GPU workarounds for OLCF Frontier

Author: henryleberre

src/common/m_phase_change.fpp

@@ -298,7 +298,11 @@ contains
         !!  @param rhoe mixture energy
         !!  @param TS equilibrium temperature at the interface
     subroutine s_infinite_pt_relaxation_k(j, k, l, MFL, pS, p_infpT, rM, q_cons_vf, rhoe, TS)
+#ifdef CRAY_ACC_WAR
+        !DIR$ INLINEALWAYS s_compute_speed_of_sound
+#else
         !$acc routine seq
+#endif
 
         ! initializing variables
         integer, intent(in) :: j, k, l, MFL
@@ -402,7 +406,11 @@ contains
         !!  @param TS equilibrium temperature at the interface
     subroutine s_infinite_ptg_relaxation_k(j, k, l, pS, p_infpT, rhoe, q_cons_vf, TS)
 
+#ifdef CRAY_ACC_WAR
+        !DIR$ INLINEALWAYS s_infinite_ptg_relaxation_k
+#else
         !$acc routine seq
+#endif
 
         integer, intent(in) :: j, k, l
         real(kind(0.0d0)), intent(inout) :: pS
@@ -522,7 +530,11 @@ contains
         !!  @param k generic loop iterator for y direction
         !!  @param l generic loop iterator for z direction
     subroutine s_correct_partial_densities(MCT, q_cons_vf, rM, j, k, l)
+#ifdef CRAY_ACC_WAR
+        !DIR$ INLINEALWAYS s_correct_partial_densities
+#else
         !$acc routine seq
+#endif
 
         !> @name variables for the correction of the reacting partial densities
         !> @{
@@ -580,7 +592,12 @@ contains
         !!  @param q_cons_vf Cell-average conservative variables
         !!  @param TJac Transpose of the Jacobian Matrix
     subroutine s_compute_jacobian_matrix(InvJac, j, Jac, k, l, mCPD, mCVGP, mCVGP2, pS, q_cons_vf, TJac)
+
+#ifdef CRAY_ACC_WAR
+        !DIR$ INLINEALWAYS s_compute_jacobian_matrix
+#else
         !$acc routine seq
+#endif
 
         real(kind(0.0d0)), dimension(2, 2), intent(out) :: InvJac
         integer, intent(in) :: j
@@ -682,7 +699,12 @@ contains
         !!  @param rhoe mixture energy
         !!  @param R2D (2D) residue array
     subroutine s_compute_pTg_residue(j, k, l, mCPD, mCVGP, mQD, q_cons_vf, pS, rhoe, R2D)
+
+#ifdef CRAY_ACC_WAR
+        !DIR$ INLINEALWAYS s_compute_pTg_residue
+#else
         !$acc routine seq
+#endif
 
         integer, intent(in) :: j, k, l
         real(kind(0.0d0)), intent(in) :: mCPD, mCVGP, mQD
@@ -728,7 +750,11 @@ contains
         !!  @param TSat Saturation Temperature
         !!  @param TSIn equilibrium Temperature
     subroutine s_TSat(pSat, TSat, TSIn)
+#ifdef CRAY_ACC_WAR
+        !DIR$ INLINEALWAYS s_compute_speed_of_sound
+#else
         !$acc routine seq
+#endif
 
         real(kind(0.0d0)), intent(in) :: pSat
         real(kind(0.0d0)), intent(out) :: TSat

src/simulation/m_chemistry.fpp

@@ -70,16 +70,19 @@ contains
 
         #:if chemistry
 
-            !$acc parallel loop collapse(3) private(Ys)
+            !$acc parallel loop collapse(3) gang vector default(present) &
+            !$acc private(Ys, omega, enthalpies)
             do x = 0, m
                 do y = 0, n
                     do z = 0, p
 
                         rho = 0d0
+                        !$acc loop seq
                         do eqn = chemxb, chemxe
                             rho = rho + q_cons_qp(eqn)%sf(x, y, z)
                         end do
 
+                        !$acc loop seq
                         do eqn = chemxb, chemxe
                             Ys(eqn - chemxb + 1) = q_prim_qp(eqn)%sf(x, y, z)
                         end do
@@ -96,7 +99,7 @@ contains
 
                         call get_net_production_rates(rho, T, Ys, omega)
 
-                        !$acc routine seq
+                        !$acc loop seq
                         do eqn = chemxb, chemxe
 
                             omega_m = mol_weights(eqn - chemxb + 1)*omega(eqn - chemxb + 1)

src/simulation/m_rhs.fpp

@@ -1000,7 +1000,7 @@ contains
                 call s_compute_chemistry_reaction_flux(rhs_vf, q_cons_qp%vf, q_prim_qp%vf)
                 call nvtxEndRange
             else
-                !$acc parallel loop default(present)
+                !$acc parallel loop gang vector default(present)
                 do i = chemxb, chemxe
                     rhs_vf(i)%sf(:, :, :) = 0d0
                 end do

src/simulation/m_riemann_solvers.fpp

@@ -363,8 +363,11 @@ contains
         #:for NORM_DIR, XYZ in [(1, 'x'), (2, 'y'), (3, 'z')]
 
             if (norm_dir == ${NORM_DIR}$) then
-                !$acc parallel loop collapse(3) gang vector default(present) private(alpha_rho_L, alpha_rho_R, vel_L, vel_R, alpha_L, alpha_R, vel_avg, tau_e_L, tau_e_R, G_L, G_R, Re_L, Re_R, &
-                !$acc rho_avg, h_avg, gamma_avg, s_L, s_R, s_S, Y_L, Ys_L, Y_R, Ys_R)
+                !$acc parallel loop collapse(3) gang vector default(present)    &
+                !$acc private(alpha_rho_L, alpha_rho_R, vel_L, vel_R, alpha_L,  &
+                !$acc alpha_R, vel_avg, tau_e_L, tau_e_R, G_L, G_R, Re_L, Re_R, &
+                !$acc rho_avg, h_avg, gamma_avg, s_L, s_R, s_S, Y_L, Ys_L, Y_R, &
+                !$acc Ys_R)
                 do l = is3%beg, is3%end
                     do k = is2%beg, is2%end
                         do j = is1%beg, is1%end

src/simulation/m_sim_helpers.f90

@@ -31,7 +31,12 @@ module m_sim_helpers
         !! @param k y index
         !! @param l z index
     subroutine s_compute_enthalpy(q_prim_vf, pres, rho, gamma, pi_inf, Re, H, alpha, vel, vel_sum, j, k, l)
+#ifdef CRAY_ACC_WAR
+        !DIR$ INLINEALWAYS s_compute_enthalpy
+#else
         !$acc routine seq
+#endif
+
         type(scalar_field), dimension(sys_size) :: q_prim_vf
         real(kind(0d0)), dimension(num_fluids) :: alpha_rho
         real(kind(0d0)), dimension(num_fluids) :: alpha

toolchain/pyproject.toml

@@ -41,4 +41,4 @@ dependencies = [
 ]
 
 [tool.hatch.metadata]
-allow-direct-references = true
+allow-direct-references = true