlong_as_double(old);
}
);

static const char *deviceFunctions = QUOTE(

//------------------------------------------------------------------------------
// Typedefs
//------------------------------------------------------------------------------
typedef struct { const CeedScalar* in[16]; CeedScalar* out[16]; } CudaFields;
typedef struct { CeedInt* in[16]; CeedInt* out[16]; } CudaFieldsInt;

typedef struct {
  CeedInt tidx;
  CeedInt tidy;
  CeedInt tidz;
  CeedInt tid;
  CeedScalar* slice;
} BackendData;

//------------------------------------------------------------------------------
// Load matrices for basis actions
//------------------------------------------------------------------------------
template <int P, int Q>
inline __device__ void loadMatrix(BackendData& data, const CeedScalar* d_B, CeedScalar* B) {
  for (CeedInt i = data.tid; i < P*Q; i += blockDim.x*blockDim.y*blockDim.z)
    B[i] = d_B[i];
}

//------------------------------------------------------------------------------
// 1D
//------------------------------------------------------------------------------

//------------------------------------------------------------------------------
// L-vector -> E-vector, offsets provided
//------------------------------------------------------------------------------
template <int NCOMP, int COMPSTRIDE, int P1d>
inline __device__ void readDofsOffset1d(BackendData& data, const CeedInt nnodes, const CeedInt elem, const CeedInt* indices, const CeedScalar* d_u, CeedScalar* r_u) {
  if (data.tidx < P1d) {
    const CeedInt node = data.tidx;
    const CeedInt ind = indices[node + elem * P1d];
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      r_u[comp] = d_u[ind + COMPSTRIDE * comp];
  }
}

//------------------------------------------------------------------------------
// L-vector -> E-vector, strided
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
inline __device__ void readDofsStrided1d(BackendData& data, const CeedInt elem, const CeedScalar* d_u, CeedScalar* r_u) {
  if (data.tidx < P1d) {
    const CeedInt node = data.tidx;
    const CeedInt ind = node * STRIDES_NODE + elem * STRIDES_ELEM;
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      r_u[comp] = d_u[ind + comp * STRIDES_COMP];
  }
}

//------------------------------------------------------------------------------
// E-vector -> L-vector, offsets provided
//------------------------------------------------------------------------------
template <int NCOMP, int COMPSTRIDE, int P1d>
inline __device__ void writeDofsOffset1d(BackendData& data, const CeedInt nnodes, const CeedInt elem, const CeedInt* indices, const CeedScalar* r_v, CeedScalar* d_v) {
  if (data.tidx < P1d) {
    const CeedInt node = data.tidx;
    const CeedInt ind = indices[node + elem * P1d];
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      atomicAdd(&d_v[ind + COMPSTRIDE * comp], r_v[comp]);
  }
}

//------------------------------------------------------------------------------
// E-vector -> L-vector, strided
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
inline __device__ void writeDofsStrided1d(BackendData& data, const CeedInt elem, const CeedScalar* r_v, CeedScalar* d_v) {
  if (data.tidx < P1d) {
    const CeedInt node = data.tidx;
    const CeedInt ind = node * STRIDES_NODE + elem * STRIDES_ELEM;
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      d_v[ind + comp * STRIDES_COMP] += r_v[comp];
  }
}

//------------------------------------------------------------------------------
// 1D tensor contraction x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractX1d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  data.slice[data.tidx] = *U;
  __syncthreads();
  *V = 0.0;
  for (CeedInt i = 0; i < P1d; ++i)
    *V += B[i + data.tidx*P1d] * data.slice[i]; // Contract x direction
  __syncthreads();
}

//------------------------------------------------------------------------------
// 1D transpose tensor contraction x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractTransposeX1d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  data.slice[data.tidx] = *U;
  __syncthreads();
  *V = 0.0;
  for (CeedInt i = 0; i < Q1d; ++i)
    *V += B[data.tidx + i*P1d] * data.slice[i]; // Contract x direction
  __syncthreads();
}

//------------------------------------------------------------------------------
// 1D interpolate to quadrature points
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void interp1d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, CeedScalar *__restrict__ r_V) {
  for (CeedInt comp = 0; comp < NCOMP; comp++)
    ContractX1d<NCOMP, P1d, Q1d>(data, r_U + comp, c_B, r_V + comp);
}

//------------------------------------------------------------------------------
// 1D interpolate transpose
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void interpTranspose1d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, CeedScalar *__restrict__ r_V) {
  for (CeedInt comp=0; comp<NCOMP; comp++)
    ContractTransposeX1d<NCOMP, P1d, Q1d>(data, r_U + comp, c_B, r_V + comp);
}

//------------------------------------------------------------------------------
// 1D derivatives at quadrature points
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void grad1d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  for (CeedInt comp = 0; comp < NCOMP; comp++)
    ContractX1d<NCOMP, P1d, Q1d>(data, r_U + comp, c_G, r_V + comp);
}

//------------------------------------------------------------------------------
// 1D derivatives transpose
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void gradTranspose1d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  for (CeedInt comp = 0; comp < NCOMP; comp++)
    ContractTransposeX1d<NCOMP, P1d, Q1d>(data, r_U + comp, c_G, r_V + comp);
}

//------------------------------------------------------------------------------
// 2D
//------------------------------------------------------------------------------

//------------------------------------------------------------------------------
// L-vector -> E-vector, offsets provided
//------------------------------------------------------------------------------
template <int NCOMP, int COMPSTRIDE, int P1d>
inline __device__ void readDofsOffset2d(BackendData& data, const CeedInt nnodes, const CeedInt elem, const CeedInt* indices, const CeedScalar* d_u, CeedScalar* r_u) {
  if (data.tidx < P1d && data.tidy < P1d) {
    const CeedInt node = data.tidx + data.tidy*P1d;
    const CeedInt ind = indices[node + elem * P1d*P1d];
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      r_u[comp] = d_u[ind + COMPSTRIDE * comp];
  }
}

//------------------------------------------------------------------------------
// L-vector -> E-vector, strided
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
inline __device__ void readDofsStrided2d(BackendData& data, const CeedInt elem, const CeedScalar* d_u, CeedScalar* r_u) {
  if (data.tidx < P1d && data.tidy < P1d) {
    const CeedInt node = data.tidx + data.tidy*P1d;
    const CeedInt ind = node * STRIDES_NODE + elem * STRIDES_ELEM;
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      r_u[comp] = d_u[ind + comp * STRIDES_COMP];
  }
}

//------------------------------------------------------------------------------
// E-vector -> L-vector, offsets provided
//------------------------------------------------------------------------------
template <int NCOMP, int COMPSTRIDE, int P1d>
inline __device__ void writeDofsOffset2d(BackendData& data, const CeedInt nnodes, const CeedInt elem, const CeedInt* indices, const CeedScalar* r_v, CeedScalar* d_v) {
  if (data.tidx < P1d && data.tidy < P1d) {
    const CeedInt node = data.tidx + data.tidy*P1d;
    const CeedInt ind = indices[node + elem * P1d*P1d];
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      atomicAdd(&d_v[ind + COMPSTRIDE * comp], r_v[comp]);
  }
}

//------------------------------------------------------------------------------
// E-vector -> L-vector, strided
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
inline __device__ void writeDofsStrided2d(BackendData& data, const CeedInt elem, const CeedScalar* r_v, CeedScalar* d_v) {
  if (data.tidx < P1d && data.tidy < P1d) {
    const CeedInt node = data.tidx + data.tidy*P1d;
    const CeedInt ind = node * STRIDES_NODE + elem * STRIDES_ELEM;
    for (CeedInt comp = 0; comp < NCOMP; ++comp)
      d_v[ind + comp * STRIDES_COMP] += r_v[comp];
  }
}

