cubecl_ir/

properties.rs

use alloc::string::String;
use core::{
    fmt,
    hash::{BuildHasher, Hash, Hasher},
    str::FromStr,
};

use crate::{
    AddressType, ElemType, EnumSet, EnumSetType, OpaqueType, SemanticType, Type, TypeHash,
    VectorSize,
    features::{AtomicUsage, ComplexUsage, Features, TypeUsage},
};
use cubecl_common::profile::TimingMethod;

/// Properties of the device related to the accelerator hardware.
///
/// # Plane size min/max
///
/// This is a range of possible values for the plane size.
///
/// For Nvidia GPUs and HIP, this is a single fixed value.
///
/// For wgpu with AMD GPUs this is a range of possible values, but the actual configured value
/// is undefined and can only be queried at runtime. Should usually be 32, but not guaranteed.
///
/// For Intel GPUs, this is variable based on the number of registers used in the kernel. No way to
/// query this at compile time is currently available. As a result, the minimum value should usually
/// be assumed.
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub struct HardwareProperties {
    /// The widest single load instruction, in bits.
    pub load_width: u32,
    /// How many `load_width`-bit vector registers a kernel may keep live, or `None` where the
    /// runtime states no budget, as a GPU does.
    pub vector_register_count: Option<u32>,
    /// The minimum size of a plane on this device
    pub plane_size_min: u32,
    /// The maximum size of a plane on this device
    pub plane_size_max: u32,
    /// minimum number of bindings for a kernel that can be used at once.
    pub max_bindings: u32,
    /// Maximum amount of shared memory, in bytes
    pub max_shared_memory_size: usize,
    /// Maximum `CubeCount` in x, y and z dimensions
    pub max_cube_count: (u32, u32, u32),
    /// Maximum number of total units in a cube
    pub max_units_per_cube: u32,
    /// Maximum `CubeDim` in x, y, and z dimensions
    pub max_cube_dim: (u32, u32, u32),
    /// Number of streaming multiprocessors (SM), if available
    pub num_streaming_multiprocessors: Option<u32>,
    /// Number of available parallel cpu units, if the runtime is CPU.
    pub num_cpu_cores: Option<u32>,
    /// Bytes of the device's last level cache, and `None`, never `Some(0)`,
    /// where the runtime cannot read one.
    ///
    /// The size a working set has to outgrow before what it reaches is set by
    /// memory rather than by the chip.
    pub last_level_cache_size: Option<usize>,
    /// Number of tensor cores per SM, if any
    pub num_tensor_cores: Option<u32>,
    /// The minimum tiling dimension for a single axis in tensor cores.
    ///
    /// For a backend that only supports 16x16x16, the value would be 16.
    /// For a backend that also supports 32x8x16, the value would be 8.
    pub min_tensor_cores_dim: Option<u32>,
    /// Maximum vector size supported by the device
    pub max_vector_size: VectorSize,
    /// Memory reserved for the driver when using cube-scoped matrices
    pub cube_mma_reserved_shared_memory: usize,
}

/// A device's vector registers, counted in elements of one type.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct VectorRegisters {
    count: usize,
    register_vector_size: VectorSize,
    max_vector_size: VectorSize,
}

impl VectorRegisters {
    /// The registers as `elem_size`-byte elements see them, or `None` without a register count.
    pub fn new(hardware: &HardwareProperties, elem_size: usize) -> Option<Self> {
        Some(VectorRegisters {
            count: hardware.vector_register_count? as usize,
            register_vector_size: hardware.vector_size_in(hardware.load_width, elem_size),
            max_vector_size: prev_power_of_two(hardware.max_vector_size.max(1)),
        })
    }

    pub fn count(&self) -> usize {
        self.count
    }

    /// Registers a vector of `vector_size` elements occupies, at least one.
    pub fn registers_for(&self, vector_size: VectorSize) -> usize {
        vector_size.div_ceil(self.register_vector_size).max(1)
    }

    /// The widest power-of-two size at which `live` vectors stay in registers, at least one.
    pub fn widest_vector_size(&self, live: usize) -> VectorSize {
        let registers = prev_power_of_two((self.count / live.max(1)).max(1));
        (registers * self.register_vector_size).min(self.max_vector_size)
    }

