internal static void fDCT2D_llm(
MutableSpan <float> s,
MutableSpan <float> d,
MutableSpan <float> temp,
bool downscaleBy8 = false,
bool offsetSourceByNeg128 = false)
{
MutableSpan <float> sWorker = offsetSourceByNeg128 ? s.AddScalarToAllValues(-128f) : s;
for (int j = 0; j < 8; j++)
{
fDCT1Dllm_32f(sWorker.Slice(j * 8), temp.Slice(j * 8));
}
Transpose8x8(temp, d);
for (int j = 0; j < 8; j++)
{
fDCT1Dllm_32f(d.Slice(j * 8), temp.Slice(j * 8));
}
Transpose8x8(temp, d);
if (downscaleBy8)
{
for (int j = 0; j < 64; j++)
{
d[j] *= 0.125f;
}
}
}
public static void fDCT8x8_llm_sse(MutableSpan <float> s, MutableSpan <float> d, MutableSpan <float> temp)
{
Transpose8x8(s, temp);
fDCT2D8x4_32f(temp, d);
fDCT2D8x4_32f(temp.Slice(4), d.Slice(4));
Transpose8x8(d, temp);
fDCT2D8x4_32f(temp, d);
fDCT2D8x4_32f(temp.Slice(4), d.Slice(4));
Vector4 c = new Vector4(0.1250f);
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //0
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //1
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //2
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //3
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //4
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //5
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //6
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //7
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //8
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //9
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //10
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //11
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //12
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //13
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //14
_mm_store_ps(d, 0, (_mm_load_ps(d, 0) * c)); d.AddOffset(4); //15
}
private static void _mm_store_ps(MutableSpan <float> dest, int offset, Vector4 src)
{
dest = dest.Slice(offset);
dest[0] = src.X;
dest[1] = src.Y;
dest[2] = src.Z;
dest[3] = src.W;
}
/// <summary>
/// Original: https://github.com/norishigefukushima/dct_simd/blob/master/dct/dct8x8_simd.cpp#L239
/// Applyies IDCT transformation on "s" copying transformed values to "d", using temporal block "temp"
/// </summary>
/// <param name="s"></param>
/// <param name="d"></param>
/// <param name="temp"></param>
internal static void iDCT2D_llm(MutableSpan <float> s, MutableSpan <float> d, MutableSpan <float> temp)
{
int j;
for (j = 0; j < 8; j++)
{
iDCT1Dllm_32f(s.Slice(j * 8), temp.Slice(j * 8));
}
Transpose8x8(temp, d);
for (j = 0; j < 8; j++)
{
iDCT1Dllm_32f(d.Slice(j * 8), temp.Slice(j * 8));
}
Transpose8x8(temp, d);
for (j = 0; j < 64; j++)
{
d[j] *= 0.125f;
}
}
public void FDCT8x4_RightPart(int seed)
{
var src = Create8x8RandomFloatData(-200, 200, seed);
var srcBlock = new Block8x8F();
srcBlock.LoadFrom(src);
var destBlock = new Block8x8F();
var expectedDest = new MutableSpan <float>(64);
ReferenceImplementations.fDCT2D8x4_32f(src.Slice(4), expectedDest.Slice(4));
DCT.FDCT8x4_RightPart(ref srcBlock, ref destBlock);
var actualDest = new MutableSpan <float>(64);
destBlock.CopyTo(actualDest);
Assert.Equal(actualDest.Data, expectedDest.Data, new ApproximateFloatComparer(1f));
}
public void iDCT2D8x4_RightPart()
{
MutableSpan <float> sourceArray = Create8x8FloatData();
MutableSpan <float> expectedDestArray = new float[64];
ReferenceImplementations.iDCT2D8x4_32f(sourceArray.Slice(4), expectedDestArray.Slice(4));
Block8x8F source = new Block8x8F();
source.LoadFrom(sourceArray);
Block8x8F dest = new Block8x8F();
DCT.IDCT8x4_RightPart(ref source, ref dest);
float[] actualDestArray = new float[64];
dest.CopyTo(actualDestArray);
this.Print8x8Data(expectedDestArray);
this.Output.WriteLine("**************");
this.Print8x8Data(actualDestArray);
Assert.Equal(expectedDestArray.Data, actualDestArray);
}
private static Vector4 _mm_load_ps(MutableSpan <float> src, int offset)
{
src = src.Slice(offset);
return(new Vector4(src[0], src[1], src[2], src[3]));
}