add new API/OP: paddle.linalg.triangular_solve (#36714) (#37551)

cherry-pick #36714

paddle/fluid/operators/math/blas.h

@@ -253,6 +253,12 @@ class Blas {
   void BatchedGETRS(CBLAS_TRANSPOSE trans, int n, int nrhs, const T** a,
                     int lda, int* ipiv, T** b, int ldb, int* info,
                     int batch_size) const;
+
+  // cuBlas triangular_solve
+  template <typename T>
+  void BatchedTRSM(CBLAS_SIDE side, CBLAS_UPLO uplo, CBLAS_TRANSPOSE transA,
+                   CBLAS_DIAG diag, int M, int N, T alpha, const T** a, int lda,
+                   T** b, int ldb, int batch_size) const;
 #endif
 
  private:
@@ -414,6 +420,12 @@ class BlasT : private Blas<DeviceContext> {
   void BatchedGETRS(ARGS... args) const {
     Base()->template BatchedGETRS<T>(args...);
   }
+
+  // triangular_solve
+  template <typename... ARGS>
+  void BatchedTRSM(ARGS... args) const {
+    Base()->template BatchedTRSM<T>(args...);
+  }
 #endif
 
  private:

paddle/fluid/operators/math/blas_impl.cu.h

@@ -120,6 +120,11 @@ struct CUBlas<float> {
     PADDLE_ENFORCE_CUDA_SUCCESS(
         platform::dynload::cublasSgetrsBatched(args...));
   }
+
+  template <typename... ARGS>
+  static void TRSM_BATCH(ARGS... args) {
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cublasStrsmBatched(args...));
+  }
 };
 
 template <>
@@ -194,6 +199,11 @@ struct CUBlas<double> {
     PADDLE_ENFORCE_CUDA_SUCCESS(
         platform::dynload::cublasDgetrsBatched(args...));
   }
+
+  template <typename... ARGS>
+  static void TRSM_BATCH(ARGS... args) {
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cublasDtrsmBatched(args...));
+  }
 };
 
 template <>
@@ -339,6 +349,19 @@ struct CUBlas<platform::complex<float>> {
         reinterpret_cast<cuFloatComplex *>(C), ldc));
   }
 
+  static void TRSM(cublasHandle_t handle, cublasSideMode_t side,
+                   cublasFillMode_t uplo, cublasOperation_t transa,
+                   cublasDiagType_t diag, int m, int n,
+                   const paddle::platform::complex<float> *alpha,
+                   const paddle::platform::complex<float> *A, int lda,
+                   paddle::platform::complex<float> *B, int ldb) {
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cublasCtrsm(
+        handle, side, uplo, transa, diag, m, n,
+        reinterpret_cast<const cuFloatComplex *>(alpha),
+        reinterpret_cast<const cuFloatComplex *>(A), lda,
+        reinterpret_cast<cuFloatComplex *>(B), ldb));
+  }
+
   // NOTES: GEMM_EX can use Tensor Core to accelerate matrix multiply.
   // https://docs.nvidia.com/cuda/cublas/index.html#cublassetmathmode
   template <typename... ARGS>
@@ -370,6 +393,20 @@ struct CUBlas<platform::complex<float>> {
         "cublasGemmEx is not supported on cuda <= 7.5"));
 #endif
   }
+
+  static void TRSM_BATCH(cublasHandle_t handle, cublasSideMode_t side,
+                         cublasFillMode_t uplo, cublasOperation_t transa,
+                         cublasDiagType_t diag, int m, int n,
+                         const paddle::platform::complex<float> *alpha,
+                         const paddle::platform::complex<float> **A, int lda,
+                         paddle::platform::complex<float> **B, int ldb,
+                         int batch_size) {
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cublasCtrsmBatched(
+        handle, side, uplo, transa, diag, m, n,
+        reinterpret_cast<const cuFloatComplex *>(alpha),
+        reinterpret_cast<const cuFloatComplex **>(A), lda,
+        reinterpret_cast<cuFloatComplex **>(B), ldb, batch_size));
+  }
 };
 
 template <>
@@ -440,6 +477,33 @@ struct CUBlas<platform::complex<double>> {
         reinterpret_cast<cuDoubleComplex *>(C), ldc));
   }
 
+  static void TRSM(cublasHandle_t handle, cublasSideMode_t side,
+                   cublasFillMode_t uplo, cublasOperation_t transa,
+                   cublasDiagType_t diag, int m, int n,
+                   const paddle::platform::complex<double> *alpha,
+                   const paddle::platform::complex<double> *A, int lda,
+                   paddle::platform::complex<double> *B, int ldb) {
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cublasZtrsm(
+        handle, side, uplo, transa, diag, m, n,
+        reinterpret_cast<const cuDoubleComplex *>(alpha),
+        reinterpret_cast<const cuDoubleComplex *>(A), lda,
+        reinterpret_cast<cuDoubleComplex *>(B), ldb));
+  }
+
+  static void TRSM_BATCH(cublasHandle_t handle, cublasSideMode_t side,
+                         cublasFillMode_t uplo, cublasOperation_t transa,
+                         cublasDiagType_t diag, int m, int n,
+                         const paddle::platform::complex<double> *alpha,
+                         const paddle::platform::complex<double> **A, int lda,
+                         paddle::platform::complex<double> **B, int ldb,
+                         int batch_size) {
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cublasZtrsmBatched(
+        handle, side, uplo, transa, diag, m, n,
+        reinterpret_cast<const cuDoubleComplex *>(alpha),
+        reinterpret_cast<const cuDoubleComplex **>(A), lda,
+        reinterpret_cast<cuDoubleComplex **>(B), ldb, batch_size));
+  }
+
   // NOTES: GEMM_EX can use Tensor Core to accelerate matrix multiply.
   // https://docs.nvidia.com/cuda/cublas/index.html#cublassetmathmode
   template <typename... ARGS>
@@ -897,6 +961,30 @@ void Blas<platform::CUDADeviceContext>::BatchedGETRS(
   });
 }
 
+template <>
+template <typename T>
+void Blas<platform::CUDADeviceContext>::BatchedTRSM(
+    CBLAS_SIDE side, CBLAS_UPLO uplo, CBLAS_TRANSPOSE transA, CBLAS_DIAG diag,
+    int M, int N, T alpha, const T **A, int lda, T **B, int ldb,
+    int batch_size) const {
+  // solve row major `op ( A ) X = α B` by taking it as `X' op ( A' )  =  α B'`
+  // where ' stands for transpose
+  cublasSideMode_t cuSide =
+      (side == CblasLeft) ? CUBLAS_SIDE_RIGHT : CUBLAS_SIDE_LEFT;
+  cublasFillMode_t cuUplo =
+      (uplo == CblasLower) ? CUBLAS_FILL_MODE_UPPER : CUBLAS_FILL_MODE_LOWER;
+  // use CUBLAS_OP_C (conjugate transpose) for complex
+  cublasOperation_t cuTransA =
+      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
+  cublasDiagType_t cuDiag =
+      (diag == CblasUnit) ? CUBLAS_DIAG_UNIT : CUBLAS_DIAG_NON_UNIT;
+
+  context_.CublasCall([&](cublasHandle_t handle) {
+    CUBlas<T>::TRSM_BATCH(handle, cuSide, cuUplo, cuTransA, cuDiag, N, M,
+                          &alpha, A, lda, B, ldb, batch_size);
+  });
+}
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/math/blas_impl.h

@@ -434,6 +434,17 @@ struct CBlas<platform::complex<float>> {
                                    a_, lda, b_, ldb, &beta, c_, ldc);
   }
 
