authorgravatar for 3575188313@qq.comZhenming-Lin <3575188313@qq.com> 2026-02-13 15:38:49+08:00
committergravatar for andrew@ziglang.orgAndrew Kelley <andrew@ziglang.org> 2026-02-13 12:20:53-08:00
log8901d38b7d3eab22af87bba9962f5a749fb31661
tree6caa7f46fc86daedcf8b6aa7de215ce8fd042f57
parent07b08b96389a1dd60bc150a5ed72586ee57b2303

std.math.atan: Add `@Vector(?,f32)` and `@Vector(?,f64)` support

1 files changed, 211 insertions(+), 8 deletions(-)

lib/std/math/atan.zig+211-8
...@@ -4,6 +4,12 @@...@@ -4,6 +4,12 @@
4// https://git.musl-libc.org/cgit/musl/tree/src/math/atanf.c4// https://git.musl-libc.org/cgit/musl/tree/src/math/atanf.c
5// https://git.musl-libc.org/cgit/musl/tree/src/math/atan.c5// https://git.musl-libc.org/cgit/musl/tree/src/math/atan.c
6// https://git.musl-libc.org/cgit/musl/tree/src/math/atanl.c6// https://git.musl-libc.org/cgit/musl/tree/src/math/atanl.c
7//
8// Ported from ARM-software, which is licensed under the MIT license:
9// https://github.com/ARM-software/optimized-routines/blob/master/LICENSE
10//
11// https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/atanf.c
12// https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/atan.c
713
8const std = @import("../std.zig");14const std = @import("../std.zig");
9const math = std.math;15const math = std.math;
...@@ -17,14 +23,22 @@ const testing = std.testing;...@@ -17,14 +23,22 @@ const testing = std.testing;
17/// - atan(+-inf) = +-pi/223/// - atan(+-inf) = +-pi/2
18pub fn atan(x: anytype) @TypeOf(x) {24pub fn atan(x: anytype) @TypeOf(x) {
19 const T = @TypeOf(x);25 const T = @TypeOf(x);
20 return switch (T) {26 switch (T) {
21 f16 => atanBinary16(x),27 f16 => return atanBinary16(x),
22 f32 => atanBinary32(x),28 f32 => return atanBinary32(x),
23 f64 => atanBinary64(x),29 f64 => return atanBinary64(x),
24 f80 => atanExtended80(x),30 f80 => return atanExtended80(x),
25 f128 => atanBinary128(x),31 f128 => return atanBinary128(x),
26 else => @compileError("atan not implemented for " ++ @typeName(T)),32 else => {
27 };33 const info = @typeInfo(T);
34 if (info == .vector) {
35 const vec_type = info.vector.child;
36 if (vec_type == f32) return atanBinary32Vec(x);
37 if (vec_type == f64) return atanBinary64Vec(x);
38 }
39 @compileError("atan not implemented for " ++ @typeName(T));
40 },
41 }
28}42}
2943
30fn atanBinary16(x: f16) f16 {44fn atanBinary16(x: f16) f16 {
...@@ -579,3 +593,192 @@ test "atanBinary128" {...@@ -579,3 +593,192 @@ test "atanBinary128" {
579 try testing.expectApproxEqAbs(atanBinary128(0x1.299d54ac7d6afc5154643b601519p1), 0x1.2a24e22d861debfd6f974500567fp0, math.floatEpsAt(f128, 0x1.2a24e22d861debfd6f974500567fp0));593 try testing.expectApproxEqAbs(atanBinary128(0x1.299d54ac7d6afc5154643b601519p1), 0x1.2a24e22d861debfd6f974500567fp0, math.floatEpsAt(f128, 0x1.2a24e22d861debfd6f974500567fp0));
580 try testing.expectApproxEqAbs(atanBinary128(-0x1.0264fb9f3d50e4f0f966f0686064p1), -0x1.1c617825f97512b7f38656ab12cdp0, math.floatEpsAt(f128, -0x1.1c617825f97512b7f38656ab12cdp0));594 try testing.expectApproxEqAbs(atanBinary128(-0x1.0264fb9f3d50e4f0f966f0686064p1), -0x1.1c617825f97512b7f38656ab12cdp0, math.floatEpsAt(f128, -0x1.1c617825f97512b7f38656ab12cdp0));
581}595}
596
597fn atanBinary32Vec(x: anytype) @TypeOf(x) {
