Struct AmdGpu 

pub struct AmdGpu {
    pub pci_slot: String,
    pub pci_device_id: String,
    pub marketing_name: Option<String>,
    pub gfx_arch: Option<String>,
    pub vram_bytes: Option<u64>,
    pub is_integrated: bool,
    pub rocm_index: Option<u32>,
}

AMD discrete GPU discovered via sysfs.

All fields optional except pci_slot and is_integrated. Detectors populate from /sys/class/drm/card*/device and PCI tables.

Fields

pci_slot: String

PCI slot in BB:DD.F form (canonicalized, no 0000: prefix).

pci_device_id: String

PCI device ID (hex string, e.g. "0x744c" for Navi 31).

marketing_name: Option<String>

Marketing name from lspci/rocminfo if available.

gfx_arch: Option<String>

gfx architecture (e.g. "gfx1100" for Navi 31).

vram_bytes: Option<u64>

Total VRAM in bytes from mem_info_vram_total.

is_integrated: bool

true if classified as integrated by device_is_integrated.

rocm_index: Option<u32>

Real ROCm (HSA) agent index from rocminfo, when available.

This is the authoritative device index for ROCR_VISIBLE_DEVICES / HIP_VISIBLE_DEVICES: those vars filter by the original enumeration order (CPU + iGPU + dGPU agents as rocminfo prints them), NOT a re-numbered post-filter sequence. On an APU+dGPU host where the iGPU is ROCr agent 2 and the dGPU is agent 3, the visibility mask must be 3, not the compacted 0 produced by enumerating the filtered dGPU list.

None when rocminfo is unavailable; callers fall back to compacted indices (the historical sysfs-only behavior) in that case.

Trait Implementations

impl Clone for AmdGpu

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for AmdGpu

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.