    /// How many vectors of `vector_size` elements fit beside `reserved` registers.
    pub fn vectors_fitting(&self, vector_size: VectorSize, reserved: usize) -> usize {
        self.count.saturating_sub(reserved) / self.registers_for(vector_size)
    }
}

/// Properties of the device related to allocation.
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
#[non_exhaustive]
pub struct MemoryDeviceProperties {
    /// The maximum nr. of bytes that can be allocated in one go.
    pub max_page_size: u64,
    /// The required memory offset alignment in bytes.
    pub alignment: u64,
    /// Private because [`set_max_memory`](Self::set_max_memory) is its only
    /// writer, and that is where a zero becomes `None`.
    max_memory: Option<u64>,
}

impl MemoryDeviceProperties {
    /// Properties that state no capacity. A runtime that can read one adds it
    /// with [`with_max_memory`](Self::with_max_memory).
    pub const fn new(max_page_size: u64, alignment: u64) -> Self {
        Self {
            max_page_size,
            alignment,
            max_memory: None,
        }
    }

    /// How many bytes this memory may be asked to hold at once, or `None`,
    /// never `Some(0)`, where the runtime has no figure.
    ///
    /// This sizes a whole workload, while
    /// [`max_page_size`](Self::max_page_size) bounds a single allocation and
    /// is often derived from it. It is a budget, not a hardware census: each
    /// runtime reports the largest figure its own API stands behind, and
    /// staying under it is what keeps the device off its paging path.
    ///
    /// A runtime with no figure leaves it `None` rather than guess, because a
    /// guess reads as a measurement to every caller downstream.
    pub const fn max_memory(&self) -> Option<u64> {
        self.max_memory
    }

    /// States the capacity, as [`set_max_memory`](Self::set_max_memory) does.
    pub const fn with_max_memory(mut self, max_memory: u64) -> Self {
        self.set_max_memory(max_memory);
        self
    }

    /// States the capacity, dropping a zero: an API with nothing to report
    /// reports `0`, and [`max_memory`](Self::max_memory) says that as `None`.
    pub const fn set_max_memory(&mut self, max_memory: u64) {
        self.max_memory = match max_memory {
            0 => None,
            size => Some(size),
        };
    }
}

/// Who a device is, and what its compiled code is keyed to.
///
/// `name` and `fingerprint` answer different questions and must not be confused. `name`
/// is for people: it names the physical part, and two machines holding the same
/// part report the same name. `fingerprint` is for correctness: it is verbatim
/// the string this runtime passes to
/// [`compilation_store`](../../cubecl_runtime/compiler/fn.compilation_store.html),
/// which is what puts a compiled artifact out of reach of a machine that cannot
/// run it.
///
/// Reporting the fingerprint here rather than recomputing it is the whole
/// point: a backend derives it once and hands it to both consumers, so the
/// identity a bundle is stamped with and the namespace its kernels live under
/// cannot drift apart.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Default)]
pub struct DeviceIdentity {
    /// The device as it names itself — `AMD Radeon 8060S Graphics`,
    /// `NVIDIA H100 PCIe`. Distinct parts may share a name, so no capability
    /// and no compiled artifact may be keyed to it.
    pub name: String,
    /// What this runtime compiles *for* — `hip-kernel_gfx1151`, `ptx_sm90`.
    /// Verbatim the `compilation_store` fingerprint, so a namespace read back
    /// out of a bundle compares against it directly.
    pub fingerprint: String,
    /// The card behind the device, `None` for a device that is no card: a CPU, or a software
    /// rasterizer.
    pub physical: Option<PhysicalDevice>,
}