+  static void TRSM(CBLAS_LAYOUT layout, CBLAS_SIDE side, CBLAS_UPLO uplo,
+                   CBLAS_TRANSPOSE trans_a, CBLAS_DIAG diag, int M, int N,
+                   paddle::platform::complex<float> alpha,
+                   const paddle::platform::complex<float> *A, int lda,
+                   paddle::platform::complex<float> *B, int ldb) {
+    const void *a_ = (const void *)(A);
+    void *b_ = static_cast<void *>(B);
+    platform::dynload::cblas_ctrsm(layout, side, uplo, trans_a, diag, M, N,
+                                   &alpha, a_, lda, b_, ldb);
+  }
+
   template <typename... ARGS>
   static void GEMM_BATCH(CBLAS_LAYOUT layout, CBLAS_TRANSPOSE *trans_a,
                          CBLAS_TRANSPOSE *trans_b, int *M, int *N, int *K,
@@ -562,6 +573,17 @@ struct CBlas<platform::complex<double>> {
                                    a_, lda, b_, ldb, &beta, c_, ldc);
   }
 
+  static void TRSM(CBLAS_LAYOUT layout, CBLAS_SIDE side, CBLAS_UPLO uplo,
+                   CBLAS_TRANSPOSE trans_a, CBLAS_DIAG diag, int M, int N,
+                   paddle::platform::complex<double> alpha,
+                   const paddle::platform::complex<double> *A, int lda,
+                   paddle::platform::complex<double> *B, int ldb) {
+    const void *a_ = (const void *)(A);
+    void *b_ = static_cast<void *>(B);
+    platform::dynload::cblas_ztrsm(layout, side, uplo, trans_a, diag, M, N,
+                                   &alpha, a_, lda, b_, ldb);
+  }
+
   template <typename... ARGS>
   static void GEMM_BATCH(CBLAS_LAYOUT layout, CBLAS_TRANSPOSE *trans_a,
                          CBLAS_TRANSPOSE *trans_b, int *M, int *N, int *K,
@@ -682,6 +704,15 @@ struct CBlas<platform::complex<float>> {
     cblas_cgemm(layout, TransA, TransB, M, N, K, &alpha, A, lda, B, ldb, &beta,
                 C, ldc);
   }
+
+  static void TRSM(const CBLAS_LAYOUT layout, const CBLAS_SIDE side,
+                   const CBLAS_UPLO uplo, const CBLAS_TRANSPOSE transA,
+                   const CBLAS_DIAG diag, const int M, const int N,
+                   const paddle::platform::complex<float> alpha,
+                   const paddle::platform::complex<float> *A, const int lda,
+                   paddle::platform::complex<double> *B, const int ldb) {
+    cblas_ctrsm(layout, side, uplo, transA, diag, M, N, &alpha, A, lda, B, ldb);
+  }
 };
 
 template <>
@@ -720,6 +751,15 @@ struct CBlas<platform::complex<double>> {
     cblas_zgemm(layout, TransA, TransB, M, N, K, &alpha, A, lda, B, ldb, &beta,
                 C, ldc);
   }
+
+  static void TRSM(const CBLAS_LAYOUT layout, const CBLAS_SIDE side,
+                   const CBLAS_UPLO uplo, const CBLAS_TRANSPOSE transA,
+                   const CBLAS_DIAG diag, const int M, const int N,
+                   const paddle::platform::complex<double> alpha,
+                   const paddle::platform::complex<double> *A, const int lda,
+                   paddle::platform::complex<double> *B, const int ldb) {
+    cblas_ztrsm(layout, side, uplo, transA, diag, M, N, &alpha, A, lda, B, ldb);
+  }
 };
 
 #endif

paddle/fluid/operators/math/blas_impl.hip.h

@@ -90,6 +90,12 @@ struct CUBlas<float> {
     PADDLE_THROW(platform::errors::Unimplemented(
         "cublasSmatinvBatched is not supported on HIP platform."));
   }
+
+  template <typename... ARGS>
+  static void TRSM_BATCH(ARGS... args) {
+    PADDLE_THROW(platform::errors::Unimplemented(
+        "cublasStrsmBatched is not supported on HIP platform."));
+  }
 };
 
 template <>
@@ -153,6 +159,12 @@ struct CUBlas<double> {
     PADDLE_THROW(platform::errors::Unimplemented(
         "cublasDmatinvBatched is not supported on HIP platform."));
   }
+
+  template <typename... ARGS>
+  static void TRSM_BATCH(ARGS... args) {
+    PADDLE_THROW(platform::errors::Unimplemented(
+        "cublasDtrsmBatched is not supported on HIP platform."));
+  }
 };
 
 template <>
@@ -730,6 +742,32 @@ void Blas<platform::CUDADeviceContext>::BatchedGETRS(
                            batch_size);
   });
 }
+
+template <>
+template <typename T>
+void Blas<platform::CUDADeviceContext>::BatchedTRSM(
+    CBLAS_SIDE side, CBLAS_UPLO uplo, CBLAS_TRANSPOSE transA, CBLAS_DIAG diag,
+    int M, int N, T alpha, const T **A, int lda, T **B, int ldb,
+    int batch_size) const {
+  // solve row major `op ( A ) X = α B` by taking it as `X' op ( A' )  =  α B'`
+  // where ' stands for transpose
+  rocblas_side cuSide =
+      (side == CblasLeft) ? rocblas_side_right : rocblas_side_left;
+  rocblas_fill cuUplo =
+      (uplo == CblasLower) ? rocblas_fill_upper : rocblas_fill_lower;
+  // use CUBLAS_OP_C (conjugate transpose) for complex
+  rocblas_operation cuTransA = (transA == CblasNoTrans)
+                                   ? rocblas_operation_none
+                                   : rocblas_operation_transpose;
+  rocblas_diagonal cuDiag =
+      (diag == CblasUnit) ? rocblas_diagonal_unit : rocblas_diagonal_non_unit;
+
+  context_.CublasCall([&](rocblas_handle handle) {
+    CUBlas<T>::TRSM_BATCH(handle, cuSide, cuUplo, cuTransA, cuDiag, N, M,
+                          &alpha, A, lda, B, ldb, batch_size);
+  });
+}
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/math/matrix_solve.cc

@@ -34,6 +34,45 @@ class MatrixSolveFunctor<platform::CPUDeviceContext, T> {
 template class MatrixSolveFunctor<platform::CPUDeviceContext, float>;
 template class MatrixSolveFunctor<platform::CPUDeviceContext, double>;
 
+template <typename T>
+class TriangularSolveFunctor<platform::CPUDeviceContext, T> {
+ public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::Tensor* a, framework::Tensor* b, bool left,
+                  bool upper, bool transpose, bool unitriangular) {
+    CBLAS_SIDE side = left ? CblasLeft : CblasRight;
+    CBLAS_UPLO uplo = upper ? CblasUpper : CblasLower;
+    CBLAS_TRANSPOSE transA = transpose ? CblasTrans : CblasNoTrans;
+    CBLAS_DIAG diag = unitriangular ? CblasUnit : CblasNonUnit;
+
+    const T* a_data = a->data<T>();
+    T* b_data = b->mutable_data<T>(context.GetPlace());
+
+    int a_dim_size = a->dims().size();
+    int b_dim_size = b->dims().size();
+
+    int M = static_cast<int>(b->dims()[b_dim_size - 2]);
+    int N = static_cast<int>(b->dims()[b_dim_size - 1]);
+    auto lda = left ? std::max(1, M) : std::max(1, N);
+    auto ldb = std::max(1, N);
+
+    int batch_size = 1;
+    auto& a_dim = a->dims();
+    for (int i = 0; i < a_dim_size - 2; i++) {
+      batch_size *= a_dim[i];
+    }
+
+    auto blas = math::GetBlas<platform::CPUDeviceContext, T>(context);
+    for (int i = 0; i < batch_size; i++) {
+      blas.TRSM(side, uplo, transA, diag, M, N, T(1), a_data + i * M * M, lda,
+                b_data + i * N * M, ldb);
+    }
+  }
+};
+
+template class TriangularSolveFunctor<platform::CPUDeviceContext, float>;
+template class TriangularSolveFunctor<platform::CPUDeviceContext, double>;
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/math/matrix_solve.cu.cc

