xref: /libCEED/backends/cuda-gen/ceed-cuda-gen-operator-build.cpp (revision eb7e6cafeb5d6cc94d59355f95e7bc9ae3fc1c25)

// Copyright (c) 2017-2022, Lawrence Livermore National Security, LLC and other CEED contributors.
// All Rights Reserved. See the top-level LICENSE and NOTICE files for details.
//
// SPDX-License-Identifier: BSD-2-Clause
//
// This file is part of CEED:  http://github.com/ceed

#define CEED_DEBUG_COLOR 12

#include <ceed.h>
#include <ceed/backend.h>
#include <ceed/jit-source/cuda/cuda-types.h>
#include <ceed/jit-tools.h>
#include <cuda_runtime.h>

#include <iostream>
#include <sstream>

#include "../cuda-ref/ceed-cuda-ref.h"
#include "../cuda-shared/ceed-cuda-shared.h"
#include "../cuda/ceed-cuda-common.h"
#include "../cuda/ceed-cuda-compile.h"
#include "ceed-cuda-gen.h"

//------------------------------------------------------------------------------
// Build singe operator kernel
//------------------------------------------------------------------------------
extern "C" int CeedOperatorBuildKernel_Cuda_gen(CeedOperator op) {
  using std::ostringstream;
  using std::string;
  bool is_setup_done;
  CeedCallBackend(CeedOperatorIsSetupDone(op, &is_setup_done));
  if (is_setup_done) return CEED_ERROR_SUCCESS;
  Ceed ceed;
  CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
  CeedOperator_Cuda_gen *data;
  CeedCallBackend(CeedOperatorGetData(op, &data));
  CeedQFunction           qf;
  CeedQFunction_Cuda_gen *qf_data;
  CeedCallBackend(CeedOperatorGetQFunction(op, &qf));
  CeedCallBackend(CeedQFunctionGetData(qf, &qf_data));
  CeedSize lsize;
  CeedInt  Q, P_1d = 0, Q_1d = 0, elem_size, num_input_fields, num_output_fields, num_comp, dim = 1;
  CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q));
  Q_1d = Q;
  CeedOperatorField *op_input_fields, *op_output_fields;
  CeedCallBackend(CeedOperatorGetFields(op, &num_input_fields, &op_input_fields, &num_output_fields, &op_output_fields));
  CeedQFunctionField *qf_input_fields, *qf_output_fields;
  CeedCallBackend(CeedQFunctionGetFields(qf, NULL, &qf_input_fields, NULL, &qf_output_fields));
  CeedEvalMode              eval_mode;
  CeedBasis                 basis;
  CeedBasis_Cuda_shared    *basis_data;
  CeedElemRestriction       Erestrict;
  CeedElemRestriction_Cuda *restr_data;

  // TODO: put in a function?
  // Check for restriction only identity operator
  bool is_identity_qf;
  CeedCallBackend(CeedQFunctionIsIdentity(qf, &is_identity_qf));
  if (is_identity_qf) {
    CeedEvalMode eval_mode_in, eval_mode_out;
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[0], &eval_mode_in));
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[0], &eval_mode_out));
    CeedCheck(eval_mode_in != CEED_EVAL_NONE || eval_mode_out != CEED_EVAL_NONE, ceed, CEED_ERROR_BACKEND,
              "Backend does not implement restriction only identity operators");
  }

  ostringstream code;

  // TODO: put in a function?
  // Add atomicAdd function for old NVidia architectures
  struct cudaDeviceProp prop;
  Ceed_Cuda            *ceed_data;
  CeedCallBackend(CeedGetData(ceed, &ceed_data));
  CeedCallBackend(cudaGetDeviceProperties(&prop, ceed_data->device_id));
  if ((prop.major < 6) && (CEED_SCALAR_TYPE != CEED_SCALAR_FP32)) {
    char *atomic_add_path, *atomic_add_source;
    CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-atomic-add-fallback.h", &atomic_add_path));
    CeedDebug256(ceed, 2, "----- Loading Atomic Add Source -----\n");
    CeedCallBackend(CeedLoadSourceToBuffer(ceed, atomic_add_path, &atomic_add_source));
    code << atomic_add_source;
    CeedCallBackend(CeedFree(&atomic_add_path));
    CeedCallBackend(CeedFree(&atomic_add_source));
  }

