Improve fastcall half-float benchmark stability

* Increase ctypes benchmark samples

Use more ASV repeats for the half-float ctypes round-trip benchmark. The extra samples reduce noise while keeping the targeted benchmark runtime well below the expected limit.

Signed-off-by: Eric Shi <ershi@nvidia.com>

* Improve half-float ctypes benchmark

Reduce the round-trip ctypes sample size so each ASV measurement is short enough to avoid millisecond-scale scheduler noise. Increase the method-specific repeat count to preserve statistical power while keeping the targeted benchmark runtime bounded.

Signed-off-by: Eric Shi <ershi@nvidia.com>

Approved-by: Eric Shi <ershi@nvidia.com>

See merge request omniverse/warp!2406

Author: shi-eric

asv/benchmarks/fastcall.py

@@ -19,9 +19,11 @@
 class HalfFloatConversion:
     """Benchmark half-float conversion via METH_FASTCALL and ctypes paths."""
 
-    # The fastcall path is short enough that a ~2 ms sample can hide scheduler
-    # noise. Keep those samples near 0.05-0.10 ms and compensate with more ASV
-    # repeats. The ctypes path is slower, so keep its sampling unchanged.
+    # Short fastcall loops can hide scheduler noise in millisecond-scale
+    # samples. Keep those samples near 0.05-0.10 ms and compensate with more
+    # ASV repeats. The single-conversion ctypes benchmarks still use larger
+    # inner loops; only the round-trip ctypes benchmark uses a shorter loop
+    # because it performs two ctypes calls per iteration.
     repeat = 300
     number = 1
     warmup_time = 0.1
@@ -63,10 +65,11 @@ def time_round_trip_fastcall(self):
     def time_round_trip_ctypes(self):
         to_half = self.ctypes.wp_float_to_half_bits
         to_float = self.ctypes.wp_half_bits_to_float
-        for _ in range(5_000):
+        for _ in range(100):
             to_float(to_half(1.0))
 
 
 HalfFloatConversion.time_float_to_half_bits_fastcall.repeat = 2_000
 HalfFloatConversion.time_half_bits_to_float_fastcall.repeat = 2_000
 HalfFloatConversion.time_round_trip_fastcall.repeat = 2_000
+HalfFloatConversion.time_round_trip_ctypes.repeat = 20_000