[ESIMD] Add lsc_slm_block_load() with merging semantics #8552
Conversation
The new lsc_slm_block_load() has an additional operand 'old_values' that contains the values returned from the function if the predicate passed to it is 0. Signed-off-by: Vyacheslav N Klochkov <vyacheslav.n.klochkov@intel.com>
| __ESIMD_NS::simd<uint32_t, N> offsets = offset; | ||
| return __esimd_lsc_load_slm<T, cache_hint::none, cache_hint::none, | ||
| _AddressScale, _ImmOffset, _DS, _VS, _Transposed, | ||
| N>(pred.data(), offsets.data()); | ||
| AddressScale, ImmOffset, FDS, VS, _Transposed, N>( |
There was a problem hiding this comment.
Not sure if I'm missing something but shouldn't _Transposed be Transposed? I don't see _Transposed defined.
There was a problem hiding this comment.
Good catch! Fixed now.
I re-started a LIT test before uploading for review, but forgot to rebuild the compiler after this last-minute change (removal of those underscores).
| /// @tparam NElts is the number of elements to load per address. | ||
| /// @tparam DS is the data size. | ||
| /// @param offset is the zero-based offset for SLM buffer in bytes. | ||
| /// @param pred is the predicate; if it contains 0, then the actual load |
There was a problem hiding this comment.
"if it contains 0, then the actual load is not performed and the returned value is undefined" is this accurate? I thought it would be old_values?
There was a problem hiding this comment.
Another good catch. It a copy-paste error. Fixed now. Thank you.
…t being tested before 'git push' Signed-off-by: Vyacheslav N Klochkov <vyacheslav.n.klochkov@intel.com>
…nsics_p3_lsc_slm_block_load
|
These timeouts on Linux are unrelated to this PR: |
|
This ESIMD PR is guaranteed to be unrelated to any potential issues in CUDA CI. |
The new lsc_slm_block_load() has an additional operand 'old_values' that contains the values returned from the function if the predicate passed to it is 0.
The corresponding LIT test: intel/llvm-test-suite#1637