//------------------------------------------------------------------------------
// 2D tensor contraction x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractX2d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  data.slice[data.tidx+data.tidy*Q1d] = *U;
  __syncthreads();
  *V = 0.0;
  for (CeedInt i = 0; i < P1d; ++i)
    *V += B[i + data.tidx*P1d] * data.slice[i + data.tidy*Q1d]; // Contract x direction
  __syncthreads();
}

//------------------------------------------------------------------------------
// 2D tensor contract y
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractY2d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  data.slice[data.tidx+data.tidy*Q1d] = *U;
  __syncthreads();
  *V = 0.0;
  for (CeedInt i = 0; i < P1d; ++i)
    *V += B[i + data.tidy*P1d] * data.slice[data.tidx + i*Q1d]; // Contract y direction
  __syncthreads();
}

//------------------------------------------------------------------------------
// 2D transpose tensor contract y
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractYTranspose2d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  data.slice[data.tidx+data.tidy*Q1d] = *U;
  __syncthreads();
  *V = 0.0;
  if (data.tidy < P1d)
    for (CeedInt i = 0; i < Q1d; ++i)
      *V += B[data.tidy + i*P1d] * data.slice[data.tidx + i*Q1d]; // Contract y direction
  __syncthreads();
}

//------------------------------------------------------------------------------
// 2D transpose tensor contract x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractXTranspose2d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  data.slice[data.tidx+data.tidy*Q1d] = *U;
  __syncthreads();
  *V = 0.0;
  if (data.tidx < P1d)
    for (CeedInt i = 0; i < Q1d; ++i)
      *V += B[data.tidx + i*P1d] * data.slice[i + data.tidy*Q1d]; // Contract x direction
  __syncthreads();
}

//------------------------------------------------------------------------------
// 2D transpose tensor contract and add x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractXTransposeAdd2d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  data.slice[data.tidx+data.tidy*Q1d] = *U;
  __syncthreads();
  if (data.tidx < P1d)
    for (CeedInt i = 0; i < Q1d; ++i)
      *V += B[data.tidx + i*P1d] * data.slice[i + data.tidy*Q1d]; // Contract x direction
  __syncthreads();
}

//------------------------------------------------------------------------------
// 2D interpolate to quadrature points
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void interp2d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t[1];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractX2d<NCOMP, P1d, Q1d>(data, r_U + comp, c_B, r_t);
    ContractY2d<NCOMP, P1d, Q1d>(data, r_t, c_B, r_V + comp);
  }
}

//------------------------------------------------------------------------------
// 2D interpolate transpose
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void interpTranspose2d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t[1];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractYTranspose2d<NCOMP, P1d, Q1d>(data, r_U + comp, c_B, r_t);
    ContractXTranspose2d<NCOMP, P1d, Q1d>(data, r_t, c_B, r_V + comp);
  }
}

//------------------------------------------------------------------------------
// 2D derivatives at quadrature points
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void grad2d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t[1];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractX2d<NCOMP, P1d, Q1d>(data, r_U + comp, c_G, r_t);
    ContractY2d<NCOMP, P1d, Q1d>(data, r_t, c_B, r_V + comp + 0*NCOMP);
    ContractX2d<NCOMP, P1d, Q1d>(data, r_U + comp, c_B, r_t);
    ContractY2d<NCOMP, P1d, Q1d>(data, r_t, c_G, r_V + comp + 1*NCOMP);
  }
}

//------------------------------------------------------------------------------
// 2D derivatives transpose
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void gradTranspose2d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t[1];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractYTranspose2d<NCOMP, P1d, Q1d>(data, r_U + comp + 0*NCOMP, c_B, r_t);
    ContractXTranspose2d<NCOMP, P1d, Q1d>(data, r_t, c_G, r_V + comp);
    ContractYTranspose2d<NCOMP, P1d, Q1d>(data, r_U + comp + 1*NCOMP, c_G, r_t);
    ContractXTransposeAdd2d<NCOMP, P1d, Q1d>(data, r_t, c_B, r_V + comp);
  }
}

//------------------------------------------------------------------------------
// 3D
//------------------------------------------------------------------------------

//------------------------------------------------------------------------------
// L-vector -> E-vector, offsets provided
//------------------------------------------------------------------------------
template <int NCOMP, int COMPSTRIDE, int P1d>
inline __device__ void readDofsOffset3d(BackendData& data, const CeedInt nnodes, const CeedInt elem, const CeedInt* indices, const CeedScalar* d_u, CeedScalar* r_u) {
  if (data.tidx < P1d && data.tidy < P1d)
    for (CeedInt z = 0; z < P1d; ++z) {
      const CeedInt node = data.tidx + data.tidy*P1d + z*P1d*P1d;
      const CeedInt ind = indices[node + elem * P1d*P1d*P1d];
      for (CeedInt comp = 0; comp < NCOMP; ++comp)
        r_u[z+comp*P1d] = d_u[ind + COMPSTRIDE * comp];
    }
}

//------------------------------------------------------------------------------
// L-vector -> E-vector, strided
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
inline __device__ void readDofsStrided3d(BackendData& data, const CeedInt elem, const CeedScalar* d_u, CeedScalar* r_u) {
  if (data.tidx < P1d && data.tidy < P1d)
    for (CeedInt z = 0; z < P1d; ++z) {
      const CeedInt node = data.tidx + data.tidy*P1d + z*P1d*P1d;
      const CeedInt ind = node * STRIDES_NODE + elem * STRIDES_ELEM;
      for (CeedInt comp = 0; comp < NCOMP; ++comp)
        r_u[z+comp*P1d] = d_u[ind + comp * STRIDES_COMP];
    }
}

//------------------------------------------------------------------------------
// E-vector -> Q-vector, offests provided
//------------------------------------------------------------------------------
template <int NCOMP, int COMPSTRIDE, int Q1d>
inline __device__ void readSliceQuadsOffset3d(BackendData& data, const CeedInt nquads, const CeedInt elem, const CeedInt q, const CeedInt* indices, const CeedScalar* d_u, CeedScalar* r_u) {
  const CeedInt node = data.tidx + data.tidy*Q1d + q*Q1d*Q1d;
  const CeedInt ind = indices[node + elem * Q1d*Q1d*Q1d];;
  for (CeedInt comp = 0; comp < NCOMP; ++comp)
    r_u[comp] = d_u[ind + COMPSTRIDE * comp];
}

//------------------------------------------------------------------------------
// E-vector -> Q-vector, strided
//------------------------------------------------------------------------------
template <int NCOMP, int Q1d, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
inline __device__ void readSliceQuadsStrided3d(BackendData& data, const CeedInt elem, const CeedInt q, const CeedScalar* d_u, CeedScalar* r_u) {
  const CeedInt node = data.tidx + data.tidy*Q1d + q*Q1d*Q1d;
  const CeedInt ind = node * STRIDES_NODE + elem * STRIDES_ELEM;
  for (CeedInt comp = 0; comp < NCOMP; ++comp)
    r_u[comp] = d_u[ind + comp * STRIDES_COMP];
}

