Skip to content

Commit 69caa49

Browse files
JimBobSquarePantsJimBobSquarePants
committed
Add Vector128 rounding
1 parent 3fbbde9 commit 69caa49

File tree

5 files changed

+199
-244
lines changed

5 files changed

+199
-244
lines changed

‎src/ImageSharp/Formats/Jpeg/Components/Block8x8F.Generated.cs

Lines changed: 130 additions & 101 deletions
Original file line numberDiff line numberDiff line change
@@ -3,13 +3,14 @@
33

44
using System.Numerics;
55
using System.Runtime.CompilerServices;
6+
using System.Runtime.Intrinsics;
67

78
// <auto-generated />
89
namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
910

1011
internal partial struct Block8x8F
1112
{
12-
/// <summary>
13+
/// <summary>
1314
/// Level shift by +maximum/2, clip to [0, maximum]
1415
/// </summary>
1516
public void NormalizeColorsInPlace(float maximum)
@@ -37,38 +38,66 @@ public void NormalizeColorsInPlace(float maximum)
3738
}
3839

3940
/// <summary>
40-
/// AVX2-only variant for executing <see cref="NormalizeColorsInPlace"/> and <see cref="RoundInPlace"/> in one step.
41+
/// <see cref="Vector256{Single}"/> version of <see cref="NormalizeColorsInPlace(float)"/> and <see cref="RoundInPlace()"/>.
4142
/// </summary>
43+
/// <param name="maximum">The maximum value to normalize to.</param>
4244
[MethodImpl(InliningOptions.ShortMethod)]
43-
public void NormalizeColorsAndRoundInPlaceVector8(float maximum)
45+
public void NormalizeColorsAndRoundInPlaceVector256(float maximum)
4446
{
45-
var off = new Vector<float>(MathF.Ceiling(maximum * 0.5F));
46-
var max = new Vector<float>(maximum);
47-
48-
ref Vector<float> row0 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V0L);
49-
row0 = NormalizeAndRound(row0, off, max);
50-
51-
ref Vector<float> row1 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V1L);
52-
row1 = NormalizeAndRound(row1, off, max);
53-
54-
ref Vector<float> row2 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V2L);
55-
row2 = NormalizeAndRound(row2, off, max);
56-
57-
ref Vector<float> row3 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V3L);
58-
row3 = NormalizeAndRound(row3, off, max);
59-
60-
ref Vector<float> row4 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V4L);
61-
row4 = NormalizeAndRound(row4, off, max);
62-
63-
ref Vector<float> row5 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V5L);
64-
row5 = NormalizeAndRound(row5, off, max);
65-
66-
ref Vector<float> row6 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V6L);
67-
row6 = NormalizeAndRound(row6, off, max);
68-
69-
ref Vector<float> row7 = ref Unsafe.As<Vector4, Vector<float>>(ref this.V7L);
70-
row7 = NormalizeAndRound(row7, off, max);
71-
47+
Vector256<float> off = Vector256.Create(MathF.Ceiling(maximum * 0.5F));
48+
Vector256<float> max = Vector256.Create(maximum);
49+
50+
ref Vector256<float> row0 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V0L);
51+
row0 = NormalizeAndRoundVector256(row0, off, max);
52+
53+
