diff --git a/src/python/magnum/math.cpp b/src/python/magnum/math.cpp
index 598083b..cb6442e 100644
--- a/src/python/magnum/math.cpp
+++ b/src/python/magnum/math.cpp
@@ -37,6 +37,56 @@
 
 namespace magnum {
 
+/* Keep in sync with math.h */
+
+const char* const FormatStrings[]{
+    /* 0. Representing bytes as unsigned. Not using 'c' because then it behaves
+       differently from bytes/bytearray, where you can do `a[0] = ord('A')`. */
+    "B",
+
+    "b", /* 1 -- std::int8_t */
+    "B", /* 2 -- std::uint8_t */
+    "i", /* 3 -- std::int32_t */
+    "I", /* 4 -- std::uint32_t */
+    "f", /* 5 -- float */
+    "d"  /* 6 -- double */
+};
+
+/* Flipped as numpy expects row-major */
+const Py_ssize_t MatrixShapes[][2]{
+    {2, 2}, /* 0 -- 2 cols, 2 rows */
+    {3, 2}, /* 1 -- 2 cols, 3 rows */
+    {4, 2}, /* 2 -- 2 cols, 4 rows */
+    {2, 3}, /* 3 -- 3 cols, 2 rows */
+    {3, 3}, /* 4 -- 3 cols, 3 rows */
+    {4, 3}, /* 5 -- 3 cols, 4 rows */
+    {2, 4}, /* 6 -- 4 cols, 2 rows */
+    {3, 4}, /* 7 -- 4 cols, 3 rows */
+    {4, 4}  /* 8 -- 4 cols, 4 rows */
+};
+const Py_ssize_t MatrixStridesFloat[][2]{
+    {4, 4*2}, /* 0 -- 2 cols, 2 rows */
+    {4, 4*3}, /* 1 -- 2 cols, 3 rows */
+    {4, 4*4}, /* 2 -- 2 cols, 4 rows */
+    {4, 4*2}, /* 3 -- 3 cols, 2 rows */
+    {4, 4*3}, /* 4 -- 3 cols, 3 rows */
+    {4, 4*4}, /* 5 -- 3 cols, 4 rows */
+    {4, 4*2}, /* 6 -- 4 cols, 2 rows */
+    {4, 4*3}, /* 7 -- 4 cols, 3 rows */
+    {4, 4*4}  /* 8 -- 4 cols, 4 rows */
+};
+const Py_ssize_t MatrixStridesDouble[][2]{
+    {8, 8*2}, /* 0 -- 2 cols, 2 rows */
+    {8, 8*3}, /* 1 -- 2 cols, 3 rows */
+    {8, 8*4}, /* 2 -- 2 cols, 4 rows */
+    {8, 8*2}, /* 3 -- 3 cols, 2 rows */
+    {8, 8*3}, /* 4 -- 3 cols, 3 rows */
+    {8, 8*4}, /* 5 -- 3 cols, 4 rows */
+    {8, 8*2}, /* 6 -- 4 cols, 2 rows */
+    {8, 8*3}, /* 7 -- 4 cols, 3 rows */
+    {8, 8*4}  /* 8 -- 4 cols, 4 rows */
+};
+
 namespace {
 
 template<class T> void angle(py::class_<T>& c) {
diff --git a/src/python/magnum/math.h b/src/python/magnum/math.h
index 3ee74f8..a28a482 100644
--- a/src/python/magnum/math.h
+++ b/src/python/magnum/math.h
@@ -26,6 +26,7 @@
 */
 
 #include <sstream>
+#include <Python.h>
 #include <Corrade/Utility/Debug.h>
 #include <Magnum/Magnum.h>
 