//------------------------------------------------------------------------------
// E-vector -> L-vector, offsets provided
//------------------------------------------------------------------------------
template <int NCOMP, int COMPSTRIDE, int P1d>
inline __device__ void writeDofsOffset3d(BackendData& data, const CeedInt nnodes, const CeedInt elem, const CeedInt* indices, const CeedScalar* r_v, CeedScalar* d_v) {
  if (data.tidx < P1d && data.tidy < P1d)
    for (CeedInt z = 0; z < P1d; ++z) {
      const CeedInt node = data.tidx + data.tidy*P1d + z*P1d*P1d;
      const CeedInt ind = indices[node + elem * P1d*P1d*P1d];
      for (CeedInt comp = 0; comp < NCOMP; ++comp)
        atomicAdd(&d_v[ind + COMPSTRIDE * comp], r_v[z+comp*P1d]);
    }
}

//------------------------------------------------------------------------------
// E-vector -> L-vector, strided
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int STRIDES_NODE, int STRIDES_COMP, int STRIDES_ELEM>
inline __device__ void writeDofsStrided3d(BackendData& data, const CeedInt elem, const CeedScalar* r_v, CeedScalar* d_v) {
  if (data.tidx < P1d && data.tidy < P1d)
    for (CeedInt z = 0; z < P1d; ++z) {
      const CeedInt node = data.tidx + data.tidy*P1d + z*P1d*P1d;
      const CeedInt ind = node * STRIDES_NODE + elem * STRIDES_ELEM;
      for (CeedInt comp = 0; comp < NCOMP; ++comp)
        d_v[ind + comp * STRIDES_COMP] += r_v[z+comp*P1d];
    }
}

//------------------------------------------------------------------------------
// 3D tensor contract x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractX3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  CeedScalar r_B[P1d];
  for (CeedInt i = 0; i < P1d; ++i)
    r_B[i] = B[i + data.tidx*P1d];

  for (CeedInt k = 0; k < P1d; ++k) {
    data.slice[data.tidx+data.tidy*Q1d] = U[k];
    __syncthreads();
    V[k] = 0.0;
    for (CeedInt i = 0; i < P1d; ++i)
      V[k] += r_B[i] * data.slice[i + data.tidy*Q1d]; // Contract x direction
    __syncthreads();
  }
}

//------------------------------------------------------------------------------
// 3D tensor contract y
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractY3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  CeedScalar r_B[P1d];
  for (CeedInt i = 0; i < P1d; ++i)
    r_B[i] = B[i + data.tidy*P1d];

  for (CeedInt k = 0; k < P1d; ++k) {
    data.slice[data.tidx+data.tidy*Q1d] = U[k];
    __syncthreads();
    V[k] = 0.0;
    for (CeedInt i = 0; i < P1d; ++i)
      V[k] += r_B[i] * data.slice[data.tidx + i*Q1d]; // Contract y direction
    __syncthreads();
  }
}

//------------------------------------------------------------------------------
// 3D tensor contract z
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractZ3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  for (CeedInt k = 0; k < Q1d; ++k) {
    V[k] = 0.0;
    for (CeedInt i = 0; i < P1d; ++i)
      V[k] += B[i + k*P1d] * U[i]; // Contract z direction
  }
}

//------------------------------------------------------------------------------
// 3D transpose tensor contract z
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractTransposeZ3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  for (CeedInt k = 0; k < Q1d; ++k) {
    V[k] = 0.0;
    if (k < P1d)
      for (CeedInt i = 0; i < Q1d; ++i)
        V[k] += B[k + i*P1d] * U[i]; // Contract z direction
  }
}

//------------------------------------------------------------------------------
// 3D transpose tensor contract y
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractTransposeY3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  CeedScalar r_B[Q1d];
  for (CeedInt i = 0; i < Q1d; ++i)
    r_B[i] = B[data.tidy + i*P1d];

  for (CeedInt k = 0; k < P1d; ++k) {
    data.slice[data.tidx+data.tidy*Q1d] = U[k];
    __syncthreads();
    V[k] = 0.0;
    if (data.tidy < P1d)
      for (CeedInt i = 0; i < Q1d; ++i)
        V[k] += r_B[i] * data.slice[data.tidx + i*Q1d]; // Contract y direction
    __syncthreads();
  }
}

//------------------------------------------------------------------------------
// 3D transpose tensor contract add y
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractTransposeAddY3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  CeedScalar r_B[Q1d];
  for (CeedInt i = 0; i < Q1d; ++i)
    r_B[i] = B[data.tidy + i*P1d];

  for (CeedInt k = 0; k < P1d; ++k) {
    data.slice[data.tidx+data.tidy*Q1d] = U[k];
    __syncthreads();
    if (data.tidy < P1d)
      for (CeedInt i = 0; i < Q1d; ++i)
        V[k] += r_B[i] * data.slice[data.tidx + i*Q1d]; // Contract y direction
    __syncthreads();
  }
}

//------------------------------------------------------------------------------
// 3D transpose tensor contract x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractTransposeX3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  CeedScalar r_B[Q1d];
  for (CeedInt i = 0; i < Q1d; ++i)
    r_B[i] = B[data.tidx + i*P1d];

  for (CeedInt k = 0; k < P1d; ++k) {
    data.slice[data.tidx+data.tidy*Q1d] = U[k];
    __syncthreads();
    V[k] = 0.0;
    if (data.tidx < P1d)
      for (CeedInt i = 0; i < Q1d; ++i)
        V[k] += r_B[i] * data.slice[i + data.tidy*Q1d]; // Contract x direction
    __syncthreads();
  }
}

//------------------------------------------------------------------------------
// 3D transpose tensor contract add x
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void ContractTransposeAddX3d(BackendData& data, const CeedScalar *U, const CeedScalar *B, CeedScalar *V) {
  CeedScalar r_B[Q1d];
  for (CeedInt i = 0; i < Q1d; ++i)
    r_B[i] = B[data.tidx + i*P1d];

  for (CeedInt k = 0; k < P1d; ++k) {
    data.slice[data.tidx+data.tidy*Q1d] = U[k];
    __syncthreads();
    if (data.tidx < P1d)
      for (CeedInt i = 0; i < Q1d; ++i)
        V[k] += r_B[i] * data.slice[i + data.tidy*Q1d]; // Contract x direction
    __syncthreads();
  }
}

//------------------------------------------------------------------------------
// 3D interpolate to quadrature points
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void interp3d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t1[Q1d];
  CeedScalar r_t2[Q1d];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractX3d<NCOMP, P1d, Q1d>(data, r_U + comp*P1d, c_B, r_t1);
    ContractY3d<NCOMP, P1d, Q1d>(data, r_t1, c_B, r_t2);
    ContractZ3d<NCOMP, P1d, Q1d>(data, r_t2, c_B, r_V + comp*Q1d);
  }
}

//------------------------------------------------------------------------------
// 3D interpolate transpose
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void interpTranspose3d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t1[Q1d];
  CeedScalar r_t2[Q1d];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractTransposeZ3d<NCOMP, P1d, Q1d>(data, r_U + comp*Q1d, c_B, r_t1);
    ContractTransposeY3d<NCOMP, P1d, Q1d>(data, r_t1, c_B, r_t2);
    ContractTransposeX3d<NCOMP, P1d, Q1d>(data, r_t2, c_B, r_V + comp*P1d);
  }
}

