projects
/
pytorch.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Measures for FP16 and FP32
[pytorch.git]
/
speed.py
diff --git
a/speed.py
b/speed.py
index
f682a16
..
9e845db
100755
(executable)
--- a/
speed.py
+++ b/
speed.py
@@
-4,25
+4,30
@@
import time, torch
if torch.cuda.is_available():
device = torch.device('cuda')
if torch.cuda.is_available():
device = torch.device('cuda')
+ sync = lambda: torch.cuda.synchronize()
else:
device = torch.device('cpu')
else:
device = torch.device('cpu')
+ sync = lambda: None
nb_runs = 10000
nb_runs = 10000
-d1, d2, d3 =
50000, 256, 512
+d1, d2, d3 =
2048, 2048, 2048
-a, b = torch.rand(d1, d2).to(device), torch.rand(d2, d3).to(device)
+for t in [ torch.float32, torch.float16 ]:
+ a = torch.rand(d1, d2, device = device, dtype = t)
+ b = torch.rand(d2, d3, device = device, dtype = t)
-start_time = time.perf_counter()
-for k in range(nb_runs):
- c = a @ b
-duration = time.perf_counter() - start_time
+ sync()
+ start_time = time.perf_counter()
+ for k in range(nb_runs):
+ c = a @ b
+ sync()
+ duration = time.perf_counter() - start_time
-nb_flop = float(nb_runs * d1 * d2 * d3)
-speed = nb_flop / duration
+ nb_flop = float(nb_runs * d1 * d2 * d3 * 2) # 1 multiply-and-add is 2 ops
+
speed = nb_flop / duration
-for u in [ '', 'K', 'M', 'G', 'T', 'P' ]:
- if speed < 1e3: break
- speed /= 1e3
-
-print(f'{speed:.02f} {u}flops on {device}')
+ for u in [ '', 'K', 'M', 'G', 'T', 'P' ]:
+ if speed < 1e3: break
+ speed /= 1e3
+ print(f'{speed:.02f} {u}flops with {t} on {device}')