/// The card a device runs on. Two runtimes report different fields for one card, so compare with
/// [`is_same_card`](Self::is_same_card), not `==`.
#[derive(Debug, Clone, PartialEq, Eq, Hash, Default)]
#[non_exhaustive]
pub struct PhysicalDevice {
    /// The key every runtime reports alike on Linux.
    pub pci_address: Option<PciAddress>,
    /// The key every runtime reports alike on Windows.
    pub luid: Option<AdapterLuid>,
    /// The key every runtime reports alike on macOS.
    pub registry_entry_id: Option<RegistryEntryId>,
    pub vendor: Option<PciVendor>,
}

impl PhysicalDevice {
    /// Whether `other` is known to be this card through another runtime: by PCI address, otherwise
    /// by LUID, otherwise by registry entry id. A card with none of them matches nothing, itself
    /// included, since two such cards of one make compare equal.
    pub fn is_same_card(&self, other: &Self) -> bool {
        if let (Some(mine), Some(theirs)) = (self.pci_address, other.pci_address) {
            return mine == theirs;
        }
        if let (Some(mine), Some(theirs)) = (self.luid, other.luid) {
            return mine == theirs;
        }
        matches!(
            (self.registry_entry_id, other.registry_entry_id),
            (Some(mine), Some(theirs)) if mine == theirs
        )
    }

    /// Takes what this runtime left out from `other`, the same card seen through another runtime.
    pub fn fill_from(&mut self, other: &Self) {
        let Self {
            pci_address,
            luid,
            registry_entry_id,
            vendor,
        } = *other;
        self.pci_address = self.pci_address.or(pci_address);
        self.luid = self.luid.or(luid);
        self.registry_entry_id = self.registry_entry_id.or(registry_entry_id);
        self.vendor = self.vendor.or(vendor);
    }
}

/// The id of the GPU's entry in the macOS I/O Registry, which Metal reports as `registryID`. Like
/// [`AdapterLuid`] it is not promised to survive a restart, so it has no serialization or text
/// form.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct RegistryEntryId(u64);

impl RegistryEntryId {
    pub fn new(id: u64) -> Self {
        Self(id)
    }

    pub fn get(self) -> u64 {
        self.0
    }
}

/// The id Windows gives a graphics adapter. It changes on restart, so it has no serialization or
/// text form: a stored key wants [`PhysicalDevice::pci_address`].
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub struct AdapterLuid([u8; 8]);

impl AdapterLuid {
    /// From the eight bytes of a Windows `LUID`, low part first.
    pub fn new(bytes: [u8; 8]) -> Self {
        Self(bytes)
    }

    pub fn from_parts(low_part: u32, high_part: i32) -> Self {
        let mut bytes = [0; 8];
        bytes[..4].copy_from_slice(&low_part.to_le_bytes());
        bytes[4..].copy_from_slice(&high_part.to_le_bytes());
        Self(bytes)
    }

    pub fn bytes(self) -> [u8; 8] {
        self.0
    }
}

/// The maker of a card, by PCI vendor id.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum PciVendor {
    Nvidia,
    Amd,
    Intel,
    Apple,
    /// Mali GPUs.
    Arm,
    /// Adreno GPUs.
    Qualcomm,
    Other(u32),
}

impl PciVendor {
    pub fn id(self) -> u32 {
        match self {
            Self::Nvidia => 0x10de,
            Self::Amd => 0x1002,
            Self::Intel => 0x8086,
            Self::Apple => 0x106b,
            Self::Arm => 0x13b5,
            Self::Qualcomm => 0x5143,
            Self::Other(id) => id,
        }
    }
}

impl From<u32> for PciVendor {
    fn from(id: u32) -> Self {
        match id {
            0x10de => Self::Nvidia,
            0x1002 => Self::Amd,
            0x8086 => Self::Intel,
            0x106b => Self::Apple,
            0x13b5 => Self::Arm,
            0x5143 => Self::Qualcomm,
            other => Self::Other(other),
        }
    }
}

impl fmt::Display for PciVendor {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::Nvidia => f.write_str("NVIDIA"),
            Self::Amd => f.write_str("AMD"),
            Self::Intel => f.write_str("Intel"),
            Self::Apple => f.write_str("Apple"),
            Self::Arm => f.write_str("Arm"),
            Self::Qualcomm => f.write_str("Qualcomm"),
            Self::Other(id) => write!(f, "{id:#06x}"),
        }
    }
}

/// `domain:bus:device.function`, written `0000:07:00.0`. CUDA and NVML call it the bus id.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
pub struct PciAddress {
    pub domain: u32,
    pub bus: u8,
    pub device: u8,
    pub function: u8,
}

impl fmt::Display for PciAddress {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(
            f,
            "{:04x}:{:02x}:{:02x}.{:x}",
            self.domain, self.bus, self.device, self.function
        )
    }
}