@@ -33,6 +34,40 @@
 
 namespace magnum {
 
+/* Keep in sync with math.cpp */
+
+extern const char* const FormatStrings[];
+template<class> constexpr std::size_t formatIndex();
+template<> constexpr std::size_t formatIndex<char>() { return 0; }
+template<> constexpr std::size_t formatIndex<Byte>() { return 1; }
+template<> constexpr std::size_t formatIndex<UnsignedByte>() { return 2; }
+template<> constexpr std::size_t formatIndex<Int>() { return 3; }
+template<> constexpr std::size_t formatIndex<UnsignedInt>() { return 4; }
+template<> constexpr std::size_t formatIndex<Float>() { return 5; }
+template<> constexpr std::size_t formatIndex<Double>() { return 6; }
+
+extern const Py_ssize_t MatrixShapes[][2];
+template<UnsignedInt cols, UnsignedInt rows> constexpr std::size_t matrixShapeStrideIndex();
+template<> constexpr std::size_t matrixShapeStrideIndex<2, 2>() { return 0; }
+template<> constexpr std::size_t matrixShapeStrideIndex<2, 3>() { return 1; }
+template<> constexpr std::size_t matrixShapeStrideIndex<2, 4>() { return 2; }
+template<> constexpr std::size_t matrixShapeStrideIndex<3, 2>() { return 3; }
+template<> constexpr std::size_t matrixShapeStrideIndex<3, 3>() { return 4; }
+template<> constexpr std::size_t matrixShapeStrideIndex<3, 4>() { return 5; }
+template<> constexpr std::size_t matrixShapeStrideIndex<4, 2>() { return 6; }
+template<> constexpr std::size_t matrixShapeStrideIndex<4, 3>() { return 7; }
+template<> constexpr std::size_t matrixShapeStrideIndex<4, 4>() { return 8; }
+
+extern const Py_ssize_t MatrixStridesFloat[][2];
+extern const Py_ssize_t MatrixStridesDouble[][2];
+template<class> constexpr const Py_ssize_t* matrixStridesFor(std::size_t i);
+template<> constexpr const Py_ssize_t* matrixStridesFor<Float>(std::size_t i) {
+    return MatrixStridesFloat[i];
+}
+template<> constexpr const Py_ssize_t* matrixStridesFor<Double>(std::size_t i) {
+    return MatrixStridesDouble[i];
+}
+
 template<class T> std::string repr(const T& value) {
     std::ostringstream out;
     Debug{&out, Debug::Flag::NoNewlineAtTheEnd} << value;
diff --git a/src/python/magnum/math.matrix.h b/src/python/magnum/math.matrix.h
index 1c1cfc9..2330699 100644
--- a/src/python/magnum/math.matrix.h
+++ b/src/python/magnum/math.matrix.h
@@ -27,11 +27,13 @@
 
 #include <pybind11/pybind11.h>
 #include <pybind11/operators.h>
+#include <Corrade/Containers/ScopeGuard.h>
 #include <Corrade/Utility/FormatStl.h>
 #include <Magnum/Math/Matrix3.h>
 #include <Magnum/Math/Matrix4.h>
 
 #include "corrade/PybindExtras.h"
+#include "corrade/PyBuffer.h"
 
 #include "magnum/math.h"
 
@@ -44,10 +46,10 @@ template<class T> struct VectorTraits<2, T> { typedef Math::Vector2<T> Type; };
 template<class T> struct VectorTraits<3, T> { typedef Math::Vector3<T> Type; };
 template<class T> struct VectorTraits<4, T> { typedef Math::Vector4<T> Type; };
 
-template<class U, class T> void initFromBuffer(T& out, const py::buffer_info& info) {
+template<class U, class T> void initFromBuffer(T& out, const Py_buffer& buffer) {
     for(std::size_t i = 0; i != T::Cols; ++i)
         for(std::size_t j = 0; j != T::Rows; ++j)
-            out[i][j] = static_cast<typename T::Type>(*reinterpret_cast<const U*>(static_cast<const char*>(info.ptr) + i*info.strides[1] + j*info.strides[0]));
+            out[i][j] = static_cast<typename T::Type>(*reinterpret_cast<const U*>(static_cast<const char*>(buffer.buf) + i*buffer.strides[1] + j*buffer.strides[0]));
 }
 
 /* Called for both Matrix3x3 and Matrix3 in order to return a proper type /
@@ -67,24 +69,30 @@ template<class T, class ...Args> void everyRectangularMatrix(py::class_<T, Args.
         .def(py::init(), "Default constructor")
         .def(py::init<typename T::Type>(), "Construct a matrix with one value for all components")
 
