Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[PHI decoupling] move softmax from fluid to phi and remove cpu_vec.h in fluid #48970

Merged
merged 16 commits into from
Dec 15, 2022
Merged

[PHI decoupling] move softmax from fluid to phi and remove cpu_vec.h in fluid #48970

Show file tree
Hide file tree
Changes from 13 commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
3648b34
move MayIUse in cpu_info from fluid to phi
huangjiyi Dec 9, 2022
bea9e9b
decouple softmax from fluid
huangjiyi Dec 9, 2022
20d33a1
decouple softmax and remove cpu_vec.h in fluid
huangjiyi Dec 9, 2022
c772b42
move softmax from fluid to phi
huangjiyi Dec 9, 2022
4246a22
replace mutable_data with Alloc
huangjiyi Dec 9, 2022
3887d8e
fix bugs
huangjiyi Dec 9, 2022
1f4db48
fix namespace error
huangjiyi Dec 9, 2022
f416f23
correct codestyle
huangjiyi Dec 9, 2022
c9d7c34
fix namespace
huangjiyi Dec 9, 2022
737ce47
fix namespace
huangjiyi Dec 9, 2022
5e4ba72
move cpu_info from fluid to phi
huangjiyi Dec 13, 2022
bc0ed4b
Merge remote-tracking branch 'upstream/develop' into decouple_softmax_h
huangjiyi Dec 13, 2022
258cff3
fix cmake
huangjiyi Dec 13, 2022
2ff9d9e
rm cpu_info in fluid
huangjiyi Dec 14, 2022
951474a
rm cpu_info in fluid
huangjiyi Dec 14, 2022
b3f7277
fix namespace
huangjiyi Dec 14, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 21 additions & 15 deletions paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,13 +13,15 @@ See the License for the specific language governing permissions and
limitations under the License. */

#include "paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.h"
#include "paddle/fluid/operators/math/softmax_impl.h"

#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/platform/collective_helper.h"
#include "paddle/fluid/platform/device/gpu/nccl_helper.h"
#include "paddle/fluid/string/string_helper.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/kernels/funcs/axis_utils.h"
#include "paddle/phi/kernels/funcs/cross_entropy.h"
#include "paddle/phi/kernels/funcs/softmax_impl.h"

namespace paddle {
namespace operators {
Expand Down Expand Up @@ -129,15 +131,15 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
softmax_2d.ShareDataWith(*softmax).Resize({N, D});
loss_2d.ShareDataWith(*loss).Resize({N, 1});

auto eigen_logits = math::EigenMatrix<T>::From(logits_2d);
auto eigen_softmax = math::EigenMatrix<T>::From(softmax_2d);
auto eigen_logits = phi::funcs::EigenMatrix<T>::From(logits_2d);
auto eigen_softmax = phi::funcs::EigenMatrix<T>::From(softmax_2d);

// step 1, obtain logit_max
phi::DenseTensor logits_max;
logits_max = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);
void* logits_max_buff = logits_max.mutable_data<T>(place);

auto eigen_logits_max = math::EigenMatrix<T>::From(logits_max);
auto eigen_logits_max = phi::funcs::EigenMatrix<T>::From(logits_max);
Eigen::DSizes<int, 1> along_axis(1);
eigen_logits_max.device(*dev_ctx.eigen_device()) =
eigen_logits.maximum(along_axis);
Expand All @@ -158,7 +160,7 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
eigen_softmax.device(*dev_ctx.eigen_device()) =
(eigen_logits -
eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class))
.unaryExpr(math::ValueClip<T>());
.unaryExpr(phi::funcs::ValueClip<T>());

// step 3, obtain predict target
phi::DenseTensor predicted_logits;
Expand Down Expand Up @@ -217,7 +219,8 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
sum_exp_logits = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);
void* sum_exp_logits_buff = sum_exp_logits.mutable_data<T>(place);

auto eigen_sum_exp_logits = math::EigenMatrix<T>::From(sum_exp_logits);
auto eigen_sum_exp_logits =
phi::funcs::EigenMatrix<T>::From(sum_exp_logits);
eigen_sum_exp_logits.device(*dev_ctx.eigen_device()) =
eigen_softmax.sum(along_axis);

