MultiShift, Masked shift and masked interleave #2431
Conversation
|
Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA). View this failed invocation of the CLA check for more information. For the most up to date status, view the checks section at the bottom of the pull request. |
|
AVX3_DL has _mm_multishift_epi64_epi8, _mm256_multishift_epi64_epi8, and _mm512_multishift_epi64_epi8 intrinsics that can be used to implement MultiShift on AVX3_DL. |
|
|
||
| * `V`: `{u,i}64`, `VI`: `{u,i}8` \ | ||
| <code>V **MultiShift**(V vals, VI indices)</code>: returns a | ||
| vector with `(vals[i] >> indices[i*8+j]) & 0xff` in byte `j` of `r[i]` for each |
There was a problem hiding this comment.
Let's define r, for example "vector r".
| for(size_t i = 0; i < N; i++) { | ||
| uint64_t shift_result = 0; | ||
| for(int j = 0; j < 8; j++) { | ||
| uint64_t rot_result = (v[i] >> indices[i*8+j]) | (v[i] << (64 - indices[i*8+j])); |
There was a problem hiding this comment.
Because we're only using the lower 8 bits, we can just talk about the right-shift and not a rotate, right?
|
|
||
| #### Masked Shifts | ||
| * `V`: `{u,i}` \ | ||
| <code>V **MaskedShiftLeftOrZero**<int>(M mask, V a)</code> returns `a[i] << int` or `0` if |
There was a problem hiding this comment.
As before, let's drop the OrZero suffix?
| @@ -2081,6 +2124,12 @@ Ops in this section are only available if `HWY_TARGET != HWY_SCALAR`: | |||
| `InterleaveOdd(d, a, b)` is usually more efficient than `OddEven(b, | |||
| DupOdd(a))`. | |||
|
|
|||
| * <code>V **InterleaveEvenOrZero**(M m, V a, V b)</code>: Performs the same | |||
There was a problem hiding this comment.
Add Masked prefix and remove OrZero suffix? I notice this doesn't have an SVE implementation yet?
| @@ -574,6 +592,27 @@ HWY_API V MaskedSatSubOr(V no, M m, V a, V b) { | |||
| } | |||
| #endif // HWY_NATIVE_MASKED_ARITH | |||
|
|
|||
| // ------------------------------ MaskedShift | |||
| template <int kshift, class V, class M> | |||
There was a problem hiding this comment.
Nit: please rename kshift -> kShift.
| @@ -7299,6 +7338,62 @@ HWY_API V BitShuffle(V v, VI idx) { | |||
|
|
|||
| #endif // HWY_NATIVE_BITSHUFFLE | |||
|
|
|||
| // ------------------------------ MultiShift (Rol) | |||
There was a problem hiding this comment.
The ops in parentheses are ops that the implementation of the one in this section uses, for purposes of sorting the ops in the source file. I think this is a copy-paste remnant? We can remove it because this implementation does not seem to use any ops defined in generic_ops-inl.h.
| const auto extract_mask = Dup128VecFromValues(du8, 0, 2, 4, 6, 8, 10, 12, 14, | ||
| 0, 0, 0, 0, 0, 0, 0, 0); | ||
| const auto even_lanes = | ||
| BitCast(d64, TableLookupBytes(extracted_even_bytes, extract_mask)); |
There was a problem hiding this comment.
I think we could use ConcatEven here?
Introduces the following operations for
generic_ops-inl.handarm_sve-inl.hwhere the use of sve intrinsics will provide a performance gain:g3doc/quick_reference.md.Tests have been introduced for all the listed operations.