Add AMD OpenCL kernel, runtime-loaded CUDA, mixed backend, portability
AMD GPU backend: - Add the GCN-tuned equihash192_7.cl kernel (clearCounter/blake/round1..7/ combine pipeline) and its host driver src/gpu_amd.rs. GpuSolver now dispatches AMD-vendor OpenCL devices to it and other devices to the existing kernel (force with ZCL_OPENCL_KERNEL=amd|legacy). Validated on an RX 9060 XT: GPU solutions match the CPU reference 1/1. - Expose BatchHasher::midstate() for the kernel's ulong8 hashState arg. Runtime-loaded GPU drivers (minimum host deps): - dlopen libcuda / libnvidia-ml via libloading instead of linking them (src/dylib.rs macro; cuda.rs, nvml.rs, gpu_probe.rs). The binary now builds and starts on hosts without an NVIDIA driver and reports no CUDA devices gracefully; remove build.rs (its only job was linking those libs). - Add Dockerfile.portable + build-portable.sh: build against Debian bullseye's glibc 2.31 for a binary that runs on older distros and drives both AMD (OpenCL) and NVIDIA (CUDA) cards. Document the build matrix in the README. Mixed backend (default): - Add --backend mixed (now the default): each card on its native backend (NVIDIA->CUDA, AMD/Intel->OpenCL), deduped so no card is mined twice. --devices indexes the unified list shown by --list-devices. Misc: - Stale-work timeout (--job-timeout) default 300s -> 600s (10 minutes). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
jackpotincorporated
+125
-15
@@ -14,6 +14,14 @@ mod tui;
|
||||
#[cfg(feature = "gpu")]
|
||||
mod gpu;
|
||||
|
||||
// AMD-tuned OpenCL kernel driver (selected by GpuSolver for AMD-vendor devices).
|
||||
#[cfg(feature = "gpu")]
|
||||
mod gpu_amd;
|
||||
|
||||
// Runtime dynamic-library loader (dlopen) for the CUDA driver + NVML.
|
||||
#[cfg(feature = "cuda")]
|
||||
mod dylib;
|
||||
|
||||
#[cfg(feature = "cuda")]
|
||||
mod cuda;
|
||||
|
||||
@@ -79,8 +87,9 @@ struct Args {
|
||||
jackpot: Option<u32>,
|
||||
|
||||
/// Pause mining if no new job arrives within this many seconds (stale work
|
||||
/// guard); resumes automatically when fresh work arrives. 0 disables.
|
||||
#[arg(long, value_name = "SECS", default_value_t = 300)]
|
||||
/// guard); resumes automatically when fresh work arrives. Default 600 (10
|
||||
/// minutes). 0 disables.
|
||||
#[arg(long, value_name = "SECS", default_value_t = 600)]
|
||||
job_timeout: u64,
|
||||
|
||||
/// Open a local control server on 127.0.0.1:<PORT> so the GUI config tool can
|
||||
@@ -139,8 +148,11 @@ struct Args {
|
||||
#[arg(long, default_value = "all")]
|
||||
devices: String,
|
||||
|
||||
/// GPU backend: "opencl" or "cuda" (for nvidia cards).
|
||||
#[arg(long, default_value = "cuda")]
|
||||
/// GPU backend: "mixed" (default — each card on its native backend: NVIDIA
|
||||
/// on CUDA, AMD/Intel on OpenCL), "opencl" (every card via OpenCL), or
|
||||
/// "cuda" (NVIDIA only). In mixed mode `--devices` indexes the combined list
|
||||
/// shown by --list-devices.
|
||||
#[arg(long, default_value = "mixed")]
|
||||
backend: String,
|
||||
|
||||
/// Force the OpenCL backend, disabling CUDA (overrides --backend).
|
||||
@@ -610,6 +622,9 @@ fn main() -> Result<()> {
|
||||
/// Which GPU backend the user selected.
|
||||
enum BackendKind {
|
||||
Cpu,
|
||||
/// Each physical card on its native backend (NVIDIA→CUDA, others→OpenCL).
|
||||
#[cfg(any(feature = "gpu", feature = "cuda"))]
|
||||
Mixed,
|
||||
#[cfg(feature = "gpu")]
|
||||
OpenCl,
|
||||
#[cfg(feature = "cuda")]
|
||||
@@ -633,6 +648,16 @@ fn backend_kind(args: &Args) -> Result<BackendKind> {
|
||||
}
|
||||
}
|
||||
match args.backend.to_ascii_lowercase().as_str() {
|
||||
"mixed" => {
|
||||
// Each card on its native backend; falls back to whatever single GPU
|
||||
// backend is compiled, or to CPU when none is.
