|
|
@ -22,52 +22,42 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters |
|
|
{ |
|
|
{ |
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
protected override void ConvertCoreVectorized(in ComponentValues values, Span<Vector4> result) |
|
|
protected override void ConvertCoreVectorizedInplace(in ComponentValues values) |
|
|
{ |
|
|
{ |
|
|
#if SUPPORTS_RUNTIME_INTRINSICS
|
|
|
#if SUPPORTS_RUNTIME_INTRINSICS
|
|
|
ref Vector256<float> yBase = |
|
|
ref Vector256<float> c0Base = |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0)); |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component0)); |
|
|
ref Vector256<float> cbBase = |
|
|
ref Vector256<float> c1Base = |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1)); |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component1)); |
|
|
ref Vector256<float> crBase = |
|
|
ref Vector256<float> c2Base = |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2)); |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component2)); |
|
|
ref Vector256<float> kBase = |
|
|
ref Vector256<float> kBase = |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3)); |
|
|
ref Unsafe.As<float, Vector256<float>>(ref MemoryMarshal.GetReference(values.Component3)); |
|
|
|
|
|
|
|
|
ref Vector256<float> resultBase = |
|
|
|
|
|
ref Unsafe.As<Vector4, Vector256<float>>(ref MemoryMarshal.GetReference(result)); |
|
|
|
|
|
|
|
|
|
|
|
// Used for the color conversion
|
|
|
// Used for the color conversion
|
|
|
var chromaOffset = Vector256.Create(-this.HalfValue); |
|
|
var chromaOffset = Vector256.Create(-this.HalfValue); |
|
|
var scale = Vector256.Create(1 / this.MaximumValue); |
|
|
var scale = Vector256.Create(1 / (this.MaximumValue * this.MaximumValue)); |
|
|
var max = Vector256.Create(this.MaximumValue); |
|
|
var max = Vector256.Create(this.MaximumValue); |
|
|
var rCrMult = Vector256.Create(1.402F); |
|
|
var rCrMult = Vector256.Create(1.402F); |
|
|
var gCbMult = Vector256.Create(-0.344136F); |
|
|
var gCbMult = Vector256.Create(-0.344136F); |
|
|
var gCrMult = Vector256.Create(-0.714136F); |
|
|
var gCrMult = Vector256.Create(-0.714136F); |
|
|
var bCbMult = Vector256.Create(1.772F); |
|
|
var bCbMult = Vector256.Create(1.772F); |
|
|
|
|
|
|
|
|
// Used for packing.
|
|
|
|
|
|
var va = Vector256.Create(1F); |
|
|
|
|
|
ref byte control = ref MemoryMarshal.GetReference(HwIntrinsics.PermuteMaskEvenOdd8x32); |
|
|
|
|
|
Vector256<int> vcontrol = Unsafe.As<byte, Vector256<int>>(ref control); |
|
|
|
|
|
|
|
|
|
|
|
// Walking 8 elements at one step:
|
|
|
// Walking 8 elements at one step:
|
|
|
int n = result.Length / 8; |
|
|
nint n = values.Component0.Length / 8; |
|
|
for (int i = 0; i < n; i++) |
|
|
for (nint i = 0; i < n; i++) |
|
|
{ |
|
|
{ |
|
|
// y = yVals[i];
|
|
|
// y = yVals[i];
|
|
|
// cb = cbVals[i] - 128F;
|
|
|
// cb = cbVals[i] - 128F;
|
|
|
// cr = crVals[i] - 128F;
|
|
|
