From 4a2885ad3ae0396486d288df94339d0c45e6db8b Mon Sep 17 00:00:00 2001
From: Adrian Kummerlaender
Date: Sun, 10 Nov 2019 21:14:07 +0100
Subject: Implement basic CUDA target

Currently only for the SSS streaming pattern.

CudaCodePrinter in `utility/printer.py` is required to add a 'f' suffix
to all single precision floating point literals. If this is not done
(when targeting single precision) most calculations happen in double
precision which destroys performance. (In OpenCL this is not necessary
as we can simply set the `-cl-single-precision-constant` flag. Sadly
such a flag doesn't seem to exist for nvcc.)
---
 README.md                                          |  2 +-
 boltzgen/kernel/memory/precision.py                |  3 +
 .../kernel/template/bounce_back_boundary.cuda.mako | 16 +++++
 boltzgen/kernel/template/collect_moments.cuda.mako | 20 ++++++
 .../kernel/template/collide_and_stream.cuda.mako   | 17 +++++
 boltzgen/kernel/template/equilibrilize.cuda.mako   | 11 +++
 .../kernel/template/momenta_boundary.cuda.mako     | 39 +++++++++++
 boltzgen/kernel/template/pattern/SSS.cuda.mako     | 81 ++++++++++++++++++++++
 .../template/update_sss_control_structure.cpp.mako |  5 +-
 .../update_sss_control_structure.cuda.mako         | 21 ++++++
 boltzgen/utility/__init__.py                       |  1 +
 boltzgen/utility/printer.py                        | 13 ++++
 12 files changed, 224 insertions(+), 5 deletions(-)
 create mode 100644 boltzgen/kernel/template/bounce_back_boundary.cuda.mako
 create mode 100644 boltzgen/kernel/template/collect_moments.cuda.mako
 create mode 100644 boltzgen/kernel/template/collide_and_stream.cuda.mako
 create mode 100644 boltzgen/kernel/template/equilibrilize.cuda.mako
 create mode 100644 boltzgen/kernel/template/momenta_boundary.cuda.mako
 create mode 100644 boltzgen/kernel/template/pattern/SSS.cuda.mako
 create mode 100644 boltzgen/kernel/template/update_sss_control_structure.cuda.mako
 create mode 100644 boltzgen/utility/printer.py

diff --git a/README.md b/README.md
index 13f4d0c..ed2c1bd 100644
--- a/README.md
+++ b/README.md
@@ -15,7 +15,7 @@ At the moment this is a more structured and cleaned up version of the OpenCL ker
 * configurable cell indexing sequence
 * static resolution of memory offsets
 * AB, AA and SSS streaming patterns
-* C++ and OpenCL targets
+* C++, OpenCL and CUDA targets
 * simple CLI frontend
 
 ## Usage
diff --git a/boltzgen/kernel/memory/precision.py b/boltzgen/kernel/memory/precision.py
index d22e922..0c52fc0 100644
--- a/boltzgen/kernel/memory/precision.py
+++ b/boltzgen/kernel/memory/precision.py
@@ -14,3 +14,6 @@ class cpp(common):
 
