Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
“dcfe198631058dbcd4fe6e887a4e514008ed1e68”上不存在“paddle/phi/kernels/reduce_sum_grad_kernel.h”
提交 a99fd15e 编写于 作者: J jim19930609
浏览文件
操作

Moved LoDTensor methods/members to Tensor

上级 19a833c8
隐藏空白更改
内联 并排
Showing with 259 addition and 254 deletion
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ a99fd15e
......@@ -151,7 +151,8 @@ bool CheckLoD(const LoD &in, int tensor_height) {
}
// check: the lowest level's last offset should equals `tensor_height` if
// tensor_height>0.
if (tensor_height > 0 && (size_t)tensor_height != in.back().back())
if (tensor_height > 0 &&
static_cast<size_t>(tensor_height) != in.back().back())
return false;
// check: the higher level's last offset should equals the lower level's
......@@ -184,42 +185,13 @@ bool CheckAbsLoD(const LoD &in, int tensor_height) {
if (level.front() != 0) return false;
if (tensor_height < 0) {
tensor_height = level.back();
} else if ((size_t)tensor_height != level.back()) {
} else if (static_cast<size_t>(tensor_height) != level.back()) {
return false;
}
}
return true;
}
using LoDAndOffset = std::pair<LoD, std::pair<size_t, size_t>>;
LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD &lod, size_t start_idx,
size_t end_idx, size_t start_level) {
LoD sub_lod;
for (size_t level_idx = start_level; level_idx < lod.size(); ++level_idx) {
PADDLE_ENFORCE_LE(start_idx, end_idx,
platform::errors::InvalidArgument(
"The start index should be less than the end index, "
"but received start index is %d, end index is %d.",
start_idx, end_idx));
PADDLE_ENFORCE_LT(
end_idx, lod[level_idx].size(),
platform::errors::InvalidArgument(
"The end index should be less than the LoD level size, but "
"received end index is %d, LoD level size is %d.",
end_idx, lod[level_idx].size()));
std::vector<size_t> level_lens;
for (size_t i = start_idx; i < end_idx; ++i) {
level_lens.push_back(lod[level_idx][i + 1] - lod[level_idx][i]);
}
sub_lod.emplace_back(level_lens);
start_idx = lod[level_idx][start_idx];
end_idx = lod[level_idx][end_idx];
}
return LoDAndOffset{sub_lod, {start_idx, end_idx}};
}
void AppendLoD(LoD *lod, const LoD &lod_length) {
PADDLE_ENFORCE(
lod->empty() || lod->size() == lod_length.size(),
......@@ -347,153 +319,6 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
}
std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
const std::vector<platform::Place> places) const {
PADDLE_ENFORCE_GT(places.size(), 0,
platform::errors::InvalidArgument(
"Place number cannot be empty when splitting."));
check_memory_size();
size_t batch_size =
lod().empty() ? static_cast<size_t>(dims()[0]) : lod()[0].size() - 1;
// if batch_size is 0, just return #places.size() copys of empty
// tensors.
if (batch_size == 0) {
std::vector<LoDTensor> empty_results;
empty_results.reserve(places.size());
for (size_t i = 0; i < places.size(); ++i) {
LoDTensor dst;
dst.Resize(dims());
dst.mutable_data(places[i], type());
if (!lod().empty()) {
dst.set_lod(lod());
}
empty_results.emplace_back(std::move(dst));
}
return empty_results;
}
auto step_width = (batch_size + places.size() - 1) / places.size();
auto result_size = (batch_size + step_width - 1) / step_width;
std::vector<LoDTensor> results;
results.reserve(result_size);
for (size_t i = 0; i < result_size; ++i) {
auto begin = i * step_width;
auto end = std::min<size_t>((i + 1) * step_width, batch_size);
PADDLE_ENFORCE_LT(begin, end,
platform::errors::InvalidArgument(
"The begin index must be less than the end index, "
"but received begin index is %d, end index is %d.",
begin, end));
LoDTensor dst;
if (lod().empty()) {
auto src = Slice(begin, end);
auto &dst_place = places[i];
