added addrspace(4)/constant memory for const uniform declarations

This commit is contained in:
Evghenii
2014-01-08 13:27:24 +01:00
parent f011b3cb22
commit de4d66c56f
6 changed files with 139 additions and 88 deletions

View File

@@ -63,11 +63,17 @@ define i32 @__nctaid_z() nounwind readnone alwaysinline
}
;;;;;;;;
declare i64* @llvm.nvvm.ptr.shared.to.gen.p0i64.p3i64(i64 addrspace(3)*)
declare i64* @llvm.nvvm.ptr.shared.to.gen.p0i64.p4i64(i64 addrspace(4)*)
define i64* @__cvt_loc2gen(i64 addrspace(3)*) nounwind readnone alwaysinline
{
%ptr = tail call i64* @llvm.nvvm.ptr.shared.to.gen.p0i64.p3i64(i64 addrspace(3)* %0)
ret i64* %ptr
}
define i64* @__cvt_const2gen(i64 addrspace(4)*) nounwind readnone alwaysinline
{
%ptr = tail call i64* @llvm.nvvm.ptr.shared.to.gen.p0i64.p4i64(i64 addrspace(4)* %0)
ret i64* %ptr
}
;;;;;;;;
define i32 @__shfl_i32_nvptx(i32, i32) nounwind readnone alwaysinline
{

View File

@@ -4543,5 +4543,6 @@ declare i32 @__nctaid_x() nounwind readnone alwaysinline
declare i32 @__nctaid_y() nounwind readnone alwaysinline
declare i32 @__nctaid_z() nounwind readnone alwaysinline
declare i64* @__cvt_loc2gen(i64 addrspace(3)*) nounwind readnone alwaysinline
declare i64* @__cvt_const2gen(i64 addrspace(4)*) nounwind readnone alwaysinline
')

