Add test to compare results from OpenCL, CPU, CUDA in a full LBM simulation

For some reason their result does not agree with the calculated reference velocity. Did not check for the analytical solution.

#9 (closed)

lbmpy_tests/test_poisuille_channel.py

+"""
+This test revealed a problem with OpenCL (https://i10git.cs.fau.de/pycodegen/lbmpy/issues/9#note_9521)
+"""
+
+
+import numpy as np
+import pytest
+
+import lbmpy
+from lbmpy.boundaries import NoSlip
+from lbmpy.boundaries.boundaryhandling import LatticeBoltzmannBoundaryHandling
+from lbmpy.macroscopic_value_kernels import macroscopic_values_setter
+from lbmpy.session import *
+from pystencils.session import *
+
+previous_vmids = {}
+previous_endresults = {}
+
+
+@pytest.mark.parametrize('target', ('cpu', 'gpu', 'opencl'))
+def test_poiseuille_channel(target):
+    # # Lattice Boltzmann
+    #
+    # ## Definitions
+    if target == 'opencl':
+        import pytest
+        pytest.importorskip("pyopencl")
+        import pystencils.opencl.autoinit
+    elif target == 'gpu':
+        import pytest
+        pytest.importorskip("pycuda")
+
+    # In[2]:
+
+    L = (34, 34)
+
+    lb_stencil = get_stencil("D2Q9")
+    eta = 1
+    omega = lbmpy.relaxationrates.relaxation_rate_from_lattice_viscosity(eta)
+
+    f_pre = 0.00001
+
+    # ## Data structures
+
+    # In[3]:
+
+    dh = ps.create_data_handling(L, periodicity=(True, False), default_target=target)
+
+    opts = {'cpu_openmp': True,
+            'cpu_vectorize_info': None,
+            'target': dh.default_target}
+
+    src = dh.add_array('src', values_per_cell=len(lb_stencil))
+    dst = dh.add_array_like('dst', 'src')
+    ρ = dh.add_array('rho', latex_name='\\rho')
+    u = dh.add_array('u', values_per_cell=dh.dim)
+
+    # In[4]:
+
+    collision = create_lb_update_rule(stencil=lb_stencil,
+                                      relaxation_rate=omega,
+                                      compressible=True,
+                                      force_model='guo',
+                                      force=sp.Matrix([f_pre, 0]),
+                                      kernel_type='collide_only',
+                                      optimization={'symbolic_field': src})
+
+    stream = create_stream_pull_with_output_kernel(collision.method, src, dst, {'velocity': u})
+
+    lbbh = LatticeBoltzmannBoundaryHandling(collision.method, dh, src.name, target=dh.default_target)
+
+    stream_kernel = ps.create_kernel(stream, **opts).compile()
+    collision_kernel = ps.create_kernel(collision, **opts).compile()
+
+    # ## Set up the simulation
+
+    # In[5]:
+
+    init = macroscopic_values_setter(collision.method, velocity=(0,)*dh.dim,
+                                     pdfs=src.center_vector, density=ρ.center)
+    init_kernel = ps.create_kernel(init, ghost_layers=0).compile()
+
+    noslip = NoSlip()
+    lbbh.set_boundary(noslip, make_slice[:, :4])
+    lbbh.set_boundary(noslip, make_slice[:, -4:])
+
+    for bh in lbbh, :
+        assert len(bh._boundary_object_to_boundary_info) == 1, "Restart kernel to clear boundaries"
+
+    def init():
+        dh.fill(ρ.name, 1)
+        dh.fill(u.name, np.nan, ghost_layers=True, inner_ghost_layers=True)
+        dh.fill(u.name, 0)
+
+        dh.run_kernel(init_kernel)
+
+    # In[6]:
+
+    sync_pdfs = dh.synchronization_function([src.name])
+
+    def time_loop(steps):
+        dh.all_to_gpu()
+        vmid = np.empty((2, steps//10+1))
+        i = -1
+        for i in range(steps):
+            dh.run_kernel(collision_kernel)
+            sync_pdfs()
+            lbbh()
+            dh.run_kernel(stream_kernel)
+
+            dh.swap(src.name, dst.name)
+
+            if i % 10 == 0:
+                if u.name in dh.gpu_arrays:
+                    dh.to_cpu(u.name)
+                uu = dh.gather_array(u.name)
+                uu = uu[L[0]//2-1:L[0]//2+1, L[1]//2-1:L[1]//2+1, 0].mean()
+                vmid[:, i//10] = [i, uu]
+                if 1/np.sqrt(3) < uu:
+                    break
+        if dh.is_on_gpu(u.name):
+            dh.to_cpu(u.name)
+
+        return vmid[:, :i//10+1]
+
+    # In[7]:
+
+    # def plot():
+
+        # plt.subplot(2, 2, 1)
+        # plt.title("$u$")
+        # plt.xlabel("$x$")
+        # plt.ylabel("$y$")
+        # plt.vector_field_magnitude(uu)
+        # plt.colorbar()
+
+        # plt.subplot(2, 2, 2)
+        # plt.title("$u$")
+        # plt.xlabel("$x/2$")
+        # plt.ylabel("$y/2$")
+        # plt.vector_field(uu, step=2)
+
+        # actualwidth = np.sum(1-np.isnan(uu[0, :, 0]))
+        # uref = f_pre*actualwidth**2/(8*(eta))
+
+        # plt.subplot(2, 2, 3)
+        # plt.title("flow profile")
+        # plt.xlabel("$y$")
+        # plt.ylabel(r"$u_x$")
+        # plt.plot((uu[L[0]//2-1, :, 0]+uu[L[0]//2, :, 0])/2)
+
+        # plt.subplot(2, 2, 4)
+        # plt.title("convergence")
+        # plt.xlabel("$t$")
+        # plt.plot()
+
+    # ## Run the simulation
+
+    # In[8]:
+
+    init()
+    vmid = time_loop(1000)
+    # plot()
+
+    uu = dh.gather_array(u.name)
+
+    for target, prev_endresult in previous_endresults.items():
+        assert np.allclose(uu[4:-4, 4:-4, 0], prev_endresult[4:-4, 4:-4, 0],
+                           atol=1e-5), f'uu does not agree with result from {target}'
+
+    for target, prev_vmid in previous_vmids.items():
+        assert np.allclose(vmid, prev_vmid, atol=1e-5), f'vmid does not agree with result from {target}'
+
+    # uref = f_pre*actualwidth**2/(8*(eta))
+    # actualwidth = np.sum(1-np.isnan(uu[0, :, 0]))
+    # assert np.allclose(vmid[1, -1]/uref, 1, atol=1e-3)
+    # print(vmid[1, -1])
+
+    previous_vmids[target] = vmid
+    previous_endresults[target] = uu