@@ -163,6 +163,68 @@ class MatrixSolveFunctor<platform::CUDADeviceContext, T> {
 template class MatrixSolveFunctor<platform::CUDADeviceContext, float>;
 template class MatrixSolveFunctor<platform::CUDADeviceContext, double>;
 
+template <typename T>
+class TriangularSolveFunctor<platform::CUDADeviceContext, T> {
+ public:
+  void operator()(const platform::CUDADeviceContext& context, const Tensor* a,
+                  Tensor* b, bool left, bool upper, bool transpose,
+                  bool unitriangular) {
+    CBLAS_SIDE side = left ? CblasLeft : CblasRight;
+    CBLAS_UPLO uplo = upper ? CblasUpper : CblasLower;
+    CBLAS_TRANSPOSE transA = transpose ? CblasTrans : CblasNoTrans;
+    CBLAS_DIAG diag = unitriangular ? CblasUnit : CblasNonUnit;
+
+    const T* a_data = a->data<T>();
+    T* b_data = b->mutable_data<T>(context.GetPlace());
+
+    int a_dim_size = a->dims().size();
+    int b_dim_size = b->dims().size();
+
+    int M = static_cast<int>(b->dims()[b_dim_size - 2]);
+    int N = static_cast<int>(b->dims()[b_dim_size - 1]);
+    auto lda = left ? std::max(1, M) : std::max(1, N);
+    auto ldb = std::max(1, N);
+
+    int batch_size = 1;
+    auto& a_dim = a->dims();
+    for (int i = 0; i < a_dim_size - 2; i++) {
+      batch_size *= a_dim[i];
+    }
+
+    auto blas = math::GetBlas<platform::CUDADeviceContext, T>(context);
+    if (batch_size <= 8 && M >= 64) {
+      for (auto i = 0; i < batch_size; i++) {
+        blas.TRSM(side, uplo, transA, diag, M, N, static_cast<T>(1.0),
+                  a_data + i * M * M, lda, b_data + i * N * M, ldb);
+      }
+    } else {
+      std::vector<const T*> cpu_ptrs(batch_size * 2);
+      for (int i = 0; i < batch_size; ++i) {
+        cpu_ptrs[i] = a_data + i * M * M;
+        cpu_ptrs[i + batch_size] = b_data + i * M * N;
+      }
+
+      // Copy the addresses of A and tmp_b from host to device.
+      memory::allocation::AllocationPtr tmp_gpu_ptrs_data =
+          memory::Alloc(context, cpu_ptrs.size() * sizeof(T*));
+      memory::Copy(BOOST_GET_CONST(platform::CUDAPlace, context.GetPlace()),
+                   tmp_gpu_ptrs_data->ptr(), platform::CPUPlace(),
+                   static_cast<void*>(cpu_ptrs.data()),
+                   cpu_ptrs.size() * sizeof(T*), context.stream());
+
+      const T** gpu_a_ptrs =
+          reinterpret_cast<const T**>(tmp_gpu_ptrs_data->ptr());
+      T** gpu_b_ptrs =
+          reinterpret_cast<T**>(tmp_gpu_ptrs_data->ptr()) + batch_size;
+      blas.BatchedTRSM(side, uplo, transA, diag, M, N, static_cast<T>(1.0),
+                       gpu_a_ptrs, lda, gpu_b_ptrs, ldb, batch_size);
+    }
+  }
+};
+
+template class TriangularSolveFunctor<platform::CUDADeviceContext, float>;
+template class TriangularSolveFunctor<platform::CUDADeviceContext, double>;
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/math/matrix_solve.h

@@ -117,6 +117,14 @@ class MatrixSolveFunctor {
                   const framework::Tensor& b, framework::Tensor* out);
 };
 
+template <typename DeviceContext, typename T>
+class TriangularSolveFunctor {
+ public:
+  void operator()(const DeviceContext& context, const framework::Tensor* a,
+                  framework::Tensor* b, bool left, bool upper, bool transpose,
+                  bool unitriangular);
+};
+
 }  // namespace math
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/solve_op.h

@@ -49,9 +49,9 @@ struct IdentityFunctor {
 };
 
 template <typename DeviceContext, typename T>
-void ReduceSumForSolveGrad(const Tensor* input, Tensor* output,
-                           const std::vector<int>& reduce_dims, bool keep_dim,
-                           const paddle::framework::ExecutionContext& ctx) {
+void ReduceSumForSolve(const Tensor* input, Tensor* output,
+                       const std::vector<int>& reduce_dims, bool keep_dim,
+                       const paddle::framework::ExecutionContext& ctx) {
 #if defined(__NVCC__) || defined(__HIPCC__)
   auto stream = ctx.cuda_device_context().stream();
   TensorReduce<T, T, cub::Sum, IdentityFunctor>(*input, output, reduce_dims,
@@ -185,36 +185,6 @@ static std::vector<int64_t> get_broadcast_batch_portion(
   return batchPortion;
 }
 
-// necessary check before expand operation
-static void expand_check(const Tensor& arg1,
-                         std::vector<int64_t> expand_shape) {
-  auto rank = arg1.dims().size();
-  PADDLE_ENFORCE_GE(
-      rank, 1, platform::errors::InvalidArgument(
-                   "The rank of the input 'X' for expand must be positive, "
-                   "but the value received is %d.",
-                   rank));
-  PADDLE_ENFORCE_LE(
-      rank, MAX_RANK_SUPPORTED,
-      platform::errors::InvalidArgument(
-          "The rank of the input 'X' for expand must be less than "
-          "or equal to %d, but the value received is %d.",
-          MAX_RANK_SUPPORTED, rank));
-  auto shape_size = static_cast<int>(expand_shape.size());
-  PADDLE_ENFORCE_GE(
-      shape_size, rank,
-      platform::errors::InvalidArgument(
-          "The number (%d) of elements of 'shape' for expand must be "
-          "greater than or equal to the rank (%d) of the input 'X'.",
-          shape_size, rank));
-  PADDLE_ENFORCE_LE(
-      shape_size, MAX_RANK_SUPPORTED,
-      platform::errors::InvalidArgument(
-          "The number (%d) of elements of 'shape' for expand must be "
-          "less than or equal to %d.",
-          shape_size, MAX_RANK_SUPPORTED));
-}
-
 // broadcast the batch dimensions of tensor x and tensor y.
 static inline std::tuple<std::vector<int64_t>, std::vector<int64_t>>
 get_broadcast_dims(const Tensor& x, const Tensor& y) {
@@ -246,15 +216,13 @@ get_broadcast_dims(const Tensor& x, const Tensor& y) {
 }
 
 template <int Rank, typename T, typename DeviceContext>
-void tensor_expand(const framework::ExecutionContext& context,
-                   const Tensor& arg1, Tensor* out0,
-                   std::vector<int64_t> expand_size) {
-  auto in_dims = arg1.dims();
-  auto expand_shape = expand_size;
-  auto vec_in_dims = framework::vectorize<int>(in_dims);
+void expand_impl(const DeviceContext& context, const Tensor& in, Tensor* out,
+                 const std::vector<int64_t>& expand_shape) {
+  auto vec_in_dims = framework::vectorize<int>(in.dims());
   auto diff = expand_shape.size() - vec_in_dims.size();
   vec_in_dims.insert(vec_in_dims.begin(), diff, 1);
   std::vector<int> repeat_times(vec_in_dims.size());
+
   for (size_t i = 0; i < vec_in_dims.size(); ++i) {
     PADDLE_ENFORCE_NE(
         expand_shape[i], 0,
@@ -301,12 +269,11 @@ void tensor_expand(const framework::ExecutionContext& context,
     out_dims[i] *= repeat_times[i];
   }
 