#[derive(Debug, Clone, PartialEq, Eq)]
pub struct PciAddressError(pub String);

impl fmt::Display for PciAddressError {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        write!(f, "not a PCI address: {}", self.0)
    }
}

impl core::error::Error for PciAddressError {}

impl FromStr for PciAddress {
    type Err = PciAddressError;

    /// Also accepts the domainless `07:00.0` CUDA emits.
    fn from_str(text: &str) -> Result<Self, Self::Err> {
        let err = || PciAddressError(String::from(text));
        let (rest, function) = text.rsplit_once('.').ok_or_else(err)?;
        let mut parts = rest.rsplitn(3, ':');
        let device = parts.next().ok_or_else(err)?;
        let bus = parts.next().ok_or_else(err)?;
        let domain = parts.next().unwrap_or("0");
        Ok(Self {
            domain: u32::from_str_radix(domain, 16).map_err(|_| err())?,
            bus: u8::from_str_radix(bus, 16).map_err(|_| err())?,
            device: u8::from_str_radix(device, 16).map_err(|_| err())?,
            function: u8::from_str_radix(function, 16).map_err(|_| err())?,
        })
    }
}

/// Properties of what the device can do, like what `Feature` are
/// supported by it and what its memory properties are.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct DeviceProperties {
    /// The features supported by the runtime.
    pub features: Features,
    /// The memory properties of this client.
    pub memory: MemoryDeviceProperties,
    /// The topology properties of this client.
    pub hardware: HardwareProperties,
    /// The method used for profiling on the device.
    pub timing_method: TimingMethod,
    /// Who the device is, and what its kernels are keyed to.
    pub identity: DeviceIdentity,
    /// Bits IO is sized to in place of the load width, where a backend measured wider as faster.
    pub io_width_override: Option<u32>,
}

impl TypeHash for DeviceProperties {
    fn write_hash(_hasher: &mut impl core::hash::Hasher) {
        // ignored.
    }
}

impl DeviceProperties {
    /// Create a new feature set with the given features and memory properties.
    ///
    /// `identity` is a required argument rather than something a backend may
    /// fill in afterwards, so a runtime cannot ship reporting an anonymous
    /// device — the failure mode that leaves a bundle unable to say what it was
    /// built for.
    pub fn new(
        features: Features,
        memory_props: MemoryDeviceProperties,
        hardware: HardwareProperties,
        timing_method: TimingMethod,
        identity: DeviceIdentity,
    ) -> Self {
        DeviceProperties {
            features,
            memory: memory_props,
            hardware,
            timing_method,
            identity,
            io_width_override: None,
        }
    }

    /// The widest vector, in bits, that reads and writes are sized to.
    pub fn io_width(&self) -> u32 {
        self.io_width_override.unwrap_or(self.hardware.load_width)
    }

    /// Vector sizes, widest first, that reads and writes of `elem_size`-byte elements are sized to.
    pub fn io_optimized_vector_sizes(
        &self,
        elem_size: usize,
    ) -> impl Iterator<Item = VectorSize> + Clone {
        vector_sizes_down_from(self.widest_io_vector_size(elem_size))
    }

    /// Vector sizes, widest first, that keep `live` vectors in registers, else the IO sizes.
    pub fn vector_sizes_in_registers(
        &self,
        elem_size: usize,
        live: usize,
    ) -> impl Iterator<Item = VectorSize> + Clone {
        let widest = match VectorRegisters::new(&self.hardware, elem_size) {
            Some(registers) => registers.widest_vector_size(live),
            None => self.widest_io_vector_size(elem_size),
        };

        vector_sizes_down_from(widest)
    }

    /// Get the usages for a type
    pub fn type_usage(&self, ty: ElemType) -> EnumSet<TypeUsage> {
        self.features.type_usage(ty)
    }

    /// Get the complex capability families for a type.
    pub fn complex_usage(&self, ty: ElemType) -> EnumSet<ComplexUsage> {
        self.features.complex_usage(ty)
    }

