Skip to content

[Bug][CodeGen,Cuda]fix cast fp16 to int8/uint8 in cuda #13641

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
masahi merged 14 commits into from
Jan 2, 2023
Merged

[Bug][CodeGen,Cuda]fix cast fp16 to int8/uint8 in cuda #13641

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
e0ad1bb
[Fix Bug]fix the bug of tensorflow frontend when parsing Range layer
Jan 20, 2022
f032b20
Merge pull request #1 from apache/main
ninesheep Jan 21, 2022
3a357a7
Merge pull request #2 from apache/main
ninesheep Jan 24, 2022
c263c02
Merge pull request #3 from apache/main
ninesheep Jan 25, 2022
feadfc9
Merge pull request #4 from apache/main
ninesheep Jan 26, 2022
b3f39b4
Merge branch 'apache:main' into main
ninesheep Oct 17, 2022
32df505
Merge branch 'apache:main' into main
ninesheep Dec 1, 2022
ff4205a
Merge branch 'apache:main' into main
ninesheep Dec 5, 2022
6a922cd
[Fix Bug]fix the bug of schedule batch_matmul_int8 on cuda
ninesheep Dec 5, 2022
c9b4178
Merge branch 'apache:main' into main
ninesheep Dec 14, 2022
795bf5c
Merge branch 'apache:main' into main
ninesheep Dec 15, 2022
899d5b3
Merge branch 'apache:main' into main
ninesheep Dec 19, 2022
8fab01e
fix cast fp16 to int8/uint8 on cuda
ninesheep Dec 19, 2022
de9bae8
Merge branch 'apache:main' into main
ninesheep Dec 21, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions src/target/source/codegen_cuda.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -579,6 +579,23 @@ void CodeGenCUDA::PrintStorageScope(const std::string& scope, std::ostream& os)
}
}

std::string CodeGenCUDA::CastFromTo(std::string value, DataType from, DataType target) {
if (from == target) return value;
std::ostringstream os;
os << "((";
this->PrintType(target, os);
os << ")";
if (from.is_float16() && (target.is_int() || target.is_uint()) && target.bits() == 8) {
os << "(";
if (target.is_uint()) {
os << "u";
}
os << "int)";
}
os << value << ")";
return os.str();
}

void CodeGenCUDA::VisitExpr_(const CastNode* op, std::ostream& os) {
DataType from_ty = op->value.dtype();
DataType target_ty = op->dtype;
Expand Down
1 change: 1 addition & 0 deletions src/target/source/codegen_cuda.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,6 +58,7 @@ class CodeGenCUDA final : public CodeGenC {
void PrintVecElemStore(const std::string& vec, DataType t, int i, const std::string& value) final;
void BindThreadIndex(const IterVar& iv) final; // NOLINT(*)
void PrintVecElemLoadExpr(DataType t, int i, const std::string& value, std::ostream& os) final;
std::string CastFromTo(std::string value, DataType from, DataType target) final;
// overload visitor
void VisitExpr_(const RampNode* op, std::ostream& os) final; // NOLINT(*)
void VisitExpr_(const ShuffleNode* op, std::ostream& os) final; // NOLINT(*)
Expand Down