Add AVX2 support #23035
Add AVX2 support #23035
Conversation
site/source/docs/porting/simd.rst
Outdated
| * - _mm_i64gather_epi64 | ||
| - ❌ scalarized | ||
|
|
||
| All the 128-bit wide instructions from AVX2 instruction set are listed. Only a small part of the 256-bit AVX2 instruction set are listed, most of the 256-bit wide AVX2 instructions are emulated by two 128-bit wide instructions. |
There was a problem hiding this comment.
Can you wrap text at 80 here
| * found in the LICENSE file. | ||
| */ | ||
|
|
||
| #ifndef __emscripten_immintrin_h__ |
There was a problem hiding this comment.
Let's use a different macro that gets undefined after all the relevant includes in immintrin.h. Otherwise, we won't emit this error if the user does something like this:
#include <immintrin.h>
#include <avxintrin.h>
There was a problem hiding this comment.
Yes, in this case, it won't emit error message, but I think it's suitable.
Actually, include avxintrin.h directly will miss prerequisites and cause compile error. Include immintrin.h can avoid this error.
immintrin.h includes the necessary prerequisites and choose the *intrin.h corresponding to options(-mavx, -mavx2...).
so in the following code, if -mavx option is specified, #include <avxintrin.h> does nothing due to header guards, it's valid code, the behavior is consistent with LLVM.
#include <immintrin.h>
#include <avxintrin.h>
system/include/compat/avx2intrin.h
Outdated
| // This may cause an out-of-bounds memory load since we first load and | ||
| // then mask, but since there are no segmentation faults in Wasm memory | ||
| // accesses, that is ok (as long as we are within the heap bounds - | ||
| // a negligible limitation in practice) | ||
| // TODO, loadu or load, 128-bit align? |
There was a problem hiding this comment.
The snippet is ported from old code in avxintrin.h, I am not sure, I will check that.
There was a problem hiding this comment.
The implementation is changed, scalarized version is used now, it's more compliant with specifications.
jiepan-intel
force-pushed
the
avx2_to_wasm
branch
from
e838094 to
1551fa2
Compare
jiepan-intel
force-pushed
the
avx2_to_wasm
branch
from
dbf0628 to
4f6db96
Compare
jiepan-intel
force-pushed
the
avx2_to_wasm
branch
from
da4ba4f to
50f8885
Compare
|
Looks like this was lgtm'd pending tests passing, and they just passed, landing. |
4 participants
It's successor of PR #22430, one 256-bit AVX2 intrinsic can be also emulated by two 128-bit intrinsics.