Files
2026-07-13 12:40:42 +08:00

920 lines
33 KiB
C++

/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/phi/api/lib/api_gen_utils.h"
#include "paddle/common/flags.h"
#include "paddle/phi/backends/gpu/gpu_info.h"
#include "paddle/phi/core/memory/malloc.h"
#include "paddle/phi/core/memory/mem_utils.h"
#include "paddle/phi/core/memory/stats.h"
#include "paddle/phi/core/visit_type.h"
#include "paddle/phi/kernels/strided_copy_kernel.h"
PHI_DECLARE_bool(use_stride_kernel);
COMMON_DECLARE_bool(enable_compact_mem);
COMMON_DECLARE_int64(max_reserved_threshold_in_gb);
COMMON_DECLARE_int64(cur_allocated_threshold_in_gb);
COMMON_DECLARE_bool(try_allocate);
#include "glog/logging.h"
#include "paddle/phi/core/distributed/auto_parallel/dist_attr.h"
#include "paddle/phi/core/distributed/auto_parallel/dist_meta_tensor.h"
#include "paddle/phi/core/distributed/auto_parallel/dist_tensor.h"
#include "paddle/phi/core/kernel_factory.h"
namespace paddle::experimental {
/* ------------------ for input ----------------------- */
std::shared_ptr<phi::DenseTensor> TensorToDenseTensor(const Tensor& tensor) {
return std::static_pointer_cast<phi::DenseTensor>(tensor.impl());
}
paddle::optional<phi::DenseTensor> TensorToDenseTensor(
const paddle::optional<Tensor>& tensor) {
if (tensor) {
return {*std::static_pointer_cast<phi::DenseTensor>(tensor->impl())};
}
return nullptr;
}
std::unique_ptr<std::vector<phi::DenseTensor*>> TensorToDenseTensor(
const std::vector<Tensor>& tensors) {
auto pt_tensors = std::make_unique<std::vector<phi::DenseTensor*>>();
pt_tensors->reserve(tensors.size());
for (const auto& t : tensors) {
pt_tensors->push_back(
std::dynamic_pointer_cast<phi::DenseTensor>(t.impl()).get());
}
return pt_tensors;
}
std::vector<const phi::DenseTensor*> TensorToConstDenseTensorPtr(
const std::vector<Tensor>& tensors) {
std::vector<const phi::DenseTensor*> pt_tensors(tensors.size());
for (size_t i = 0; i < tensors.size(); ++i) {
pt_tensors[i] = static_cast<phi::DenseTensor*>(tensors[i].impl().get());
}
return pt_tensors;
}
paddle::optional<std::vector<const phi::DenseTensor*>>
TensorToConstDenseTensorPtr(
const paddle::optional<std::vector<Tensor>>& tensors) {
paddle::optional<std::vector<const phi::DenseTensor*>> pt_tensors;
if (tensors) {
pt_tensors =
paddle::optional<std::vector<const phi::DenseTensor*>>(tensors->size());
for (size_t i = 0; i < tensors->size(); ++i) {
pt_tensors->at(i) =
static_cast<phi::DenseTensor*>(tensors->at(i).impl().get());
}
}
return pt_tensors;
}
std::shared_ptr<phi::SelectedRows> TensorToSelectedRows(const Tensor& tensor) {
return std::static_pointer_cast<phi::SelectedRows>(tensor.impl());
}
paddle::optional<phi::SelectedRows> TensorToSelectedRows(
const paddle::optional<Tensor>& tensor) {
if (tensor) {
return {*std::static_pointer_cast<phi::SelectedRows>(tensor->impl())};
}
return nullptr;
}
std::shared_ptr<phi::StringTensor> TensorToStringTensor(const Tensor& tensor) {
return std::dynamic_pointer_cast<phi::StringTensor>(tensor.impl());
}
std::shared_ptr<phi::SparseCooTensor> TensorToSparseCooTensor(
const Tensor& tensor) {
return std::static_pointer_cast<phi::SparseCooTensor>(tensor.impl());