    /// Whether a complex type supports the requested capability family.
    pub fn supports_complex_usage(&self, ty: ElemType, usage: ComplexUsage) -> bool {
        self.features.supports_complex_usage(ty, usage)
    }

    /// Get the usages for an atomic type on global memory
    pub fn atomic_type_usage(&self, ty: Type) -> EnumSet<AtomicUsage> {
        self.features.atomic_type_usage(ty)
    }

    /// Get the usages for an atomic type on shared memory
    pub fn shared_atomic_type_usage(&self, ty: Type) -> EnumSet<AtomicUsage> {
        self.features.shared_atomic_type_usage(ty)
    }

    /// Whether the type is supported in any way
    pub fn supports_type(&self, ty: impl Into<Type>) -> bool {
        self.features.supports_type(ty)
    }

    /// Whether the address type is supported in any way
    pub fn supports_address(&self, ty: impl Into<AddressType>) -> bool {
        self.features.supports_address(ty)
    }

    /// Register an address type to the features
    pub fn register_address_type(&mut self, ty: impl Into<AddressType>) {
        self.features.types.address.insert(ty.into());
    }

    /// Register the usages of an atomic type on global memory
    pub fn register_atomic_type_usage(&mut self, ty: Type, uses: impl Into<EnumSet<AtomicUsage>>) {
        *self.features.types.atomic.entry(ty).or_default() |= uses.into();
    }

    /// Register the usages of an atomic type on shared memory
    pub fn register_shared_atomic_type_usage(
        &mut self,
        ty: Type,
        uses: impl Into<EnumSet<AtomicUsage>>,
    ) {
        *self.features.types.shared_atomic.entry(ty).or_default() |= uses.into();
    }

    /// Register a storage type to the features
    pub fn register_type_usage(
        &mut self,
        ty: impl Into<ElemType>,
        uses: impl Into<EnumSet<TypeUsage>>,
    ) {
        *self.features.types.elem.entry(ty.into()).or_default() |= uses.into();
    }

    /// Register `ty`'s conversion as emulated in software ([`Types::emulated_conversion`]).
    ///
    /// [`Types::emulated_conversion`]: crate::Types::emulated_conversion
    pub fn register_emulated_conversion(&mut self, ty: impl Into<ElemType>) {
        self.features.types.emulated_conversion.insert(ty.into());
    }

    /// Register `ty`'s arithmetic as emulated through a wider type
    /// ([`Types::emulated_arithmetic`]).
    ///
    /// [`Types::emulated_arithmetic`]: crate::Types::emulated_arithmetic
    pub fn register_emulated_arithmetic(&mut self, ty: impl Into<ElemType>) {
        self.features.types.emulated_arithmetic.insert(ty.into());
    }

    /// Register complex capability families for an element type.
    pub fn register_complex_usage(
        &mut self,
        ty: impl Into<ElemType>,
        uses: impl Into<EnumSet<ComplexUsage>>,
    ) {
        *self.features.types.complex.entry(ty.into()).or_default() |= uses.into();
    }

    /// Register a semantic type to the features
    pub fn register_semantic_type(&mut self, ty: SemanticType) {
        self.features.types.semantic.insert(ty);
    }

    /// Register an opaque type to the features
    pub fn register_opaque_type(&mut self, ty: OpaqueType) {
        self.features.types.opaque.insert(ty);
    }

    /// Create a stable hash of all device properties relevant to kernel compilation. Can be used
    /// as a stable checksum for a compilation cache.
    pub fn checksum(&self) -> u64 {
        let state = foldhash::fast::FixedState::default();
        let mut hasher = state.build_hasher();
        self.features.hash(&mut hasher);
        self.hardware.hash(&mut hasher);
        hasher.finish()
    }

    /// A power of two, and at least one.
    fn widest_io_vector_size(&self, elem_size: usize) -> VectorSize {
        self.hardware.vector_size_in(self.io_width(), elem_size)
    }
}

