xref: /libCEED/rust/libceed-sys/c-src/backends/magma/ceed-magma-basis.c (revision 00fb7a044a7fd1c8bfdb0605078b0c7ba7a4ad58)

// 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

#include <ceed.h>
#include <ceed/backend.h>
#include <ceed/jit-tools.h>
#include <string.h>

#ifdef CEED_MAGMA_USE_HIP
#include "../hip/ceed-hip-common.h"
#include "../hip/ceed-hip-compile.h"
#else
#include "../cuda/ceed-cuda-common.h"
#include "../cuda/ceed-cuda-compile.h"
#endif
#include "ceed-magma-common.h"
#include "ceed-magma.h"

#ifdef __cplusplus
CEED_INTERN "C"
#endif
    int
    CeedBasisApply_Magma(CeedBasis basis, CeedInt nelem, CeedTransposeMode tmode, CeedEvalMode emode, CeedVector U, CeedVector V) {
  Ceed ceed;
  CeedCallBackend(CeedBasisGetCeed(basis, &ceed));
  CeedInt dim, ncomp, ndof;
  CeedCallBackend(CeedBasisGetDimension(basis, &dim));
  CeedCallBackend(CeedBasisGetNumComponents(basis, &ncomp));
  CeedCallBackend(CeedBasisGetNumNodes(basis, &ndof));

  Ceed_Magma *data;
  CeedCallBackend(CeedGetData(ceed, &data));

  const CeedScalar *du;
  CeedScalar       *dv;
  if (U != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(U, CEED_MEM_DEVICE, &du));
  else CeedCheck(emode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode");
  CeedCallBackend(CeedVectorGetArrayWrite(V, CEED_MEM_DEVICE, &dv));

  CeedBasis_Magma *impl;
  CeedCallBackend(CeedBasisGetData(basis, &impl));

  CeedInt P1d, Q1d;
  CeedCallBackend(CeedBasisGetNumNodes1D(basis, &P1d));
  CeedCallBackend(CeedBasisGetNumQuadraturePoints1D(basis, &Q1d));

  CeedDebug256(ceed, 4, "[CeedBasisApply_Magma] vsize=%" CeedInt_FMT ", comp = %" CeedInt_FMT, ncomp * CeedIntPow(P1d, dim), ncomp);

  if (tmode == CEED_TRANSPOSE) {
    CeedSize length;
    CeedCallBackend(CeedVectorGetLength(V, &length));
    if (CEED_SCALAR_TYPE == CEED_SCALAR_FP32) {
      magmablas_slaset(MagmaFull, length, 1, 0., 0., (float *)dv, length, data->queue);
    } else {
      magmablas_dlaset(MagmaFull, length, 1, 0., 0., (double *)dv, length, data->queue);
    }
    ceed_magma_queue_sync(data->queue);
  }

  switch (emode) {
    case CEED_EVAL_INTERP: {
      CeedInt P = P1d, Q = Q1d;
      if (tmode == CEED_TRANSPOSE) {
        P = Q1d;
        Q = P1d;
      }

      // Define element sizes for dofs/quad
      CeedInt elquadsize = CeedIntPow(Q1d, dim);
      CeedInt eldofssize = CeedIntPow(P1d, dim);

      // E-vector ordering -------------- Q-vector ordering
      //  component                        component
      //    elem                             elem
      //       node                            node

      // ---  Define strides for NOTRANSPOSE mode: ---
      // Input (du) is E-vector, output (dv) is Q-vector

      // Element strides
      CeedInt u_elstride = eldofssize;
      CeedInt v_elstride = elquadsize;
      // Component strides
      CeedInt u_compstride = nelem * eldofssize;
      CeedInt v_compstride = nelem * elquadsize;

      // ---  Swap strides for TRANSPOSE mode: ---
      if (tmode == CEED_TRANSPOSE) {
        // Input (du) is Q-vector, output (dv) is E-vector
        // Element strides
        v_elstride = eldofssize;
        u_elstride = elquadsize;
        // Component strides
        v_compstride = nelem * eldofssize;
        u_compstride = nelem * elquadsize;
      }