|
||||
#[cfg(any(feature = "gpu", feature = "cuda"))]
|
||||
{
|
||||
Ok(BackendKind::Mixed)
|
||||
}
|
||||
#[cfg(not(any(feature = "gpu", feature = "cuda")))]
|
||||
Ok(BackendKind::Cpu)
|
||||
}
|
||||
"cuda" => {
|
||||
#[cfg(feature = "cuda")]
|
||||
{
|
||||
@@ -649,7 +674,7 @@ fn backend_kind(args: &Args) -> Result<BackendKind> {
|
||||
#[cfg(not(feature = "gpu"))]
|
||||
Ok(BackendKind::Cpu)
|
||||
}
|
||||
other => Err(anyhow!("unknown --backend '{other}' (expected opencl or cuda)")),
|
||||
other => Err(anyhow!("unknown --backend '{other}' (expected mixed, opencl, or cuda)")),
|
||||
}
|
||||
}
|
||||
|
||||
@@ -707,6 +732,8 @@ fn backend_specs(args: &Args, gpu_devices: &[GpuDeviceCfg]) -> Result<Vec<Backen
|
||||
let clamp = (args.cpu_clamp != 0).then_some(args.cpu_clamp);
|
||||
return Ok(vec![BackendSpec::Cpu(clamp)]);
|
||||
}
|
||||
// Mixed builds its own unified list (each card on its native backend).
|
||||
BackendKind::Mixed => return mixed_specs(args),
|
||||
#[cfg(feature = "cuda")]
|
||||
BackendKind::Cuda => (cuda::device_count()?, true),
|
||||
#[cfg(feature = "gpu")]
|
||||
@@ -735,6 +762,71 @@ fn backend_specs(args: &Args, gpu_devices: &[GpuDeviceCfg]) -> Result<Vec<Backen
|
||||
}
|
||||
}
|
||||
|
||||
/// The unified device list for the `mixed` backend, as `(label, spec)`: each
|
||||
/// physical GPU on its native backend, with no card mined twice. NVIDIA cards go
|
||||
/// to CUDA (listed first); the remaining OpenCL devices (AMD/Intel, plus NVIDIA
|
||||
/// when CUDA is unavailable) go to OpenCL. Shared by [`mixed_specs`] and
|
||||
/// [`list_devices`]; `--devices` indexes into this list.
|
||||
#[cfg(any(feature = "gpu", feature = "cuda"))]
|
||||
fn mixed_plan() -> Vec<(String, BackendSpec)> {
|
||||
/// Drop a leading `"[<n>] "` index prefix from a backend's device label, so
|
||||
/// the mixed list shows its own single index instead of two.
|
||||
fn strip_index(label: &str) -> &str {
|
||||
label
|
||||
.strip_prefix('[')
|
||||
.and_then(|s| s.split_once("] "))
|
||||
.map(|(_, rest)| rest)
|
||||
.unwrap_or(label)
|
||||
}
|
||||
|
||||
#[allow(unused_mut)]
|
||||
let mut plan: Vec<(String, BackendSpec)> = Vec::new();
|
||||
|
||||
// NVIDIA cards via CUDA, when the backend is compiled and the driver loads.
|
||||
#[cfg(feature = "cuda")]
|
||||
let cuda_has_nvidia = {
|
||||
let names = cuda::list_devices().unwrap_or_default();
|
||||
for (i, label) in names.iter().enumerate() {
|
||||
plan.push((format!("{} (CUDA)", strip_index(label)), BackendSpec::Cuda(i)));
|
||||
}
|
||||
!names.is_empty()
|
||||
};
|
||||
#[cfg(not(feature = "cuda"))]
|
||||
let cuda_has_nvidia = false;
|
||||
|
||||
// Remaining OpenCL cards via OpenCL; skip NVIDIA ones already on CUDA.
|
||||
#[cfg(feature = "gpu")]
|
||||
{
|
||||
let names = gpu::list_devices().unwrap_or_default();
|
||||
let nvidia = gpu::device_is_nvidia();
|
||||
for (j, label) in names.iter().enumerate() {
|
||||
if nvidia.get(j).copied().unwrap_or(false) && cuda_has_nvidia {
|
||||
continue;
|
||||
}
|
||||
plan.push((format!("{} (OpenCL)", strip_index(label)), BackendSpec::Gpu(j)));
|
||||
}
|
||||
}
|
||||
// `cuda_has_nvidia` is only consumed by the OpenCL branch above.
