未验证 提交 b9e4aaa5 编写于 作者: X XiangGao 提交者: GitHub

fix trace offset out of shape (#33922)

上级 fcdbc8de
...@@ -14,6 +14,7 @@ ...@@ -14,6 +14,7 @@
#include <thrust/device_vector.h> #include <thrust/device_vector.h>
#include <thrust/host_vector.h> #include <thrust/host_vector.h>
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/reduce_ops/cub_reduce.h" #include "paddle/fluid/operators/reduce_ops/cub_reduce.h"
#include "paddle/fluid/operators/trace_op.h" #include "paddle/fluid/operators/trace_op.h"
...@@ -50,6 +51,9 @@ class TraceCUDAKernel : public framework::OpKernel<T> { ...@@ -50,6 +51,9 @@ class TraceCUDAKernel : public framework::OpKernel<T> {
TensorReduce<T, T, cub::Sum, IdentityFunctor>( TensorReduce<T, T, cub::Sum, IdentityFunctor>(
diag, out, reduce_dims, static_cast<T>(0), cub::Sum(), diag, out, reduce_dims, static_cast<T>(0), cub::Sum(),
IdentityFunctor(), stream); IdentityFunctor(), stream);
} else {
math::SetConstant<DeviceContext, T> functor;
functor(context.device_context<DeviceContext>(), out, static_cast<T>(0));
} }
} }
}; };
......
...@@ -179,7 +179,7 @@ class TraceKernel : public framework::OpKernel<T> { ...@@ -179,7 +179,7 @@ class TraceKernel : public framework::OpKernel<T> {
auto output_dims = out->dims(); auto output_dims = out->dims();
out->mutable_data<T>(context.GetPlace()); T* out_data = out->mutable_data<T>(context.GetPlace());
const framework::Tensor diag = const framework::Tensor diag =
Diagonal<DeviceContext, T>(context, input, offset, dim1, dim2); Diagonal<DeviceContext, T>(context, input, offset, dim1, dim2);
...@@ -191,6 +191,8 @@ class TraceKernel : public framework::OpKernel<T> { ...@@ -191,6 +191,8 @@ class TraceKernel : public framework::OpKernel<T> {
auto reduce_dim = Eigen::array<int, 1>({1}); auto reduce_dim = Eigen::array<int, 1>({1});
output.device(place) = x.sum(reduce_dim); output.device(place) = x.sum(reduce_dim);
out->Resize(output_dims); out->Resize(output_dims);
} else {
std::fill(out_data, out_data + out->numel(), static_cast<T>(0));
} }
} }
}; };
......
Markdown is supported
0% .
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