nv-sgl 0.6.0__cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

sgl/include/sgl/device/python/cursor_utils.h

@@ -0,0 +1,853 @@
+#pragma once
+
+#include <optional>
+
+#include "nanobind.h"
+
+#include "sgl/device/reflection.h"
+#include "sgl/device/cursor_utils.h"
+
+#include "sgl/math/vector_types.h"
+#include "sgl/math/matrix_types.h"
+
+namespace sgl {
+
+/// Helper to convert from numpy type mask to slang scalar type.
+inline std::optional<TypeReflection::ScalarType> dtype_to_scalar_type(nb::dlpack::dtype dtype)
+{
+    switch (dtype.code) {
+    case uint8_t(nb::dlpack::dtype_code::Int):
+        switch (dtype.bits) {
+        case 8:
+            return TypeReflection::ScalarType::int8;
+        case 16:
+            return TypeReflection::ScalarType::int16;
+        case 32:
+            return TypeReflection::ScalarType::int32;
+        case 64:
+            return TypeReflection::ScalarType::int64;
+        }
+        break;
+    case uint8_t(nb::dlpack::dtype_code::UInt):
+        switch (dtype.bits) {
+        case 8:
+            return TypeReflection::ScalarType::uint8;
+        case 16:
+            return TypeReflection::ScalarType::uint16;
+        case 32:
+            return TypeReflection::ScalarType::uint32;
+        case 64:
+            return TypeReflection::ScalarType::uint64;
+        }
+        break;
+    case uint8_t(nb::dlpack::dtype_code::Float):
+        switch (dtype.bits) {
+        case 16:
+            return TypeReflection::ScalarType::float16;
+        case 32:
+            return TypeReflection::ScalarType::float32;
+        case 64:
+            return TypeReflection::ScalarType::float64;
+        }
+        break;
+    }
+    return {};
+}
+
+/// Enforces properties needed for converting to/from vector.
+template<typename T>
+concept IsSpecializationOfVector = requires {
+    T::dimension;
+    typename T::value_type;
+};
+
+/// Enforces properties needed for converting to/from matrix.
+template<typename T>
+concept IsSpecializationOfMatrix = requires {
+    T::rows;
+    T::cols;
+    typename T::value_type;
+};
+
+#define scalar_case(c_type, scalar_type)                                                                               \
+    m_read_scalar[(int)TypeReflection::ScalarType::scalar_type]                                                        \
+        = [](const CursorType& self) { return _read_scalar<c_type>(self); };
+
+#define vector_case(c_type, scalar_type)                                                                               \
+    m_read_vector[(int)TypeReflection::ScalarType::scalar_type][c_type::dimension]                                     \
+        = [](const CursorType& self) { return _read_vector<c_type>(self); };
+
+#define matrix_case(c_type, scalar_type)                                                                               \
+    m_read_matrix[(int)TypeReflection::ScalarType::scalar_type][c_type::rows][c_type::cols]                            \
+        = [](const CursorType& self) { return _read_matrix<c_type>(self); };
+
+/// Table of converters based on slang scalar type and shape.
+template<typename CursorType>
+class ReadConverterTable {
+public:
+    ReadConverterTable()
+    {
+        // Initialize all entries to an error function that throws an exception.
+        auto read_err_func = [](const CursorType&) -> nb::object { SGL_THROW("Unsupported element type"); };
+        for (int i = 0; i < (int)TypeReflection::ScalarType::COUNT; i++) {
+            m_read_scalar[i] = read_err_func;
+            for (int j = 0; j < 5; ++j) {
+                m_read_vector[i][j] = read_err_func;
+                for (int k = 0; k < 5; ++k) {
+                    m_read_matrix[i][j][k] = read_err_func;
+                }
+            }
+        }
+
+        // Register converters for all supported scalar types.
+        scalar_case(bool, bool_);
+        scalar_case(int8_t, int8);
+        scalar_case(uint8_t, uint8);
+        scalar_case(int16_t, int16);
+        scalar_case(uint16_t, uint16);
+        scalar_case(int32_t, int32);
+        scalar_case(uint32_t, uint32);
+        scalar_case(int64_t, int64);
+        scalar_case(uint64_t, uint64);
+        scalar_case(float16_t, float16);
+        scalar_case(float, float32);
+        scalar_case(double, float64);
+        scalar_case(intptr_t, intptr);
+        scalar_case(uintptr_t, uintptr);
+
+        // Register converters for all supported vector types.
