Refactor libCEED QFunction source code to accelerate JIT compilation

palace/fem/integ/curlcurl.cpp

@@ -10,9 +10,7 @@
 PalacePragmaDiagnosticPush
 PalacePragmaDiagnosticDisableUnused
 
-#include "fem/qfunctions/hdiv_build_qf.h"
 #include "fem/qfunctions/hdiv_qf.h"
-#include "fem/qfunctions/l2_build_qf.h"
 #include "fem/qfunctions/l2_qf.h"
 
 PalacePragmaDiagnosticPop

palace/fem/integ/curlcurlmass.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/hdivmass_build_qf.h"
 #include "fem/qfunctions/hdivmass_qf.h"
 
 namespace palace

palace/fem/integ/diffusion.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/hcurl_build_qf.h"
 #include "fem/qfunctions/hcurl_qf.h"
 
 namespace palace

palace/fem/integ/diffusionmass.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/hcurlmass_build_qf.h"
 #include "fem/qfunctions/hcurlmass_qf.h"
 
 namespace palace

palace/fem/integ/divdiv.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/l2_build_qf.h"
 #include "fem/qfunctions/l2_qf.h"
 
 namespace palace

palace/fem/integ/divdivmass.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/l2mass_build_qf.h"
 #include "fem/qfunctions/l2mass_qf.h"
 
 namespace palace

palace/fem/integ/grad.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/hcurlh1d_build_qf.h"
 #include "fem/qfunctions/hcurlh1d_qf.h"
 
 namespace palace

palace/fem/integ/mass.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/h1_build_qf.h"
 #include "fem/qfunctions/h1_qf.h"
 
 namespace palace

palace/fem/integ/mixedveccurl.cpp

@@ -10,9 +10,7 @@
 PalacePragmaDiagnosticPush
 PalacePragmaDiagnosticDisableUnused
 
-#include "fem/qfunctions/hcurlhdiv_build_qf.h"
 #include "fem/qfunctions/hcurlhdiv_qf.h"
-#include "fem/qfunctions/hdiv_build_qf.h"
 #include "fem/qfunctions/hdiv_qf.h"
 
 PalacePragmaDiagnosticPop

palace/fem/integ/mixedvecgrad.cpp

@@ -6,7 +6,6 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/hcurl_build_qf.h"
 #include "fem/qfunctions/hcurl_qf.h"
 
 namespace palace

palace/fem/integ/vecfemass.cpp

@@ -6,11 +6,8 @@
 #include "fem/libceed/coefficient.hpp"
 #include "fem/libceed/integrator.hpp"
 
-#include "fem/qfunctions/hcurl_build_qf.h"
 #include "fem/qfunctions/hcurl_qf.h"
-#include "fem/qfunctions/hcurlhdiv_build_qf.h"
 #include "fem/qfunctions/hcurlhdiv_qf.h"
-#include "fem/qfunctions/hdiv_build_qf.h"
 #include "fem/qfunctions/hdiv_qf.h"
 
 namespace palace

palace/fem/libceed/coefficient.cpp

@@ -7,7 +7,7 @@
 #include "fem/libceed/ceed.hpp"
 #include "models/materialoperator.hpp"
 
-#include "fem/qfunctions/coeff_qf.h"
+#include "fem/qfunctions/coeff/coeff_qf.h"
 
 namespace palace::ceed
 {
@@ -36,26 +36,15 @@ auto InitDefaultCoefficient(int dim)
 
 void MakeDiagonalCoefficient(int dim, CeedIntScalar *mat_coeff, CeedScalar a, CeedInt k)
 {
-  switch (dim)
+  const int coeff_dim = CoeffDim(dim);
+  for (int i = 0; i < coeff_dim; i++)
   {
-    case 1:
-      mat_coeff[k].second = a;
-      break;
-    case 2:
-      mat_coeff[3 * k + 0].second = a;
-      mat_coeff[3 * k + 1].second = 0.0;
-      mat_coeff[3 * k + 2].second = a;
-      break;
-    case 3:
-      mat_coeff[6 * k + 0].second = a;
-      mat_coeff[6 * k + 1].second = 0.0;
-      mat_coeff[6 * k + 2].second = 0.0;
-      mat_coeff[6 * k + 3].second = a;
-      mat_coeff[6 * k + 4].second = 0.0;
-      mat_coeff[6 * k + 5].second = a;
-      break;
-    default:
-      MFEM_ABORT("Unsupported dimension for diagonal coefficient!");
+    mat_coeff[coeff_dim * k + i].second = 0.0;
+  }
+  for (int di = 0; di < dim; ++di)
+  {
+    const int idx = (di * dim) - (((di - 1) * di) / 2);
+    mat_coeff[coeff_dim * k + idx].second = a;
   }
 }
 
