benchmark single kernel launch (#8921)

* benchmark kernel launch

* don't realize unneeded

* faster

* faster metal

* fix mypy

* without sync

* no div 0

* lru cache that

* no sync in the profile

test/external/external_benchmark_kernel_launch.py

@@ -0,0 +1,38 @@
+import time
+from tinygrad import Tensor, TinyJit, Device, Context
+from tinygrad.helpers import Profiling, Timing, GlobalCounters
+
+# python3 test/test_speed_v_torch.py TestSpeed.test_add_a
+
+@TinyJit
+def plus(a:Tensor, b:Tensor): return a+b
+
+if __name__ == "__main__":
+  a = Tensor([1]).realize()
+  b = Tensor([1]).realize()
+  for i in range(5):
+    with Timing(prefix=f"{i}:"):
+      c = plus(a,b)
+      Device[c.device].synchronize()
+  assert c.item() == 2
+  for i in range(5):
+    st = time.perf_counter()
+    c = plus(a,b)
+    et = time.perf_counter() - st
+    print(f"nosync  {i}: {et*1e6:.2f} us")
+  Device[c.device].synchronize()
+  for i in range(5):
+    st = time.perf_counter()
+    c = plus(a,b)
+    Device[c.device].synchronize()
+    et = time.perf_counter() - st
+    print(f"precise {i}: {et*1e6:.2f} us")
+  assert GlobalCounters.time_sum_s == 0
+  with Context(DEBUG=2):
+    st = time.perf_counter()
+    c = plus(a,b)
+    Device[c.device].synchronize()
+    et = time.perf_counter() - st
+  print(f"kernel {GlobalCounters.time_sum_s*1e3:.2f} ms / full {et*1e3:.2f} ms -- {et/(GlobalCounters.time_sum_s+1e-12):.2f} x")
+  with Profiling():
+    c = plus(a,b)

test/test_speed_v_torch.py

@@ -202,8 +202,12 @@ class TestSpeed(unittest.TestCase):
     def f(a, b): return (a*b).sum()
     helper_test_generic_square('mul_sum', 4096, f, f)
 
-  def test_add(self):
-    for N in [1, 1024, 4096]:
+  def test_add_a(self):
+    def f(a, b): return a + b
+    helper_test_generic_square('add', 1, f, f)
+
+  def test_add_big(self):
+    for N in [1024, 4096]:
       def f(a, b): return a + b
       helper_test_generic_square('add', N, f, f)