Files
uu-remote-linux/benches/nvenc_pipeline_bench.rs
059a3b1bb8 inproc-av: 4:4:4 on the native NVENC engine (#76, #77)
H.264 High 4:4:4 / HEVC FRExt with chromaFormatIDC 3 and AQ off when 真彩 is requested; GPUs without YUV444 encode fail the capability check and fall back. full_chroma_available() no longer excludes the in-process path. UU_PROBE_CHROMA=444 for the probe.

Co-authored-by: ferstar <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-21 08:24:32 -07:00

134 lines
5.7 KiB
Rust

//! Standalone benchmark tool measuring:
//! 1. Host Memory Copy Bandwidth & CPU Cost (What FFmpeg/Sway CPU swizzling does)
//! 2. NVENC H.265 Hardware Throughput & Latency (1920x1200 @ ultra-low-latency)
//! 3. NVENC reading a registered (pinned) host buffer
use std::time::Instant;
fn main() {
println!("=========================================================");
println!(" UU Remote Linux: Encoder Pipeline Benchmark Suite ");
println!("=========================================================");
println!();
let width: u32 = 1920;
let height: u32 = 1200;
let frames: usize = 300; // 5 seconds of 60fps video
let frame_bytes: usize = (width * height * 4) as usize;
println!("Test Configuration:");
println!(" Resolution : {}x{}", width, height);
println!(" Frame Size : {:.2} MB (ARGB32)", frame_bytes as f64 / 1024.0 / 1024.0);
println!(" Iterations : {} frames (Equivalent to 5.0s @ 60 FPS)", frames);
println!();
// Simulating XShm framebuffer
let mut host_pixels = vec![0x55u8; frame_bytes];
// Benchmark 1: CPU Memcpy (External FFmpeg stdin / Sway wlr-screencopy CPU copy baseline)
println!("[1] Baseline: CPU Frame Memcpy Overhead (IPC / Process Pipe Cost)");
let mut sink = vec![0u8; frame_bytes];
let start_memcpy = Instant::now();
for _ in 0..frames {
sink.copy_from_slice(&host_pixels);
std::hint::black_box(&sink);
}
let duration_memcpy = start_memcpy.elapsed();
let memcpy_per_frame = duration_memcpy.as_secs_f64() * 1000.0 / frames as f64;
let memcpy_gbps = (frame_bytes * frames) as f64 / duration_memcpy.as_secs_f64() / 1e9;
let cpu_load_60fps = (memcpy_per_frame / (1000.0 / 60.0)) * 100.0;
println!(" -> Total CPU copy time : {:.2?}", duration_memcpy);
println!(" -> Per-frame copy time : {:.3} ms", memcpy_per_frame);
println!(" -> Memory Bus Bandwidth : {:.2} GB/s", memcpy_gbps);
println!(" -> CPU Core Saturation : {:.1}% of 1 core consumed solely by memcpy at 60 FPS", cpu_load_60fps);
println!();
// Benchmark 2: in-process NVENC reading a registered host buffer (device copy when registration fails)
#[cfg(feature = "inproc-av")]
{
extern "C" {
fn nvenc_create(config: *const NvencConfigFFI) -> *mut std::ffi::c_void;
fn nvenc_register_host_buffer(
engine: *mut std::ffi::c_void,
host_ptr: *mut std::ffi::c_void,
size: usize,
) -> i32;
fn nvenc_encode_frame(
engine: *mut std::ffi::c_void,
input_argb: *const u8,
pitch: u32,
force_idr: i32,
out_data: *mut *const u8,
out_len: *mut u32,
) -> i32;
fn nvenc_destroy(engine: *mut std::ffi::c_void);
}
// Must match NvencConfig in c_engine/nvenc_engine.h.
#[repr(C)]
struct NvencConfigFFI {
width: u32,
height: u32,
fps: u32,
gop: u32,
cq: u32,
target_bps: i64,
max_bps: i64,
buffer_bits: i64,
is_hevc: i32,
chroma_444: i32,
}
let cfg = NvencConfigFFI {
width,
height,
fps: 60,
gop: 120,
cq: 18,
target_bps: 16_000_000,
max_bps: 30_000_000,
buffer_bits: 500_000,
is_hevc: 1, // H.265 / HEVC
chroma_444: i32::from(std::env::var("UU_BENCH_CHROMA").as_deref() == Ok("444")),
};
println!("[2] In-process NVENC H.265 encoding");
unsafe {
let engine = nvenc_create(&cfg);
if engine.is_null() {
println!(" [SKIP] NVENC driver not available on this system.");
} else {
let registered = nvenc_register_host_buffer(engine, host_pixels.as_mut_ptr() as *mut _, frame_bytes) == 0;
println!(" -> Input path : {}", if registered { "registered host buffer (DMA)" } else { "cuMemcpyHtoD copy" });
let input = if registered { std::ptr::null() } else { host_pixels.as_ptr() };
let mut out_data: *const u8 = std::ptr::null();
let mut out_len: u32 = 0;
// Warmup
let _ = nvenc_encode_frame(engine, input, width * 4, 1, &mut out_data, &mut out_len);
let start = Instant::now();
let (mut encoded_bytes, mut encoded_frames) = (0u64, 0usize);
for i in 0..frames {
if nvenc_encode_frame(engine, input, width * 4, i32::from(i == 0), &mut out_data, &mut out_len) == 0 {
encoded_bytes += u64::from(out_len);
encoded_frames += 1;
}
}
let elapsed = start.elapsed();
if encoded_frames == 0 {
println!(" [FAIL] no frame was encoded");
} else {
let per_frame = elapsed.as_secs_f64() * 1000.0 / encoded_frames as f64;
println!(" -> Encoded frames : {encoded_frames}/{frames} in {elapsed:.2?}");
println!(" -> Per-frame latency : {per_frame:.3} ms ({:.1} FPS max)", 1000.0 / per_frame);
println!(" -> Output bitrate : {:.2} Mbps average", encoded_bytes as f64 * 8.0 / elapsed.as_secs_f64() / 1e6);
if registered {
println!(" -> Host-to-device copy : none ({:.0} MB/s at 60 fps avoided)", frame_bytes as f64 * 60.0 / 1024.0 / 1024.0);
}
}
nvenc_destroy(engine);
}
}
}
}