  // Load basis source files
  // TODO: generalize to accept different device functions?
  {
    char *tensor_basis_kernel_path, *tensor_basis_kernel_source;
    CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-shared-basis-tensor-templates.h", &tensor_basis_kernel_path));
    CeedDebug256(ceed, 2, "----- Loading Tensor Basis Kernel Source -----\n");
    CeedCallBackend(CeedLoadSourceToBuffer(ceed, tensor_basis_kernel_path, &tensor_basis_kernel_source));
    code << tensor_basis_kernel_source;
    CeedCallBackend(CeedFree(&tensor_basis_kernel_path));
    CeedCallBackend(CeedFree(&tensor_basis_kernel_source));
  }
  {
    char *cuda_gen_template_path, *cuda_gen_template_source;
    CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/cuda/cuda-gen-templates.h", &cuda_gen_template_path));
    CeedDebug256(ceed, 2, "----- Loading Cuda-Gen Template Source -----\n");
    CeedCallBackend(CeedLoadSourceToBuffer(ceed, cuda_gen_template_path, &cuda_gen_template_source));
    code << cuda_gen_template_source;
    CeedCallBackend(CeedFree(&cuda_gen_template_path));
    CeedCallBackend(CeedFree(&cuda_gen_template_source));
  }

  // Get QFunction source and name
  string q_function_source(qf_data->q_function_source);
  string q_function_name(qf_data->q_function_name);
  string operator_name;
  operator_name = "CeedKernelCudaGenOperator_" + q_function_name;

  // Find dim, P_1d, Q_1d
  data->max_P_1d = 0;
  for (CeedInt i = 0; i < num_input_fields; i++) {
    CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis));
    if (basis != CEED_BASIS_COLLOCATED) {
      CeedCallBackend(CeedBasisGetData(basis, &basis_data));
      CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));

      // Collect dim, P_1d, and Q_1d
      CeedCallBackend(CeedBasisGetDimension(basis, &dim));
      bool isTensor;
      CeedCallBackend(CeedBasisIsTensor(basis, &isTensor));
      CeedCheck(isTensor, ceed, CEED_ERROR_BACKEND, "Backend does not implement operators with non-tensor basis");
      CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d));
      CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d));
      if (P_1d > data->max_P_1d) data->max_P_1d = P_1d;
    }
  }
  // Check output bases for Q_1d, dim as well
  //   The only input basis might be CEED_BASIS_COLLOCATED
  for (CeedInt i = 0; i < num_output_fields; i++) {
    CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis));

    if (basis != CEED_BASIS_COLLOCATED) {
      CeedCallBackend(CeedBasisGetData(basis, &basis_data));
      CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));

      // Collect Q_1d
      CeedCallBackend(CeedBasisGetDimension(basis, &dim));
      bool isTensor;
      CeedCallBackend(CeedBasisIsTensor(basis, &isTensor));
      CeedCheck(isTensor, ceed, CEED_ERROR_BACKEND, "Backend does not implement operators with non-tensor basis");
      CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q_1d));
    }
  }
  data->dim  = dim;
  data->Q_1d = Q_1d;

  // Only use 3D collocated gradient parallelization strategy when gradient is computed
  // TODO: put in a function?
  bool use_collograd_parallelization = false;
  if (dim == 3) {
    bool was_grad_found = false;
    for (CeedInt i = 0; i < num_input_fields; i++) {
      CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
      if (eval_mode == CEED_EVAL_GRAD) {
        CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis));
        CeedCallBackend(CeedBasisGetData(basis, &basis_data));
        use_collograd_parallelization = !!basis_data->d_collo_grad_1d && (was_grad_found ? use_collograd_parallelization : true);
        was_grad_found                = true;
      }
    }
    for (CeedInt i = 0; i < num_output_fields; i++) {
      CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
      if (eval_mode == CEED_EVAL_GRAD) {
        CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis));
        CeedCallBackend(CeedBasisGetData(basis, &basis_data));
        use_collograd_parallelization = !!basis_data->d_collo_grad_1d && (was_grad_found ? use_collograd_parallelization : true);
        was_grad_found                = true;
      }
    }
  }

