Skip to content

Commit

Permalink
support adagrad sparse update (#5272)
Browse files Browse the repository at this point in the history 
* adam sparse support

* fix gpu build error

* fix ci

* fix ci

* fix adagrad sparse update bug

* fix gpu build error
  • Loading branch information
@QiJune
QiJune committed Nov 16, 2017
1 parent e0e3a8a commit d7bf372
Show file tree
Hide file tree
Showing 7 changed files with 386 additions and 38 deletions.
9 changes: 7 additions & 2 deletions paddle/operators/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -174,13 +174,18 @@ set(DEPS_OPS
array_to_lod_tensor_op
lstm_op
tensor_array_read_write_op
gru_op)
gru_op
adagrad_op
sgd_op)


op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
op_library(cross_entropy_op DEPS cross_entropy)
op_library(softmax_with_cross_entropy_op DEPS cross_entropy softmax)
op_library(sum_op DEPS selected_rows_functor)
op_library(sgd_op DEPS selected_rows_functor)
op_library(adagrad_op DEPS selected_rows_functor)
op_library(conv_op DEPS vol2col)
op_library(sum_op DEPS net_op selected_rows_functor)
op_library(pool_op DEPS pooling)
op_library(pool_with_index_op DEPS pooling)
op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table)
Expand Down
90 changes: 85 additions & 5 deletions paddle/operators/adagrad_op.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,14 +14,19 @@ limitations under the License. */

#include "paddle/operators/adagrad_op.h"

#include <cmath>

#include "paddle/operators/math/math_function.h"
#include "paddle/operators/math/selected_rows_functor.h"

namespace paddle {
namespace operators {

class AdagradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;

void InferShape(framework::InferShapeContext *ctx) const override {
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Param"),
"Input(Param) of AdagradOp should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Grad"),
Expand Down Expand Up @@ -54,8 +59,8 @@ class AdagradOp : public framework::OperatorWithKernel {

class AdagradOpMaker : public framework::OpProtoAndCheckerMaker {
public:
AdagradOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
AdagradOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Param", "(Tensor) Input parameter");
AddInput("Grad", "(Tensor) Input gradient");
Expand Down Expand Up @@ -87,10 +92,85 @@ for numerical stability to avoid the division by zero error.
)DOC");
}
};

namespace {
size_t FindPos(const std::vector<int64_t>& rows, int64_t value) {
return std::find(rows.begin(), rows.end(), value) - rows.begin();
}
} // namespace

template <typename T>
struct SparseAdagradFunctor<platform::CPUPlace, T> {
void operator()(const platform::DeviceContext& context,
const framework::SelectedRows& grad,
const framework::Tensor& learning_rate, T epsilon,
framework::Tensor* moment, framework::Tensor* param) {
// 1. g_m.rows = set(g.rows)
auto grad_rows = grad.rows();
std::set<int64_t> row_set(grad_rows.begin(), grad_rows.end());
std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());

auto grad_width = grad.value().dims()[1];
std::unique_ptr<framework::SelectedRows> grad_merge{
new framework::SelectedRows()};
grad_merge->set_rows(merge_rows);
grad_merge->set_height(grad.height());
grad_merge->mutable_value()->mutable_data<T>(
framework::make_ddim(
{static_cast<int64_t>(merge_rows.size()), grad_width}),
context.GetPlace());

math::SetConstant<platform::CPUPlace, T> constant_functor;
constant_functor(context, grad_merge->mutable_value(), 0.0);

auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
auto* grad_data = grad.value().data<T>();

for (size_t i = 0; i < grad_rows.size(); i++) {
size_t grad_merge_i = FindPos(merge_rows, grad_rows[i]);
for (int64_t j = 0; j < grad_width; j++) {
grad_merge_data[grad_merge_i * grad_width + j] +=
grad_data[i * grad_width + j];
}
}

// 2. m += g_m * g_m
std::unique_ptr<framework::SelectedRows> grad_square{
new framework::SelectedRows()};
grad_square->set_rows(grad_merge->rows());
grad_square->set_height(grad_merge->height());
grad_square->mutable_value()->mutable_data<T>(grad_merge->value().dims(),
context.GetPlace());
auto gs =
framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
gs.device(*context.GetEigenDevice<platform::CPUPlace>()) = gm * gm;

math::SelectedRowsAddToTensor<platform::CPUPlace, T> functor;
functor(context, *grad_square, moment);

