[SYCL][ESIMD] Add support for lsc mem access APIs #5512
Conversation
Signed-off-by: Sergey Dmitriev <[email protected]>
| @@ -176,11 +176,279 @@ enum class atomic_op : uint8_t { | |||
| /// Compare and exchange (floating point). | |||
| /// <code>if (*addr == src0) *addr = src1;</code> | |||
| fcmpwr = 0x12, | |||
| fadd = 0x13, | |||
There was a problem hiding this comment.
Please follow up with a PR adding comments to these new operations.
| /// Data size or format to read or store | ||
| enum class lsc_data_size : uint8_t { | ||
| default_size = 0, | ||
| u8 = 1, |
There was a problem hiding this comment.
please add documentation comments to undocumented elements in the follow-up PR.
| vals.data()); | ||
| } | ||
|
|
||
| /// 2D flat-address block load. |
There was a problem hiding this comment.
Some usage examples are highly desirable for this and few other lsc APIs. Please add a TODO marker in the follow-up PR.
|
@bader, @vladimirlaz - CUDA testing stalled. This patch does can't affect CUDA - could you please merge? |
| check_lsc_data_size<T, DS>(); | ||
| if (DS != lsc_data_size::default_size) | ||
| return DS; | ||
| else if (sizeof(T) == 1) |
There was a problem hiding this comment.
Actually U8 and U16 data size values are only supported by 2d block messages. D32U8 and D32U16 should be used instead
There was a problem hiding this comment.
I think this has been fixed in f922b01
There was a problem hiding this comment.
Probably the fixed code will not work. When sizeof(T) is 1 or 2, values should be also bit-casted to uint32_t.
There was a problem hiding this comment.
Please review the updated patch. I think this problem has been fixed.
| /// @tparam BlockHeight is the block height in number of elements. | ||
| /// @tparam NBlocks is the number of blocks. | ||
| /// @tparam Transposed is the transposed version or not. | ||
| /// @tparam Transformed is apply VNNI transform or not. |
There was a problem hiding this comment.
VNNI transform is only supported when data size is U8 or U16
There was a problem hiding this comment.
I have added static asserts to check supported data sizes in e9e39f7
| /// @tparam BlockWidth is the block width in number of elements. | ||
| /// @tparam BlockHeight is the block height in number of elements. | ||
| /// @tparam NBlocks is the number of blocks. | ||
| /// @tparam Transposed is the transposed version or not. |
There was a problem hiding this comment.
Transposed messages are only supported for U32 and U64 (with some constraints)
There was a problem hiding this comment.
I have added static asserts to check supported data sizes when Transposed is true in e9e39f7
I need to take this request back - few more comments arrived. |
… with regular atomics
| /// | ||
| /// @tparam T is element type. | ||
| /// @tparam NElts is the number of elements to load per address. | ||
| /// @tparam DS is the data size. |
There was a problem hiding this comment.
All the Transposed (block) messages support only u32 and u64 values for DS
There was a problem hiding this comment.
Added static assert to all Transposed functions in f922b01
|
The precommit failures is unrelated - timeouts on two tests: Probably caused by enormous amount of warning messages. |
Signed-off-by: Sergey Dmitriev [email protected]