@@ -97,8 +86,8 @@ PopulateCoefficientContext(int dim, const MaterialPropertyCoefficient *Q, double
 
   // Map unassigned attributes to zero material property coefficient (the last material
   // property is reserved for zero).
-  std::vector<CeedIntScalar> ctx(2 + attr_mat.Size() +
-                                 CoeffDim(dim) * (mat_coeff.SizeK() + 1));
+  const int coeff_dim = CoeffDim(dim);
+  std::vector<CeedIntScalar> ctx(2 + attr_mat.Size() + coeff_dim * (mat_coeff.SizeK() + 1));
   ctx[0].first = attr_mat.Size();
   const int zero_mat = mat_coeff.SizeK();
   for (int i = 0; i < attr_mat.Size(); i++)
@@ -123,16 +112,16 @@ PopulateCoefficientContext(int dim, const MaterialPropertyCoefficient *Q, double
       {
         for (int di = dj; di < dim; ++di)
         {
+          // Column-major ordering.
           const int idx = (dj * dim) - (((dj - 1) * dj) / 2) + di - dj;
-          MatCoeff(ctx.data())[CoeffDim(dim) * k + idx].second =
-              a * mat_coeff(di, dj, k);  // Column-major
+          MatCoeff(ctx.data())[coeff_dim * k + idx].second = a * mat_coeff(di, dj, k);
         }
       }
     }
   }
-  for (int d = 0; d < CoeffDim(dim); d++)
+  for (int d = 0; d < coeff_dim; d++)
   {
-    MatCoeff(ctx.data())[CoeffDim(dim) * zero_mat + d].second = 0.0;
+    MatCoeff(ctx.data())[coeff_dim * zero_mat + d].second = 0.0;
   }
 
   return ctx;

palace/fem/qfunctions/1/h1_1_qf.h

@@ -0,0 +1,24 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_H1_1_QF_H
+#define PALACE_LIBCEED_H1_1_QF_H
+
+#include "../coeff/coeff_1_qf.h"
+
+CEED_QFUNCTION(f_apply_h1_1)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *u = in[1];
+  CeedScalar *v = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    const CeedScalar coeff = CoeffUnpack1((const CeedIntScalar *)ctx, (CeedInt)attr[i]);
+
+    v[i] = coeff * wdetJ[i] * u[i];
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_H1_1_QF_H

palace/fem/qfunctions/1/h1_build_1_qf.h

@@ -0,0 +1,24 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_H1_BUILD_1_QF_H
+#define PALACE_LIBCEED_H1_BUILD_1_QF_H
+
+#include "../coeff/coeff_1_qf.h"
+
+CEED_QFUNCTION(f_build_h1_1)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q;
+  CeedScalar *qd = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    const CeedScalar coeff = CoeffUnpack1((const CeedIntScalar *)ctx, (CeedInt)attr[i]);
+
+    qd[i] = coeff * wdetJ[i];
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_H1_BUILD_1_QF_H

palace/fem/qfunctions/1/l2_1_qf.h

@@ -0,0 +1,24 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_L2_1_QF_H
+#define PALACE_LIBCEED_L2_1_QF_H
+
+#include "../coeff/coeff_1_qf.h"
+
+CEED_QFUNCTION(f_apply_l2_1)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *qw = in[1], *u = in[2];
+  CeedScalar *v = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    const CeedScalar coeff = CoeffUnpack1((const CeedIntScalar *)ctx, (CeedInt)attr[i]);
+
+    v[i] = (coeff * qw[i] * qw[i] / wdetJ[i]) * u[i];
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_L2_1_QF_H