  // Define CEED_Q_VLA
  code << "\n#undef CEED_Q_VLA\n";
  if (dim != 3 || use_collograd_parallelization) {
    code << "#define CEED_Q_VLA 1\n\n";
  } else {
    code << "#define CEED_Q_VLA " << Q_1d << "\n\n";
  }

  code << q_function_source;

  // Setup
  code << "\n// -----------------------------------------------------------------------------\n";
  code << "\nextern \"C\" __global__ void " << operator_name
       << "(CeedInt num_elem, void* ctx, FieldsInt_Cuda indices, Fields_Cuda fields, Fields_Cuda B, Fields_Cuda G, CeedScalar* W) {\n";
  for (CeedInt i = 0; i < num_input_fields; i++) {
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
    if (eval_mode != CEED_EVAL_WEIGHT) {  // Skip CEED_EVAL_WEIGHT
      code << "  const CeedScalar* d_u_" << i << " = fields.inputs[" << i << "];\n";
    }
  }

  for (CeedInt i = 0; i < num_output_fields; i++) {
    code << "  CeedScalar* d_v_" << i << " = fields.outputs[" << i << "];\n";
  }

  code << "  const CeedInt dim = " << dim << ";\n";
  code << "  const CeedInt Q_1d = " << Q_1d << ";\n";

  code << "  extern __shared__ CeedScalar slice[];\n";
  // TODO put in a function? InitSharedData_Cuda?
  code << "  SharedData_Cuda data;\n";
  code << "  data.t_id_x = threadIdx.x;\n";
  code << "  data.t_id_y = threadIdx.y;\n";
  code << "  data.t_id_z = threadIdx.z;\n";
  code << "  data.t_id  = threadIdx.x + threadIdx.y*blockDim.x + threadIdx.z*blockDim.y*blockDim.x;\n";
  code << "  data.slice = slice+data.t_id_z*T_1D" << (dim > 1 ? "*T_1D" : "") << ";\n";

  code << "\n  // -- Input field constants and basis data --\n";
  // TODO: Put in a function?
  // Initialize constants, and matrices B and G
  for (CeedInt i = 0; i < num_input_fields; i++) {
    code << "  // ---- Input field " << i << " ----\n";
    // Get elem_size, eval_mode, num_comp
    CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &Erestrict));
    CeedCallBackend(CeedElemRestrictionGetElementSize(Erestrict, &elem_size));
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
    CeedCallBackend(CeedElemRestrictionGetNumComponents(Erestrict, &num_comp));

    // Set field constants
    if (eval_mode != CEED_EVAL_WEIGHT) {
      CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis));
      if (basis != CEED_BASIS_COLLOCATED) {
        CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d));
        code << "  const CeedInt P_in_" << i << " = " << P_1d << ";\n";
      } else {
        code << "  const CeedInt P_in_" << i << " = " << Q_1d << ";\n";
      }
      code << "  const CeedInt num_comp_in_" << i << " = " << num_comp << ";\n";
    }

    // Load basis data
    code << "  // EvalMode: " << CeedEvalModes[eval_mode] << "\n";
    switch (eval_mode) {
      case CEED_EVAL_NONE:
        break;
      case CEED_EVAL_INTERP:
        CeedCallBackend(CeedBasisGetData(basis, &basis_data));
        data->B.inputs[i] = basis_data->d_interp_1d;
        code << "  __shared__ CeedScalar s_B_in_" << i << "[" << P_1d * Q_1d << "];\n";
        code << "  loadMatrix<P_in_" << i << ",Q_1d>(data, B.inputs[" << i << "], s_B_in_" << i << ");\n";
        break;
      case CEED_EVAL_GRAD:
        CeedCallBackend(CeedBasisGetData(basis, &basis_data));
        data->B.inputs[i] = basis_data->d_interp_1d;
        code << "  __shared__ CeedScalar s_B_in_" << i << "[" << P_1d * Q_1d << "];\n";
        code << "  loadMatrix<P_in_" << i << ",Q_1d>(data, B.inputs[" << i << "], s_B_in_" << i << ");\n";
        if (use_collograd_parallelization) {
          data->G.inputs[i] = basis_data->d_collo_grad_1d;
          code << "  __shared__ CeedScalar s_G_in_" << i << "[" << Q_1d * Q_1d << "];\n";
          code << "  loadMatrix<Q_1d,Q_1d>(data, G.inputs[" << i << "], s_G_in_" << i << ");\n";
        } else {
          bool has_collo_grad = !!basis_data->d_collo_grad_1d;
          data->G.inputs[i]   = has_collo_grad ? basis_data->d_collo_grad_1d : basis_data->d_grad_1d;
          code << "  __shared__ CeedScalar s_G_in_" << i << "[" << Q_1d * (has_collo_grad ? Q_1d : P_1d) << "];\n";
          code << "  loadMatrix<" << (has_collo_grad ? "Q_1d" : ("P_in_" + std::to_string(i))) << ",Q_1d>(data, G.inputs[" << 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 < num_output_fields; i++) {
    code << "  // ---- Output field " << i << " ----\n";
    // Get elem_size, eval_mode, num_comp
    CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &Erestrict));
    CeedCallBackend(CeedElemRestrictionGetElementSize(Erestrict, &elem_size));
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
    CeedCallBackend(CeedElemRestrictionGetNumComponents(Erestrict, &num_comp));