 class cl(common):
     pass
+
+class cuda(common):
+    pass
diff --git a/boltzgen/kernel/template/bounce_back_boundary.cuda.mako b/boltzgen/kernel/template/bounce_back_boundary.cuda.mako
new file mode 100644
index 0000000..09499e2
--- /dev/null
+++ b/boltzgen/kernel/template/bounce_back_boundary.cuda.mako
@@ -0,0 +1,16 @@
+<%namespace name="pattern" file="${'/pattern/%s.cuda.mako' % context['streaming']}"/>
+<%
+from boltzgen.utility.printer import CudaCodePrinter
+ccode = CudaCodePrinter(float_type).doprint
+subexpr, assignment = model.collision(f_eq = model.equilibrium(resolve_moments = True))
+%>
+
+<%call expr="pattern.operator('bounce_back_boundary')">
+% for i, expr in enumerate(subexpr):
+    const ${float_type} ${expr[0]} = ${ccode(expr[1])};
+% endfor
+
+% for i, c_i in enumerate(descriptor.c):
+    const ${float_type} ${assignment[i].lhs} = ${ccode(assignment[descriptor.c.index(-c_i)].rhs)};
+% endfor
+</%call>
diff --git a/boltzgen/kernel/template/collect_moments.cuda.mako b/boltzgen/kernel/template/collect_moments.cuda.mako
new file mode 100644
index 0000000..36f03b0
--- /dev/null
+++ b/boltzgen/kernel/template/collect_moments.cuda.mako
@@ -0,0 +1,20 @@
+<%namespace name="pattern" file="${'/pattern/%s.cuda.mako' % context['streaming']}"/>
+<%
+from boltzgen.utility.printer import CudaCodePrinter
+ccode = CudaCodePrinter(float_type).doprint
+moments_subexpr, moments_assignment = model.moments()
+%>
+
+<%call expr="pattern.functor('collect_moments', [('%s*' % float_type, 'rho'), ('%s*' % float_type, 'u')])">
+% for i, expr in enumerate(moments_subexpr):
+    const ${float_type} ${expr[0]} = ${ccode(expr[1])};
+% endfor
+
+% for i, expr in enumerate(moments_assignment):
+%   if i == 0:
+    rho[gid] = ${ccode(expr.rhs)};
+%   else:
+    u[gid*${descriptor.d} + ${i-1}] = ${ccode(expr.rhs)};
+%   endif
+% endfor
+</%call>
diff --git a/boltzgen/kernel/template/collide_and_stream.cuda.mako b/boltzgen/kernel/template/collide_and_stream.cuda.mako
new file mode 100644
index 0000000..b46cf0f
--- /dev/null
+++ b/boltzgen/kernel/template/collide_and_stream.cuda.mako
@@ -0,0 +1,17 @@
+<%namespace name="pattern" file="${'/pattern/%s.cuda.mako' % context['streaming']}"/>
+<%
+from boltzgen.utility.printer import CudaCodePrinter
+ccode = CudaCodePrinter(float_type).doprint
+subexpr, assignment = model.collision(f_eq = model.equilibrium(resolve_moments = True))
+
+%>
+
+<%call expr="pattern.operator('collide_and_stream')">
+% for i, expr in enumerate(subexpr):
+    const ${float_type} ${expr[0]} = ${ccode(expr[1])};
+% endfor
+
+% for i, expr in enumerate(assignment):
+    const ${float_type} ${ccode(expr)}
+% endfor
+</%call>
diff --git a/boltzgen/kernel/template/equilibrilize.cuda.mako b/boltzgen/kernel/template/equilibrilize.cuda.mako
new file mode 100644
index 0000000..2f4cdee
--- /dev/null
+++ b/boltzgen/kernel/template/equilibrilize.cuda.mako
@@ -0,0 +1,11 @@
+<%namespace name="pattern" file="${'/pattern/%s.cuda.mako' % context['streaming']}"/>
+<%
+from boltzgen.utility.printer import CudaCodePrinter
+ccode = CudaCodePrinter(float_type).doprint
+%>
+
+<%call expr="pattern.operator('equilibrilize')">
+% for i, w_i in enumerate(descriptor.w):
+    const ${float_type} f_next_${i} = ${ccode(w_i.evalf())};
+% endfor
+</%call>
diff --git a/boltzgen/kernel/template/momenta_boundary.cuda.mako b/boltzgen/kernel/template/momenta_boundary.cuda.mako
