Pull in C++ template from symlbm_playground's standalone branch

3 files changed, 68 insertions, 2 deletions

diff --git a/boltzgen/kernel/generator.py b/boltzgen/kernel/generator.py
index bd1bb86..59654e2 100644
--- a/boltzgen/kernel/generator.py
+++ b/boltzgen/kernel/generator.py
@@ -3,8 +3,8 @@ import sympy
 from mako.template import Template
 from pathlib import Path
 
-def source(descriptor, moments, collide, boundary_src, float_type, geometry):
-    return Template(filename = str(Path(__file__).parent/'template/kernel.mako')).render(
+def source(target, descriptor, moments, collide, boundary_src, float_type, geometry):
+    return Template(filename = str(Path(__file__).parent/('template/basic.' + target + '.mako'))).render(
         descriptor = descriptor,
         geometry   = geometry,
 
diff --git a/boltzgen/kernel/template/basic.cpp.mako b/boltzgen/kernel/template/basic.cpp.mako
new file mode 100644
index 0000000..1879c35
--- /dev/null
+++ b/boltzgen/kernel/template/basic.cpp.mako
@@ -0,0 +1,66 @@
+<%
+def pop_offset(i):
+    return i * geometry.volume
+%>
+
+void equilibrilize(${float_type}*    f_next,
+                   ${float_type}*    f_prev,
+                   const std::size_t gid)
+{
+    ${float_type}* preshifted_f_next = f_next + gid;
+    ${float_type}* preshifted_f_prev = f_prev + gid;
+
+% for i, w_i in enumerate(descriptor.w):
+    preshifted_f_next[${pop_offset(i)}] = ${w_i.evalf()};
+    preshifted_f_prev[${pop_offset(i)}] = ${w_i.evalf()};
+% endfor
+}
+
+<%
+def neighbor_offset(c_i):
+    return {
+        2: lambda:                                          c_i[1]*geometry.size_x + c_i[0],
+        3: lambda: c_i[2]*geometry.size_x*geometry.size_y + c_i[1]*geometry.size_x + c_i[0]
+    }.get(descriptor.d)()
+
+def padding():
+    return {
+        2: lambda:                                     1*geometry.size_x + 1,
+        3: lambda: 1*geometry.size_x*geometry.size_y + 1*geometry.size_x + 1
+    }.get(descriptor.d)()
+%>
+
+void collide_and_stream(      ${float_type}* f_next,
+                        const ${float_type}* f_prev,
+                        const int*           material,
+                        const std::size_t    gid)
+{
+    const int m = material[gid];
+
+          ${float_type}* preshifted_f_next = f_next + gid;
+    const ${float_type}* preshifted_f_prev = f_prev + gid;
+
+% for i, c_i in enumerate(descriptor.c):
+    const ${float_type} f_curr_${i} = preshifted_f_prev[${pop_offset(i) + neighbor_offset(-c_i)}];
+% endfor
+
+% for i, expr in enumerate(moments_subexpr):
+    const ${float_type} ${expr[0]} = ${ccode(expr[1])};
+% endfor
+
+% for i, expr in enumerate(moments_assignment):
+    ${float_type} ${ccode(expr)}
+% endfor
+
+% for i, expr in enumerate(collide_subexpr):
+    const ${float_type} ${expr[0]} = ${ccode(expr[1])};
+% endfor
+
+% for i, expr in enumerate(collide_assignment):
+    const ${float_type} ${ccode(expr)}
+% endfor
+
+% for i, expr in enumerate(collide_assignment):
+    preshifted_f_next[${pop_offset(i)}] = m*f_next_${i} + (1.0-m)*${descriptor.w[i].evalf()};
+% endfor
+}
diff --git a/boltzgen/kernel/template/kernel.mako b/boltzgen/kernel/template/basic.opencl.mako
index 5ddf64c..5ddf64c 100644
--- a/boltzgen/kernel/template/kernel.mako
+++ b/boltzgen/kernel/template/basic.opencl.mako