rlx_runtime/

cost.rs

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.

//! Cross-backend cost interface.
//!
//! Each backend implements `BackendCostModel` to expose its execution
//! characteristics (kernel throughput, dispatch overhead, memory bw).
//! The runtime can then estimate the cost of running a graph on each
//! available backend and pick the fastest.
//!
//! This is what enables "auto device" — given a graph, pick CPU or
//! Metal automatically based on which is faster for THIS workload on
//! THIS hardware.

use crate::Device;
use rlx_ir::{Graph, Node, Op};

/// Hardware-aware cost characteristics for a backend on the current machine.
pub trait BackendCostModel: Send + Sync {
    /// Identify which device this model is for.
    fn device(&self) -> Device;

    /// Effective f32 sgemm throughput in GFLOP/s for the most-used kernel
    /// path at the given dimensions. Backends should return their best
    /// sustained rate (not peak).
    fn sgemm_gflops(&self, m: usize, k: usize, n: usize) -> f64;

    /// Cost to dispatch one kernel (function call, BLAS setup, etc.) in ns.
    fn dispatch_overhead_ns(&self) -> f64;

    /// Cost to commit + wait for a command buffer / forward pass in ns.
    /// Roughly amortized per-forward overhead independent of kernel count.
    fn roundtrip_overhead_ns(&self) -> f64;

    /// Memory bandwidth in bytes/ns (== GB/s).
    fn memory_bw(&self) -> f64;

    /// Number of compute threads available.
    fn num_threads(&self) -> usize;
}

/// Estimate forward-pass time (ns) for a graph on the given backend.
/// Uses node-level cost contributions; conservative — actual time may
/// be lower due to hardware parallelism we don't model.
pub fn estimate_graph_cost(graph: &Graph, model: &dyn BackendCostModel) -> f64 {
    let mut total = model.roundtrip_overhead_ns();
    for node in graph.nodes() {
        total += node_cost(node, graph, model);
    }
    total
}

fn node_cost(node: &Node, graph: &Graph, model: &dyn BackendCostModel) -> f64 {
    let dispatch = model.dispatch_overhead_ns();
    match &node.op {
        Op::Input { .. } | Op::Param { .. } | Op::Constant { .. } => 0.0,
        Op::MatMul | Op::FusedMatMulBiasAct { .. } => {
            let n = node.shape.dim(node.shape.rank() - 1).unwrap_static();
            let total = node.shape.num_elements().unwrap_or(0);
            let m = total / n.max(1);
            let a_total = graph.node(node.inputs[0]).shape.num_elements().unwrap_or(0);
            let k = a_total / m.max(1);
            let flops = 2.0 * m as f64 * k as f64 * n as f64;
            flops / (model.sgemm_gflops(m, k, n) + 1.0) + dispatch
        }
        Op::Attention {
            num_heads,
            head_dim,
            ..
        } => {
            let q_shape = &graph.node(node.inputs[0]).shape;
            let seq = q_shape.dim(q_shape.rank() - 2).unwrap_static();
            let batch = q_shape.num_elements().unwrap_or(0) / (seq * num_heads * head_dim).max(1);
            let flops = (batch * num_heads * seq * seq * head_dim * 2) as f64;
            flops / (model.sgemm_gflops(seq, *head_dim, seq) + 1.0) + dispatch
        }
        // Element-wise + small ops: bounded by memory bandwidth.
        _ => {
            let bytes = node.shape.num_elements().unwrap_or(0) * 4;
            (bytes as f64) / model.memory_bw().max(1.0) + dispatch
        }
    }
}

/// Pick the device with the lowest predicted cost for this graph.
pub fn pick_best_device(graph: &Graph, models: &[&dyn BackendCostModel]) -> Device {
    let mut best = (Device::Cpu, f64::INFINITY);
    for &m in models {
        let cost = estimate_graph_cost(graph, m);
        if cost < best.1 {
            best = (m.device(), cost);
        }
    }
    best.0
}

// ── Backend adapters (plan #29) ─────────────────────────────────
//
// The CPU and Metal crates own their own internal cost models for
// kernel-selection decisions. These thin adapters wrap them in
// `BackendCostModel` so `pick_best_device` can compare both with a
// single uniform interface.

/// `BackendCostModel` impl backed by `rlx_cpu::cost::HwModel`.
#[cfg(feature = "cpu")]
pub struct CpuCostModel(rlx_cpu::cost::HwModel);

#[cfg(feature = "cpu")]
impl CpuCostModel {
    pub fn new() -> Self {
        let cfg = rlx_cpu::config::RuntimeConfig::global();
        Self(rlx_cpu::cost::HwModel::from_config(cfg))
    }
}

#[cfg(feature = "cpu")]
impl Default for CpuCostModel {
    fn default() -> Self {
        Self::new()
    }
}