+        vector_case(bool1, bool_);
+        vector_case(float1, float32);
+        vector_case(float16_t1, float16);
+        vector_case(int1, int32);
+        vector_case(uint1, uint32);
+        vector_case(bool2, bool_);
+        vector_case(float2, float32);
+        vector_case(float16_t2, float16);
+        vector_case(int2, int32);
+        vector_case(uint2, uint32);
+        vector_case(bool3, bool_);
+        vector_case(float3, float32);
+        vector_case(float16_t3, float16);
+        vector_case(int3, int32);
+        vector_case(uint3, uint32);
+        vector_case(bool4, bool_);
+        vector_case(float4, float32);
+        vector_case(float16_t4, float16);
+        vector_case(int4, int32);
+        vector_case(uint4, uint32);
+
+        // Register converters for all supported matrix types.
+        matrix_case(float2x2, float32);
+        matrix_case(float3x3, float32);
+        matrix_case(float2x4, float32);
+        matrix_case(float3x4, float32);
+        matrix_case(float4x4, float32);
+    }
+
+    /// Read function inspects the slang type and attempts to convert it
+    /// to a matching python type. For structs and arrays, generates
+    /// a nested dictionary or list and recurses.
+    nb::object read(const CursorType& self)
+    {
+        m_stack.clear();
+        try {
+            return read_internal(self);
+        } catch (const std::exception& err) {
+            SGL_THROW("{}: {}", build_error(), err.what());
+        }
+    }
+
+private:
+    std::function<nb::object(const CursorType&)> m_read_scalar[(int)TypeReflection::ScalarType::COUNT];
+    std::function<nb::object(const CursorType&)> m_read_vector[(int)TypeReflection::ScalarType::COUNT][5];
+    std::function<nb::object(const CursorType&)> m_read_matrix[(int)TypeReflection::ScalarType::COUNT][5][5];
+    std::vector<const char*> m_stack;
+
+    std::string build_error() { return fmt::format("{}", fmt::join(m_stack, ".")); }
+
+    nb::object read_internal(const CursorType& self)
+    {
+        if (!self.is_valid())
+            return nb::none();
+        auto type = self.type();
+        if (type) {
+            switch (type->kind()) {
+            case TypeReflection::Kind::scalar: {
+                return m_read_scalar[(int)type->scalar_type()](self);
+            }
+            case TypeReflection::Kind::vector: {
+                return m_read_vector[(int)type->scalar_type()][type->col_count()](self);
+            }
+            case TypeReflection::Kind::matrix: {
+                return m_read_matrix[(int)type->scalar_type()][type->row_count()][type->col_count()](self);
+            }
+            case TypeReflection::Kind::struct_: {
+                nb::dict res;
+                for (uint32_t i = 0; i < type->field_count(); i++) {
+                    auto field = type->get_field_by_index(i);
+                    const char* name = field->name();
+                    m_stack.push_back(name);
+                    res[name] = read_internal(self[name]);
+                    m_stack.pop_back();
+                }
+                return res;
+            }
+            case TypeReflection::Kind::array: {
+                nb::list res;
+                m_stack.push_back("[]");
+                for (uint32_t i = 0; i < type->element_count(); i++) {
+                    res.append(read_internal(self[i]));
+                }
+                m_stack.pop_back();
+                return res;
+            }
+            default:
+                break;
+            }
+        }
+        SGL_THROW("Unsupported element type");
+    }
+
+    /// Read scalar value from buffer element cursor and convert to Python object.
+    template<typename ValType>
+    inline static nb::object _read_scalar(const CursorType& self)
+    {
+        ValType res;
+        self.get(res);
+        return nb::cast(res);
+    }
+
+    /// Read vector value from buffer element cursor and convert to Python object.
+    template<typename ValType>
+        requires IsSpecializationOfVector<ValType>
+    inline static nb::object _read_vector(const CursorType& self)
+    {
+        ValType res;
+        self.get(res);
+        return nb::cast(res);
+    }
+
+    /// Read matrix value from buffer element cursor and convert to Python object.