39
ctx.cpp
View File

@@ -1833,8 +1833,9 @@ FunctionEmitContext::BitCastInst(llvm::Value *value, llvm::Type *type,
}
/* NVPTX:
* this is a helper function which adds a warp offset to a base pointer in local memory
* either in addrspace(3) or converted from addrspace(3) to addrspace(0)
* this is a helper function which adds a warp offset to a base pointer
* pointer must either be in local memory addrspace(3)
* or the one just converted from addrspace(3) to addrspace(0) in lConvertToGenericPtr
*/
static llvm::Value* lAddWarpOffset(FunctionEmitContext *ctx, llvm::Value *value)
{
@@ -1848,7 +1849,8 @@ static llvm::Value* lAddWarpOffset(FunctionEmitContext *ctx, llvm::Value *value)
}
/* NVPTX:
* this function compute correct address in local memory for load/store operations*/
* this function compute correct address in local memory for load/store operations
*/
static llvm::Value* lCorrectLocalPtr(FunctionEmitContext *ctx, llvm::Value* value)
{
assert(value->getType()->isPointerTy());
@@ -1858,32 +1860,37 @@ static llvm::Value* lCorrectLocalPtr(FunctionEmitContext *ctx, llvm::Value* valu
}
/* NVPTX:
* this function converts pointers from addrspace(3) to addrspace(0)
* this function converts a pointer in addrspace(3 or 4) to addrspace(0)
*/
static llvm::Value* lConvertLocal2GenericPtr(FunctionEmitContext *ctx, llvm::Value *value)
static llvm::Value* lConvertToGenericPtr(FunctionEmitContext *ctx, llvm::Value *value, const SourcePos &currentPos)
{
if (!value->getType()->isPointerTy() || g->target->getISA() != Target::NVPTX) return value;
llvm::PointerType *pt = llvm::dyn_cast<llvm::PointerType>(value->getType());
if (pt->getAddressSpace() != 3) return value;
/* if array, extracts its element type */
/* make sure addrspace corresponds to either local or constant memories */
const int addressSpace = pt->getAddressSpace();
if (addressSpace != 3 && addressSpace != 4) return value;
/* if array, extracts element type */
llvm::Type *type = pt->getElementType();
llvm::Type *typeEl = type;
if (type->isArrayTy())
{
typeEl = type->getArrayElementType();
assert(!typeEl->isArrayTy()); /* currently we don't support array-of-array in uniform */
if (typeEl->isArrayTy())
Error(currentPos, "Currently \"nvptx\" target doesn't support array-of-array");
}
/* convert elTy addrspace(3)* to i64* addrspace(3)* */
llvm::PointerType *Int64Ptr3 = llvm::PointerType::get(LLVMTypes::Int64Type, 3);
value = ctx->BitCastInst(value, Int64Ptr3, "cvtLog2Gen_i64ptr");
llvm::PointerType *Int64Ptr3 = llvm::PointerType::get(LLVMTypes::Int64Type, addressSpace);
value = ctx->BitCastInst(value, Int64Ptr3, "cvt2gen_i64ptr");
/* convert i64* addrspace(3) to i64* */
llvm::Function *__cvt_loc2gen = m->module->getFunction("__cvt_loc2gen");
std::vector<llvm::Value *> __cvt_loc2gen_args;
__cvt_loc2gen_args.push_back(value);
value = ctx->CallInst(__cvt_loc2gen, NULL, __cvt_loc2gen_args, "cvtLoc2Gen");
llvm::Function *__cvt2gen = m->module->getFunction(
addressSpace == 3 ? "__cvt_loc2gen" : "__cvt_const2gen");
std::vector<llvm::Value *> __cvt2gen_args;
__cvt2gen_args.push_back(value);
value = ctx->CallInst(__cvt2gen, NULL, __cvt2gen_args, "cvt2gen_call");
/* convert i64* to elTy* */
llvm::PointerType *typeElPtr = llvm::PointerType::get(typeEl, 0);
@@ -1907,7 +1914,7 @@ FunctionEmitContext::PtrToIntInst(llvm::Value *value, const char *name) {
if (name == NULL)
name = LLVMGetName(value, "_ptr2int");
value = lConvertLocal2GenericPtr(this, value); /* NVPTX : convert addrspace 3->0 before converting pointer */
value = lConvertToGenericPtr(this, value, currentPos); /* NVPTX : convert to addrspace(0) */
llvm::Type *type = LLVMTypes::PointerIntType;
llvm::Instruction *inst = new llvm::PtrToIntInst(value, type, name, bblock);
AddDebugPos(inst);
@@ -1941,7 +1948,7 @@ FunctionEmitContext::PtrToIntInst(llvm::Value *value, llvm::Type *toType,
}
}
value = lConvertLocal2GenericPtr(this, value); /* NVPTX : convert addrspace 3->0 before converting pointer */
value = lConvertToGenericPtr(this, value, currentPos); /* NVPTX : convert to addrspace(0) */
llvm::Instruction *inst = new llvm::PtrToIntInst(value, toType, name, bblock);
AddDebugPos(inst);
return inst;

View File

@@ -2353,25 +2353,25 @@ Module::CompileAndOutput(const char *srcFile,
m = new Module(srcFile);
if (m->CompileFile() == 0) {
/* NVPTX:
* for PTX target replace '.' with '_' in all global variables
* a PTX identifier name must match [a-zA-Z$_][a-zA-Z$_0-9]*
*/
if (g->target->getISA() == Target::NVPTX)
{
llvm::Module::global_iterator
I = m->module->global_begin(),
E = m->module->global_end();
for (; I != E; I++)
/* NVPTX:
* for PTX target replace '.' with '_' in all global variables
* a PTX identifier name must match [a-zA-Z$_][a-zA-Z$_0-9]*
*/
if (g->target->getISA() == Target::NVPTX)
{
std::string name = I->getName();
for (int i = 0; i < name.length(); i++)
if (name[i] == '.')
name[i] = '_';
I->setName(name);
llvm::Module::global_iterator
I = m->module->global_begin(),
E = m->module->global_end();
for (; I != E; I++)
{
std::string name = I->getName();
for (int i = 0; i < name.length(); i++)
if (name[i] == '.')
name[i] = '_';
I->setName(name);
}
}
}
if (outputType == CXX) {
if (target == NULL || strncmp(target, "generic-", 8) != 0) {
Error(SourcePos(), "When generating C++ output, one of the \"generic-*\" "

