mirror of
https://github.com/saymrwulf/pytorch.git
synced 2026-05-15 21:00:47 +00:00
e8c274cf16
Summary: Use ElementwiseLinearOps instead of manual Mul + Sum. That saves intermediate blobs. For NMT use case Before: https://our.intern.facebook.com/intern/fblearner/details/18060753 Time per step: 0.072 memory usage (per each of 2 GPUs): 9041MiB After:https://our.intern.facebook.com/intern/fblearner/details/18107583 Time per step: 0.0715 Memory (per each GPU): 8560MiB Reviewed By: akyrola Differential Revision: D5038785 fbshipit-source-id: 4bc8155dbd0c87729e17236d68d62ca530aadb53
39 lines
1.1 KiB
C++
39 lines
1.1 KiB
C++
#ifndef CAFFE2_OPERATORS_ELEMENTWISE_LINEAR_OP_H_
|
|
#define CAFFE2_OPERATORS_ELEMENTWISE_LINEAR_OP_H_
|
|
|
|
#include "caffe2/core/context.h"
|
|
#include "caffe2/core/operator.h"
|
|
#include "caffe2/utils/math.h"
|
|
|
|
namespace caffe2 {
|
|
template <typename T, class Context, class Engine = DefaultEngine>
|
|
class ElementwiseLinearOp final : public Operator<Context> {
|
|
public:
|
|
ElementwiseLinearOp(const OperatorDef& operator_def, Workspace* ws)
|
|
: Operator<Context>(operator_def, ws),
|
|
axis_(OperatorBase::GetSingleArgument<int>("axis", 1)) {}
|
|
|
|
USE_OPERATOR_CONTEXT_FUNCTIONS;
|
|
bool RunOnDevice() override;
|
|
|
|
protected:
|
|
int axis_;
|
|
};
|
|
|
|
template <typename T, class Context, class Engine = DefaultEngine>
|
|
class ElementwiseLinearGradientOp final : public Operator<Context> {
|
|
public:
|
|
ElementwiseLinearGradientOp(const OperatorDef& operator_def, Workspace* ws)
|
|
: Operator<Context>(operator_def, ws),
|
|
axis_(OperatorBase::GetSingleArgument<int>("axis", 1)) {}
|
|
|
|
USE_OPERATOR_CONTEXT_FUNCTIONS;
|
|
bool RunOnDevice() override;
|
|
|
|
protected:
|
|
int axis_;
|
|
};
|
|
|
|
} // namespace caffe2
|
|
|
|
#endif // CAFFE2_OPERATORS_ELEMENTWISE_LINEAR_OP_H_
|