Skip to content
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

Jump to bottom

register fluid kerenls to phi [part 2] #52044

Merged
merged 9 commits into from
Mar 31, 2023
Merged

register fluid kerenls to phi [part 2] #52044

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
64c55b3
update bipartite_match
huangjiyi Mar 23, 2023
478b3e8
update
huangjiyi Mar 23, 2023
c0a7200
fix bug
huangjiyi Mar 23, 2023
83f337d
fix test
huangjiyi Mar 23, 2023
c7ff9db
fix bug
huangjiyi Mar 24, 2023
cc95cad
Merge remote-tracking branch 'Paddle/develop' into unify_kernel_5
huangjiyi Mar 26, 2023
ceab63c
fix Kunlun-KP-Build
huangjiyi Mar 26, 2023
ab847aa
Revert "fix Kunlun-KP-Build"
huangjiyi Mar 28, 2023
a422943
update
huangjiyi Mar 31, 2023
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
16 changes: 9 additions & 7 deletions paddle/fluid/operators/bpr_loss_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -174,17 +174,19 @@ class BprLossGradMaker : public framework::SingleGradOpMaker<T> {
} // namespace paddle

namespace ops = paddle::operators;
using CPUCtx = phi::CPUContext;

REGISTER_OPERATOR(bpr_loss,
ops::BprLossOp,
ops::BprLossOpMaker,
ops::BprLossGradMaker<paddle::framework::OpDesc>,
ops::BprLossGradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(bpr_loss_grad, ops::BprLossGradientOp);
REGISTER_OP_CPU_KERNEL(bpr_loss,
ops::BprLossOpKernel<CPUCtx, float>,
ops::BprLossOpKernel<CPUCtx, double>);
REGISTER_OP_CPU_KERNEL(bpr_loss_grad,
ops::BprLossGradientOpKernel<CPUCtx, float>,
ops::BprLossGradientOpKernel<CPUCtx, double>);

PD_REGISTER_STRUCT_KERNEL(
bpr_loss, CPU, ALL_LAYOUT, ops::BprLossOpKernel, float, double) {}
PD_REGISTER_STRUCT_KERNEL(bpr_loss_grad,
CPU,
ALL_LAYOUT,
ops::BprLossGradientOpKernel,
float,
double) {}
4 changes: 2 additions & 2 deletions paddle/fluid/operators/bpr_loss_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,7 +35,7 @@ struct TolerableValue {
}
};

template <typename DeviceContext, typename T>
template <typename T, typename DeviceContext>
class BprLossOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
Expand Down Expand Up @@ -83,7 +83,7 @@ class BprLossOpKernel : public framework::OpKernel<T> {
}
};

template <typename DeviceContext, typename T>
template <typename T, typename DeviceContext>
class BprLossGradientOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
Expand Down
21 changes: 12 additions & 9 deletions paddle/fluid/operators/collective/c_allgather_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,12 +73,15 @@ REGISTER_OP_WITHOUT_GRADIENT(c_allgather,
ops::CAllGatherOp,
ops::CAllGatherOpMaker);

REGISTER_OP_CPU_KERNEL(c_allgather,
ops::CAllGatherOpCPUKernel<float>,
ops::CAllGatherOpCPUKernel<double>,
ops::CAllGatherOpCPUKernel<int>,
ops::CAllGatherOpCPUKernel<int64_t>,
ops::CAllGatherOpCPUKernel<uint8_t>,
ops::CAllGatherOpCPUKernel<int8_t>,
ops::CAllGatherOpCPUKernel<bool>,
ops::CAllGatherOpCPUKernel<plat::float16>);
PD_REGISTER_STRUCT_KERNEL(c_allgather,
CPU,
ALL_LAYOUT,
ops::CAllGatherOpCPUKernel,
float,
double,
int,
int8_t,
int64_t,
uint8_t,
huangjiyi marked this conversation as resolved.
Show resolved Hide resolved
bool,
plat::float16) {}
26 changes: 15 additions & 11 deletions paddle/fluid/operators/collective/c_allgather_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ limitations under the License. */
namespace paddle {
namespace operators {

template <typename T>
template <typename T, typename DeviceContext>
class CAllGatherOpCUDAKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
Expand Down Expand Up @@ -93,15 +93,19 @@ class CAllGatherOpCUDAKernel : public framework::OpKernel<T> {
namespace ops = paddle::operators;
namespace plat = paddle::platform;