+    template<typename ValType>
+        requires IsSpecializationOfMatrix<ValType>
+    inline static nb::object _read_matrix(const CursorType& self)
+    {
+        ValType res;
+        self.get(res);
+        return nb::cast(res);
+    }
+};
+
+#undef scalar_case
+#undef vector_case
+#undef matrix_case
+
+#define scalar_case(c_type, scalar_type)                                                                               \
+    m_write_scalar[(int)TypeReflection::ScalarType::scalar_type]                                                       \
+        = [](CursorType& self, nb::object nbval) { _write_scalar<c_type>(self, nbval); };
+
+#define vector_case(c_type, scalar_type)                                                                               \
+    m_write_vector[(int)TypeReflection::ScalarType::scalar_type][c_type::dimension]                                    \
+        = [](CursorType& self, nb::object nbval) { _write_vector<c_type>(self, nbval); };
+
+#define bool_vector_case(c_type, scalar_type)                                                                          \
+    m_write_vector[(int)TypeReflection::ScalarType::scalar_type][c_type::dimension]                                    \
+        = [](CursorType& self, nb::object nbval) { _write_bool_vector<c_type>(self, nbval); };
+
+#define matrix_case(c_type, scalar_type)                                                                               \
+    m_write_matrix[(int)TypeReflection::ScalarType::scalar_type][c_type::rows][c_type::cols]                           \
+        = [](CursorType& self, nb::object nbval) { _write_matrix<c_type>(self, nbval); };
+
+/// Table of converters based on slang scalar type and shape.
+template<typename CursorType>
+class WriteConverterTable {
+public:
+    WriteConverterTable()
+    {
+        // Initialize all entries to an error function that throws an exception.
+        auto write_err_func = [](const CursorType&, nb::object) { SGL_THROW("Unsupported element type"); };
+        for (int i = 0; i < (int)TypeReflection::ScalarType::COUNT; i++) {
+            m_write_scalar[i] = write_err_func;
+            for (int j = 0; j < 5; ++j) {
+                m_write_vector[i][j] = write_err_func;
+                for (int k = 0; k < 5; ++k) {
+                    m_write_matrix[i][j][k] = write_err_func;
+                }
+            }
+        }
+
+        // Register converters for all supported scalar types.
+        scalar_case(bool, bool_);
+        scalar_case(int8_t, int8);
+        scalar_case(uint8_t, uint8);
+        scalar_case(int16_t, int16);
+        scalar_case(uint16_t, uint16);
+        scalar_case(int32_t, int32);
+        scalar_case(uint32_t, uint32);
+        scalar_case(int64_t, int64);
+        scalar_case(uint64_t, uint64);
+        scalar_case(float16_t, float16);
+        scalar_case(float, float32);
+        scalar_case(double, float64);
+        scalar_case(intptr_t, intptr);
+        scalar_case(uintptr_t, uintptr);
+
+        // Register converters for all supported vector types.
+        bool_vector_case(bool1, bool_);
+        vector_case(float1, float32);
+        vector_case(float16_t1, float16);
+        vector_case(int1, int32);
+        vector_case(uint1, uint32);
+        bool_vector_case(bool2, bool_);
+        vector_case(float2, float32);
+        vector_case(float16_t2, float16);
+        vector_case(int2, int32);
+        vector_case(uint2, uint32);
+        bool_vector_case(bool3, bool_);
+        vector_case(float3, float32);
+        vector_case(float16_t3, float16);
+        vector_case(int3, int32);
+        vector_case(uint3, uint32);
+        bool_vector_case(bool4, bool_);
+        vector_case(float4, float32);
+        vector_case(float16_t4, float16);
+        vector_case(int4, int32);
+        vector_case(uint4, uint32);
+
+        // Register converters for all supported matrix types.
+        matrix_case(float2x2, float32);
+        matrix_case(float3x3, float32);
+        matrix_case(float2x4, float32);
+        matrix_case(float3x4, float32);
+        matrix_case(float4x4, float32);
+    }
+
+    /// Virtual for writing none-basic value types.
+    virtual bool write_value(CursorType& self, nb::object nbval)
+    {
+        SGL_UNUSED(self);
+        SGL_UNUSED(nbval);
+        return false;
+    }
+
+    /// Write function inspects the slang type and uses it to try
+    /// and convert a Python input to the correct c++ type. For structs
+    /// and arrays, expects a dict, sequence type or numpy array.