-        /* Buffer protocol, needed in order to make numpy treat the matric
-           correctly as column-major. Has to be defined *before* the from-tuple
-           constructor so it gets precedence for types that implement the
-           buffer protocol. */
-        .def(py::init([](py::buffer buffer) {
-            py::buffer_info info = buffer.request();
+        /* Buffer protocol, needed in order to properly detect row-major
+           layouts. Has to be defined *before* the from-tuple constructor so it
+           gets precedence for types that implement the buffer protocol. */
+        .def(py::init([](py::buffer other) {
+            Py_buffer buffer{};
+            if(PyObject_GetBuffer(other.ptr(), &buffer, PyBUF_FORMAT|PyBUF_STRIDES) != 0)
+                throw py::error_already_set{};
 
-            if(info.ndim != 2)
-                throw py::buffer_error{Utility::formatString("expected 2 dimensions but got {}", info.ndim)};
+            Containers::ScopeGuard e{&buffer, PyBuffer_Release};
 
-            if(info.shape[0] != T::Rows ||info.shape[1] != T::Cols)
-                throw py::buffer_error{Utility::formatString("expected {}x{} elements but got {}x{}", T::Cols, T::Rows, info.shape[1], info.shape[0])};
+            if(buffer.ndim != 2)
+                throw py::buffer_error{Utility::formatString("expected 2 dimensions but got {}", buffer.ndim)};
+
+            if(buffer.shape[0] != T::Rows || buffer.shape[1] != T::Cols)
+                throw py::buffer_error{Utility::formatString("expected {}x{} elements but got {}x{}", T::Cols, T::Rows, buffer.shape[1], buffer.shape[0])};
 
             T out{Math::NoInit};
 
-            if(info.format == "f") initFromBuffer<Float>(out, info);
-            else if(info.format == "d") initFromBuffer<Double>(out, info);
-            else throw py::buffer_error{Utility::formatString("expected format f or d but got {}", info.format)};
+            /* Expecting just an one-letter format */
+            if(buffer.format[0] == 'f' && !buffer.format[1])
+                initFromBuffer<Float>(out, buffer);
+            else if(buffer.format[0] == 'd' && !buffer.format[1])
+                initFromBuffer<Double>(out, buffer);
+            else throw py::buffer_error{Utility::formatString("expected format f or d but got {}", buffer.format)};
 
             return out;
         }), "Construct from a buffer")
@@ -113,6 +121,31 @@ template<class T, class ...Args> void everyRectangularMatrix(py::class_<T, Args.
         }, "Values on diagonal");
 }
 
+template<class T> bool rectangularMatrixBufferProtocol(T& self, Py_buffer& buffer, int flags) {
+    /* I hate the const_casts but I assume this is to make editing easier, NOT
+       to make it possible for users to stomp on these values. */
+    buffer.ndim = 2;
+    buffer.itemsize = sizeof(typename T::Type);
+    buffer.len = sizeof(T);
+    buffer.buf = self.data();
+    buffer.readonly = false;
+    if((flags & PyBUF_FORMAT) == PyBUF_FORMAT)
+        buffer.format = const_cast<char*>(FormatStrings[formatIndex<typename T::Type>()]);
+    if(flags != PyBUF_SIMPLE) {
+        /* Reusing shape definitions from matrices because I don't want to
+           create another useless array for that and reinterpret_cast on the
+           buffer.internal is UGLY. It's flipped from column-major to
+           row-major, so adjusting the row instead. */
+        buffer.shape = const_cast<Py_ssize_t*>(MatrixShapes[matrixShapeStrideIndex<T::Cols, T::Rows>()]);
+        CORRADE_INTERNAL_ASSERT(buffer.shape[0] == T::Rows);
+        CORRADE_INTERNAL_ASSERT(buffer.shape[1] == T::Cols);
+        if((flags & PyBUF_STRIDES) == PyBUF_STRIDES)
+            buffer.strides = const_cast<Py_ssize_t*>(matrixStridesFor<typename T::Type>(matrixShapeStrideIndex<T::Cols, T::Rows>()));
+    }
+
+    return true;
+}
+
 template<class T> void rectangularMatrix(py::class_<T>& c) {
     /*
         Missing APIs:
@@ -128,21 +161,6 @@ template<class T> void rectangularMatrix(py::class_<T>& c) {
     */
 