    // Set field constants
    CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis));
    if (basis != CEED_BASIS_COLLOCATED) {
      CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d));
      code << "  const CeedInt P_out_" << i << " = " << P_1d << ";\n";
    } else {
      code << "  const CeedInt P_out_" << i << " = " << Q_1d << ";\n";
    }
    code << "  const CeedInt num_comp_out_" << i << " = " << num_comp << ";\n";

    // Load basis data
    code << "  // EvalMode: " << CeedEvalModes[eval_mode] << "\n";
    switch (eval_mode) {
      case CEED_EVAL_NONE:
        break;  // No action
      case CEED_EVAL_INTERP:
        CeedCallBackend(CeedBasisGetData(basis, &basis_data));
        data->B.outputs[i] = basis_data->d_interp_1d;
        code << "  __shared__ CeedScalar s_B_out_" << i << "[" << P_1d * Q_1d << "];\n";
        code << "  loadMatrix<P_out_" << i << ",Q_1d>(data, B.outputs[" << i << "], s_B_out_" << i << ");\n";
        break;
      case CEED_EVAL_GRAD:
        CeedCallBackend(CeedBasisGetData(basis, &basis_data));
        data->B.outputs[i] = basis_data->d_interp_1d;
        code << "  __shared__ CeedScalar s_B_out_" << i << "[" << P_1d * Q_1d << "];\n";
        code << "  loadMatrix<P_out_" << i << ",Q_1d>(data, B.outputs[" << i << "], s_B_out_" << i << ");\n";
        if (use_collograd_parallelization) {
          data->G.outputs[i] = basis_data->d_collo_grad_1d;
          code << "  __shared__ CeedScalar s_G_out_" << i << "[" << Q_1d * Q_1d << "];\n";
          code << "  loadMatrix<Q_1d,Q_1d>(data, G.outputs[" << i << "], s_G_out_" << i << ");\n";
        } else {
          bool has_collo_grad = !!basis_data->d_collo_grad_1d;
          data->G.outputs[i]  = has_collo_grad ? basis_data->d_collo_grad_1d : basis_data->d_grad_1d;
          code << "  __shared__ CeedScalar s_G_out_" << i << "[" << Q_1d * (has_collo_grad ? Q_1d : P_1d) << "];\n";
          code << "  loadMatrix<" << (has_collo_grad ? "Q_1d" : ("P_out_" + std::to_string(i))) << ",Q_1d>(data, G.outputs[" << i << "], s_G_out_"
               << i << ");\n";
        }
        break;
      // LCOV_EXCL_START
      case CEED_EVAL_WEIGHT: {
        Ceed ceed;
        CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
        return CeedError(ceed, CEED_ERROR_BACKEND, "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
                // LCOV_EXCL_STOP
    }
  }
  code << "\n  // -- Element loop --\n";
  code << "  __syncthreads();\n";
  code << "  for (CeedInt elem = blockIdx.x*blockDim.z + threadIdx.z; elem < num_elem; elem += gridDim.x*blockDim.z) {\n";
  // Input basis apply if needed
  // Generate the correct eval mode code for each input
  code << "    // -- Input field restrictions and basis actions --\n";
  for (CeedInt i = 0; i < num_input_fields; i++) {
    code << "    // ---- Input field " << i << " ----\n";
    // Get elem_size, eval_mode, num_comp
    CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &Erestrict));
    CeedCallBackend(CeedElemRestrictionGetElementSize(Erestrict, &elem_size));
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
    CeedCallBackend(CeedElemRestrictionGetNumComponents(Erestrict, &num_comp));