//------------------------------------------------------------------------------
// 3D derivatives at quadrature points
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void grad3d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t1[Q1d];
  CeedScalar r_t2[Q1d];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractX3d<NCOMP, P1d, Q1d>(data, r_U + comp*P1d, c_G, r_t1);
    ContractY3d<NCOMP, P1d, Q1d>(data, r_t1, c_B, r_t2);
    ContractZ3d<NCOMP, P1d, Q1d>(data, r_t2, c_B, r_V + comp*Q1d + 0*NCOMP*Q1d);
    ContractX3d<NCOMP, P1d, Q1d>(data, r_U + comp*P1d, c_B, r_t1);
    ContractY3d<NCOMP, P1d, Q1d>(data, r_t1, c_G, r_t2);
    ContractZ3d<NCOMP, P1d, Q1d>(data, r_t2, c_B, r_V + comp*Q1d + 1*NCOMP*Q1d);
    ContractX3d<NCOMP, P1d, Q1d>(data, r_U + comp*P1d, c_B, r_t1);
    ContractY3d<NCOMP, P1d, Q1d>(data, r_t1, c_B, r_t2);
    ContractZ3d<NCOMP, P1d, Q1d>(data, r_t2, c_G, r_V + comp*Q1d + 2*NCOMP*Q1d);
  }
}

//------------------------------------------------------------------------------
// 3D derivatives transpose
//------------------------------------------------------------------------------
template <int NCOMP, int P1d, int Q1d>
inline __device__ void gradTranspose3d(BackendData& data, const CeedScalar *__restrict__ r_U, const CeedScalar *c_B, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  CeedScalar r_t1[Q1d];
  CeedScalar r_t2[Q1d];
  for (CeedInt comp = 0; comp < NCOMP; comp++) {
    ContractTransposeZ3d<NCOMP, P1d, Q1d>(data, r_U + comp*Q1d + 0*NCOMP*Q1d, c_B, r_t1);
    ContractTransposeY3d<NCOMP, P1d, Q1d>(data, r_t1, c_B, r_t2);
    ContractTransposeX3d<NCOMP, P1d, Q1d>(data, r_t2, c_G, r_V + comp*P1d);
    ContractTransposeZ3d<NCOMP, P1d, Q1d>(data, r_U + comp*Q1d + 1*NCOMP*Q1d, c_B, r_t1);
    ContractTransposeY3d<NCOMP, P1d, Q1d>(data, r_t1, c_G, r_t2);
    ContractTransposeAddX3d<NCOMP,P1d, Q1d>(data, r_t2, c_B, r_V + comp*P1d);
    ContractTransposeZ3d<NCOMP, P1d, Q1d>(data, r_U + comp*Q1d + 2*NCOMP*Q1d, c_G, r_t1);
    ContractTransposeY3d<NCOMP, P1d, Q1d>(data, r_t1, c_B, r_t2);
    ContractTransposeAddX3d<NCOMP, P1d, Q1d>(data, r_t2, c_B, r_V + comp*P1d);
  }
}

//------------------------------------------------------------------------------
// 3D collocated derivatives computation
//------------------------------------------------------------------------------
template <int NCOMP, int Q1d>
inline __device__ void gradCollo3d(BackendData& data, const CeedInt q, const CeedScalar *__restrict__ r_U, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  for (CeedInt comp = 0; comp < NCOMP; ++comp) {
    data.slice[data.tidx + data.tidy*Q1d] = r_U[q + comp*Q1d];
    __syncthreads();
    // X derivative
    r_V[comp+0*NCOMP] = 0.0;
    for (CeedInt i = 0; i < Q1d; ++i)
      r_V[comp+0*NCOMP] += c_G[i + data.tidx*Q1d] * data.slice[i + data.tidy*Q1d]; // Contract x direction (X derivative)
    // Y derivative
    r_V[comp+1*NCOMP] = 0.0;
    for (CeedInt i = 0; i < Q1d; ++i)
      r_V[comp+1*NCOMP] += c_G[i + data.tidy*Q1d] * data.slice[data.tidx + i*Q1d]; // Contract y direction (Y derivative)
    // Z derivative
    r_V[comp+2*NCOMP] = 0.0;
    for (CeedInt i = 0; i < Q1d; ++i)
      r_V[comp+2*NCOMP] += c_G[i + q*Q1d] * r_U[i + comp*Q1d]; // Contract z direction (Z derivative)
    __syncthreads();
  }
}

//------------------------------------------------------------------------------
// 3D collocated derivatives transpose
//------------------------------------------------------------------------------
template <int NCOMP, int Q1d>
inline __device__ void gradColloTranspose3d(BackendData& data, const CeedInt q, const CeedScalar *__restrict__ r_U, const CeedScalar *c_G, CeedScalar *__restrict__ r_V) {
  for (CeedInt comp = 0; comp < NCOMP; ++comp) {
    // X derivative
    data.slice[data.tidx + data.tidy*Q1d] = r_U[comp + 0*NCOMP];
    __syncthreads();
    for (CeedInt i = 0; i < Q1d; ++i)
      r_V[q+comp*Q1d] += c_G[data.tidx + i*Q1d] * data.slice[i + data.tidy*Q1d]; // Contract x direction (X derivative)
    __syncthreads();
    // Y derivative
    data.slice[data.tidx + data.tidy*Q1d] = r_U[comp + 1*NCOMP];
    __syncthreads();
    for (CeedInt i = 0; i < Q1d; ++i)
      r_V[q+comp*Q1d] += c_G[data.tidy + i*Q1d] * data.slice[data.tidx + i*Q1d]; // Contract y direction (Y derivative)
    __syncthreads();
    // Z derivative
    for (CeedInt i = 0; i < Q1d; ++i)
      r_V[i+comp*Q1d] += c_G[i + q*Q1d] * r_U[comp + 2*NCOMP]; // PARTIAL contract z direction (Z derivative)
  }
}

//------------------------------------------------------------------------------
// 1D quadrature weights
//------------------------------------------------------------------------------
template <int Q1d>
inline __device__ void weight1d(BackendData& data, const CeedScalar *qweight1d, CeedScalar *w) {
  *w = qweight1d[data.tidx];
}

//------------------------------------------------------------------------------
// 2D quadrature weights
//------------------------------------------------------------------------------
template <int Q1d>
inline __device__ void weight2d(BackendData& data, const CeedScalar *qweight1d, CeedScalar *w) {
  *w = qweight1d[data.tidx]*qweight1d[data.tidy];
}

//------------------------------------------------------------------------------
// 3D quadrature weights
//------------------------------------------------------------------------------
template <int Q1d>
inline __device__ void weight3d(BackendData& data, const CeedScalar *qweight1d, CeedScalar *w) {
  const CeedScalar pw = qweight1d[data.tidx]*qweight1d[data.tidy];
  for (CeedInt z = 0; z < Q1d; ++z)
    w[z] = pw*qweight1d[z];
}

);
//------------------------------------------------------------------------------
// Build singe operator kernel
//------------------------------------------------------------------------------
extern "C" int CeedCudaGenOperatorBuild(CeedOperator op) {

  using std::ostringstream;
  using std::string;
  int ierr;
  bool setupdone;
  ierr = CeedOperatorIsSetupDone(op, &setupdone); CeedChk(ierr);
  if (setupdone) return 0;
  Ceed ceed;
  ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr);
  CeedOperator_Cuda_gen *data;
  ierr = CeedOperatorGetData(op, (void**)&data); CeedChk(ierr);
  CeedQFunction qf;
  CeedQFunction_Cuda_gen *qf_data;
  ierr = CeedOperatorGetQFunction(op, &qf); CeedChk(ierr);
  ierr = CeedQFunctionGetData(qf, (void **)&qf_data); CeedChk(ierr);
  CeedInt Q, P1d, Q1d = 0, numelements, elemsize, numinputfields,
          numoutputfields, ncomp, dim = 0, lsize;
  ierr = CeedOperatorGetNumQuadraturePoints(op, &Q); CeedChk(ierr);
  ierr = CeedOperatorGetNumElements(op, &numelements); CeedChk(ierr);
  ierr = CeedQFunctionGetNumArgs(qf, &numinputfields, &numoutputfields);
  CeedChk(ierr);
  CeedOperatorField *opinputfields, *opoutputfields;
  ierr = CeedOperatorGetFields(op, &opinputfields, &opoutputfields);
  CeedChk(ierr);
  CeedQFunctionField *qfinputfields, *qfoutputfields;
  ierr = CeedQFunctionGetFields(qf, &qfinputfields, &qfoutputfields);
  CeedChk(ierr);
  CeedEvalMode emode;
  CeedBasis basis;
  CeedBasis_Cuda_shared *basis_data;
  CeedElemRestriction Erestrict;
  CeedElemRestriction_Cuda_reg *restr_data;

  ostringstream code;
  string devFunctions(deviceFunctions);

