diff --git a/doc/Doxyfile.in b/doc/Doxyfile.in
index 9069c299e8..931b2d09c6 100644
--- a/doc/Doxyfile.in
+++ b/doc/Doxyfile.in
@@ -814,7 +814,6 @@ WARN_LOGFILE           =
 # Note: If this tag is empty the current directory is searched.
 INPUT                  =  ${DLAF_SOURCE_DIR}/doc/mainpage.dox
 INPUT                  += ${DLAF_SOURCE_DIR}/include
-INPUT                  += ${DLAF_SOURCE_DIR}/src
 
 # This tag can be used to specify the character encoding of the source files
 # that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses
@@ -2149,6 +2148,7 @@ PREDEFINED             =  DLAF_DOXYGEN
 PREDEFINED             += DLAF_ASSERT_ENABLE
 PREDEFINED             += DLAF_ASSERT_MODERATE_ENABLE
 PREDEFINED             += DLAF_ASSERT_HEAVY_ENABLE
+PREDEFINED             += DLAF_WITH_GPU
 
 # If the MACRO_EXPANSION and EXPAND_ONLY_PREDEF tags are set to YES then this
 # tag can be used to specify a list of macro names that should be expanded. The
diff --git a/include/dlaf/auxiliary.h b/include/dlaf/auxiliary.h
index b711c7a99c..bec74a7f29 100644
--- a/include/dlaf/auxiliary.h
+++ b/include/dlaf/auxiliary.h
@@ -9,4 +9,7 @@
 //
 #pragma once
 
+/// @file
+/// Includes all auxiliary algorithms such as norms.
+
 #include "dlaf/auxiliary/norm.h"
diff --git a/include/dlaf/auxiliary/norm.h b/include/dlaf/auxiliary/norm.h
index 91f6848f8d..0d0e96baae 100644
--- a/include/dlaf/auxiliary/norm.h
+++ b/include/dlaf/auxiliary/norm.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include "dlaf/auxiliary/norm/api.h"
 #include "dlaf/blas/enum_output.h"
 #include "dlaf/common/assert.h"
diff --git a/include/dlaf/auxiliary/norm/api.h b/include/dlaf/auxiliary/norm/api.h
index 4f020fc068..21393c81f0 100644
--- a/include/dlaf/auxiliary/norm/api.h
+++ b/include/dlaf/auxiliary/norm/api.h
@@ -27,7 +27,7 @@ struct Norm<Backend::MC, Device::CPU, T> {
                                  Matrix<const T, Device::CPU>& matrix);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_NORM_ETI(KWORD, DATATYPE) KWORD template struct Norm<Backend::MC, Device::CPU, DATATYPE>;
 
 DLAF_NORM_ETI(extern, float)
diff --git a/include/dlaf/blas/enum_output.h b/include/dlaf/blas/enum_output.h
index 63820d8679..7adc873019 100644
--- a/include/dlaf/blas/enum_output.h
+++ b/include/dlaf/blas/enum_output.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <ostream>
 
 #include <blas.hh>
diff --git a/include/dlaf/blas/scal.h b/include/dlaf/blas/scal.h
index da0ddbbfc3..028480482a 100644
--- a/include/dlaf/blas/scal.h
+++ b/include/dlaf/blas/scal.h
@@ -13,13 +13,12 @@
 #include <complex>
 
 /// @file
-
-namespace blas {
-
-/// Provides overloads for mixed real complex variants missing in blaspp.
+/// Provides overloads for mixed real complex variants of scal missing in blaspp.
 /// - csscal
 /// - zdscal
 
+namespace blas {
+
 void scal(std::int64_t n, float a, std::complex<float>* x, std::int64_t incx) noexcept;
 void scal(std::int64_t n, double a, std::complex<double>* x, std::int64_t incx) noexcept;
 
diff --git a/include/dlaf/blas/tile.h b/include/dlaf/blas/tile.h
index 0f36218da8..1d5dafce43 100644
--- a/include/dlaf/blas/tile.h
+++ b/include/dlaf/blas/tile.h
@@ -9,6 +9,9 @@
 //
 #pragma once
 
+/// @file
+/// Provides `Tile` wrappers for BLAS operations.
+
 #include <cstddef>
 
 #include <blas.hh>
@@ -169,16 +172,12 @@ template <Backend B, class T, Device D>
 void gemm(const blas::Op op_a, const blas::Op op_b, const T alpha, const Tile<const T, D>& a,
           const Tile<const T, D>& b, const T beta, const Tile<T, D>& c);
 
-/// \overload gemm
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto gemm(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload gemm
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
@@ -191,16 +190,12 @@ template <Backend B, class T, Device D>
 void hemm(const blas::Side side, const blas::Uplo uplo, const T alpha, const Tile<const T, D>& a,
           const Tile<const T, D>& b, const T beta, const Tile<T, D>& c);
 
-/// \overload hemm
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto hemm(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload hemm
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
@@ -213,16 +208,12 @@ template <Backend B, class T, Device D>
 void her2k(const blas::Uplo uplo, const blas::Op op, const T alpha, const Tile<const T, D>& a,
            const Tile<const T, D>& b, const BaseType<T> beta, const Tile<T, D>& c);
 
-/// \overload her2k
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto her2k(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload her2k
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
@@ -235,16 +226,12 @@ template <Backend B, class T, Device D>
 void herk(const blas::Uplo uplo, const blas::Op op, const BaseType<T> alpha, const Tile<const T, D>& a,
           const BaseType<T> beta, const Tile<T, D>& c);
 
-/// \overload herk
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto herk(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload herk
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
@@ -258,16 +245,12 @@ void trmm(const dlaf::internal::Policy<B>& policy, const blas::Side side, const
           const blas::Op op, const blas::Diag diag, const T alpha, const Tile<const T, D>& a,
           const Tile<T, D>& b);
 
-/// \overload trmm
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto trmm(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload trmm
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
@@ -282,16 +265,12 @@ void trmm3(const dlaf::internal::Policy<B>& policy, const blas::Side side, const
            const blas::Op op, const blas::Diag diag, const T alpha, const Tile<const T, D>& a,
            const Tile<const T, D>& b, const Tile<T, D>& c);
 
-/// \overload trmm3
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto trmm3(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload trmm3
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
@@ -305,16 +284,12 @@ void trsm(const dlaf::internal::Policy<B>& policy, const blas::Side side, const
           const blas::Op op, const blas::Diag diag, const T alpha, const Tile<const T, D>& a,
           const Tile<T, D>& b);
 
