[session-cpu] Subtle change for profiling result

maekawatoshiki · Aug 4, 2024 · c4240f9 · c4240f9
1 parent d48361d
commit c4240f9
Show file tree

Hide file tree

Showing 4 changed files with 17 additions and 24 deletions.
diff --git a/crates/session-cpu/examples/deit_cpu.rs b/crates/session-cpu/examples/deit_cpu.rs
@@ -1,6 +1,5 @@
 use std::{cmp::Ordering, fs::read_to_string, path::Path, time::Instant};
 
-use altius_core::flops::compute_flops;
 use ndarray::CowArray;
 use ort::{Environment, GraphOptimizationLevel, SessionBuilder, Value};
 use structopt::StructOpt;
@@ -70,11 +69,9 @@ fn main() {
         let mut out = out[0].data::<f32>().iter().enumerate().collect::<Vec<_>>();
         out.sort_by(|(_, a), (_, b)| b.partial_cmp(a).unwrap_or(Ordering::Equal));
 
-        log::info!("inference result: {}", classes[out[0].0]);
-        log::info!("top5: {:?}", &out[..5]);
+        println!("inference result: {}", classes[out[0].0]);
+        println!("top5: {:?}", &out[..5]);
     }
-    let flops = compute_flops(i.model()).unwrap();
-    log::info!("Model FLOPs: {} ({}M)", flops, flops / 1_000_000);
 }
 
 fn run_on_ort(opt: &Opt) {

diff --git a/crates/session-cpu/examples/mobilenet_cpu.rs b/crates/session-cpu/examples/mobilenet_cpu.rs
@@ -1,4 +1,3 @@
-use altius_core::flops::compute_flops;
 use ndarray::CowArray;
 use ort::{Environment, ExecutionProvider, SessionBuilder, Value};
 use structopt::StructOpt;
@@ -77,8 +76,8 @@ fn main() {
             let mut out = out.iter().enumerate().collect::<Vec<_>>();
             out.sort_by(|(_, a), (_, b)| b.partial_cmp(a).unwrap_or(Ordering::Equal));
 
-            log::info!("prediction: {}", classes[out[0].0]);
-            log::info!("top5: {:?}", &out[..5]);
+            println!("prediction: {}", classes[out[0].0]);
+            println!("top5: {:?}", &out[..5]);
         }
     } else {
         let model = load_onnx(root.join("mobilenetv3.onnx")).unwrap();
@@ -94,10 +93,8 @@ fn main() {
             let mut out = out[0].data::<f32>().iter().enumerate().collect::<Vec<_>>();
             out.sort_by(|(_, a), (_, b)| b.partial_cmp(a).unwrap_or(Ordering::Equal));
 
-            log::info!("prediction: {}", classes[out[0].0]);
-            log::info!("top5: {:?}", &out[..5]);
+            println!("prediction: {}", classes[out[0].0]);
+            println!("top5: {:?}", &out[..5]);
         }
-        let flops = compute_flops(session.model()).unwrap();
-        log::info!("Model FLOPs: {} ({}M)", flops, flops / 1_000_000);
     }
 }
diff --git a/crates/session-cpu/src/session.rs b/crates/session-cpu/src/session.rs
@@ -1,4 +1,5 @@
 use altius_core::{
+    flops::compute_flops,
     model::Model,
     tensor::{Tensor, TypedFixedShape},
     value::ValueId,
@@ -65,12 +66,18 @@ impl CPUSession {
                 })
                 .collect::<Vec<_>>();
             let sum_durations = durations.iter().map(|(_, d)| d).sum::<f32>();
-            durations.push(("All", entire_duration));
             durations.push(("All (Kernel)", sum_durations));
             durations.sort_by(|(_, b), (_, a)| a.partial_cmp(b).unwrap());
             let width = durations.iter().map(|(op, _)| op.len()).max().unwrap();
             for (op, duration) in durations {
-                log::info!("| {op:width$}: {duration:.5} [ms]");
+                log::info!("{op:width$}: {duration:.5} ms");
+            }
+            if let Ok(flops) = compute_flops(&self.model) {
+                log::info!(
+                    "[ {:.5} ms, {:.5} GFLOPS ]",
+                    entire_duration,
+                    flops as f32 / (entire_duration / 1000.0) / 1_000_000_000.0
+                );
             }
         }
 

diff --git a/crates/session-cpu/src/translator.rs b/crates/session-cpu/src/translator.rs
@@ -234,8 +234,8 @@ impl<'a> Translator<'a> {
                             ));
                         }
                         num_compilied_kernels.fetch_add(num_kernels, Ordering::SeqCst);
-                        log::debug!(
-                            "Compiled {}/{} kernels",
+                        eprint!(
+                            "[{:03}/{:03}] Compilation in progress\r",
                             num_compilied_kernels.load(Ordering::SeqCst),
                             num_kernels_to_compile
                         );
@@ -726,8 +726,6 @@ elapsed_{opname} += end_in_sec - start_in_sec;",
     ) -> Result<String, SessionError> {
         let input_names = &args[..inputs.len()];
         let output_names = &args[inputs.len()..];
-        log::debug!("input names: {:?}", input_names);
-        log::debug!("output names: {:?}", output_names);
 
         let input = &inputs[Op::CONV2D_IN];
         let _weight = &inputs[Op::CONV2D_WEIGHT];
@@ -764,8 +762,6 @@ elapsed_{opname} += end_in_sec - start_in_sec;",
         let _pad_b = padding[2];
         let _pad_r = padding[3];
 
-        log::debug!("kernel: {:?}", kernel);
-
         let code_fill_bias = if let Some(bias) = input_names.get(Op::CONV2D_BIAS) {
             let output_name = &output_names[0];
             format!(
@@ -3358,10 +3354,6 @@ impl Regions {
     fn find_first_free_region(&self, size: usize) -> Range<usize> {
         // Sorted by start offset
         let regions = self.start_to_region.values().collect::<Vec<_>>();
-        log::debug!(
-            "regions: {regions:?} (count: {count})",
-            count = regions.len()
-        );
         fn roundup(x: usize) -> usize {
             let mask = 32 - 1;
             (x + mask) & !mask