  // Add atomicAdd function for old NVidia architectures
  struct cudaDeviceProp prop;
  Ceed_Cuda *ceed_data;
  ierr = CeedGetData(ceed, (void **)&ceed_data); CeedChk(ierr);
  ierr = cudaGetDeviceProperties(&prop, ceed_data->deviceId);
  if (prop.major<6){
    code << atomicAdd;
  }

  code << devFunctions;

  string qFunction(qf_data->qFunctionSource);

  code << "\n#define CEED_QFUNCTION(name) inline __device__ int name\n";
  code << "\n#define CeedPragmaSIMD\n";

  // Find dim and Q1d
  bool collograd = false;
  for (CeedInt i = 0; i < numinputfields; i++) {
    ierr = CeedOperatorFieldGetBasis(opinputfields[i], &basis); CeedChk(ierr);
    if (basis != CEED_BASIS_COLLOCATED) {
      ierr = CeedBasisGetData(basis, (void **)&basis_data); CeedChk(ierr);

      // Check for collocated gradient
      if (basis_data->d_collograd1d)
        collograd = true;

      // Collect dim and Q1d
      ierr = CeedBasisGetDimension(basis, &dim); CeedChk(ierr);
      bool isTensor;
      ierr = CeedBasisIsTensor(basis, &isTensor); CeedChk(ierr);
      if (isTensor) {
        ierr = CeedBasisGetNumQuadraturePoints1D(basis, &Q1d); CeedChk(ierr);
      } else {
        return CeedError(ceed, 1, "Backend does not implement operators with non-tensor basis");
        }
    }
  }
  // Check output bases for Q1d, dim as well
  //   The only imput basis might be CEED_BASIS_COLLOCATED
  for (CeedInt i = 0; i < numoutputfields; i++) {
    ierr = CeedOperatorFieldGetBasis(opoutputfields[i], &basis); CeedChk(ierr);
    if (basis != CEED_BASIS_COLLOCATED) {
      ierr = CeedBasisGetData(basis, (void **)&basis_data); CeedChk(ierr);
      // Collect dim and Q1d
      ierr = CeedBasisGetDimension(basis, &dim); CeedChk(ierr);
      bool isTensor;
      ierr = CeedBasisIsTensor(basis, &isTensor); CeedChk(ierr);
      if (isTensor) {
        ierr = CeedBasisGetNumQuadraturePoints1D(basis, &Q1d); CeedChk(ierr);
      } else {
        return CeedError(ceed, 1, "Backend does not implement operators with non-tensor basis");
        }
    }
  }
  data->dim = dim;
  data->Q1d = Q1d;

  // Define CEED_Q_VLA
  if (dim != 3 || collograd) {
    code << "\n#define CEED_Q_VLA 1\n\n";
  } else {
    code << "\n#define CEED_Q_VLA "<<Q1d<<"\n\n";
  }

  code << qFunction;

  // Setup
  code << "\nextern \"C\" __global__ void oper(CeedInt nelem, void* ctx, CudaFieldsInt indices, CudaFields fields, CudaFields B, CudaFields G, CeedScalar* W) {\n";
  // Input Evecs and Restriction
  for (CeedInt i = 0; i < numinputfields; i++) {
    ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode);
    CeedChk(ierr);
    if (emode != CEED_EVAL_WEIGHT) { // Skip CEED_EVAL_WEIGHT
      code << "const CeedScalar* d_u" <<i<<" = fields.in["<<i<<"];\n";
    }
  }

  for (CeedInt i = 0; i < numoutputfields; i++) {
    code << "CeedScalar* d_v"<<i<<" = fields.out["<<i<<"];\n";
  }

  code << "const CeedInt Dim = "<<dim<<";\n";
  code << "const CeedInt Q1d = "<<Q1d<<";\n";

  code << "extern __shared__ CeedScalar slice[];\n";
  code << "BackendData data;\n";
  code << "data.tidx = threadIdx.x;\n";
  code << "data.tidy = threadIdx.y;\n";
  code << "data.tidz = threadIdx.z;\n";
  code << "data.tid  = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.y*blockDim.x;\n";
  code << "data.slice = slice+data.tidz*Q1d"<<(dim>1?"*Q1d":"")<<";\n";

  code << "\n// Input field constants and basis data\n";
  //Initialize constants, and matrices B and G
  for (CeedInt i = 0; i < numinputfields; i++) {
    code << "// -- Input field "<<i<<" --\n";
    // Get elemsize, emode, ncomp
    ierr = CeedOperatorFieldGetElemRestriction(opinputfields[i], &Erestrict);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetElementSize(Erestrict, &elemsize);
    CeedChk(ierr);
    ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetNumComponents(Erestrict, &ncomp);
    CeedChk(ierr);

    // Set field constants
    if (emode != CEED_EVAL_WEIGHT) {
      ierr = CeedOperatorFieldGetBasis(opinputfields[i], &basis); CeedChk(ierr);
      if (basis != CEED_BASIS_COLLOCATED) {
        ierr = CeedBasisGetNumNodes1D(basis, &P1d); CeedChk(ierr);
        code << "  const CeedInt P_in_"<<i<<" = "<<P1d<<";\n";
      } else {
        code << "  const CeedInt P_in_"<<i<<" = "<<Q1d<<";\n";
      }
      code << "  const CeedInt ncomp_in_"<<i<<" = "<<ncomp<<";\n";
    }

    // Load basis data
    code << "// EvalMode: "<<CeedEvalModes[emode]<<"\n";
    switch (emode) {
    case CEED_EVAL_NONE:
      break;
    case CEED_EVAL_INTERP:
      ierr = CeedBasisGetData(basis, (void **)&basis_data); CeedChk(ierr);
      data->B.in[i] = basis_data->d_interp1d;
      code << "__shared__ double s_B_in_"<<i<<"["<<P1d*Q1d<<"];\n";
      code << "loadMatrix<P_in_"<<i<<",Q1d>(data, B.in["<<i<<"], s_B_in_"<<i<<");\n";
      break;
    case CEED_EVAL_GRAD:
      ierr = CeedBasisGetData(basis, (void **)&basis_data); CeedChk(ierr);
      data->B.in[i] = basis_data->d_interp1d;
      code << "__shared__ double s_B_in_"<<i<<"["<<P1d*Q1d<<"];\n";
      code << "loadMatrix<P_in_"<<i<<",Q1d>(data, B.in["<<i<<"], s_B_in_"<<i<<");\n";
      if (basis_data->d_collograd1d) {
        data->G.in[i] = basis_data->d_collograd1d;
        code << "__shared__ double s_G_in_"<<i<<"["<<Q1d*Q1d<<"];\n";
        code << "loadMatrix<Q1d,Q1d>(data, G.in["<<i<<"], s_G_in_"<<i<<");\n";
      } else {
        data->G.in[i] = basis_data->d_grad1d;
        code << "__shared__ double s_G_in_"<<i<<"["<<P1d*Q1d<<"];\n";
        code << "loadMatrix<P_in_"<<i<<",Q1d>(data, G.in["<<i<<"], s_G_in_"<<i<<");\n";
      }
      break;
    case CEED_EVAL_WEIGHT:
      break; // No action
    case CEED_EVAL_DIV:
      break; // TODO: Not implemented
    case CEED_EVAL_CURL:
      break; // TODO: Not implemented
    }
  }

  code << "\n// Output field constants and basis data\n";
  for (CeedInt i = 0; i < numoutputfields; i++) {
    code << "// -- Output field "<<i<<" --\n";
    // Get elemsize, emode, ncomp
    ierr = CeedOperatorFieldGetElemRestriction(opoutputfields[i], &Erestrict);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetElementSize(Erestrict, &elemsize);
    CeedChk(ierr);
    ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetNumComponents(Erestrict, &ncomp);
    CeedChk(ierr);

