src/simulation/m_weno.fpp (2)

1107-1109: Optional: hoist the shared denominator and consider applying the same unroll to order-3/order-7.

Two minor points on the order-5 omega normalization changes at 1107-1109 and 1148-1150:

1. The expression (alpha(0) + alpha(1) + alpha(2)) is recomputed three times per side. A compiler will likely CSE, but an explicit temporary makes intent clearer and is cheaper to read/verify:

   -                                        omega(0) = alpha(0)/(alpha(0) + alpha(1) + alpha(2))
   -                                        omega(1) = alpha(1)/(alpha(0) + alpha(1) + alpha(2))
   -                                        omega(2) = alpha(2)/(alpha(0) + alpha(1) + alpha(2))
   +                                        tau = alpha(0) + alpha(1) + alpha(2)   ! reuse scratch or add a local
   +                                        omega(0) = alpha(0)/tau
   +                                        omega(1) = alpha(1)/tau
   +                                        omega(2) = alpha(2)/tau

   (Use a dedicated local, not tau, if it's still live elsewhere — here it is not after this point in the left block. For the right block at 1148-1150, tau is still reusable by wenoz above, so add a fresh scalar and declare it private in the outer GPU_PARALLEL_LOOP.)

2. The same omega = alpha/sum(alpha) pattern still exists at lines 986, 1015 (order-3), 1304, 1366 (order-7), and inside the order-5 mapped_weno/teno branches (1070, 1094). If the AMD Flang perf regression you targeted is specifically the array sum() reduction, consider applying the unroll uniformly (or justify in a comment why only the order-5 final normalization was unrolled). As-is, the half-done refactor is a small consistency wart.

---

965-983: Add GPU_LOOP(parallelism='[seq]') markers to new q loops for consistency with adjacent code.

The newly introduced do q = 0, weno_num_stencils loops inside GPU_PARALLEL_LOOP kernels (order-3 at lines 965–983 and 996–1012, order-5 at lines 1063–1075 and 1123–1135, order-7 at lines 1264–1276 and 1340–1352) lack explicit $:GPU_LOOP(parallelism='[seq]') annotations, while sibling wenoz/teno q loops in the same kernels carry these markers (e.g., lines 1080, 1088, 1096, 1137, 1142, 1280, 1292, 1354, 1360). Although inner loops default to sequential execution when unmarked (confirmed across nvfortran, Cray ftn, and AMD Flang), adding explicit seq markers maintains code clarity and consistency with established patterns in this function.

♻️ Example for order-3 left block

                                     if (wenojs) then
+                                        $:GPU_LOOP(parallelism='[seq]')
                                         do q = 0, weno_num_stencils
                                             alpha(q) = d_cbL_${XYZ}$ (q, j)/(beta(q)**2._wp)
                                         end do
                                     else if (mapped_weno) then
+                                        $:GPU_LOOP(parallelism='[seq]')
                                         do q = 0, weno_num_stencils
                                             alpha(q) = d_cbL_${XYZ}$ (q, j)/(beta(q)**2._wp)
                                         end do
                                         omega = alpha/sum(alpha)
+                                        $:GPU_LOOP(parallelism='[seq]')
                                         do q = 0, weno_num_stencils
                                             alpha(q) = (d_cbL_${XYZ}$ (q, j)*(1._wp + d_cbL_${XYZ}$ (q, &
                                                   & j) - 3._wp*omega(q)) + omega(q)**2._wp)*(omega(q)/(d_cbL_${XYZ}$ (q, &
                                                   & j)**2._wp + omega(q)*(1._wp - 2._wp*d_cbL_${XYZ}$ (q, j))))
                                         end do
                                     else if (wenoz) then
                                         ! Borges, et al. (2008)
                                         tau = abs(beta(1) - beta(0))
                                         do q = 0, weno_num_stencils
                                             alpha(q) = d_cbR_${XYZ}$ (q, j)*(1._wp + tau/beta(q))
                                         end do

Apply the same treatment to the mirrored right-side blocks (lines 996–1012, 1123–1135, 1340–1352) and order-5/7 counterparts (lines 1063–1075, 1264–1276).