598 const type_info = @typeInfo(@TypeOf(x));
599 comptime std.debug.assert(type_info.vector.child == f32);
600 const vec_len = type_info.vector.len;
601 const sign_mask: @Vector(vec_len, u32) = @splat(0x80000000);
602 const neg_one: @Vector(vec_len, f32) = @splat(-1.0);
603 const pi_over_2: @Vector(vec_len, u32) = @splat(0x3fc90fdb);
604 const zero: @Vector(vec_len, u32) = @splat(0);
605 const c0: @Vector(vec_len, f32) = @splat(-0x1.5554dcp-2);
606 const c1: @Vector(vec_len, f32) = @splat(0x1.9978ecp-3);
607 const c2: @Vector(vec_len, f32) = @splat(-0x1.230a94p-3);
608 const c3: @Vector(vec_len, f32) = @splat(0x1.b4debp-4);
609 const c4: @Vector(vec_len, f32) = @splat(-0x1.3550dap-4);
610 const c5: @Vector(vec_len, f32) = @splat(0x1.61eebp-5);
611 const c6: @Vector(vec_len, f32) = @splat(-0x1.0c17d4p-6);
612 const c7: @Vector(vec_len, f32) = @splat(0x1.7ea694p-9);
613
614 const ix: @Vector(vec_len, u32) = @bitCast(x);
615 const sign = ix & sign_mask;
616 const pred = @abs(x) > @abs(neg_one);
617 const z = @select(f32, pred, neg_one / x, x);
618 const shift: @Vector(vec_len, f32) = @bitCast(@select(u32, pred, pi_over_2 ^ sign, zero));
619 const z2 = z * z;
620 const z3 = z * z2;
621 const z4 = z2 * z2;
622 const z8 = z4 * z4;
623 const p0_1 = @mulAdd(@Vector(vec_len, f32), z2, c1, c0);
624 const p2_3 = @mulAdd(@Vector(vec_len, f32), z2, c3, c2);
625 const p4_5 = @mulAdd(@Vector(vec_len, f32), z2, c5, c4);
626 const p6_7 = @mulAdd(@Vector(vec_len, f32), z2, c7, c6);
627 const p0_3 = @mulAdd(@Vector(vec_len, f32), z4, p2_3, p0_1);
628 const p4_7 = @mulAdd(@Vector(vec_len, f32), z4, p6_7, p4_5);
629 const p0_7 = @mulAdd(@Vector(vec_len, f32), z8, p4_7, p0_3);
630 return @mulAdd(@Vector(vec_len, f32), z3, p0_7, shift + z);
631}
632
633fn atanBinary64Vec(x: anytype) @TypeOf(x) {
634 const type_info = @typeInfo(@TypeOf(x));
635 comptime std.debug.assert(type_info.vector.child == f64);
636 const vec_len = type_info.vector.len;
637 const sign_mask: @Vector(vec_len, u64) = @splat(0x8000000000000000);
638 const neg_one: @Vector(vec_len, f64) = @splat(-1.0);
639 const pi_over_2: @Vector(vec_len, u64) = @splat(0x3ff921fb54442d18);
640 const zero: @Vector(vec_len, u64) = @splat(0);
641 const c0: @Vector(vec_len, f64) = @splat(-0x1.555555555552ap-2);
642 const c1: @Vector(vec_len, f64) = @splat(0x1.9999999995aebp-3);
643 const c2: @Vector(vec_len, f64) = @splat(-0x1.24924923923f6p-3);
644 const c3: @Vector(vec_len, f64) = @splat(0x1.c71c7184288a2p-4);
645 const c4: @Vector(vec_len, f64) = @splat(-0x1.745d11fb3d32bp-4);
646 const c5: @Vector(vec_len, f64) = @splat(0x1.3b136a18051b9p-4);
647 const c6: @Vector(vec_len, f64) = @splat(-0x1.110e6d985f496p-4);
648 const c7: @Vector(vec_len, f64) = @splat(0x1.e1bcf7f08801dp-5);
649 const c8: @Vector(vec_len, f64) = @splat(-0x1.ae644e28058c3p-5);
650 const c9: @Vector(vec_len, f64) = @splat(0x1.82eeb1fed85c6p-5);
651 const c10: @Vector(vec_len, f64) = @splat(-0x1.59d7f901566cbp-5);
652 const c11: @Vector(vec_len, f64) = @splat(0x1.2c982855ab069p-5);
653 const c12: @Vector(vec_len, f64) = @splat(-0x1.eb49592998177p-6);
654 const c13: @Vector(vec_len, f64) = @splat(0x1.69d8b396e3d38p-6);
655 const c14: @Vector(vec_len, f64) = @splat(-0x1.ca980345c4204p-7);
