Skip to content

Commit

Permalink
Refactor: extract horizontal and vertical pass into methods
Browse files Browse the repository at this point in the history
  • Loading branch information
@brianpopow
brianpopow committed Nov 12, 2021
1 parent 544319e commit 5074ee6
Showing 1 changed file with 63 additions and 98 deletions.
161 changes: 63 additions & 98 deletions src/ImageSharp/Formats/Webp/Lossy/Vp8Encoding.cs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -136,61 +136,14 @@ public static void ITransform(Span<byte> reference, Span<short> input, Span<byte

// Vertical pass and subsequent transpose.
// First pass, c and d calculations are longer because of the "trick" multiplications.
Vector128<short> a = Sse2.Add(in0.AsInt16(), in2.AsInt16());
Vector128<short> b = Sse2.Subtract(in0.AsInt16(), in2.AsInt16());

// c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
Vector128<short> c1 = Sse2.MultiplyHigh(in1.AsInt16(), K2);
Vector128<short> c2 = Sse2.MultiplyHigh(in3.AsInt16(), K1);
Vector128<short> c3 = Sse2.Subtract(in1.AsInt16(), in3.AsInt16());
Vector128<short> c4 = Sse2.Subtract(c1, c2);
Vector128<short> c = Sse2.Add(c3, c4);

// d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
Vector128<short> d1 = Sse2.MultiplyHigh(in1.AsInt16(), K1);
Vector128<short> d2 = Sse2.MultiplyHigh(in3.AsInt16(), K2);
Vector128<short> d3 = Sse2.Add(in1.AsInt16(), in3.AsInt16());
Vector128<short> d4 = Sse2.Add(d1, d2);
Vector128<short> d = Sse2.Add(d3, d4);

// Second pass.
Vector128<short> tmp0 = Sse2.Add(a, d);
Vector128<short> tmp1 = Sse2.Add(b, c);
Vector128<short> tmp2 = Sse2.Subtract(b, c);
Vector128<short> tmp3 = Sse2.Subtract(a, d);
InverseTransformVerticalPass(in0, in2, in1, in3, out Vector128<short> tmp0, out Vector128<short> tmp1, out Vector128<short> tmp2, out Vector128<short> tmp3);

// Transpose the two 4x4.
LossyUtils.Vp8Transpose_2_4x4_16b(tmp0, tmp1, tmp2, tmp3, out Vector128<long> t0, out Vector128<long> t1, out Vector128<long> t2, out Vector128<long> t3);

// Horizontal pass and subsequent transpose.
// First pass, c and d calculations are longer because of the "trick" multiplications.
Vector128<short> dc = Sse2.Add(t0.AsInt16(), Four);
a = Sse2.Add(dc, t2.AsInt16());
b = Sse2.Subtract(dc, t2.AsInt16());

// c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
c1 = Sse2.MultiplyHigh(t1.AsInt16(), K2);
c2 = Sse2.MultiplyHigh(t3.AsInt16(), K1);
c3 = Sse2.Subtract(t1.AsInt16(), t3.AsInt16());
c4 = Sse2.Subtract(c1, c2);
c = Sse2.Add(c3, c4);

// d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
d1 = Sse2.MultiplyHigh(t1.AsInt16(), K1);
d2 = Sse2.MultiplyHigh(t3.AsInt16(), K2);
d3 = Sse2.Add(t1.AsInt16(), t3.AsInt16());
d4 = Sse2.Add(d1, d2);
d = Sse2.Add(d3, d4);

// Second pass.
tmp0 = Sse2.Add(a, d);
tmp1 = Sse2.Add(b, c);
tmp2 = Sse2.Subtract(b, c);
tmp3 = Sse2.Subtract(a, d);
Vector128<short> shifted0 = Sse2.ShiftRightArithmetic(tmp0, 3);
Vector128<short> shifted1 = Sse2.ShiftRightArithmetic(tmp1, 3);
Vector128<short> shifted2 = Sse2.ShiftRightArithmetic(tmp2, 3);
Vector128<short> shifted3 = Sse2.ShiftRightArithmetic(tmp3, 3);
InverseTransformHorizontalPass(t0, t2, t1, t3, out Vector128<short> shifted0, out Vector128<short> shifted1, out Vector128<short> shifted2, out Vector128<short> shifted3);