    // Set field constants
    ierr = CeedOperatorFieldGetBasis(opoutputfields[i], &basis); CeedChk(ierr);
    if (basis != CEED_BASIS_COLLOCATED) {
      ierr = CeedBasisGetNumNodes1D(basis, &P1d); CeedChk(ierr);
      code << "  const CeedInt P_out_"<<i<<" = "<<P1d<<";\n";
    } else {
      code << "  const CeedInt P_out_"<<i<<" = "<<Q1d<<";\n";
    }
    code << "  const CeedInt ncomp_out_"<<i<<" = "<<ncomp<<";\n";

    // Load basis data
    code << "// EvalMode: "<<CeedEvalModes[emode]<<"\n";
    switch (emode) {
    case CEED_EVAL_NONE:
      break; // No action
    case CEED_EVAL_INTERP:
      ierr = CeedBasisGetData(basis, (void **)&basis_data); CeedChk(ierr);
      data->B.out[i] = basis_data->d_interp1d;
      code << "  __shared__ double s_B_out_"<<i<<"["<<P1d*Q1d<<"];\n";
      code << "  loadMatrix<P_out_"<<i<<",Q1d>(data, B.out["<<i<<"], s_B_out_"<<i<<");\n";
      break;
    case CEED_EVAL_GRAD:
      ierr = CeedBasisGetData(basis, (void **)&basis_data); CeedChk(ierr);
      data->B.out[i] = basis_data->d_interp1d;
      code << "__shared__ double s_B_out_"<<i<<"["<<P1d*Q1d<<"];\n";
      code << "loadMatrix<P_out_"<<i<<",Q1d>(data, B.out["<<i<<"], s_B_out_"<<i<<");\n";
      if (basis_data->d_collograd1d) {
        data->G.out[i] = basis_data->d_collograd1d;
        code << "__shared__ double s_G_out_"<<i<<"["<<Q1d*Q1d<<"];\n";
        code << "loadMatrix<Q1d,Q1d>(data, G.out["<<i<<"], s_G_out_"<<i<<");\n";
      } else {
        data->G.out[i] = basis_data->d_grad1d;
        code << "__shared__ double s_G_out_"<<i<<"["<<P1d*Q1d<<"];\n";
        code << "loadMatrix<P_out_"<<i<<",Q1d>(data, G.out["<<i<<"], s_G_out_"<<i<<");\n";
      }
      break;
    case CEED_EVAL_WEIGHT: {
      Ceed ceed;
      ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr);
      return CeedError(ceed, 1,
                       "CEED_EVAL_WEIGHT cannot be an output evaluation mode");
      break; // Should not occur
    }
    case CEED_EVAL_DIV:
      break; // TODO: Not implemented
    case CEED_EVAL_CURL:
      break; // TODO: Not implemented
    }
  }
  code << "\n";
  code << "__syncthreads();\n";
  code << "for (CeedInt elem = blockIdx.x*blockDim.z + threadIdx.z; elem < nelem; elem += gridDim.x*blockDim.z) {\n";
  // Input basis apply if needed
  // Generate the correct eval mode code for each input
  code << "\n// Input field restrictions and basis actions\n";
  for (CeedInt i = 0; i < numinputfields; i++) {
    code << "  // -- Input field "<<i<<" --\n";
    // Get elemsize, emode, ncomp
    ierr = CeedOperatorFieldGetElemRestriction(opinputfields[i], &Erestrict);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetElementSize(Erestrict, &elemsize);
    CeedChk(ierr);
    ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetNumComponents(Erestrict, &ncomp);
    CeedChk(ierr);

    // Restriction
    if (emode != CEED_EVAL_WEIGHT &&
        !((emode == CEED_EVAL_NONE) && basis_data->d_collograd1d)) {
      code << "  CeedScalar r_u"<<i<<"[ncomp_in_"<<i<<"*P_in_"<<i<<"];\n";

      bool isStrided;
      ierr = CeedElemRestrictionIsStrided(Erestrict, &isStrided); CeedChk(ierr);
      if (!isStrided) {
        ierr = CeedElemRestrictionGetLVectorSize(Erestrict, &lsize);
        CeedChk(ierr);
        code << "  const CeedInt lsize_in_"<<i<<" = "<<lsize<<";\n";
        CeedInt compstride;
        ierr = CeedElemRestrictionGetCompStride(Erestrict, &compstride); CeedChk(ierr);
        code << "  // CompStride: "<<compstride<<"\n";
        ierr = CeedElemRestrictionGetData(Erestrict, (void **)&restr_data); CeedChk(ierr);
        data->indices.in[i] = restr_data->d_ind;
        code << "  readDofsOffset"<<dim<<"d<ncomp_in_"<<i<<", "<<compstride<<", P_in_"<<i<<">(data, lsize_in_"<<i<<", elem, indices.in["<<i<<"], d_u"<<i<<", r_u"<<i<<");\n";
      } else {
        bool backendstrides;
        ierr = CeedElemRestrictionHasBackendStrides(Erestrict, &backendstrides);
        CeedChk(ierr);
        CeedInt strides[3] = {1, elemsize, elemsize*ncomp};
        if (!backendstrides) {
          ierr = CeedElemRestrictionGetStrides(Erestrict, &strides);
          CeedChk(ierr);
        }
        code << "  // Strides: {"<<strides[0]<<", "<<strides[1]<<", "<<strides[2]<<"}\n";
        code << "  readDofsStrided"<<dim<<"d<ncomp_in_"<<i<<",P_in_"<<i<<","<<strides[0]<<","<<strides[1]<<","<<strides[2]<<">(data, elem, d_u"<<i<<", r_u"<<i<<");\n";
      }
    }

