Avx #90
Avx #90
Conversation
With assert_instr: too many instructions in the disassembly
|
I don't know why but |
|
Thanks for the PR! The error is indeed curious here. I wonder if the ABI for calling functions wrt SIMD registers is slightly different on Windows? In any case it looks like the right instructions are being generated, so want to up the instruction limit to maybe 30 instructions? |
| /// lanes using the control in `imm8`. | ||
| #[inline(always)] | ||
| #[target_feature = "+avx"] | ||
| //#[cfg_attr(test, assert_instr(vshufpd, imm8 = 0x0))] // FIXME |
There was a problem hiding this comment.
Should this be enabled now?
There was a problem hiding this comment.
The doc says that vshufpd should be generated but clang does not assert that it is the case. Maybe the expanded code is wrong ?
match (imm8 >> 0) & 1 {
0 =>
match (imm8 >> 1) & 1 {
0 =>
match (imm8 >> 2) & 1 {
0 =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [0, 4, 2, 6]),
_ => simd_shuffle4(a, b, [0, 4, 2, 7]),
},
_ =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [0, 4, 3, 6]),
_ => simd_shuffle4(a, b, [0, 4, 3, 7]),
},
},
_ =>
match (imm8 >> 2) & 1 {
0 =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [0, 5, 2, 6]),
_ => simd_shuffle4(a, b, [0, 5, 2, 7]),
},
_ =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [0, 5, 3, 6]),
_ => simd_shuffle4(a, b, [0, 5, 3, 7]),
},
},
},
_ =>
match (imm8 >> 1) & 1 {
0 =>
match (imm8 >> 2) & 1 {
0 =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [1, 4, 2, 6]),
_ => simd_shuffle4(a, b, [1, 4, 2, 7]),
},
_ =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [1, 4, 3, 6]),
_ => simd_shuffle4(a, b, [1, 4, 3, 7]),
},
},
_ =>
match (imm8 >> 2) & 1 {
0 =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [1, 5, 2, 6]),
_ => simd_shuffle4(a, b, [1, 5, 2, 7]),
},
_ =>
match (imm8 >> 3) & 1 {
0 => simd_shuffle4(a, b, [1, 5, 3, 6]),
_ => simd_shuffle4(a, b, [1, 5, 3, 7]),
},
},
},
}| /// Extract an 8-bit integer from `a`, selected with `imm8`. | ||
| #[inline(always)] | ||
| #[target_feature = "+avx"] | ||
| pub unsafe fn _mm256_extract_epi8(a: i8x32, imm8: i32) -> i32 { |
There was a problem hiding this comment.
Should this and some of the extractions below assert for a particular instruction?
There was a problem hiding this comment.
It appears that there is no instrunction specified for this intrinsic.
There was a problem hiding this comment.
As I had some duplicated work in #95, I've tried to suggest places where some work could still be used :)
src/x86/avx.rs
Outdated
| } | ||
|
|
||
| /// Horizontally add adjacent pairs of double-precision (64-bit) floating-point | ||
| /// elements in `a` and `b`, and pack the results. |
There was a problem hiding this comment.
Is it possible to include documentation for the location of the results, as in #95 ? Eg:
/// In the result, sums of elements from `a` are returned in even locations,
/// while sums of elements from `b` are returned in odd locations.
There was a problem hiding this comment.
I will try to merge your changes in my branch.
src/x86/avx.rs
Outdated
| } | ||
|
|
||
| /// Horizontally add adjacent pairs of single-precision (32-bit) floating-point | ||
| /// elements in `a` and `b`, and pack the results. |
There was a problem hiding this comment.
Similarly, you could add
/// In the result, sums of elements from `a` are returned in locations of
/// indices 0, 1, 4, 5; while sums of elements from `b` are in locations
/// 2, 3, 6, 7.
src/x86/avx.rs
Outdated
| } | ||
|
|
||
| /// Horizontally subtract adjacent pairs of single-precision (32-bit) floating-point | ||
| /// elements in `a` and `b`, and pack the results. |
There was a problem hiding this comment.
/// In the result, sums of elements from `a` are returned in locations of
/// indices 0, 1, 4, 5; while sums of elements from `b` are in locations
/// 2, 3, 6, 7.
src/x86/avx.rs
Outdated
| } | ||
|
|
||
| /// Horizontally subtract adjacent pairs of double-precision (64-bit) floating-point | ||
| /// elements in `a` and `b`, and pack the results. |
There was a problem hiding this comment.
/// In the result, sums of elements from `a` are returned in even locations,
/// while sums of elements from `b` are returned in odd locations.
|
Thanks @gwenn! |
I had trouble implementing
_mm256_dp_ps,_mm_cmp_pd,_mm_cmp_sd(I will try again).I submit this PR because I don't want to prevent you from implementing other avx intrinsics but I will probably continue to commit in this branch until it is merged.