make multithread-mm a feature (dis by default)

Author: kali

cli/Cargo.toml

@@ -63,3 +63,4 @@ pulse = [ "tract-pulse", "tract-pulse-opl" ]
 tf = [ "tract-tensorflow", "tract-libcli/hir" ]
 tflite = [ "tract-tflite" ]
 conform = [ "tract-tensorflow/conform"  ]
+multithread-mm = [ "tract-linalg/multithread-mm" ]

cli/src/main.rs

@@ -706,11 +706,16 @@ fn handle(matches: clap::ArgMatches, probe: Option<&Probe>) -> TractResult<()> {
 
     let mut need_optimisations = false;
 
+    #[cfg(feature = "multithread-mm")]
     if let Some(threads) = matches.value_of("threads") {
         let threads: usize = threads.parse()?;
         let threads = if threads == 0 { num_cpus::get_physical() } else { threads };
         multithread::set_default_executor(multithread::Executor::multithread(threads));
     }
+    #[cfg(not(feature = "multithread-mm"))]
+    if let Some(_) = matches.value_of("threads") {
+        bail!("tract is compiled without multithread support")
+    }
 
     match matches.subcommand() {
         Some(("bench", m)) => {

linalg/Cargo.toml

@@ -23,7 +23,7 @@ lazy_static.workspace = true
 log.workspace = true
 num-traits.workspace = true
 paste.workspace = true
-rayon.workspace = true
+rayon = { workspace = true, optional = true }
 scan_fmt.workspace = true
 tract-data.workspace = true
 
@@ -61,7 +61,8 @@ proptest = { version = "1.0.0", default-features = false, features = ["std", "bi
 # preferred.
 no_fp16 = []
 apple-amx-ios = []
-default = []
+default = [ ]
+multithread-mm = [ "rayon" ]
 complex = [ "tract-data/complex" ]
 
 [[bench]]

linalg/src/frame/mmm/mod.rs

@@ -16,6 +16,7 @@ mod storage;
 pub mod tests;
 
 use crate::multithread::Executor;
+#[cfg(feature = "multithread-mm")]
 use rayon::prelude::*;
 use std::borrow::Cow;
 use std::cmp::Ordering;
@@ -223,6 +224,7 @@ unsafe fn run_with_scratch_space_vec<K: MatMatMulKer>(
             }
             Ok(())
         }
+        #[cfg(feature = "multithread-mm")]
         Executor::MultiThread(pool) => pool.install(|| {
             (0..m.div_ceil(ker.mr()))
                 .into_par_iter()
@@ -247,6 +249,7 @@ unsafe fn run_with_scratch_space_col_outer<K: MatMatMulKer>(
             }
             Ok(())
         }
+        #[cfg(feature = "multithread-mm")]
         Executor::MultiThread(pool) => pool.install(|| {
             (0..n.div_ceil(ker.nr())).into_par_iter().try_for_each(|ib| {
                 for ia in 0..m.divceil(ker.mr()) {
@@ -274,6 +277,7 @@ unsafe fn run_with_scratch_space_row_outer<K: MatMatMulKer>(
             }
             Ok(())
         }
+        #[cfg(feature = "multithread-mm")]
         Executor::MultiThread(pool) => pool.install(|| {
             pool.install(|| {
                 (0..m.div_ceil(ker.mr())).into_par_iter().try_for_each(|ia| {

linalg/src/multithread.rs

@@ -1,20 +1,25 @@
 use std::cell::RefCell;
+#[allow(unused_imports)]
 use std::sync::{Arc, Mutex};
 
+#[cfg(feature = "multithread-mm")]
 use rayon::{ThreadPool, ThreadPoolBuilder};
 
 #[derive(Debug, Clone, Default)]
 pub enum Executor {
     #[default]
     SingleThread,
+    #[cfg(feature = "multithread-mm")]
     MultiThread(Arc<ThreadPool>),
 }
 
 impl Executor {
+    #[cfg(feature = "multithread-mm")]
     pub fn multithread(n: usize) -> Executor {
         Executor::multithread_with_name(n, "tract-default")
     }
 
+    #[cfg(feature = "multithread-mm")]
     pub fn multithread_with_name(n: usize, name: &str) -> Executor {
         let name = name.to_string();
         let pool = ThreadPoolBuilder::new()