Avx512f #912

minybot · 2020-09-17T20:55:43Z

knot, kandn, kxornd, kmov
permute: epi32
extractf32x4_ps not mask and maskz
permute_f32x4
permute_f64x2
permute_i32x4
permute_i64x2
moveldup_ps
movehdup_ps
movedup_pd

merge

merge from base

merge base

rust-highfive · 2020-09-17T20:55:47Z

r? @Amanieu

(rust_highfive has picked a reviewer for you, use r? to override)

Amanieu · 2020-09-19T13:50:58Z

crates/core_arch/src/x86/avx512f.rs

+/// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=knot_mask16&expand=3233)
+#[inline]
+#[target_feature(enable = "avx512f")]
+#[cfg_attr(all(test, not(target_os = "macos")), assert_instr(not))] // generate normal not code instead of knotw


Why are you special-casing macos here?

In osX, it does not generate any "not" or "xor" instructions. It generates "vmovaps" tested in CI and osX with clang.

Can you post the fully assembly output you are getting on macOS? I find it extremely strange that it is behaving differently from Linux.

osX Rust

#[target_feature(enable = "avx512f")]
unsafe fn avx512() {
let c = _mm512_knot(0b00000000_00000000);
}

fn main() {
unsafe { avx512(); }
}

.loc 6 10538 0 .cfi_startproc pushq %rbp .cfi_def_cfa_offset 16 .cfi_offset %rbp, -16 movq %rsp, %rbp .cfi_def_cfa_register %rbp subq $16, %rsp movw %di, -4(%rbp)

Ltmp21:
.loc 6 10539 15 prologue_end
movzwl %di, %edi
movl $65535, %esi
callq __ZN9core_arch9core_arch3x867avx512f11_mm512_kxor17hbdf54d0540836621E
movw %ax, -6(%rbp)
.loc 6 0 15 is_stmt 0
movw -6(%rbp), %ax
.loc 6 10539 5
movw %ax, -2(%rbp)
movw -2(%rbp), %ax
movw %ax, -8(%rbp)
.loc 6 0 5
movw -8(%rbp), %ax
.loc 6 10540 2 is_stmt 1
addq $16, %rsp
popq %rbp
retq

minybot · 2020-09-19T16:39:13Z

Clang on osX:
There is a knot now.

mask_0 = _mm512_int2mask( (1<<0) | (1<<1) | (1<<2) | (1<<3) | (1<<4) | (1<<5) | (1<<6) | (1<<7) | (1<<8) | (1<<9) | (1<<10) | (1<<11) | (1<<12) | (1<<13) | (1<<14) | (1<<15) );

mask_2 = _mm512_knot(mask_0);

%bb.0:

    pushq   %rbp
    .cfi_def_cfa_offset 16
    .cfi_offset %rbp, -16
    movq    %rsp, %rbp
    .cfi_def_cfa_register %rbp
    andq    $-64, %rsp
    subq    $1152, %rsp             ## imm = 0x480
    xorl    %eax, %eax
    movl    $65535, 1128(%rsp)      ## imm = 0xFFFF
    movw    1128(%rsp), %cx
    movw    %cx, 30(%rsp)
    movw    30(%rsp), %cx
    movw    %cx, 1134(%rsp)
    kmovw   1134(%rsp), %k0
    knotw   %k0, %k0
    kmovw   %k0, %edx
                                    ## kill: def $dx killed $dx killed $edx
    movw    %dx, 26(%rsp)
    movq    %rbp, %rsp
    popq    %rbp
    retq
    .cfi_endproc
                                    ## -- End function

.subsections_via_symbols

Amanieu · 2020-09-19T17:14:59Z

For the k{not,and,xor, etc} function I would prefer if we used simple bitwise operations on integers rather than intrinsics. As a bonus it would allow the compiler to optimize these functions better.

We could then remove the mostly useless assert_instr on these intrinsics.

minybot · 2020-09-19T17:18:13Z

For the k{not,and,xor, etc} function I would prefer if we used simple bitwise operations on integers rather than intrinsics. As a bonus it would allow the compiler to optimize these functions better.

We could then remove the mostly useless assert_instr on these intrinsics.
Ok. I will update a new version use bitwise operations.

Amanieu · 2020-09-19T17:43:10Z

knot doesn't use a not instruction because it only inverts a 16-bit value. It uses xor 0xffff instead.

minybot · 2020-09-19T17:45:55Z

knot doesn't use a not instruction because it only inverts a 16-bit value. It uses xor 0xffff instead.
Yes, I just remove assert_instr on these intrinsics?