     c
-        /* Buffer protocol, needed in order to make numpy treat the matrix
-           correctly as column-major. The constructor is defined in
-           everyRectangularMatrix(). */
-        .def_buffer([](const T& self) -> py::buffer_info {
-            // TODO: ownership?
-            return py::buffer_info{
-                const_cast<typename T::Type*>(self.data()),
-                sizeof(typename T::Type),
-                py::format_descriptor<typename T::Type>::format(),
-                2,
-                {T::Rows, T::Cols},
-                {sizeof(typename T::Type), sizeof(typename T::Type)*T::Rows}
-            };
-        })
-
         /* Comparison */
         .def(py::self == py::self, "Equality comparison")
         .def(py::self != py::self, "Non-equality comparison")
@@ -170,6 +188,11 @@ template<class T> void rectangularMatrix(py::class_<T>& c) {
 
         .def("__repr__", repr<T>, "Object representation");
 
+    /* Buffer protocol, needed in order to make numpy treat the matrix
+       correctly as column-major. The constructor is defined in
+       everyRectangularMatrix(). */
+    corrade::enableBetterBufferProtocol<T, rectangularMatrixBufferProtocol>(c);
+
     /* Matrix column count */
     char lenDocstring[] = "Matrix column count. Returns _.";
     lenDocstring[sizeof(lenDocstring) - 3] = '0' + T::Cols;
diff --git a/src/python/magnum/math.matrixdouble.cpp b/src/python/magnum/math.matrixdouble.cpp
index 173ce98..27a99a0 100644
--- a/src/python/magnum/math.matrixdouble.cpp
+++ b/src/python/magnum/math.matrixdouble.cpp
@@ -40,8 +40,13 @@ void mathMatrixDouble(py::module& root) {
     py::class_<Matrix4x3d> matrix4x3d{root, "Matrix4x3d", "4x3 double matrix", py::buffer_protocol{}};
     py::class_<Matrix4x4d> matrix4x4d{root, "Matrix4x4d", "4x4 double matrix", py::buffer_protocol{}};
 
-    py::class_<Matrix3d, Matrix3x3d> matrix3d{root, "Matrix3d", "2D double transformation matrix", py::buffer_protocol{}};
-    py::class_<Matrix4d, Matrix4x4d> matrix4d{root, "Matrix4d", "3D double transformation matrix", py::buffer_protocol{}};
+    /* The subclasses don't have buffer protocol enabled, as that's already
+       done by the base classes. Moreover, just adding py::buffer_protocol{}
+       would cause it to not find the buffer functions as we don't add them
+       anywhere, thus failing with `pybind11_getbuffer(): Internal error`. */
+
+    py::class_<Matrix3d, Matrix3x3d> matrix3d{root, "Matrix3d", "2D double transformation matrix"};
+    py::class_<Matrix4d, Matrix4x4d> matrix4d{root, "Matrix4d", "3D double transformation matrix"};
 
     matrices<Double>(
         matrix2x2d, matrix2x3d, matrix2x4d,
diff --git a/src/python/magnum/math.matrixfloat.cpp b/src/python/magnum/math.matrixfloat.cpp
index a453fbe..75d8a41 100644
--- a/src/python/magnum/math.matrixfloat.cpp
+++ b/src/python/magnum/math.matrixfloat.cpp
@@ -40,8 +40,13 @@ void mathMatrixFloat(py::module& root) {
     py::class_<Matrix4x3> matrix4x3{root, "Matrix4x3", "4x3 float matrix", py::buffer_protocol{}};
     py::class_<Matrix4x4> matrix4x4{root, "Matrix4x4", "4x4 float matrix", py::buffer_protocol{}};
 
