add my code in svd_helper.h

Author: 0x45f

paddle/fluid/operators/matrix_rank_op.cc

@@ -213,7 +213,7 @@ class MatrixRankCPUKernel : public framework::OpKernel<T> {
 
     Tensor temp_rtol_tensor;
     framework::TensorFromVector<T>(std::vector<T>{rtol_T}, &temp_rtol_tensor);
-    Tensor rtol_tensor = dito_T.mul(temp_rtol_tensor, max_eigenvalue_tensor);
+    Tensor rtol_tensor = dito_T.Mul(temp_rtol_tensor, max_eigenvalue_tensor);
     Tensor tol_tensor;
     tol_tensor.mutable_data<T>(dim_out, context.GetPlace());
     ElementwiseComputeEx<GreaterElementFunctor<T>, platform::CPUDeviceContext,
@@ -239,9 +239,9 @@ class MatrixRankCPUKernel : public framework::OpKernel<T> {
     auto dito_int =
         math::DeviceIndependenceTensorOperations<platform::CPUDeviceContext,
                                                  int64_t>(context);
-    std::vector<int> res_shape = framework::vectorize<int>(dim_out);
-    Tensor res = dito_int.ReduceSum(compare_result, res_shape);
-    out->ShareDataWith(res);
+    std::vector<int> result_shape = framework::vectorize<int>(dim_out);
+    Tensor result = dito_int.ReduceSum(compare_result, result_shape);
+    out->ShareDataWith(result);
   }
 };
 

paddle/fluid/operators/matrix_rank_op.cu

@@ -116,7 +116,7 @@ class MatrixRankGPUKernel : public framework::OpKernel<T> {
     Tensor temp_rtol_tensor;
     framework::TensorFromVector<T>(std::vector<T>{rtol_T},
                                    context.device_context(), &temp_rtol_tensor);
-    Tensor rtol_tensor = dito_T.mul(temp_rtol_tensor, max_eigenvalue_tensor);
+    Tensor rtol_tensor = dito_T.Mul(temp_rtol_tensor, max_eigenvalue_tensor);
     Tensor tol_tensor;
     tol_tensor.mutable_data<T>(dim_out, context.GetPlace());
     ElementwiseComputeEx<GreaterElementFunctor<T>, platform::CUDADeviceContext,
@@ -143,9 +143,9 @@ class MatrixRankGPUKernel : public framework::OpKernel<T> {
     auto dito_int =
         math::DeviceIndependenceTensorOperations<platform::CUDADeviceContext,
                                                  int64_t>(context);
-    std::vector<int> res_shape = framework::vectorize<int>(dim_out);
-    Tensor res = dito_int.ReduceSum(compare_result, res_shape);
-    out->ShareDataWith(res);
+    std::vector<int> result_shape = framework::vectorize<int>(dim_out);
+    Tensor result = dito_int.ReduceSum(compare_result, result_shape);
+    out->ShareDataWith(result);
   }
 
   void GesvdjBatched(const platform::CUDADeviceContext& dev_ctx, int batchSize,

paddle/fluid/operators/svd_helper.h

@@ -48,7 +48,6 @@ void EigenSvd(const T* X, T* U, T* VH, T* S, int rows, int cols,
       svd(2, 2, flag);
   /*NOTE(xiongkun03) Eigen::Matrix API need non-const pointer.*/
   T* input = const_cast<T*>(X);
-This conversation was marked as resolved by Aurelius84
   auto m = Eigen::Map<
       Eigen::Matrix<T, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>(
       input, rows, cols);
@@ -314,6 +313,22 @@ struct DeviceIndependenceTensorOperations {
     return CreateOpRunAndReturnTensor("slice", inputs, attrs, out_shape);
   }
 
+  framework::Tensor ReduceSum(const framework::Tensor& x,
+                              std::vector<int> out_dim) {
+    framework::AttributeMap attrs;
+    attrs["dim"] = std::vector<int>{-1};
+    NameInTensorMap inputs({{"X", {&x}}});
+    return CreateOpRunAndReturnTensor("reduce_sum", inputs, attrs, out_dim);
+  }
+
+  framework::Tensor ReduceMax(const framework::Tensor& x,
+                              std::vector<int> out_dim) {
+    framework::AttributeMap attrs;
+    attrs["dim"] = std::vector<int>{-1};
+    NameInTensorMap inputs({{"X", {&x}}});
+    return CreateOpRunAndReturnTensor("reduce_max", inputs, attrs, out_dim);
+  }
+
  private:
   const framework::ExecutionContext& context;
   BlasT<DeviceContext, T> GetBlas() {
@@ -370,4 +385,4 @@ struct DeviceIndependenceTensorOperations {
 };
 }  // namespace math
 }  // namespace operators
-}  // namespace paddle
+}  // namespace paddle

paddle/fluid/platform/dynload/cusolver.h

@@ -68,12 +68,6 @@ CUSOLVER_ROUTINE_EACH_R1(DECLARE_DYNAMIC_LOAD_CUSOLVER_WRAP)
 
 #if CUDA_VERSION >= 9020
 #define CUSOLVER_ROUTINE_EACH_R2(__macro) \
-  __macro(cusolverDnCreateGesvdjInfo);    \
-  __macro(cusolverDnSgesvdj_bufferSize);  \
-  __macro(cusolverDnDgesvdj_bufferSize);  \
-  __macro(cusolverDnSgesvdj);             \
-  __macro(cusolverDnDgesvdj);             \
-  __macro(cusolverDnDestroyGesvdjInfo);   \
   __macro(cusolverDnCreateSyevjInfo);     \
   __macro(cusolverDnSsyevj_bufferSize);   \
   __macro(cusolverDnDsyevj_bufferSize);   \