refactor: use time.perf_counter for consistent and accurate benchmarking (#2943)

benchmarks/python/comparative/bench_mlx.py

@@ -38,10 +38,10 @@ def bench(f, *args):
     for i in range(10):
         f(*args)
 
-    s = time.time()
+    s = time.perf_counter()
     for i in range(100):
         f(*args)
-    e = time.time()
+    e = time.perf_counter()
     return e - s
 
 

benchmarks/python/comparative/bench_torch.py

@@ -37,10 +37,10 @@ def bench(f, *args):
     for i in range(10):
         f(*args)
 
-    s = time.time()
+    s = time.perf_counter()
     for i in range(100):
         f(*args)
-    e = time.time()
+    e = time.perf_counter()
     return e - s
 
 

benchmarks/python/time_utils.py

@@ -31,8 +31,8 @@ def measure_runtime(fn, **kwargs):
     for _ in range(5):
         fn(**kwargs)
 
-    tic = time.time()
+    tic = time.perf_counter()
     iters = 100
     for _ in range(iters):
         fn(**kwargs)
-    return (time.time() - tic) * 1000 / iters
+    return (time.perf_counter() - tic) * 1000 / iters

docs/src/dev/extensions.rst

@@ -777,11 +777,11 @@ with the naive :meth:`simple_axpby` we first defined.
             mx.eval(z)
 
         # Timed run
-        s = time.time()
+        s = time.perf_counter()
         for i in range(100):
             z = f(x, y, alpha, beta)
             mx.eval(z)
-        e = time.time()
+        e = time.perf_counter()
         return 1000 * (e - s) / 100
 
     simple_time = bench(simple_axpby)

examples/python/linear_regression.py

@@ -29,12 +29,12 @@ def loss_fn(w):
 
 grad_fn = mx.grad(loss_fn)
 
-tic = time.time()
+tic = time.perf_counter()
 for _ in range(num_iters):
     grad = grad_fn(w)
     w = w - lr * grad
     mx.eval(w)
-toc = time.time()
+toc = time.perf_counter()
 
 loss = loss_fn(w)
 error_norm = mx.sum(mx.square(w - w_star)).item() ** 0.5