palace/fem/qfunctions/1/l2_build_1_qf.h

@@ -0,0 +1,24 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_L2_BUILD_1_QF_H
+#define PALACE_LIBCEED_L2_BUILD_1_QF_H
+
+#include "../coeff/coeff_1_qf.h"
+
+CEED_QFUNCTION(f_build_l2_1)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *qw = in[1];
+  CeedScalar *qd = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    const CeedScalar coeff = CoeffUnpack1((const CeedIntScalar *)ctx, (CeedInt)attr[i]);
+
+    qd[i] = coeff * qw[i] * qw[i] / wdetJ[i];
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_L2_BUILD_1_QF_H

palace/fem/qfunctions/2/h1_2_qf.h

@@ -0,0 +1,28 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_H1_2_QF_H
+#define PALACE_LIBCEED_H1_2_QF_H
+
+#include "../coeff/coeff_2_qf.h"
+
+CEED_QFUNCTION(f_apply_h1_2)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *u = in[1];
+  CeedScalar *v = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    CeedScalar coeff[3];
+    CoeffUnpack2((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
+
+    const CeedScalar u0 = u[i + Q * 0];
+    const CeedScalar u1 = u[i + Q * 1];
+    v[i + Q * 0] = wdetJ[i] * (coeff[0] * u0 + coeff[1] * u1);
+    v[i + Q * 1] = wdetJ[i] * (coeff[1] * u0 + coeff[2] * u1);
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_H1_2_QF_H

palace/fem/qfunctions/2/h1_build_2_qf.h

@@ -0,0 +1,27 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_H1_BUILD_2_QF_H
+#define PALACE_LIBCEED_H1_BUILD_2_QF_H
+
+#include "../coeff/coeff_2_qf.h"
+
+CEED_QFUNCTION(f_build_h1_2)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q;
+  CeedScalar *qd = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    CeedScalar coeff[3];
+    CoeffUnpack2((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
+
+    qd[i + Q * 0] = wdetJ[i] * coeff[0];
+    qd[i + Q * 1] = wdetJ[i] * coeff[1];
+    qd[i + Q * 2] = wdetJ[i] * coeff[2];
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_H1_BUILD_2_QF_H

palace/fem/qfunctions/2/l2_2_qf.h

@@ -0,0 +1,29 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_L2_2_QF_H
+#define PALACE_LIBCEED_L2_2_QF_H
+
+#include "../coeff/coeff_2_qf.h"
+
+CEED_QFUNCTION(f_apply_l2_2)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *qw = in[1], *u = in[2];
+  CeedScalar *v = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    CeedScalar coeff[3];
+    CoeffUnpack2((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
+    const CeedScalar w = qw[i] * qw[i] / wdetJ[i];
+
+    const CeedScalar u0 = u[i + Q * 0];
+    const CeedScalar u1 = u[i + Q * 1];
+    v[i + Q * 0] = w * (coeff[0] * u0 + coeff[1] * u1);
+    v[i + Q * 1] = w * (coeff[1] * u0 + coeff[2] * u1);
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_L2_2_QF_H

palace/fem/qfunctions/2/l2_build_2_qf.h

@@ -0,0 +1,28 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+#ifndef PALACE_LIBCEED_L2_BUILD_2_QF_H
+#define PALACE_LIBCEED_L2_BUILD_2_QF_H
+
+#include "../coeff/coeff_2_qf.h"
+
+CEED_QFUNCTION(f_build_l2_2)(void *__restrict__ ctx, CeedInt Q, const CeedScalar *const *in,
+                             CeedScalar *const *out)
+{
+  const CeedScalar *attr = in[0], *wdetJ = in[0] + Q, *qw = in[1];
+  CeedScalar *qd = out[0];
+
+  CeedPragmaSIMD for (CeedInt i = 0; i < Q; i++)
+  {
+    CeedScalar coeff[3];
+    CoeffUnpack2((const CeedIntScalar *)ctx, (CeedInt)attr[i], coeff);
+    const CeedScalar w = qw[i] * qw[i] / wdetJ[i];
+
+    qd[i + Q * 0] = w * coeff[0];
+    qd[i + Q * 1] = w * coeff[1];
+    qd[i + Q * 2] = w * coeff[2];
+  }
+  return 0;
+}
+
+#endif  // PALACE_LIBCEED_L2_BUILD_2_QF_H