-/// \overload trsm
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto trsm(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload trsm
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
diff --git a/include/dlaf/blas/tile_extensions.h b/include/dlaf/blas/tile_extensions.h
index d5d92caaf7..ac8666b608 100644
--- a/include/dlaf/blas/tile_extensions.h
+++ b/include/dlaf/blas/tile_extensions.h
@@ -9,6 +9,9 @@
 //
 #pragma once
 
+/// @file
+/// Provides `Tile` wrappers for extra basic linear algebra operations not covered by BLAS.
+
 #include <blas.hh>
 
 #include "dlaf/blas/tile.h"
@@ -39,16 +42,12 @@ using matrix::Tile;
 template <Backend B, class T, Device D>
 void add(T alpha, const matrix::Tile<const T, D>& tile_b, const matrix::Tile<T, D>& tile_a);
 
-/// \overload add
-///
 /// This overload takes a policy argument and a sender which must send all required arguments for the
 /// algorithm. Returns a sender which signals a connected receiver when the algorithm is done.
 template <Backend B, typename Sender,
           typename = std::enable_if_t<pika::execution::experimental::is_sender_v<Sender>>>
 auto add(const dlaf::internal::Policy<B>& p, Sender&& s);
 
-/// \overload add
-///
 /// This overload partially applies the algorithm with a policy for later use with operator| with a
 /// sender on the left-hand side.
 template <Backend B>
diff --git a/include/dlaf/common/assert.h b/include/dlaf/common/assert.h
index 301db28b5d..bf25b5e678 100644
--- a/include/dlaf/common/assert.h
+++ b/include/dlaf/common/assert.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #include <exception>
 #include <iostream>
 #include <memory>
@@ -130,7 +128,7 @@ void silenceUnusedWarningFor(Args&&...) {}
 
 #define DLAF_STATIC_FAIL(DummyType, Msg) static_assert(sizeof(DummyType) == 0, Msg)
 
-/// Returns a fake object of type T (specified by __VA_ARGS__) dereferencing a null pointer.
+/// Returns a fake object of type T (specified by @p __VA_ARGS__ ) dereferencing a null pointer.
 ///
 /// At runtime the program is exited with an error before dereferencing the null pointer.
 /// This macro is useful when a return statement is needed in unreachable branches,
diff --git a/include/dlaf/common/callable_object.h b/include/dlaf/common/callable_object.h
index aa8cb8a238..22b310b0c7 100644
--- a/include/dlaf/common/callable_object.h
+++ b/include/dlaf/common/callable_object.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 /// Given a function name @fname generates a constexpr object with name fname_o
 /// with type fname_t. The generated type has one static operator() which
 /// transparently forwards all arguments to a call to fname. This macro is
diff --git a/include/dlaf/common/data.h b/include/dlaf/common/data.h
index e2dc27b387..08c6ea6091 100644
--- a/include/dlaf/common/data.h
+++ b/include/dlaf/common/data.h
@@ -23,7 +23,15 @@
 namespace dlaf {
 namespace common {
 
+#ifdef DLAF_DOXYGEN
+
 /// Traits for verifying if the given type is an implementation of the Data concept.
+///
+/// Derive from @p std::true_type or @p std::false_type.
+template <class Data>
+struct is_data {};
+
+#else
 template <class Data, class = void>
 struct is_data : std::false_type {};
 
@@ -37,6 +45,7 @@ struct is_data<
                      std::is_same_v<decltype(data_count(std::declval<const Data&>())), SizeType> &&
                      std::is_same_v<decltype(data_iscontiguous(std::declval<const Data&>())), bool>>>
     : std::true_type {};
+#endif
 
 template <class T>
 inline constexpr bool is_data_v = is_data<T>::value;
diff --git a/include/dlaf/common/data_descriptor.h b/include/dlaf/common/data_descriptor.h
index 28a5c1716b..bea6a53b8d 100644
--- a/include/dlaf/common/data_descriptor.h
+++ b/include/dlaf/common/data_descriptor.h
@@ -138,7 +138,7 @@ struct DataDescriptor {
   SizeType stride_;
 };
 
-/// Helper class for creatig a DataDescriptor from a bounded C-array.
+/// Helper class for creating a DataDescriptor from a bounded C-array.
 template <class T, std::size_t N>
 struct DataDescriptor<T[N]> : DataDescriptor<T> {
   /// Create a Data from a given bounded C-array.
diff --git a/include/dlaf/common/index2d.h b/include/dlaf/common/index2d.h
index 28e2238a44..39df3017ea 100644
--- a/include/dlaf/common/index2d.h
+++ b/include/dlaf/common/index2d.h
@@ -26,7 +26,7 @@ namespace dlaf {
 
 enum class Coord { Row, Col };
 
-// Given a Coord, returns its orthogonal
+/// Given a Coord, returns its orthogonal
 constexpr Coord orthogonal(const Coord rc) {
   return rc == Coord::Row ? Coord::Col : Coord::Row;
 }
diff --git a/include/dlaf/common/pipeline.h b/include/dlaf/common/pipeline.h
index 2b8aa8f300..ccaa12ef2c 100644
--- a/include/dlaf/common/pipeline.h
+++ b/include/dlaf/common/pipeline.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <pika/async_rw_mutex.hpp>
 #include <pika/execution.hpp>
 #include <dlaf/common/assert.h>
diff --git a/include/dlaf/common/round_robin.h b/include/dlaf/common/round_robin.h
index 309742153b..c4ed913aa3 100644
--- a/include/dlaf/common/round_robin.h
+++ b/include/dlaf/common/round_robin.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <vector>
 
 namespace dlaf {
diff --git a/include/dlaf/common/single_threaded_blas.h b/include/dlaf/common/single_threaded_blas.h
index 452f14b0a0..1d925d5b2d 100644
--- a/include/dlaf/common/single_threaded_blas.h
+++ b/include/dlaf/common/single_threaded_blas.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #ifdef DLAF_ASSERT_MODERATE_ENABLE
 #include <thread>
 #endif
diff --git a/include/dlaf/common/timer.h b/include/dlaf/common/timer.h
index 693574a259..d38aeaa6fb 100644
--- a/include/dlaf/common/timer.h
+++ b/include/dlaf/common/timer.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <chrono>
 
 namespace dlaf {
diff --git a/include/dlaf/common/unwrap.h b/include/dlaf/common/unwrap.h
index 8992660288..f427491212 100644
--- a/include/dlaf/common/unwrap.h
+++ b/include/dlaf/common/unwrap.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <pika/async_rw_mutex.hpp>
 #include <pika/execution.hpp>
 
diff --git a/include/dlaf/common/vector.h b/include/dlaf/common/vector.h
index b0bad77175..7174148179 100644
--- a/include/dlaf/common/vector.h
+++ b/include/dlaf/common/vector.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <vector>
 #include "dlaf/types.h"
 
diff --git a/include/dlaf/communication/communicator.h b/include/dlaf/communication/communicator.h
index 948c70e877..2a688630ab 100644
--- a/include/dlaf/communication/communicator.h
+++ b/include/dlaf/communication/communicator.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <memory>
 #include "dlaf/communication/error.h"
 
diff --git a/include/dlaf/communication/datatypes.h b/include/dlaf/communication/datatypes.h
index c5193b1f03..360ec0b851 100644
--- a/include/dlaf/communication/datatypes.h
+++ b/include/dlaf/communication/datatypes.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/communication/error.h"
 