+    void write(CursorType& self, nb::object nbval)
+    {
+        m_stack.clear();
+        try {
+            write_internal(self, nbval);
+        } catch (const std::exception& err) {
+            SGL_THROW("{}: {}", build_error(), err.what());
+        }
+    }
+
+private:
+    std::function<void(CursorType&, nb::object)> m_write_scalar[(int)TypeReflection::ScalarType::COUNT];
+    std::function<void(CursorType&, nb::object)> m_write_vector[(int)TypeReflection::ScalarType::COUNT][5];
+    std::function<void(CursorType&, nb::object)> m_write_matrix[(int)TypeReflection::ScalarType::COUNT][5][5];
+    std::vector<const char*> m_stack;
+
+    std::string build_error() { return fmt::format("{}", fmt::join(m_stack, ".")); }
+
+    void write_internal(CursorType& self, nb::object nbval)
+    {
+        if (!self.is_valid())
+            return;
+
+        ref<const TypeLayoutReflection> type_layout = self.type_layout();
+        auto kind = type_layout->kind();
+
+        switch (kind) {
+        case TypeReflection::Kind::scalar: {
+            auto type = type_layout->type();
+            SGL_ASSERT(type);
+            return m_write_scalar[(int)type->scalar_type()](self, nbval);
+        }
+        case TypeReflection::Kind::vector: {
+            auto type = type_layout->type();
+            SGL_ASSERT(type);
+            return m_write_vector[(int)type->scalar_type()][type->col_count()](self, nbval);
+        }
+        case TypeReflection::Kind::matrix: {
+            auto type = type_layout->type();
+            SGL_ASSERT(type);
+            return m_write_matrix[(int)type->scalar_type()][type->row_count()][type->col_count()](self, nbval);
+        }
+        case TypeReflection::Kind::constant_buffer:
+        case TypeReflection::Kind::parameter_block:
+        case TypeReflection::Kind::struct_: {
+            // Unwrap constant buffers or parameter blocks
+            if (kind != TypeReflection::Kind::struct_)
+                type_layout = type_layout->element_type_layout();
+
+            // Handle shader object if possible.
+            if constexpr (requires { self.set_object(nullptr); }) {
+                if (nb::isinstance<MutableShaderObject>(nbval)) {
+                    self.set_object(nb::cast<ref<MutableShaderObject>>(nbval));
+                    return;
+                }
+            }
+
+            // Expect a dict for a slang struct.
+            if (nb::isinstance<nb::dict>(nbval)) {
+                auto dict = nb::cast<nb::dict>(nbval);
+                for (uint32_t i = 0; i < type_layout->field_count(); i++) {
+                    auto field = type_layout->get_field_by_index(i);
+                    const char* name = field->name();
+                    auto child = self[name];
+                    if (dict.contains(name)) {
+                        m_stack.push_back(name);
+                        write_internal(child, dict[name]);
+                        m_stack.pop_back();
+                    }
+                }
+                return;
+            } else {
+                SGL_THROW("Expected dict");
+            }
+        }
+        case TypeReflection::Kind::array: {
+            // Expect numpy array or sequence for a slang array.
+            if (nb::isinstance<nb::ndarray<nb::numpy>>(nbval)) {
+                // TODO: Should be able to do better job of interpreting nb array values by reading
+                // data type and extracting individual elements.
+                auto nbarray = nb::cast<nb::ndarray<nb::numpy>>(nbval);
+                SGL_CHECK(nbarray.ndim() == 1, "numpy array must have 1 dimension.");
+                SGL_CHECK(nbarray.shape(0) == type_layout->element_count(), "numpy array is the wrong length.");
+                SGL_CHECK(is_ndarray_contiguous(nbarray), "data is not contiguous");
+                self._set_array(
+                    nbarray.data(),
+                    nbarray.nbytes(),
+                    type_layout->element_type_layout()->type()->scalar_type(),
+                    narrow_cast<int>(nbarray.shape(0))
+                );
+                return;
+            } else if (nb::isinstance<nb::sequence>(nbval)) {
+                auto seq = nb::cast<nb::sequence>(nbval);
+                SGL_CHECK(
+                    nb::len(seq) == type_layout->element_count(),
+                    "sequence is the wrong length accessing type {}: {} != {}.",
+                    type_layout->type()->full_name(),
+                    nb::len(seq),
+                    type_layout->element_count()
+                );
+                m_stack.push_back("[]");
+                for (uint32_t i = 0; i < type_layout->element_count(); i++) {
+                    auto child = self[i];
+                    write_internal(child, seq[i]);
+                }
+                m_stack.pop_back();
+                return;
+            } else {
+                SGL_THROW("Expected list");
+            }
+        }
+        default:
+            break;
+        }
+
+        // In default case call the virtual write_value, and fail if it returns false.