REGISTER_OP_CUDA_KERNEL(c_allgather,
ops::CAllGatherOpCUDAKernel<float>,
ops::CAllGatherOpCUDAKernel<double>,
PD_REGISTER_STRUCT_KERNEL(c_allgather,
GPU,
ALL_LAYOUT,
ops::CAllGatherOpCUDAKernel,
float,
double,
#if NCCL_VERSION_CODE >= 21000
ops::CAllGatherOpCUDAKernel<plat::bfloat16>,
plat::bfloat16,
#endif
ops::CAllGatherOpCUDAKernel<int>,
ops::CAllGatherOpCUDAKernel<uint8_t>,
ops::CAllGatherOpCUDAKernel<int8_t>,
ops::CAllGatherOpCUDAKernel<int64_t>,
ops::CAllGatherOpCUDAKernel<bool>,
ops::CAllGatherOpCUDAKernel<plat::float16>);
int,
uint8_t,
int8_t,
int64_t,
bool,
plat::float16) {
}
2 changes: 1 addition & 1 deletion paddle/fluid/operators/collective/c_allgather_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,7 @@ limitations under the License. */
namespace paddle {
namespace operators {

template <typename T>
template <typename T, typename DeviceContext>
class CAllGatherOpCPUKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
Expand Down
18 changes: 11 additions & 7 deletions paddle/fluid/operators/collective/c_allreduce_max_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,8 @@ class CAllReduceMaxOpMaker : public CAllReduceOpMaker {

DECLARE_INPLACE_OP_INFERER(AllreduceMaxInplaceInferer, {"X", "Out"});

DEFINE_C_ALLREDUCE_CPU_KERNEL(CAllReduceMax, kRedMax)

} // namespace operators
} // namespace paddle

Expand All @@ -45,10 +47,12 @@ REGISTER_OP_WITHOUT_GRADIENT(c_allreduce_max,
ops::CAllReduceOp,
ops::CAllReduceMaxOpMaker,
ops::AllreduceMaxInplaceInferer)

REGISTER_OP_CPU_KERNEL(c_allreduce_max,
ops::CAllReduceOpCPUKernel<ops::kRedMax, float>,
ops::CAllReduceOpCPUKernel<ops::kRedMax, double>,
ops::CAllReduceOpCPUKernel<ops::kRedMax, int>,
ops::CAllReduceOpCPUKernel<ops::kRedMax, int64_t>,
ops::CAllReduceOpCPUKernel<ops::kRedMax, plat::float16>);
PD_REGISTER_STRUCT_KERNEL(c_allreduce_max,
CPU,
ALL_LAYOUT,
ops::CAllReduceMaxCPUKernel,
float,
double,
int,
int64_t,
plat::float16) {}
22 changes: 15 additions & 7 deletions paddle/fluid/operators/collective/c_allreduce_max_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,13 +14,21 @@ limitations under the License. */

#include "paddle/fluid/operators/collective/c_allreduce_op.h"

namespace paddle {
namespace operators {
DEFINE_C_ALLREDUCE_CUDA_KERNEL(CAllReduceMax, kRedMax)
} // namespace operators
} // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;

REGISTER_OP_CUDA_KERNEL(
c_allreduce_max,
ops::CAllReduceOpCUDAKernel<ops::kRedMax, float>,
ops::CAllReduceOpCUDAKernel<ops::kRedMax, double>,
ops::CAllReduceOpCUDAKernel<ops::kRedMax, int>,
ops::CAllReduceOpCUDAKernel<ops::kRedMax, int64_t>,
ops::CAllReduceOpCUDAKernel<ops::kRedMax, plat::float16>)
PD_REGISTER_STRUCT_KERNEL(c_allreduce_max,
GPU,
ALL_LAYOUT,
ops::CAllReduceMaxCUDAKernel,
float,
double,
int,
int64_t,
plat::float16) {}
17 changes: 11 additions & 6 deletions paddle/fluid/operators/collective/c_allreduce_min_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,8 @@ class CAllReduceMinOpMaker : public CAllReduceOpMaker {

DECLARE_INPLACE_OP_INFERER(AllreduceMinInplaceInferer, {"X", "Out"});

DEFINE_C_ALLREDUCE_CPU_KERNEL(CAllReduceMin, kRedMin)

} // namespace operators
} // namespace paddle

Expand All @@ -46,9 +48,12 @@ REGISTER_OP_WITHOUT_GRADIENT(c_allreduce_min,
ops::CAllReduceMinOpMaker,
ops::AllreduceMinInplaceInferer)