 #include <complex>
diff --git a/include/dlaf/communication/functions_sync.h b/include/dlaf/communication/functions_sync.h
index ff0bebdc6e..0881962588 100644
--- a/include/dlaf/communication/functions_sync.h
+++ b/include/dlaf/communication/functions_sync.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/communication/sync/all_reduce.h"
 #include "dlaf/communication/sync/basic.h"
 #include "dlaf/communication/sync/broadcast.h"
diff --git a/include/dlaf/communication/init.h b/include/dlaf/communication/init.h
index e72881fb36..772bc2d56c 100644
--- a/include/dlaf/communication/init.h
+++ b/include/dlaf/communication/init.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <iostream>
 #include <mutex>
 
diff --git a/include/dlaf/communication/kernels.h b/include/dlaf/communication/kernels.h
index f39cfc28d5..062a2f13a1 100644
--- a/include/dlaf/communication/kernels.h
+++ b/include/dlaf/communication/kernels.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/communication/kernels/all_reduce.h"
 #include "dlaf/communication/kernels/broadcast.h"
 #include "dlaf/communication/kernels/p2p.h"
diff --git a/include/dlaf/communication/type_handler.h b/include/dlaf/communication/type_handler.h
index d4e8f3e6d7..81853c46e0 100644
--- a/include/dlaf/communication/type_handler.h
+++ b/include/dlaf/communication/type_handler.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <type_traits>
 
 #include "dlaf/communication/datatypes.h"
diff --git a/include/dlaf/eigensolver.h b/include/dlaf/eigensolver.h
index 6cffb134ed..48cec0e6fd 100644
--- a/include/dlaf/eigensolver.h
+++ b/include/dlaf/eigensolver.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include "dlaf/eigensolver/band_to_tridiag.h"
 #include "dlaf/eigensolver/bt_band_to_tridiag.h"
 #include "dlaf/eigensolver/bt_reduction_to_band.h"
diff --git a/include/dlaf/eigensolver/band_to_tridiag.h b/include/dlaf/eigensolver/band_to_tridiag.h
index a31d03fbc0..b64e2106c0 100644
--- a/include/dlaf/eigensolver/band_to_tridiag.h
+++ b/include/dlaf/eigensolver/band_to_tridiag.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include "dlaf/common/assert.h"
 #include "dlaf/communication/communicator_grid.h"
diff --git a/include/dlaf/eigensolver/band_to_tridiag/api.h b/include/dlaf/eigensolver/band_to_tridiag/api.h
index f5964fde1d..3450f01977 100644
--- a/include/dlaf/eigensolver/band_to_tridiag/api.h
+++ b/include/dlaf/eigensolver/band_to_tridiag/api.h
@@ -45,7 +45,7 @@ struct BandToTridiag<Backend::MC, D, T> {
                                               Matrix<const T, D>& mat_a) noexcept;
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_EIGENSOLVER_B2T_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct BandToTridiag<BACKEND, DEVICE, DATATYPE>;
 #define DLAF_EIGENSOLVER_B2T_DISTR_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
diff --git a/include/dlaf/eigensolver/band_to_tridiag/mc.h b/include/dlaf/eigensolver/band_to_tridiag/mc.h
index 4a5a23221b..ee0a0c61e3 100644
--- a/include/dlaf/eigensolver/band_to_tridiag/mc.h
+++ b/include/dlaf/eigensolver/band_to_tridiag/mc.h
@@ -28,7 +28,7 @@
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/communication/kernels.h"
 #include "dlaf/eigensolver/band_to_tridiag/api.h"
-#include "dlaf/eigensolver/get_1d_block_size.h"
+#include "dlaf/eigensolver/internal/get_1d_block_size.h"
 #include "dlaf/lapack/gpu/lacpy.h"
 #include "dlaf/lapack/gpu/laset.h"
 #include "dlaf/lapack/tile.h"
diff --git a/include/dlaf/eigensolver/bt_band_to_tridiag.h b/include/dlaf/eigensolver/bt_band_to_tridiag.h
index 0f48cafe48..bf6cf3780e 100644
--- a/include/dlaf/eigensolver/bt_band_to_tridiag.h
+++ b/include/dlaf/eigensolver/bt_band_to_tridiag.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/eigensolver/bt_band_to_tridiag/api.h"
 #include "dlaf/matrix/matrix.h"
 #include "dlaf/types.h"
diff --git a/include/dlaf/eigensolver/bt_band_to_tridiag/api.h b/include/dlaf/eigensolver/bt_band_to_tridiag/api.h
index 31a8dcab62..05bf10fad7 100644
--- a/include/dlaf/eigensolver/bt_band_to_tridiag/api.h
+++ b/include/dlaf/eigensolver/bt_band_to_tridiag/api.h
@@ -23,7 +23,7 @@ struct BackTransformationT2B {
                    Matrix<const T, Device::CPU>& mat_hh);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_EIGENSOLVER_BT_BAND_TO_TRIDIAGONAL_ETI(KWORD, BACKEND, DEVICE, T) \
   KWORD template struct BackTransformationT2B<BACKEND, DEVICE, T>;
 
