use super::inference::Inference;
use super::scheduler::LpuScheduler;
use crate::common::once::Once;
use core::sync::atomic::{AtomicUsize, Ordering};
static LPU_INFERENCE: Once<Inference> = Once::new();
static LPU_SCHEDULER: Once<LpuScheduler> = Once::new();
pub fn init() {
match crate::arch::detect_arch() {
crate::arch::Architecture::X86_64 => {
if let Some(ctx) = crate::arch::x86_64::lpu::init_lpu(0, 0, 0) {
static X86_LPU_SIG: AtomicUsize = AtomicUsize::new(0);
X86_LPU_SIG.store(
ctx.bus as usize
^ ctx.device as usize
^ ctx.function as usize
^ ctx.vendor_id as usize
^ ctx.device_id as usize
^ ctx.bar0_base
^ ctx.bar0_size
^ ctx.msi_vector as usize
^ ctx.coherent_dma_base,
Ordering::Release,
);
}
static X86_LPU_DIAG: AtomicUsize = AtomicUsize::new(0);
X86_LPU_DIAG.store(
crate::arch::x86_64::lpu::diagnostics(0, 0, 0),
Ordering::Release,
);
}
crate::arch::Architecture::AArch64 => {
let (dt_base, dt_size, dt_irq) =
crate::firmware::devicetree::find_device_by_compatible(b"arm,lpu")
.unwrap_or((0, 0, 0));
if dt_base == 0 {
} else if let Some(ctx) = crate::arch::aarch64::lpu::init_lpu(dt_base, dt_size, dt_irq)
{
static ARM_LPU_SIG: AtomicUsize = AtomicUsize::new(0);
ARM_LPU_SIG.store(
ctx.mmio_base
^ ctx.mmio_size
^ ctx.device_id as usize
^ ctx.spi_id as usize
^ ctx.smmu_stream_id as usize
^ ctx.dma_region,
Ordering::Release,
);
}
static ARM_LPU_DIAG: AtomicUsize = AtomicUsize::new(0);
ARM_LPU_DIAG.store(
crate::arch::aarch64::lpu::diagnostics(dt_base),
Ordering::Release,
);
}
_ => {}
}
crate::lpu::device::init();
if LPU_INFERENCE.get().is_none() {
let ok_infer = LPU_INFERENCE.set(Inference::new());
if !ok_infer {}
}
if LPU_SCHEDULER.get().is_none() {
let ok_sched = LPU_SCHEDULER.set(LpuScheduler::new());
if !ok_sched {}
}
let base = crate::firmware::devicetree::find_device_by_compatible(b"arm,lpu")
.map(|(b, _, _)| b)
.unwrap_or(0);
if let Some(accel) = super::drivers::generic::Accelerator::probe(base) {
accel.init();
accel.submit_workload(0x5000, 16);
let ai = accel.is_initialized();
let ver = accel.read_version();
static LPU_ACCEL_SIG: AtomicUsize = AtomicUsize::new(0);
LPU_ACCEL_SIG.store(
accel.workload_count() ^ (ai as usize) ^ (ver as usize),
Ordering::Release,
);
}
if let Some(apple) = super::drivers::apple::AppleLpu::probe(base) {
apple.init();
apple.submit_tokens(0x6000, 8);
static APPLE_SIG: AtomicUsize = AtomicUsize::new(0);
APPLE_SIG.store(
apple.inference_count() ^ (apple.is_initialized() as usize),
Ordering::Release,
);
}
if let Some(edge) = super::drivers::edge::EdgeLpu::probe(base) {
edge.init();
edge.run_inference(0x7000, 0x8000, 0x9000);
static EDGE_SIG: AtomicUsize = AtomicUsize::new(0);
EDGE_SIG.store(
edge.inference_count() ^ (edge.is_initialized() as usize),
Ordering::Release,
);
}
if let Some(qc) = super::drivers::qualcomm::Qualcomm::probe(base) {
qc.init();
qc.submit_workload(0xA000, 4);
let ver = qc.read_version();
static QC_SIG: AtomicUsize = AtomicUsize::new(0);
QC_SIG.store(
qc.workload_count() ^ (qc.is_initialized() as usize) ^ (ver as usize),
Ordering::Release,
);
}
}
pub fn inference() -> Option<&'static Inference> {
LPU_INFERENCE.get()
}
pub fn scheduler() -> Option<&'static LpuScheduler> {
LPU_SCHEDULER.get()
}