    // Basis action
    code << "// EvalMode: "<<CeedEvalModes[emode]<<"\n";
    switch (emode) {
    case CEED_EVAL_NONE:
      if (!basis_data->d_collograd1d) {
        code << "  CeedScalar* r_t"<<i<<" = r_u"<<i<<";\n";
      }
      break;
    case CEED_EVAL_INTERP:
      code << "  CeedScalar r_t"<<i<<"[ncomp_in_"<<i<<"*Q1d];\n";
      code << "  interp"<<dim<<"d<ncomp_in_"<<i<<",P_in_"<<i<<",Q1d>(data, r_u"<<i<<", s_B_in_"<<i<<", r_t"<<i<<");\n";
      break;
    case CEED_EVAL_GRAD:
      if (basis_data->d_collograd1d) {
        code << "  CeedScalar r_t"<<i<<"[ncomp_in_"<<i<<"*Q1d];\n";
        code << "  interp"<<dim<<"d<ncomp_in_"<<i<<",P_in_"<<i<<",Q1d>(data, r_u"<<i<<", s_B_in_"<<i<<", r_t"<<i<<");\n";
      } else {
        code << "  CeedScalar r_t"<<i<<"[ncomp_in_"<<i<<"*Dim*Q1d];\n";
        code << "  grad"<<dim<<"d<ncomp_in_"<<i<<",P_in_"<<i<<",Q1d>(data, r_u"<<i<<", s_B_in_"<<i<<", s_G_in_"<<i<<", r_t"<<i<<");\n";
      }
      break;
    case CEED_EVAL_WEIGHT:
      code << "  CeedScalar r_t"<<i<<"[Q1d];\n";
      ierr = CeedOperatorFieldGetBasis(opinputfields[i], &basis); CeedChk(ierr);
      ierr = CeedBasisGetData(basis, (void **)&basis_data); CeedChk(ierr);
      data->W = basis_data->d_qweight1d;
      code << "  weight"<<dim<<"d<Q1d>(data, W, r_t"<<i<<");\n";
      break; // No action
    case CEED_EVAL_DIV:
      break; // TODO: Not implemented
    case CEED_EVAL_CURL:
      break; // TODO: Not implemented
    }
  }

  // Q function
  code << "\n// Output field setup\n";
  for (CeedInt i = 0; i < numoutputfields; i++) {
      code << "  // -- Output field "<<i<<" --\n";
    ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode);
    CeedChk(ierr);
    if (emode==CEED_EVAL_GRAD)
    {
      if (basis_data->d_collograd1d) {
        //Accumulator for gradient slices
        code << "  CeedScalar r_tt"<<i<<"[ncomp_out_"<<i<<"*Q1d];\n";
        code << "  for (CeedInt i = 0; i < ncomp_out_"<<i<<"; ++i) {\n";
        code << "    for (CeedInt j = 0; j < Q1d; ++j) {\n";
        code << "      r_tt"<<i<<"[j + i*Q1d] = 0.0;\n";
        code << "    }\n";
        code << "  }\n";
      } else {
        code << "  CeedScalar r_tt"<<i<<"[ncomp_out_"<<i<<"*Dim*Q1d];\n";
      }
    }
    if (emode==CEED_EVAL_NONE || emode==CEED_EVAL_INTERP)
    {
      code << "  CeedScalar r_tt"<<i<<"[ncomp_out_"<<i<<"*Q1d];\n";
    }
  }
  // We treat quadrature points per slice in 3d to save registers
  if (basis_data->d_collograd1d) {
    code << "\n  // Note: Collocated Gradient\n";
    code << "#pragma unroll\n";
    code << "for (CeedInt q=0; q<Q1d; q++) {\n";
    for (CeedInt i = 0; i < numinputfields; i++) {
      code << "  // -- Input field "<<i<<" --\n";
      // Get elemsize, emode, ncomp
      ierr = CeedQFunctionFieldGetEvalMode(qfinputfields[i], &emode);
      CeedChk(ierr);
      // Basis action
      code << "// EvalMode: "<<CeedEvalModes[emode]<<"\n";
      switch (emode) {
      case CEED_EVAL_NONE:
        code << "  CeedScalar r_q"<<i<<"[ncomp_in_"<<i<<"];\n";

        bool isStrided;
        ierr = CeedElemRestrictionIsStrided(Erestrict, &isStrided); CeedChk(ierr);
        if (!isStrided) {
          ierr = CeedElemRestrictionGetLVectorSize(Erestrict, &lsize);
          CeedChk(ierr);
          code << "  const CeedInt lsize_in_"<<i<<" = "<<lsize<<";\n";
          CeedInt compstride;
          ierr = CeedElemRestrictionGetCompStride(Erestrict, &compstride); CeedChk(ierr);
          code << "  // CompStride: "<<compstride<<"\n";
          ierr = CeedElemRestrictionGetData(Erestrict, (void **)&restr_data); CeedChk(ierr);
          data->indices.in[i] = restr_data->d_ind;
          code << "  readSliceQuadsOffset"<<"3d<ncomp_in_"<<i<<", "<<compstride<<", Q1d>(data, lsize_in_"<<i<<", elem, q, indices.in["<<i<<"], d_u"<<i<<", r_q"<<i<<");\n";
        } else {
          bool backendstrides;
          ierr = CeedElemRestrictionHasBackendStrides(Erestrict, &backendstrides);
          CeedChk(ierr);
          CeedInt strides[3] = {1, elemsize, elemsize*ncomp};
          if (!backendstrides) {
            ierr = CeedElemRestrictionGetStrides(Erestrict, &strides);
            CeedChk(ierr);
          }
          code << "  // Strides: {"<<strides[0]<<", "<<strides[1]<<", "<<strides[2]<<"}\n";
          code << "  readSliceQuadsStrided"<<"3d<ncomp_in_"<<i<<",Q1d"","<<strides[0]<<","<<strides[1]<<","<<strides[2]<<">(data, elem, q, d_u"<<i<<", r_q"<<i<<");\n";
        }
        break;
      case CEED_EVAL_INTERP:
        code << "  CeedScalar r_q"<<i<<"[ncomp_in_"<<i<<"];\n";
        code << "  for (CeedInt j = 0; j < ncomp_in_"<<i<<" ; ++j) {\n";
        code << "    r_q"<<i<<"[j] = r_t"<<i<<"[q + j*Q1d];\n";
        code << "  }\n";
        break;
      case CEED_EVAL_GRAD:
        code << "  CeedScalar r_q"<<i<<"[ncomp_in_"<<i<<"*Dim];\n";
        code << "  gradCollo3d<ncomp_in_"<<i<<",Q1d>(data, q, r_t"<<i<<", s_G_in_"<<i<<", r_q"<<i<<");\n";
        break;
      case CEED_EVAL_WEIGHT:
        code << "  CeedScalar r_q"<<i<<"[1];\n";
        code << "  r_q"<<i<<"[0] = r_t"<<i<<"[q];\n";
        break; // No action
      case CEED_EVAL_DIV:
        break; // TODO: Not implemented
      case CEED_EVAL_CURL:
        break; // TODO: Not implemented
      }
    }
    for (CeedInt i = 0; i < numoutputfields; i++) {
      code << "  // -- Output field "<<i<<" --\n";
      ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode);
      CeedChk(ierr);
      // Basis action
      switch (emode) {
      case CEED_EVAL_NONE:
        code << "  CeedScalar r_qq"<<i<<"[ncomp_out_"<<i<<"];\n";
        break; // No action
      case CEED_EVAL_INTERP:
        code << "  CeedScalar r_qq"<<i<<"[ncomp_out_"<<i<<"];\n";
        break;
      case CEED_EVAL_GRAD:
        code << "  CeedScalar r_qq"<<i<<"[ncomp_out_"<<i<<"*Dim];\n";
        break;
      case CEED_EVAL_WEIGHT:
        break; // Should not occur
      case CEED_EVAL_DIV:
        break; // TODO: Not implemented
      case CEED_EVAL_CURL:
        break; // TODO: Not implemented
      }
    }
  } else {
      code << "\n  // Note: No Collocated Gradient\n";
    for (CeedInt i = 0; i < numinputfields; i++) {
      code << "  // -- Input field "<<i<<" --\n";
      code << "  CeedScalar* r_q"<<i<<" = r_t"<<i<<";\n";
    }
    for (CeedInt i = 0; i < numoutputfields; i++) {
      code << "  // -- Output field "<<i<<" --\n";
      code << "  CeedScalar* r_qq"<<i<<" = r_tt"<<i<<";\n";
    }
  }
  code << "  // QFunction Inputs and outputs\n";
  code << "  CeedScalar* in["<<numinputfields<<"];\n";
  for (CeedInt i = 0; i < numinputfields; i++) {
    code << "  // -- Input field "<<i<<" --\n";
    code << "  in["<<i<<"] = r_q"<<i<<";\n";
  }
  code << "  CeedScalar* out["<<numoutputfields<<"];\n";
  for (CeedInt i = 0; i < numoutputfields; i++) {
    code << "  // -- Output field "<<i<<" --\n";
    code << "  out["<<i<<"] = r_qq"<<i<<";\n";
  }
  code << "\n  // Apply QFunction\n";
  string qFunctionName(qf_data->qFunctionName);
  code << "  "<<qFunctionName<<"(ctx, ";
  if (dim != 3 || basis_data->d_collograd1d) {
    code << "1 ";
  } else {
    code << "Q1d ";
  }
  code << ", in, out);\n";
  if (basis_data->d_collograd1d) {
    code << "\n  // Note: Collocated Gradient\n";
    for (CeedInt i = 0; i < numoutputfields; i++) {
      code << "  // -- Output field "<<i<<" --\n";
      ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode);
      CeedChk(ierr);
      // Basis action
      code << "  // EvalMode: "<<CeedEvalModes[emode]<<"\n";
      switch (emode) {
      case CEED_EVAL_NONE:
        code << "  for (CeedInt j = 0; j < ncomp_out_"<<i<<" ; ++j) {\n";
        code << "    r_tt"<<i<<"[q + j*Q1d] = r_qq"<<i<<"[j];\n";
        code << "  }\n";
        break; // No action
      case CEED_EVAL_INTERP:
        code << "  for (CeedInt j = 0; j < ncomp_out_"<<i<<" ; ++j) {\n";
        code << "    r_tt"<<i<<"[q + j*Q1d] = r_qq"<<i<<"[j];\n";
        code << "  }\n";
        break;
      case CEED_EVAL_GRAD:
        code << "  gradColloTranspose3d<ncomp_out_"<<i<<",Q1d>(data, q, r_qq"<<i<<", s_G_out_"<<i<<", r_tt"<<i<<");\n";
        break;
      case CEED_EVAL_WEIGHT:
        break; // Should not occur
      case CEED_EVAL_DIV:
        break; // TODO: Not implemented
      case CEED_EVAL_CURL:
        break; // TODO: Not implemented
      }
    }
    code << "}\n";
  }