      CeedInt nthreads = 1;
      CeedInt ntcol    = 1;
      CeedInt shmem    = 0;
      CeedInt maxPQ    = CeedIntMax(P, Q);

      switch (dim) {
        case 1:
          nthreads = maxPQ;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_1D);
          shmem += sizeof(CeedScalar) * ntcol * (ncomp * (1 * P + 1 * Q));
          shmem += sizeof(CeedScalar) * (P * Q);
          break;
        case 2:
          nthreads = maxPQ;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_2D);
          shmem += P * Q * sizeof(CeedScalar);                // for sT
          shmem += ntcol * (P * maxPQ * sizeof(CeedScalar));  // for reforming rU we need PxP, and for the intermediate output we need PxQ
          break;
        case 3:
          nthreads = maxPQ * maxPQ;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_3D);
          shmem += sizeof(CeedScalar) * (P * Q);  // for sT
          shmem += sizeof(CeedScalar) * ntcol *
                   (CeedIntMax(P * P * maxPQ,
                               P * Q * Q));  // rU needs P^2xP, the intermediate output needs max(P^2xQ,PQ^2)
      }
      CeedInt grid   = (nelem + ntcol - 1) / ntcol;
      void   *args[] = {&impl->dinterp1d, &du, &u_elstride, &u_compstride, &dv, &v_elstride, &v_compstride, &nelem};

      if (tmode == CEED_TRANSPOSE) {
        CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->magma_interp_tr, grid, nthreads, ntcol, 1, shmem, args));
      } else {
        CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->magma_interp, grid, nthreads, ntcol, 1, shmem, args));
      }
    } break;
    case CEED_EVAL_GRAD: {
      CeedInt P = P1d, Q = Q1d;
      // In CEED_NOTRANSPOSE mode:
      // du is (P^dim x nc), column-major layout (nc = ncomp)
      // dv is (Q^dim x nc x dim), column-major layout (nc = ncomp)
      // In CEED_TRANSPOSE mode, the sizes of du and dv are switched.
      if (tmode == CEED_TRANSPOSE) {
        P = Q1d, Q = P1d;
      }

      // Define element sizes for dofs/quad
      CeedInt elquadsize = CeedIntPow(Q1d, dim);
      CeedInt eldofssize = CeedIntPow(P1d, dim);

      // E-vector ordering -------------- Q-vector ordering
      //                                  dim
      //  component                        component
      //    elem                              elem
      //       node                            node

      // ---  Define strides for NOTRANSPOSE mode: ---
      // Input (du) is E-vector, output (dv) is Q-vector

      // Element strides
      CeedInt u_elstride = eldofssize;
      CeedInt v_elstride = elquadsize;
      // Component strides
      CeedInt u_compstride = nelem * eldofssize;
      CeedInt v_compstride = nelem * elquadsize;
      // Dimension strides
      CeedInt u_dimstride = 0;
      CeedInt v_dimstride = nelem * elquadsize * ncomp;

      // ---  Swap strides for TRANSPOSE mode: ---
      if (tmode == CEED_TRANSPOSE) {
        // Input (du) is Q-vector, output (dv) is E-vector
        // Element strides
        v_elstride = eldofssize;
        u_elstride = elquadsize;
        // Component strides
        v_compstride = nelem * eldofssize;
        u_compstride = nelem * elquadsize;
        // Dimension strides
        v_dimstride = 0;
        u_dimstride = nelem * elquadsize * ncomp;
      }

      CeedInt nthreads = 1;
      CeedInt ntcol    = 1;
      CeedInt shmem    = 0;
      CeedInt maxPQ    = CeedIntMax(P, Q);

