@@ -28,8 +28,6 @@ class SequenceReshapeKernel : public framework::OpKernel<T> {
2828 auto * out = context.Output <LoDTensor>(" Out"
2929 int out_width = context.Attr <int >(" new_dim"
3030
31- const T* p_in_data = in->data <T>();
32-
3331 auto in_dims = in->dims ();
3432 int64_t in_width = in_dims[1 ];
3533 auto & in_lod = in->lod ();
@@ -43,53 +41,29 @@ class SequenceReshapeKernel : public framework::OpKernel<T> {
4341 auto in_lod_l0 = in_lod[0 ];
4442 int seq_num = in_lod_l0.size () - 1 ;
4543
46- auto & out_lod = *out-> mutable_lod ();
47- out_lod. resize ( 1 );
48- out_lod[ 0 ]. clear ();
49- out_lod[ 0 ]. push_back ( 0 );
50- for ( int i = 0 ; i < seq_num; ++i) {
51- size_t seq_len = in_lod_l0[i + 1 ] - in_lod_l0[i] ;
52- size_t offset = 0 ;
53- offset = (seq_len * in_width) / out_width;
54- PADDLE_ENFORCE_EQ (offset * out_width, seq_len * in_width,
55- " Please make sure (sequence_length * dimension) can be "
56- " divided by new_dim with no remainder for each "
57- " sequence. The %dth sequence is invalid. "
58- i + 1 );
59- PADDLE_ENFORCE_GT (offset, 0 ,
60- " Illegal operation, length of the %dth sequence become "
61- " to 0 after reshaped. " ,
62- i + 1 );
63- out_lod[ 0 ]. push_back (out_lod[ 0 ]. back () + offset);
44+ if (in_width == out_width) {
45+ out-> set_lod (in-> lod () );
46+ } else {
47+ auto & out_lod = *out-> mutable_lod ( );
48+ out_lod. resize ( 1 );
49+ out_lod[ 0 ]. clear () ;
50+ out_lod[ 0 ]. push_back ( 0 ) ;
51+ for ( int i = 0 ; i < seq_num; ++i) {
52+ size_t seq_len = in_lod_l0[i + 1 ] - in_lod_l0[i];
53+ size_t offset = 0 ;
54+ offset = (seq_len * in_width) / out_width;
55+ PADDLE_ENFORCE_EQ (offset * out_width, seq_len * in_width ,
56+ " Please make sure (sequence_length * dimension) can "
57+ " be divided by new_dim with no remainder for each "
58+ " sequence. The %dth sequence is invalid. " ,
59+ i + 1 );
60+ out_lod[ 0 ]. push_back (out_lod[ 0 ]. back () + offset );
61+ }
6462 }
6563
6664 out->mutable_data <T>(context.GetPlace ());
67- out->Resize ({static_cast <int64_t >(out_lod[0 ].back ()), out_width});
68- T* p_out_data = out->mutable_data <T>(context.GetPlace ());
69- math::set_constant (context.device_context (), out, 0 .0f );
70-
71- for (int i = 0 ; i < seq_num; ++i) {
72- size_t in_offset = in_lod_l0[i] * in_width;
73- size_t out_offset = out_lod[0 ][i] * out_width;
74- size_t in_count = (in_lod_l0[i + 1 ] - in_lod_l0[i]) * in_width;
75- size_t out_count = (out_lod[0 ][i + 1 ] - out_lod[0 ][i]) * out_width;
76- size_t bytes = sizeof (T) * std::min (in_count, out_count);
77- if (platform::is_cpu_place (context.GetPlace ())) {
78- memory::Copy (boost::get<platform::CPUPlace>(context.
GetPlace ()),
79- p_out_data + out_offset,
80- boost::get<platform::CPUPlace>(context.GetPlace ()),
81- p_in_data + in_offset, bytes);
82- } else {
83- #ifdef PADDLE_WITH_CUDA
84- auto & dev_ctx =
85- context.
template device_context <platform::CUDADeviceContext>();
86- memory::Copy (boost::get<platform::CUDAPlace>(context.
GetPlace ()),
87- p_out_data + out_offset,
88- boost::get<platform::CUDAPlace>(context.GetPlace ()),
89- p_in_data + in_offset, bytes, dev_ctx.stream ());
90- #endif
91- }
92- }
65+ framework::Copy (*in, context.GetPlace (), out);
66+ out->Resize ({static_cast <int64_t >(out->lod ()[0 ].back ()), out_width});
9367 }
9468};
9569
@@ -98,45 +72,14 @@ class SequenceReshapeGradKernel : public framework::OpKernel<T> {
9872 public:
9973 void Compute (const framework::ExecutionContext& context) const override {
10074 auto * x_tensor_ptr = context.Input <LoDTensor>(" X"
101- auto * out_tensor_ptr = context.Input <LoDTensor>(" Out"
102- auto * out_grad_tensor_ptr =
75+ auto * outg_tensor_ptr =
10376 context.Input <LoDTensor>(framework::GradVarName (" Out"
104- auto * x_grad_tensor_ptr =
77+ auto * xg_tensor_ptr =
10578 context.Output <LoDTensor>(framework::GradVarName (" X"
10679
107- T* p_x_grad_data = x_grad_tensor_ptr->mutable_data <T>(context.GetPlace ());
108- const T* p_out_grad_data = out_grad_tensor_ptr->data <T>();
109-
110- auto & x_lod = x_tensor_ptr->lod ();
111- int seq_num = x_lod[0 ].size () - 1 ;
112- int x_width = x_tensor_ptr->dims ()[1 ];
113- auto & out_lod = out_tensor_ptr->lod ();
114- int out_width = out_tensor_ptr->dims ()[1 ];
115-
116- math::set_constant (context.device_context (), x_grad_tensor_ptr, 0 .0f );
117-
118- for (int i = 0 ; i < seq_num; ++i) {
119- size_t src_offset = out_lod[0 ][i] * out_width;
120- size_t dst_offset = x_lod[0 ][i] * x_width;
121- size_t src_count = (out_lod[0 ][i + 1 ] - out_lod[0 ][i]) * out_width;
122- size_t dst_count = (x_lod[0 ][i + 1 ] - x_lod[0 ][i]) * x_width;
123- size_t bytes = sizeof (T) * std::min (src_count, dst_count);
124- if (platform::is_cpu_place (context.GetPlace ())) {
125- memory::Copy (boost::get<platform::CPUPlace>(context.
GetPlace ()),
126- p_x_grad_data + dst_offset,
127- boost::get<platform::CPUPlace>(context.GetPlace ()),
128- p_out_grad_data + src_offset, bytes);
129- } else {
130- #ifdef PADDLE_WITH_CUDA
131- auto & dev_ctx =
132- context.
template device_context <platform::CUDADeviceContext>();
133- memory::Copy (boost::get<platform::CUDAPlace>(context.
GetPlace ()),
134- p_x_grad_data + dst_offset,
135- boost::get<platform::CUDAPlace>(context.GetPlace ()),
136- p_out_grad_data + src_offset, bytes, dev_ctx.stream ());
137- #endif
138- }
139- }
80+ xg_tensor_ptr->mutable_data <T>(context.GetPlace ());
81+ framework::Copy (*outg_tensor_ptr, context.GetPlace (), xg_tensor_ptr);
82+ xg_tensor_ptr->Resize (x_tensor_ptr->dims ());
14083 }
14184};
14285
0 commit comments