/// Unchecked optimizations for float operations. May cause precision differences, or undefined
/// behaviour if the relevant conditions are not followed.
#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, TypeHash, EnumSetType)]
pub enum FastMath {
    /// Assume values are never `NaN`. If they are, the result is considered undefined behaviour.
    NotNaN,
    /// Assume values are never `Inf`/`-Inf`. If they are, the result is considered undefined
    /// behaviour.
    NotInf,
    /// Ignore sign on zero values.
    UnsignedZero,
    /// Allow swapping float division with a reciprocal, even if that swap would change precision.
    AllowReciprocal,
    /// Allow contracting float operations into fewer operations, even if the precision could
    /// change.
    AllowContraction,
    /// Allow reassociation for float operations, even if the precision could change.
    AllowReassociation,
    /// Allow all mathematical transformations for float operations, including contraction and
    /// reassociation, even if the precision could change.
    AllowTransform,
    /// Allow using lower precision intrinsics
    ReducedPrecision,
}

impl FastMath {
    pub const fn all() -> EnumSet<FastMath> {
        EnumSet::all()
    }
}

impl HardwareProperties {
    /// `width` bits of `elem_size`-byte elements, as a power of two the device accepts.
    fn vector_size_in(&self, width: u32, elem_size: usize) -> VectorSize {
        let elems = width as usize / (elem_size * 8);
        prev_power_of_two(elems.min(self.max_vector_size).max(1))
    }
}

fn prev_power_of_two(value: usize) -> usize {
    1 << value.ilog2()
}

fn vector_sizes_down_from(widest: usize) -> impl Iterator<Item = VectorSize> + Clone {
    (0..=widest.ilog2()).rev().map(|exponent| 1 << exponent)
}

#[cfg(test)]
mod tests {
    use super::*;
    use alloc::string::ToString;

    #[test]
    fn a_device_without_a_fixed_register_set_has_no_budget() {
        assert_eq!(VectorRegisters::new(&hardware(128, None), 4), None);
    }

    #[test]
    fn the_widest_vector_size_keeps_every_live_vector_in_registers() {
        let f32 = registers(AVX2, 4);
        assert_eq!(f32.widest_vector_size(6), 16);
        assert_eq!(f32.widest_vector_size(8), 16);
        assert_eq!(f32.widest_vector_size(9), 8);
        assert_eq!(registers(AVX2, 8).widest_vector_size(6), 8);
        assert_eq!(registers(AVX512, 4).widest_vector_size(6), 64);
    }

    #[test]
    fn neon_and_avx2_budget_the_same_vector_sizes_from_equal_register_files() {
        let (neon, avx2) = (registers(NEON, 4), registers(AVX2, 4));
        assert_eq!(neon.widest_vector_size(3), avx2.widest_vector_size(3));
        assert_eq!(neon.widest_vector_size(6), avx2.widest_vector_size(6));
    }

    #[test]
    fn more_live_vectors_than_registers_still_get_one_register_each() {
        assert_eq!(registers(AVX2, 4).widest_vector_size(40), 8);
        assert_eq!(registers(NEON, 4).widest_vector_size(40), 4);
    }

    #[test]
    fn vectors_fit_in_the_registers_left_unreserved() {
        let f32 = registers(AVX2, 4);
        assert_eq!(f32.registers_for(16), 2);
        assert_eq!(f32.vectors_fitting(16, 0), 8);
        assert_eq!(f32.vectors_fitting(8, 4), 12);
        assert_eq!(f32.vectors_fitting(8, 20), 0);
    }

    #[test]
    fn a_capped_vector_size_caps_the_register_vector_size() {
        let mut capped = hardware(256, Some(16));
        capped.max_vector_size = 4;
        let f32 = VectorRegisters::new(&capped, 4).unwrap();
        assert_eq!(f32.registers_for(8), 2);
        assert_eq!(f32.widest_vector_size(1), 4);
    }

    #[test]
    fn io_follows_a_load_width_raised_after_construction() {
        let mut device = device(128);
        device.hardware.load_width = 256;
        assert_eq!(device.io_width(), 256);
    }

    #[test]
    fn a_stated_io_width_outlasts_a_load_width_change() {
        let mut device = device(256);
        device.io_width_override = Some(512);
        device.hardware.load_width = 128;
        assert_eq!(device.io_width(), 512);
    }