    // TODO: put in a function?
    // Restriction
    if (eval_mode != CEED_EVAL_WEIGHT && !((eval_mode == CEED_EVAL_NONE) && use_collograd_parallelization)) {
      code << "    CeedScalar r_u_" << i << "[num_comp_in_" << i << "*P_in_" << i << "];\n";

      bool is_strided;
      CeedCallBackend(CeedElemRestrictionIsStrided(Erestrict, &is_strided));
      if (!is_strided) {
        CeedCallBackend(CeedElemRestrictionGetLVectorSize(Erestrict, &lsize));
        code << "    const CeedInt lsize_in_" << i << " = " << lsize << ";\n";
        CeedInt comp_stride;
        CeedCallBackend(CeedElemRestrictionGetCompStride(Erestrict, &comp_stride));
        code << "    // CompStride: " << comp_stride << "\n";
        CeedCallBackend(CeedElemRestrictionGetData(Erestrict, &restr_data));
        data->indices.inputs[i] = restr_data->d_ind;
        code << "    readDofsOffset" << dim << "d<num_comp_in_" << i << ", " << comp_stride << ", P_in_" << i << ">(data, lsize_in_" << i
             << ", elem, indices.inputs[" << i << "], d_u_" << i << ", r_u_" << i << ");\n";
      } else {
        bool backendstrides;
        CeedCallBackend(CeedElemRestrictionHasBackendStrides(Erestrict, &backendstrides));
        CeedInt num_elem;
        CeedCallBackend(CeedElemRestrictionGetNumElements(Erestrict, &num_elem));
        CeedInt strides[3] = {1, elem_size * num_elem, elem_size};
        if (!backendstrides) {
          CeedCallBackend(CeedElemRestrictionGetStrides(Erestrict, &strides));
        }
        code << "    // Strides: {" << strides[0] << ", " << strides[1] << ", " << strides[2] << "}\n";
        code << "    readDofsStrided" << dim << "d<num_comp_in_" << i << ",P_in_" << i << "," << strides[0] << "," << strides[1] << "," << strides[2]
             << ">(data, elem, d_u_" << i << ", r_u_" << i << ");\n";
      }
    }

    // TODO: put in a function?
    // Basis action
    code << "    // EvalMode: " << CeedEvalModes[eval_mode] << "\n";
    switch (eval_mode) {
      case CEED_EVAL_NONE:
        if (!use_collograd_parallelization) {
          code << "    CeedScalar* r_t_" << i << " = r_u_" << i << ";\n";
        }
        break;
      case CEED_EVAL_INTERP:
        code << "    CeedScalar r_t_" << i << "[num_comp_in_" << i << "*Q_1d];\n";
        code << "    Interp" << (dim > 1 ? "Tensor" : "") << dim << "d<num_comp_in_" << i << ",P_in_" << i << ",Q_1d>(data, r_u_" << i << ", s_B_in_"
             << i << ", r_t_" << i << ");\n";
        break;
      case CEED_EVAL_GRAD:
        if (use_collograd_parallelization) {
          code << "    CeedScalar r_t_" << i << "[num_comp_in_" << i << "*Q_1d];\n";
          code << "    Interp" << (dim > 1 ? "Tensor" : "") << dim << "d<num_comp_in_" << i << ",P_in_" << i << ",Q_1d>(data, r_u_" << i
               << ", s_B_in_" << i << ", r_t_" << i << ");\n";
        } else {
          CeedInt P_1d;
          CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis));
          CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d));
          code << "    CeedScalar r_t_" << i << "[num_comp_in_" << i << "*dim*Q_1d];\n";
          code << "    Grad" << (dim > 1 ? "Tensor" : "") << (dim == 3 && Q_1d >= P_1d ? "Collocated" : "") << dim << "d<num_comp_in_" << i
               << ",P_in_" << i << ",Q_1d>(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 << "[Q_1d];\n";
        CeedCallBackend(CeedOperatorFieldGetBasis(op_input_fields[i], &basis));
        CeedCallBackend(CeedBasisGetData(basis, &basis_data));
        data->W = basis_data->d_q_weight_1d;
        code << "    Weight" << (dim > 1 ? "Tensor" : "") << dim << "d<Q_1d>(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
    }
  }