diff --git a/include/dlaf/eigensolver/bt_reduction_to_band.h b/include/dlaf/eigensolver/bt_reduction_to_band.h
index 28d82ab31b..7740754629 100644
--- a/include/dlaf/eigensolver/bt_reduction_to_band.h
+++ b/include/dlaf/eigensolver/bt_reduction_to_band.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 
 #include "dlaf/communication/communicator_grid.h"
diff --git a/include/dlaf/eigensolver/bt_reduction_to_band/api.h b/include/dlaf/eigensolver/bt_reduction_to_band/api.h
index d82c33e2b3..1101bfea91 100644
--- a/include/dlaf/eigensolver/bt_reduction_to_band/api.h
+++ b/include/dlaf/eigensolver/bt_reduction_to_band/api.h
@@ -25,7 +25,7 @@ struct BackTransformationReductionToBand {
                    common::internal::vector<pika::shared_future<common::internal::vector<T>>> taus);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_EIGENSOLVER_BT_REDUCTION_TO_BAND_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct BackTransformationReductionToBand<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/eigensolver/eigensolver.h b/include/dlaf/eigensolver/eigensolver.h
index e86868333d..5446fb7ae5 100644
--- a/include/dlaf/eigensolver/eigensolver.h
+++ b/include/dlaf/eigensolver/eigensolver.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 
 #include "dlaf/communication/communicator_grid.h"
diff --git a/include/dlaf/eigensolver/eigensolver/api.h b/include/dlaf/eigensolver/eigensolver/api.h
index 9fb039cf3e..28effdf1c6 100644
--- a/include/dlaf/eigensolver/eigensolver/api.h
+++ b/include/dlaf/eigensolver/eigensolver/api.h
@@ -34,7 +34,7 @@ struct Eigensolver {
                    Matrix<BaseType<T>, D>& evals, Matrix<T, D>& mat_e);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_EIGENSOLVER_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct Eigensolver<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/eigensolver/eigensolver/impl.h b/include/dlaf/eigensolver/eigensolver/impl.h
index 91afb7da92..65e180e9bb 100644
--- a/include/dlaf/eigensolver/eigensolver/impl.h
+++ b/include/dlaf/eigensolver/eigensolver/impl.h
@@ -20,7 +20,7 @@
 #include "dlaf/eigensolver/band_to_tridiag.h"
 #include "dlaf/eigensolver/bt_band_to_tridiag.h"
 #include "dlaf/eigensolver/bt_reduction_to_band.h"
-#include "dlaf/eigensolver/get_band_size.h"
+#include "dlaf/eigensolver/internal/get_band_size.h"
 #include "dlaf/eigensolver/reduction_to_band.h"
 #include "dlaf/eigensolver/tridiag_solver.h"
 #include "dlaf/lapack/tile.h"
diff --git a/include/dlaf/eigensolver/gen_eigensolver.h b/include/dlaf/eigensolver/gen_eigensolver.h
index b9d7480b6a..94087ec8b5 100644
--- a/include/dlaf/eigensolver/gen_eigensolver.h
+++ b/include/dlaf/eigensolver/gen_eigensolver.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include "dlaf/eigensolver/gen_eigensolver/api.h"
 #include "dlaf/matrix/matrix.h"
@@ -25,8 +27,8 @@ namespace dlaf::eigensolver {
 /// - the lower triangle or the upper triangle (depending on @p uplo) of @p mat_a,
 /// including the diagonal, is destroyed.
 /// - @p mat_b contains the Cholesky decomposition of B
-/// - @eigenvalues contains all the eigenvalues lambda
-/// - @eigenvectors contains all the eigenvectors x
+/// - @p eigenvalues contains all the eigenvalues lambda
+/// - @p eigenvectors contains all the eigenvectors x
 ///
 /// Implementation on local memory.
 ///
@@ -104,8 +106,8 @@ EigensolverResult<T, D> genEigensolver(blas::Uplo uplo, Matrix<T, D>& mat_a, Mat
 /// - the lower triangle or the upper triangle (depending on @p uplo) of @p mat_a,
 /// including the diagonal, is destroyed.
 /// - @p mat_b contains the Cholesky decomposition of B
-/// - @eigenvalues contains all the eigenvalues lambda
-/// - @eigenvectors contains all the eigenvectors x
+/// - @p eigenvalues contains all the eigenvalues lambda
+/// - @p eigenvectors contains all the eigenvectors x
 ///
 /// Implementation on distributed memory.
 ///
diff --git a/include/dlaf/eigensolver/gen_eigensolver/api.h b/include/dlaf/eigensolver/gen_eigensolver/api.h
index c44c55c1d6..34f9b485be 100644
--- a/include/dlaf/eigensolver/gen_eigensolver/api.h
+++ b/include/dlaf/eigensolver/gen_eigensolver/api.h
@@ -25,7 +25,7 @@ struct GenEigensolver {
                    Matrix<T, device>& eigenvectors);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_EIGENSOLVER_GEN_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct GenEigensolver<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/eigensolver/gen_to_std.h b/include/dlaf/eigensolver/gen_to_std.h
index 2939dbde59..09e8b2c54a 100644
--- a/include/dlaf/eigensolver/gen_to_std.h
+++ b/include/dlaf/eigensolver/gen_to_std.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/eigensolver/gen_to_std/api.h"
diff --git a/include/dlaf/eigensolver/gen_to_std/api.h b/include/dlaf/eigensolver/gen_to_std/api.h
index 49257060dd..1d766bd851 100644
--- a/include/dlaf/eigensolver/gen_to_std/api.h
+++ b/include/dlaf/eigensolver/gen_to_std/api.h
@@ -24,7 +24,7 @@ struct GenToStd {
   static void call_U(comm::CommunicatorGrid grid, Matrix<T, device>& mat_a, Matrix<T, device>& mat_u);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_GENTOSTD_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct GenToStd<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/eigensolver/get_1d_block_size.h b/include/dlaf/eigensolver/internal/get_1d_block_size.h
similarity index 100%
rename from include/dlaf/eigensolver/get_1d_block_size.h
rename to include/dlaf/eigensolver/internal/get_1d_block_size.h
diff --git a/include/dlaf/eigensolver/get_band_size.h b/include/dlaf/eigensolver/internal/get_band_size.h
similarity index 100%
rename from include/dlaf/eigensolver/get_band_size.h
rename to include/dlaf/eigensolver/internal/get_band_size.h
diff --git a/include/dlaf/eigensolver/get_red2band_panel_nworkers.h b/include/dlaf/eigensolver/internal/get_red2band_panel_nworkers.h
similarity index 100%
rename from include/dlaf/eigensolver/get_red2band_panel_nworkers.h
rename to include/dlaf/eigensolver/internal/get_red2band_panel_nworkers.h
diff --git a/include/dlaf/eigensolver/get_tridiag_rank1_nworkers.h b/include/dlaf/eigensolver/internal/get_tridiag_rank1_nworkers.h
similarity index 100%
rename from include/dlaf/eigensolver/get_tridiag_rank1_nworkers.h
rename to include/dlaf/eigensolver/internal/get_tridiag_rank1_nworkers.h
diff --git a/include/dlaf/eigensolver/reduction_to_band.h b/include/dlaf/eigensolver/reduction_to_band.h
index 24c34373ef..e32582e517 100644
--- a/include/dlaf/eigensolver/reduction_to_band.h
+++ b/include/dlaf/eigensolver/reduction_to_band.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/common/vector.h"
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/matrix/matrix.h"
@@ -107,12 +109,15 @@ common::internal::vector<pika::shared_future<common::internal::vector<T>>> reduc
 ///
 /// In particular, @p mat_a will look like this (tile representation) if band_size == blocksize
 ///
-/// B ~ ~ ~ ~ ~
-/// * B ~ ~ ~ ~
-/// v * B ~ ~ ~
-/// v v * B ~ ~
-/// v v v * B ~
-/// v v v v * B
+/** @verbatim
+B ~ ~ ~ ~ ~
+* B ~ ~ ~ ~
+v * B ~ ~ ~
+v v * B ~ ~
+v v v * B ~
+v v v v * B
+@endverbatim
+*/
 ///
 /// where each column of `v` is an elementary reflector without its first element (which is always equal
 /// to 1), `B` are the tiles containg the band-diagonal form, while `*` tiles contain both elements
@@ -120,14 +125,15 @@ common::internal::vector<pika::shared_future<common::internal::vector<T>>> reduc
 /// diagonal excluded).
 ///
 /// In case band_size < blocksize:
-///
-/// * ~ ~ ~ ~ ~
-/// * * ~ ~ ~ ~
-/// v * * ~ ~ ~
-/// v v * * ~ ~
-/// v v v * * ~
-/// v v v v * *
-///
+/** @verbatim
+* ~ ~ ~ ~ ~
+* * ~ ~ ~ ~
+v * * ~ ~ ~
+v v * * ~ ~
+v v v * * ~
+v v v v * *
+@endverbatim
+*/
 /// @param grid is the CommunicatorGrid on which @p mat_a is distributed
 /// @param mat_a on entry it contains an Hermitian matrix, on exit it is overwritten with the
 /// band-diagonal result together with the elementary reflectors as described above. Just the tiles of
diff --git a/include/dlaf/eigensolver/reduction_to_band/api.h b/include/dlaf/eigensolver/reduction_to_band/api.h
index 5e53dcf564..0b6ea8588c 100644
--- a/include/dlaf/eigensolver/reduction_to_band/api.h
+++ b/include/dlaf/eigensolver/reduction_to_band/api.h
@@ -26,7 +26,7 @@ struct ReductionToBand {
       comm::CommunicatorGrid grid, Matrix<T, D>& mat_a, const SizeType band_size);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_EIGENSOLVER_REDUCTION_TO_BAND_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct ReductionToBand<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/eigensolver/reduction_to_band/impl.h b/include/dlaf/eigensolver/reduction_to_band/impl.h
index 805c770fc4..f2ec116445 100644
--- a/include/dlaf/eigensolver/reduction_to_band/impl.h
+++ b/include/dlaf/eigensolver/reduction_to_band/impl.h
@@ -32,7 +32,7 @@
 #include "dlaf/communication/kernels/all_reduce.h"
 #include "dlaf/communication/kernels/reduce.h"
 #include "dlaf/communication/rdma.h"
-#include "dlaf/eigensolver/get_red2band_panel_nworkers.h"
+#include "dlaf/eigensolver/internal/get_red2band_panel_nworkers.h"
 #include "dlaf/lapack/tile.h"
 #include "dlaf/matrix/copy_tile.h"
 #include "dlaf/matrix/distribution.h"
diff --git a/include/dlaf/eigensolver/tridiag_solver.h b/include/dlaf/eigensolver/tridiag_solver.h
index c812869cb9..a6b39d9af7 100644
--- a/include/dlaf/eigensolver/tridiag_solver.h
+++ b/include/dlaf/eigensolver/tridiag_solver.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include "dlaf/common/assert.h"
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/eigensolver/tridiag_solver/api.h"
diff --git a/include/dlaf/eigensolver/tridiag_solver/api.h b/include/dlaf/eigensolver/tridiag_solver/api.h
index b122b1767c..3066dc8c4b 100644
--- a/include/dlaf/eigensolver/tridiag_solver/api.h
+++ b/include/dlaf/eigensolver/tridiag_solver/api.h
@@ -25,7 +25,7 @@ struct TridiagSolver {
                    Matrix<T, device>& evals, Matrix<std::complex<T>, device>& evecs);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_TRIDIAGONAL_EIGENSOLVER_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct TridiagSolver<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/eigensolver/tridiag_solver/merge.h b/include/dlaf/eigensolver/tridiag_solver/merge.h
index bf4d5169bf..0e19d76985 100644
--- a/include/dlaf/eigensolver/tridiag_solver/merge.h
+++ b/include/dlaf/eigensolver/tridiag_solver/merge.h
@@ -17,7 +17,7 @@
 #include "dlaf/common/range2d.h"
 #include "dlaf/common/single_threaded_blas.h"
 #include "dlaf/communication/kernels.h"
-#include "dlaf/eigensolver/get_tridiag_rank1_nworkers.h"
+#include "dlaf/eigensolver/internal/get_tridiag_rank1_nworkers.h"
 #include "dlaf/eigensolver/tridiag_solver/coltype.h"
 #include "dlaf/eigensolver/tridiag_solver/index_manipulation.h"
 #include "dlaf/eigensolver/tridiag_solver/kernels.h"
diff --git a/include/dlaf/factorization.h b/include/dlaf/factorization.h
index 9c71c30daa..ae53af7ced 100644
--- a/include/dlaf/factorization.h
+++ b/include/dlaf/factorization.h
@@ -9,5 +9,7 @@
 //
 #pragma once
 