+        if (write_value(self, nbval))
+            return;
+
+        SGL_THROW("Unsupported element type");
+    }
+
+
+    /// Write scalar value to buffer element cursor from Python object.
+    template<typename ValType>
+    inline static void _write_scalar(CursorType& self, nb::object nbval)
+    {
+        auto val = nb::cast<ValType>(nbval);
+        self.set(val);
+    }
+
+    /// Default implementation of write vector from numpy array.
+    template<typename ValType>
+        requires IsSpecializationOfVector<ValType>
+    inline static void _write_vector_from_numpy(CursorType& self, nb::ndarray<nb::numpy> nbarray)
+    {
+        SGL_CHECK(nbarray.nbytes() == sizeof(ValType), "numpy array has wrong size.");
+        auto val = *reinterpret_cast<const ValType*>(nbarray.data());
+        self.set(val);
+    }
+
+    /// Version of vector write specifically for bool vectors (which are stored as uint32_t)
+    template<typename ValType>
+        requires IsSpecializationOfVector<ValType>
+    inline static void _write_bool_vector_from_numpy(CursorType& self, nb::ndarray<nb::numpy> nbarray)
+    {
+        SGL_CHECK(nbarray.nbytes() == ValType::dimension * 4, "numpy array has wrong size.");
+        self._set_vector(nbarray.data(), nbarray.nbytes(), TypeReflection::ScalarType::bool_, ValType::dimension);
+    }
+
+    /// Write vector value to buffer element cursor from Python object
+    template<typename ValType>
+        requires IsSpecializationOfVector<ValType>
+    inline static void _write_vector(CursorType& self, nb::object nbval)
+    {
+        if (nb::isinstance<ValType>(nbval)) {
+            // A vector of the correct type - just convert it.
+            auto val = nb::cast<ValType>(nbval);
+            self.set(val);
+        } else if (nb::isinstance<nb::ndarray<nb::numpy>>(nbval)) {
+            // A numpy array. Reinterpret numpy memory as vector type.
+            nb::ndarray<nb::numpy> nbarray = nb::cast<nb::ndarray<nb::numpy>>(nbval);
+            SGL_CHECK(is_ndarray_contiguous(nbarray), "data is not contiguous");
+            SGL_CHECK(nbarray.ndim() == 1 || nbarray.ndim() == 2, "numpy array must have 1 or 2 dimensions.");
+            size_t dimension = 1;
+            for (size_t i = 0; i < nbarray.ndim(); ++i)
+                dimension *= nbarray.shape(i);
+            SGL_CHECK(dimension == ValType::dimension, "numpy array has wrong dimension.");
+            _write_vector_from_numpy<ValType>(self, nbarray);
+        } else if (nb::isinstance<nb::sequence>(nbval)) {
+            // A list or tuple. Attempt to cast each element of list to element of vector.
+            auto seq = nb::cast<nb::sequence>(nbval);
+            SGL_CHECK(nb::len(seq) == ValType::dimension, "sequence has wrong dimension.");
+            ValType val;
+            for (int i = 0; i < ValType::dimension; i++) {
+                val[i] = nb::cast<typename ValType::value_type>(seq[i]);
+            }
+            self.set(val);
+        } else {
+            SGL_THROW("Expected numpy array or vector");
+        }
+    }
+
+    /// Bespoke vector implementation for bools.
+    template<typename ValType>
+        requires IsSpecializationOfVector<ValType>
+    inline static void _write_bool_vector(CursorType& self, nb::object nbval)
+    {
+        if (nb::isinstance<ValType>(nbval)) {
+            // A vector of the correct type - just convert it.
+            auto val = nb::cast<ValType>(nbval);
+            self.set(val);
+        } else if (nb::isinstance<nb::ndarray<nb::numpy>>(nbval)) {
+            // A numpy array. Reinterpret numpy memory as vector type.