}
/* ----------------- for infer_meta --------------------- */
phi::MetaTensor MakeMetaTensor(const phi::TensorBase& tensor) {
return phi::MetaTensor(tensor);
}
std::vector<phi::MetaTensor> MakeMetaTensor(
const std::vector<const phi::TensorBase*>& tensors) {
std::vector<phi::MetaTensor> meta_tensors;
meta_tensors.reserve(tensors.size());
for (const auto* t : tensors) {
meta_tensors.emplace_back(*t);
}
return meta_tensors;
}
phi::MetaTensor MakeMetaTensor(
const paddle::optional<phi::DenseTensor>& tensor) {
if (tensor) {
return {phi::MetaTensor(*tensor)};
}
return phi::MetaTensor();
}
std::vector<phi::MetaTensor> MakeMetaTensor(
const std::vector<const phi::DenseTensor*>& tensors) {
std::vector<phi::MetaTensor> meta_tensors;
meta_tensors.reserve(tensors.size());
for (const auto* t : tensors) {
meta_tensors.emplace_back(*t);
}
return meta_tensors;
}
std::vector<phi::MetaTensor> MakeMetaTensor(
const std::vector<const phi::SelectedRows*>& tensors) {
std::vector<phi::MetaTensor> meta_tensors;
meta_tensors.reserve(tensors.size());
for (const auto* t : tensors) {
meta_tensors.emplace_back(*t);
}
return meta_tensors;
}
std::vector<phi::MetaTensor> MakeMetaTensor(
const std::vector<phi::DenseTensor*>& tensors) {
std::vector<phi::MetaTensor> meta_tensors;
meta_tensors.reserve(tensors.size());
for (auto* t : tensors) {
meta_tensors.emplace_back(*t);
}
return meta_tensors;
}
phi::MetaTensor MakeMetaTensor(
const paddle::optional<phi::SelectedRows>& tensor) {
if (tensor) {
return {phi::MetaTensor(*tensor)};
}
return phi::MetaTensor();
}
phi::MetaTensor MakeMetaTensor(
const paddle::optional<phi::SparseCooTensor>& tensor) {
if (tensor) {
return {phi::MetaTensor(*tensor)};
}
return phi::MetaTensor();
}
phi::MetaTensor MakeMetaTensor(
const paddle::optional<phi::SparseCsrTensor>& tensor) {
if (tensor) {
return {phi::MetaTensor(*tensor)};
}
return phi::MetaTensor();
}
std::vector<phi::MetaTensor> MakeMetaTensor(
const paddle::optional<std::vector<const phi::DenseTensor*>>& tensors) {
std::vector<phi::MetaTensor> meta_tensors;
if (tensors) {
meta_tensors.reserve(tensors->size());
for (auto* t : tensors.get()) {
meta_tensors.emplace_back(*t);
}
}
return meta_tensors;
}
phi::DenseTensor* SetKernelOutput(Tensor* out) {
if (out) {
if (out->impl() == nullptr) {
out->set_impl(std::make_shared<phi::DenseTensor>());
}
return static_cast<phi::DenseTensor*>(out->impl().get());
}
return nullptr;
}
std::vector<phi::DenseTensor*> SetKernelOutput(size_t out_size,
std::vector<Tensor>* out) {
out->reserve(out_size);
std::vector<phi::DenseTensor*> results(out_size);
for (size_t i = 0; i < out_size; ++i) {
auto tensor_ptr = std::make_shared<phi::DenseTensor>();
results[i] = tensor_ptr.get();
out->emplace_back();
out->back().set_impl(tensor_ptr);
}
return results;
}
std::vector<phi::DenseTensor*> SetInplaceVectorKernelOutput(
size_t out_size, std::vector<Tensor>* out) {
std::vector<phi::DenseTensor*> results(out->size(), nullptr);
for (size_t i = 0; i < out->size(); ++i) {
results[i] = static_cast<phi::DenseTensor*>(out->at(i).impl().get());
}
return results;
}
std::vector<phi::DenseTensor*> SetInplaceOptionalVectorKernelOutput(
