Skip to content

Commit

Permalink
[Type Casting Pass] Add support for fp64->fp32 conversion
Browse files Browse the repository at this point in the history
  • Loading branch information
@weimingzha0
Weiming Zhao authored and weimingzha0 committed Nov 30, 2021
1 parent edf5770 commit 0750995
Show file tree
Hide file tree
Showing 7 changed files with 66 additions and 44 deletions.
2 changes: 1 addition & 1 deletion include/halo/lib/transforms/typecast.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ namespace halo {
/// This pass eliminates usage of int64 by casting down.
class TypeCast final : public FunctionPass {
public:
TypeCast() : FunctionPass("Cast down int64 to int32") {}
TypeCast() : FunctionPass("Cast down int64/fp64 to int32/fp32") {}

bool RunOnFunction(Function* func) override;
};
Expand Down
56 changes: 39 additions & 17 deletions lib/transforms/typecast.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,22 @@

namespace halo {

template <typename Tsrc, typename Tdst>
static bool ReplaceConstant(Constant* c, DataType src_type, DataType new_type) {
const auto& orig_type = c->GetResultType(0);
if (orig_type.GetDataType() != src_type) {
return false;
}
std::vector<Tdst> ret;
auto n = orig_type.GetTotalNumOfElements();
ret.reserve(n);
for (unsigned int i = 0; i < n; ++i) {
ret.push_back(static_cast<float>(c->GetData<Tsrc>(i)));
}
c->SetData(halo::Type{new_type, orig_type.GetDimSizes()}, ret.data());
return true;
}

bool TypeCast::RunOnFunction(Function* func) {
bool changed = false;
// Replace arguments.
Expand All @@ -32,25 +48,27 @@ bool TypeCast::RunOnFunction(Function* func) {
halo::Type new_ty{DataType::INT32, ty.GetDimSizes()};
arg->SetType(new_ty);
changed |= true;
} else if (ty.GetDataType() == DataType::FLOAT64) {
halo::Type new_ty{DataType::FLOAT32, ty.GetDimSizes()};
arg->SetType(new_ty);
changed |= true;
}
}

// Replace constants.
ConstantBuilder cb(func);
Function::ConstantList& constants = func->Constants();
for (auto it = constants.begin(), ie = constants.end(); it != ie; ++it) {
const auto& orig_type = (*it)->GetResultType(0);
if (orig_type.GetDataType() == DataType::INT64) {
std::vector<int> ret;
ret.reserve(orig_type.GetTotalNumOfElements());
for (unsigned int i = 0; i < orig_type.GetTotalNumOfElements(); ++i) {
ret.push_back(static_cast<int>((*it)->GetData<int64_t>(i)));
}
Constant* c_ret = cb.CreateConstant(
(*it)->GetName() + "_castdown",
halo::Type{DataType::INT32, orig_type.GetDimSizes()}, ret.data());
(*it)->ReplaceAllUsesWith(0, *c_ret);
changed = true;
}
Module* m = func->GetParent();
for (auto& c : m->Constants()) {
changed |= ReplaceConstant<double, float>(c.get(), DataType::FLOAT64,
DataType::FLOAT32);
changed |= ReplaceConstant<int64_t, int32_t>(c.get(), DataType::INT64,
DataType::INT32);
}

for (auto& c : func->Constants()) {
changed |= ReplaceConstant<double, float>(c.get(), DataType::FLOAT64,
DataType::FLOAT32);
changed |= ReplaceConstant<int64_t, int32_t>(c.get(), DataType::INT64,
DataType::INT32);
}

for (auto& bb : *func) {
Expand All @@ -60,11 +78,15 @@ bool TypeCast::RunOnFunction(Function* func) {
if (orig_type.IsValid() && orig_type.GetDataType() == DataType::INT64) {
inst->GetResultsTypes()[i] =
halo::Type{DataType::INT32, orig_type.GetDimSizes()};
} else if (orig_type.IsValid() &&
orig_type.GetDataType() == DataType::FLOAT64) {
inst->GetResultsTypes()[i] =
halo::Type{DataType::FLOAT32, orig_type.GetDimSizes()};
}
}
}
}
return changed;
} // namespace halo
}