656 const c15: @Vector(vec_len, f64) = @splat(0x1.dc050eafde0b3p-8);
657 const c16: @Vector(vec_len, f64) = @splat(-0x1.7ea70755b8eccp-9);
658 const c17: @Vector(vec_len, f64) = @splat(0x1.ba3da3de903e8p-11);
659 const c18: @Vector(vec_len, f64) = @splat(-0x1.44a4b059b6f67p-13);
660 const c19: @Vector(vec_len, f64) = @splat(0x1.c4a45029e5a91p-17);
661
662 const ix: @Vector(vec_len, u64) = @bitCast(x);
663 const sign = ix & sign_mask;
664 const pred = @abs(x) > @abs(neg_one);
665 const shift: @Vector(vec_len, f64) = @bitCast(@select(u64, pred, pi_over_2 ^ sign, zero));
666 const z = @select(f64, pred, neg_one / x, x);
667 const z2 = z * z;
668 const z3 = z * z2;
669 const z4 = z2 * z2;
670 const z8 = z4 * z4;
671 const z16 = z8 * z8;
672 const p0_1 = @mulAdd(@Vector(vec_len, f64), z2, c1, c0);
673 const p2_3 = @mulAdd(@Vector(vec_len, f64), z2, c3, c2);
674 const p0_3 = @mulAdd(@Vector(vec_len, f64), z4, p2_3, p0_1);
675 const p4_5 = @mulAdd(@Vector(vec_len, f64), z2, c5, c4);
676 const p6_7 = @mulAdd(@Vector(vec_len, f64), z2, c7, c6);
677 const p4_7 = @mulAdd(@Vector(vec_len, f64), z4, p6_7, p4_5);
678 const p0_7 = @mulAdd(@Vector(vec_len, f64), z8, p4_7, p0_3);
679 const p8_9 = @mulAdd(@Vector(vec_len, f64), z2, c9, c8);
680 const p10_11 = @mulAdd(@Vector(vec_len, f64), z2, c11, c10);
681 const p8_11 = @mulAdd(@Vector(vec_len, f64), z4, p10_11, p8_9);
682 const p12_13 = @mulAdd(@Vector(vec_len, f64), z2, c13, c12);
683 const p14_15 = @mulAdd(@Vector(vec_len, f64), z2, c15, c14);
684 const p12_15 = @mulAdd(@Vector(vec_len, f64), z4, p14_15, p12_13);
685 const p16_17 = @mulAdd(@Vector(vec_len, f64), z2, c17, c16);
686 const p18_19 = @mulAdd(@Vector(vec_len, f64), z2, c19, c18);
687 const p16_19 = @mulAdd(@Vector(vec_len, f64), z4, p18_19, p16_17);
688 const p8_15 = @mulAdd(@Vector(vec_len, f64), z8, p12_15, p8_11);
689 const p8_19 = @mulAdd(@Vector(vec_len, f64), z16, p16_19, p8_15);
690 const p0_19 = @mulAdd(@Vector(vec_len, f64), p8_19, z16, p0_7);
691 return @mulAdd(@Vector(vec_len, f64), z3, p0_19, shift + z);
692}
693
694test "atanBinary32Vec.special" {
695 const input: @Vector(7, f32) = .{
696 0x0p+0,
697 -0x0p+0,
698 0x1p+0,
699 -0x1p+0,
700 math.inf(f32),
701 -math.inf(f32),
702 math.nan(f32),
703 };
704 const output = atanBinary32Vec(input);
705 try testing.expectEqual(output[0], 0x0p+0);
706 try testing.expectEqual(output[1], -0x0p+0);
707 try testing.expectApproxEqAbs(output[2], 0x1.921fb6p-1, math.floatEpsAt(f32, 0x1.921fb6p-1));
708 try testing.expectApproxEqAbs(output[3], -0x1.921fb6p-1, math.floatEpsAt(f32, -0x1.921fb6p-1));
709 try testing.expectApproxEqAbs(output[4], 0x1.921fb6p+0, math.floatEpsAt(f32, 0x1.921fb6p+0));
710 try testing.expectApproxEqAbs(output[5], -0x1.921fb6p+0, math.floatEpsAt(f32, -0x1.921fb6p+0));
711 try testing.expect(math.isNan(output[6]));
712}
713
714test "atanBinary32Vec" {
715 const input: @Vector(10, f32) = .{
716 -0x1.8629dp-2,
717 -0x1.59d42ep1,
718 -0x1.d2dbe2p0,
719 -0x1.5f314ep-1,
720 0x1.5869bp1,
721 -0x1.b13a06p-2,
722 0x1.3cb0f2p1,
723 -0x1.0ed746p-2,
724 0x1.299d54p1,
725 -0x1.0264fcp1,
726 };
727 const output = atanBinary32Vec(input);
728 try testing.expectApproxEqAbs(output[0], -0x1.74c62p-2, math.floatEpsAt(f32, -0x1.74c62p-2));