-    py::class_<Matrix3, Matrix3x3> matrix3{root, "Matrix3", "2D float transformation matrix", py::buffer_protocol{}};
-    py::class_<Matrix4, Matrix4x4> matrix4{root, "Matrix4", "3D float transformation matrix", py::buffer_protocol{}};
+    /* The subclasses don't have buffer protocol enabled, as that's already
+       done by the base classes. Moreover, just adding py::buffer_protocol{}
+       would cause it to not find the buffer functions as we don't add them
+       anywhere, thus failing with `pybind11_getbuffer(): Internal error`. */
+
+    py::class_<Matrix3, Matrix3x3> matrix3{root, "Matrix3", "2D float transformation matrix"};
+    py::class_<Matrix4, Matrix4x4> matrix4{root, "Matrix4", "3D float transformation matrix"};
 
     matrices<Float>(
         matrix2x2, matrix2x3, matrix2x4,
diff --git a/src/python/magnum/math.vector.h b/src/python/magnum/math.vector.h
index 00dbf11..2a3f4a8 100644
--- a/src/python/magnum/math.vector.h
+++ b/src/python/magnum/math.vector.h
@@ -26,53 +26,55 @@
 */
 
 #include <pybind11/operators.h>
+#include <Corrade/Containers/ScopeGuard.h>
 #include <Corrade/Utility/FormatStl.h>
 #include <Magnum/Math/Color.h>
 #include <Magnum/Math/Vector4.h>
 
 #include "corrade/PybindExtras.h"
+#include "corrade/PyBuffer.h"
 
 #include "magnum/math.h"
 
 namespace magnum {
 
-template<class> bool isTypeCompatible(const std::string&);
-template<> inline bool isTypeCompatible<Float>(const std::string& format) {
-    return format == "f" || format == "d";
+template<class> constexpr bool isTypeCompatible(char);
+template<> constexpr bool isTypeCompatible<Float>(char format) {
+    return format == 'f' || format == 'd';
 }
-template<> inline bool isTypeCompatible<Double>(const std::string& format) {
-    return format == "f" || format == "d";
+template<> constexpr bool isTypeCompatible<Double>(char format) {
+    return format == 'f' || format == 'd';
 }
-template<> inline bool isTypeCompatible<Int>(const std::string& format) {
-    return format == "i" || format == "l";
+template<> constexpr bool isTypeCompatible<Int>(char format) {
+    return format == 'i' || format == 'l';
 }
-template<> inline bool isTypeCompatible<UnsignedInt>(const std::string& format) {
-    return format == "I" || format == "L";
+template<> constexpr bool isTypeCompatible<UnsignedInt>(char format) {
+    return format == 'I' || format == 'L';
 }
 
-template<class U, class T> void initFromBuffer(T& out, const py::buffer_info& info) {
+template<class U, class T> void initFromBuffer(T& out, const Py_buffer& buffer) {
     for(std::size_t i = 0; i != T::Size; ++i)
-        out[i] = static_cast<typename T::Type>(*reinterpret_cast<const U*>(static_cast<const char*>(info.ptr) + i*info.strides[0]));
+        out[i] = static_cast<typename T::Type>(*reinterpret_cast<const U*>(static_cast<const char*>(buffer.buf) + i*buffer.strides[0]));
 }
 
 /* Floating-point init */
-template<class T> void initFromBuffer(T& out, const py::buffer_info& info, std::true_type, std::true_type) {
-    if(info.format == "f") initFromBuffer<Float>(out, info);
-    else if(info.format == "d") initFromBuffer<Double>(out, info);
+template<class T> void initFromBuffer(typename std::enable_if<std::is_floating_point<typename T::Type>::value, T>::type& out, const Py_buffer& buffer) {
+    if(buffer.format[0] == 'f') initFromBuffer<Float>(out, buffer);
+    else if(buffer.format[0] == 'd') initFromBuffer<Double>(out, buffer);
     else CORRADE_ASSERT_UNREACHABLE(); /* LCOV_EXCL_LINE */
 }
 
-/* Signed integeral init */
-template<class T> void initFromBuffer(T& out, const py::buffer_info& info, std::false_type, std::true_type) {
-    if(info.format == "i") initFromBuffer<Int>(out, info);
-    else if(info.format == "l") initFromBuffer<Long>(out, info);
+/* Signed integral init */
+template<class T> void initFromBuffer(typename std::enable_if<std::is_integral<typename T::Type>::value && std::is_signed<typename T::Type>::value, T>::type& out, const Py_buffer& buffer) {
+    if(buffer.format[0] == 'i') initFromBuffer<Int>(out, buffer);
+    else if(buffer.format[0] == 'l') initFromBuffer<Long>(out, buffer);
     else CORRADE_ASSERT_UNREACHABLE(); /* LCOV_EXCL_LINE */
 }
 