size_t out_size, const paddle::optional<std::vector<Tensor>>& out) {
std::vector<phi::DenseTensor*> results;
if (out) {
results = std::vector<phi::DenseTensor*>(out->size(), nullptr);
for (size_t i = 0; i < out->size(); ++i) {
results[i] = static_cast<phi::DenseTensor*>(out->at(i).impl().get());
}
}
return results;
}
std::vector<phi::DenseTensor*> SetKernelOutput(std::vector<Tensor*>* out) {
std::vector<phi::DenseTensor*> results(out->size(), nullptr);
for (size_t i = 0; i < out->size(); ++i) {
if (out->at(i)) {
auto tensor_ptr = std::make_shared<phi::DenseTensor>();
results[i] = tensor_ptr.get();
(*out)[i]->set_impl(tensor_ptr);
}
}
return results;
}
phi::SelectedRows* SetSelectedRowsKernelOutput(Tensor* out) {
if (!out->initialized()) {
auto select_rows = std::make_shared<phi::SelectedRows>();
out->set_impl(select_rows);
return select_rows.get();
}
return static_cast<phi::SelectedRows*>(out->impl().get());
}
phi::TensorBase* SetSparseKernelOutput(Tensor* out, TensorType type) {
if (!out) {
return nullptr;
}
if (!out->initialized()) {
if (type == TensorType::SPARSE_COO) {
auto sparse_tensor = std::make_shared<phi::SparseCooTensor>(
phi::DenseTensor(), phi::DenseTensor(), phi::DDim{-1});
out->set_impl(sparse_tensor);
return sparse_tensor.get();
} else if (type == TensorType::SPARSE_CSR) {
auto sparse_tensor =
std::make_shared<phi::SparseCsrTensor>(phi::DenseTensor(),
phi::DenseTensor(),
phi::DenseTensor(),
phi::DDim{-1, -1});
out->set_impl(sparse_tensor);
return sparse_tensor.get();
} else {
auto dense_tensor = std::make_shared<phi::DenseTensor>();
out->set_impl(dense_tensor);
return dense_tensor.get();
}
}
return out->impl().get();
}
phi::TensorBase* SetStringsKernelOutput(Tensor* out, TensorType type) {
if (!out->initialized()) {
if (type == TensorType::STRING_TENSOR) {
if (out->impl() == nullptr) {
auto strings_tensor = std::make_shared<phi::StringTensor>();
out->set_impl(strings_tensor);
}
return out->impl().get();
}
}
return out->impl().get();
}
phi::DenseTensor* ProcessStrideBackup(phi::DenseTensor** tensor) {
if (!FLAGS_use_stride_kernel || *tensor == nullptr ||
!(*tensor)->IsInitialized() || (*tensor)->meta().is_contiguous()) {
return nullptr;
} else {
phi::DenseTensor* backup = *tensor;
*tensor = new phi::DenseTensor();
return backup;
}
}
std::vector<phi::DenseTensor*> ProcessStrideBackup(
std::vector<phi::DenseTensor*>* tensor) {
std::vector<phi::DenseTensor*> backup;
backup.reserve(tensor->size());
for (auto& t : *tensor) {
if (!FLAGS_use_stride_kernel || t == nullptr || !t->IsInitialized() ||
t->meta().is_contiguous()) {
backup.emplace_back(nullptr);
} else {
backup.emplace_back(t);
t = new phi::DenseTensor();
}
}
return backup;
}
phi::SelectedRows* ProcessStrideBackup(phi::SelectedRows** tensor) {
return nullptr;
}
template <typename Context>
void TransStride(const Context& dev_ctx,
phi::DenseTensor* from,
phi::DenseTensor* to) {
if (to) {
PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, Context>(
dev_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
}));
delete from;
}
}
template <typename Context>
void TransStride(const Context& dev_ctx,