  // TODO: put in a function + separate collograd logic
  // Q function
  code << "\n    // -- Output field setup --\n";
  for (CeedInt i = 0; i < num_output_fields; i++) {
    code << "\n    // ---- Output field " << i << " ----\n";
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
    if (eval_mode == CEED_EVAL_GRAD) {
      if (use_collograd_parallelization) {
        // Accumulator for gradient slices
        code << "    CeedScalar r_tt_" << i << "[num_comp_out_" << i << "*Q_1d];\n";
        code << "    for (CeedInt i = 0; i < num_comp_out_" << i << "; i++) {\n";
        code << "      for (CeedInt j = 0; j < Q_1d; ++j) {\n";
        code << "        r_tt_" << i << "[j + i*Q_1d] = 0.0;\n";
        code << "      }\n";
        code << "    }\n";
      } else {
        code << "    CeedScalar r_tt_" << i << "[num_comp_out_" << i << "*dim*Q_1d];\n";
      }
    }
    if (eval_mode == CEED_EVAL_NONE || eval_mode == CEED_EVAL_INTERP) {
      code << "    CeedScalar r_tt_" << i << "[num_comp_out_" << i << "*Q_1d];\n";
    }
  }
  // We treat quadrature points per slice in 3d to save registers
  if (use_collograd_parallelization) {
    code << "\n    // Note: Using planes of 3D elements\n";
    code << "#pragma unroll\n";
    code << "    for (CeedInt q = 0; q < Q_1d; q++) {\n";
    code << "      // -- Input fields --\n";
    for (CeedInt i = 0; i < num_input_fields; i++) {
      code << "      // ---- Input field " << i << " ----\n";
      // Get elem_size, eval_mode, num_comp
      CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_input_fields[i], &eval_mode));
      // Basis action
      code << "      // EvalMode: " << CeedEvalModes[eval_mode] << "\n";
      switch (eval_mode) {
        case CEED_EVAL_NONE:
          code << "      CeedScalar r_q_" << i << "[num_comp_in_" << i << "];\n";