+            nb::ndarray<nb::numpy> nbarray = nb::cast<nb::ndarray<nb::numpy>>(nbval);
+            SGL_CHECK(is_ndarray_contiguous(nbarray), "data is not contiguous");
+            SGL_CHECK(nbarray.ndim() == 1 || nbarray.ndim() == 2, "numpy array must have 1 or 2 dimensions.");
+            size_t dimension = 1;
+            for (size_t i = 0; i < nbarray.ndim(); ++i)
+                dimension *= nbarray.shape(i);
+            SGL_CHECK(dimension == ValType::dimension, "numpy array has wrong dimension.");
+            _write_bool_vector_from_numpy<ValType>(self, nbarray);
+        } else if (nb::isinstance<nb::sequence>(nbval)) {
+            // A list or tuple. Attempt to cast each element of list to element of vector.
+            auto seq = nb::cast<nb::sequence>(nbval);
+            SGL_CHECK(nb::len(seq) == ValType::dimension, "sequence has wrong dimension.");
+            ValType val;
+            for (int i = 0; i < ValType::dimension; i++) {
+                val[i] = nb::cast<typename ValType::value_type>(seq[i]);
+            }
+            self.set(val);
+        } else {
+            SGL_THROW("Expected numpy array or vector");
+        }
+    }
+
+    /// Write matrix value to buffer element cursor from Python object.
+    template<typename ValType>
+        requires IsSpecializationOfMatrix<ValType>
+    inline static void _write_matrix(CursorType& self, nb::object nbval)
+    {
+        if (nb::isinstance<ValType>(nbval)) {
+            // Matrix of correct type
+            auto val = nb::cast<ValType>(nbval);
+            self.set(val);
+        } else if (nb::isinstance<nb::ndarray<nb::numpy>>(nbval)) {
+            // A numpy array. We have a python cast from numpy->matrix,
+            // so can just call it here to convert properly.
+            auto val = nb::cast<ValType>(nbval);
+            self.set(val);
+        } else {
+            SGL_THROW("Expected numpy array or matrix");
+        }
+    }
+};
+
+#undef scalar_case
+#undef vector_case
+#undef bool_vector_case
+#undef matrix_case
+
+
+template<typename CursorType>
+    requires TraversableCursor<CursorType>
+inline void bind_traversable_cursor(nanobind::class_<CursorType>& cursor)
+{
+    cursor //
+        .def_prop_ro("_type_layout", &CursorType::type_layout, D_NA(CursorType, type_layout))
+        .def_prop_ro("_type", &CursorType::type, D_NA(CursorType, type))
+        .def("is_valid", &CursorType::is_valid, D_NA(CursorType, is_valid))
+        .def("find_field", &CursorType::find_field, "name"_a, D_NA(CursorType, find_field))
+        .def("find_element", &CursorType::find_element, "index"_a, D_NA(CursorType, find_element))
+        .def("has_field", &CursorType::has_field, "name"_a, D_NA(CursorType, has_field))
+        .def("has_element", &CursorType::has_element, "index"_a, D_NA(CursorType, has_element))
+        .def("__getitem__", [](CursorType& self, std::string_view name) { return self[name]; })
+        .def("__getitem__", [](CursorType& self, int index) { return self[index]; })
+        // note: __getattr__ should not except if field is not found
+        .def("__getattr__", [](CursorType& self, std::string_view name) { return self.find_field(name); });
+}
+
+template<typename CursorType>
+inline void bind_writable_cursor(WriteConverterTable<CursorType>& table, nanobind::class_<CursorType>& cursor)
+{
+    // __setitem__ and __setattr__ functions are overloaded to allow direct setting
+    // of fields and elements.
