void GpuRule::generateCallCode(const std::string& name,
Transform& trans,
CodeGenerator& o,
const SimpleRegionPtr& region,
RuleFlavor flavor,
bool,
std::vector<RegionNodeGroup>& regionNodesGroups,
int nodeID,
int gpuCopyOut,
SpatialCallType)
{
o.comment("from GpuRule::generateCallCode():");
switch(flavor) {
case RuleFlavor::SEQUENTIAL:
o.callSpatial(_rule->trampcodename(trans)+TX_OPENCL_POSTFIX, region);
break;
case RuleFlavor::WORKSTEALING:
o.mkCreateGpuSpatialMethodCallTask(trans.name(), name, trans.instClassName() + "_workstealing", _rule->trampcodename(trans), region, regionNodesGroups, nodeID, gpuCopyOut, _rule->getToRegions(), _rule->isDivisible());
break;
case RuleFlavor::DISTRIBUTED:
o.comment("flavor distributed");
UNIMPLEMENTED();
break;
default:
UNIMPLEMENTED();
}
}
void petabricks::IterationDefinition::unpackargs(CodeGenerator& o) const {
o.comment("Define iteration bounds");
for(size_t i=0; i<_var.size(); ++i){
o.write("const IndexT "+_begin[i]->toString()+" = "COORD_BEGIN_STR"["+jalib::XToString(i)+"];");
o.write("const IndexT "+_end[i]->toString()+" = "COORD_END_STR"["+jalib::XToString(i)+"];");
}
}
void petabricks::IterationDefinition::genScratchRegionLoopBegin(CodeGenerator& o){
if(isSingleCall()){
genLoopBegin(o);
}else{
o.comment("Iterate along all the directions");
// Compute size
for(size_t i=0; i<_size.size(); ++i){
o.write("int " + _size[i]->toString() + " = " + _end[i]->toString() +
" - " + _begin[i]->toString() + ";");
}
for(size_t i=0; i<_var.size(); ++i){
FormulaPtr b= new FormulaLiteral<int>(0);
FormulaPtr e=_size[i];
FormulaPtr s=_step[i];
FormulaPtr v=_var[i];
//TODO: expand to reorder dimensions
if(_order.canIterateForward(i) || !_order.canIterateBackward(i)){
JWARNING(_order.canIterateForward(i))(_order).Text("couldn't find valid iteration order, assuming forward");
o.beginFor(v->toString(), b, e, s);
} else {
o.beginReverseFor(v->toString(), b, e, s);
}
}
}
}
void petabricks::IterationDefinition::genLoopBegin(CodeGenerator& o){
if(isSingleCall()){
o.write("{");
for(size_t i=0; i<_var.size(); ++i){
o.varDecl("const IndexT "+_var[i]->toString()+" = "+_begin[i]->toString());
}
}else{
o.comment("Iterate along all the directions");
for(size_t i=0; i<_var.size(); ++i){
FormulaPtr b=_begin[i];
FormulaPtr e=_end[i];
FormulaPtr s=_step[i];
FormulaPtr v=_var[i];
//TODO: expand to reorder dimensions
if(_order.canIterateForward(i) || !_order.canIterateBackward(i)){
JWARNING(_order.canIterateForward(i))(_order).Text("couldn't find valid iteration order, assuming forward");
o.beginFor(v->toString(), b, e, s);
} else {
o.beginReverseFor(v->toString(), b, e, s);
}
}
}
}
void
GpuRule::generateCallTaskCode(const std::string& name, Transform& trans, CodeGenerator& o, const SimpleRegionPtr& region)
{
o.comment( "GENERATECALLTASKCODE" );
o.mkSpatialTask(name, trans.instClassName(), codename(), region, RuleFlavor::WORKSTEALING);
}
void
GpuRule::generateCallCodeSimple(Transform& /*trans*/, CodeGenerator& o, const SimpleRegionPtr& region)
{
o.comment( "GENERATECALLCODESIMPLE" );
o.callSpatial(codename(), region);
}
void petabricks::CodeGenerator::mkCreateGpuSpatialMethodCallTask(
const std::string& transname,
const std::string& taskname,
const std::string& objname,
const std::string& methodname,
const SimpleRegion& region,
std::vector<RegionNodeGroup>& regionNodesGroups,
int nodeID,
int gpuCopyOut,
RegionList to,
bool divisible) {
std::string taskclass
;
int dim_int = region.totalDimensions();
std::string lastdim = jalib::XToString(dim_int - 1);
std::string max = region.maxCoord()[dim_int - 1]->toString();
std::string min = region.minCoord()[dim_int - 1]->toString();
if(!divisible) {
taskclass = "petabricks::CreateGpuSpatialMethodCallTask<"+objname
