Skip to content

Commit

Permalink
Runtime-configurable symbolic kernel
Browse files Browse the repository at this point in the history 
See #205
  • Loading branch information
@ddemidov
ddemidov committed May 7, 2016
1 parent cb6f9eb commit e2fad48
Show file tree
Hide file tree
Showing 4 changed files with 99 additions and 110 deletions.
5 changes: 5 additions & 0 deletions vexcl/backend/compute/kernel.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -206,6 +206,11 @@ class kernel {
/// Get reference to the underlying object.
const boost::compute::kernel& get() const { return K; }
boost::compute::kernel& get() { return K; }

/// Reset argument counter.
void reset() {
argpos = 0;
}
private:
unsigned argpos;

Expand Down
9 changes: 7 additions & 2 deletions vexcl/backend/cuda/kernel.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -137,8 +137,7 @@ class kernel {
)
);

stack.clear();
prm_pos.clear();
reset();
}

#ifndef BOOST_NO_VARIADIC_TEMPLATES
Expand Down Expand Up @@ -206,6 +205,12 @@ class kernel {

/// Get the underlying CUfunction
CUfunction get() const { return K; }

/// Reset argument counter.
void reset() {
stack.clear();
prm_pos.clear();
}
private:
context ctx;
program P;
Expand Down
5 changes: 5 additions & 0 deletions vexcl/backend/opencl/kernel.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -231,6 +231,11 @@ class kernel {
/// Get reference to the underlying object.
const cl::Kernel& get() const { return K; }
cl::Kernel& get() { return K; }

/// Reset argument counter.
void reset() {
argpos = 0;
}
private:
unsigned argpos;

Expand Down
190 changes: 82 additions & 108 deletions vexcl/generator.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -502,38 +502,34 @@ std::ostream& operator<<(std::ostream &os, const symbolic<T> &sym) {
namespace generator {

/// Autogenerated kernel.
template <size_t NP>
class Kernel {
class kernel {
public:
template <class ArgTuple>
Kernel(
kernel(
const std::vector<backend::command_queue> &queue,
const std::string &name, const std::string &body,
const ArgTuple& args
) : queue(queue)
{
static_assert(
boost::fusion::result_of::size<ArgTuple>::value == NP,
"Wrong number of kernel parameters"
);
const std::string &name
) : queue(queue), name(name), psize(queue.size(), 0)
{}

template <class SymVar>
void add_param(const SymVar &var) {
prm_decl << "\t" << var.prmdecl() << ",\n";
prm_read << var.init();
prm_save << var.write();
}

void build(const std::string &body) {
for(auto q = queue.begin(); q != queue.end(); q++) {
backend::source_generator source(*q);

source << get_preamble().str();

source.kernel(name).open("(")
.parameter<size_t>("n");
source.kernel(name).open("(");

boost::fusion::for_each(args, declare_params(source));
source << prm_decl.str() << "\t" << type_name<size_t>() << " n";

source.close(")").open("{").grid_stride_loop().open("{");

boost::fusion::for_each(args, read_params(source));

source << body;

boost::fusion::for_each(args, write_params(source));
source.new_line() << prm_read.str() << body << prm_save.str();

source.close("}").close("}");

Expand All @@ -545,117 +541,85 @@ class Kernel {
}
}

template <class T>
void push_arg(const T &v) {
for(unsigned d = 0; d < queue.size(); d++) {
cache.find(backend::get_context_id(queue[d]))->second.push_arg(v);
}
}

template <class T>
void push_arg(const vector<T> &v) {
for(unsigned d = 0; d < queue.size(); d++) {
cache.find(backend::get_context_id(queue[d]))->second.push_arg(v(d));
psize[d] = std::max(psize[d], v.part_size(d));
}
}

void operator()() {
for(unsigned d = 0; d < queue.size(); d++) {
auto &K = cache.find(backend::get_context_id(queue[d]))->second;

if (psize[d]) {
K.push_arg(psize[d]);
K(queue[d]);

psize[d] = 0;
} else {
K.reset();
}
}
}

#ifndef BOOST_NO_VARIADIC_TEMPLATES
/// Launches the kernel with the provided parameters.
template <class... Param>
void operator()(const Param&... param) {
launch(boost::fusion::vector_tie(param...));
boost::fusion::for_each(boost::fusion::vector_tie(param...), push_args(*this));
(*this)();
}
#else