      switch (dim) {
        case 1:
          nthreads = maxPQ;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_1D);
          shmem += sizeof(CeedScalar) * ntcol * (ncomp * (1 * P + 1 * Q));
          shmem += sizeof(CeedScalar) * (P * Q);
          break;
        case 2:
          nthreads = maxPQ;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_2D);
          shmem += sizeof(CeedScalar) * 2 * P * Q;            // for sTinterp and sTgrad
          shmem += sizeof(CeedScalar) * ntcol * (P * maxPQ);  // for reforming rU we need PxP, and for the intermediate output we need PxQ
          break;
        case 3:
          nthreads = maxPQ * maxPQ;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_3D);
          shmem += sizeof(CeedScalar) * 2 * P * Q;  // for sTinterp and sTgrad
          shmem += sizeof(CeedScalar) * ntcol *
                   CeedIntMax(P * P * P,
                              (P * P * Q) + (P * Q * Q));  // rU needs P^2xP, the intermediate outputs need (P^2.Q + P.Q^2)
      }
      CeedInt grid   = (nelem + ntcol - 1) / ntcol;
      void   *args[] = {&impl->dinterp1d, &impl->dgrad1d, &du,          &u_elstride, &u_compstride, &u_dimstride, &dv,
                        &v_elstride,      &v_compstride,  &v_dimstride, &nelem};

      if (tmode == CEED_TRANSPOSE) {
        CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->magma_grad_tr, grid, nthreads, ntcol, 1, shmem, args));
      } else {
        CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->magma_grad, grid, nthreads, ntcol, 1, shmem, args));
      }
    } break;
    case CEED_EVAL_WEIGHT: {
      CeedCheck(tmode != CEED_TRANSPOSE, ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT incompatible with CEED_TRANSPOSE");
      CeedInt Q          = Q1d;
      CeedInt eldofssize = CeedIntPow(Q, dim);
      CeedInt nthreads   = 1;
      CeedInt ntcol      = 1;
      CeedInt shmem      = 0;

      switch (dim) {
        case 1:
          nthreads = Q;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_1D);
          shmem += sizeof(CeedScalar) * Q;          // for dqweight1d
          shmem += sizeof(CeedScalar) * ntcol * Q;  // for output
          break;
        case 2:
          nthreads = Q;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_2D);
          shmem += sizeof(CeedScalar) * Q;  // for dqweight1d
          break;
        case 3:
          nthreads = Q * Q;
          ntcol    = MAGMA_BASIS_NTCOL(nthreads, MAGMA_MAXTHREADS_3D);
          shmem += sizeof(CeedScalar) * Q;  // for dqweight1d
      }
      CeedInt grid   = (nelem + ntcol - 1) / ntcol;
      void   *args[] = {&impl->dqweight1d, &dv, &eldofssize, &nelem};

      CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, impl->magma_weight, grid, nthreads, ntcol, 1, shmem, args));
    } break;
    // LCOV_EXCL_START
    case CEED_EVAL_DIV:
      return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported");
    case CEED_EVAL_CURL:
      return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported");
    case CEED_EVAL_NONE:
      return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context");
      // LCOV_EXCL_STOP
  }

  // must sync to ensure completeness
  ceed_magma_queue_sync(data->queue);

  if (emode != CEED_EVAL_WEIGHT) {
    CeedCallBackend(CeedVectorRestoreArrayRead(U, &du));
  }
  CeedCallBackend(CeedVectorRestoreArray(V, &dv));
  return CEED_ERROR_SUCCESS;
}

#ifdef __cplusplus
CEED_INTERN "C"
#endif
    int
    CeedBasisApplyNonTensor_Magma(CeedBasis basis, CeedInt nelem, CeedTransposeMode tmode, CeedEvalMode emode, CeedVector U, CeedVector V) {
  Ceed ceed;
  CeedCallBackend(CeedBasisGetCeed(basis, &ceed));

  Ceed_Magma *data;
  CeedCallBackend(CeedGetData(ceed, &data));

  magma_int_t arch = magma_getdevice_arch();

  CeedInt dim, ncomp, ndof, nqpt;
  CeedCallBackend(CeedBasisGetDimension(basis, &dim));
  CeedCallBackend(CeedBasisGetNumComponents(basis, &ncomp));
  CeedCallBackend(CeedBasisGetNumNodes(basis, &ndof));
  CeedCallBackend(CeedBasisGetNumQuadraturePoints(basis, &nqpt));
  const CeedScalar *du;
  CeedScalar       *dv;
  if (U != CEED_VECTOR_NONE) CeedCallBackend(CeedVectorGetArrayRead(U, CEED_MEM_DEVICE, &du));
  else CeedCheck(emode == CEED_EVAL_WEIGHT, ceed, CEED_ERROR_BACKEND, "An input vector is required for this CeedEvalMode");
  CeedCallBackend(CeedVectorGetArrayWrite(V, CEED_MEM_DEVICE, &dv));