-/* Unsigned integeral init */
-template<class T> void initFromBuffer(T& out, const py::buffer_info& info, std::false_type, std::false_type) {
-    if(info.format == "I") initFromBuffer<UnsignedInt>(out, info);
-    else if(info.format == "L") initFromBuffer<UnsignedLong>(out, info);
+/* Unsigned integral init */
+template<class T> void initFromBuffer(typename std::enable_if<std::is_integral<typename T::Type>::value && std::is_unsigned<typename T::Type>::value, T>::type& out, const Py_buffer& buffer) {
+    if(buffer.format[0] == 'I') initFromBuffer<UnsignedInt>(out, buffer);
+    else if(buffer.format[0] == 'L') initFromBuffer<UnsignedLong>(out, buffer);
     else CORRADE_ASSERT_UNREACHABLE(); /* LCOV_EXCL_LINE */
 }
 
@@ -91,25 +93,6 @@ template<class T, class ...Args> void everyVector(py::class_<T, Args...>& c) {
         }, "Construct a zero vector")
         .def(py::init(), "Default constructor")
 
-        /* Ideally, only the constructor (in vectorBuffer()) would be needed
-           (and thus also no py::buffer_protocol() specified for the class),
-           but conversion of vectors to lists is extremely slow due to pybind
-           exceptions being somehow extra heavy compared to native python ones,
-           so in order to have acceptable performance we need the buffer
-           protocol on the other side as well. See test/benchmark_math.py for
-           more information. */
-        .def_buffer([](const T& self) -> py::buffer_info {
-            // TODO: ownership?
-            return py::buffer_info{
-                const_cast<typename T::Type*>(self.data()),
-                sizeof(typename T::Type),
-                py::format_descriptor<typename T::Type>::format(),
-                1,
-                {T::Size},
-                {sizeof(typename T::Type)}
-            };
-        })
-
         /* Operators */
         .def(-py::self, "Negated vector")
         .def(py::self += py::self, "Add and assign a vector")
@@ -135,24 +118,53 @@ template<class T, class ...Args> void vectorBuffer(py::class_<T, Args...>& c) {
         /* Buffer protocol. If not present, implicit conversion from numpy
            arrays of non-default types somehow doesn't work. There's also the
            other part in vectorBuffer(). */
-        .def(py::init([](py::buffer buffer) {
-            py::buffer_info info = buffer.request();
+        .def(py::init([](py::buffer other) {
+            Py_buffer buffer{};
+            if(PyObject_GetBuffer(other.ptr(), &buffer, PyBUF_FORMAT|PyBUF_STRIDES) != 0)
+                throw py::error_already_set{};
+
+            Containers::ScopeGuard e{&buffer, PyBuffer_Release};
 
-            if(info.ndim != 1)
-                throw py::buffer_error{Utility::formatString("expected 1 dimension but got {}", info.ndim)};
+            if(buffer.ndim != 1)
+                throw py::buffer_error{Utility::formatString("expected 1 dimension but got {}", buffer.ndim)};
 
-            if(info.shape[0] != T::Size)
-                throw py::buffer_error{Utility::formatString("expected {} elements but got {}", T::Size, info.shape[0])};
+            if(buffer.shape[0] != T::Size)
+                throw py::buffer_error{Utility::formatString("expected {} elements but got {}", T::Size, buffer.shape[0])};
 
-            if(!isTypeCompatible<typename T::Type>(info.format))
-                throw py::buffer_error{Utility::formatString("unexpected format {} for a {} vector", info.format, py::format_descriptor<typename T::Type>::format())};
+            /* Expecting just an one-letter format */
+            if(!buffer.format[0] || buffer.format[1] || !isTypeCompatible<typename T::Type>(buffer.format[0]))
+                throw py::buffer_error{Utility::formatString("unexpected format {} for a {} vector", buffer.format, FormatStrings[formatIndex<typename T::Type>()])};
 
             T out{Math::NoInit};
-            initFromBuffer(out, info, std::is_floating_point<typename T::Type>{}, std::is_signed<typename T::Type>{});
+            initFromBuffer<T>(out, buffer);
             return out;
         }), "Construct from a buffer");
 }
 