-  out0->Resize(out_dims);
-  auto x = EigenTensor<T, Rank>::From(arg1, new_in_dims);
-  out0->mutable_data<T>(context.GetPlace());
-  auto y = EigenTensor<T, Rank>::From(*out0, out_dims);
-  auto& place =
-      *context.template device_context<DeviceContext>().eigen_device();
+  out->Resize(out_dims);
+  out->mutable_data<T>(context.GetPlace());
+  auto x = EigenTensor<T, Rank>::From(in, new_in_dims);
+  auto y = EigenTensor<T, Rank>::From(*out, out_dims);
+  auto& place = *context.eigen_device();
   // use 32-bit index to speed up
   bool use_32bit_index = y.size() < Eigen::NumTraits<int>::highest();
   if (use_32bit_index) {
@@ -318,6 +285,41 @@ void tensor_expand(const framework::ExecutionContext& context,
   }
 }
 
+template <typename T, typename DeviceContext>
+void TensorExpand(const DeviceContext& context, const Tensor& in, Tensor* out,
+                  const std::vector<int64_t>& expand_shape) {
+  // necessary check before expand operation
+  PADDLE_ENFORCE_GE(expand_shape.size(), in.dims().size(),
+                    platform::errors::InvalidArgument(
+                        "The size of 'expand_shape' (%d) should >= the input "
+                        "Tensor's rank (%d).",
+                        expand_shape.size(), in.dims().size()));
+  PADDLE_ENFORCE_LE(expand_shape.size(), MAX_RANK_SUPPORTED,
+                    platform::errors::InvalidArgument(
+                        "The size of 'expand_shape' (%d) should be <= %d",
+                        expand_shape.size(), MAX_RANK_SUPPORTED));
+  switch (expand_shape.size()) {
+    case 1:
+      expand_impl<1, T, DeviceContext>(context, in, out, expand_shape);
+      break;
+    case 2:
+      expand_impl<2, T, DeviceContext>(context, in, out, expand_shape);
+      break;
+    case 3:
+      expand_impl<3, T, DeviceContext>(context, in, out, expand_shape);
+      break;
+    case 4:
+      expand_impl<4, T, DeviceContext>(context, in, out, expand_shape);
+      break;
+    case 5:
+      expand_impl<5, T, DeviceContext>(context, in, out, expand_shape);
+      break;
+    case 6:
+      expand_impl<6, T, DeviceContext>(context, in, out, expand_shape);
+      break;
+  }
+}
+
 template <typename DeviceContext, typename T>
 static void linalg_solve(const framework::ExecutionContext& context,
                          const framework::Tensor* x, const framework::Tensor* y,
@@ -356,69 +358,11 @@ static void linalg_solve(const framework::ExecutionContext& context,
   std::tie(x_broadcast_dims, y_broadcast_dims) =
       get_broadcast_dims(tmp_x, tmp_y);
 
-  expand_check(tmp_x, x_broadcast_dims);
-  expand_check(tmp_y, y_broadcast_dims);
-
   Tensor tmp_x_bc;
-  Tensor tmp_y_bc;
-  auto tmp_x_rank = tmp_x.dims().size();
-  auto tmp_y_rank = tmp_y.dims().size();
-
-  auto rank_0 = std::max(tmp_x_rank, static_cast<int>(x_broadcast_dims.size()));
-  switch (rank_0) {
-    case 1:
-      tensor_expand<1, T, DeviceContext>(context, tmp_x, &tmp_x_bc,
-                                         x_broadcast_dims);
-      break;
-    case 2:
-      tensor_expand<2, T, DeviceContext>(context, tmp_x, &tmp_x_bc,
-                                         x_broadcast_dims);
-      break;
-    case 3:
-      tensor_expand<3, T, DeviceContext>(context, tmp_x, &tmp_x_bc,
-                                         x_broadcast_dims);
-      break;
-    case 4:
-      tensor_expand<4, T, DeviceContext>(context, tmp_x, &tmp_x_bc,
-                                         x_broadcast_dims);
-      break;
-    case 5:
-      tensor_expand<5, T, DeviceContext>(context, tmp_x, &tmp_x_bc,
-                                         x_broadcast_dims);
-      break;
-    case 6:
-      tensor_expand<6, T, DeviceContext>(context, tmp_x, &tmp_x_bc,
-                                         x_broadcast_dims);
-      break;
-  }
+  TensorExpand<T, DeviceContext>(dev_ctx, tmp_x, &tmp_x_bc, x_broadcast_dims);
 
-  auto rank_1 = std::max(tmp_y_rank, static_cast<int>(y_broadcast_dims.size()));
-  switch (rank_1) {
-    case 1:
-      tensor_expand<1, T, DeviceContext>(context, tmp_y, &tmp_y_bc,
-                                         y_broadcast_dims);
-      break;
-    case 2:
-      tensor_expand<2, T, DeviceContext>(context, tmp_y, &tmp_y_bc,
-                                         y_broadcast_dims);
-      break;
-    case 3:
-      tensor_expand<3, T, DeviceContext>(context, tmp_y, &tmp_y_bc,
-                                         y_broadcast_dims);
-      break;
-    case 4:
-      tensor_expand<4, T, DeviceContext>(context, tmp_y, &tmp_y_bc,
-                                         y_broadcast_dims);
-      break;
-    case 5:
-      tensor_expand<5, T, DeviceContext>(context, tmp_y, &tmp_y_bc,
-                                         y_broadcast_dims);
-      break;
-    case 6:
-      tensor_expand<6, T, DeviceContext>(context, tmp_y, &tmp_y_bc,
-                                         y_broadcast_dims);
-      break;
-  }
+  Tensor tmp_y_bc;
+  TensorExpand<T, DeviceContext>(dev_ctx, tmp_y, &tmp_y_bc, y_broadcast_dims);
 
   auto x_dim = x->dims();
   auto y_dim = y->dims();
@@ -658,8 +602,8 @@ class SolveGradKernel : public framework::OpKernel<T> {
           if (dy_help.dims().size() != dy->dims().size()) {
             keep_dim = false;
           }
-          ReduceSumForSolveGrad<DeviceContext, T>(&dy_help, dy, dy_reduce_dims,
-                                                  keep_dim, ctx);
+          ReduceSumForSolve<DeviceContext, T>(&dy_help, dy, dy_reduce_dims,
+                                              keep_dim, ctx);
         }
         dy->Resize(y->dims());
       }
@@ -708,8 +652,8 @@ class SolveGradKernel : public framework::OpKernel<T> {
           if (dx_help.dims().size() != dx->dims().size()) {
             keep_dim = false;
           }
-          ReduceSumForSolveGrad<DeviceContext, T>(&dx_help, dx, dx_reduce_dims,
-                                                  keep_dim, ctx);
+          ReduceSumForSolve<DeviceContext, T>(&dx_help, dx, dx_reduce_dims,
+                                              keep_dim, ctx);
         }
         dx->Resize(input->dims());
       }