    #[test]
    fn a_cap_between_powers_of_two_rounds_down_alike() {
        let mut device = device(512);
        device.hardware.max_vector_size = 12;
        device.hardware.vector_register_count = Some(32);

        assert_eq!(
            device
                .io_optimized_vector_sizes(4)
                .collect::<alloc::vec::Vec<_>>(),
            alloc::vec![8, 4, 2, 1]
        );
        assert_eq!(device.vector_sizes_in_registers(4, 1).next(), Some(8));
    }

    #[test]
    fn a_device_without_registers_sizes_by_its_io() {
        let device = device(128);
        assert!(
            device
                .vector_sizes_in_registers(4, 6)
                .eq(device.io_optimized_vector_sizes(4))
        );
    }

    /// A capacity is stated only by the runtime that read one.
    ///
    /// Properties built without one must answer `None`, or a caller would
    /// size a workload against a figure nobody measured.
    #[test]
    fn a_memory_states_no_capacity_until_one_is_read() {
        let props = MemoryDeviceProperties::new(1024, 32);
        assert_eq!(props.max_memory(), None);
        assert_eq!(props.clone().with_max_memory(4096).max_memory(), Some(4096));
    }

    /// A zero from the runtime's API reads as no capacity.
    ///
    /// An API with nothing to report reports `0`, which must not reach a
    /// caller as a device that holds nothing.
    #[test]
    fn a_capacity_of_zero_is_no_capacity() {
        let mut props = MemoryDeviceProperties::new(1024, 32).with_max_memory(4096);
        props.set_max_memory(0);
        assert_eq!(props.max_memory(), None);
    }

    #[test]
    fn a_vendor_keeps_its_id_whether_named_or_not() {
        for id in [0x10de, 0x1002, 0x8086, 0x106b, 0x13b5, 0x5143, 0x1af4] {
            assert_eq!(PciVendor::from(id).id(), id);
        }
        assert_eq!(PciVendor::from(0x10de), PciVendor::Nvidia);
        assert_eq!(PciVendor::from(0x1af4), PciVendor::Other(0x1af4));
        assert_eq!(PciVendor::Other(0x1af4).to_string(), "0x1af4");
    }

    #[test]
    fn a_luid_from_parts_is_the_bytes_vulkan_reports() {
        let bytes = [0x8a, 0x1d, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00];
        assert_eq!(
            AdapterLuid::from_parts(0x0001_1d8a, 0),
            AdapterLuid::new(bytes)
        );
        assert_eq!(
            AdapterLuid::from_parts(1, -1).bytes(),
            [1, 0, 0, 0, 0xff, 0xff, 0xff, 0xff]
        );
    }

    #[test]
    fn a_card_is_matched_by_address_then_by_luid() {
        let address = |bus| {
            Some(PciAddress {
                domain: 0,
                bus,
                device: 0,
                function: 0,
            })
        };
        let luid = |low| Some(AdapterLuid::from_parts(low, 0));
        let card = |pci_address, luid| PhysicalDevice {
            pci_address,
            luid,
            registry_entry_id: None,
            vendor: None,
        };

        assert!(card(address(7), None).is_same_card(&card(address(7), luid(1))));
        assert!(!card(address(7), luid(1)).is_same_card(&card(address(8), luid(1))));
        assert!(card(None, luid(1)).is_same_card(&card(address(7), luid(1))));
        assert!(!card(None, luid(1)).is_same_card(&card(None, luid(2))));
        assert!(!card(None, None).is_same_card(&card(None, None)));
    }

    #[test]
    fn a_card_filled_from_another_runtime_keeps_what_it_reported() {
        let address = |bus| {
            Some(PciAddress {
                domain: 0,
                bus,
                device: 0,
                function: 0,
            })
        };
        let luid = Some(AdapterLuid::from_parts(1, 0));
        let mut card = PhysicalDevice {
            pci_address: address(7),
            luid: None,
            registry_entry_id: None,
            vendor: None,
        };

        card.fill_from(&PhysicalDevice {
            pci_address: address(8),
            luid,
            registry_entry_id: None,
            vendor: Some(PciVendor::Nvidia),
        });

        assert_eq!(
            card,
            PhysicalDevice {
                pci_address: address(7),
                luid,
                registry_entry_id: None,
                vendor: Some(PciVendor::Nvidia),
            }
        );
    }