REGISTER_OP_CPU_KERNEL(c_allreduce_min,
ops::CAllReduceOpCPUKernel<ops::kRedMin, float>,
ops::CAllReduceOpCPUKernel<ops::kRedMin, double>,
ops::CAllReduceOpCPUKernel<ops::kRedMin, int>,
ops::CAllReduceOpCPUKernel<ops::kRedMin, int64_t>,
ops::CAllReduceOpCPUKernel<ops::kRedMin, plat::float16>);
PD_REGISTER_STRUCT_KERNEL(c_allreduce_min,
CPU,
ALL_LAYOUT,
ops::CAllReduceMinCPUKernel,
float,
double,
int,
int64_t,
plat::float16) {}
22 changes: 15 additions & 7 deletions paddle/fluid/operators/collective/c_allreduce_min_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,13 +14,21 @@ limitations under the License. */

#include "paddle/fluid/operators/collective/c_allreduce_op.h"

namespace paddle {
namespace operators {
DEFINE_C_ALLREDUCE_CUDA_KERNEL(CAllReduceMin, kRedMin)
} // namespace operators
} // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;

REGISTER_OP_CUDA_KERNEL(
c_allreduce_min,
ops::CAllReduceOpCUDAKernel<ops::kRedMin, float>,
ops::CAllReduceOpCUDAKernel<ops::kRedMin, double>,
ops::CAllReduceOpCUDAKernel<ops::kRedMin, int>,
ops::CAllReduceOpCUDAKernel<ops::kRedMin, int64_t>,
ops::CAllReduceOpCUDAKernel<ops::kRedMin, plat::float16>)
PD_REGISTER_STRUCT_KERNEL(c_allreduce_min,
GPU,
ALL_LAYOUT,
ops::CAllReduceMinCUDAKernel,
float,
double,
int,
int64_t,
plat::float16) {}
8 changes: 8 additions & 0 deletions paddle/fluid/operators/collective/c_allreduce_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,6 +148,10 @@ class CAllReduceOpCPUKernel : public framework::OpKernel<T> {
}
};

#define DEFINE_C_ALLREDUCE_CPU_KERNEL(op_name, red_type) \
template <typename T, typename DeviceContext> \
class op_name##CPUKernel : public CAllReduceOpCPUKernel<red_type, T> {};

#if defined(PADDLE_WITH_ASCEND_CL)
// return true if found_nan or return false;
inline bool ContainsNan(const paddle::platform::NPUDeviceContext& dev_ctx,
Expand Down Expand Up @@ -527,6 +531,10 @@ class CAllReduceOpCUDAKernel : public framework::OpKernel<T> {
}
};

#define DEFINE_C_ALLREDUCE_CUDA_KERNEL(op_name, red_type) \
template <typename T, typename DeviceContext> \
class op_name##CUDAKernel : public CAllReduceOpCUDAKernel<red_type, T> {};

template <ReduceType red_type, typename T>
class CAllReduceOpMLUKernel : public framework::OpKernel<T> {
public:
Expand Down
17 changes: 11 additions & 6 deletions paddle/fluid/operators/collective/c_allreduce_prod_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,8 @@ class CAllReduceProdOpMaker : public CAllReduceOpMaker {

DECLARE_INPLACE_OP_INFERER(AllreduceProdInplaceInferer, {"X", "Out"});

DEFINE_C_ALLREDUCE_CPU_KERNEL(CAllReduceProd, kRedProd)

} // namespace operators
} // namespace paddle

Expand All @@ -46,9 +48,12 @@ REGISTER_OP_WITHOUT_GRADIENT(c_allreduce_prod,
ops::CAllReduceProdOpMaker,
ops::AllreduceProdInplaceInferer)

REGISTER_OP_CPU_KERNEL(c_allreduce_prod,
ops::CAllReduceOpCPUKernel<ops::kRedProd, float>,
ops::CAllReduceOpCPUKernel<ops::kRedProd, double>,
ops::CAllReduceOpCPUKernel<ops::kRedProd, int>,
ops::CAllReduceOpCPUKernel<ops::kRedProd, int64_t>,
ops::CAllReduceOpCPUKernel<ops::kRedProd, plat::float16>)
PD_REGISTER_STRUCT_KERNEL(c_allreduce_prod,
CPU,
ALL_LAYOUT,
ops::CAllReduceProdCPUKernel,
float,
double,
int,
int64_t,
plat::float16) {}
22 changes: 15 additions & 7 deletions paddle/fluid/operators/collective/c_allreduce_prod_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,13 +14,21 @@ limitations under the License. */

#include "paddle/fluid/operators/collective/c_allreduce_op.h"

namespace paddle {
namespace operators {
DEFINE_C_ALLREDUCE_CUDA_KERNEL(CAllReduceProd, kRedProd)
} // namespace operators
} // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;

