[SYCL][ESIMD][EMU] LSC support for ESIMD_EMULATOR backend #6099
Conversation
36 out of 40 LSC tests are passing
| static_assert(ImmOffset == 0); | ||
| static_assert(DS != __ESIMD_ENS::lsc_data_size::u16u32h); | ||
|
|
||
| constexpr uint MASK = loadstoreAlignMask<Ty, VS, DS, N>(); |
There was a problem hiding this comment.
Please move this down to the place of use. Same is true for Output, ElemCount
| constexpr uint MASK = loadstoreAlignMask<Ty, VS, DS, N>(); | ||
| __ESIMD_DNS::vector_type_t<Ty, N * __ESIMD_EDNS::to_int<VS>()> Output = 0; | ||
|
|
||
| constexpr int ElemCount = __ESIMD_EDNS::to_int<VS>(); |
There was a problem hiding this comment.
I think N's definition /// @tparam N is the number of channels (platform dependent). is wrong here and in many other places in this file. This is actually the simd size - the number of chunks of VS*sizeof(T) size. If we take rgba load/strore as reference, then VS would be the number of channels, not N. So ElemCount is actually the number of channels. It would be great if you could change N's definition to something like "the SIMD size of operation (the number of addresses to access)".
There was a problem hiding this comment.
Comment updated.
| ByteDistance += rawAddressIncrement<Ty, DS>(), | ||
| VecIdx += vectorIndexIncrement<N, _Transposed>()) { | ||
|
|
||
| Output[VecIdx] = *((Ty *)(addrs[AddrIdx] + ByteDistance)); |
There was a problem hiding this comment.
Please hoist addrs[AddrIdx] out of the loop
BaseAddr = addrs[AddrIdx];
assert((BaseAddr & MASK) == 0 && "Address Alignment Error!!");
for (...) {
Output[VecIdx] = *((Ty *)(BaseAddr + ByteDistance));
...
- Updating comments for definition of 'N' : SIMD operationsize - addrs[] to BaseAddr - Relocating variables close to the place of its use - Elem* to Chanl*
| @@ -260,7 +657,8 @@ __ESIMD_INTRIN void __esimd_raw_send_nbarrier_signal( | |||
| /// @tparam DS is the data size. | |||
| /// @tparam VS is the number of elements to load per address. | |||
| /// @tparam Transposed indicates if the data is transposed during the transfer. | |||
| /// @tparam N is the number of channels (platform dependent). | |||
| /// @tparam N is the SIMD size of operation (the number of addresses to access, | |||
| /// platform dependent). | |||
There was a problem hiding this comment.
I don't quite understand what "platform-dependent" means for N, as this is what user passes in the code. Restrictions maybe platform-dependent, but that's normal situation for many other ESIMD API parameters, so I suggest to remove it here and on all other similar places
There was a problem hiding this comment.
Removed 'platform dependent' comment.
|
Unrelated failures from HIP AMDGPU LLVM Test Suite |
|
@dongkyunahn-intel, @kbobrovs, please check post-commit fails on Windows: https://github.com/intel/llvm/actions/runs/2416580740 |
|
Posted PR for fix - #6223 |
36 out of 40 LSC tests are passing