138
stmt.cpp
View File

@@ -207,10 +207,16 @@ DeclStmt::EmitCode(FunctionEmitContext *ctx) const {
if (sym->storageClass == SC_STATIC) {
if (g->target->getISA() == Target::NVPTX && !sym->type->IsConstType())
Error(sym->pos,
"Non-constant static variable ""\"%s\" is not supported with ""\"nvptx\" target.",
sym->name.c_str());
if (g->target->getISA() == Target::NVPTX && !sym->type->IsConstType())
Error(sym->pos,
"Non-constant static variable ""\"%s\" is not supported with ""\"nvptx\" target.",
sym->name.c_str());
#if 0
if (g->target->getISA() == Target::NVPTX && sym->type->IsVaryingType())
Error(sym->pos,
"const static varying variable ""\"%s\" is not supported with ""\"nvptx\" target.",
sym->name.c_str());
#endif
// For static variables, we need a compile-time constant value
// for its initializer; if there's no initializer, we use a
@@ -248,59 +254,93 @@ DeclStmt::EmitCode(FunctionEmitContext *ctx) const {
llvm::Twine("static_") +
llvm::Twine(sym->pos.first_line) +
llvm::Twine("_") + sym->name.c_str());
#if 0
NULL,
llvm::GlobalVariable::NotThreadLocal,
3);
#endif
// Tell the FunctionEmitContext about the variable
ctx->EmitVariableDebugInfo(sym);
}
else {
if (sym->type->IsArrayType() && sym->type->IsUniformType()
&& g->target->getISA() == Target::NVPTX)
if (sym->type->IsArrayType() &&
sym->type->IsUniformType() &&
g->target->getISA() == Target::NVPTX)
{
if (initExpr != NULL)
Error(initExpr->pos,
"It is not possible to initialize \"uniform\" array \"%s\" with \"nvptx\" target. "
"Use \"varying\" or \"const static uniform\" if possible.",
sym->name.c_str());
/* deal with "const uniform" or "uniform" arrays for nvptx target */
PerformanceWarning(sym->pos,
"\"uniform\" arrays might be slow with \"nvptx\" target. "
"Unless data sharing between program instances is required, use \"varying\" if possible.");
if (initExpr != NULL && !sym->type->IsConstType())
Error(initExpr->pos,
"It is not possible to initialize non-constant \"uniform\" array \"%s\" with \"nvptx\" target. "
"Please use \"varying\" or \"const static uniform\".",
sym->name.c_str());
PerformanceWarning(sym->pos,
"\"uniform\" arrays might be slow with \"nvptx\" target."
" Unless data sharing between program instances is required, use \"varying\" instead.");
llvm::Constant *cinit = NULL;
llvm::Type *llvmTypeUn;
int addressSpace;
if (sym->type->IsConstType())
{
addressSpace = 4; /* constant */
if (initExpr != NULL) {
if (PossiblyResolveFunctionOverloads(initExpr, sym->type) == false)
continue;
// FIXME: we only need this for function pointers; it was
// already done for atomic types and enums in
// DeclStmt::TypeCheck()...
if (dynamic_cast<ExprList *>(initExpr) == NULL) {
initExpr = TypeConvertExpr(initExpr, sym->type,
"initializer");
// FIXME: and this is only needed to re-establish
// constant-ness so that GetConstant below works for
// constant artithmetic expressions...
initExpr = ::Optimize(initExpr);
}
const ArrayType *at = CastType<ArrayType>(sym->type);
const int nel = at->GetElementCount();
/* we must scale # elements by 4, because a thread-block will run 4 warps