+/// @file
+
 #include "dlaf/factorization/cholesky.h"
 #include "dlaf/factorization/qr.h"
diff --git a/include/dlaf/factorization/cholesky.h b/include/dlaf/factorization/cholesky.h
index 79157b62a5..fa6f397b57 100644
--- a/include/dlaf/factorization/cholesky.h
+++ b/include/dlaf/factorization/cholesky.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/factorization/cholesky/api.h"
diff --git a/include/dlaf/factorization/cholesky/api.h b/include/dlaf/factorization/cholesky/api.h
index 8916051833..3ea92edcac 100644
--- a/include/dlaf/factorization/cholesky/api.h
+++ b/include/dlaf/factorization/cholesky/api.h
@@ -23,7 +23,7 @@ struct Cholesky {
   static void call_U(comm::CommunicatorGrid grid, Matrix<T, device>& mat_a);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_FACTORIZATION_CHOLESKY_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct Cholesky<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/factorization/qr.h b/include/dlaf/factorization/qr.h
index 0f0e0a9b6a..e4d17e19ec 100644
--- a/include/dlaf/factorization/qr.h
+++ b/include/dlaf/factorization/qr.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include "dlaf/factorization/qr/api.h"
 #include "dlaf/matrix/index.h"
 #include "dlaf/matrix/tile.h"
diff --git a/include/dlaf/factorization/qr/api.h b/include/dlaf/factorization/qr/api.h
index 46ea0a7e82..d089f1ff3e 100644
--- a/include/dlaf/factorization/qr/api.h
+++ b/include/dlaf/factorization/qr/api.h
@@ -85,7 +85,7 @@ struct QR_Tfactor {
                    common::Pipeline<comm::Communicator>& mpi_col_task_chain);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_FACTORIZATION_QR_TFACTOR_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct QR_Tfactor<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/gpu/blas/api.h b/include/dlaf/gpu/blas/api.h
index 95ba808692..b0a09b2d5d 100644
--- a/include/dlaf/gpu/blas/api.h
+++ b/include/dlaf/gpu/blas/api.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #ifdef DLAF_WITH_HIP
 