framework::TensorCopy(src, dst_place, &dst);
} else {
auto lod_and_offset = GetSubLoDAndAbsoluteOffset(lod(), begin, end, 0);
auto &offset = lod_and_offset.second;
auto src = Slice(offset.first, offset.second);
auto &dst_place = places[i];
framework::TensorCopy(src, dst_place, &dst);
LoD my_lod;
for (auto &l : lod_and_offset.first) {
std::vector<size_t> v{0};
for (auto &ll : l) {
v.push_back(ll + v.back());
}
my_lod.emplace_back(v);
}
dst.set_lod(my_lod);
}
results.emplace_back(std::move(dst));
}
return results;
}
void LoDTensor::MergeLoDTensor(
const std::vector<const LoDTensor *> &lod_tensors,
platform::Place dst_place) {
PADDLE_ENFORCE_EQ(lod_tensors.empty(), false,
platform::errors::InvalidArgument(
"The LoDTensors to be merged are empty."));
framework::DDim new_dim = lod_tensors[0]->dims();
proto::VarType::Type new_type = proto::VarType::FP32;
framework::DataLayout new_layout = lod_tensors[0]->layout();
for (auto *t : lod_tensors) {
if (t->numel() && t->IsInitialized()) {
new_dim = t->dims();
new_type = t->type();
new_layout = t->layout();
break;
}
}
LoD new_lod = lod_tensors[0]->lod();
for (size_t i = 1; i < lod_tensors.size(); ++i) {
auto *t = lod_tensors[i];
if (t->numel() && t->IsInitialized()) {
PADDLE_ENFORCE_EQ(
new_type, t->type(),
platform::errors::InvalidArgument(
"LoDTensor data type does not match, expected type is %s, actual "
"type is %s.",
DataTypeToString(new_type), DataTypeToString(t->type())));
PADDLE_ENFORCE_EQ(
new_layout, t->layout(),
platform::errors::InvalidArgument(
"LoDTensor layout does not match, expected layout is %s, "
"actual layout is %s.",
DataLayoutToString(new_layout), DataLayoutToString(t->layout())));
PADDLE_ENFORCE_EQ(
framework::product(new_dim) / new_dim[0],
framework::product(t->dims()) / t->dims()[0],
platform::errors::InvalidArgument(
"LoDTensor dimension does not match, all dimensions except the "
"first dimension need to be equal,"
"but expected dimension is %s, actual dimension is %s.",
new_dim, t->dims()));
new_dim[0] += t->dims()[0];
}
auto &lod = t->lod();
PADDLE_ENFORCE_EQ(new_lod.size(), lod.size(),
platform::errors::InvalidArgument(
"The LoD information of LoDTensor does not match, "
"expected LoD is %s, actual LoD is %s.",
new_lod, lod));
for (size_t j = 0; j < lod.size(); ++j) {
auto &sub_lod = new_lod[j];
size_t offset = sub_lod.back();
for (size_t k = 1; k < lod[j].size(); ++k) {
sub_lod.push_back(lod[j][k] + offset);
}
}
}
Resize(new_dim);
set_layout(new_layout);
set_lod(new_lod);
mutable_data(dst_place, new_type);
int begin = 0;
for (auto *src : lod_tensors) {
int end = begin + src->dims()[0];
if (end == begin) {
continue;
}
auto dst = Slice(begin, end);
framework::TensorCopy(*src, dst_place, &dst);
begin = end;
}
}
LoD ConvertToLengthBasedLoD(const LoD &offset_lod) {
LoD length_lod;
length_lod.reserve(offset_lod.size());
......
paddle/fluid/framework/lod_tensor.h
浏览文件 @ a99fd15e
......@@ -39,6 +39,8 @@ class DeviceContext;
namespace paddle {
namespace framework {
using LoD = std::vector<Vector<size_t>>;
/*
* LoD is short for Level of Details.
*
......@@ -54,7 +56,6 @@ namespace framework {
* 0 2 4 7
* 0 2 5 7 10 12 15 20
*/
using LoD = std::vector<Vector<size_t>>;
std::ostream& operator<<(std::ostream& os, const LoD& lod);
std::ostream& operator<<(std::ostream& os, const LoDTensor& t);
......@@ -108,64 +109,14 @@ bool CheckAbsLoD(const LoD& in, int tensor_height = -1);
*/
class LoDTensor : public Tensor {
public:
LoDTensor() : Tensor() {}
explicit LoDTensor(const LoD& lod) : lod_(lod) {}
void set_lod(const LoD& lod) { lod_ = lod; }
const LoD& lod() const { return lod_; }
LoD* mutable_lod() { return &lod_; }
/*
* Get the start offset and end offset of an element from LoD.