* or 128 threads.
* ***note-to-me***:please define these value (128threads/4warps)
* in nvptx-target definition
* instead of compile-time constants
*/
const int nel4 = nel*4;
ArrayType nat(at->GetElementType(), nel4);
llvm::Type *llvmType = nat.LLVMType(g->ctx);
cinit = initExpr->GetConstant(sym->type);
if (cinit == NULL)
Error(initExpr->pos, "Initializer for static variable "
"\"%s\" must be a constant.", sym->name.c_str());
}
else
{
Error(sym->pos, "Missing initializer for const variable "
"\"%s\".", sym->name.c_str());
}
if (cinit == NULL)
cinit = llvm::Constant::getNullValue(llvmType);
llvmTypeUn = llvmType;
}
else
{
addressSpace = 3; /* local */
const ArrayType *at = CastType<ArrayType>(sym->type);
const int nel = at->GetElementCount();
/* we must scale # elements by 4, because a thread-block will run 4 warps
* or 128 threads.
* ***note-to-me***:please define these value (128threads/4warps)
* in nvptx-target definition
* instead of compile-time constants
*/
const int nel4 = nel*4;
ArrayType nat(at->GetElementType(), nel4);
llvmTypeUn = nat.LLVMType(g->ctx);
cinit = llvm::UndefValue::get(llvmTypeUn);
}
// addrspace(3) must be undefined at initialization
llvm::Constant *cinit = llvm::UndefValue::get(llvmType);
sym->storagePtr =
new llvm::GlobalVariable(*m->module, llvmType,
sym->type->IsConstType(),
llvm::GlobalValue::PrivateLinkage,
cinit,
llvm::Twine("local_") +
llvm::Twine(sym->pos.first_line) +
llvm::Twine("_") + sym->name.c_str(),
NULL,
llvm::GlobalVariable::NotThreadLocal,
/*AddressSpace=*/ 3);
sym->storagePtr =
new llvm::GlobalVariable(*m->module, llvmTypeUn,
sym->type->IsConstType(),
llvm::GlobalValue::PrivateLinkage,
cinit,
llvm::Twine("local_") +
llvm::Twine(sym->pos.first_line) +
llvm::Twine("_") + sym->name.c_str(),
NULL,
llvm::GlobalVariable::NotThreadLocal,
addressSpace);
#if 0
llvm::GlobalVariable *var = llvm::dyn_cast<llvm::GlobalVariable>(sym->storagePtr);
var->setAlignment(128);
llvm::GlobalVariable *var = llvm::dyn_cast<llvm::GlobalVariable>(sym->storagePtr);
var->setAlignment(128);
#endif
// Tell the FunctionEmitContext about the variable
ctx->EmitVariableDebugInfo(sym);
// Tell the FunctionEmitContext about the variable
ctx->EmitVariableDebugInfo(sym);
}
else
{
@@ -311,6 +351,10 @@ DeclStmt::EmitCode(FunctionEmitContext *ctx) const {
// this before the initializer stuff.
ctx->EmitVariableDebugInfo(sym);
if (initExpr == 0 && sym->type->IsConstType())
Error(sym->pos, "Missing initializer for const variable "
"\"%s\".", sym->name.c_str());
// And then get it initialized...
sym->parentFunction = ctx->GetFunction();
InitSymbol(sym->storagePtr, sym->type, initExpr, ctx, sym->pos);

View File

@@ -24,14 +24,7 @@ export void f_f(uniform float RET[], uniform float aFOO[])
}
task void result_task(uniform float RET[])
{
uniform float ret = -1;
switch (taskIndex)
{
case 0: ret = 496; break;
case 1: ret = 1520; break;
case 2: ret = 2544; break;
case 3: ret = 3568; break;
}
const uniform float ret = reduce_add(programIndex + programCount*taskIndex);
if (programIndex < 32/4)
RET[programCount/4*taskIndex + programIndex] = ret;
}