-
Notifications
You must be signed in to change notification settings - Fork 113
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 code change workaround for 64-bit reduce_by_segment bug #1791
base: main
Are you sure you want to change the base?
Conversation
Filling the SYCL private memory array with the identity prior to loading data works around the encountered IGC bug. No real performance impact can be measured with this change. The current macro workaround is also removed Signed-off-by: Matthew Michel <[email protected]>
Signed-off-by: Matthew Michel <[email protected]>
@@ -34,21 +34,11 @@ namespace unseq_backend | |||
//This optimization depends on Intel(R) oneAPI DPC++ Compiler implementation such as support of binary operators from std namespace. | |||
//We need to use defined(SYCL_IMPLEMENTATION_INTEL) macro as a guard. | |||
|
|||
template <typename _Tp> | |||
inline constexpr bool __can_use_known_identity = | |||
# if ONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION |
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.
Is it okay if we directly remove this macro check, or should it be deprecated first with a #warning
or something similar? From the user's perspective, they should just see reduce_by_segment
speedup with 64-bit types.
// TODO: Remove this initialization to the identity when possible. We load real data to __loc_partials | ||
// in the first loop below but this initialization to the identity works around an IGC register | ||
// filling bug. | ||
std::array<__val_type, __vals_per_item> __loc_partials = {__identity}; |
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.
Is this meant to fill the array with the identity value? Because I believe as it is currently written, only the first value in the array would be populated and the rest will be uninitialized. If the intent is for all of the elements to be the identity, then this can be written as:
std::array<__val_type, __vals_per_item> __loc_partials = {__identity}; | |
std::array<__val_type, __vals_per_item> __loc_partials; | |
std::fill(__loc_partials.begin(), __loc_partials.end(), __identity); |
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.
Thanks, it looks like the rest of the elements may be initialized to 0: https://en.cppreference.com/w/c/language/array_initialization.
The way the fix was implemented still worked since it does not matter what is loaded into the array as long as it's something. However, I switched to your suggestion to be consistent.
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.
Filling the array after its definition seems to reintroduce the bug. I will see if I can find a better solution. I suppose what we originally had adds a default constructability requirement we do not want.
Signed-off-by: Matthew Michel <[email protected]>
@@ -351,7 +351,12 @@ __sycl_reduce_by_segment(__internal::__hetero_tag<_BackendTag>, _ExecutionPolicy | |||
__seg_reduce_wg_kernel, | |||
#endif | |||
sycl::nd_range<1>{__n_groups * __wgroup_size, __wgroup_size}, [=](sycl::nd_item<1> __item) { | |||
::std::array<__val_type, __vals_per_item> __loc_partials; | |||
auto __identity = unseq_backend::__known_identity<_BinaryOperator, __val_type>; |
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.
Let's use __val_type
instead of auto
- it's will more readable, I think.
There is an IGC bug that affects
reduce_by_segment
with 64-bit types on GPU Series Max devices which has previously required us to provide theONEDPL_WORKAROUND_FOR_IGPU_64BIT_REDUCTION
macro workaround. This workaround invokes the legacy implementation which is around ~3x slower but produces correct results.The IGC bug still exists, but I have a found a workaround with negligible performance impact within our
reduce_by_segment
implementation. This enables users to invoke the fasterreduce_by_segment
implementation without correctness issues.By first initializing the private memory arrays to the known identity element prior to loading real data into some of the array indices, the register filling bug is avoided. I have verified with oneDPL tests (which previously caught this issue) and with external tests.
I have also removed the macro workaround and additional test.
I've collected information on the performance impact which is negligible. Feel free to request if you would like to see it.