Amanieu · 2020-09-19T17:46:47Z

You can change it to xor. CI will complain if you have an intrinsic without assert_instr.

Amanieu · 2020-09-19T18:53:49Z

crates/core_arch/src/x86/avx512f.rs

@@ -10524,7 +10524,7 @@ pub unsafe fn _mm512_kxor(a: __mmask16, b: __mmask16) -> __mmask16 {
 /// [Intel's documentation](https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=knot_mask16&expand=3233)
 #[inline]
 #[target_feature(enable = "avx512f")]
-#[cfg_attr(test, assert_instr(not))] // generate normal not code instead of knotw
+//#[cfg_attr(test, assert_instr(xor))] // generate normal not code instead of knotw


Please don't leave commented-out code. Either keep the line with xor or delete it entirely.

minybot · 2020-09-19T20:08:12Z

For osX, it uses xor. For Linux, it uses not. So, I remove assert_instr test for _mm512_not?

Amanieu · 2020-09-19T20:22:01Z

Yes, in that case just remove the assert_instr. I think this happens because the darwin target optimizes for the core2 CPU while other targets optimize for the generic cpu.

minybot added 15 commits September 4, 2020 20:36

Merge pull request #1 from rust-lang/master

731dc70

merge

Merge pull request #3 from rust-lang/master

72ff3d7

merge from base

Merge pull request #4 from rust-lang/master

bcc2a2c

merge base

Merge remote-tracking branch 'upstream/master'

9b729dc

modify avx512f.md

8c2e3d8

knot, kandn, kxornd, kmov

da1fbac

permute: epi32

8c8af48

extractf32x4_ps

cd94300

permute_f32x4

03e2f89

permute_f64x2

edc3886

permute_i32x4, permute_i64x2

65323da

moveldup_ps

7b642db

movehdup_ps

cc8495a

movedup_pd

9fe7441

cargo fmt

e0f44bf

rust-highfive assigned Amanieu Sep 17, 2020

minybot added 5 commits September 17, 2020 21:02

change mask not,xnor instruction to pass test

64f1f70

change mask not instruction to pass test

30b68e1

change mask not instruction to pass test v2

396830d

change mask not instruction to pass test v3

96359ba

change mask not instruction to pass test v4

499854b

Amanieu reviewed Sep 19, 2020

View reviewed changes

use simple bitwise operations on masks

a5f3235

change knot assert_instr from not to xor

5d0b923

Amanieu reviewed Sep 19, 2020

View reviewed changes

remove comment out

260fe3b

remove mask not assert_instr

c6fa7d5

Amanieu merged commit 20f9e2f into rust-lang:master Sep 19, 2020

minybot deleted the avx512f branch September 19, 2020 23:27

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Avx512f #912

Avx512f #912

Uh oh!

minybot commented Sep 17, 2020

Uh oh!

rust-highfive commented Sep 17, 2020

Uh oh!

Amanieu Sep 19, 2020

Uh oh!

minybot Sep 19, 2020

Uh oh!

Amanieu Sep 19, 2020

Uh oh!

minybot Sep 19, 2020

Uh oh!

minybot commented Sep 19, 2020

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

minybot commented Sep 19, 2020

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

minybot commented Sep 19, 2020

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

Amanieu Sep 19, 2020

Uh oh!

minybot commented Sep 19, 2020

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

Uh oh!

Avx512f #912

Avx512f #912

Uh oh!

Conversation

minybot commented Sep 17, 2020

Uh oh!

rust-highfive commented Sep 17, 2020

Uh oh!

Amanieu Sep 19, 2020

Choose a reason for hiding this comment

Uh oh!

minybot Sep 19, 2020

Choose a reason for hiding this comment

Uh oh!

Amanieu Sep 19, 2020

Choose a reason for hiding this comment

Uh oh!

minybot Sep 19, 2020

Choose a reason for hiding this comment

fn main() { unsafe { avx512(); } }

Uh oh!

minybot commented Sep 19, 2020

Clang on osX: There is a knot now.

%bb.0:

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

minybot commented Sep 19, 2020

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

minybot commented Sep 19, 2020

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

Amanieu Sep 19, 2020

Choose a reason for hiding this comment

Uh oh!

minybot commented Sep 19, 2020

Uh oh!

Amanieu commented Sep 19, 2020

Uh oh!

Uh oh!

fn main() {
unsafe { avx512(); }
}

Clang on osX:
There is a knot now.