+template<class T> bool vectorBufferProtocol(T& self, Py_buffer& buffer, int flags) {
+    /* I hate the const_casts but I assume this is to make editing easier, NOT
+       to make it possible for users to stomp on these values. */
+    buffer.ndim = 1;
+    buffer.itemsize = sizeof(typename T::Type);
+    buffer.len = sizeof(T);
+    buffer.buf = self.data();
+    buffer.readonly = false;
+    if((flags & PyBUF_FORMAT) == PyBUF_FORMAT)
+        buffer.format = const_cast<char*>(FormatStrings[formatIndex<typename T::Type>()]);
+    if(flags != PyBUF_SIMPLE) {
+        /* Reusing shape definitions from matrices because I don't want to
+           create another useless array for that and reinterpret_cast on the
+           buffer.internal is UGLY. It's flipped from column-major to
+           row-major, so adjusting the row instead. */
+        buffer.shape = const_cast<Py_ssize_t*>(MatrixShapes[matrixShapeStrideIndex<2, T::Size>()]);
+        CORRADE_INTERNAL_ASSERT(buffer.shape[0] == T::Size);
+        if((flags & PyBUF_STRIDES) == PyBUF_STRIDES)
+            buffer.strides = &buffer.itemsize;
+    }
+
+    return true;
+}
+
 /* Things common for vectors of all sizes and types */
 template<class T> void vector(py::module& m, py::class_<T>& c) {
     /*
@@ -203,6 +215,15 @@ template<class T> void vector(py::module& m, py::class_<T>& c) {
 
         .def("__repr__", repr<T>, "Object representation");
 
+    /* Ideally, only the constructor (in vectorBuffer()) would be needed
+       (and thus also no py::buffer_protocol() specified for the class),
+       but conversion of vectors to lists is extremely slow due to pybind
+       exceptions being somehow extra heavy compared to native python ones,
+       so in order to have acceptable performance we need the buffer
+       protocol on the other side as well. See test/benchmark_math.py for more
+       information. */
+    corrade::enableBetterBufferProtocol<T, vectorBufferProtocol>(c);
+
     /* Vector length */
     char lenDocstring[] = "Vector size. Returns _.";
     lenDocstring[sizeof(lenDocstring) - 3] = '0' + T::Size;
diff --git a/src/python/magnum/math.vectorfloat.cpp b/src/python/magnum/math.vectorfloat.cpp
index de9c6db..e31a94d 100644
--- a/src/python/magnum/math.vectorfloat.cpp
+++ b/src/python/magnum/math.vectorfloat.cpp
@@ -89,12 +89,17 @@ void mathVectorFloat(py::module& root, py::module& m) {
     vectorsFloat<Float>(m, vector2, vector3, vector4);
     vectorsFloat<Double>(m, vector2d, vector3d, vector4d);
 
-    py::class_<Color3, Vector3> color3_{root, "Color3", "Color in linear RGB color space", py::buffer_protocol{}};
+    /* The subclasses don't have buffer protocol enabled, as that's already
+       done by the base classes. Moreover, just adding py::buffer_protocol{}
+       would cause it to not find the buffer functions as we don't add them
+       anywhere, thus failing with `pybind11_getbuffer(): Internal error`. */
+
+    py::class_<Color3, Vector3> color3_{root, "Color3", "Color in linear RGB color space"};
     everyVector(color3_);
     color(color3_);
     color3(color3_);
 
-    py::class_<Color4, Vector4> color4_{root, "Color4", "Color in linear RGBA color space", py::buffer_protocol{}};
+    py::class_<Color4, Vector4> color4_{root, "Color4", "Color in linear RGBA color space"};
     everyVector(color4_);
     color(color4_);
     color4(color4_);
diff --git a/src/python/magnum/test/test_math.py b/src/python/magnum/test/test_math.py
index aaf2515..df97c13 100644
--- a/src/python/magnum/test/test_math.py
+++ b/src/python/magnum/test/test_math.py
@@ -23,6 +23,7 @@
 #   DEALINGS IN THE SOFTWARE.
 #
 