#define VEXCL_FUNCALL_OPERATOR(z, n, data) \
template <BOOST_PP_ENUM_PARAMS(n, class Param)> \
void operator()(BOOST_PP_ENUM_BINARY_PARAMS(n, const Param, &param)) { \
launch(boost::fusion::vector_tie(BOOST_PP_ENUM_PARAMS(n, param))); \
boost::fusion::for_each( \
boost::fusion::vector_tie(BOOST_PP_ENUM_PARAMS(n, param)), \
push_args(*this) \
); \
(*this)(); \
}

BOOST_PP_REPEAT_FROM_TO(1, VEXCL_MAX_ARITY, VEXCL_FUNCALL_OPERATOR, ~)

#undef VEXCL_FUNCALL_OPERATOR

#endif
private:

template <class ParamTuple>
void launch(const ParamTuple &param) {
static_assert(
boost::fusion::result_of::size<ParamTuple>::value == NP,
"Wrong number of kernel parameters"
);

for(unsigned d = 0; d < queue.size(); d++) {
if (size_t psize = boost::fusion::fold(param, 0, param_size(d))) {
auto key = backend::get_context_id(queue[d]);
auto krn = cache.find(key);
krn->second.push_arg(psize);

set_params setprm(krn->second, d);
boost::fusion::for_each(param, setprm);

krn->second(queue[d]);
}
}
}

struct declare_params {
backend::source_generator &src;

declare_params(backend::source_generator &src) : src(src) {}
struct add_params {
kernel &K;

template <class T>
void operator()(const T &v) const {
src << ",\n\t" << v.prmdecl();
}
};

struct read_params {
backend::source_generator &src;

read_params(backend::source_generator &src) : src(src) {}

template <class T>
void operator()(const T &v) const {
src << v.init();
}
};

struct write_params {
backend::source_generator &src;

write_params(backend::source_generator &src) : src(src) {}
add_params(kernel &K) : K(K) {}

template <class T>
void operator()(const T &v) const {
src << v.write();
K.add_param(v);
}
};

struct set_params {
backend::kernel &krn;
unsigned d;

set_params(backend::kernel &krn, unsigned d)
: krn(krn), d(d) {};

template <class T>
void operator()(const T &v) const {
krn.push_arg(v);
}
template <class T>
void operator()(const vector<T> &v) const {
krn.push_arg(v(d));
}
};

private:
std::vector<backend::command_queue> queue;

std::string name;
std::vector<size_t> psize;
std::ostringstream prm_decl, prm_read, prm_save;
std::map<vex::backend::context_id, vex::backend::kernel> cache;

struct param_size {
unsigned device;

param_size(unsigned device) : device(device) {}

typedef size_t result_type;

template <class T>
size_t operator()(size_t s, const T&) const {
return s;
}
struct push_args {
kernel &K;
push_args(kernel &K) : K(K) {}

template <class T>
size_t operator()(size_t s, const vector<T> &v) const {
return std::max(s, v.part_size(device));
void operator()(const T &p) const {
K.push_arg(p);
}
};
};
Expand Down Expand Up @@ -700,12 +664,18 @@ class Function {
* arguments.
*/
template <class... Args>
auto build_kernel(
kernel build_kernel(
const std::vector<backend::command_queue> &queue,
const std::string &name, const std::string& body, const Args&... args
) -> Kernel<sizeof...(Args)>
)
{
return Kernel<sizeof...(Args)>(queue, name, body, boost::fusion::vector_tie(args...));
kernel K(queue, name);
boost::fusion::for_each(
boost::fusion::vector_tie(args...),
kernel::add_params(K)
);
K.build(body);
return K;
}

/// Builds function body from the recorded expression.
Expand All @@ -721,11 +691,15 @@ std::string make_function(std::string body, const Ret &ret, const Args&... args)

#define VEXCL_BUILD_KERNEL(z, n, data) \
template <BOOST_PP_ENUM_PARAMS(n, class Arg)> \
Kernel<n> build_kernel(const std::vector<backend::command_queue> & queue, \
kernel build_kernel(const std::vector<backend::command_queue> & queue, \
const std::string & name, const std::string & body, \
BOOST_PP_ENUM_BINARY_PARAMS(n, const Arg, &arg)) { \
return Kernel<n>(queue, name, body, \
boost::fusion::vector_tie(BOOST_PP_ENUM_PARAMS(n, arg))); \
kernel K(queue, name); \
boost::fusion::for_each( \
boost::fusion::vector_tie(BOOST_PP_ENUM_PARAMS(n, arg)), \
detail::kernel_add_param(K)); \
K.build(body); \
return K; \
}

#define VEXCL_MAKE_FUNCTION(z, n, data) \
Expand Down

0 comments on commit e2fad48

Please sign in to comment.