// Transpose the two 4x4.
LossyUtils.Vp8Transpose_2_4x4_16b(shifted0, shifted1, shifted2, shifted3, out t0, out t1, out t2, out t3);
Expand Down Expand Up @@ -266,61 +219,14 @@ public static void ITransformOne(Span<byte> reference, Span<short> input, Span<b

// Vertical pass and subsequent transpose.
// First pass, c and d calculations are longer because of the "trick" multiplications.
Vector128<short> a = Sse2.Add(in0.AsInt16(), in2.AsInt16());
Vector128<short> b = Sse2.Subtract(in0.AsInt16(), in2.AsInt16());

// c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
Vector128<short> c1 = Sse2.MultiplyHigh(in1.AsInt16(), K2);
Vector128<short> c2 = Sse2.MultiplyHigh(in3.AsInt16(), K1);
Vector128<short> c3 = Sse2.Subtract(in1.AsInt16(), in3.AsInt16());
Vector128<short> c4 = Sse2.Subtract(c1, c2);
Vector128<short> c = Sse2.Add(c3, c4);

// d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
Vector128<short> d1 = Sse2.MultiplyHigh(in1.AsInt16(), K1);
Vector128<short> d2 = Sse2.MultiplyHigh(in3.AsInt16(), K2);
Vector128<short> d3 = Sse2.Add(in1.AsInt16(), in3.AsInt16());
Vector128<short> d4 = Sse2.Add(d1, d2);
Vector128<short> d = Sse2.Add(d3, d4);

// Second pass.
Vector128<short> tmp0 = Sse2.Add(a, d);
Vector128<short> tmp1 = Sse2.Add(b, c);
Vector128<short> tmp2 = Sse2.Subtract(b, c);
Vector128<short> tmp3 = Sse2.Subtract(a, d);
InverseTransformVerticalPass(in0, in2, in1, in3, out Vector128<short> tmp0, out Vector128<short> tmp1, out Vector128<short> tmp2, out Vector128<short> tmp3);

// Transpose the two 4x4.
LossyUtils.Vp8Transpose_2_4x4_16b(tmp0, tmp1, tmp2, tmp3, out Vector128<long> t0, out Vector128<long> t1, out Vector128<long> t2, out Vector128<long> t3);

// Horizontal pass and subsequent transpose.
// First pass, c and d calculations are longer because of the "trick" multiplications.
Vector128<short> dc = Sse2.Add(t0.AsInt16(), Four);
a = Sse2.Add(dc, t2.AsInt16());
b = Sse2.Subtract(dc, t2.AsInt16());

// c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
c1 = Sse2.MultiplyHigh(t1.AsInt16(), K2);
c2 = Sse2.MultiplyHigh(t3.AsInt16(), K1);
c3 = Sse2.Subtract(t1.AsInt16(), t3.AsInt16());
c4 = Sse2.Subtract(c1, c2);
c = Sse2.Add(c3, c4);

// d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
d1 = Sse2.MultiplyHigh(t1.AsInt16(), K1);
d2 = Sse2.MultiplyHigh(t3.AsInt16(), K2);
d3 = Sse2.Add(t1.AsInt16(), t3.AsInt16());
d4 = Sse2.Add(d1, d2);
d = Sse2.Add(d3, d4);

// Second pass.
tmp0 = Sse2.Add(a, d);
tmp1 = Sse2.Add(b, c);
tmp2 = Sse2.Subtract(b, c);
tmp3 = Sse2.Subtract(a, d);
Vector128<short> shifted0 = Sse2.ShiftRightArithmetic(tmp0, 3);
Vector128<short> shifted1 = Sse2.ShiftRightArithmetic(tmp1, 3);
Vector128<short> shifted2 = Sse2.ShiftRightArithmetic(tmp2, 3);
Vector128<short> shifted3 = Sse2.ShiftRightArithmetic(tmp3, 3);
InverseTransformHorizontalPass(t0, t2, t1, t3, out Vector128<short> shifted0, out Vector128<short> shifted1, out Vector128<short> shifted2, out Vector128<short> shifted3);