+import array
 import unittest
 
 from magnum import *
@@ -263,6 +264,14 @@ class Vector(unittest.TestCase):
     def test_repr(self):
         self.assertEqual(repr(Vector3(1.0, 3.14, -13.37)), 'Vector(1, 3.14, -13.37)')
 
+    def test_from_buffer(self):
+        a = Vector3i(array.array('i', [2, 3, 5]))
+        self.assertEqual(a, Vector3i(2, 3, 5))
+
+    def test_to_buffer(self):
+        a = memoryview(Vector4(1.0, 2.0, 3.0, 4.0))
+        self.assertEqual(a.tolist(), [1.0, 2.0, 3.0, 4.0])
+
 class Color3_(unittest.TestCase):
     def test_init(self):
         a1 = Color3()
@@ -350,6 +359,16 @@ class Color4_(unittest.TestCase):
         self.assertIsInstance(Color4().rgb, Color3)
         self.assertIsInstance(Color4().xyz, Color3)
 
+    def test_from_buffer(self):
+        a = Color3(array.array('f', [2.0, 3.0, 5.0]))
+        self.assertEqual(a, Color3(2.0, 3.0, 5.0))
+
+    def test_to_buffer(self):
+        # Color4 doesn't define py::buffer_protocol(), the one from base should
+        # "just work"
+        a = memoryview(Color4(1.0, 2.0, 3.0, 4.0))
+        self.assertEqual(a.tolist(), [1.0, 2.0, 3.0, 4.0])
+
 class Matrix(unittest.TestCase):
     def test_init(self):
         a = Matrix3x2()
@@ -582,6 +601,13 @@ class Matrix(unittest.TestCase):
                                   '       2, 5,\n'
                                   '       3, 6)')
 
+    # conversion from buffer is tested in test_math_numpy, array.array is
+    # one-dimensional and I don't want to drag numpy here just for one test
+
+    def test_to_buffer(self):
+        a = memoryview(Matrix2x2((1.0, 2.0), (3.0, 4.0)))
+        self.assertEqual(a.tolist(), [[1.0, 3.0], [2.0, 4.0]])
+
 class Matrix3_(unittest.TestCase):
     def test_init(self):
         a = Matrix3()
@@ -673,6 +699,20 @@ class Matrix3_(unittest.TestCase):
         self.assertIsInstance(Matrix3().transposed(), Matrix3)
         self.assertIsInstance(Matrix3().inverted(), Matrix3)
 
+    # conversion from buffer is tested in test_math_numpy, array.array is
+    # one-dimensional and I don't want to drag numpy here just for one test
+
+    def test_to_buffer(self):
+        # Matrix3 doesn't define py::buffer_protocol(), the one from base
+        # should "just work"
+        a = memoryview(Matrix3((1.0, 2.0, 3.0),
+                               (4.0, 5.0, 6.0),
+                               (7.0, 8.0, 9.0)))
+        self.assertEqual(a.tolist(), [
+            [1.0, 4.0, 7.0],
+            [2.0, 5.0, 8.0],
+            [3.0, 6.0, 9.0]])
+
 class Matrix4_(unittest.TestCase):
     def test_init(self):
         a = Matrix4()
@@ -778,6 +818,22 @@ class Matrix4_(unittest.TestCase):
         self.assertIsInstance(Matrix4().transposed(), Matrix4)
         self.assertIsInstance(Matrix4().inverted(), Matrix4)
 
+    # conversion from buffer is tested in test_math_numpy, array.array is
+    # one-dimensional and I don't want to drag numpy here just for one test
+
+    def test_to_buffer(self):
+        # Matrix3 doesn't define py::buffer_protocol(), the one from base
+        # should "just work"
+        a = memoryview(Matrix4((1.0, 2.0, 3.0, 4.0),
+                               (5.0, 6.0, 7.0, 8.0),
+                               (9.0, 10.0, 11.0, 12.0),
+                               (13.0, 14.0, 15.0, 16.0)))
+        self.assertEqual(a.tolist(), [
+            [1.0, 5.0, 9.0, 13.0],
+            [2.0, 6.0, 10.0, 14.0],
+            [3.0, 7.0, 11.0, 15.0],
+            [4.0, 8.0, 12.0, 16.0]])
+
 class Quaternion_(unittest.TestCase):
     def test_init(self):
         a = Quaternion()