#[cfg(feature = "cpu")]
impl BackendCostModel for CpuCostModel {
    fn device(&self) -> Device {
        Device::Cpu
    }
    fn sgemm_gflops(&self, m: usize, k: usize, n: usize) -> f64 {
        // Take the better of NEON / BLAS at this shape.
        let flops = 2.0 * m as f64 * k as f64 * n as f64;
        let neon_time = flops / self.0.neon_flops.max(1.0);
        let blas_time = flops / self.0.blas_flops.max(1.0);
        let pick = neon_time.min(blas_time);
        if pick > 0.0 {
            flops / (pick * 1e9)
        } else {
            0.0
        }
    }
    fn dispatch_overhead_ns(&self) -> f64 {
        self.0.blas_overhead_ns
    }
    fn roundtrip_overhead_ns(&self) -> f64 {
        self.0.par_for_overhead_ns
    }
    fn memory_bw(&self) -> f64 {
        self.0.mem_bw
    }
    fn num_threads(&self) -> usize {
        self.0.num_threads
    }
}

/// `BackendCostModel` impl backed by `rlx_metal::cost`. Reads from
/// the on-disk calibration cache so the numbers reflect what this
/// machine actually measured.
#[cfg(feature = "metal")]
pub struct MetalCostModel {
    sgemm_gflops_avg: f64,
    roundtrip_ns: f64,
    memory_bw: f64,
}

#[cfg(feature = "metal")]
impl MetalCostModel {
    pub fn new() -> Self {
        let cal = rlx_metal::calibrate::Calibration::load_or_measure();
        // Effective single-shape sgemm: best of the calibrated paths.
        let best = cal
            .sgemm_simd_4x4_flops
            .max(cal.sgemm_simd_flops)
            .max(cal.sgemm_padded_flops);
        Self {
            sgemm_gflops_avg: best,
            roundtrip_ns: cal.roundtrip_overhead_ns,
            // Apple Silicon unified memory bandwidth (rough): ~200 GB/s
            // on M-series base, much higher on Pro/Max. The calibrator
            // doesn't measure pure mem-bw yet, so we hard-code a
            // floor that makes mem-bound ops not look free.
            memory_bw: 200.0,
        }
    }
}

#[cfg(feature = "metal")]
impl Default for MetalCostModel {
    fn default() -> Self {
        Self::new()
    }
}

#[cfg(feature = "metal")]
impl BackendCostModel for MetalCostModel {
    fn device(&self) -> Device {
        Device::Metal
    }
    fn sgemm_gflops(&self, _m: usize, _k: usize, _n: usize) -> f64 {
        self.sgemm_gflops_avg
    }
    fn dispatch_overhead_ns(&self) -> f64 {
        // Per-kernel encode cost — small relative to the round-trip.
        2_000.0
    }
    fn roundtrip_overhead_ns(&self) -> f64 {
        self.roundtrip_ns
    }
    fn memory_bw(&self) -> f64 {
        self.memory_bw
    }
    fn num_threads(&self) -> usize {
        1
    } // single command queue
}

/// `BackendCostModel` impl backed by `rlx_mlx::calibrate`. Reads from
/// the on-disk MLX calibration cache. The first construction on a
/// fresh machine pays a one-time measurement cost (tens of ms);
/// subsequent constructions read the cache.
#[cfg(all(feature = "mlx", target_os = "macos"))]
pub struct MlxCostModel {
    sgemm_large_flops: f64,
    sgemm_small_flops: f64,
    roundtrip_ns: f64,
    memory_bw: f64,
}

#[cfg(all(feature = "mlx", target_os = "macos"))]
impl MlxCostModel {
    pub fn new() -> Self {
        let cal = rlx_mlx::calibrate::Calibration::load_or_measure();
        // Use measured memory bandwidth when available (post-PR16
        // calibrators record it); fall back to the Apple-Silicon
        // unified-memory floor otherwise so old caches still produce
        // sane numbers.
        let memory_bw = if cal.memory_bw_gbps > 0.0 {
            cal.memory_bw_gbps
        } else {
            200.0
        };
        Self {
            sgemm_large_flops: cal.sgemm_large_flops,
            sgemm_small_flops: cal.sgemm_small_flops,
            roundtrip_ns: cal.roundtrip_overhead_ns,
            memory_bw,
        }
    }
}

#[cfg(all(feature = "mlx", target_os = "macos"))]
impl Default for MlxCostModel {
    fn default() -> Self {
        Self::new()
    }
}

#[cfg(all(feature = "mlx", target_os = "macos"))]
impl BackendCostModel for MlxCostModel {
    fn device(&self) -> Device {
        Device::Mlx
    }
    fn sgemm_gflops(&self, m: usize, k: usize, n: usize) -> f64 {
        // Crossover heuristic: small shapes pay the per-op overhead;
        // large shapes hit the optimized path. The cutoff is rough —
        // matches the calibrator's "small" / "large" probe sizes.
        let total = m as f64 * k as f64 * n as f64;
        if total < 32_768.0 {
            self.sgemm_small_flops
        } else {
            self.sgemm_large_flops
        }
    }
    fn dispatch_overhead_ns(&self) -> f64 {
        // MLX's lazy-eval keeps per-op encode cost low; trace
        // construction in Rust is the dominant per-op cost.
        2_000.0
    }
    fn roundtrip_overhead_ns(&self) -> f64 {
        self.roundtrip_ns
    }
    fn memory_bw(&self) -> f64 {
        self.memory_bw
    }
    fn num_threads(&self) -> usize {
        1
    }
}