shallow water: add partial gpu support, warm up jit cache

Author: tkarna

examples/shallow_water.py

@@ -111,28 +111,32 @@ def run(n, backend, datatype, benchmark_mode):
     t_end = 1.0
 
     # coordinate arrays
+    sync()
     x_t_2d = fromfunction(
-        lambda i, j: xmin + i * dx + dx / 2,
-        (nx, ny),
-        dtype=dtype,
+        lambda i, j: xmin + i * dx + dx / 2, (nx, ny), dtype=dtype, device=""
     )
     y_t_2d = fromfunction(
-        lambda i, j: ymin + j * dy + dy / 2,
-        (nx, ny),
-        dtype=dtype,
+        lambda i, j: ymin + j * dy + dy / 2, (nx, ny), dtype=dtype, device=""
+    )
+    x_u_2d = fromfunction(
+        lambda i, j: xmin + i * dx, (nx + 1, ny), dtype=dtype, device=""
     )
-    x_u_2d = fromfunction(lambda i, j: xmin + i * dx, (nx + 1, ny), dtype=dtype)
     y_u_2d = fromfunction(
         lambda i, j: ymin + j * dy + dy / 2,
         (nx + 1, ny),
         dtype=dtype,
+        device="",
     )
     x_v_2d = fromfunction(
         lambda i, j: xmin + i * dx + dx / 2,
         (nx, ny + 1),
         dtype=dtype,
+        device="",
     )
-    y_v_2d = fromfunction(lambda i, j: ymin + j * dy, (nx, ny + 1), dtype=dtype)
+    y_v_2d = fromfunction(
+        lambda i, j: ymin + j * dy, (nx, ny + 1), dtype=dtype, device=""
+    )
+    sync()
 
     T_shape = (nx, ny)
     U_shape = (nx + 1, ny)
@@ -157,7 +161,7 @@ def run(n, backend, datatype, benchmark_mode):
     q = create_full(F_shape, 0.0, dtype)
 
     # bathymetry
-    h = create_full(T_shape, 0.0, dtype)
+    h = create_full(T_shape, 1.0, dtype)  # HACK init with 1
 
     hu = create_full(U_shape, 0.0, dtype)
     hv = create_full(V_shape, 0.0, dtype)
@@ -205,22 +209,16 @@ def bathymetry(x_t_2d, y_t_2d, lx, ly):
         bath = 1.0
         return bath * create_full(T_shape, 1.0, dtype)
 
-    # inital elevation
-    u0, v0, e0 = exact_solution(
-        0, x_t_2d, y_t_2d, x_u_2d, y_u_2d, x_v_2d, y_v_2d
-    )
-    e[:, :] = e0
-    u[:, :] = u0
-    v[:, :] = v0
-
     # set bathymetry
-    h[:, :] = bathymetry(x_t_2d, y_t_2d, lx, ly)
+    # h[:, :] = bathymetry(x_t_2d, y_t_2d, lx, ly).to_device(device)
     # steady state potential energy
-    pe_offset = 0.5 * g * float(np.sum(h**2.0, all_axes)) / nx / ny
+    # pe_offset = 0.5 * g * float(np.sum(h**2.0, all_axes)) / nx / ny
+    pe_offset = 0.5 * g * float(1.0) / nx / ny
 
     # compute time step
     alpha = 0.5
-    h_max = float(np.max(h, all_axes))
+    # h_max = float(np.max(h, all_axes))
+    h_max = float(1.0)
     c = (g * h_max) ** 0.5
     dt = alpha * dx / c
     dt = t_export / int(math.ceil(t_export / dt))
@@ -329,6 +327,18 @@ def step(u, v, e, u1, v1, e1, u2, v2, e2):
         v[:, 1:-1] = v[:, 1:-1] / 3.0 + 2.0 / 3.0 * (v2[:, 1:-1] + dt * dvdt)
         e[:, :] = e[:, :] / 3.0 + 2.0 / 3.0 * (e2[:, :] + dt * dedt)
 
+    # warm jit cache
+    step(u, v, e, u1, v1, e1, u2, v2, e2)
+    sync()
+
+    # initial solution
+    u0, v0, e0 = exact_solution(
+        0, x_t_2d, y_t_2d, x_u_2d, y_u_2d, x_v_2d, y_v_2d
+    )
+    e[:, :] = e0.to_device(device)
+    u[:, :] = u0.to_device(device)
+    v[:, :] = v0.to_device(device)
+
     t = 0
     i_export = 0
     next_t_export = 0
@@ -341,30 +351,41 @@ def step(u, v, e, u1, v1, e1, u2, v2, e2):
         t = i * dt
 