const std::vector<phi::DenseTensor*>& from,
const std::vector<phi::DenseTensor*>& to) {
for (size_t i = 0; i < to.size(); i++) {
if (to[i]) {
PD_VISIT_ALL_TYPES(to[i]->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, Context>(
dev_ctx,
*from[i],
common::vectorize<int64_t>(to[i]->dims()),
common::vectorize<int64_t>(to[i]->strides()),
to[i]->offset(),
to[i]);
}));
delete from[i];
}
}
}
void TransStride(phi::DeviceContext* dev_ctx,
phi::DenseTensor* from,
phi::DenseTensor* to) {
if (to) {
auto* cpu_ctx = dynamic_cast<phi::CPUContext*>(dev_ctx);
if (cpu_ctx) {
PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::CPUContext>(
*cpu_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
}));
delete from;
return;
}
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
auto* gpu_ctx = dynamic_cast<phi::GPUContext*>(dev_ctx);
if (gpu_ctx) {
PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::GPUContext>(
*gpu_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
}));
delete from;
return;
}
#endif
#ifdef PADDLE_WITH_XPU
auto* xpu_ctx = dynamic_cast<phi::XPUContext*>(dev_ctx);
if (xpu_ctx) {
PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::XPUContext>(
*xpu_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
}));
delete from;
return;
}
#endif
#ifdef PADDLE_WITH_CUSTOM_DEVICE
auto* custom_ctx = dynamic_cast<phi::CustomContext*>(dev_ctx);
if (custom_ctx) {
const phi::KernelKey& kernel_key = {phi::TransToPhiBackend(to->place()),
phi::DataLayout::ALL_LAYOUT,
to->dtype()};
using kernel_signature = void (*)(const phi::DeviceContext&,
const phi::DenseTensor&,
const std::vector<int64_t>&,
const std::vector<int64_t>&,
int64_t,
phi::DenseTensor*);
PD_VISIT_KERNEL("strided_copy",
kernel_key,
kernel_signature,
false,
*custom_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
delete from;
return;
}
#endif
}
}
void TransStrideLegacy(phi::DeviceContext* dev_ctx,
phi::DenseTensor* from,
phi::DenseTensor* to) {
if (to) {
auto* cpu_ctx = dynamic_cast<phi::CPUContext*>(dev_ctx);
if (cpu_ctx) {
PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::CPUContext>(
*cpu_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
}));
return;
}
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
auto* gpu_ctx = dynamic_cast<phi::GPUContext*>(dev_ctx);
if (gpu_ctx) {
PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::GPUContext>(
*gpu_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
}));
return;
}
#endif
#ifdef PADDLE_WITH_XPU
auto* xpu_ctx = dynamic_cast<phi::XPUContext*>(dev_ctx);
if (xpu_ctx) {
PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::XPUContext>(
*xpu_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
}));
return;
}
#endif
#ifdef PADDLE_WITH_CUSTOM_DEVICE
auto* custom_ctx = dynamic_cast<phi::CustomContext*>(dev_ctx);
if (custom_ctx) {
const phi::KernelKey& kernel_key = {phi::TransToPhiBackend(to->place()),
phi::DataLayout::ALL_LAYOUT,
to->dtype()};
using kernel_signature = void (*)(const phi::DeviceContext&,
const phi::DenseTensor&,
const std::vector<int64_t>&,
const std::vector<int64_t>&,
int64_t,
phi::DenseTensor*);
PD_VISIT_KERNEL("strided_copy",
kernel_key,