ref Vector256<float> row1 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V1L);
54+
row1 = NormalizeAndRoundVector256(row1, off, max);
55+
56+
ref Vector256<float> row2 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V2L);
57+
row2 = NormalizeAndRoundVector256(row2, off, max);
58+
59+
ref Vector256<float> row3 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V3L);
60+
row3 = NormalizeAndRoundVector256(row3, off, max);
61+
62+
ref Vector256<float> row4 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V4L);
63+
row4 = NormalizeAndRoundVector256(row4, off, max);
64+
65+
ref Vector256<float> row5 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V5L);
66+
row5 = NormalizeAndRoundVector256(row5, off, max);
67+
68+
ref Vector256<float> row6 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V6L);
69+
row6 = NormalizeAndRoundVector256(row6, off, max);
70+
71+
ref Vector256<float> row7 = ref Unsafe.As<Vector4, Vector256<float>>(ref this.V7L);
72+
row7 = NormalizeAndRoundVector256(row7, off, max);
73+
}
74+
75+
/// <summary>
76+
/// <see cref="Vector128{Single}"/> version of <see cref="NormalizeColorsInPlace(float)"/> and <see cref="RoundInPlace()"/>.
77+
/// </summary>
78+
/// <param name="maximum">The maximum value to normalize to.</param>
79+
[MethodImpl(InliningOptions.ShortMethod)]
80+
public void NormalizeColorsAndRoundInPlaceVector128(float maximum)
81+
{
82+
Vector128<float> off = Vector128.Create(MathF.Ceiling(maximum * 0.5F));
83+
Vector128<float> max = Vector128.Create(maximum);
84+
85+
this.V0L = NormalizeAndRoundVector128(this.V0L.AsVector128(), off, max).AsVector4();
86+
this.V0R = NormalizeAndRoundVector128(this.V0R.AsVector128(), off, max).AsVector4();
87+
this.V1L = NormalizeAndRoundVector128(this.V1L.AsVector128(), off, max).AsVector4();
88+
this.V1R = NormalizeAndRoundVector128(this.V1R.AsVector128(), off, max).AsVector4();
89+
this.V2L = NormalizeAndRoundVector128(this.V2L.AsVector128(), off, max).AsVector4();
90+
this.V2R = NormalizeAndRoundVector128(this.V2R.AsVector128(), off, max).AsVector4();
91+
this.V3L = NormalizeAndRoundVector128(this.V3L.AsVector128(), off, max).AsVector4();
92+
this.V3R = NormalizeAndRoundVector128(this.V3R.AsVector128(), off, max).AsVector4();
93+
this.V4L = NormalizeAndRoundVector128(this.V4L.AsVector128(), off, max).AsVector4();
94+
this.V4R = NormalizeAndRoundVector128(this.V4R.AsVector128(), off, max).AsVector4();
95+
this.V5L = NormalizeAndRoundVector128(this.V5L.AsVector128(), off, max).AsVector4();
96+
this.V5R = NormalizeAndRoundVector128(this.V5R.AsVector128(), off, max).AsVector4();
97+
this.V6L = NormalizeAndRoundVector128(this.V6L.AsVector128(), off, max).AsVector4();
98+
this.V6R = NormalizeAndRoundVector128(this.V6R.AsVector128(), off, max).AsVector4();
99+
this.V7L = NormalizeAndRoundVector128(this.V7L.AsVector128(), off, max).AsVector4();
100+
this.V7R = NormalizeAndRoundVector128(this.V7R.AsVector128(), off, max).AsVector4();
72101
}
73102