  CeedBasisNonTensor_Magma *impl;
  CeedCallBackend(CeedBasisGetData(basis, &impl));

  CeedDebug256(ceed, 4, "[CeedBasisApplyNonTensor_Magma] vsize=%" CeedInt_FMT ", comp = %" CeedInt_FMT, ncomp * ndof, ncomp);

  if (tmode == CEED_TRANSPOSE) {
    CeedSize length;
    CeedCallBackend(CeedVectorGetLength(V, &length));
    if (CEED_SCALAR_TYPE == CEED_SCALAR_FP32) {
      magmablas_slaset(MagmaFull, length, 1, 0., 0., (float *)dv, length, data->queue);
    } else {
      magmablas_dlaset(MagmaFull, length, 1, 0., 0., (double *)dv, length, data->queue);
    }
    ceed_magma_queue_sync(data->queue);
  }

  CeedInt            P = ndof, Q = nqpt, N = nelem * ncomp;
  CeedInt            NB = 1;
  CeedMagmaFunction *interp, *grad;

  CeedInt Narray[MAGMA_NONTENSOR_KERNEL_INSTANCES] = {MAGMA_NONTENSOR_N_VALUES};
  CeedInt iN                                       = 0;
  CeedInt diff                                     = abs(Narray[iN] - N);
  for (CeedInt in = iN + 1; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) {
    CeedInt idiff = abs(Narray[in] - N);
    if (idiff < diff) {
      iN   = in;
      diff = idiff;
    }
  }

  NB     = nontensor_rtc_get_nb(arch, 'd', emode, tmode, P, Narray[iN], Q);
  interp = (tmode == CEED_TRANSPOSE) ? &impl->magma_interp_tr_nontensor[iN] : &impl->magma_interp_nontensor[iN];
  grad   = (tmode == CEED_TRANSPOSE) ? &impl->magma_grad_tr_nontensor[iN] : &impl->magma_grad_nontensor[iN];

  switch (emode) {
    case CEED_EVAL_INTERP: {
      CeedInt P = ndof, Q = nqpt;
      if (P < MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P && Q < MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) {
        CeedInt M     = (tmode == CEED_TRANSPOSE) ? P : Q;
        CeedInt K     = (tmode == CEED_TRANSPOSE) ? Q : P;
        CeedInt ntcol = MAGMA_NONTENSOR_BASIS_NTCOL(M);
        CeedInt shmem = 0, shmemA = 0, shmemB = 0;
        shmemB += ntcol * K * NB * sizeof(CeedScalar);
        shmemA += (tmode == CEED_TRANSPOSE) ? 0 : K * M * sizeof(CeedScalar);
        shmem = (tmode == CEED_TRANSPOSE) ? (shmemA + shmemB) : CeedIntMax(shmemA, shmemB);

        CeedInt       grid   = MAGMA_CEILDIV(MAGMA_CEILDIV(N, NB), ntcol);
        magma_trans_t transA = (tmode == CEED_TRANSPOSE) ? MagmaNoTrans : MagmaTrans;
        magma_trans_t transB = MagmaNoTrans;
        CeedScalar    alpha = 1.0, beta = 0.0;

        void *args[] = {&transA, &transB, &N, &alpha, &impl->dinterp, &P, &du, &K, &beta, &dv, &M};
        CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, *interp, grid, M, ntcol, 1, shmem, args));
      } else {
        if (tmode == CEED_TRANSPOSE)
          magma_gemm_nontensor(MagmaNoTrans, MagmaNoTrans, P, nelem * ncomp, Q, 1.0, impl->dinterp, P, du, Q, 0.0, dv, P, data->queue);
        else magma_gemm_nontensor(MagmaTrans, MagmaNoTrans, Q, nelem * ncomp, P, 1.0, impl->dinterp, P, du, P, 0.0, dv, Q, data->queue);
      }
    } break;