729 try testing.expectApproxEqAbs(output[1], -0x1.375fd8p0, math.floatEpsAt(f32, -0x1.375fd8p0));
730 try testing.expectApproxEqAbs(output[2], -0x1.11b8aep0, math.floatEpsAt(f32, -0x1.11b8aep0));
731 try testing.expectApproxEqAbs(output[3], -0x1.33d28cp-1, math.floatEpsAt(f32, -0x1.33d28cp-1));
732 try testing.expectApproxEqAbs(output[4], 0x1.37082ep0, math.floatEpsAt(f32, 0x1.37082ep0));
733 try testing.expectApproxEqAbs(output[5], -0x1.99d7cap-2, math.floatEpsAt(f32, -0x1.99d7cap-2));
734 try testing.expectApproxEqAbs(output[6], 0x1.2fcb12p0, math.floatEpsAt(f32, 0x1.2fcb12p0));
735 try testing.expectApproxEqAbs(output[7], -0x1.08c71ap-2, math.floatEpsAt(f32, -0x1.08c71ap-2));
736 try testing.expectApproxEqAbs(output[8], 0x1.2a24e2p0, math.floatEpsAt(f32, 0x1.2a24e2p0));
737 try testing.expectApproxEqAbs(output[9], -0x1.1c6178p0, math.floatEpsAt(f32, -0x1.1c6178p0));
738}
739
740test "atanBinary64Vec.special" {
741 const input: @Vector(7, f64) = .{
742 0x0p+0,
743 -0x0p+0,
744 0x1p+0,
745 -0x1p+0,
746 math.inf(f64),
747 -math.inf(f64),
748 math.nan(f64),
749 };
750 const output = atanBinary64Vec(input);
751 try testing.expectEqual(output[0], 0x0p+0);
752 try testing.expectEqual(output[1], -0x0p+0);
753 try testing.expectApproxEqAbs(output[2], 0x1.921fb54442d18p-1, math.floatEpsAt(f64, 0x1.921fb54442d18p-1));
754 try testing.expectApproxEqAbs(output[3], -0x1.921fb54442d18p-1, math.floatEpsAt(f64, -0x1.921fb54442d18p-1));
755 try testing.expectApproxEqAbs(output[4], 0x1.921fb54442d18p+0, math.floatEpsAt(f64, 0x1.921fb54442d18p+0));
756 try testing.expectApproxEqAbs(output[5], -0x1.921fb54442d18p+0, math.floatEpsAt(f64, -0x1.921fb54442d18p+0));
757 try testing.expect(math.isNan(output[6]));
758}
759
760test "atanBinary64Vec" {
761 const input: @Vector(10, f64) = .{
762 -0x1.8629d0244cdccp-2,
763 -0x1.59d42d4659937p1,
764 -0x1.d2dbe23d04f06p0,
765 -0x1.5f314e72398e8p-1,
766 0x1.5869af37b7d08p1,
767 -0x1.b13a05a662618p-2,
768 0x1.3cb0f12f39d8ap1,
769 -0x1.0ed746b39cbb7p-2,
770 0x1.299d54ac7d6bp1,
771 -0x1.0264fb9f3d50ep1,
772 };
773 const output = atanBinary64Vec(input);
774 try testing.expectApproxEqAbs(output[0], -0x1.74c61f4377016p-2, math.floatEpsAt(f64, -0x1.74c61f4377016p-2));
775 try testing.expectApproxEqAbs(output[1], -0x1.375fd7987cc2p0, math.floatEpsAt(f64, -0x1.375fd7987cc2p0));
776 try testing.expectApproxEqAbs(output[2], -0x1.11b8adeba5616p0, math.floatEpsAt(f64, -0x1.11b8adeba5616p0));
777 try testing.expectApproxEqAbs(output[3], -0x1.33d28ca762539p-1, math.floatEpsAt(f64, -0x1.33d28ca762539p-1));
778 try testing.expectApproxEqAbs(output[4], 0x1.37082ce2dd03p0, math.floatEpsAt(f64, 0x1.37082ce2dd03p0));
779 try testing.expectApproxEqAbs(output[5], -0x1.99d7cac66dd44p-2, math.floatEpsAt(f64, -0x1.99d7cac66dd44p-2));
780 try testing.expectApproxEqAbs(output[6], 0x1.2fcb120468e8ep0, math.floatEpsAt(f64, 0x1.2fcb120468e8ep0));
781 try testing.expectApproxEqAbs(output[7], -0x1.08c71aa0e509p-2, math.floatEpsAt(f64, -0x1.08c71aa0e509p-2));
782 try testing.expectApproxEqAbs(output[8], 0x1.2a24e22d861dfp0, math.floatEpsAt(f64, 0x1.2a24e22d861dfp0));
783 try testing.expectApproxEqAbs(output[9], -0x1.1c617825f9751p0, math.floatEpsAt(f64, -0x1.1c617825f9751p0));
784}