// cr = crVals[i] - 128F;
|
|
|
// k = kVals[i] / 256F;
|
|
|
// k = kVals[i] / 256F;
|
|
|
Vector256<float> y = Unsafe.Add(ref yBase, i); |
|
|
ref Vector256<float> c0 = ref Unsafe.Add(ref c0Base, i); |
|
|
Vector256<float> cb = Avx.Add(Unsafe.Add(ref cbBase, i), chromaOffset); |
|
|
ref Vector256<float> c1 = ref Unsafe.Add(ref c1Base, i); |
|
|
Vector256<float> cr = Avx.Add(Unsafe.Add(ref crBase, i), chromaOffset); |
|
|
ref Vector256<float> c2 = ref Unsafe.Add(ref c2Base, i); |
|
|
Vector256<float> k = Avx.Divide(Unsafe.Add(ref kBase, i), max); |
|
|
Vector256<float> y = c0; |
|
|
|
|
|
Vector256<float> cb = Avx.Add(c1, chromaOffset); |
|
|
y = Avx2.PermuteVar8x32(y, vcontrol); |
|
|
Vector256<float> cr = Avx.Add(c2, chromaOffset); |
|
|
cb = Avx2.PermuteVar8x32(cb, vcontrol); |
|
|
Vector256<float> scaledK = Avx.Multiply(Unsafe.Add(ref kBase, i), scale); |
|
|
cr = Avx2.PermuteVar8x32(cr, vcontrol); |
|
|
|
|
|
k = Avx2.PermuteVar8x32(k, vcontrol); |
|
|
|
|
|
|
|
|
|
|
|
// r = y + (1.402F * cr);
|
|
|
// r = y + (1.402F * cr);
|
|
|
// g = y - (0.344136F * cb) - (0.714136F * cr);
|
|
|
// g = y - (0.344136F * cb) - (0.714136F * cr);
|
|
|
@ -82,29 +72,19 @@ namespace SixLabors.ImageSharp.Formats.Jpeg.Components.Decoder.ColorConverters |
|
|
g = Avx.Subtract(max, Avx.RoundToNearestInteger(g)); |
|
|
g = Avx.Subtract(max, Avx.RoundToNearestInteger(g)); |
|
|
b = Avx.Subtract(max, Avx.RoundToNearestInteger(b)); |
|
|
b = Avx.Subtract(max, Avx.RoundToNearestInteger(b)); |
|
|
|
|
|
|
|
|
r = Avx.Multiply(Avx.Multiply(r, k), scale); |
|
|
r = Avx.Multiply(r, scaledK); |
|
|
g = Avx.Multiply(Avx.Multiply(g, k), scale); |
|
|
g = Avx.Multiply(g, scaledK); |
|
|
b = Avx.Multiply(Avx.Multiply(b, k), scale); |
|
|
b = Avx.Multiply(b, scaledK); |
|
|
|
|
|
|
|
|
Vector256<float> vte = Avx.UnpackLow(r, b); |
|
|
|
|
|
Vector256<float> vto = Avx.UnpackLow(g, va); |
|
|
|
|
|
|
|
|
|
|
|
ref Vector256<float> destination = ref Unsafe.Add(ref resultBase, i * 4); |
|
|
|
|
|
|
|
|
|
|
|
destination = Avx.UnpackLow(vte, vto); |
|
|
|
|
|
Unsafe.Add(ref destination, 1) = Avx.UnpackHigh(vte, vto); |
|
|
|
|
|
|
|
|
|
|
|
vte = Avx.UnpackHigh(r, b); |
|
|
|
|
|
vto = Avx.UnpackHigh(g, va); |
|
|
|
|
|
|
|
|
|
|
|
Unsafe.Add(ref destination, 2) = Avx.UnpackLow(vte, vto); |
|
|
c0 = r; |
|
|
Unsafe.Add(ref destination, 3) = Avx.UnpackHigh(vte, vto); |
|
|
c1 = g; |
|
|
|
|
|
c2 = b; |
|
|
} |
|
|
} |
|
|
#endif
|
|
|
#endif
|
|
|
} |
|
|
} |
|
|
|
|
|
|
|
|
protected override void ConvertCore(in ComponentValues values, Span<Vector4> result) => |
|
|
protected override void ConvertCoreInplace(in ComponentValues values) => |
|
|
FromYccKBasic.ConvertCore(values, result, this.MaximumValue, this.HalfValue); |
|
|
FromYccKBasic.ConvertCoreInplace(values, this.MaximumValue, this.HalfValue); |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
|
} |
|
|
|
Ildar Khayrutdinov
5 years ago