kernel_signature,
false,
*custom_ctx,
*from,
common::vectorize<int64_t>(to->dims()),
common::vectorize<int64_t>(to->strides()),
to->offset(),
to);
return;
}
#endif
}
}
void TransStride(phi::DeviceContext* dev_ctx,
const std::vector<phi::DenseTensor*>& from,
const std::vector<phi::DenseTensor*>& to) {
for (size_t i = 0; i < to.size(); i++) {
if (to[i]) {
auto* cpu_ctx = dynamic_cast<phi::CPUContext*>(dev_ctx);
if (cpu_ctx) {
PD_VISIT_ALL_TYPES(to[i]->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::CPUContext>(
*cpu_ctx,
*from[i],
common::vectorize<int64_t>(to[i]->dims()),
common::vectorize<int64_t>(to[i]->strides()),
to[i]->offset(),
to[i]);
}));
delete from[i];
continue;
}
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
auto* gpu_ctx = dynamic_cast<phi::GPUContext*>(dev_ctx);
if (gpu_ctx) {
PD_VISIT_ALL_TYPES(to[i]->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::GPUContext>(
*gpu_ctx,
*from[i],
common::vectorize<int64_t>(to[i]->dims()),
common::vectorize<int64_t>(to[i]->strides()),
to[i]->offset(),
to[i]);
}));
delete from[i];
continue;
}
#endif
#ifdef PADDLE_WITH_XPU
auto* xpu_ctx = dynamic_cast<phi::XPUContext*>(dev_ctx);
if (xpu_ctx) {
PD_VISIT_ALL_TYPES(to[i]->dtype(), "StridedCopyKernel", ([&] {
phi::StridedCopyKernel<data_t, phi::XPUContext>(
*xpu_ctx,
*from[i],
common::vectorize<int64_t>(to[i]->dims()),
common::vectorize<int64_t>(to[i]->strides()),
to[i]->offset(),
to[i]);
}));
delete from[i];
continue;
}
#endif
#ifdef PADDLE_WITH_CUSTOM_DEVICE
auto* custom_ctx = dynamic_cast<phi::CustomContext*>(dev_ctx);
if (custom_ctx) {
const phi::KernelKey& kernel_key = {
phi::TransToPhiBackend(to[i]->place()),
phi::DataLayout::ALL_LAYOUT,
to[i]->dtype()};
using kernel_signature = void (*)(const phi::DeviceContext&,
const phi::DenseTensor&,
const std::vector<int64_t>&,
const std::vector<int64_t>&,
int64_t,
phi::DenseTensor*);
PD_VISIT_KERNEL("strided_copy",
kernel_key,
kernel_signature,
false,
*custom_ctx,
*from[i],
common::vectorize<int64_t>(to[i]->dims()),
common::vectorize<int64_t>(to[i]->strides()),
to[i]->offset(),
to[i]);
delete from[i];
continue;
}
#endif
}
}
}
void TransStride(phi::DeviceContext* dev_ctx,
phi::SelectedRows* from,
phi::SelectedRows* to) {}
/* ------------------ for auto parallel ----------------------- */
phi::distributed::DistMetaTensor MakeDistMetaTensor(
const phi::TensorBase& tensor) {
return phi::distributed::DistMetaTensor(tensor);
}
std::vector<phi::distributed::DistMetaTensor> MakeDistMetaTensor(
const std::vector<const phi::TensorBase*>& tensors) {
std::vector<phi::distributed::DistMetaTensor> meta_tensors;
meta_tensors.reserve(tensors.size());
for (const auto* t : tensors) {
meta_tensors.emplace_back(*t);
}
return meta_tensors;
}
phi::distributed::DistTensor* SetKernelDistOutput(
Tensor* out, const phi::distributed::ArgDistAttr& dist_attr) {
PADDLE_ENFORCE_EQ(
paddle::holds_alternative<phi::distributed::TensorDistAttr>(dist_attr),
true,
common::errors::PreconditionNotMet(
"Arg must be a single TensorDistAttr"));
if (out) {
if (out->impl() == nullptr) {
auto dist_t = std::make_shared<phi::distributed::DistTensor>(