+ ", " + jalib::XToString(region.totalDimensions())
+ ", &" + objname + "::" + methodname + TX_OPENCL_POSTFIX + "_createtasks"
+ ">";
//beginIf(min+"<"+max);
comment("MARKER 6");
write("IndexT _tmp_begin[] = {" + region.getIterationLowerBounds() + "};");
write("IndexT _tmp_end[] = {" + region.getIterationUpperBounds() + "};");
write("RegionNodeGroupMapPtr groups = new RegionNodeGroupMap();");
for(std::vector<RegionNodeGroup>::iterator group = regionNodesGroups.begin(); group != regionNodesGroups.end(); ++group){
write("{");
incIndent();
write("std::set<int> ids;");
for(std::vector<int>::iterator id = group->nodeIDs().begin(); id != group->nodeIDs().end(); ++id){
write("ids.insert("+jalib::XToString(*id)+");");
}
write("groups->insert(RegionNodeGroup(\""+group->matrixName()+"\",ids));");
decIndent();
write("}");
}
write(taskname+" = new "+taskclass+"(this,_tmp_begin, _tmp_end, "+jalib::XToString(nodeID)+", groups, "+jalib::XToString(gpuCopyOut)+");");
//endIf();
return;
}
std::string n_div = "cont_" + jalib::XToString(_contCounter++);
write(taskname + " = new petabricks::MethodCallTask<"+_curClass+", &"+_curClass+"::"+n_div+">( this );");
// Add divider function
CodeGenerator helper = forkhelper();
helper.beginFunc("DynamicTaskPtr", n_div);
helper.write("DynamicTaskPtr _fini = new NullDynamicTask();");
// Assign the gpu-cpu division point.
helper.write("ElementT gpu_ratio = "+transname+"_gpuratio/8.0;");
std::string div = "div";
RegionPtr proxy = to.front();
helper.write("IndexT totalRow = "+proxy->matrix()->name()+".size("+jalib::XToString(dim_int - 1)+");");
helper.write("IndexT div = ceil(gpu_ratio * totalRow);");
helper.beginIf("div > " + max);
helper.write("div = "+max+";");
helper.endIf();
// GPU
taskclass = "petabricks::CreateGpuSpatialMethodCallTask<"+objname
+ ", " + jalib::XToString(dim_int)
+ ", &" + objname + "::" + methodname + TX_OPENCL_POSTFIX + "_createtasks"
+ ">";
helper.comment("MARKER 6");
//helper.beginIf(min+" < div");
helper.write("IndexT _gpu_begin[] = {" + region.getIterationLowerBounds() + "};");
helper.write("IndexT _gpu_end[] = {" + region.getIterationMiddleEnd(div) + "};");
helper.write("RegionNodeGroupMapPtr groups = new RegionNodeGroupMap();");
for(std::vector<RegionNodeGroup>::iterator group = regionNodesGroups.begin(); group != regionNodesGroups.end(); ++group){
helper.write("{");
helper.incIndent();
helper.write("std::set<int> ids;");
for(std::vector<int>::iterator id = group->nodeIDs().begin(); id != group->nodeIDs().end(); ++id){
helper.write("ids.insert("+jalib::XToString(*id)+");");
}
helper.write("groups->insert(RegionNodeGroup(\""+group->matrixName()+"\",ids));");
helper.decIndent();
helper.write("}");
}
helper.write("DynamicTaskPtr gpu_task = new "+taskclass+"(this,_gpu_begin, _gpu_end, "+jalib::XToString(nodeID)+", groups, "+jalib::XToString(gpuCopyOut)+");");
helper.write("gpu_task->enqueue();");
helper.write("_fini->dependsOn(gpu_task);");
//helper.endIf();
// CPU
taskclass = "petabricks::SpatialMethodCallTask<CLASS"
", " + jalib::XToString(dim_int)
+ ", &CLASS::" + methodname + "_workstealing_wrap"
//+ ", &CLASS::" + methodname + "_workstealing"
+ ">";
helper.beginIf("div < " + max);
helper.beginIf("div < " + min);
helper.write("div = "+min+";");
helper.endIf();
helper.comment("MARKER 6");
helper.write("IndexT _cpu_begin[] = {" + region.getIterationMiddleBegin(div) + "};");
helper.write("IndexT _cpu_end[] = {" + region.getIterationUpperBounds() + "};");
helper.write("DynamicTaskPtr cpu_task = new "+taskclass+"(this,_cpu_begin, _cpu_end);");
helper.write("cpu_task->enqueue();");
helper.write("_fini->dependsOn(cpu_task);");
helper.endIf();
helper.write("return _fini;");
helper.endFunc();
}
