[SYCL][ESIMD] Implement accessor-based gather/scatter and scalar I/O #2700
Conversation
| /// Load a scalar value from an accessor. | ||
| template <typename T, typename AccessorTy, CacheHint L1H = CacheHint::None, | ||
| CacheHint L3H = CacheHint::None> | ||
| ESIMD_INLINE ESIMD_NODEBUG T load(AccessorTy acc, uint32_t offset) { |
There was a problem hiding this comment.
This is basically redundant to SIMD1 gather, which is not commonly used in CM for performance reason. Do we really need to expose a dedicated API for this? The naming is too general and could be a source of confusion, perhaps change to scalar_load to make a clear distinction with block_load.
There was a problem hiding this comment.
This is basically redundant to SIMD1 gather
This is a convenient shortcut to calling generic gather/scatter. So I don't think it is redundant. scalar_load sounds good.
is not commonly used in CM for performance reason. Do we really need to expose a dedicated API for this?
there are cases when scalar access is needed by algorithm (e.g. haar), so yes, I think we need this API for scalar access via accessors.
There was a problem hiding this comment.
OK. Please update name to scalar_load/scalar_store.
|
|
||
| // surface index-based gather/scatter: | ||
| // num blocks, scale, surface index, global offset, elem offsets | ||
| {"surf_read", {"gather.scaled2", {t(3), c16(0), aSI(1), a(2), a(3)}}}, |
There was a problem hiding this comment.
Why not map to gather.scaled? It may need general predicate support.
There was a problem hiding this comment.
Good question. I'm actually following advice from the BE team (@aus-intel) who recommended to use this one for now. When predication is needed, we can add this w/o breaking user code.
kbobrovs
force-pushed
the
esimd_acc_gather_scatter
branch
from
607cb77 to
198e799
Compare
…access. Signed-off-by: Konstantin S Bobrovsky <[email protected]>
kbobrovs
force-pushed
the
esimd_acc_gather_scatter
branch
from
198e799 to
4883ae0
Compare
|
Doing compatibility testing. Please do not merge. |
This PR implements accessor-based gather/scatter and scalar memory access.
Functional tests: intel/llvm-test-suite#43
Signed-off-by: Konstantin S Bobrovsky [email protected]