         if t >= next_t_export - 1e-8:
-            _elev_max = np.max(e, all_axes)
-            _u_max = np.max(u, all_axes)
-            _q_max = np.max(q, all_axes)
-            _total_v = np.sum(e + h, all_axes)
-
-            # potential energy
-            _pe = 0.5 * g * (e + h) * (e - h) + pe_offset
-            _total_pe = np.sum(_pe, all_axes)
-
-            # kinetic energy
-            u2 = u * u
-            v2 = v * v
-            u2_at_t = 0.5 * (u2[1:, :] + u2[:-1, :])
-            v2_at_t = 0.5 * (v2[:, 1:] + v2[:, :-1])
-            _ke = 0.5 * (u2_at_t + v2_at_t) * (e + h)
-            _total_ke = np.sum(_ke, all_axes)
-
-            total_pe = float(_total_pe) * dx * dy
-            total_ke = float(_total_ke) * dx * dy
-            total_e = total_ke + total_pe
-            elev_max = float(_elev_max)
-            u_max = float(_u_max)
-            q_max = float(_q_max)
-            total_v = float(_total_v) * dx * dy
+            if device:
+                # FIXME gpu.memcpy to host requires identity layout
+                # FIXME reduction on gpu
+                elev_max = 0
+                u_max = 0
+                q_max = 0
+                diff_e = 0
+                diff_v = 0
+                total_pe = 0
+                total_ke = 0
+            else:
+                _elev_max = np.max(e, all_axes)
+                _u_max = np.max(u, all_axes)
+                _q_max = np.max(q, all_axes)
+                _total_v = np.sum(e + h, all_axes)
+
+                # potential energy
+                _pe = 0.5 * g * (e + h) * (e - h) + pe_offset
+                _total_pe = np.sum(_pe, all_axes)
+
+                # kinetic energy
+                u2 = u * u
+                v2 = v * v
+                u2_at_t = 0.5 * (u2[1:, :] + u2[:-1, :])
+                v2_at_t = 0.5 * (v2[:, 1:] + v2[:, :-1])
+                _ke = 0.5 * (u2_at_t + v2_at_t) * (e + h)
+                _total_ke = np.sum(_ke, all_axes)
+
+                total_pe = float(_total_pe) * dx * dy
+                total_ke = float(_total_ke) * dx * dy
+                total_e = total_ke + total_pe
+                elev_max = float(_elev_max)
+                u_max = float(_u_max)
+                q_max = float(_q_max)
+                total_v = float(_total_v) * dx * dy
 
             if i_export == 0:
                 initial_v = total_v
@@ -399,35 +420,40 @@ def step(u, v, e, u1, v1, e1, u2, v2, e2):
     duration = time_mod.perf_counter() - tic
     info(f"Duration: {duration:.2f} s")
 
-    e_exact = exact_solution(t, x_t_2d, y_t_2d, x_u_2d, y_u_2d, x_v_2d, y_v_2d)[
-        2
-    ]
-    err2 = (e_exact - e) * (e_exact - e) * dx * dy / lx / ly
-    err_L2 = math.sqrt(float(np.sum(err2, all_axes)))
-    info(f"L2 error: {err_L2:7.15e}")
-
-    if nx < 128 or ny < 128:
-        info("Skipping correctness test due to small problem size.")
-    elif not benchmark_mode:
-        tolerance_ene = 1e-7 if datatype == "f32" else 1e-9
-        assert (
-            diff_e < tolerance_ene
-        ), f"Energy error exceeds tolerance: {diff_e} > {tolerance_ene}"
-        if nx == 128 and ny == 128:
-            if datatype == "f32":
-                assert numpy.allclose(
-                    err_L2, 4.3127859e-05, rtol=1e-5
-                ), "L2 error does not match"
-            else:
-                assert numpy.allclose(
-                    err_L2, 4.315799035627906e-05
-                ), "L2 error does not match"
-        else:
-            tolerance_l2 = 1e-4
+    if device:
+        # FIXME gpu.memcpy to host requires identity layout
+        # FIXME reduction on gpu
+        pass
+    else:
+        e_exact = exact_solution(
+            t, x_t_2d, y_t_2d, x_u_2d, y_u_2d, x_v_2d, y_v_2d
+        )[2]
+        err2 = (e_exact - e) * (e_exact - e) * dx * dy / lx / ly
+        err_L2 = math.sqrt(float(np.sum(err2, all_axes)))
+        info(f"L2 error: {err_L2:7.15e}")
+
+        if nx < 128 or ny < 128:
+            info("Skipping correctness test due to small problem size.")
+        elif not benchmark_mode:
+            tolerance_ene = 1e-7 if datatype == "f32" else 1e-9
             assert (
-                err_L2 < tolerance_l2
-            ), f"L2 error exceeds tolerance: {err_L2} > {tolerance_l2}"
-        info("SUCCESS")
+                diff_e < tolerance_ene
+            ), f"Energy error exceeds tolerance: {diff_e} > {tolerance_ene}"
+            if nx == 128 and ny == 128:
+                if datatype == "f32":
+                    assert numpy.allclose(
+                        err_L2, 4.3127859e-05, rtol=1e-5
+                    ), "L2 error does not match"
+                else:
+                    assert numpy.allclose(
+                        err_L2, 4.315799035627906e-05
+                    ), "L2 error does not match"
+            else:
+                tolerance_l2 = 1e-4
+                assert (
+                    err_L2 < tolerance_l2
+                ), f"L2 error exceeds tolerance: {err_L2} > {tolerance_l2}"
+            info("SUCCESS")
 
     fini()