} // end namespace halo
8 changes: 4 additions & 4 deletions tests/unittests/lit_cases/test_tensorrt/test_einsum_batch_diagonal_tensorrt.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,12 +17,12 @@

// clang-format off
// Testing CXX Code Gen using ODLA API on tensorrt
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_batch_diagonal/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_diagonal/test_data_set_0/output_0.pb
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_batch_diagonal/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_diagonal/test_data_set_0/input_0.pb
// RUN: %halo_compiler -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_batch_diagonal/model.onnx -o %t.cc
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_batch_diagonal/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_diagonal/test_data_set_0/output_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_batch_diagonal/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_diagonal/test_data_set_0/input_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_batch_diagonal/model.onnx -o %t.cc
// RUN: %cxx -c -fPIC -o %t.o %t.cc -I%odla_path/include
// RUN: %cxx -g %s %t.o %t.bin -I%T -I%odla_path/include -I%unittests_path -I%data_path/test_einsum_batch_diagonal/test_data_set_0 %odla_link %device_link -lodla_tensorrt -o %t_tensorrt.exe -Wno-deprecated-declarations
// RUN: %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_batch_diagonal | FileCheck %s
// RUN: ODLA_TRT_USE_EXPLICIT_BATCH=1 %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_batch_diagonal | FileCheck %s
// CHECK: Result Pass
// clang-format on
// XFAIL: *
Expand Down
12 changes: 6 additions & 6 deletions tests/unittests/lit_cases/test_tensorrt/test_einsum_batch_matmul_tensorrt.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,14 +17,14 @@

// clang-format off
// Testing CXX Code Gen using ODLA API on tensorrt
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_batch_matmul/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_matmul/test_data_set_0/input_0.pb
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_batch_matmul/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_matmul/test_data_set_0/output_0.pb
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_batch_matmul/test_data_set_0/input_1.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_matmul/test_data_set_0/input_1.pb
// RUN: %halo_compiler -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_batch_matmul/model.onnx -o %t.cc
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_batch_matmul/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_matmul/test_data_set_0/input_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_batch_matmul/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_matmul/test_data_set_0/output_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_batch_matmul/test_data_set_0/input_1.cc -x onnx -emit-data-as-c %data_path/test_einsum_batch_matmul/test_data_set_0/input_1.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_batch_matmul/model.onnx -o %t.cc
// RUN: %cxx -c -fPIC -o %t.o %t.cc -I%odla_path/include
// RUN: %cxx -g %s %t.o %t.bin -I%T -I%odla_path/include -I%unittests_path -I%data_path/test_einsum_batch_matmul/test_data_set_0 %odla_link %device_link -lodla_tensorrt -o %t_tensorrt.exe -Wno-deprecated-declarations
// RUN: %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_batch_matmul | FileCheck %s
// RUN: ODLA_TRT_USE_EXPLICIT_BATCH=1 %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_batch_matmul | FileCheck %s
// CHECK: Result Pass
// clang-format on
// XFAIL: *

#include "test_einsum_batch_matmul_tensorrt.cc.tmp.main.cc.in"
12 changes: 6 additions & 6 deletions tests/unittests/lit_cases/test_tensorrt/test_einsum_inner_prod_tensorrt.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,14 +17,14 @@

