x64: pool: Accuracy issue in benchnn pooling test #2678
Conversation
github-actions
bot
added
the
platform:cpu-x64
jstachowintel
changed the title
jstachowintel
force-pushed
the
fix_overflowing
branch
from
fb69a22 to
f81b3a7
Compare
jstachowintel
force-pushed
the
fix_overflowing
branch
from
da5e72e to
ec29352
Compare
| @@ -75,6 +75,13 @@ struct jit_uni_pooling_fwd_t : public primitive_t { | |||
| CHECK(jit_uni_pool_kernel<isa>::init_conf( | |||
| jpp_, scratchpad, attr_, this)); | |||
|
|
|||
| dim_t src_size = static_cast<dim_t>(jpp_.mb) * jpp_.c * jpp_.id | |||
There was a problem hiding this comment.
Why do we need to type cast? Is this check just going to fail if we happen to get problem dimensions exceeding INT_MAX?
There was a problem hiding this comment.
If the product of our input sizes exceeds INT_MAX, an overflow will occur, and we want to skip this implementation. We need to use casting to get the correct, not overflowed number that can be compared with INT_MAX.
There was a problem hiding this comment.
This issue is that jpp_.mb may have already overflowed. As I understand this code, all the problem dimensions are set here. The src_d memory descriptor returns a dim_t which is truncated for storing in jpp. If we set one of those dimensions to something like jpp.mb UINT_MAX + 1, then jpp_.mb will be 1 and the check may inadvertently pass. It seem better to just directly check the memory descriptor sizes similar to here or to change the dimensions types in jpp to dim_t so that no cast is necessary.
It is similar use case as #2627. The problem is with huge input sizes, so I decided disable
src_size <= INT_MAXfor jit:avx512_core implementation.