paddle/fluid/operators/triangular_solve_op.cc

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/triangular_solve_op.h"
+#include "paddle/fluid/operators/solve_op.h"
+
+namespace paddle {
+namespace operators {
+
+class TriangularSolveOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "TriangularSolve");
+    OP_INOUT_CHECK(ctx->HasInput("Y"), "Input", "Y", "TriangularSolve");
+    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "TriangularSolve");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+
+    auto x_dims_n = x_dims.size();
+    auto y_dims_n = y_dims.size();
+
+    PADDLE_ENFORCE_GE(
+        x_dims_n, 2, platform::errors::InvalidArgument(
+                         "The input tensor X's dimensions of TriangularSolveOp "
+                         "should be >= 2. But received X's "
+                         "dimensions = %d, X's shape = [%s]",
+                         x_dims.size(), x_dims));
+
+    PADDLE_ENFORCE_GE(
+        y_dims_n, 2, platform::errors::InvalidArgument(
+                         "The input tensor Y's dimensions of TriangularSolveOp "
+                         "should be >=2. But received Y's "
+                         "dimensions = %d, Y's shape = [%s]",
+                         y_dims.size(), y_dims));
+
+    PADDLE_ENFORCE_EQ(x_dims[x_dims_n - 2], x_dims[x_dims_n - 1],
+                      platform::errors::InvalidArgument(
+                          "The inner-most 2 dimensions of Input(X) all should "
+                          "be square matrices "
+                          "But received X's shape[-2] = %d and shape[-1] = %d.",
+                          x_dims[x_dims_n - 2], x_dims[x_dims_n - 1]));
+
+    std::vector<int64_t> x_dims_vec = paddle::framework::vectorize(x_dims);
+    std::vector<int64_t> y_dims_vec = paddle::framework::vectorize(y_dims);
+
+    std::vector<int64_t> x_dims_vec_cut(x_dims_vec.begin(),
+                                        x_dims_vec.end() - 2);
+    std::vector<int64_t> y_dims_vec_cut(y_dims_vec.begin(),
+                                        y_dims_vec.end() - 2);
+
+    std::vector<int64_t> expand_batch_portion =
+        get_broadcast_batch_portion(x_dims_vec_cut, y_dims_vec_cut);
+
+    std::vector<int64_t> y_broadcast_dims({expand_batch_portion});
+    y_broadcast_dims.insert(y_broadcast_dims.end(), {y_dims_vec[y_dims_n - 2],
+                                                     y_dims_vec[y_dims_n - 1]});
+
+    // dim of 'Out' is the same with 'Y' after broadcast
+    ctx->SetOutputDim("Out", framework::make_ddim(y_broadcast_dims));
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const {
+    return framework::OpKernelType(
+        OperatorWithKernel::IndicateVarDataType(ctx, "X"), ctx.GetPlace());
+  }
+};
+
+class TriangularSolveOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "(Tensor), The first input tensor of triangular solve op, which "
+             "is the triangular coefficient matrix.");
+    AddInput("Y",
+             "(Tensor), The second input tensor of triangular solve op, which "
+             "is multiple right-hand.");
+    AddOutput("Out", "(Tensor), The solution tensor of triangular solve op.");
+    AddAttr<bool>("upper",
+                  "whether to solve the upper-triangular or the "
+                  "lower-triangular system of equations")
+        .SetDefault(true);
+    AddAttr<bool>("transpose", "whether X should be transposed firstly.")
+        .SetDefault(false);
+    AddAttr<bool>("unitriangular", "whether X is unit triangular.")
+        .SetDefault(false);
+    AddComment(R"DOC(
+          Triangular Solve Operator.
+          This operator is used to computes the solution of equations with a triangular coefficient matrix.
+
+          The equation is:
+          $$Out = X^-1 * Y$$
+)DOC");
+  }
+};
+
+class TriangularSolveOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string>& GetInputOutputWithSameType()
+      const override {
+    static std::unordered_map<std::string, std::string> m{{"X", /*->*/ "Out"}};
+    return m;
+  }
+};
+
+class TriangularSolveGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "triangular_solve");
+    OP_INOUT_CHECK(ctx->HasInput("Y"), "Input", "Y", "triangular_solve");
+    OP_INOUT_CHECK(ctx->HasInput("Out"), "Input", "Out", "triangular_solve");
+    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
+                   "Out@GRAD", "triangular_solve");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+
+    auto x_grad_name = framework::GradVarName("X");
+    auto y_grad_name = framework::GradVarName("Y");
+
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+    if (ctx->HasOutput(y_grad_name)) {
+      ctx->SetOutputDim(y_grad_name, y_dims);
+    }
+  }
+};
+
+template <typename T>
+class TriangularSolveOpGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> retv) const override {
+    retv->SetType("triangular_solve_grad");
+    retv->SetInput("X", this->Input("X"));
+    retv->SetInput("Y", this->Input("Y"));
+    retv->SetInput("Out", this->Output("Out"));
+    retv->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+
+    retv->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    retv->SetOutput(framework::GradVarName("Y"), this->InputGrad("Y"));
+    retv->SetAttrMap(this->Attrs());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(triangular_solve, ops::TriangularSolveOp,
+                  ops::TriangularSolveOpMaker,
+                  ops::TriangularSolveOpInferVarType,
+                  ops::TriangularSolveOpGradMaker<paddle::framework::OpDesc>,
+                  ops::TriangularSolveOpGradMaker<paddle::imperative::OpBase>);
+
+REGISTER_OPERATOR(triangular_solve_grad, ops::TriangularSolveGradOp);
+
+REGISTER_OP_CPU_KERNEL(
+    triangular_solve,
+    ops::TriangularSolveKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::TriangularSolveKernel<paddle::platform::CPUDeviceContext, double>);
+
+REGISTER_OP_CPU_KERNEL(
+    triangular_solve_grad,
+    ops::TriangularSolveGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::TriangularSolveGradKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/triangular_solve_op.cu

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/reduce_ops/reduce_functor_op.h"
+#include "paddle/fluid/operators/reduce_ops/reduce_op.h"
+#include "paddle/fluid/operators/triangular_solve_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+struct MatrixReduceSumFunctor<platform::CUDADeviceContext, T> {
+  void operator()(const Tensor& in, Tensor* out,
+                  const framework::ExecutionContext& ctx) {
+    // For example: in's dim = [5, 3, 2, 7, 3] ; out's dim = [3, 1, 7, 3]
+    // out_reduce_dim should be [0, 2]
+    const std::vector<std::int64_t> in_dims = framework::vectorize(in.dims());
+    auto in_size = in_dims.size();
+    const std::vector<std::int64_t> out_dims =
+        framework::vectorize(out->dims());
+    auto out_size = out_dims.size();
+
+    std::vector<std::int64_t> out_bst_dims(in_size);
+
+    std::fill(out_bst_dims.data(), out_bst_dims.data() + in_size - out_size, 1);
+    std::copy(out_dims.data(), out_dims.data() + out_size,
+              out_bst_dims.data() + in_size - out_size);
+
+    std::vector<int> out_reduce_dims;
+    for (size_t idx = 0; idx <= in_size - 3; idx++) {
+      if (in_dims[idx] != 1 && out_bst_dims[idx] == 1) {
+        out_reduce_dims.push_back(idx);
+      }
+    }
+    gpuStream_t stream = ctx.cuda_device_context().stream();
+    TensorReduceFunctorImpl<T, T, CustomSum>(in, out, out_reduce_dims, stream);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    triangular_solve,
+    ops::TriangularSolveKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TriangularSolveKernel<paddle::platform::CUDADeviceContext, double>);
+
+REGISTER_OP_CUDA_KERNEL(
+    triangular_solve_grad,
+    ops::TriangularSolveGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::TriangularSolveGradKernel<paddle::platform::CUDADeviceContext,
+                                   double>);