    case CEED_EVAL_GRAD: {
      CeedInt P = ndof, Q = nqpt;
      if (P < MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P && Q < MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q) {
        CeedInt M     = (tmode == CEED_TRANSPOSE) ? P : Q;
        CeedInt K     = (tmode == CEED_TRANSPOSE) ? Q : P;
        CeedInt ntcol = MAGMA_NONTENSOR_BASIS_NTCOL(M);
        CeedInt shmem = 0, shmemA = 0, shmemB = 0;
        shmemB += ntcol * K * NB * sizeof(CeedScalar);
        shmemA += (tmode == CEED_TRANSPOSE) ? 0 : K * M * sizeof(CeedScalar);
        shmem = shmemA + shmemB;

        CeedInt       grid   = MAGMA_CEILDIV(MAGMA_CEILDIV(N, NB), ntcol);
        magma_trans_t transA = (tmode == CEED_TRANSPOSE) ? MagmaNoTrans : MagmaTrans;
        magma_trans_t transB = MagmaNoTrans;

        void *args[] = {&transA, &transB, &N, &impl->dgrad, &P, &du, &K, &dv, &M};
        CeedCallBackend(CeedRunKernelDimSharedMagma(ceed, *grad, grid, M, ntcol, 1, shmem, args));
      } else {
        if (tmode == CEED_TRANSPOSE) {
          CeedScalar beta = 0.0;
          for (int d = 0; d < dim; d++) {
            if (d > 0) beta = 1.0;
            magma_gemm_nontensor(MagmaNoTrans, MagmaNoTrans, P, nelem * ncomp, Q, 1.0, impl->dgrad + d * P * Q, P, du + d * nelem * ncomp * Q, Q,
                                 beta, dv, P, data->queue);
          }
        } else {
          for (int d = 0; d < dim; d++)
            magma_gemm_nontensor(MagmaTrans, MagmaNoTrans, Q, nelem * ncomp, P, 1.0, impl->dgrad + d * P * Q, P, du, P, 0.0,
                                 dv + d * nelem * ncomp * Q, Q, data->queue);
        }
      }
    } break;

    case CEED_EVAL_WEIGHT: {
      CeedCheck(tmode != CEED_TRANSPOSE, ceed, CEED_ERROR_BACKEND, "CEED_EVAL_WEIGHT incompatible with CEED_TRANSPOSE");

      int elemsPerBlock = 1;  // basis->Q1d < 7 ? optElems[basis->Q1d] : 1;
      int grid          = nelem / elemsPerBlock + ((nelem / elemsPerBlock * elemsPerBlock < nelem) ? 1 : 0);
      magma_weight_nontensor(grid, nqpt, nelem, nqpt, impl->dqweight, dv, data->queue);
    } break;

    // LCOV_EXCL_START
    case CEED_EVAL_DIV:
      return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_DIV not supported");
    case CEED_EVAL_CURL:
      return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_CURL not supported");
    case CEED_EVAL_NONE:
      return CeedError(ceed, CEED_ERROR_BACKEND, "CEED_EVAL_NONE does not make sense in this context");
      // LCOV_EXCL_STOP
  }

  // must sync to ensure completeness
  ceed_magma_queue_sync(data->queue);

  if (emode != CEED_EVAL_WEIGHT) {
    CeedCallBackend(CeedVectorRestoreArrayRead(U, &du));
  }
  CeedCallBackend(CeedVectorRestoreArray(V, &dv));
  return CEED_ERROR_SUCCESS;
}

#ifdef __cplusplus
CEED_INTERN "C"
#endif
    int
    CeedBasisDestroy_Magma(CeedBasis basis) {
  CeedBasis_Magma *impl;
  CeedCallBackend(CeedBasisGetData(basis, &impl));

  CeedCallBackend(magma_free(impl->dqref1d));
  CeedCallBackend(magma_free(impl->dinterp1d));
  CeedCallBackend(magma_free(impl->dgrad1d));
  CeedCallBackend(magma_free(impl->dqweight1d));
  Ceed ceed;
  CeedCallBackend(CeedBasisGetCeed(basis, &ceed));
#ifdef CEED_MAGMA_USE_HIP
  CeedCallHip(ceed, hipModuleUnload(impl->module));
#else
  CeedCallCuda(ceed, cuModuleUnload(impl->module));
#endif

