szha commented on a change in pull request #20338:
URL: https://github.com/apache/incubator-mxnet/pull/20338#discussion_r657103406
##########
File path: python/mxnet/util.py
##########
@@ -848,3 +848,20 @@ def setenv(name, value):
"""
passed_value = None if value is None else c_str(value)
check_call(_LIB.MXSetEnv(c_str(name), passed_value))
+
+def set_flush_denorms(value):
+ """Change floating-point calculations when dealing with denormalized
values.
+ This is only applicable to architectures which supports flush-to-zero.
+ Denormalized values are positive and negative values that are very
close to 0
+ (exponent is the smallest possible value).
+ Flushing denormalized values to 0 can speedup calculations if such
values occurs,
+ but if fulfilling whole IEEE 754 standard is required this option
should be disabled.
+ Parameters
+ ----------
+ value : bool
+ State of flush-to-zero and denormals-are-zero in MXCSR register
Review comment:
@ptrendx is there an equivalent setting for CUDA? What behavior should
we expect for denormal values in GPU calculation?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