paddle/fluid/operators/triangular_solve_op.h

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "glog/logging.h"
+#include "paddle/fluid/framework/ddim.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/complex_functors.h"
+#include "paddle/fluid/operators/reduce_ops/reduce_op.h"
+#include "paddle/fluid/operators/solve_op.h"
+#include "paddle/fluid/operators/tril_triu_op.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+static void triangular_solve(const DeviceContext& context, const Tensor& x,
+                             const Tensor& y, Tensor* out, bool upper,
+                             bool transpose, bool unitriangular) {
+  // Tensor broadcast use eigen
+  std::vector<int64_t> x_bst_dims_vec;
+  std::vector<int64_t> y_bst_dims_vec;
+  std::tie(x_bst_dims_vec, y_bst_dims_vec) = get_broadcast_dims(x, y);
+
+  Tensor x_bst(x.type());
+  TensorExpand<T, DeviceContext>(context, x, &x_bst, x_bst_dims_vec);
+
+  Tensor y_bst(y.type());
+  TensorExpand<T, DeviceContext>(context, y, &y_bst, y_bst_dims_vec);
+
+  // TriangularSolveFunctor performs calculations in-place
+  // x_clone should be a copy of 'x' after broadcast
+  // out should be a copy of 'y' after broadcast
+  Tensor x_clone(x.type());
+  x_clone.Resize(framework::make_ddim(x_bst_dims_vec));
+  x_clone.mutable_data<T>(context.GetPlace());
+  framework::TensorCopy(x_bst, context.GetPlace(), context, &x_clone);
+
+  out->Resize(framework::make_ddim(y_bst_dims_vec));
+  out->mutable_data<T>(context.GetPlace());
+  framework::TensorCopy(y_bst, context.GetPlace(), context, out);
+
+  math::TriangularSolveFunctor<DeviceContext, T> functor;
+  functor(context, &x_clone, out, /*left=*/true, upper, transpose,
+          unitriangular);
+}
+
+template <typename DeviceContext, typename T>
+class MatrixReduceSumFunctor {
+ public:
+  void operator()(const Tensor& input, Tensor* output,
+                  const framework::ExecutionContext& ctx);
+};
+
+template <typename T>
+class MatrixReduceSumFunctor<platform::CPUDeviceContext, T> {
+ public:
+  void operator()(const Tensor& in, Tensor* out,
+                  const framework::ExecutionContext& ctx) {
+    // For example: in's dim = [5, 3, 2, 7, 3] ; out's dim = [3, 1, 7, 3]
+    // out_reduce_dim should be [0, 2]
+    const std::vector<std::int64_t> in_dims = framework::vectorize(in.dims());
+    auto in_size = in_dims.size();
+    const std::vector<std::int64_t> out_dims =
+        framework::vectorize(out->dims());
+    auto out_size = out_dims.size();
+
+    std::vector<std::int64_t> out_bst_dims(in_size);
+
+    std::fill(out_bst_dims.data(), out_bst_dims.data() + in_size - out_size, 1);
+    std::copy(out_dims.data(), out_dims.data() + out_size,
+              out_bst_dims.data() + in_size - out_size);
+    out->Resize(framework::make_ddim(out_bst_dims));
+
+    std::vector<int> out_reduce_dims;
+    for (size_t idx = 0; idx <= in_size - 3; idx++) {
+      if (in_dims[idx] != 1 && out_bst_dims[idx] == 1) {
+        out_reduce_dims.push_back(idx);
+      }
+    }
+
+    ReduceKernelFunctor<platform::CPUDeviceContext, T, SumFunctor>(
+        &in, out, out_reduce_dims, true, false, ctx)
+        .template apply<T>();
+    out->Resize(framework::make_ddim(out_dims));
+  }
+};
+
+template <typename DeviceContext, typename T>
+class TriangularSolveKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* x = ctx.Input<framework::Tensor>("X");
+    const auto* y = ctx.Input<framework::Tensor>("Y");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+
+    bool upper = ctx.template Attr<bool>("upper");
+    bool transpose = ctx.template Attr<bool>("transpose");
+    bool unitriangular = ctx.template Attr<bool>("unitriangular");
+
+    const auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    triangular_solve<DeviceContext, T>(dev_ctx, *x, *y, out, upper, transpose,
+                                       unitriangular);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class TriangularSolveGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    const auto* x = ctx.Input<framework::Tensor>("X");
+    const auto* y = ctx.Input<framework::Tensor>("Y");
+    const auto* out = ctx.Input<framework::Tensor>("Out");
+    const auto* dout =
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+
+    auto* dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
+
+    bool upper = ctx.template Attr<bool>("upper");
+    bool transpose = ctx.template Attr<bool>("transpose");
+    bool unitriangular = ctx.template Attr<bool>("unitriangular");
+
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+
+    std::vector<int64_t> x_bst_dims_vec;
+    std::vector<int64_t> y_bst_dims_vec;
+    std::tie(x_bst_dims_vec, y_bst_dims_vec) = get_broadcast_dims(*x, *y);
+
+    Tensor dy_bst(y->type());
+    if (dy) {
+      dy->mutable_data<T>(y->dims(), dev_ctx.GetPlace());
+      dy_bst.Resize(framework::make_ddim(y_bst_dims_vec));
+      dy_bst.mutable_data<T>(dev_ctx.GetPlace());
+
+      // calculate x's conjugate for complex
+      Tensor x_conj(x->type());
+      platform::ForRange<DeviceContext> x_for_range(dev_ctx, x->numel());
+      math::ConjFunctor<T> x_functor(
+          x->data<T>(), x->numel(),
+          x_conj.mutable_data<T>(x->dims(), dev_ctx.GetPlace()));
+      x_for_range(x_functor);
+
+      // reuse forward to get dy_bst, and the result has been broadcated.
+      triangular_solve<DeviceContext, T>(dev_ctx, x_conj, *dout, &dy_bst, upper,
+                                         !transpose, unitriangular);
+
+      if (dy_bst.dims() == dy->dims()) {
+        framework::TensorCopy(dy_bst, dev_ctx.GetPlace(), dev_ctx, dy);
+      } else {
+        MatrixReduceSumFunctor<DeviceContext, T> functor;
+        functor(dy_bst, dy, ctx);
+        dy->Resize(y->dims());
+      }
+    }
+
+    Tensor dx_bst(x->type());
+    if (dx) {
+      dx->mutable_data<T>(x->dims(), dev_ctx.GetPlace());
+      dx_bst.Resize(framework::make_ddim(x_bst_dims_vec));
+      dx_bst.mutable_data<T>(dev_ctx.GetPlace());
+
+      // calculate out's conjugate for complex
+      Tensor out_conj(out->type());
+      platform::ForRange<DeviceContext> out_for_range(dev_ctx, out->numel());
+      math::ConjFunctor<T> out_functor(
+          out->data<T>(), out->numel(),
+          out_conj.mutable_data<T>(out->dims(), dev_ctx.GetPlace()));
+      out_for_range(out_functor);
+
+      auto blas = math::GetBlas<DeviceContext, T>(ctx);
+      if (transpose) {
+        auto mat_dim_a =
+            math::CreateMatrixDescriptor(out_conj.dims(), 0, false);
+        auto mat_dim_b = math::CreateMatrixDescriptor(dy_bst.dims(), 0, true);
+        blas.MatMul(out_conj, mat_dim_a, dy_bst, mat_dim_b, static_cast<T>(-1),
+                    &dx_bst, static_cast<T>(0));
+      } else {
+        auto mat_dim_a = math::CreateMatrixDescriptor(dy_bst.dims(), 0, false);
+        auto mat_dim_b = math::CreateMatrixDescriptor(out_conj.dims(), 0, true);
+        blas.MatMul(dy_bst, mat_dim_a, out_conj, mat_dim_b, static_cast<T>(-1),
+                    &dx_bst, static_cast<T>(0));
+      }
+
+      Tensor dx_bst_upper(x->type());
+      // get upper or lower triangular
+      dx_bst_upper.Resize(dx_bst.dims());
+      dx_bst_upper.mutable_data<T>(dev_ctx.GetPlace());
+
+      const auto& dims = dx_bst.dims();
+      const auto H = dims[dims.size() - 2];
+      const auto W = dims[dims.size() - 1];
+      platform::ForRange<DeviceContext> x_for_range(dev_ctx, dx_bst.numel());
+      TrilTriuCompute<T> tril_triu_computer(dx_bst.data<T>(), unitriangular,
+                                            !upper, H, W,
+                                            dx_bst_upper.data<T>());
+      x_for_range(tril_triu_computer);
+
+      if (dx_bst_upper.dims() == dx->dims()) {
+        framework::TensorCopy(dx_bst_upper, dev_ctx.GetPlace(), dev_ctx, dx);
+      } else {
+        MatrixReduceSumFunctor<DeviceContext, T> functor;
+        functor(dx_bst_upper, dx, ctx);
+        dx->Resize(x->dims());
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/fluid/platform/dynload/cublas.h