  CeedCallBackend(CeedFree(&impl));

  return CEED_ERROR_SUCCESS;
}

#ifdef __cplusplus
CEED_INTERN "C"
#endif
    int
    CeedBasisDestroyNonTensor_Magma(CeedBasis basis) {
  CeedBasisNonTensor_Magma *impl;
  CeedCallBackend(CeedBasisGetData(basis, &impl));

  CeedCallBackend(magma_free(impl->dqref));
  CeedCallBackend(magma_free(impl->dinterp));
  CeedCallBackend(magma_free(impl->dgrad));
  CeedCallBackend(magma_free(impl->dqweight));
  Ceed ceed;
  CeedCallBackend(CeedBasisGetCeed(basis, &ceed));
#ifdef CEED_MAGMA_USE_HIP
  for (CeedInt in = 0; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) {
    CeedCallHip(ceed, hipModuleUnload(impl->module[in]));
  }
#else
  for (CeedInt in = 0; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) {
    CeedCallCuda(ceed, cuModuleUnload(impl->module[in]));
  }
#endif
  CeedCallBackend(CeedFree(&impl));

  return CEED_ERROR_SUCCESS;
}

#ifdef __cplusplus
CEED_INTERN "C"
#endif
    int
    CeedBasisCreateTensorH1_Magma(CeedInt dim, CeedInt P1d, CeedInt Q1d, const CeedScalar *interp1d, const CeedScalar *grad1d,
                                  const CeedScalar *qref1d, const CeedScalar *qweight1d, CeedBasis basis) {
  CeedBasis_Magma *impl;
  CeedCallBackend(CeedCalloc(1, &impl));
  Ceed ceed;
  CeedCallBackend(CeedBasisGetCeed(basis, &ceed));

  // Check for supported parameters
  CeedInt ncomp = 0;
  CeedCallBackend(CeedBasisGetNumComponents(basis, &ncomp));
  Ceed_Magma *data;
  CeedCallBackend(CeedGetData(ceed, &data));

  // Compile kernels
  char *magma_common_path;
  char *interp_path, *grad_path, *weight_path;
  char *basis_kernel_source;
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma_common_defs.h", &magma_common_path));
  CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n");
  CeedCallBackend(CeedLoadSourceToBuffer(ceed, magma_common_path, &basis_kernel_source));
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma_common_tensor.h", &magma_common_path));
  CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, magma_common_path, &basis_kernel_source));
  char   *interp_name_base = "ceed/jit-source/magma/interp";
  CeedInt interp_name_len  = strlen(interp_name_base) + 6;
  char    interp_name[interp_name_len];
  snprintf(interp_name, interp_name_len, "%s-%" CeedInt_FMT "d.h", interp_name_base, dim);
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, interp_name, &interp_path));
  CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, interp_path, &basis_kernel_source));
  char   *grad_name_base = "ceed/jit-source/magma/grad";
  CeedInt grad_name_len  = strlen(grad_name_base) + 6;
  char    grad_name[grad_name_len];
  snprintf(grad_name, grad_name_len, "%s-%" CeedInt_FMT "d.h", grad_name_base, dim);
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, grad_name, &grad_path));
  CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, grad_path, &basis_kernel_source));
  char   *weight_name_base = "ceed/jit-source/magma/weight";
  CeedInt weight_name_len  = strlen(weight_name_base) + 6;
  char    weight_name[weight_name_len];
  snprintf(weight_name, weight_name_len, "%s-%" CeedInt_FMT "d.h", weight_name_base, dim);
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, weight_name, &weight_path));
  CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, weight_path, &basis_kernel_source));
  CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source Complete! -----\n");
  // The RTC compilation code expects a Ceed with the common Ceed_Cuda or Ceed_Hip
  // data
  Ceed delegate;
  CeedCallBackend(CeedGetDelegate(ceed, &delegate));
  CeedCallBackend(CeedCompileMagma(delegate, basis_kernel_source, &impl->module, 5, "DIM", dim, "NCOMP", ncomp, "P", P1d, "Q", Q1d, "MAXPQ",
                                   CeedIntMax(P1d, Q1d)));