*/
std::pair<size_t, size_t> lod_element(size_t level, size_t elem) const {
PADDLE_ENFORCE_LT(
level, NumLevels(),
platform::errors::InvalidArgument(
"The input level of LoD is invalid, it should be less than LoD "
"size. The input level is %zu, the LoD size is %zu.",
level, NumLevels()));
PADDLE_ENFORCE_LT(elem, NumElements(level),
platform::errors::InvalidArgument(
"The input element of LoD is invalid, it should be "
"less than the number of elements in its level."
"The input element is %zu, the number of elements in "
"its level is %zu.",
elem, NumElements(level)));
return std::make_pair((lod_)[level][elem], (lod_)[level][elem + 1]);
}
/*
* Number of LoDTensor's levels, each level has units of data, for example,
* in the sentence's view, article, paragraph, sentence are 3 levels.
*/
size_t NumLevels() const { return lod_.size(); }
/*
* Number of elements in a level.
*/
size_t NumElements(size_t level = 0) const {
PADDLE_ENFORCE_LT(
level, NumLevels(),
platform::errors::InvalidArgument(
"The input level of LoD is invalid, it should be less than LoD "
"size. The input level is %zu, the LoD size is %zu.",
level, NumLevels()));
// the last offset is the end of last element
return (lod_)[level].size() - 1;
}
// Split LoDTensor and copy to each place specified in places.
std::vector<LoDTensor> SplitLoDTensor(
const std::vector<platform::Place> places) const;
void MergeLoDTensor(const std::vector<const LoDTensor*>& lod_tensors,
platform::Place place);
private:
LoD lod_;
platform::Place place) {
std::vector<const Tensor*> tmp;
for (const LoDTensor* lod_tensor : lod_tensors) {
tmp.push_back(lod_tensor);
}
Tensor::MergeLoDTensor(tmp, place);
}
};
/*
......@@ -210,21 +161,6 @@ LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
return tensor;
}
// Get the absolute offset of a lod[start_level][start_idx:end_idx] and
// relative length of details for every levels(i.e., [start_level: ]).
//
// For example,
// lod = [[0, 3, 4, 8], [0, 9, 10, 11, 13, 17, 19, 22, 24]]
// start_level = 0
// start_idx = 1
// end_idx = 3
//
// Returns:
// LoD = [[1, 4], [2, 4, 2, 3, 2]]
// pair<size_t, size_t> = {11, 24}
std::pair<LoD, std::pair<size_t, size_t>> GetSubLoDAndAbsoluteOffset(
const LoD& lod, size_t start_idx, size_t end_idx, size_t start_level);
void AppendLoD(LoD* lod, const LoD& lod_length);
/*
......
paddle/fluid/framework/mixed_vector.h
浏览文件 @ a99fd15e
......@@ -23,10 +23,9 @@ limitations under the License. */
#include "glog/logging.h"
#include "paddle/fluid/framework/details/cow_ptr.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/memory/malloc.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/utils/none.h"
#include "paddle/utils/optional.h"
......