Expand All @@ -231,8 +234,9 @@ struct CSoftmaxWithCrossEntropyFunctor<phi::GPUContext, T> {
comm->comm(),
stream));

auto eigen_loss = math::EigenMatrix<T>::From(loss_2d);
auto eigen_predicted_logits = math::EigenMatrix<T>::From(predicted_logits);
auto eigen_loss = phi::funcs::EigenMatrix<T>::From(loss_2d);
auto eigen_predicted_logits =
phi::funcs::EigenMatrix<T>::From(predicted_logits);

eigen_loss.device(*dev_ctx.eigen_device()) =
(eigen_sum_exp_logits.log().unaryExpr(phi::funcs::TolerableValue<T>()) -
Expand Down Expand Up @@ -281,14 +285,14 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
softmax_2d.ShareDataWith(*softmax).Resize({N, D});
loss_2d.ShareDataWith(*loss).Resize({N, 1});

auto eigen_logits = math::EigenMatrix<T>::From(logits_2d);
auto eigen_softmax = math::EigenMatrix<T>::From(softmax_2d);
auto eigen_logits = phi::funcs::EigenMatrix<T>::From(logits_2d);
auto eigen_softmax = phi::funcs::EigenMatrix<T>::From(softmax_2d);

// step 1, obtain logit_max
phi::DenseTensor logits_max;
logits_max = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);

auto eigen_logits_max = math::EigenMatrix<T>::From(logits_max);
auto eigen_logits_max = phi::funcs::EigenMatrix<T>::From(logits_max);
Eigen::DSizes<int, 1> along_axis(1);
eigen_logits_max.device(*dev_ctx.eigen_device()) =
eigen_logits.maximum(along_axis);
Expand All @@ -304,7 +308,7 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
eigen_softmax.device(*dev_ctx.eigen_device()) =
(eigen_logits -
eigen_logits_max.reshape(batch_by_one).broadcast(one_by_class))
.unaryExpr(math::ValueClip<T>());
.unaryExpr(phi::funcs::ValueClip<T>());

// step 3, obtain predict target
phi::DenseTensor predicted_logits;
Expand Down Expand Up @@ -357,7 +361,8 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
sum_exp_logits = ctx.AllocateTmpTensor<T, phi::GPUContext>({N, 1}, dev_ctx);
void* sum_exp_logits_buff = sum_exp_logits.mutable_data<T>(place);

auto eigen_sum_exp_logits = math::EigenMatrix<T>::From(sum_exp_logits);
auto eigen_sum_exp_logits =
phi::funcs::EigenMatrix<T>::From(sum_exp_logits);
eigen_sum_exp_logits.device(*dev_ctx.eigen_device()) =
eigen_softmax.sum(along_axis);

Expand All @@ -366,8 +371,9 @@ struct CSoftmaxWithCrossEntropyProcessGroupFunctor<phi::GPUContext, T> {
opts.reduce_op = distributed::ReduceOp::SUM;
pg->AllReduce(in_out, in_out, opts)->Synchronize();

auto eigen_loss = math::EigenMatrix<T>::From(loss_2d);
auto eigen_predicted_logits = math::EigenMatrix<T>::From(predicted_logits);
auto eigen_loss = phi::funcs::EigenMatrix<T>::From(loss_2d);
auto eigen_predicted_logits =
phi::funcs::EigenMatrix<T>::From(predicted_logits);

eigen_loss.device(*dev_ctx.eigen_device()) =
(eigen_sum_exp_logits.log().unaryExpr(phi::funcs::TolerableValue<T>()) -
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/operators/collective/c_softmax_with_cross_entropy_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,9 +22,9 @@ limitations under the License. */
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/softmax.h"
#include "paddle/phi/api/include/tensor.h"
#include "paddle/phi/kernels/funcs/cross_entropy.h"
#include "paddle/phi/kernels/funcs/softmax.h"

namespace paddle {
namespace operators {
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/operators/jit/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ file(APPEND ${jit_file} "\#include \"paddle/fluid/operators/jit/helper.h\"\n")
file(APPEND ${jit_file}
"\#include \"paddle/fluid/operators/jit/registry.h\"\n\n")

set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce place xxhash)
set(JIT_KERNEL_DEPS device_context cblas gflags enforce place xxhash)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

这里device_context能否改成phi_backends?

Copy link
Member Author

@huangjiyi huangjiyi Dec 14, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@YuanRisheng 这里我最开始用的是 phi_backends，然后编译报了找不到 device_context 中一些定义的错误，然后我就改成了 device_context（device_context 也依赖了 phi_backends）


file(
GLOB jit_kernel_cc_srcs
Expand Down
1 change: 0 additions & 1 deletion paddle/fluid/operators/math/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,6 @@ math_library(maxouting)
math_library(sequence_padding)
math_library(sequence_pooling DEPS math_function jit_kernel_helper)
math_library(sequence_scale)
math_library(softmax DEPS math_function jit_kernel_helper)
if(WITH_ASCEND_CL)
math_library(beam_search DEPS math_function beam_search_npu)
elseif(WITH_XPU)
Expand Down
Loading