// Transpose the two 4x4.
LossyUtils.Vp8Transpose_2_4x4_16b(shifted0, shifted1, shifted2, shifted3, out t0, out t1, out t2, out t3);
Expand Down Expand Up @@ -409,6 +315,65 @@ public static void ITransformOne(Span<byte> reference, Span<short> input, Span<b
}
}

#if SUPPORTS_RUNTIME_INTRINSICS
private static void InverseTransformVerticalPass(Vector128<long> in0, Vector128<long> in2, Vector128<long> in1, Vector128<long> in3, out Vector128<short> tmp0, out Vector128<short> tmp1, out Vector128<short> tmp2, out Vector128<short> tmp3)
{
Vector128<short> a = Sse2.Add(in0.AsInt16(), in2.AsInt16());
Vector128<short> b = Sse2.Subtract(in0.AsInt16(), in2.AsInt16());

// c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
Vector128<short> c1 = Sse2.MultiplyHigh(in1.AsInt16(), K2);
Vector128<short> c2 = Sse2.MultiplyHigh(in3.AsInt16(), K1);
Vector128<short> c3 = Sse2.Subtract(in1.AsInt16(), in3.AsInt16());
Vector128<short> c4 = Sse2.Subtract(c1, c2);
Vector128<short> c = Sse2.Add(c3, c4);

// d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
Vector128<short> d1 = Sse2.MultiplyHigh(in1.AsInt16(), K1);
Vector128<short> d2 = Sse2.MultiplyHigh(in3.AsInt16(), K2);
Vector128<short> d3 = Sse2.Add(in1.AsInt16(), in3.AsInt16());
Vector128<short> d4 = Sse2.Add(d1, d2);
Vector128<short> d = Sse2.Add(d3, d4);

// Second pass.
tmp0 = Sse2.Add(a, d);
tmp1 = Sse2.Add(b, c);
tmp2 = Sse2.Subtract(b, c);
tmp3 = Sse2.Subtract(a, d);
}

private static void InverseTransformHorizontalPass(Vector128<long> t0, Vector128<long> t2, Vector128<long> t1, Vector128<long> t3, out Vector128<short> shifted0, out Vector128<short> shifted1, out Vector128<short> shifted2, out Vector128<short> shifted3)
{
Vector128<short> dc = Sse2.Add(t0.AsInt16(), Four);
Vector128<short> a = Sse2.Add(dc, t2.AsInt16());
Vector128<short> b = Sse2.Subtract(dc, t2.AsInt16());

// c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
Vector128<short> c1 = Sse2.MultiplyHigh(t1.AsInt16(), K2);
Vector128<short> c2 = Sse2.MultiplyHigh(t3.AsInt16(), K1);
Vector128<short> c3 = Sse2.Subtract(t1.AsInt16(), t3.AsInt16());
Vector128<short> c4 = Sse2.Subtract(c1, c2);
Vector128<short> c = Sse2.Add(c3, c4);

// d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
Vector128<short> d1 = Sse2.MultiplyHigh(t1.AsInt16(), K1);
Vector128<short> d2 = Sse2.MultiplyHigh(t3.AsInt16(), K2);
Vector128<short> d3 = Sse2.Add(t1.AsInt16(), t3.AsInt16());
Vector128<short> d4 = Sse2.Add(d1, d2);
Vector128<short> d = Sse2.Add(d3, d4);

// Second pass.
Vector128<short> tmp0 = Sse2.Add(a, d);
Vector128<short> tmp1 = Sse2.Add(b, c);
Vector128<short> tmp2 = Sse2.Subtract(b, c);
Vector128<short> tmp3 = Sse2.Subtract(a, d);
shifted0 = Sse2.ShiftRightArithmetic(tmp0, 3);
shifted1 = Sse2.ShiftRightArithmetic(tmp1, 3);
shifted2 = Sse2.ShiftRightArithmetic(tmp2, 3);
shifted3 = Sse2.ShiftRightArithmetic(tmp3, 3);
}
#endif

public static void FTransform2(Span<byte> src, Span<byte> reference, Span<short> output, Span<short> output2, Span<int> scratch)
{
FTransform(src, reference, output, scratch);
Expand Down

0 comments on commit 5074ee6

Please sign in to comment.