validate: profile CUDA commands using nvprof

Pipfile

@@ -11,10 +11,12 @@ matplotlib = "*"
 numpy = "*"
 scipy = "*"
 pandas = "*"
-jupyterlab = "*"
 wasabi = "*"
+click = "*"
 
 [dev-packages]
 invoke = "*"
 flake8 = "*"
 black = "*"
+jupyterlab = "*"
+mypy = "*"

WIP.md

@@ -1,16 +1,12 @@
 #### TODO
 
-- notes
-
-  - l2 cache is shared in sub partitions, we get different l2 metrics
-  - l1i are the same always, as they are local to the core
-
 - today:
 
-  - record mem fetch latency in playground and box
-  - write trait for tag array
+  - convert, match and plot statistics
   - rename crates and github repo
   - publish to crates.io
+  - record mem fetch latency in playground and box
+  - DONE: write trait for tag array
 
 - TODO:
 
@@ -21,7 +17,6 @@
   - use traits for common components
   - record mem fetch latency
   - add a few more stats
-  - plot statistics
   - publish python package to pip
 
 - tomorrow:

accelsim/src/stats.rs

@@ -1,5 +1,6 @@
 use color_eyre::eyre;
 use std::collections::HashMap;
+use utils::box_slice;
 
 pub type Stat = (String, u16, String);
 pub type Map = indexmap::IndexMap<Stat, f64>;
@@ -102,7 +103,7 @@ impl TryFrom<Stats> for stats::Stats {
         .into_iter()
         .collect();
 
-        let mut l2d_stats = stats::PerCache(vec![stats::Cache::default(); 1].into_boxed_slice());
+        let mut l2d_stats = stats::PerCache(box_slice![stats::Cache::default(); 1]);
         let l2d_total = &mut l2d_stats[0];
         for kind in AccessKind::iter() {
             for reservation_failure in ReservationFailure::iter() {
@@ -160,10 +161,11 @@ impl TryFrom<Stats> for stats::Stats {
             },
             accesses: stats::Accesses(accesses),
             dram: stats::DRAM {
-                bank_writes: vec![vec![vec![total_dram_writes]]],
-                bank_reads: vec![vec![vec![total_dram_reads]]],
-                total_bank_writes: vec![vec![total_dram_writes]],
-                total_bank_reads: vec![vec![total_dram_reads]],
+                bank_writes: box_slice![box_slice![box_slice![total_dram_writes]]],
+                bank_reads: box_slice![box_slice![box_slice![total_dram_reads]]],
+                total_bank_writes: box_slice![box_slice![total_dram_writes]],
+                total_bank_reads: box_slice![box_slice![total_dram_reads]],
+                ..stats::DRAM::default()
             },
             instructions: stats::InstructionCounts::default(),
             l1i_stats: stats::PerCache::new(0),

benches/vectoradd.rs

@@ -111,10 +111,10 @@ pub fn box_benchmark(c: &mut Criterion) {
 }
 
 criterion::criterion_group!(benches, box_benchmark, play_benchmark, accelsim_benchmark);
-// criterion::criterion_main!(benches);
+criterion::criterion_main!(benches);
 
 #[allow(dead_code)]
-fn main() -> eyre::Result<()> {
+fn main_other() -> eyre::Result<()> {
     use itertools::Itertools;
     #[allow(unused_imports)]
     use std::io::Write;

gpucachesim/__init__.py

@@ -0,0 +1,3 @@
+from pathlib import Path
+
+ROOT_DIR = Path(__file__).parent

gpucachesim/benchmarks.py

@@ -0,0 +1,42 @@
+import click
+import yaml
+from pathlib import Path
+from os import PathLike
+from typing import Optional
+
+from gpucachesim import ROOT_DIR
+
+REPO_ROOT_DIR = ROOT_DIR.parent
+DEFAULT_BENCH_FILE = REPO_ROOT_DIR / "test-apps/test-apps-materialized.yml"
+
+
+class Benchmarks:
+    def __init__(self, path: PathLike) -> None:
+        """load the materialized benchmark config"""
+
+        with open(path or DEFAULT_BENCH_FILE, "rb") as f:
+            benchmarks = yaml.safe_load(f)
+
+        self.benchmarks = benchmarks["benchmarks"]
+
+    def __getitem__(self, bench_name: str):
+        return self.benchmarks[bench_name]
+
+    def get_bench_config(self, bench_name: str, input_idx: int):
+        return self.benchmarks[bench_name][input_idx]
+
+
+@click.command()
+@click.option("--path", default=DEFAULT_BENCH_FILE, help="Path to materialized benchmark config")
+def main(path):
+    from pprint import pprint
+
+    print(path)
+    b = Benchmarks(path)
+
+    benchmark_names = list(b.benchmarks.keys())
+    pprint(benchmark_names)
+
+
+if __name__ == "__main__":
+    main()

gpucachesim/stats/__init__.py

@@ -0,0 +1,27 @@
+import click
+
+import gpucachesim.stats.stats as stats
+import gpucachesim.stats.native as native
+from gpucachesim.benchmarks import Benchmarks
+
+
+@click.command()
+@click.option("--path", help="Path to materialized benchmark config")
+@click.option("--bench", help="Benchmark name")
+@click.option("--input", default=0, help="Input index")
+def main(path, bench, input):
+    from pprint import pprint
+
+    b = Benchmarks(path)
+    if bench is None:
+        raise NotImplemented
+    print(bench, input)
+    bench_config = b.get_bench_config(bench, input)
+    # pprint(bench_config)
+
+    our_stats = stats.Stats(bench_config["simulate"])
+    native_stats = native.Stats(bench_config["simulate"])
+
+
+if __name__ == "__main__":
+    main()

gpucachesim/stats/accelsim.py

@@ -0,0 +1,2 @@
+class AccelsimStats:
+    pass

gpucachesim/stats/native.py

@@ -0,0 +1,6 @@
+from os import PathLike
+
+
+class Stats:
+    def __init__(self, result_dir: PathLike) -> None:
+        self.path = result_dir

gpucachesim/stats/stats.py

@@ -0,0 +1,6 @@
+from os import PathLike
+
+
+class Stats:
+    def __init__(self, result_dir: PathLike) -> None:
+        self.path = result_dir