REGISTER_OP_CUDA_KERNEL(
c_allreduce_prod,
ops::CAllReduceOpCUDAKernel<ops::kRedProd, float>,
ops::CAllReduceOpCUDAKernel<ops::kRedProd, double>,
ops::CAllReduceOpCUDAKernel<ops::kRedProd, int>,
ops::CAllReduceOpCUDAKernel<ops::kRedProd, int64_t>,
ops::CAllReduceOpCUDAKernel<ops::kRedProd, plat::float16>)
PD_REGISTER_STRUCT_KERNEL(c_allreduce_prod,
GPU,
ALL_LAYOUT,
ops::CAllReduceProdCUDAKernel,
float,
double,
int,
int64_t,
plat::float16) {}
17 changes: 11 additions & 6 deletions paddle/fluid/operators/collective/c_allreduce_sum_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,6 +56,8 @@ class CAllReduceSumOpMaker : public CAllReduceOpMaker {

DECLARE_INPLACE_OP_INFERER(AllreduceSumInplaceInferer, {"X", "Out"});

DEFINE_C_ALLREDUCE_CPU_KERNEL(CAllReduceSum, kRedSum)

} // namespace operators
} // namespace paddle

Expand All @@ -67,9 +69,12 @@ REGISTER_OP_WITHOUT_GRADIENT(c_allreduce_sum,
ops::CAllReduceSumOpMaker,
ops::AllreduceSumInplaceInferer)

REGISTER_OP_CPU_KERNEL(c_allreduce_sum,
ops::CAllReduceOpCPUKernel<ops::kRedSum, float>,
ops::CAllReduceOpCPUKernel<ops::kRedSum, double>,
ops::CAllReduceOpCPUKernel<ops::kRedSum, int>,
ops::CAllReduceOpCPUKernel<ops::kRedSum, int64_t>,
ops::CAllReduceOpCPUKernel<ops::kRedSum, plat::float16>)
PD_REGISTER_STRUCT_KERNEL(c_allreduce_sum,
CPU,
ALL_LAYOUT,
ops::CAllReduceSumCPUKernel,
float,
double,
int,
int64_t,
plat::float16) {}
25 changes: 17 additions & 8 deletions paddle/fluid/operators/collective/c_allreduce_sum_op.cu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,16 +14,25 @@ limitations under the License. */

#include "paddle/fluid/operators/collective/c_allreduce_op.h"

namespace paddle {
namespace operators {
DEFINE_C_ALLREDUCE_CUDA_KERNEL(CAllReduceSum, kRedSum)
} // namespace operators
} // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;

REGISTER_OP_CUDA_KERNEL(
c_allreduce_sum,
ops::CAllReduceOpCUDAKernel<ops::kRedSum, float>,
PD_REGISTER_STRUCT_KERNEL(c_allreduce_sum,
GPU,
ALL_LAYOUT,
ops::CAllReduceSumCUDAKernel,
float,
#if NCCL_VERSION_CODE >= 21000
ops::CAllReduceOpCUDAKernel<ops::kRedSum, plat::bfloat16>,
plat::bfloat16,
#endif
ops::CAllReduceOpCUDAKernel<ops::kRedSum, double>,
ops::CAllReduceOpCUDAKernel<ops::kRedSum, int>,
ops::CAllReduceOpCUDAKernel<ops::kRedSum, int64_t>,
ops::CAllReduceOpCUDAKernel<ops::kRedSum, plat::float16>)
double,
int,
int64_t,
plat::float16) {
}
12 changes: 8 additions & 4 deletions paddle/fluid/operators/detection/bipartite_match_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,7 +64,7 @@ bool DistPairDescend(std::tuple<int, int, T> pair1,
return std::get<2>(pair1) > std::get<2>(pair2);
}

template <typename T>
template <typename T, typename DeviceContext>
class BipartiteMatchKernel : public framework::OpKernel<T> {
public:
// The match_indices must be initialized to -1 at first.
Expand Down Expand Up @@ -318,6 +318,10 @@ REGISTER_OPERATOR(
ops::BipartiteMatchOpMaker,
paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,
paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>);
REGISTER_OP_CPU_KERNEL(bipartite_match,
ops::BipartiteMatchKernel<float>,
ops::BipartiteMatchKernel<double>);

PD_REGISTER_STRUCT_KERNEL(bipartite_match,
CPU,
ALL_LAYOUT,
ops::BipartiteMatchKernel,
float,
double) {}
Loading