// 3. update parameter
auto* lr = learning_rate.data<T>();
auto* param_data = param->data<T>();
auto* moment_data = moment->data<T>();

for (size_t i = 0; i < merge_rows.size(); i++) {
for (int64_t j = 0; j < grad_width; j++) {
param_data[merge_rows[i] * grad_width + j] -=
lr[0] * grad_merge_data[i * grad_width + j] /
(std::sqrt(moment_data[merge_rows[i] * grad_width + j]) + epsilon);
}
}
}
};

template struct SparseAdagradFunctor<platform::CPUPlace, float>;
template struct SparseAdagradFunctor<platform::CPUPlace, double>;
} // namespace operators
} // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker);
REGISTER_OP_CPU_KERNEL(adagrad,
ops::AdagradOpKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
adagrad, ops::AdagradOpKernel<paddle::platform::CPUPlace, float>,
ops::AdagradOpKernel<paddle::platform::CPUPlace, double>);
135 changes: 133 additions & 2 deletions paddle/operators/adagrad_op.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,138 @@

#define EIGEN_USE_GPU
#include "paddle/operators/adagrad_op.h"
#include "paddle/operators/math/selected_rows_functor.h"
#include "paddle/operators/math/math_function.h"
#include "paddle/platform/cuda_helper.h"

namespace paddle {
namespace operators {

namespace {

template <typename T, int block_size>
__global__ void MergeGradKernel(const T* grad, const int64_t* grad_rows,
T* grad_merge, const int64_t* grad_merge_rows,
size_t grad_merge_rows_size,
int64_t row_numel) {
const int ty = blockIdx.y;
int tid = threadIdx.x;
__shared__ size_t grad_merge_idx;

if (tid == 0) {
for (size_t i = 0; i < grad_merge_rows_size; i++) {
if (grad_rows[ty] == grad_merge_rows[i]) {
grad_merge_idx = i;
}
}
}

__syncthreads();

grad += ty * row_numel;
grad_merge += grad_merge_idx * row_numel;
for (int index = tid; index < row_numel; index += block_size) {
paddle::platform::CudaAtomicAdd(grad_merge + index, grad[index]);
}
}

template <typename T, int block_size>
__global__ void SparseAdagradFunctorKernel(const T* grad, const int64_t* rows,
const T* learning_rate, T* param,
T* moment, int64_t row_numel,
T epsilon) {
const int ty = blockIdx.y;
int tid = threadIdx.x;

grad += ty * row_numel;
param += rows[ty] * row_numel;
moment += rows[ty] * row_numel;

for (int index = tid; index < row_numel; index += block_size) {
// Since index in rows of SelectedRows can be duplicate, we have to use
// Atomic Operation to avoid concurrent write error.
paddle::platform::CudaAtomicAdd(param + index,
-1.0 * learning_rate[0] * grad[index] /
(sqrt(moment[index]) + epsilon));
}
}
} // namespace

template <typename T>
struct SparseAdagradFunctor<platform::GPUPlace, T> {
void operator()(const platform::DeviceContext& context,
const framework::SelectedRows& grad,
const framework::Tensor& learning_rate, T epsilon,
framework::Tensor* moment, framework::Tensor* param) {
// 1. g_m.rows = set(g.rows)
auto grad_rows = grad.rows();
std::set<int64_t> row_set(grad_rows.begin(), grad_rows.end());
std::vector<int64_t> merge_rows(row_set.begin(), row_set.end());

auto grad_width = grad.value().dims()[1];
std::unique_ptr<framework::SelectedRows> grad_merge{
new framework::SelectedRows()};
grad_merge->set_rows(merge_rows);
grad_merge->set_height(grad.height());
grad_merge->mutable_value()->mutable_data<T>(
framework::make_ddim(
{static_cast<int64_t>(merge_rows.size()), grad_width}),
context.GetPlace());

math::SetConstant<platform::GPUPlace, T> constant_functor;
constant_functor(context, grad_merge->mutable_value(), 0.0);

auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
auto* grad_data = grad.value().data<T>();

const int block_size = 256;
dim3 threads(block_size, 1);
dim3 grid1(1, grad_rows.size());

MergeGradKernel<
T, 256><<<grid1, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(grad_data, grad.rows().data(),
grad_merge_data, grad_merge->rows().data(),
grad_merge->rows().size(), grad_width);

// 2. m += g_m * g_m
std::unique_ptr<framework::SelectedRows> grad_square{
new framework::SelectedRows()};
grad_square->set_rows(grad_merge->rows());
grad_square->set_height(grad_merge->height());
grad_square->mutable_value()->mutable_data<T>(grad_merge->value().dims(),
context.GetPlace());
auto gs =
framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
gs.device(*context.GetEigenDevice<platform::GPUPlace>()) = gm * gm;

math::SelectedRowsAddToTensor<platform::GPUPlace, T> functor;
functor(context, *grad_square, moment);