  // Kernel setup
  switch (dim) {
    case 1:
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_1d_kernel", &impl->magma_interp));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_1d_kernel", &impl->magma_interp_tr));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_1d_kernel", &impl->magma_grad));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_1d_kernel", &impl->magma_grad_tr));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_1d_kernel", &impl->magma_weight));
      break;
    case 2:
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_2d_kernel", &impl->magma_interp));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_2d_kernel", &impl->magma_interp_tr));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_2d_kernel", &impl->magma_grad));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_2d_kernel", &impl->magma_grad_tr));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_2d_kernel", &impl->magma_weight));
      break;
    case 3:
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpn_3d_kernel", &impl->magma_interp));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_interpt_3d_kernel", &impl->magma_interp_tr));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradn_3d_kernel", &impl->magma_grad));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_gradt_3d_kernel", &impl->magma_grad_tr));
      CeedCallBackend(CeedGetKernelMagma(ceed, impl->module, "magma_weight_3d_kernel", &impl->magma_weight));
  }

  CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApply_Magma));
  CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroy_Magma));

  // Copy qref1d to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dqref1d, Q1d * sizeof(qref1d[0])));
  magma_setvector(Q1d, sizeof(qref1d[0]), qref1d, 1, impl->dqref1d, 1, data->queue);

  // Copy interp1d to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dinterp1d, Q1d * P1d * sizeof(interp1d[0])));
  magma_setvector(Q1d * P1d, sizeof(interp1d[0]), interp1d, 1, impl->dinterp1d, 1, data->queue);

  // Copy grad1d to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dgrad1d, Q1d * P1d * sizeof(grad1d[0])));
  magma_setvector(Q1d * P1d, sizeof(grad1d[0]), grad1d, 1, impl->dgrad1d, 1, data->queue);

  // Copy qweight1d to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dqweight1d, Q1d * sizeof(qweight1d[0])));
  magma_setvector(Q1d, sizeof(qweight1d[0]), qweight1d, 1, impl->dqweight1d, 1, data->queue);

  CeedCallBackend(CeedBasisSetData(basis, impl));
  CeedCallBackend(CeedFree(&magma_common_path));
  CeedCallBackend(CeedFree(&interp_path));
  CeedCallBackend(CeedFree(&grad_path));
  CeedCallBackend(CeedFree(&weight_path));
  CeedCallBackend(CeedFree(&basis_kernel_source));

  return CEED_ERROR_SUCCESS;
}

#ifdef __cplusplus
CEED_INTERN "C"
#endif
    int
    CeedBasisCreateH1_Magma(CeedElemTopology topo, CeedInt dim, CeedInt ndof, CeedInt nqpts, const CeedScalar *interp, const CeedScalar *grad,
                            const CeedScalar *qref, const CeedScalar *qweight, CeedBasis basis) {
  CeedBasisNonTensor_Magma *impl;
  Ceed                      ceed;
  CeedCallBackend(CeedBasisGetCeed(basis, &ceed));

  Ceed_Magma *data;
  CeedCallBackend(CeedGetData(ceed, &data));
  magma_int_t arch = magma_getdevice_arch();
  CeedCallBackend(CeedCalloc(1, &impl));
  // Compile kernels
  char *magma_common_path;
  char *interp_path, *grad_path;
  char *basis_kernel_source;
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma_common_defs.h", &magma_common_path));
  CeedDebug256(ceed, CEED_DEBUG_COLOR_SUCCESS, "----- Loading Basis Kernel Source -----\n");
  CeedCallBackend(CeedLoadSourceToBuffer(ceed, magma_common_path, &basis_kernel_source));
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/magma_common_nontensor.h", &magma_common_path));
  CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, magma_common_path, &basis_kernel_source));
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/interp-nontensor.h", &interp_path));
  CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, interp_path, &basis_kernel_source));
  CeedCallBackend(CeedGetJitAbsolutePath(ceed, "ceed/jit-source/magma/grad-nontensor.h", &grad_path));
  CeedCallBackend(CeedLoadSourceToInitializedBuffer(ceed, grad_path, &basis_kernel_source));