74103
/// <summary>
@@ -78,76 +107,76 @@ public void LoadFromInt16Scalar(ref Block8x8 source)
78107
{
79108
ref short selfRef = ref Unsafe.As<Block8x8, short>(ref source);
80109

81-
this.V0L.X = Unsafe.Add(ref selfRef, 0);
82-
this.V0L.Y = Unsafe.Add(ref selfRef, 1);
83-
this.V0L.Z = Unsafe.Add(ref selfRef, 2);
84-
this.V0L.W = Unsafe.Add(ref selfRef, 3);
85-
this.V0R.X = Unsafe.Add(ref selfRef, 4);
86-
this.V0R.Y = Unsafe.Add(ref selfRef, 5);
87-
this.V0R.Z = Unsafe.Add(ref selfRef, 6);
88-
this.V0R.W = Unsafe.Add(ref selfRef, 7);
89-
90-
this.V1L.X = Unsafe.Add(ref selfRef, 8);
91-
this.V1L.Y = Unsafe.Add(ref selfRef, 9);
92-
this.V1L.Z = Unsafe.Add(ref selfRef, 10);
93-
this.V1L.W = Unsafe.Add(ref selfRef, 11);
94-
this.V1R.X = Unsafe.Add(ref selfRef, 12);
95-
this.V1R.Y = Unsafe.Add(ref selfRef, 13);
96-
this.V1R.Z = Unsafe.Add(ref selfRef, 14);
97-
this.V1R.W = Unsafe.Add(ref selfRef, 15);
98-
99-
this.V2L.X = Unsafe.Add(ref selfRef, 16);
100-
this.V2L.Y = Unsafe.Add(ref selfRef, 17);
101-
this.V2L.Z = Unsafe.Add(ref selfRef, 18);
102-
this.V2L.W = Unsafe.Add(ref selfRef, 19);
103-
this.V2R.X = Unsafe.Add(ref selfRef, 20);
104-
this.V2R.Y = Unsafe.Add(ref selfRef, 21);
105-
this.V2R.Z = Unsafe.Add(ref selfRef, 22);
106-
this.V2R.W = Unsafe.Add(ref selfRef, 23);
107-
108-
this.V3L.X = Unsafe.Add(ref selfRef, 24);
109-
this.V3L.Y = Unsafe.Add(ref selfRef, 25);
110-
this.V3L.Z = Unsafe.Add(ref selfRef, 26);
111-
this.V3L.W = Unsafe.Add(ref selfRef, 27);
112-
this.V3R.X = Unsafe.Add(ref selfRef, 28);
113-
this.V3R.Y = Unsafe.Add(ref selfRef, 29);
114-
this.V3R.Z = Unsafe.Add(ref selfRef, 30);
115-
this.V3R.W = Unsafe.Add(ref selfRef, 31);
116-
117-
this.V4L.X = Unsafe.Add(ref selfRef, 32);
118-
this.V4L.Y = Unsafe.Add(ref selfRef, 33);
119-
this.V4L.Z = Unsafe.Add(ref selfRef, 34);
120-
this.V4L.W = Unsafe.Add(ref selfRef, 35);
121-
this.V4R.X = Unsafe.Add(ref selfRef, 36);
122-
this.V4R.Y = Unsafe.Add(ref selfRef, 37);
123-
this.V4R.Z = Unsafe.Add(ref selfRef, 38);
124-
this.V4R.W = Unsafe.Add(ref selfRef, 39);
125-
126-
this.V5L.X = Unsafe.Add(ref selfRef, 40);
127-
this.V5L.Y = Unsafe.Add(ref selfRef, 41);
128-
this.V5L.Z = Unsafe.Add(ref selfRef, 42);
129-
this.V5L.W = Unsafe.Add(ref selfRef, 43);
130-
this.V5R.X = Unsafe.Add(ref selfRef, 44);
131-
this.V5R.Y = Unsafe.Add(ref selfRef, 45);
132-
this.V5R.Z = Unsafe.Add(ref selfRef, 46);
133-
this.V5R.W = Unsafe.Add(ref selfRef, 47);
134-
135-
this.V6L.X = Unsafe.Add(ref selfRef, 48);
136-
this.V6L.Y = Unsafe.Add(ref selfRef, 49);
137-
this.V6L.Z = Unsafe.Add(ref selfRef, 50);
138-
this.V6L.W = Unsafe.Add(ref selfRef, 51);
139-
this.V6R.X = Unsafe.Add(ref selfRef, 52);
140-
this.V6R.Y = Unsafe.Add(ref selfRef, 53);
141-
this.V6R.Z = Unsafe.Add(ref selfRef, 54);
142-
this.V6R.W = Unsafe.Add(ref selfRef, 55);
143-
144-
this.V7L.X = Unsafe.Add(ref selfRef, 56);
145-
this.V7L.Y = Unsafe.Add(ref selfRef, 57);
146-
this.V7L.Z = Unsafe.Add(ref selfRef, 58);
147-
this.V7L.W = Unsafe.Add(ref selfRef, 59);
148-
this.V7R.X = Unsafe.Add(ref selfRef, 60);
149-
this.V7R.Y = Unsafe.Add(ref selfRef, 61);
150-
this.V7R.Z = Unsafe.Add(ref selfRef, 62);
151-
this.V7R.W = Unsafe.Add(ref selfRef, 63);
110+