// 3. update parameter
auto* lr = learning_rate.data<T>();
auto* param_data = param->data<T>();
auto* moment_data = moment->data<T>();

dim3 grid2(1, merge_rows.size());
SparseAdagradFunctorKernel<
T, 256><<<grid2, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(grad_merge_data, grad_merge->rows().data(),
lr, param_data,
moment_data, grad_width, epsilon);
}
};

template struct SparseAdagradFunctor<platform::GPUPlace, float>;
template struct SparseAdagradFunctor<platform::GPUPlace, double>;

} // namespace operators
} // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(adagrad,
ops::AdagradOpKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
adagrad, ops::AdagradOpKernel<paddle::platform::GPUPlace, float>,
ops::AdagradOpKernel<paddle::platform::GPUPlace, double>);
66 changes: 45 additions & 21 deletions paddle/operators/adagrad_op.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,35 +19,59 @@ limitations under the License. */
namespace paddle {
namespace operators {

template <typename Place, typename T>
struct SparseAdagradFunctor {
void operator()(const platform::DeviceContext& context,
const framework::SelectedRows& grad,
const framework::Tensor& learning_rate, T epsilon,
framework::Tensor* moment, framework::Tensor* param);
};

template <typename Place, typename T>
class AdagradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
auto moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");
auto* param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
auto* moment_out_tensor = ctx.Output<framework::Tensor>("MomentOut");

param_out_tensor->mutable_data<T>(ctx.GetPlace());
moment_out_tensor->mutable_data<T>(ctx.GetPlace());

float epsilon = ctx.Attr<float>("epsilon");

auto param = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("Param"));
auto grad = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("Grad"));
auto moment = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("Moment"));
auto lr = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("LearningRate"));

auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
auto place = ctx.GetEigenDevice<Place>();

moment_out.device(place) = moment + grad * grad;
Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
param_out.device(place) =
param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));

auto* grad_var = ctx.InputVar("Grad");
if (grad_var->IsType<framework::LoDTensor>()) {
auto param = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("Param"));
auto grad = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("Grad"));
auto moment = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("Moment"));
auto lr = framework::EigenVector<T>::Flatten(
*ctx.Input<framework::Tensor>("LearningRate"));

auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
auto place = ctx.GetEigenDevice<Place>();

moment_out.device(place) = moment + grad * grad;
Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
param_out.device(place) =
param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
} else if (grad_var->IsType<framework::SelectedRows>()) {
auto* param_tensor = ctx.Input<framework::Tensor>("Param");
PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor);

auto* moment_tensor = ctx.Input<framework::Tensor>("Moment");
PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor);

SparseAdagradFunctor<Place, T> functor;
functor(ctx.device_context(), *ctx.Input<framework::SelectedRows>("Grad"),
*ctx.Input<framework::Tensor>("LearningRate"), epsilon,
moment_out_tensor, param_out_tensor);
} else {
PADDLE_THROW("Unsupported Variable Type of Grad");
}
}
};

Expand Down
15 changes: 8 additions & 7 deletions paddle/operators/sgd_op.cu
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,11 +20,11 @@ namespace paddle {
namespace operators {

namespace {
template <typename T>
template <typename T, int block_size>
__global__ void SparseSGDFunctorKernel(const T* selected_rows,
const int64_t* rows,
const T* learning_rate, T* tensor_out,
int64_t row_numel, int block_size) {
int64_t row_numel) {
const int ty = blockIdx.y;
int tid = threadIdx.x;

Expand Down Expand Up @@ -59,14 +59,15 @@ struct SparseSGDFunctor<platform::GPUPlace, T> {
auto* in_data = in_value.data<T>();
auto* out_data = output->data<T>();

int block_size = 256;
const int block_size = 256;
dim3 threads(block_size, 1);
dim3 grid(1, in_rows.size());
SparseSGDFunctorKernel<
T><<<grid, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(in_data, in_rows.data(), learning_rate.data<T>(),
out_data, in_row_numel, block_size);
T, 256><<<grid, threads, 0,
reinterpret_cast<const platform::CUDADeviceContext&>(context)
.stream()>>>(in_data, in_rows.data(),
learning_rate.data<T>(), out_data,
in_row_numel);
}
};

Expand Down
Loading

0 comments on commit d7bf372

Please sign in to comment.