mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 21:14:14 +08:00
1433160a36
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/66742 Modified loops in files under fbsource/fbcode/caffe2/ from the format `for(TYPE var=x0;var<x_max;x++)` to the format `for(const auto var: irange(xmax))` This was achieved by running r-barnes's loop upgrader script (D28874212) with some modification to exclude all files under /torch/jit and a number of reversions or unused variable suppression warnings added by hand. Test Plan: Sandcastle Reviewed By: malfet Differential Revision: D31705366 fbshipit-source-id: be58222426c192406a7f93c21582c3f6f2082401
126 lines
3.8 KiB
C++
126 lines
3.8 KiB
C++
#ifndef CAFFE2_OPERATORS_EXPAND_OP_H_
|
|
#define CAFFE2_OPERATORS_EXPAND_OP_H_
|
|
|
|
#include <vector>
|
|
|
|
#include "caffe2/core/context.h"
|
|
#include "caffe2/core/operator.h"
|
|
#include "caffe2/core/types.h"
|
|
#include "caffe2/utils/math.h"
|
|
#include "c10/util/irange.h"
|
|
|
|
namespace caffe2 {
|
|
|
|
template <typename InputTypes, class Context>
|
|
class ExpandOp final : public Operator<Context> {
|
|
public:
|
|
USE_OPERATOR_CONTEXT_FUNCTIONS;
|
|
|
|
template <class... Args>
|
|
explicit ExpandOp(Args&&... args)
|
|
: Operator<Context>(std::forward<Args>(args)...),
|
|
OP_SINGLE_ARG(bool, "allow_broadcast_fastpath", allow_broadcast_fastpath_, false) {}
|
|
|
|
bool RunOnDevice() override {
|
|
return DispatchHelper<InputTypes>::call(this, Input(0));
|
|
}
|
|
template <typename T>
|
|
bool DoRunWithType() {
|
|
const auto& X = Input(0);
|
|
const auto& Y_shape_tensor = Input(1);
|
|
std::vector<int64_t> shape_dims(Y_shape_tensor.numel());
|
|
context_.template CopyToCPU<int64_t>(
|
|
Y_shape_tensor.numel(),
|
|
Y_shape_tensor.template data<int64_t>(),
|
|
shape_dims.data());
|
|
|
|
const int ndim = shape_dims.size();
|
|
const std::vector<int> X_dims(X.sizes().cbegin(), X.sizes().cend());
|
|
std::vector<int> Y_dims;
|
|
Y_dims.reserve(std::max(ndim, X.dim()));
|
|
// ndim, X.ndim() might equal to 0
|
|
for (int i = ndim - 1, j = X.dim() - 1; i >= 0 || j >= 0; --i, --j) {
|
|
const int shape_x = (j >= 0 ? X_dims[j] : 1);
|
|
// In PyTorch expand treats -1 as a special value to indicate
|
|
// preserving the size of that dimension.
|
|
const int shape_y = ((i >= 0 && shape_dims[i] > 0) ? shape_dims[i] : 1);
|
|
|
|
CAFFE_ENFORCE(
|
|
shape_x == 1 || shape_y == 1 || shape_x == shape_y,
|
|
"Dimensions format invalid.");
|
|
Y_dims.push_back(std::max(shape_x, shape_y));
|
|
}
|
|
std::reverse(Y_dims.begin(), Y_dims.end());
|
|
// TODO: remove when the function in math are changed to use vector<int64_t>
|
|
std::vector<int64_t> Y_dims_int64;
|
|
std::copy(Y_dims.begin(), Y_dims.end(), std::back_inserter(Y_dims_int64));
|
|
auto* Y = Output(0, Y_dims_int64, at::dtype<T>());
|
|
math::Broadcast<T, Context>(
|
|
X_dims.size(),
|
|
X_dims.data(),
|
|
Y_dims.size(),
|
|
Y_dims.data(),
|
|
T(1),
|
|
X.template data<T>(),
|
|
Y->template mutable_data<T>(),
|
|
&context_,
|
|
allow_broadcast_fastpath_);
|
|
return true;
|
|
}
|
|
|
|
const bool allow_broadcast_fastpath_;
|
|
};
|
|
|
|
template <typename InputTypes, class Context>
|
|
class ExpandGradientOp final : public Operator<Context> {
|
|
public:
|
|
USE_OPERATOR_CONTEXT_FUNCTIONS;
|
|
|
|
template <class... Args>
|
|
explicit ExpandGradientOp(Args&&... args)
|
|
: Operator<Context>(std::forward<Args>(args)...),
|
|
OP_SINGLE_ARG(bool, "allow_broadcast_fastpath", allow_broadcast_fastpath_, false) {}
|
|
|
|
bool RunOnDevice() override {
|
|
return DispatchHelper<InputTypes>::call(this, Input(0));
|
|
}
|
|
|
|
template <typename T>
|
|
bool DoRunWithType() {
|
|
const auto& dY = Input(0);
|
|
const auto& X = Input(1);
|
|
|
|
const int ndim = dY.dim();
|
|
const std::vector<int> dX_dims(X.sizes().cbegin(), X.sizes().cend());
|
|
const std::vector<int> dY_dims(dY.sizes().cbegin(), dY.sizes().cend());
|
|
auto* dX = Output(0, X.sizes(), at::dtype<T>());
|
|
std::vector<int> axes;
|
|
const int offset = ndim - X.dim();
|
|
for (const auto i : c10::irange(ndim)) {
|
|
if (i < offset || dX_dims[i - offset] == 1) {
|
|
axes.push_back(i);
|
|
}
|
|
}
|
|
std::vector<int> X_dims = dY_dims;
|
|
for (const int axis : axes) {
|
|
X_dims[axis] = 1;
|
|
}
|
|
math::ReduceSum<T, Context>(
|
|
dY_dims.size(),
|
|
dY_dims.data(),
|
|
X_dims.data(),
|
|
T(1),
|
|
dY.template data<T>(),
|
|
dX->template mutable_data<T>(),
|
|
&context_,
|
|
allow_broadcast_fastpath_);
|
|
return true;
|
|
}
|
|
|
|
const bool allow_broadcast_fastpath_;
|
|
};
|
|
|
|
} // namespace caffe2
|
|
|
|
#endif // CAFFE2_OPERATORS_REDUCE_OPS_H_
|