 #include <rocblas/rocblas.h>
diff --git a/include/dlaf/gpu/blas/error.h b/include/dlaf/gpu/blas/error.h
index b087ae64dc..891d63a667 100644
--- a/include/dlaf/gpu/blas/error.h
+++ b/include/dlaf/gpu/blas/error.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #include <exception>
 #include <iostream>
 
diff --git a/include/dlaf/gpu/cublas/error.h b/include/dlaf/gpu/cublas/error.h
index 17caee86d8..4b78a173ac 100644
--- a/include/dlaf/gpu/cublas/error.h
+++ b/include/dlaf/gpu/cublas/error.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #include <string>
 
 #include <dlaf/gpu/blas/api.h>
diff --git a/include/dlaf/gpu/cusolver/error.h b/include/dlaf/gpu/cusolver/error.h
index 1482a38074..ec70d58842 100644
--- a/include/dlaf/gpu/cusolver/error.h
+++ b/include/dlaf/gpu/cusolver/error.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #include <string>
 
 #include "dlaf/gpu/lapack/api.h"
diff --git a/include/dlaf/gpu/lapack/api.h b/include/dlaf/gpu/lapack/api.h
index 755c228b37..6d24dfc167 100644
--- a/include/dlaf/gpu/lapack/api.h
+++ b/include/dlaf/gpu/lapack/api.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #ifdef DLAF_WITH_HIP
 
 #include <rocblas/rocblas.h>
diff --git a/include/dlaf/gpu/lapack/assert_info.h b/include/dlaf/gpu/lapack/assert_info.h
index 560cddac92..af9a3357ed 100644
--- a/include/dlaf/gpu/lapack/assert_info.h
+++ b/include/dlaf/gpu/lapack/assert_info.h
@@ -14,8 +14,6 @@
 
 #include <whip.hpp>
 
-/// @file
-
 #define DLAF_DECLARE_CUSOLVER_ASSERT_INFO(func) void assertInfo##func(whip::stream_t stream, int* info)
 
 namespace dlaf::gpulapack::internal {
diff --git a/include/dlaf/gpu/lapack/error.h b/include/dlaf/gpu/lapack/error.h
index 8595cde3a4..9c31a6d160 100644
--- a/include/dlaf/gpu/lapack/error.h
+++ b/include/dlaf/gpu/lapack/error.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #include <exception>
 #include <iostream>
 
diff --git a/include/dlaf/gpu/rocblas/error.h b/include/dlaf/gpu/rocblas/error.h
index f92107d990..407d4d5c7a 100644
--- a/include/dlaf/gpu/rocblas/error.h
+++ b/include/dlaf/gpu/rocblas/error.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #include <string>
 
 #include <dlaf/gpu/blas/api.h>
diff --git a/include/dlaf/gpu/rocsolver/error.h b/include/dlaf/gpu/rocsolver/error.h
index eac12ba7ef..8c34a32d2c 100644
--- a/include/dlaf/gpu/rocsolver/error.h
+++ b/include/dlaf/gpu/rocsolver/error.h
@@ -10,8 +10,6 @@
 
 #pragma once
 
-/// @file
-
 #include <string>
 
 #include "dlaf/gpu/lapack/api.h"
diff --git a/include/dlaf/init.h b/include/dlaf/init.h
index e176be61a9..f3bd09f712 100644
--- a/include/dlaf/init.h
+++ b/include/dlaf/init.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <iostream>
 
 #include <pika/program_options.hpp>
diff --git a/include/dlaf/matrix/copy.h b/include/dlaf/matrix/copy.h
index 24bb0d9951..2a4a5c9724 100644
--- a/include/dlaf/matrix/copy.h
+++ b/include/dlaf/matrix/copy.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <pika/execution.hpp>
 
 #include "dlaf/common/range2d.h"
diff --git a/include/dlaf/matrix/copy_tile.h b/include/dlaf/matrix/copy_tile.h
index f910e889e7..adaa61d5e2 100644
--- a/include/dlaf/matrix/copy_tile.h
+++ b/include/dlaf/matrix/copy_tile.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <type_traits>
 
 #if DLAF_WITH_GPU
diff --git a/include/dlaf/matrix/distribution.h b/include/dlaf/matrix/distribution.h
index 3dfd7bfc75..5baedced53 100644
--- a/include/dlaf/matrix/distribution.h
+++ b/include/dlaf/matrix/distribution.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/matrix/index.h"
 #include "dlaf/matrix/util_distribution.h"
diff --git a/include/dlaf/matrix/index.h b/include/dlaf/matrix/index.h
index c94a032531..0f9b4a7a43 100644
--- a/include/dlaf/matrix/index.h
+++ b/include/dlaf/matrix/index.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/common/index2d.h"
 #include "dlaf/types.h"
 
diff --git a/include/dlaf/matrix/layout_info.h b/include/dlaf/matrix/layout_info.h
index d9fb95c4ee..0a8d7029da 100644
--- a/include/dlaf/matrix/layout_info.h
+++ b/include/dlaf/matrix/layout_info.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include "dlaf/common/assert.h"
 #include "dlaf/matrix/distribution.h"
 #include "dlaf/matrix/index.h"
diff --git a/include/dlaf/matrix/matrix.h b/include/dlaf/matrix/matrix.h
index 0e48a87022..8704feea74 100644
--- a/include/dlaf/matrix/matrix.h
+++ b/include/dlaf/matrix/matrix.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <exception>
 #include <vector>
 
@@ -375,7 +377,7 @@ auto select(MatrixLike& matrix, common::IterableRange2D<SizeType, LocalTile_TAG>
   return internal::selectGeneric([&](auto index) { return matrix.readwrite(index); }, range);
 }
 
