Profiling integer dtype ops
zaccharieramzi opened this issue · comments
I would be interested in profiling ops that act on integer tensors on GPU.
For example I would like to profile the addition in the following fashion (see this colab with tboard):
import tensorflow as tf
chosen_dtype = tf.int32
a = tf.constant([7] * 100000, dtype=chosen_dtype)
with tf.profiler.experimental.Profile("logs"):
for i in range(1):
a + 3However, doing this the profiler doesn't record any GPU operation. If I just change the chosen_dtype to tf.float32, then I have the correct GPU operations (in addition to the eager execution operations).
How can I profile integer dtype ops?
It doesn't seem to be a profiler issue. Apparently, when integer is used,
the ops are scheduled on the CPU instead of GPU (you should see those CPU
ops in the profile).
So, you probably need to ask this question in the general TF forum for how
to force those integer ops to run on the GPU.
…On Thu, Jul 2, 2020 at 7:57 AM Zaccharie Ramzi ***@***.***> wrote:
I would be interested in profiling ops that act on integer tensors on GPU.
For example I would like to profile the addition in the following fashion
(see this colab with tboard
<https://colab.research.google.com/drive/1uuvPs68dObcQ0HyAocbjr7MJHxqhnwdW?usp=sharing>
):
import tensorflow as tfchosen_dtype = tf.int32a = tf.constant([7] * 100000, dtype=chosen_dtype)
with tf.profiler.experimental.Profile("logs"):
for i in range(1):
a + 3
However, doing this the profiler doesn't record any GPU operation. If I
just change the chosen_dtype to tf.float32, then I have the correct GPU
operations (in addition to the eager execution operations).
How can I profile integer dtype ops?
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#85>, or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AE33L3J2GMNVMLDYSP62SUTRZSN7LANCNFSM4OO7BAYA>
.
You are right