phi::DDim(), paddle::get<0>(dist_attr));
out->set_impl(dist_t);
}
return static_cast<phi::distributed::DistTensor*>(out->impl().get());
}
return nullptr;
}
std::vector<phi::distributed::DistTensor*> SetKernelDistOutput(
size_t out_size, std::vector<Tensor>* out) {
std::vector<phi::distributed::DistTensor*> results(out_size);
if (out->size() != out_size) {
// Empty out vector
out->reserve(out_size);
}
for (size_t i = 0; i < out_size; ++i) {
if (out->size() != out_size) {
auto dist_t = std::make_shared<phi::distributed::DistTensor>();
out->emplace_back();
out->back().set_impl(dist_t);
}
results[i] =
static_cast<phi::distributed::DistTensor*>(out->at(i).impl().get());
}
return results;
}
std::vector<phi::distributed::DistTensor*> SetKernelDistOutput(
const phi::distributed::ArgDistAttr& dist_attr, std::vector<Tensor>* out) {
PADDLE_ENFORCE_EQ(
paddle::holds_alternative<std::vector<phi::distributed::TensorDistAttr>>(
dist_attr),
true,
common::errors::PreconditionNotMet(
"Arg must be a vector of TensorDistAttr"));
const std::vector<phi::distributed::TensorDistAttr>& dist_attrs =
PADDLE_GET_CONST(std::vector<phi::distributed::TensorDistAttr>,
dist_attr);
auto out_size = dist_attrs.size();
std::vector<phi::distributed::DistTensor*> results(out_size);
// TODO(GhostScreaming): Inplace outputs are initialized, just set their
// dist_attr.
if (out->size() == out_size) {
VLOG(3) << "Outputs are inplace vector Tensors, SKIP set dist_attr for out "
<< "to avoid changing the inplaced input";
for (size_t i = 0; i < out_size; ++i) {
results[i] =
static_cast<phi::distributed::DistTensor*>(out->at(i).impl().get());
continue;
// auto t =
// static_cast<phi::distributed::DistTensor*>(out->at(i).impl().get());
// auto dist_t = std::make_shared<phi::distributed::DistTensor>(
// t->shared_value(), t->dims(), dist_attrs[i]);
// out->at(i) = Tensor();
// out->at(i).set_impl(dist_t);
// results[i] = dist_t.get();
}
} else {
out->reserve(out_size);
for (size_t i = 0; i < out_size; ++i) {
auto dist_t = std::make_shared<phi::distributed::DistTensor>(
phi::DDim(), dist_attrs[i]);
results[i] = dist_t.get();
out->emplace_back();
out->back().set_impl(dist_t);
}
}
return results;
}
// For backward
std::vector<phi::distributed::DistTensor*> SetKernelDistOutput(
std::vector<Tensor*> out) {
std::vector<phi::distributed::DistTensor*> result;
for (auto tmp : out) {
if (tmp) {
// TODO(GhostScreaming): now all dist case are nullptr
if (tmp->impl() == nullptr) {
auto dist_t = std::make_shared<phi::distributed::DistTensor>();
tmp->set_impl(dist_t);
}
result.emplace_back(
static_cast<phi::distributed::DistTensor*>(tmp->impl().get()));
} else {
result.emplace_back(nullptr);
}
}
return result;
}
std::shared_ptr<phi::distributed::DistTensor> CreateKernelDistOutput(
Tensor* out,
bool set_dist_output_as_tensor_impl,
const phi::distributed::TensorDistAttr& dist_attr) {
if (out) {
auto dist_output =
std::make_shared<phi::distributed::DistTensor>(phi::DDim(), dist_attr);
if (set_dist_output_as_tensor_impl) {
VLOG(3) << "CreateKernelDistOutput function set generated output "
"dist_tensor as Tensor's impl";
if (out->is_dist_tensor()) {