this.V0L.X = Unsafe.Add(ref selfRef, 0);
111+
this.V0L.Y = Unsafe.Add(ref selfRef, 1);
112+
this.V0L.Z = Unsafe.Add(ref selfRef, 2);
113+
this.V0L.W = Unsafe.Add(ref selfRef, 3);
114+
this.V0R.X = Unsafe.Add(ref selfRef, 4);
115+
this.V0R.Y = Unsafe.Add(ref selfRef, 5);
116+
this.V0R.Z = Unsafe.Add(ref selfRef, 6);
117+
this.V0R.W = Unsafe.Add(ref selfRef, 7);
118+
119+
this.V1L.X = Unsafe.Add(ref selfRef, 8);
120+
this.V1L.Y = Unsafe.Add(ref selfRef, 9);
121+
this.V1L.Z = Unsafe.Add(ref selfRef, 10);
122+
this.V1L.W = Unsafe.Add(ref selfRef, 11);
123+
this.V1R.X = Unsafe.Add(ref selfRef, 12);
124+
this.V1R.Y = Unsafe.Add(ref selfRef, 13);
125+
this.V1R.Z = Unsafe.Add(ref selfRef, 14);
126+
this.V1R.W = Unsafe.Add(ref selfRef, 15);
127+
128+
this.V2L.X = Unsafe.Add(ref selfRef, 16);
129+
this.V2L.Y = Unsafe.Add(ref selfRef, 17);
130+
this.V2L.Z = Unsafe.Add(ref selfRef, 18);
131+
this.V2L.W = Unsafe.Add(ref selfRef, 19);
132+
this.V2R.X = Unsafe.Add(ref selfRef, 20);
133+
this.V2R.Y = Unsafe.Add(ref selfRef, 21);
134+
this.V2R.Z = Unsafe.Add(ref selfRef, 22);
135+
this.V2R.W = Unsafe.Add(ref selfRef, 23);
136+
137+
this.V3L.X = Unsafe.Add(ref selfRef, 24);
138+
this.V3L.Y = Unsafe.Add(ref selfRef, 25);
139+
this.V3L.Z = Unsafe.Add(ref selfRef, 26);
140+
this.V3L.W = Unsafe.Add(ref selfRef, 27);
141+
this.V3R.X = Unsafe.Add(ref selfRef, 28);
142+
this.V3R.Y = Unsafe.Add(ref selfRef, 29);
143+
this.V3R.Z = Unsafe.Add(ref selfRef, 30);
144+
this.V3R.W = Unsafe.Add(ref selfRef, 31);
145+
146+
this.V4L.X = Unsafe.Add(ref selfRef, 32);
147+
this.V4L.Y = Unsafe.Add(ref selfRef, 33);
148+
this.V4L.Z = Unsafe.Add(ref selfRef, 34);
149+
this.V4L.W = Unsafe.Add(ref selfRef, 35);
150+
this.V4R.X = Unsafe.Add(ref selfRef, 36);
151+
this.V4R.Y = Unsafe.Add(ref selfRef, 37);
152+
this.V4R.Z = Unsafe.Add(ref selfRef, 38);
153+
this.V4R.W = Unsafe.Add(ref selfRef, 39);
154+
155+
this.V5L.X = Unsafe.Add(ref selfRef, 40);
156+
this.V5L.Y = Unsafe.Add(ref selfRef, 41);
157+
this.V5L.Z = Unsafe.Add(ref selfRef, 42);
158+
this.V5L.W = Unsafe.Add(ref selfRef, 43);
159+
this.V5R.X = Unsafe.Add(ref selfRef, 44);
160+
this.V5R.Y = Unsafe.Add(ref selfRef, 45);
161+
this.V5R.Z = Unsafe.Add(ref selfRef, 46);
162+
this.V5R.W = Unsafe.Add(ref selfRef, 47);
163+
164+
this.V6L.X = Unsafe.Add(ref selfRef, 48);
165+
this.V6L.Y = Unsafe.Add(ref selfRef, 49);
166+
this.V6L.Z = Unsafe.Add(ref selfRef, 50);
167+
this.V6L.W = Unsafe.Add(ref selfRef, 51);
168+
this.V6R.X = Unsafe.Add(ref selfRef, 52);
169+
this.V6R.Y = Unsafe.Add(ref selfRef, 53);
170+
this.V6R.Z = Unsafe.Add(ref selfRef, 54);
171+
this.V6R.W = Unsafe.Add(ref selfRef, 55);
172+
173+
this.V7L.X = Unsafe.Add(ref selfRef, 56);
174+
this.V7L.Y = Unsafe.Add(ref selfRef, 57);
175+
this.V7L.Z = Unsafe.Add(ref selfRef, 58);
176+
this.V7L.W = Unsafe.Add(ref selfRef, 59);
177+
this.V7R.X = Unsafe.Add(ref selfRef, 60);
178+
this.V7R.Y = Unsafe.Add(ref selfRef, 61);
179+
this.V7R.Z = Unsafe.Add(ref selfRef, 62);
180+
this.V7R.W = Unsafe.Add(ref selfRef, 63);
152181
}
153182
}

‎src/ImageSharp/Formats/Jpeg/Components/Block8x8F.Generated.tt

Lines changed: 0 additions & 103 deletions
This file was deleted.

0 commit comments

Comments
 (0)