paddle/fluid/framework/tensor.cc
浏览文件 @ a99fd15e
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/tensor_util.h"
DECLARE_bool(use_stream_safe_cuda_allocator);
......@@ -235,5 +236,181 @@ void Tensor::ResetHolderWithType(std::shared_ptr<memory::Allocation> holder,
void Tensor::set_type(const proto::VarType::Type& type) { type_ = type; }
/* ---------------------------------------- */
/* -------------- LoDTensor --------------- */
/* ---------------------------------------- */
using LoDAndOffset = std::pair<LoD, std::pair<size_t, size_t>>;
LoDAndOffset GetSubLoDAndAbsoluteOffset(const LoD& lod, size_t start_idx,
size_t end_idx, size_t start_level) {
LoD sub_lod;
for (size_t level_idx = start_level; level_idx < lod.size(); ++level_idx) {
PADDLE_ENFORCE_LE(start_idx, end_idx,
platform::errors::InvalidArgument(
"The start index should be less than the end index, "
"but received start index is %d, end index is %d.",
start_idx, end_idx));
PADDLE_ENFORCE_LT(
end_idx, lod[level_idx].size(),
platform::errors::InvalidArgument(
"The end index should be less than the LoD level size, but "
"received end index is %d, LoD level size is %d.",
end_idx, lod[level_idx].size()));
std::vector<size_t> level_lens;
for (size_t i = start_idx; i < end_idx; ++i) {
level_lens.push_back(lod[level_idx][i + 1] - lod[level_idx][i]);
}
sub_lod.emplace_back(level_lens);
start_idx = lod[level_idx][start_idx];
end_idx = lod[level_idx][end_idx];
}
return LoDAndOffset{sub_lod, {start_idx, end_idx}};
}
std::vector<Tensor> Tensor::SplitLoDTensor(
const std::vector<platform::Place> places) const {
PADDLE_ENFORCE_GT(places.size(), 0,
platform::errors::InvalidArgument(
"Place number cannot be empty when splitting."));
check_memory_size();
size_t batch_size =
lod().empty() ? static_cast<size_t>(dims()[0]) : lod()[0].size() - 1;
// if batch_size is 0, just return #places.size() copys of empty
// tensors.
if (batch_size == 0) {
std::vector<Tensor> empty_results;
empty_results.reserve(places.size());
for (size_t i = 0; i < places.size(); ++i) {
Tensor dst;
dst.Resize(dims());
dst.mutable_data(places[i], type());
if (!lod().empty()) {
dst.set_lod(lod());
}
empty_results.emplace_back(std::move(dst));
}
return empty_results;
}
auto step_width = (batch_size + places.size() - 1) / places.size();
auto result_size = (batch_size + step_width - 1) / step_width;
std::vector<Tensor> results;
results.reserve(result_size);
for (size_t i = 0; i < result_size; ++i) {
auto begin = i * step_width;
auto end = std::min<size_t>((i + 1) * step_width, batch_size);
PADDLE_ENFORCE_LT(begin, end,
platform::errors::InvalidArgument(
"The begin index must be less than the end index, "
"but received begin index is %d, end index is %d.",
begin, end));
Tensor dst;
if (lod().empty()) {
auto src = Slice(begin, end);
auto& dst_place = places[i];
framework::TensorCopy(src, dst_place, &dst);
} else {
auto lod_and_offset = GetSubLoDAndAbsoluteOffset(lod(), begin, end, 0);
auto& offset = lod_and_offset.second;
auto src = Slice(offset.first, offset.second);
auto& dst_place = places[i];
framework::TensorCopy(src, dst_place, &dst);
LoD my_lod;
for (auto& l : lod_and_offset.first) {
std::vector<size_t> v{0};
for (auto& ll : l) {
v.push_back(ll + v.back());
}
my_lod.emplace_back(v);
}
dst.set_lod(my_lod);
}
results.emplace_back(std::move(dst));
}
return results;
}
void Tensor::MergeLoDTensor(const std::vector<const Tensor*>& lod_tensors,
platform::Place dst_place) {
PADDLE_ENFORCE_EQ(lod_tensors.empty(), false,
platform::errors::InvalidArgument(
"The LoDTensors to be merged are empty."));
framework::DDim new_dim = lod_tensors[0]->dims();
proto::VarType::Type new_type = proto::VarType::FP32;
framework::DataLayout new_layout = lod_tensors[0]->layout();
for (auto* t : lod_tensors) {
if (t->numel() && t->IsInitialized()) {
new_dim = t->dims();
new_type = t->type();
new_layout = t->layout();
break;
}
}
LoD new_lod = lod_tensors[0]->lod();
for (size_t i = 1; i < lod_tensors.size(); ++i) {
auto* t = lod_tensors[i];
if (t->numel() && t->IsInitialized()) {
PADDLE_ENFORCE_EQ(
new_type, t->type(),
platform::errors::InvalidArgument(
"LoDTensor data type does not match, expected type is %s, actual "