-/// ---- ETI
+// ETI
 
 #define DLAF_MATRIX_ETI(KWORD, DATATYPE, DEVICE) \
   KWORD template class Matrix<DATATYPE, DEVICE>; \
diff --git a/include/dlaf/matrix/matrix_base.h b/include/dlaf/matrix/matrix_base.h
index 664711b010..2d5638e0e1 100644
--- a/include/dlaf/matrix/matrix_base.h
+++ b/include/dlaf/matrix/matrix_base.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <cstddef>
 #include <ostream>
 
diff --git a/include/dlaf/matrix/matrix_mirror.h b/include/dlaf/matrix/matrix_mirror.h
index 65c2963b36..541fbb524c 100644
--- a/include/dlaf/matrix/matrix_mirror.h
+++ b/include/dlaf/matrix/matrix_mirror.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <exception>
 #include <vector>
 
@@ -170,7 +172,7 @@ class MatrixMirror : public MatrixMirror<const T, Target, Source> {
   }
 };
 
-/// ---- ETI
+// ETI
 
 #define DLAF_MATRIX_MIRROR_ETI(KWORD, DATATYPE, TARGETDEVICE, SOURCEDEVICE) \
   KWORD template class MatrixMirror<DATATYPE, TARGETDEVICE, SOURCEDEVICE>;  \
diff --git a/include/dlaf/matrix/panel.h b/include/dlaf/matrix/panel.h
index ba3f323f49..0af7091eeb 100644
--- a/include/dlaf/matrix/panel.h
+++ b/include/dlaf/matrix/panel.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <pika/execution.hpp>
 
 #include "dlaf/common/index2d.h"
diff --git a/include/dlaf/matrix/print_csv.h b/include/dlaf/matrix/print_csv.h
index dbf6d99f35..532773bb3c 100644
--- a/include/dlaf/matrix/print_csv.h
+++ b/include/dlaf/matrix/print_csv.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include "dlaf/common/index2d.h"
 #include "dlaf/common/range2d.h"
 #include "dlaf/matrix/matrix.h"
diff --git a/include/dlaf/matrix/print_gpu.h b/include/dlaf/matrix/print_gpu.h
index a9245bbb3b..070074c68c 100644
--- a/include/dlaf/matrix/print_gpu.h
+++ b/include/dlaf/matrix/print_gpu.h
@@ -11,6 +11,8 @@
 #ifdef DLAF_WITH_GPU
 #pragma once
 
+/// @file
+
 #include <iostream>
 
 #include <whip.hpp>
diff --git a/include/dlaf/matrix/print_numpy.h b/include/dlaf/matrix/print_numpy.h
index 0acde1611e..dd72bd49d7 100644
--- a/include/dlaf/matrix/print_numpy.h
+++ b/include/dlaf/matrix/print_numpy.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <string>
 
 #include "dlaf/common/index2d.h"
diff --git a/include/dlaf/matrix/retiled_matrix.h b/include/dlaf/matrix/retiled_matrix.h
index e550b9bd5f..454b1ecbc3 100644
--- a/include/dlaf/matrix/retiled_matrix.h
+++ b/include/dlaf/matrix/retiled_matrix.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <cstddef>
 #include <vector>
 
@@ -161,7 +163,7 @@ class RetiledMatrix : public internal::MatrixBase {
 template <class T, Device D>
 class RetiledMatrix<const T, D>;
 
-/// ---- ETI
+// ETI
 
 #define DLAF_RETILED_MATRIX_ETI(KWORD, DATATYPE, DEVICE) \
   KWORD template class RetiledMatrix<DATATYPE, DEVICE>;
diff --git a/include/dlaf/matrix/tile.h b/include/dlaf/matrix/tile.h
index ef31e0e2ff..dbba5eb434 100644
--- a/include/dlaf/matrix/tile.h
+++ b/include/dlaf/matrix/tile.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <exception>
 #include <ostream>
 #include <tuple>
@@ -612,7 +614,7 @@ std::vector<ReadWriteTileSender<T, D>> splitTileDisjoint(ReadWriteTileSender<T,
   return senders;
 }
 
-/// ---- ETI
+// ETI
 
 #define DLAF_TILE_ETI(KWORD, DATATYPE, DEVICE) \
   KWORD template class Tile<DATATYPE, DEVICE>; \
diff --git a/include/dlaf/matrix/views.h b/include/dlaf/matrix/views.h
index fd72f7135c..3101601a0a 100644
--- a/include/dlaf/matrix/views.h
+++ b/include/dlaf/matrix/views.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <ios>
 #include "dlaf/common/assert.h"
 #include "dlaf/common/index2d.h"
diff --git a/include/dlaf/memory/memory_chunk.h b/include/dlaf/memory/memory_chunk.h
index 8b48ee7dde..40905273c4 100644
--- a/include/dlaf/memory/memory_chunk.h
+++ b/include/dlaf/memory/memory_chunk.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <cstdlib>
 #include <memory>
 
diff --git a/include/dlaf/memory/memory_view.h b/include/dlaf/memory/memory_view.h
index 756d2af676..89068d1b86 100644
--- a/include/dlaf/memory/memory_view.h
+++ b/include/dlaf/memory/memory_view.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <cstdlib>
 #include <memory>
 
@@ -152,7 +154,7 @@ class MemoryView {
   SizeType size_;
 };
 
-/// ---- ETI
+// ETI
 
 #define DLAF_MEMVIEW_ETI(KWORD, DATATYPE, DEVICE) KWORD template class MemoryView<DATATYPE, DEVICE>;
 
