gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Some renaming; lower overhead in benchmark

Browse Source
This commit is contained in:
Philippe Tillet
2015-02-01 22:28:49 -05:00
parent f0bb130416
commit 535706f35a
23 changed files with 239 additions and 213 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
lib/model/model.cpp
View File

@@ -28,9 +28,9 @@ std::string model::define_extension(std::string const & extensions, std::string
return std::string("");
}
void model::fill_program_name(char* program_name, array_expressions_container const & array_expressions, binding_policy_t binding_policy)
void model::fill_program_name(char* program_name, expressions_tuple const & expressions, binding_policy_t binding_policy)
{
if (array_expressions.order()==array_expressions_container::INDEPENDENT)
if (expressions.order()==expressions_tuple::INDEPENDENT)
*program_name++='i';
else
*program_name++='s';
@@ -39,31 +39,39 @@ void model::fill_program_name(char* program_name, array_expressions_container co
binder = new bind_to_handle();
else
binder = new bind_all_unique();
for (array_expressions_container::data_type::const_iterator it = array_expressions.data().begin(); it != array_expressions.data().end(); ++it)
for (expressions_tuple::data_type::const_iterator it = expressions.data().begin(); it != expressions.data().end(); ++it)
traverse(**it, (*it)->root(), array_expression_representation_functor(*binder, program_name),true);
*program_name='\0';
delete binder;
}
std::vector<cl_ext::lazy_compiler>& model::init(array_expressions_container const & array_expressions, cl::Context const & context, cl::Device const & device, bool force_recompilation)
std::vector<cl_ext::lazy_compiler>& model::init(expressions_tuple const & expressions, runtime_options const & opt)
{
char program_name[256];
fill_program_name(program_name, array_expressions, BIND_TO_HANDLE);
std::string pname(program_name);
cl::Context const & context = expressions.context();
std::string pname;
if(opt.program_name.empty())
{
char program_name[256];
fill_program_name(program_name, expressions, BIND_TO_HANDLE);
pname = std::string(program_name);
}
else
pname = opt.program_name;
std::vector<cl_ext::lazy_compiler> & to_init = lazy_programs_[context()][pname];
if(to_init.empty())
{
cl::Device device = queue_.getInfo<CL_QUEUE_DEVICE>();
std::string extensions = device.getInfo<CL_DEVICE_EXTENSIONS>();
to_init.push_back(cl_ext::lazy_compiler(context, pname, force_recompilation));
to_init.push_back(cl_ext::lazy_compiler(context, pname, opt.recompile));
to_init.back().add(define_extension(extensions, "cl_khr_fp64"));
to_init.push_back(cl_ext::lazy_compiler(context, pname + "_fb", force_recompilation));
to_init.push_back(cl_ext::lazy_compiler(context, pname + "_fb", opt.recompile));
to_init.back().add(define_extension(extensions, "cl_khr_fp64"));
for(size_t i = 0 ; i < templates_.size() ; ++i)
{
std::vector<std::string> cur = templates_[i]->generate(i, array_expressions, device);
std::vector<std::string> cur = templates_[i]->generate(i, expressions, device);
for(size_t j = 0 ; j < cur.size() ; ++j){
to_init[j].add(cur[j]);
}
@@ -82,42 +90,37 @@ model::model(std::vector< tools::shared_ptr<base> > const & templates, cl::Comma
model::model(base const & tp, cl::CommandQueue & queue) : templates_(1,tp.clone()), queue_(queue)
{}
void model::execute(array_expressions_container const & array_expressions, bool bypass_predictor, bool force_recompilation)
void model::execute(expressions_tuple const & expressions, runtime_options const & opt)
{
bypass_predictor = bypass_predictor || predictor_.get()==NULL;
cl::Context const & context = array_expressions.context();
assert(context() == queue_.getInfo<CL_QUEUE_CONTEXT>()());
cl::Device const & device = queue_.getInfo<CL_QUEUE_DEVICE>();
std::vector<cl_ext::lazy_compiler> & compilers = init(array_expressions, context, device, force_recompilation);
std::vector<cl_ext::lazy_compiler> & compilers = init(expressions, opt);
//Prediction
std::vector<int_t> x = templates_[0]->input_sizes(array_expressions);
int label;
//The user tuned the model specifically for this input size
if(hardcoded_.find(x)!=hardcoded_.end())
label = hardcoded_.at(x);
//The user bypasses the random forest
else if(bypass_predictor)
label = 0;
//Default
int label = 0;
if(opt.label>=0)
{
label = opt.label;
}
else
{
std::vector<float> predictions = predictor_->predict(x);
label = std::distance(predictions.begin(),std::min_element(predictions.begin(), predictions.end()));
std::vector<int_t> x = templates_[0]->input_sizes(expressions);
//The user tuned the model specifically for this input size
if(hardcoded_.find(x)!=hardcoded_.end())
label = hardcoded_.at(x);
//The user bypasses the random forest
else if(predictor_.get())
{
std::vector<float> predictions = predictor_->predict(x);
label = std::distance(predictions.begin(),std::min_element(predictions.begin(), predictions.end()));
}
}
//Execution
templates_[label]->enqueue(queue_, compilers, label, array_expressions);
templates_[label]->enqueue(queue_, compilers, label, expressions);
}
void model::tune(array_expressions_container const & array_expressions)
void model::tune(expressions_tuple const & expressions)
{
cl::Context const & context = array_expressions.context();
assert(context() == queue_.getInfo<CL_QUEUE_CONTEXT>()());
cl::Device device = queue_.getInfo<CL_QUEUE_DEVICE>();
std::vector<cl_ext::lazy_compiler> & compilers = init(array_expressions, context, device, false);
std::vector<cl_ext::lazy_compiler> & compilers = init(expressions);
//Collect the timings
std::vector<float> timings(templates_.size());
@@ -125,13 +128,13 @@ void model::tune(array_expressions_container const & array_expressions)
for(size_t i = 0 ; i < templates_.size() ; ++i)
{
timer.start();
templates_[i]->enqueue(queue_, compilers, i, array_expressions);
templates_[i]->enqueue(queue_, compilers, i, expressions);
queue_.finish();
timings[i] = timer.get();
}
//Fill the override
std::vector<int_t> x = templates_[0]->input_sizes(array_expressions);
std::vector<int_t> x = templates_[0]->input_sizes(expressions);
hardcoded_[x] = std::distance(timings.begin(),std::min_element(timings.begin(), timings.end()));
}