new file mode 100644
index 0000000..408a12d
--- /dev/null
+++ b/boltzgen/kernel/template/momenta_boundary.cuda.mako
@@ -0,0 +1,39 @@
+<%namespace name="pattern" file="${'/pattern/%s.cuda.mako' % context['streaming']}"/>
+<%
+from boltzgen.utility.printer import CudaCodePrinter
+ccode = CudaCodePrinter(float_type).doprint
+moments_subexpr, moments_assignment = model.moments()
+collision_subexpr, collision_assignment = model.collision(f_eq = model.equilibrium(resolve_moments = False))
+%>
+
+<%def name="momenta_boundary(name, params)">
+<%call expr="pattern.operator('%s_momenta_boundary' % name, params)">
+% for i, expr in enumerate(moments_subexpr):
+    const ${float_type} ${expr[0]} = ${ccode(expr[1])};
+% endfor
+
+    ${caller.body()}
+
+% for i, expr in enumerate(collision_subexpr):
+    const ${float_type} ${expr[0]} = ${ccode(expr[1])};
+% endfor
+
+% for i, expr in enumerate(collision_assignment):
+    const ${float_type} ${ccode(expr)}
+% endfor
+</%call>
+</%def>
+
+<%call expr="momenta_boundary('velocity', list(map(lambda i: (float_type, 'velocity_%d' % i), range(descriptor.d))))">
+    const ${float_type} ${ccode(moments_assignment[0])}
+% for i, expr in enumerate(moments_assignment[1:]):
+    const ${float_type} ${expr.lhs} = velocity_${i};
+% endfor
+</%call>
+
+<%call expr="momenta_boundary('density', [(float_type, 'density')])">
+    const ${float_type} ${moments_assignment[0].lhs} = density;
+% for i, expr in enumerate(moments_assignment[1:]):
+    const ${float_type} ${ccode(expr)}
+% endfor
+</%call>
diff --git a/boltzgen/kernel/template/pattern/SSS.cuda.mako b/boltzgen/kernel/template/pattern/SSS.cuda.mako
new file mode 100644
index 0000000..8f571c7
--- /dev/null
+++ b/boltzgen/kernel/template/pattern/SSS.cuda.mako
@@ -0,0 +1,81 @@
+<%def name="operator(name, params = None)">
+<%
+if layout.__class__.__name__ != 'SOA':
+    raise Exception('SSS pattern only works for the AOS memory layout')
+%>
+__global__ void ${name}(
+      ${float_type}** f
+% if 'cell_list_dispatch' in extras:
+    , std::size_t* cells
+    , std::size_t  cell_count
+% else:
+    , std::size_t gid
+% endif
+% if params is not None:
+% for param_type, param_name in params:
+    , ${param_type} ${param_name}
+% endfor
+% endif
+) {
+% if 'cell_list_dispatch' in extras:
+    const std::size_t index = blockIdx.x * blockDim.x + threadIdx.x;
+    if (!(index < cell_count)) {
+        return;
+    }
+    const std::size_t gid = cells[index];
+% endif
+
+% for i, c_i in enumerate(descriptor.c):
+    ${float_type}* preshifted_f_${i} = f[${i}] + ${layout.cell_preshift('gid')};
+% endfor
+
+% for i, c_i in enumerate(descriptor.c):
+    const ${float_type} f_curr_${i} = *preshifted_f_${i};
+% endfor
+
+    ${caller.body()}
+
+% for i, c_i in enumerate(descriptor.c):
+    *preshifted_f_${i} = f_next_${descriptor.c.index(-c_i)};
+% endfor
+}
+</%def>
+
+<%def name="functor(name, params = None)">
+<%
+if layout.__class__.__name__ != 'SOA':
+    raise Exception('SSS pattern only works for the AOS memory layout')
+%>
+__global__ void ${name}(
+      ${float_type}** f
+% if 'cell_list_dispatch' in extras:
+    , std::size_t* cells
+    , std::size_t  cell_count
+% else:
+    , std::size_t gid
+% endif
+% if params is not None:
+% for param_type, param_name in params:
+    , ${param_type} ${param_name}
+% endfor
+% endif
+) {
+% if 'cell_list_dispatch' in extras:
+    const std::size_t index = blockIdx.x * blockDim.x + threadIdx.x;
+    if (!(index < cell_count)) {
+        return;
+    }
+    const std::size_t gid = cells[index];
+% endif
+
+% for i, c_i in enumerate(descriptor.c):
+    const ${float_type}* preshifted_f_${i} = f[${i}] + ${layout.cell_preshift('gid')};
+% endfor
+
+% for i, c_i in enumerate(descriptor.c):
+    const ${float_type} f_curr_${i} = *preshifted_f_${descriptor.c.index(-c_i)};
+% endfor
+
+    ${caller.body()}
+}
+</%def>
diff --git a/boltzgen/kernel/template/update_sss_control_structure.cpp.mako b/boltzgen/kernel/template/update_sss_control_structure.cpp.mako
index 9197022..b7661a3 100644
--- a/boltzgen/kernel/template/update_sss_control_structure.cpp.mako
+++ b/boltzgen/kernel/template/update_sss_control_structure.cpp.mako
@@ -7,9 +7,6 @@ void update_sss_control_structure(${float_type}** f) {
     ${float_type}* f_old_${i} = f[${i}];
 % endfor
 % for i, c_i in enumerate(descriptor.c):
-    f[${i}]  = f_old_${descriptor.c.index(-c_i)};
-% endfor
-% for i, c_i in enumerate(descriptor.c):
-    f[${i}] += ${layout.neighbor_offset(-c_i)};
+    f[${i}]  = f_old_${descriptor.c.index(-c_i)} + ${layout.neighbor_offset(-c_i)};
 % endfor
 }
diff --git a/boltzgen/kernel/template/update_sss_control_structure.cuda.mako b/boltzgen/kernel/template/update_sss_control_structure.cuda.mako
new file mode 100644
index 0000000..2a98b94
--- /dev/null
+++ b/boltzgen/kernel/template/update_sss_control_structure.cuda.mako
@@ -0,0 +1,21 @@
+<%
+if streaming != 'SSS':
+    raise Exception('"update_sss_control_structure" function only makes sense for the SSS pattern')
+
+padding = (max(geometry.size_x,geometry.size_y,geometry.size_z)+1)**(descriptor.d-1)
+%>
+
+__global__ void init_sss_control_structure(${float_type}* f, ${float_type}** control) {
+% for i, c_i in enumerate(descriptor.c):
+    control[${i}]  = f + ${padding + layout.pop_offset(i, 2*padding)};
+% endfor
+}
+
+__global__ void update_sss_control_structure(${float_type}** f) {
+% for i, c_i in enumerate(descriptor.c):
+    ${float_type}* f_old_${i} = f[${i}];
+% endfor
+% for i, c_i in enumerate(descriptor.c):
+    f[${i}]  = f_old_${descriptor.c.index(-c_i)} + ${layout.neighbor_offset(-c_i)};
+% endfor
+}
diff --git a/boltzgen/utility/__init__.py b/boltzgen/utility/__init__.py
index fa9c760..5905c36 100644
--- a/boltzgen/utility/__init__.py
+++ b/boltzgen/utility/__init__.py
@@ -1,5 +1,6 @@
 from . import optimizations
 from . import ndindex
+from . import printer
 
 from sympy.codegen.ast import Assignment
 
diff --git a/boltzgen/utility/printer.py b/boltzgen/utility/printer.py
new file mode 100644
index 0000000..b52a5e0
--- /dev/null
+++ b/boltzgen/utility/printer.py
@@ -0,0 +1,13 @@
+from sympy.printing.ccode import C99CodePrinter
+from sympy.codegen.ast import float32, float64
+
+class CudaCodePrinter(C99CodePrinter):
+    pass
+
+    def __init__(self, float_type, **args):
+        super(CudaCodePrinter, self).__init__(**args)
+        if float_type == 'float':
+            self.type_func_suffixes[float32]    = 'f'
+            self.type_func_suffixes[float64]    = 'f'
+            self.type_literal_suffixes[float32] = 'f'
+            self.type_literal_suffixes[float64] = 'f'
-- 
cgit v1.2.3