From aa509dd4ebbb9d1d8ad6ebfe05111228fd9ae7c0 Mon Sep 17 00:00:00 2001
From: Adrian Kummerlaender
Date: Tue, 12 Nov 2019 22:54:11 +0100
Subject: Match OpenCL and CUDA cell list dispatch templates

---
 boltzgen/kernel/template/collect_moments.cl.mako | 12 ++++++++++++
 1 file changed, 12 insertions(+)

(limited to 'boltzgen/kernel/template/collect_moments.cl.mako')

diff --git a/boltzgen/kernel/template/collect_moments.cl.mako b/boltzgen/kernel/template/collect_moments.cl.mako
index 63387b7..e3ce166 100644
--- a/boltzgen/kernel/template/collect_moments.cl.mako
+++ b/boltzgen/kernel/template/collect_moments.cl.mako
@@ -16,6 +16,18 @@ moments_subexpr, moments_assignment = model.moments()
 % endfor
 </%call>
 
+<%call expr="pattern.functor_with_domain_dispatch('collect_moments_domain', [('__global %s*' % float_type, 'm')])">
+% for i, expr in enumerate(moments_subexpr):
+    const ${float_type} ${expr[0]} = ${sympy.ccode(expr[1])};
+% endfor
+
+    __global ${float_type}* preshifted_m = m + gid*${descriptor.d+1};
+
+% for i, expr in enumerate(moments_assignment):
+    preshifted_m[${i}] = ${sympy.ccode(expr.rhs)};
+% endfor
+</%call>
+
 % if 'opencl_gl_interop' in extras:
 <%call expr="pattern.functor_with_domain_dispatch('collect_moments_to_texture', [('__write_only %s' % {2: 'image2d_t', 3: 'image3d_t'}.get(descriptor.d), 'm')])">
 % for i, expr in enumerate(moments_subexpr):
-- 
cgit v1.2.3