VLOG(3) << "out is DistTensor, set DistAttr:" << dist_attr
<< " to generated DistOutput.";
dist_output->unsafe_set_dist_attr(dist_attr);
}
out->set_impl(dist_output);
}
return dist_output;
}
VLOG(4) << "CreateKernelDistOutput with NULL out";
return nullptr;
}
std::shared_ptr<phi::distributed::DistTensor> CreateKernelDistOutput(
Tensor* out,
bool set_dist_output_as_tensor_impl,
const phi::distributed::ArgDistAttr& dist_attr) {
auto& tensor_dist_attr =
PADDLE_GET_CONST(phi::distributed::TensorDistAttr, dist_attr);
return CreateKernelDistOutput(
out, set_dist_output_as_tensor_impl, tensor_dist_attr);
}
std::shared_ptr<phi::distributed::DistTensor> CreateKernelDistOutput(
Tensor* out, const phi::distributed::ArgDistAttr& dist_attr) {
auto& tensor_dist_attr =
PADDLE_GET_CONST(phi::distributed::TensorDistAttr, dist_attr);
return CreateKernelDistOutput(out, false, tensor_dist_attr);
}
std::vector<std::shared_ptr<phi::distributed::DistTensor>>
CreateKernelDistOutput(std::vector<Tensor*> out,
bool set_dist_output_as_tensor_impl,
const phi::distributed::ArgDistAttr& dist_attr) {
auto tensor_dist_attrs = PADDLE_GET_CONST(
std::vector<phi::distributed::TensorDistAttr>, dist_attr);
PADDLE_ENFORCE_EQ(
out.size(),
tensor_dist_attrs.size(),
common::errors::PreconditionNotMet(
"out.size() [%d] and tensor_dist_attrs.size() [%d] not match",
out.size(),
tensor_dist_attrs.size()));
auto size = tensor_dist_attrs.size();
std::vector<std::shared_ptr<phi::distributed::DistTensor>> results;
results.reserve(size);
for (size_t i = 0; i < size; i++) {
results.emplace_back(CreateKernelDistOutput(
out[i], set_dist_output_as_tensor_impl, tensor_dist_attrs[i]));
}
return results;
}
std::vector<std::shared_ptr<phi::distributed::DistTensor>>
CreateKernelDistOutput(std::vector<Tensor*> out,
bool set_dist_output_as_tensor_impl) {
auto size = out.size();
std::vector<std::shared_ptr<phi::distributed::DistTensor>> results;
results.reserve(size);
for (size_t i = 0; i < size; i++) {
results.emplace_back(
CreateKernelDistOutput(out[i], set_dist_output_as_tensor_impl));
}
return results;
}
void SetReplicatedDistAttrForOutput(
phi::distributed::DistTensor* out,
const phi::distributed::ProcessMesh& process_mesh) {
if (out) {
if (out->dims().size() == -1 || out->dims().size() == 0) {
if (out->local_dims().size() != -1 && out->local_dims().size() != 0) {
out->unsafe_set_dims(out->local_dims());
VLOG(3)
<< "DistTensor out has empty shape, use its local value's shape";
}
}
// For inplace output, we also need to set replicated dist attr
auto dist_attr =
phi::distributed::TensorDistAttr(common::vectorize(out->dims()));
dist_attr.set_process_mesh(process_mesh);
out->unsafe_set_dist_attr(dist_attr);
}
}
/* ------------------ for Allocator ----------------------- */
void CheckAndDoCompact(const std::vector<phi::MetaTensor*>& meta_tensors,
std::string api) {
if (!FLAGS_enable_compact_mem) return;
#if defined(PADDLE_WITH_CUDA)
const auto current_device_id = phi::backends::gpu::GetCurrentDeviceId();
const auto max_reserved =
paddle::memory::DeviceMemoryStatPeakValue("Reserved", current_device_id);
const auto max_allocated =