"type is %s.",
DataTypeToString(new_type), DataTypeToString(t->type())));
PADDLE_ENFORCE_EQ(
new_layout, t->layout(),
platform::errors::InvalidArgument(
"LoDTensor layout does not match, expected layout is %s, "
"actual layout is %s.",
DataLayoutToString(new_layout), DataLayoutToString(t->layout())));
PADDLE_ENFORCE_EQ(
framework::product(new_dim) / new_dim[0],
framework::product(t->dims()) / t->dims()[0],
platform::errors::InvalidArgument(
"LoDTensor dimension does not match, all dimensions except the "
"first dimension need to be equal,"
"but expected dimension is %s, actual dimension is %s.",
new_dim, t->dims()));
new_dim[0] += t->dims()[0];
}
auto& lod = t->lod();
PADDLE_ENFORCE_EQ(new_lod.size(), lod.size(),
platform::errors::InvalidArgument(
"The LoD information of LoDTensor does not match"));
for (size_t j = 0; j < lod.size(); ++j) {
auto& sub_lod = new_lod[j];
size_t offset = sub_lod.back();
for (size_t k = 1; k < lod[j].size(); ++k) {
sub_lod.push_back(lod[j][k] + offset);
}
}
}
Resize(new_dim);
set_layout(new_layout);
set_lod(new_lod);
mutable_data(dst_place, new_type);
int begin = 0;
for (auto* src : lod_tensors) {
int end = begin + src->dims()[0];
if (end == begin) {
continue;
}
auto dst = Slice(begin, end);
framework::TensorCopy(*src, dst_place, &dst);
begin = end;
}
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/tensor.h
浏览文件 @ a99fd15e
......@@ -24,6 +24,7 @@ limitations under the License. */
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/mixed_vector.h"
#include "paddle/fluid/memory/memory.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
......@@ -41,7 +42,7 @@ namespace paddle {
namespace framework {
class LoDTensor;
using LoD = std::vector<Vector<size_t>>;
/*
NOTE(liym27): [ What is TensorInplaceVersion used for? ]
......@@ -326,8 +327,75 @@ class Tensor {
*/
size_t offset_;
std::shared_ptr<TensorInplaceVersion> inplace_version_counter_;
/* ---------------------------------------------------------- */
/* --------------- Reserved for LoDTensor ------------------- */
/* ---------------------------------------------------------- */
public:
explicit Tensor(const LoD& lod) : lod_(lod) {}
void set_lod(const LoD& lod) { lod_ = lod; }
const LoD& lod() const { return lod_; }
LoD* mutable_lod() { return &lod_; }
std::pair<size_t, size_t> lod_element(size_t level, size_t elem) const {
PADDLE_ENFORCE_LT(
level, NumLevels(),
platform::errors::InvalidArgument(
"The input level of LoD is invalid, it should be less than LoD "
"size. The input level is %zu, the LoD size is %zu.",
level, NumLevels()));
PADDLE_ENFORCE_LT(elem, NumElements(level),
platform::errors::InvalidArgument(
"The input element of LoD is invalid, it should be "
"less than the number of elements in its level."
"The input element is %zu, the number of elements in "
"its level is %zu.",
elem, NumElements(level)));
return std::make_pair((lod_)[level][elem], (lod_)[level][elem + 1]);
}
size_t NumLevels() const { return lod_.size(); }
size_t NumElements(size_t level = 0) const {
PADDLE_ENFORCE_LT(
level, NumLevels(),
platform::errors::InvalidArgument(
"The input level of LoD is invalid, it should be less than LoD "
"size. The input level is %zu, the LoD size is %zu.",
level, NumLevels()));
// the last offset is the end of last element
return (lod_)[level].size() - 1;
}
// Split LoDTensor and copy to each place specified in places.
std::vector<Tensor> SplitLoDTensor(
const std::vector<platform::Place> places) const;
void MergeLoDTensor(const std::vector<const Tensor*>& lod_tensors,
platform::Place place);
private:
LoD lod_;
};
// Get the absolute offset of a lod[start_level][start_idx:end_idx] and
// relative length of details for every levels(i.e., [start_level: ]).
//
// For example,
// lod = [[0, 3, 4, 8], [0, 9, 10, 11, 13, 17, 19, 22, 24]]
// start_level = 0
// start_idx = 1
// end_idx = 3
//
// Returns:
// LoD = [[1, 4], [2, 4, 2, 3, 2]]
// pair<size_t, size_t> = {11, 24}
std::pair<LoD, std::pair<size_t, size_t>> GetSubLoDAndAbsoluteOffset(
const LoD& lod, size_t start_idx, size_t end_idx, size_t start_level);
} // namespace framework
} // namespace paddle
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或