+    cursor //
+        .def(
+            "__setattr__",
+            [&table](CursorType& self, std::string_view name, nb::object nbval)
+            {
+                auto child = self[name];
+                table.write(child, nbval);
+            },
+            "name"_a,
+            "val"_a,
+            D_NA(CursorType, write)
+        )
+        .def(
+            "__setitem__",
+            [&table](CursorType& self, std::string_view name, nb::object nbval)
+            {
+                auto child = self[name];
+                table.write(child, nbval);
+            },
+            "index"_a,
+            "val"_a,
+            D_NA(CursorType, write)
+        )
+        .def(
+            "__setitem__",
+            [&table](CursorType& self, int index, nb::object nbval)
+            {
+                auto child = self[index];
+                table.write(child, nbval);
+            },
+            "index"_a,
+            "val"_a,
+            D_NA(CursorType, write)
+        )
+        .def(
+            "set_data",
+            [](CursorType& self, nb::ndarray<nb::device::cpu> data)
+            {
+                SGL_CHECK(is_ndarray_contiguous(data), "data is not contiguous");
+                self.set_data(data.data(), data.nbytes());
+            },
+            "data"_a,
+            D(ShaderCursor, set_data)
+        )
+        .def(
+            "write",
+            [&table](CursorType& self, nb::object nbval) { table.write(self, nbval); },
+            "val"_a,
+            D_NA(CursorType, write)
+        );
+}
+
+template<typename CursorType>
+inline void bind_readable_cursor(ReadConverterTable<CursorType>& table, nanobind::class_<CursorType>& cursor)
+{
+    // __setitem__ and __setattr__ functions are overloaded to allow direct setting
+    // of fields and elements.
+    cursor //
+        .def(
+            "read",
+            [&table](CursorType& self) { return table.read(self); },
+            D_NA(CursorType, read)
+        );
+}
+
+template<typename CursorType>
+    requires WritableCursor<CursorType>
+inline void bind_writable_cursor_basic_types(nanobind::class_<CursorType>& cursor)
+{
+#define def_setter(type)                                                                                               \
+    cursor.def("__setitem__", [](CursorType& self, std::string_view name, type value) { self[name] = value; });        \
+    cursor.def("__setattr__", [](CursorType& self, std::string_view name, type value) { self[name] = value; });
+
+    def_setter(bool);
+    def_setter(bool1);
+    def_setter(bool2);
+    def_setter(bool3);
+    def_setter(bool4);
+
+    def_setter(uint1);
+    def_setter(uint2);
+    def_setter(uint3);
+    def_setter(uint4);
+
+    def_setter(int1);
+    def_setter(int2);
+    def_setter(int3);
+    def_setter(int4);
+
+    def_setter(float1);
+    def_setter(float2);
+    def_setter(float3);
+    def_setter(float4);
+
+    def_setter(float2x2);
+    def_setter(float3x3);
+    def_setter(float2x4);
+    def_setter(float3x4);
+    def_setter(float4x4);
+
+    def_setter(float16_t2);
+    def_setter(float16_t3);
+    def_setter(float16_t4);
+
+#undef def_setter
+
+    auto set_int_field = [](CursorType& self, std::string_view name, nb::int_ value)
+    {
+        ref<const TypeReflection> type = self[name].type();
+        SGL_CHECK(type->kind() == TypeReflection::Kind::scalar, "Field \"{}\" is not a scalar type.", name);
+        switch (type->scalar_type()) {
+        case TypeReflection::ScalarType::int16:
+            self[name] = nb::cast<int16_t>(value);
+            break;
+        case TypeReflection::ScalarType::int32:
+            self[name] = nb::cast<int32_t>(value);
+            break;
+        case TypeReflection::ScalarType::int64:
+            self[name] = nb::cast<int64_t>(value);
+            break;
+        case TypeReflection::ScalarType::uint16:
+            self[name] = nb::cast<uint16_t>(value);
+            break;
+        case TypeReflection::ScalarType::uint32:
+            self[name] = nb::cast<uint32_t>(value);
+            break;
+        case TypeReflection::ScalarType::uint64:
+            self[name] = nb::cast<uint64_t>(value);
+            break;
+        default:
+            SGL_THROW("Field \"{}\" is not an integer type.");
+            break;
+        }
+    };
+
+    auto set_int_element = [](CursorType& self, int index, nb::int_ value)
+    {
+        ref<const TypeReflection> type = self[index].type();
+        SGL_CHECK(type->kind() == TypeReflection::Kind::scalar, "Element {} is not a scalar type.", index);