// clang-format off
// Testing CXX Code Gen using ODLA API on tensorrt
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_inner_prod/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_inner_prod/test_data_set_0/input_0.pb
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_inner_prod/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_inner_prod/test_data_set_0/output_0.pb
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_inner_prod/test_data_set_0/input_1.cc -x onnx -emit-data-as-c %data_path/test_einsum_inner_prod/test_data_set_0/input_1.pb
// RUN: %halo_compiler -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_inner_prod/model.onnx -o %t.cc
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_inner_prod/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_inner_prod/test_data_set_0/input_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_inner_prod/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_inner_prod/test_data_set_0/output_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_inner_prod/test_data_set_0/input_1.cc -x onnx -emit-data-as-c %data_path/test_einsum_inner_prod/test_data_set_0/input_1.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_inner_prod/model.onnx -o %t.cc
// RUN: %cxx -c -fPIC -o %t.o %t.cc -I%odla_path/include
// RUN: %cxx -g %s %t.o %t.bin -I%T -I%odla_path/include -I%unittests_path -I%data_path/test_einsum_inner_prod/test_data_set_0 %odla_link %device_link -lodla_tensorrt -o %t_tensorrt.exe -Wno-deprecated-declarations
// RUN: %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_inner_prod | FileCheck %s
// RUN: ODLA_TRT_USE_EXPLICIT_BATCH=1 %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_inner_prod | FileCheck %s
// CHECK: Result Pass
// clang-format on
// XFAIL: *

#include "test_einsum_inner_prod_tensorrt.cc.tmp.main.cc.in"
10 changes: 5 additions & 5 deletions tests/unittests/lit_cases/test_tensorrt/test_einsum_sum_tensorrt.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,13 +17,13 @@

// clang-format off
// Testing CXX Code Gen using ODLA API on tensorrt
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_sum/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_sum/test_data_set_0/output_0.pb
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_sum/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_sum/test_data_set_0/input_0.pb
// RUN: %halo_compiler -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_sum/model.onnx -o %t.cc
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_sum/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_sum/test_data_set_0/output_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_sum/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_sum/test_data_set_0/input_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_sum/model.onnx -o %t.cc
// RUN: %cxx -c -fPIC -o %t.o %t.cc -I%odla_path/include
// RUN: %cxx -g %s %t.o %t.bin -I%T -I%odla_path/include -I%unittests_path -I%data_path/test_einsum_sum/test_data_set_0 %odla_link %device_link -lodla_tensorrt -o %t_tensorrt.exe -Wno-deprecated-declarations
// RUN: %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_sum | FileCheck %s
// RUN: ODLA_TRT_USE_EXPLICIT_BATCH=1 %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_sum | FileCheck %s
// CHECK: Result Pass
// clang-format on
// XFAIL: *

#include "test_einsum_sum_tensorrt.cc.tmp.main.cc.in"
10 changes: 5 additions & 5 deletions tests/unittests/lit_cases/test_tensorrt/test_einsum_transpose_tensorrt.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,13 +17,13 @@

// clang-format off
// Testing CXX Code Gen using ODLA API on tensorrt
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_transpose/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_transpose/test_data_set_0/output_0.pb
// RUN: %halo_compiler -target cxx -o %data_path/test_einsum_transpose/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_transpose/test_data_set_0/input_0.pb
// RUN: %halo_compiler -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_transpose/model.onnx -o %t.cc
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_transpose/test_data_set_0/output_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_transpose/test_data_set_0/output_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -o %data_path/test_einsum_transpose/test_data_set_0/input_0.cc -x onnx -emit-data-as-c %data_path/test_einsum_transpose/test_data_set_0/input_0.pb
// RUN: %halo_compiler --disable-type-cast=false -target cxx -batch-size 1 %halo_compile_flags %data_path/test_einsum_transpose/model.onnx -o %t.cc
// RUN: %cxx -c -fPIC -o %t.o %t.cc -I%odla_path/include
// RUN: %cxx -g %s %t.o %t.bin -I%T -I%odla_path/include -I%unittests_path -I%data_path/test_einsum_transpose/test_data_set_0 %odla_link %device_link -lodla_tensorrt -o %t_tensorrt.exe -Wno-deprecated-declarations
// RUN: %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_transpose | FileCheck %s
// RUN: ODLA_TRT_USE_EXPLICIT_BATCH=1 %t_tensorrt.exe 0.0001 0 tensorrt %data_path/test_einsum_transpose | FileCheck %s
// CHECK: Result Pass
// clang-format on
// XFAIL: *

#include "test_einsum_transpose_tensorrt.cc.tmp.main.cc.in"

0 comments on commit 0750995

Please sign in to comment.