          bool is_strided;
          CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_input_fields[i], &Erestrict));
          CeedCallBackend(CeedElemRestrictionIsStrided(Erestrict, &is_strided));
          if (!is_strided) {
            CeedCallBackend(CeedElemRestrictionGetLVectorSize(Erestrict, &lsize));
            code << "      const CeedInt lsize_in_" << i << " = " << lsize << ";\n";
            CeedInt comp_stride;
            CeedCallBackend(CeedElemRestrictionGetCompStride(Erestrict, &comp_stride));
            code << "      // CompStride: " << comp_stride << "\n";
            CeedCallBackend(CeedElemRestrictionGetData(Erestrict, &restr_data));
            data->indices.inputs[i] = restr_data->d_ind;
            code << "      readSliceQuadsOffset"
                 << "3d<num_comp_in_" << i << ", " << comp_stride << ", Q_1d>(data, lsize_in_" << i << ", elem, q, indices.inputs[" << i << "], d_u_"
                 << i << ", r_q_" << i << ");\n";
          } else {
            CeedCallBackend(CeedElemRestrictionGetElementSize(Erestrict, &elem_size));
            bool backendstrides;
            CeedCallBackend(CeedElemRestrictionHasBackendStrides(Erestrict, &backendstrides));
            CeedInt num_elem;
            CeedCallBackend(CeedElemRestrictionGetNumElements(Erestrict, &num_elem));
            CeedInt strides[3] = {1, elem_size * num_elem, elem_size};
            if (!backendstrides) {
              CeedCallBackend(CeedElemRestrictionGetStrides(Erestrict, &strides));
            }
            code << "      // Strides: {" << strides[0] << ", " << strides[1] << ", " << strides[2] << "}\n";
            code << "      readSliceQuadsStrided"
                 << "3d<num_comp_in_" << i
                 << ",Q_1d"
                    ","
                 << 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 << "[num_comp_in_" << i << "];\n";
          code << "      for (CeedInt j = 0; j < num_comp_in_" << i << " ; ++j) {\n";
          code << "        r_q_" << i << "[j] = r_t_" << i << "[q + j*Q_1d];\n";
          code << "      }\n";
          break;
        case CEED_EVAL_GRAD:
          code << "      CeedScalar r_q_" << i << "[num_comp_in_" << i << "*dim];\n";
          code << "      gradCollo3d<num_comp_in_" << i << ",Q_1d>(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
      }
    }
    code << "\n      // -- Output fields --\n";
    for (CeedInt i = 0; i < num_output_fields; i++) {
      code << "      // ---- Output field " << i << " ----\n";
      CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
      // Basis action
      switch (eval_mode) {
        case CEED_EVAL_NONE:
          code << "      CeedScalar r_qq_" << i << "[num_comp_out_" << i << "];\n";
          break;  // No action
        case CEED_EVAL_INTERP:
          code << "      CeedScalar r_qq_" << i << "[num_comp_out_" << i << "];\n";
          break;
        case CEED_EVAL_GRAD:
          code << "      CeedScalar r_qq_" << i << "[num_comp_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: Using full elements\n";
    code << "      // -- Input fields --\n";
    for (CeedInt i = 0; i < num_input_fields; i++) {
      code << "      // ---- Input field " << i << " ----\n";
      code << "      CeedScalar* r_q_" << i << " = r_t_" << i << ";\n";
    }
    code << "      // -- Output fields --\n";
    for (CeedInt i = 0; i < num_output_fields; i++) {
      code << "      // ---- Output field " << i << " ----\n";
      code << "      CeedScalar* r_qq_" << i << " = r_tt_" << i << ";\n";
    }
  }
  code << "\n      // -- QFunction Inputs and outputs --\n";
  code << "      CeedScalar* in[" << num_input_fields << "];\n";
  for (CeedInt i = 0; i < num_input_fields; i++) {
    code << "      // ---- Input field " << i << " ----\n";
    code << "      in[" << i << "] = r_q_" << i << ";\n";
  }
  code << "      CeedScalar* out[" << num_output_fields << "];\n";
  for (CeedInt i = 0; i < num_output_fields; i++) {
    code << "      // ---- Output field " << i << " ----\n";
    code << "      out[" << i << "] = r_qq_" << i << ";\n";
  }
  code << "\n      // -- Apply QFunction --\n";
  code << "      " << q_function_name << "(ctx, ";
  if (dim != 3 || use_collograd_parallelization) {
    code << "1";
  } else {
    code << "Q_1d";
  }
  code << ", in, out);\n";
  if (use_collograd_parallelization) {
    code << "      // -- Output fields --\n";
    for (CeedInt i = 0; i < num_output_fields; i++) {
      code << "      // ---- Output field " << i << " ----\n";
      CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
      // Basis action
      code << "      // EvalMode: " << CeedEvalModes[eval_mode] << "\n";
      switch (eval_mode) {
        case CEED_EVAL_NONE:
          code << "      for (CeedInt j = 0; j < num_comp_out_" << i << " ; ++j) {\n";