|
||||
#[cfg(not(feature = "gpu"))]
|
||||
let _ = cuda_has_nvidia;
|
||||
|
||||
plan
|
||||
}
|
||||
|
||||
/// Build the worker list for `--backend mixed`: each card on its native backend.
|
||||
/// `--devices` selects into [`mixed_plan`]'s unified list.
|
||||
#[cfg(any(feature = "gpu", feature = "cuda"))]
|
||||
fn mixed_specs(args: &Args) -> Result<Vec<BackendSpec>> {
|
||||
let plan = mixed_plan();
|
||||
if plan.is_empty() {
|
||||
return Err(anyhow!(
|
||||
"no GPUs found for the mixed backend — none detected via CUDA or OpenCL"
|
||||
));
|
||||
}
|
||||
let selected = parse_devices(&args.devices, plan.len())?;
|
||||
Ok(selected.into_iter().map(|i| plan[i].1).collect())
|
||||
}
|
||||
|
||||
/// Build a single GPU worker spec for `idx`, choosing CUDA or OpenCL, erroring if
|
||||
/// the requested backend wasn't compiled in.
|
||||
#[cfg(any(feature = "gpu", feature = "cuda"))]
|
||||
@@ -821,6 +913,18 @@ fn list_devices() {
|
||||
Ok(_) => println!("no CUDA devices found"),
|
||||
Err(e) => println!("error listing CUDA devices: {e}"),
|
||||
}
|
||||
// What the default `mixed` backend will mine, and the indices `--devices`
|
||||
// selects from in that mode.
|
||||
#[cfg(any(feature = "gpu", feature = "cuda"))]
|
||||
{
|
||||
let plan = mixed_plan();
|
||||
if !plan.is_empty() {
|
||||
println!("\nMixed backend (--backend mixed, the default) — `--devices` indexes this list:");
|
||||
for (i, (label, _)) in plan.iter().enumerate() {
|
||||
println!(" [{i}] {label}");
|
||||
}
|
||||
}
|
||||
}
|
||||
#[cfg(not(any(feature = "gpu", feature = "cuda")))]
|
||||
println!("built without GPU support (rebuild with the `gpu` or `cuda` feature)");
|
||||
}
|
||||
@@ -886,18 +990,24 @@ fn selftest(gpu_device: usize) -> Result<()> {
|
||||
let solver = gpu::GpuSolver::new(gpu_device)
|
||||
.with_context(|| format!("init OpenCL device {gpu_device}"))?;
|
||||
|
||||
// Spot-check the BLAKE2b kernel against the CPU reference.
|
||||
let outputs = solver.hash_all(&header)?;
|
||||
let step = params::BLAKE_CALLS / 64;
|
||||
for k in 0..64 {
|
||||
let g = (k * step) as u32;
|
||||
let cpu = blake::generate_hash(&base, g);
|
||||
let off = g as usize * params::HASH_OUTPUT;
|
||||
if cpu != outputs[off..off + params::HASH_OUTPUT] {
|
||||
return Err(anyhow!("GPU BLAKE2b mismatch at g={g}"));
|
||||
// Spot-check the BLAKE2b kernel against the CPU reference. The AMD kernel
|
||||
// buckets its round-0 output instead of exposing per-index digests, so
|
||||
// the probe is skipped there (the solve-vs-CPU check below still runs).
|
||||
if solver.supports_blake_probe() {
|
||||
let outputs = solver.hash_all(&header)?;
|
||||
let step = params::BLAKE_CALLS / 64;
|
||||
for k in 0..64 {
|
||||
let g = (k * step) as u32;
|
||||
let cpu = blake::generate_hash(&base, g);
|
||||
let off = g as usize * params::HASH_OUTPUT;
|
||||
if cpu != outputs[off..off + params::HASH_OUTPUT] {
|
||||
return Err(anyhow!("GPU BLAKE2b mismatch at g={g}"));
|
||||
}
|
||||
}
|
||||
info!("GPU BLAKE2b kernel matches CPU");
|
||||
} else {
|
||||
info!("skipping BLAKE2b kernel probe (AMD kernel buckets round-0 output)");
|
||||
}
|
||||
info!("GPU BLAKE2b kernel matches CPU");
|
||||
|
||||
let gpu_solutions = solver.solve(&header)?;
|
||||
info!("GPU found {} valid solution(s)", gpu_solutions.len());
|
||||
|
||||
Reference in New Issue
Block a user