  // Output basis apply if needed
  // Generate the correct eval mode code for each output
  code << "\n// Output field basis action and restrictions\n";
  for (CeedInt i = 0; i < numoutputfields; i++) {
    code << "  // -- Output field "<<i<<" --\n";
    // Get elemsize, emode, ncomp
    ierr = CeedOperatorFieldGetElemRestriction(opoutputfields[i], &Erestrict);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetElementSize(Erestrict, &elemsize);
    CeedChk(ierr);
    ierr = CeedQFunctionFieldGetEvalMode(qfoutputfields[i], &emode);
    CeedChk(ierr);
    ierr = CeedElemRestrictionGetNumComponents(Erestrict, &ncomp);
    CeedChk(ierr);
    // Basis action
    code << "  // EvalMode: "<<CeedEvalModes[emode]<<"\n";
    switch (emode) {
    case CEED_EVAL_NONE:
      code << "  CeedScalar* r_v"<<i<<" = r_tt"<<i<<";\n";
      break; // No action
    case CEED_EVAL_INTERP:
      code << "  CeedScalar r_v"<<i<<"[ncomp_out_"<<i<<"*P_out_"<<i<<"];\n";
      code << "  interpTranspose"<<dim<<"d<ncomp_out_"<<i<<",P_out_"<<i<<",Q1d>(data, r_tt"<<i<<", s_B_out_"<<i<<", r_v"<<i<<");\n";
      break;
    case CEED_EVAL_GRAD:
      code << "  CeedScalar r_v"<<i<<"[ncomp_out_"<<i<<"*P_out_"<<i<<"];\n";
      if (basis_data->d_collograd1d) {
        code << "  interpTranspose"<<dim<<"d<ncomp_out_"<<i<<",P_out_"<<i<<",Q1d>(data, r_tt"<<i<<", s_B_out_"<<i<<", r_v"<<i<<");\n";
      } else {
        code << "  gradTranspose"<<dim<<"d<ncomp_out_"<<i<<",P_out_"<<i<<",Q1d>(data, r_tt"<<i<<", s_B_out_"<<i<<", s_G_out_"<<i<<", r_v"<<i<<");\n";
      }
      break;
    case CEED_EVAL_WEIGHT: {
      Ceed ceed;
      ierr = CeedOperatorGetCeed(op, &ceed); CeedChk(ierr);
      return CeedError(ceed, 1,
                       "CEED_EVAL_WEIGHT cannot be an output evaluation mode");
      break; // Should not occur
    }
    case CEED_EVAL_DIV:
      break; // TODO: Not implemented
    case CEED_EVAL_CURL:
      break; // TODO: Not implemented
    }
    // Restriction
      bool isStrided;
      ierr = CeedElemRestrictionIsStrided(Erestrict, &isStrided); CeedChk(ierr);
    if (!isStrided) {
      ierr = CeedElemRestrictionGetLVectorSize(Erestrict, &lsize);
      CeedChk(ierr);
      code << "  const CeedInt lsize_out_"<<i<<" = "<<lsize<<";\n";
      CeedInt compstride;
      ierr = CeedElemRestrictionGetCompStride(Erestrict, &compstride); CeedChk(ierr);
      code << "  // CompStride: "<<compstride<<"\n";
      ierr = CeedElemRestrictionGetData(Erestrict, (void **)&restr_data); CeedChk(ierr);
      data->indices.out[i] = restr_data->d_ind;
      code << "  writeDofsOffset"<<dim<<"d<ncomp_out_"<<i<<", "<<compstride<<", P_out_"<<i<<">(data, lsize_out_"<<i<<", elem, indices.out["<<i<<"], r_v"<<i<<", d_v"<<i<<");\n";
    } else {
      bool backendstrides;
      ierr = CeedElemRestrictionHasBackendStrides(Erestrict, &backendstrides);
      CeedChk(ierr);
      CeedInt strides[3] = {1, elemsize, elemsize*ncomp};
      if (!backendstrides) {
        ierr = CeedElemRestrictionGetStrides(Erestrict, &strides);
        CeedChk(ierr);
      }
      code << "  // Strides: {"<<strides[0]<<", "<<strides[1]<<", "<<strides[2]<<"}\n";
      code << "  writeDofsStrided"<<dim<<"d<ncomp_out_"<<i<<",P_out_"<<i<<","<<strides[0]<<","<<strides[1]<<","<<strides[2]<<">(data, elem, r_v"<<i<<", d_v"<<i<<");\n";
    }
  }

  code << "  }\n";
  code << "}\n\n";

  // View kernel for debugging
  // std::cout << code.str();

  ierr = CeedCompileCuda(ceed, code.str().c_str(), &data->module, 0);
  CeedChk(ierr);
  ierr = CeedGetKernelCuda(ceed, data->module, "oper", &data->op);
  CeedChk(ierr);

  ierr = CeedOperatorSetSetupDone(op); CeedChk(ierr);
  return 0;
}
//------------------------------------------------------------------------------