          code << "        r_tt_" << i << "[q + j*Q_1d] = r_qq_" << i << "[j];\n";
          code << "      }\n";
          break;  // No action
        case CEED_EVAL_INTERP:
          code << "      for (CeedInt j = 0; j < num_comp_out_" << i << " ; ++j) {\n";
          code << "        r_tt_" << i << "[q + j*Q_1d] = r_qq_" << i << "[j];\n";
          code << "      }\n";
          break;
        case CEED_EVAL_GRAD:
          code << "      gradColloTranspose3d<num_comp_out_" << i << ",Q_1d>(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 < num_output_fields; i++) {
    code << "    // ---- Output field " << i << " ----\n";
    // Get elem_size, eval_mode, num_comp
    CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_output_fields[i], &Erestrict));
    CeedCallBackend(CeedElemRestrictionGetElementSize(Erestrict, &elem_size));
    CeedCallBackend(CeedQFunctionFieldGetEvalMode(qf_output_fields[i], &eval_mode));
    CeedCallBackend(CeedElemRestrictionGetNumComponents(Erestrict, &num_comp));
    // TODO put in a function
    // Basis action
    code << "    // EvalMode: " << CeedEvalModes[eval_mode] << "\n";
    switch (eval_mode) {
      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 << "[num_comp_out_" << i << "*P_out_" << i << "];\n";
        code << "    InterpTranspose" << (dim > 1 ? "Tensor" : "") << dim << "d<num_comp_out_" << i << ",P_out_" << i << ",Q_1d>(data, r_tt_" << i
             << ", s_B_out_" << i << ", r_v_" << i << ");\n";
        break;
      case CEED_EVAL_GRAD:
        code << "    CeedScalar r_v_" << i << "[num_comp_out_" << i << "*P_out_" << i << "];\n";
        if (use_collograd_parallelization) {
          code << "    InterpTranspose" << (dim > 1 ? "Tensor" : "") << dim << "d<num_comp_out_" << i << ",P_out_" << i << ",Q_1d>(data, r_tt_" << i
               << ", s_B_out_" << i << ", r_v_" << i << ");\n";
        } else {
          CeedInt P_1d;
          CeedCallBackend(CeedOperatorFieldGetBasis(op_output_fields[i], &basis));
          CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P_1d));
          code << "    GradTranspose" << (dim > 1 ? "Tensor" : "") << (dim == 3 && Q_1d >= P_1d ? "Collocated" : "") << dim << "d<num_comp_out_" << i
               << ",P_out_" << i << ",Q_1d>(data, r_tt_" << i << ", s_B_out_" << i << ", s_G_out_" << i << ", r_v_" << i << ");\n";
        }
        break;
      // LCOV_EXCL_START
      case CEED_EVAL_WEIGHT: {
        Ceed ceed;
        CeedCallBackend(CeedOperatorGetCeed(op, &ceed));
        return CeedError(ceed, CEED_ERROR_BACKEND, "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
                // LCOV_EXCL_STOP
    }
    // TODO put in a function
    // Restriction
    bool is_strided;
    CeedCallBackend(CeedElemRestrictionIsStrided(Erestrict, &is_strided));
    if (!is_strided) {
      CeedCallBackend(CeedElemRestrictionGetLVectorSize(Erestrict, &lsize));
      code << "    const CeedInt lsize_out_" << i << " = " << lsize << ";\n";
      CeedInt comp_stride;
      CeedCallBackend(CeedElemRestrictionGetCompStride(Erestrict, &comp_stride));
      code << "    // CompStride: " << comp_stride << "\n";
      CeedCallBackend(CeedElemRestrictionGetData(Erestrict, &restr_data));
      data->indices.outputs[i] = restr_data->d_ind;
      code << "    writeDofsOffset" << dim << "d<num_comp_out_" << i << ", " << comp_stride << ", P_out_" << i << ">(data, lsize_out_" << i
           << ", elem, indices.outputs[" << i << "], r_v_" << i << ", d_v_" << i << ");\n";
    } else {
      bool has_backend_strides;
      CeedCallBackend(CeedElemRestrictionHasBackendStrides(Erestrict, &has_backend_strides));
      CeedInt num_elem;
      CeedCallBackend(CeedElemRestrictionGetNumElements(Erestrict, &num_elem));
      CeedInt strides[3] = {1, elem_size * num_elem, elem_size};
      if (!has_backend_strides) {
        CeedCallBackend(CeedElemRestrictionGetStrides(Erestrict, &strides));
      }
      code << "    // Strides: {" << strides[0] << ", " << strides[1] << ", " << strides[2] << "}\n";
      code << "    writeDofsStrided" << dim << "d<num_comp_out_" << i << ",P_out_" << i << "," << strides[0] << "," << strides[1] << "," << strides[2]
           << ">(data, elem, r_v_" << i << ", d_v_" << i << ");\n";
    }
  }

  code << "  }\n";
  code << "}\n";
  code << "// -----------------------------------------------------------------------------\n\n";

  // View kernel for debugging
  CeedDebug256(ceed, 2, "Generated Operator Kernels:\n");
  CeedDebug(ceed, code.str().c_str());

  CeedCallBackend(CeedCompile_Cuda(ceed, code.str().c_str(), &data->module, 1, "T_1D", CeedIntMax(Q_1d, data->max_P_1d)));
  CeedCallBackend(CeedGetKernel_Cuda(ceed, data->module, operator_name.c_str(), &data->op));

  CeedCallBackend(CeedOperatorSetSetupDone(op));
  return CEED_ERROR_SUCCESS;
}

//------------------------------------------------------------------------------