paddle::memory::DeviceMemoryStatPeakValue("Allocated", current_device_id);
const auto cur_allocated = paddle::memory::DeviceMemoryStatCurrentValue(
"Allocated", current_device_id);
float divisor = 1 << 30;
// calculate total size by meta information
auto CalTensorSize = [&](const std::vector<phi::MetaTensor*>& meta_tensors)
-> std::pair<size_t, std::vector<size_t>> {
size_t req_total_size = 0;
size_t tensor_size = 0;
std::vector<size_t> sizes;
for (auto& meta_tensor : meta_tensors) {
if (meta_tensor == nullptr) continue;
if (meta_tensor->numel() == 0) continue;
if (meta_tensor->numel() < 0) {
VLOG(1) << "meta_tensor->numel():" << meta_tensor->numel()
<< " < 0, skip this tensor in " << api;
continue;
}
tensor_size = meta_tensor->numel() * phi::SizeOf(meta_tensor->dtype());
sizes.push_back(tensor_size);
req_total_size += tensor_size;
}
return {req_total_size, sizes};
};
// judge whether compact is needed according to the following conditions in
// sequence.
// 1. mem_max_reserved < max_reserved_threshold ==> dont need compact
// 2. mem_cur_allocated < cur_allocated_threshold ==> dont need compact
// 3. max_free_size > req_total_size ==> dont need compact
// 4. large_N_free_size < req_total_size ==> need compact
// 5. try_allocate result ==> need compact
auto NeedCompact = [&](const std::vector<phi::MetaTensor*>& meta_tensors) {
if (max_reserved < FLAGS_max_reserved_threshold_in_gb << 30) return false;
if (cur_allocated < FLAGS_cur_allocated_threshold_in_gb << 30) return false;
const auto [max_free_size, large_N_free_size] =
paddle::memory::VmmMaxFreeSize(phi::GPUPlace(current_device_id),
meta_tensors.size());
const auto& [req_total_size, size_vec] = CalTensorSize(meta_tensors);
VLOG(10) << "run api: " << api << " req_total_size: " << req_total_size
<< ", max_free_size: " << max_free_size
<< ", large_N_free_size: " << large_N_free_size
<< ", max_reserved: " << max_reserved
<< ", max_allocated: " << max_allocated
<< ", cur_allocated: " << cur_allocated;
if (req_total_size < max_free_size) return false;
if (req_total_size > large_N_free_size) {
VLOG(1) << "Need Compact in api: " << api
<< " req_total_size: " << req_total_size
<< ", large_N_free_size: " << large_N_free_size
<< ", max_free_size: " << max_free_size
<< ", max_reserved: " << max_reserved
<< ", max_allocated: " << max_allocated
<< ", cur_allocated: " << cur_allocated;
return true;
}
if (FLAGS_try_allocate) {
auto alloc_succ = paddle::memory::TryAllocBatch(
phi::GPUPlace(current_device_id), size_vec);
VLOG(1) << "TryAllocBatch api: " << api << " ret: " << alloc_succ
<< ", req_total_size: " << req_total_size
<< ", large_N_free_size: " << large_N_free_size
<< ", max_free_size: " << max_free_size
<< ", max_reserved: " << max_reserved
<< ", max_allocated: " << max_allocated
<< ", cur_allocated: " << cur_allocated;
return !alloc_succ;
}
return false;
};
if (NeedCompact(meta_tensors)) {
VLOG(1) << "Before Compact max_reserved: " << max_reserved / divisor
<< "GB, max_allocated: " << max_allocated / divisor
<< "GB, cur_allocated: " << cur_allocated / divisor << "GB";
paddle::memory::Compact(phi::GPUPlace(current_device_id));
}
#endif
}
} // namespace paddle::experimental