/// @brief Read the new scattering from the scoplib description.
///
/// @S The Scop to update
/// @OScop The ScopLib data structure describing the new scattering.
/// @return A map that contains for each Statement the new scattering.
StatementToIslMapTy *readScattering(Scop *S, scoplib_scop_p OScop) {
StatementToIslMapTy &NewScattering = *(new StatementToIslMapTy());
scoplib_statement_p stmt = OScop->statement;
// Check if we have dimensions for each scattering or if each row
// represents a scattering dimension.
int numScatteringDims = -1;
ScopStmt *pollyStmt = *S->begin();
if (stmt->schedule->NbColumns
== 2 + pollyStmt->getNumParams() + pollyStmt->getNumIterators()) {
numScatteringDims = maxScattering(stmt);
}
for (Scop::iterator SI = S->begin(), SE = S->end(); SI != SE; ++SI) {
if (!stmt) {
errs() << "Not enough statements available in OpenScop file\n";
freeStmtToIslMap(&NewScattering);
return NULL;
}
NewScattering[*SI] = scatteringForStmt(stmt->schedule, *SI,
numScatteringDims);
stmt = stmt->next;
}
if (stmt) {
errs() << "Too many statements in OpenScop file\n";
freeStmtToIslMap(&NewScattering);
return NULL;
}
return &NewScattering;
}
void ClastStmtCodeGen::codegenForGPGPU(const clast_for *F) {
BasicBlock::iterator LoopBody;
SetVector<Value *> Values;
SetVector<Value *> IVS;
std::vector<int> NumIterations;
PTXGenerator::ValueToValueMapTy VMap;
assert(!GPUTriple.empty() &&
"Target triple should be set properly for GPGPU code generation.");
PTXGenerator PTXGen(Builder, P, GPUTriple);
// Get original IVS and ScopStmt
unsigned TiledLoopDepth, NonPLoopDepth;
const clast_stmt *InnerStmt =
getScheduleInfo(F, NumIterations, TiledLoopDepth, NonPLoopDepth);
const clast_stmt *TmpStmt;
const clast_user_stmt *U;
const clast_for *InnerFor;
if (CLAST_STMT_IS_A(InnerStmt, stmt_for)) {
InnerFor = (const clast_for *)InnerStmt;
TmpStmt = InnerFor->body;
} else
TmpStmt = InnerStmt;
U = (const clast_user_stmt *)TmpStmt;
ScopStmt *Statement = (ScopStmt *)U->statement->usr;
for (unsigned i = 0; i < Statement->getNumIterators() - NonPLoopDepth; i++) {
const Value *IV = Statement->getInductionVariableForDimension(i);
IVS.insert(const_cast<Value *>(IV));
}
unsigned OutBytes;
Values = getGPUValues(OutBytes);
PTXGen.setOutputBytes(OutBytes);
PTXGen.startGeneration(Values, IVS, VMap, &LoopBody);
BasicBlock::iterator AfterLoop = Builder.GetInsertPoint();
Builder.SetInsertPoint(LoopBody);
BasicBlock *AfterBB = 0;
if (NonPLoopDepth) {
Value *LowerBound, *UpperBound, *IV, *Stride;
Type *IntPtrTy = getIntPtrTy();
LowerBound = ExpGen.codegen(InnerFor->LB, IntPtrTy);
UpperBound = ExpGen.codegen(InnerFor->UB, IntPtrTy);
Stride = Builder.getInt(APInt_from_MPZ(InnerFor->stride));
IV = createLoop(LowerBound, UpperBound, Stride, Builder, P, AfterBB,
CmpInst::ICMP_SLE);
const Value *OldIV_ = Statement->getInductionVariableForDimension(2);
Value *OldIV = const_cast<Value *>(OldIV_);
VMap.insert(std::make_pair<Value *, Value *>(OldIV, IV));
}
updateWithValueMap(VMap);
BlockGenerator::generate(Builder, *Statement, ValueMap, P);
if (AfterBB)
Builder.SetInsertPoint(AfterBB->begin());
// FIXME: The replacement of the host base address with the parameter of ptx
// subfunction should have been done by updateWithValueMap. We use the
// following codes to avoid affecting other parts of Polly. This should be
// fixed later.
Function *FN = Builder.GetInsertBlock()->getParent();
for (unsigned j = 0; j < Values.size(); j++) {
Value *baseAddr = Values[j];
for (Function::iterator B = FN->begin(); B != FN->end(); ++B) {
for (BasicBlock::iterator I = B->begin(); I != B->end(); ++I)
I->replaceUsesOfWith(baseAddr, ValueMap[baseAddr]);
}
}
Builder.SetInsertPoint(AfterLoop);
PTXGen.setLaunchingParameters(NumIterations[0], NumIterations[1],
NumIterations[2], NumIterations[3]);
PTXGen.finishGeneration(FN);
}