Add support for SSS propagation in OpenCL example

3 files changed, 96 insertions, 4 deletions

diff --git a/boltzgen.nix b/boltzgen.nix
index 74ba512..dcbc7e0 100644
--- a/boltzgen.nix
+++ b/boltzgen.nix
@@ -7,7 +7,7 @@ let in {
 
     src = builtins.fetchGit {
       url = "https://code.kummerlaender.eu/boltzgen/";
-      rev = "aa509dd4ebbb9d1d8ad6ebfe05111228fd9ae7c0";
+      rev = "b5a24f31871d900342a3c47398cc75e22bad0b6f";
     };
 
     propagatedBuildInputs = with pkgs.python37Packages; [
diff --git a/lid_driven_cavity/opencl/SSS.py b/lid_driven_cavity/opencl/SSS.py
new file mode 100644
index 0000000..86b42a9
--- /dev/null
+++ b/lid_driven_cavity/opencl/SSS.py
@@ -0,0 +1,85 @@
+import pyopencl as cl
+mf = cl.mem_flags
+
+import numpy
+
+class Memory:
+    def __init__(self, descriptor, geometry, context, float_type):
+        self.context    = context
+        self.float_type = float_type
+
+        self.size_x = geometry.size_x
+        self.size_y = geometry.size_y
+        self.size_z = geometry.size_z
+        self.volume = self.size_x * self.size_y * self.size_z
+
+        padding = (max(geometry.size_x,geometry.size_y,geometry.size_z)+1)**(descriptor.d-1)
+        self.pop_size = (geometry.volume+2*padding) * descriptor.q * self.float_type(0).nbytes
+
+        self.moments_size = 3 * self.volume * self.float_type(0).nbytes
+
+        self.cl_pop = cl.Buffer(self.context, mf.READ_WRITE, size=self.pop_size)
+        self.cl_control = cl.Buffer(self.context, mf.READ_WRITE, size=descriptor.q * 8)
+
+        self.cl_moments = cl.Buffer(self.context, mf.WRITE_ONLY, size=self.moments_size)
+
+    def gid(self, x, y, z = 0):
+        return z * (self.size_x*self.size_y) + y * self.size_x + x;
+
+class Lattice:
+    def __init__(self, geometry, kernel_src, descriptor, context, queue, precision = 'single'):
+        self.geometry = geometry
+        self.descriptor = descriptor
+
+        self.float_type = {
+            'single': (numpy.float32, 'float'),
+            'double': (numpy.float64, 'double'),
+        }.get(precision, None)
+
+        self.layout = None
+
+        self.context = context
+        self.queue   = queue
+
+        self.memory = Memory(descriptor, self.geometry, self.context, self.float_type[0])
+
+        self.compiler_args = {
+            'single': '-cl-single-precision-constant -cl-fast-relaxed-math',
+            'double': '-cl-fast-relaxed-math'
+        }.get(precision, None)
+
+        self.build_kernel(kernel_src)
+
+        self.program.init_sss_control_structure(
+            self.queue, (1,1,1), None, self.memory.cl_pop, self.memory.cl_control).wait()
+
+        self.program.equilibrilize_domain(
+            self.queue, self.geometry.size(), self.layout, self.memory.cl_control).wait()
+
+        self.tasks = []
+
+    def build_kernel(self, src):
+        self.program = cl.Program(self.context, src).build(self.compiler_args)
+
+    def schedule(self, f, cells, *params):
+        self.tasks += [ (eval("self.program.%s" % f), cells, params) ]
+
+    def evolve(self):
+        for f, cells, params in self.tasks:
+            f(self.queue, cells.size(), self.layout, self.memory.cl_control, cells.get(), *params)
+
+        self.program.update_sss_control_structure(
+            self.queue, (1,1,1), None, self.memory.cl_control)
+
+    def sync(self):
+        self.queue.finish()
+
+    def get_moments(self):
+        moments = numpy.ndarray(shape=(self.memory.volume*(self.descriptor.d+1),1), dtype=self.float_type[0])
+
+        self.program.collect_moments_domain(
+            self.queue, self.geometry.inner_size(), self.layout, self.memory.cl_control, self.memory.cl_moments)
+
+        cl.enqueue_copy(self.queue, moments, self.memory.cl_moments).wait();
+
+        return moments
diff --git a/lid_driven_cavity/opencl/ldc_2d.py b/lid_driven_cavity/opencl/ldc_2d.py
index 1ea80aa..8f9a0c3 100644
--- a/lid_driven_cavity/opencl/ldc_2d.py
+++ b/lid_driven_cavity/opencl/ldc_2d.py
@@ -20,10 +20,14 @@ functions = ['collide_and_stream', 'equilibrilize', 'collect_moments', 'momenta_
 extras    = ['cell_list_dispatch']
 
 precision = 'single'
-streaming = 'AA'
+streaming = 'SSS'
+
+if streaming == 'SSS':
+    functions = functions + ['update_sss_control_structure']
 
-import AA
 import AB
+import AA
+import SSS
 
 Lattice = eval('%s.Lattice' % streaming)
 
@@ -60,7 +64,10 @@ wall_cells = CellList(lattice.context, lattice.queue, lattice.float_type,
 lid_cells = CellList(lattice.context, lattice.queue, lattice.float_type,
     [ gid(x,y) for x, y in geometry.inner_cells() if y == geometry.size_y-2 ])
 
-if streaming == 'AB':
+if streaming == 'SSS':
+    lattice.schedule('equilibrilize', ghost_cells)
+
+if streaming in ['AB', 'SSS']:
     lattice.schedule('collide_and_stream', bulk_cells)
     lattice.schedule('velocity_momenta_boundary', wall_cells, numpy.array([0.0, 0.0], dtype=lattice.float_type[0]))
     lattice.schedule('velocity_momenta_boundary', lid_cells,  numpy.array([0.1, 0.0], dtype=lattice.float_type[0]))