+        switch (type->scalar_type()) {
+        case TypeReflection::ScalarType::int16:
+            self[index] = nb::cast<int16_t>(value);
+            break;
+        case TypeReflection::ScalarType::int32:
+            self[index] = nb::cast<int32_t>(value);
+            break;
+        case TypeReflection::ScalarType::int64:
+            self[index] = nb::cast<int64_t>(value);
+            break;
+        case TypeReflection::ScalarType::uint16:
+            self[index] = nb::cast<uint16_t>(value);
+            break;
+        case TypeReflection::ScalarType::uint32:
+            self[index] = nb::cast<uint32_t>(value);
+            break;
+        case TypeReflection::ScalarType::uint64:
+            self[index] = nb::cast<uint64_t>(value);
+            break;
+        default:
+            SGL_THROW("Element {} is not an integer type.");
+            break;
+        }
+    };
+
+    cursor.def("__setitem__", set_int_field);
+    cursor.def("__setitem__", set_int_element);
+    cursor.def("__setattr__", set_int_field);
+
+    auto set_float_field = [](CursorType& self, std::string_view name, nb::float_ value)
+    {
+        ref<const TypeReflection> type = self[name].type();
+        SGL_CHECK(type->kind() == TypeReflection::Kind::scalar, "Field \"{}\" is not a scalar type.", name);
+        switch (type->scalar_type()) {
+        case TypeReflection::ScalarType::float16:
+            self[name] = float16_t(nb::cast<float>(value));
+            break;
+        case TypeReflection::ScalarType::float32:
+            self[name] = nb::cast<float>(value);
+            break;
+        case TypeReflection::ScalarType::float64:
+            self[name] = nb::cast<double>(value);
+            break;
+        default:
+            SGL_THROW("Field \"{}\" is not a floating point type.");
+            break;
+        }
+    };
+
+    auto set_float_element = [](CursorType& self, int index, nb::float_ value)
+    {
+        ref<const TypeReflection> type = self[index].type();
+        SGL_CHECK(type->kind() == TypeReflection::Kind::scalar, "Element {} is not a scalar type.", index);
+        switch (type->scalar_type()) {
+        case TypeReflection::ScalarType::float16:
+            self[index] = float16_t(nb::cast<float>(value));
+            break;
+        case TypeReflection::ScalarType::float32:
+            self[index] = nb::cast<float>(value);
+            break;
+        case TypeReflection::ScalarType::float64:
+            self[index] = nb::cast<double>(value);
+            break;
+        default:
+            SGL_THROW("Element {} is not a floating point type.");
+            break;
+        }
+    };
+
+    cursor.def("__setitem__", set_float_field);
+    cursor.def("__setitem__", set_float_element);
+    cursor.def("__setattr__", set_float_field);
+
+    auto set_numpy_field = [](CursorType& self, std::string_view name, nb::ndarray<nb::numpy> value)
+    {
+        ref<const TypeReflection> type = self[name].type();
+        auto src_scalar_type = dtype_to_scalar_type(value.dtype());
+        SGL_CHECK(src_scalar_type, "numpy array has unsupported dtype.");
+        SGL_CHECK(is_ndarray_contiguous(value), "numpy array is not contiguous.");
+
+        switch (type->kind()) {
+        case TypeReflection::Kind::array:
+            SGL_CHECK(value.ndim() == 1, "numpy array must have 1 dimension.");
+            self[name]._set_array(value.data(), value.nbytes(), *src_scalar_type, narrow_cast<int>(value.shape(0)));
+            break;
+        case TypeReflection::Kind::matrix:
+            SGL_CHECK(value.ndim() == 2, "numpy array must have 2 dimensions.");
+            self[name]._set_matrix(
+                value.data(),
+                value.nbytes(),
+                *src_scalar_type,
+                narrow_cast<int>(value.shape(0)),
+                narrow_cast<int>(value.shape(1))
+            );
+            break;
+        case TypeReflection::Kind::vector: {
+            SGL_CHECK(value.ndim() == 1 || value.ndim() == 2, "numpy array must have 1 or 2 dimensions.");
+            size_t dimension = 1;
+            for (size_t i = 0; i < value.ndim(); ++i)
+                dimension *= value.shape(i);
+            self[name]._set_vector(value.data(), value.nbytes(), *src_scalar_type, narrow_cast<int>(dimension));
+            break;
+        }
+        default:
+            SGL_THROW("Field \"{}\" is not a vector, matrix, or array type.", name);
+        }
+    };
+
+    cursor.def("__setitem__", set_numpy_field);
+    cursor.def("__setattr__", set_numpy_field);
+}
+
+} // namespace sgl