    #[test]
    fn a_card_with_neither_address_nor_luid_is_matched_by_registry_entry_id() {
        let card = |registry_entry_id| PhysicalDevice {
            pci_address: None,
            luid: None,
            registry_entry_id: Some(RegistryEntryId::new(registry_entry_id)),
            vendor: None,
        };

        assert!(card(0x1_0000_04c8).is_same_card(&card(0x1_0000_04c8)));
        assert!(!card(0x1_0000_04c8).is_same_card(&card(0x1_0000_04c9)));
    }

    #[test]
    fn the_address_and_the_luid_outrank_the_registry_entry_id() {
        let address = |bus| {
            Some(PciAddress {
                domain: 0,
                bus,
                device: 0,
                function: 0,
            })
        };
        let luid = |low| Some(AdapterLuid::from_parts(low, 0));
        let card = |pci_address, luid| PhysicalDevice {
            pci_address,
            luid,
            registry_entry_id: Some(RegistryEntryId::new(0x1_0000_04c8)),
            vendor: None,
        };

        assert!(!card(address(7), None).is_same_card(&card(address(8), None)));
        assert!(!card(None, luid(1)).is_same_card(&card(None, luid(2))));
    }

    #[test]
    fn a_card_filled_from_another_runtime_takes_the_registry_entry_id_it_left_out() {
        let registry_entry_id = |id| Some(RegistryEntryId::new(id));
        let card = |registry_entry_id| PhysicalDevice {
            pci_address: None,
            luid: None,
            registry_entry_id,
            vendor: None,
        };

        let mut without_one = card(None);
        without_one.fill_from(&card(registry_entry_id(0x1_0000_04c8)));
        let mut with_one = card(registry_entry_id(0x1_0000_04c8));
        with_one.fill_from(&card(registry_entry_id(0x1_0000_04c9)));

        assert_eq!(
            without_one.registry_entry_id,
            registry_entry_id(0x1_0000_04c8)
        );
        assert_eq!(with_one.registry_entry_id, registry_entry_id(0x1_0000_04c8));
    }

    #[test]
    fn a_pci_address_round_trips_and_defaults_its_domain() {
        let id = PciAddress {
            domain: 0,
            bus: 7,
            device: 0,
            function: 0,
        };
        assert_eq!(id.to_string(), "0000:07:00.0");
        assert_eq!("0000:07:00.0".parse::<PciAddress>(), Ok(id));
        assert_eq!("07:00.0".parse::<PciAddress>(), Ok(id));
        assert_eq!(
            "0001:a3:1f.7".parse::<PciAddress>(),
            Ok(PciAddress {
                domain: 1,
                bus: 0xa3,
                device: 0x1f,
                function: 7
            })
        );
        assert!("07:00".parse::<PciAddress>().is_err());
        assert!("gpu".parse::<PciAddress>().is_err());
    }

    const AVX2: (u32, u32) = (256, 16);
    const AVX512: (u32, u32) = (512, 32);
    const NEON: (u32, u32) = (128, 32);

    fn hardware(load_width: u32, vector_register_count: Option<u32>) -> HardwareProperties {
        HardwareProperties {
            load_width,
            vector_register_count,
            plane_size_min: 1,
            plane_size_max: 1,
            max_bindings: u32::MAX,
            max_shared_memory_size: 32 * 1024,
            max_cube_count: (u32::MAX, u32::MAX, u32::MAX),
            max_units_per_cube: 16,
            max_cube_dim: (16, 16, 16),
            num_streaming_multiprocessors: None,
            num_cpu_cores: Some(16),
            last_level_cache_size: None,
            num_tensor_cores: None,
            min_tensor_cores_dim: None,
            max_vector_size: VectorSize::MAX,
            cube_mma_reserved_shared_memory: 0,
        }
    }

    fn registers((width, count): (u32, u32), elem_size: usize) -> VectorRegisters {
        VectorRegisters::new(&hardware(width, Some(count)), elem_size).unwrap()
    }

    fn device(load_width: u32) -> DeviceProperties {
        DeviceProperties::new(
            Features::default(),
            MemoryDeviceProperties::new(1024, 32),
            hardware(load_width, None),
            TimingMethod::System,
            DeviceIdentity::default(),
        )
    }
}