Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add support for
org.apache.spark.sql.catalyst.expressions.Bin
#2760base: branch-25.02
Are you sure you want to change the base?
Add support for
org.apache.spark.sql.catalyst.expressions.Bin
#2760Changes from 1 commit
13874a5
9c59f74
51a0d87
d6d5d4c
6fb3589
a7da182
File filter
Filter by extension
Conversations
Jump to
There are no files selected for viewing
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Useless include (not sure)?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit: we can first check if the value == 0 to save a
__clzll
call in this case.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
But wouldn't that introduce thread divergence? I understand how that can be an advantage on a CPU, but I don't really see it on a GPU.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The PTX translated from the code is as follows:
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, I was thinking there must be an if else to check size == 0 case so we can put it earlier than clz to save some calls without introducing new branch. But it looks like the compiler will optimize the
size > 0 ? size : 1;
to amax.s32
so it's branch less then the original approach looks better in anyway.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I have now changed it to this:
return max(64 - __clzll(value), 1);
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
*d_buffer++ = '0' + ((value & (1LL << i)) >> i);
perhaps this approach is more efficient since it avoids branching, which might degrade performance on GPUs with warp divergence.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
But I am not sure if it is a good practice which is really effective. I would like to hear your opinions on this issue @res-life @ttnghia .
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
maybe
*d_buffer++ = '0' + ((value & (1LL << i)) != 0);
? It will be (very slightly) cheaper and easier to read.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@thirtiseven
Yes, this one is also a Branch-Free expression since the compiler shall use
setne
instruction avoids branching by directly setting a register based on the zero flag (ZF):The corresponding codes of my alternative would be translated into:
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, I think this approach is more efficient.
You may conduct a benchmark test to double confirm.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
NIT: Just some nice-to-have improvement, use
constexpr if
instead ofif
and add an extra template variablenullable
for this functor. Because we already knew whether the column_view is nullable or NOT.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Could we add some edge case like null, LONG_MAX, LONG_MIN?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
do we have a test case at plugin level to make sure Bin(13.3) returns 1101 ?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I tested it locally, and
Bin(13.3)
indeed returns1101
. I will soon submit the plugin PR.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is really a Spark test not a good test for binary ops. Spark only accepts a Long as the input to
bin
https://github.com/apache/spark/blob/3569e768e657d4e28ee7520808ec910cdff2b099/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/mathExpressions.scala#L1010
So any floating point input gets a cast to long inserted in before bin is called. So that test is really a test that Spark is doing the right thing. Even then it would probably be something that we would want to put in the integration tests if we did test it at all.
Note that you can also pass in a string as an input and it will still try to cast it to a long before calling
bin
.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
there you go: https://github.com/apache/spark/blob/master/sql/catalyst/src/test/scala/org/apache/spark/sql/catalyst/expressions/MathExpressionsSuite.scala#L470-L490.