diff --git a/include/dlaf/multiplication/general.h b/include/dlaf/multiplication/general.h
index 8a0fc58e4b..ccce201482 100644
--- a/include/dlaf/multiplication/general.h
+++ b/include/dlaf/multiplication/general.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 
 #include "dlaf/common/assert.h"
diff --git a/include/dlaf/multiplication/general/api.h b/include/dlaf/multiplication/general/api.h
index d1f8372dd5..16b59d7ea5 100644
--- a/include/dlaf/multiplication/general/api.h
+++ b/include/dlaf/multiplication/general/api.h
@@ -30,7 +30,7 @@ struct GeneralSub {
                      Matrix<const T, D>& mat_b, const T beta, Matrix<T, D>& mat_c);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_MULTIPLICATION_GENERAL_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct GeneralSub<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/multiplication/hermitian.h b/include/dlaf/multiplication/hermitian.h
index 745f8f918f..1c87e40de5 100644
--- a/include/dlaf/multiplication/hermitian.h
+++ b/include/dlaf/multiplication/hermitian.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/matrix/matrix.h"
diff --git a/include/dlaf/multiplication/hermitian/api.h b/include/dlaf/multiplication/hermitian/api.h
index 7aa900a122..f20583ce4b 100644
--- a/include/dlaf/multiplication/hermitian/api.h
+++ b/include/dlaf/multiplication/hermitian/api.h
@@ -22,7 +22,7 @@ struct Hermitian {
                       Matrix<const T, D>& mat_b, const T beta, Matrix<T, D>& mat_c);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_MULTIPLICATION_HERMITIAN_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct Hermitian<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/multiplication/triangular.h b/include/dlaf/multiplication/triangular.h
index c9cfd5577c..de30f334a0 100644
--- a/include/dlaf/multiplication/triangular.h
+++ b/include/dlaf/multiplication/triangular.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/matrix/matrix.h"
diff --git a/include/dlaf/multiplication/triangular/api.h b/include/dlaf/multiplication/triangular/api.h
index 9a46800860..528840ef25 100644
--- a/include/dlaf/multiplication/triangular/api.h
+++ b/include/dlaf/multiplication/triangular/api.h
@@ -45,7 +45,7 @@ struct Triangular {
                        Matrix<const T, device>& mat_a, Matrix<T, device>& mat_b);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_MULTIPLICATION_TRIANGULAR_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct Triangular<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/permutations/general.h b/include/dlaf/permutations/general.h
index 9c8cbb4bcf..a27e6b1adb 100644
--- a/include/dlaf/permutations/general.h
+++ b/include/dlaf/permutations/general.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 
 #include "dlaf/common/assert.h"
diff --git a/include/dlaf/permutations/general/api.h b/include/dlaf/permutations/general/api.h
index 3a75a29e89..3ac178fb5f 100644
--- a/include/dlaf/permutations/general/api.h
+++ b/include/dlaf/permutations/general/api.h
@@ -27,7 +27,7 @@ struct Permutations {
                    Matrix<T, D>& mat_out);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_PERMUTATIONS_GENERAL_ETI(KWORD, BACKEND, DEVICE, DATATYPE)      \
   KWORD template struct Permutations<BACKEND, DEVICE, DATATYPE, Coord::Col>; \
   KWORD template struct Permutations<BACKEND, DEVICE, DATATYPE, Coord::Row>;
diff --git a/include/dlaf/schedulers.h b/include/dlaf/schedulers.h
index 204c7d2e00..b231940942 100644
--- a/include/dlaf/schedulers.h
+++ b/include/dlaf/schedulers.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <pika/execution.hpp>
 #include <pika/runtime.hpp>
 #include <pika/thread.hpp>
diff --git a/include/dlaf/solver.h b/include/dlaf/solver.h
index 60cc2a7c37..93cbfd22e1 100644
--- a/include/dlaf/solver.h
+++ b/include/dlaf/solver.h
@@ -9,4 +9,6 @@
 //
 #pragma once
 
+/// @file
+
 #include "dlaf/solver/triangular.h"
diff --git a/include/dlaf/solver/triangular.h b/include/dlaf/solver/triangular.h
index 25ccc67a24..60ad73a05f 100644
--- a/include/dlaf/solver/triangular.h
+++ b/include/dlaf/solver/triangular.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/matrix/matrix.h"
diff --git a/include/dlaf/solver/triangular/api.h b/include/dlaf/solver/triangular/api.h
index 10a2ebc690..10e40d36a0 100644
--- a/include/dlaf/solver/triangular/api.h
+++ b/include/dlaf/solver/triangular/api.h
@@ -53,7 +53,7 @@ struct Triangular {
                        Matrix<const T, device>& mat_a, Matrix<T, device>& mat_b);
 };
 
-/// ---- ETI
+// ETI
 #define DLAF_SOLVER_TRIANGULAR_ETI(KWORD, BACKEND, DEVICE, DATATYPE) \
   KWORD template struct Triangular<BACKEND, DEVICE, DATATYPE>;
 
diff --git a/include/dlaf/tune.h b/include/dlaf/tune.h
index 4aa83a509a..54531596c6 100644
--- a/include/dlaf/tune.h
+++ b/include/dlaf/tune.h
@@ -9,6 +9,8 @@
 //
 #pragma once
 
+/// @file
+
 #include <cstdint>
 
 #include <pika/runtime.hpp>
diff --git a/include/dlaf/util_blas.h b/include/dlaf/util_blas.h
index adac196856..6d1fb3b5d1 100644
--- a/include/dlaf/util_blas.h
+++ b/include/dlaf/util_blas.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 
 #include <dlaf/common/assert.h>
diff --git a/include/dlaf/util_lapack.h b/include/dlaf/util_lapack.h
index 08b3b443ae..1beea431a3 100644
--- a/include/dlaf/util_lapack.h
+++ b/include/dlaf/util_lapack.h
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <blas.hh>
 #include <lapack/util.hh>
 #include "dlaf/common/assert.h"
diff --git a/include/dlaf/version.h.in b/include/dlaf/version.h.in
index b849812295..04383cab09 100644
--- a/include/dlaf/version.h.in
+++ b/include/dlaf/version.h.in
@@ -10,6 +10,8 @@
 
 #pragma once
 
+/// @file
+
 #include <string>
 
 namespace dlaf {
diff --git a/miniapp/miniapp_eigensolver.cpp b/miniapp/miniapp_eigensolver.cpp
index 2d5168e34e..43d4238fe9 100644
--- a/miniapp/miniapp_eigensolver.cpp
+++ b/miniapp/miniapp_eigensolver.cpp
@@ -24,7 +24,7 @@
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/communication/init.h"
 #include "dlaf/eigensolver/eigensolver.h"
-#include "dlaf/eigensolver/get_band_size.h"
+#include "dlaf/eigensolver/internal/get_band_size.h"
 #include "dlaf/init.h"
 #include "dlaf/matrix/copy.h"
 #include "dlaf/matrix/matrix.h"
diff --git a/miniapp/miniapp_gen_eigensolver.cpp b/miniapp/miniapp_gen_eigensolver.cpp
index a7c328a246..c83e564925 100644
--- a/miniapp/miniapp_gen_eigensolver.cpp
+++ b/miniapp/miniapp_gen_eigensolver.cpp
@@ -24,7 +24,7 @@
 #include "dlaf/communication/communicator_grid.h"
 #include "dlaf/communication/init.h"
 #include "dlaf/eigensolver/gen_eigensolver.h"
-#include "dlaf/eigensolver/get_band_size.h"
+#include "dlaf/eigensolver/internal/get_band_size.h"
 #include "dlaf/init.h"
 #include "dlaf/matrix/copy.h"
 #include "dlaf/matrix/matrix.h"
diff --git a/src/communication/communicator_impl.h b/src/communication/communicator_impl.h
index 3e10028a01..5834141861 100644
--- a/src/communication/communicator_impl.h
+++ b/src/communication/communicator_impl.h
@@ -12,8 +12,6 @@
 
 #include "dlaf/communication/error.h"
 
-/// @file
-
 namespace dlaf {
 namespace comm {