@@ -75,6 +75,8 @@ extern void *cublas_dso_handle;
   __macro(cublasDgeam);                   \
   __macro(cublasStrsm_v2);                \
   __macro(cublasDtrsm_v2);                \
+  __macro(cublasCtrsm_v2);                \
+  __macro(cublasZtrsm_v2);                \
   __macro(cublasCreate_v2);               \
   __macro(cublasDestroy_v2);              \
   __macro(cublasSetStream_v2);            \
@@ -84,6 +86,10 @@ extern void *cublas_dso_handle;
   __macro(cublasDgemmBatched);            \
   __macro(cublasCgemmBatched);            \
   __macro(cublasZgemmBatched);            \
+  __macro(cublasStrsmBatched);            \
+  __macro(cublasDtrsmBatched);            \
+  __macro(cublasCtrsmBatched);            \
+  __macro(cublasZtrsmBatched);            \
   __macro(cublasSgetrfBatched);           \
   __macro(cublasSgetriBatched);           \
   __macro(cublasDgetrfBatched);           \

paddle/fluid/platform/dynload/mklml.h

@@ -25,7 +25,7 @@ namespace platform {
 namespace dynload {
 
 extern std::once_flag mklml_dso_flag;
-extern void* mklml_dso_handle;
+extern void *mklml_dso_handle;
 
 /**
  * The following macro definition can generate structs
@@ -40,7 +40,7 @@ extern void* mklml_dso_handle;
       std::call_once(mklml_dso_flag, []() {                                \
         mklml_dso_handle = paddle::platform::dynload::GetMKLMLDsoHandle(); \
       });                                                                  \
-      static void* p_##_name = dlsym(mklml_dso_handle, #__name);           \
+      static void *p_##_name = dlsym(mklml_dso_handle, #__name);           \
       return reinterpret_cast<mklmlFunc>(p_##_name)(args...);              \
     }                                                                      \
   };                                                                       \
@@ -67,6 +67,8 @@ extern void* mklml_dso_handle;
   __macro(cblas_zgemv);             \
   __macro(cblas_strsm);             \
   __macro(cblas_dtrsm);             \
+  __macro(cblas_ctrsm);             \
+  __macro(cblas_ztrsm);             \
   __macro(cblas_sgemm_alloc);       \
   __macro(cblas_dgemm_alloc);       \
   __macro(cblas_sgemm_pack);        \

python/paddle/fluid/tests/unittests/test_triangular_solve_op.py

+#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.w
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+
+import sys
+sys.path.append("..")
+import paddle
+from op_test import OpTest
+import paddle.fluid as fluid
+from paddle.fluid import Program, program_guard, core
+
+paddle.enable_static()
+
+
+# 2D + 2D , test 'upper'
+class TestTriangularSolveOp(OpTest):
+    """
+    case 1
+    """
+
+    def config(self):
+        self.x_shape = [12, 12]
+        self.y_shape = [12, 10]
+        self.upper = True
+        self.transpose = False
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        self.output = np.linalg.solve(
+            np.triu(self.inputs['X']), self.inputs['Y'])
+
+    def setUp(self):
+        self.op_type = "triangular_solve"
+        self.config()
+
+        self.inputs = {
+            'X': np.random.random(self.x_shape).astype(self.dtype),
+            'Y': np.random.random(self.y_shape).astype(self.dtype)
+        }
+        self.attrs = {
+            'upper': self.upper,
+            'transpose': self.transpose,
+            'unitriangular': self.unitriangular,
+        }
+        self.set_output()
+        self.outputs = {'Out': self.output}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out')
+
+
+# 2D(broadcast) + 3D, test 'transpose'
+class TestTriangularSolveOp2(TestTriangularSolveOp):
+    """
+    case 2
+    """
+
+    def config(self):
+        self.x_shape = [10, 10]
+        self.y_shape = [3, 10, 8]
+        self.upper = False
+        self.transpose = True
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.tril(self.inputs['X']).transpose(1, 0)
+        y = self.inputs['Y']
+        self.output = np.linalg.solve(x, y)
+
+
+# 3D(broadcast) + 3D
+class TestTriangularSolveOp3(TestTriangularSolveOp):
+    """
+    case 3
+    """
+
+    def config(self):
+        self.x_shape = [1, 10, 10]
+        self.y_shape = [6, 10, 12]
+        self.upper = False
+        self.transpose = False
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.tril(self.inputs['X'])
+        y = self.inputs['Y']
+        self.output = np.linalg.solve(x, y)
+
+
+# 3D + 3D(broadcast), test 'transpose'
+class TestTriangularSolveOp4(TestTriangularSolveOp):
+    """
+    case 4
+    """
+
+    def config(self):
+        self.x_shape = [3, 10, 10]
+        self.y_shape = [1, 10, 12]
+        self.upper = True
+        self.transpose = True
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.triu(self.inputs['X']).transpose(0, 2, 1)
+        y = self.inputs['Y']
+        self.output = np.linalg.solve(x, y)
+
+
+# 2D + 2D , test 'unitriangular' specially
+class TestTriangularSolveOp5(TestTriangularSolveOp):
+    """
+    case 5
+    """
+
+    def config(self):
+        self.x_shape = [10, 10]
+        self.y_shape = [10, 10]
+        self.upper = True
+        self.transpose = False
+        self.unitriangular = True
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.triu(self.inputs['X'])
+        np.fill_diagonal(x, 1.)
+        y = self.inputs['Y']
+        self.output = np.linalg.solve(x, y)
+
+    def test_check_grad_normal(self):
+        x = np.triu(self.inputs['X'])
+        np.fill_diagonal(x, 1.)
+        grad_out = np.ones([10, 10]).astype('float64')
+        grad_y = np.linalg.solve(x.transpose(1, 0), grad_out)
+
+        grad_x = -np.matmul(grad_y, self.output.transpose(1, 0))
+        grad_x = np.triu(grad_x)
+        np.fill_diagonal(grad_x, 0.)
+
+        self.check_grad(
+            ['X', 'Y'],
+            'Out',
+            user_defined_grads=[grad_x, grad_y],
+            user_defined_grad_outputs=[grad_out])
+
+
+# 4D(broadcast) + 4D(broadcast)
+class TestTriangularSolveOp6(TestTriangularSolveOp):
+    """
+    case 6
+    """
+
+    def config(self):
+        self.x_shape = [1, 3, 10, 10]
+        self.y_shape = [2, 1, 10, 5]
+        self.upper = False
+        self.transpose = False
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.tril(self.inputs['X'])
+        y = self.inputs['Y']
+        self.output = np.linalg.solve(x, y)
+
+
+# 3D(broadcast) + 4D(broadcast), test 'upper'
+class TestTriangularSolveOp7(TestTriangularSolveOp):
+    """
+    case 7
+    """
+
+    def config(self):
+        self.x_shape = [2, 10, 10]
+        self.y_shape = [5, 1, 10, 2]
+        self.upper = True
+        self.transpose = True
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.triu(self.inputs['X']).transpose(0, 2, 1)
+        y = self.inputs['Y']
+        self.output = np.linalg.solve(x, y)
+
+
+# 3D(broadcast) + 5D
+class TestTriangularSolveOp8(TestTriangularSolveOp):
+    """
+    case 8
+    """
+
+    def config(self):
+        self.x_shape = [12, 3, 3]
+        self.y_shape = [2, 3, 12, 3, 2]
+        self.upper = False
+        self.transpose = False
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.tril(self.inputs['X'])
+        y = self.inputs['Y']
+        self.output = np.linalg.solve(x, y)
+
+
+# 5D + 4D(broadcast)
+class TestTriangularSolveOp9(TestTriangularSolveOp):
+    """
+    case 9
+    """
+
+    def config(self):
+        self.x_shape = [2, 4, 2, 3, 3]
+        self.y_shape = [4, 1, 3, 10]
+        self.upper = False
+        self.transpose = False
+        self.unitriangular = False
+        self.dtype = "float64"
+
+    def set_output(self):
+        x = np.tril(self.inputs['X'])
+        y = self.inputs['Y']
+        self.output = np.matmul(np.linalg.inv(x), y)
+
+
+class TestTriangularSolveAPI(unittest.TestCase):
+    def setUp(self):
+        np.random.seed(2021)
+        self.place = [paddle.CPUPlace()]
+        self.dtype = "float64"
+        if core.is_compiled_with_cuda():
+            self.place.append(paddle.CUDAPlace(0))
+
+    def check_static_result(self, place):
+        with fluid.program_guard(fluid.Program(), fluid.Program()):
+            x = fluid.data(name="x", shape=[3, 3], dtype=self.dtype)
+            y = fluid.data(name="y", shape=[3, 2], dtype=self.dtype)
+            z = paddle.linalg.triangular_solve(x, y)
+
+            x_np = np.random.random([3, 3]).astype(self.dtype)
+            y_np = np.random.random([3, 2]).astype(self.dtype)
+            z_np = np.linalg.solve(np.triu(x_np), y_np)
+
+            exe = fluid.Executor(place)
+            fetches = exe.run(fluid.default_main_program(),
+                              feed={"x": x_np,
+                                    "y": y_np},
+                              fetch_list=[z])
+            self.assertTrue(np.allclose(fetches[0], z_np))
+
+    def test_static(self):
+        for place in self.place:
+            self.check_static_result(place=place)
+
+    def test_dygraph(self):
+        def run(place):
+            paddle.disable_static(place)
+            x_np = np.random.random([3, 3]).astype(self.dtype)
+            y_np = np.random.random([3, 2]).astype(self.dtype)
+            z_np = np.linalg.solve(np.tril(x_np), y_np)
+
+            x = paddle.to_tensor(x_np)
+            y = paddle.to_tensor(y_np)
+            z = paddle.linalg.triangular_solve(x, y, upper=False)
+
+            self.assertTrue(np.allclose(z_np, z.numpy()))
+            self.assertEqual(z_np.shape, z.numpy().shape)
+            paddle.enable_static()
+
+        for place in self.place:
+            run(place)
+
+
+class TestTriangularSolveOpError(unittest.TestCase):
+    def test_errors(self):
+        with program_guard(Program(), Program()):
+            # The input type of solve_op must be Variable.
+            x1 = fluid.create_lod_tensor(
+                np.array([[-1]]), [[1]], fluid.CPUPlace())
+            y1 = fluid.create_lod_tensor(
+                np.array([[-1]]), [[1]], fluid.CPUPlace())
+            self.assertRaises(TypeError, paddle.linalg.triangular_solve, x1, y1)
+
+            # The data type of input must be float32 or float64.        
+            x2 = fluid.data(name="x2", shape=[30, 30], dtype="bool")
+            y2 = fluid.data(name="y2", shape=[30, 10], dtype="bool")
+            self.assertRaises(TypeError, paddle.linalg.triangular_solve, x2, y2)
+
+            x3 = fluid.data(name="x3", shape=[30, 30], dtype="int32")
+            y3 = fluid.data(name="y3", shape=[30, 10], dtype="int32")
+            self.assertRaises(TypeError, paddle.linalg.triangular_solve, x3, y3)
+
+            x4 = fluid.data(name="x4", shape=[30, 30], dtype="float16")
+            y4 = fluid.data(name="y4", shape=[30, 10], dtype="float16")
+            self.assertRaises(TypeError, paddle.linalg.triangular_solve, x4, y4)
+
+            # The number of dimensions of input'X must be >= 2.
+            x5 = fluid.data(name="x5", shape=[30], dtype="float64")
+            y5 = fluid.data(name="y5", shape=[30, 30], dtype="float64")
+            self.assertRaises(ValueError, paddle.linalg.triangular_solve, x5,
+                              y5)
+
+            # The number of dimensions of input'Y must be >= 2.
+            x6 = fluid.data(name="x6", shape=[30, 30], dtype="float64")
+            y6 = fluid.data(name="y6", shape=[30], dtype="float64")
+            self.assertRaises(ValueError, paddle.linalg.triangular_solve, x6,
+                              y6)
+
+            # The inner-most 2 dimensions of input'X should be equal to each other
+            x7 = fluid.data(name="x7", shape=[2, 3, 4], dtype="float64")
+            y7 = fluid.data(name="y7", shape=[2, 4, 3], dtype="float64")
+            self.assertRaises(ValueError, paddle.linalg.triangular_solve, x7,
+                              y7)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/linalg.py