  // tuning parameters for nb
  CeedInt nb_interp_n[MAGMA_NONTENSOR_KERNEL_INSTANCES];
  CeedInt nb_interp_t[MAGMA_NONTENSOR_KERNEL_INSTANCES];
  CeedInt nb_grad_n[MAGMA_NONTENSOR_KERNEL_INSTANCES];
  CeedInt nb_grad_t[MAGMA_NONTENSOR_KERNEL_INSTANCES];
  CeedInt P = ndof, Q = nqpts;
  CeedInt Narray[MAGMA_NONTENSOR_KERNEL_INSTANCES] = {MAGMA_NONTENSOR_N_VALUES};
  for (CeedInt in = 0; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) {
    nb_interp_n[in] = nontensor_rtc_get_nb(arch, 'd', CEED_EVAL_INTERP, CEED_NOTRANSPOSE, P, Narray[in], Q);
    nb_interp_t[in] = nontensor_rtc_get_nb(arch, 'd', CEED_EVAL_INTERP, CEED_TRANSPOSE, P, Narray[in], Q);
    nb_grad_n[in]   = nontensor_rtc_get_nb(arch, 'd', CEED_EVAL_GRAD, CEED_NOTRANSPOSE, P, Narray[in], Q);
    nb_grad_t[in]   = nontensor_rtc_get_nb(arch, 'd', CEED_EVAL_GRAD, CEED_TRANSPOSE, P, Narray[in], Q);
  }

  // compile
  Ceed delegate;
  CeedCallBackend(CeedGetDelegate(ceed, &delegate));
  for (CeedInt in = 0; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) {
    CeedCallBackend(CeedCompileMagma(delegate, basis_kernel_source, &impl->module[in], 7, "DIM", dim, "P", P, "Q", Q, "NB_INTERP_N", nb_interp_n[in],
                                     "NB_INTERP_T", nb_interp_t[in], "NB_GRAD_N", nb_grad_n[in], "NB_GRAD_T", nb_grad_t[in]));
  }

  // get kernels
  for (CeedInt in = 0; in < MAGMA_NONTENSOR_KERNEL_INSTANCES; in++) {
    CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[in], "magma_interp_nontensor_n", &impl->magma_interp_nontensor[in]));
    CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[in], "magma_interp_nontensor_t", &impl->magma_interp_tr_nontensor[in]));
    CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[in], "magma_grad_nontensor_n", &impl->magma_grad_nontensor[in]));
    CeedCallBackend(CeedGetKernelMagma(ceed, impl->module[in], "magma_grad_nontensor_t", &impl->magma_grad_tr_nontensor[in]));
  }

  CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Apply", CeedBasisApplyNonTensor_Magma));
  CeedCallBackend(CeedSetBackendFunction(ceed, "Basis", basis, "Destroy", CeedBasisDestroyNonTensor_Magma));

  // Copy qref to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dqref, nqpts * sizeof(qref[0])));
  magma_setvector(nqpts, sizeof(qref[0]), qref, 1, impl->dqref, 1, data->queue);

  // Copy interp to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dinterp, nqpts * ndof * sizeof(interp[0])));
  magma_setvector(nqpts * ndof, sizeof(interp[0]), interp, 1, impl->dinterp, 1, data->queue);

  // Copy grad to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dgrad, nqpts * ndof * dim * sizeof(grad[0])));
  magma_setvector(nqpts * ndof * dim, sizeof(grad[0]), grad, 1, impl->dgrad, 1, data->queue);

  // Copy qweight to the GPU
  CeedCallBackend(magma_malloc((void **)&impl->dqweight, nqpts * sizeof(qweight[0])));
  magma_setvector(nqpts, sizeof(qweight[0]), qweight, 1, impl->dqweight, 1, data->queue);

  CeedCallBackend(CeedBasisSetData(basis, impl));
  CeedCallBackend(CeedFree(&magma_common_path));
  CeedCallBackend(CeedFree(&interp_path));
  CeedCallBackend(CeedFree(&grad_path));
  CeedCallBackend(CeedFree(&basis_kernel_source));
  return CEED_ERROR_SUCCESS;
}