@@ -29,6 +29,7 @@ from .tensor.linalg import eigvalsh
 from .tensor.linalg import det
 from .tensor.linalg import slogdet
 from .tensor.linalg import pinv
+from .tensor.linalg import triangular_solve
 
 __all__ = [
     'cholesky',  #noqa
@@ -47,5 +48,6 @@ __all__ = [
     'eigh',
     'eigvalsh',
     'pinv',
-    'solve'
+    'solve',
+    'triangular_solve',
 ]

python/paddle/tensor/__init__.py

@@ -397,6 +397,7 @@ tensor_method_func  = [ #noqa
            'uniform_',
            'multi_dot',
            'solve',
+           'triangular_solve'
 ]
 
 #this list used in math_op_patch.py for magic_method bind

python/paddle/tensor/linalg.py

@@ -2315,6 +2315,79 @@ def solve(x, y, name=None):
     return out
 
 
+def triangular_solve(x,
+                     y,
+                     upper=True,
+                     transpose=False,
+                     unitriangular=False,
+                     name=None):
+    r"""
+    Computes the solution of a system of equations with a triangular coefficient matrix `x` and
+    multiple right-hand sides `y` .
+
+    Input `x` and `y` is 2D matrices or batches of 2D matrices. If the inputs are batches, the outputs
+    is also batches.
+
+    Args:
+        x (Tensor): The input triangular coefficient matrix. Its shape should be `[*, M, M]`, where `*` is zero or
+            more batch dimensions. Its data type should be float32 or float64.
+        y (Tensor): Multiple right-hand sides of system of equations. Its shape should be `[*, M, K]`, where `*` is 
+            zero or more batch dimensions. Its data type should be float32 or float64.
+        upper (bool, optional): Whether to solve the upper-triangular system of equations (default) or the lower-triangular 
+            system of equations. Default: True.
+        transpose (bool, optional): whether `x` should be transposed before calculation. Default: False.
+        unitriangular (bool, optional): whether `x` is unit triangular. If True, the diagonal elements of `x` are assumed 
+            to be 1 and not referenced from `x` . Default: False.
+        name(str, optional): Name for the operation (optional, default is None).
+            For more information, please refer to :ref:`api_guide_Name`.
+
+    Returns:
+        Tensor: The solution of the system of equations. Its data type should be the same as that of `x`.
+
+    Examples:
+    .. code-block:: python
+
+        # a square system of linear equations:
+        # x1 +   x2  +   x3 = 0
+        #      2*x2  +   x3 = -9
+        #               -x3 = 5
+
+        import paddle
+        import numpy as np
+
+        x = paddle.to_tensor([[1, 1, 1], 
+                              [0, 2, 1],
+                              [0, 0,-1]], dtype="float64")
+        y = paddle.to_tensor([[0], [-9], [5]], dtype="float64")
+        out = paddle.linalg.triangular_solve(x, y, upper=True)
+
+        print(out)
+        # [7, -2, -5]
+    """
+    if in_dygraph_mode():
+        return _C_ops.triangular_solve(x, y, 'upper', upper, 'transpose',
+                                       transpose, 'unitriangular',
+                                       unitriangular)
+
+    inputs = {"X": [x], "Y": [y]}
+    helper = LayerHelper("triangular_solve", **locals())
+    check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'triangular_solve')
+    check_variable_and_dtype(y, 'y', ['float32', 'float64'], 'triangular_solve')
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
+
+    helper.append_op(
+        type='triangular_solve',
+        inputs={'X': x,
+                'Y': y},
+        outputs={'Out': out},
+        attrs={
+            'upper': upper,
+            'transpose': transpose,
+            'unitriangular': unitriangular
+        })
+    return out
+
+
 def eigvalsh(x, UPLO='L', name=None):
     """
     Computes the eigenvalues of a