Skip to content

Jpeg encoder optimization via removing second transpose call from FDCT #1958

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
JimBobSquarePants merged 5 commits into from
Jan 30, 2022
Merged

Jpeg encoder optimization via removing second transpose call from FDCT #1958

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
477716c
Fused transpose/zigzag implementation
Jan 25, 2022
219119a
Tests fix
Jan 25, 2022
69458f4
gfoidl len check removal
Jan 25, 2022
b6400c2
Fixed compilation error
Jan 26, 2022
4ac5cc5
Merge branch 'master' into dp/jpeg-encoder-zigzag
Jan 26, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions src/ImageSharp/Formats/Jpeg/Components/Block8x8F.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -280,7 +280,7 @@ public void AddInPlace(float value)
}

/// <summary>
/// Quantize input block, apply zig-zag ordering and store result as 16bit integers.
/// Quantize input block, transpose, apply zig-zag ordering and store as <see cref="Block8x8"/>.
/// </summary>
/// <param name="block">Source block.</param>
/// <param name="dest">Destination block.</param>
Expand All @@ -291,19 +291,19 @@ public static void Quantize(ref Block8x8F block, ref Block8x8 dest, ref Block8x8
if (Avx2.IsSupported)
{
MultiplyIntoInt16_Avx2(ref block, ref qt, ref dest);
ZigZag.ApplyZigZagOrderingAvx2(ref dest);
ZigZag.ApplyTransposingZigZagOrderingAvx2(ref dest);
}
else if (Ssse3.IsSupported)
{
MultiplyIntoInt16_Sse2(ref block, ref qt, ref dest);
ZigZag.ApplyZigZagOrderingSsse3(ref dest);
ZigZag.ApplyTransposingZigZagOrderingSsse3(ref dest);
}
else
#endif
{
for (int i = 0; i < Size; i++)
{
int idx = ZigZag.ZigZagOrder[i];
int idx = ZigZag.TransposingOrder[i];
float quantizedVal = block[idx] * qt[idx];
quantizedVal += quantizedVal < 0 ? -0.5f : 0.5f;
dest[i] = (short)quantizedVal;
Expand Down
5 changes: 2 additions & 3 deletions src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.Intrinsic.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,11 +29,10 @@ private static void FDCT8x8_Avx(ref Block8x8F block)
{
DebugGuard.IsTrue(Avx.IsSupported, "Avx support is required to execute this operation.");

// First pass - process rows
block.TransposeInplace();
// First pass - process columns
FDCT8x8_1D_Avx(ref block);

// Second pass - process columns
// Second pass - process rows
block.TransposeInplace();
FDCT8x8_1D_Avx(ref block);

Expand Down
133 changes: 7 additions & 126 deletions src/ImageSharp/Formats/Jpeg/Components/FastFloatingPointDCT.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,11 @@ public static void AdjustToFDCT(ref Block8x8F quantTable)
tableRef = 0.125f / (tableRef * Unsafe.Add(ref multipliersRef, i));
tableRef = ref Unsafe.Add(ref tableRef, 1);
}

// Spectral macroblocks are not transposed before quantization
// Transpose is done after quantization at zig-zag stage
// so we must transpose quantization table
quantTable.TransposeInplace();
}

/// <summary>
Expand Down Expand Up @@ -133,14 +138,9 @@ public static void TransformFDCT(ref Block8x8F block)
}
else
#endif
if (Vector.IsHardwareAccelerated)
{
FDCT_Vector4(ref block);
}
else
{
FDCT_Scalar(ref block);
}
}

/// <summary>
Expand Down Expand Up @@ -217,136 +217,17 @@ static void IDCT8x4_Vector4(ref Vector4 vecRef)
}
}

/// <summary>
/// Apply 2D floating point FDCT inplace using scalar operations.
/// </summary>
/// <remarks>
/// Ported from libjpeg-turbo https://github.com/libjpeg-turbo/libjpeg-turbo/blob/main/jfdctflt.c.
/// </remarks>
/// <param name="block">Input block.</param>
private static void FDCT_Scalar(ref Block8x8F block)
{
const int dctSize = 8;

float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
float tmp10, tmp11, tmp12, tmp13;
float z1, z2, z3, z4, z5, z11, z13;

// First pass - process rows
ref float blockRef = ref Unsafe.As<Block8x8F, float>(ref block);
for (int ctr = 7; ctr >= 0; ctr--)
{
tmp0 = Unsafe.Add(ref blockRef, 0) + Unsafe.Add(ref blockRef, 7);
tmp7 = Unsafe.Add(ref blockRef, 0) - Unsafe.Add(ref blockRef, 7);
tmp1 = Unsafe.Add(ref blockRef, 1) + Unsafe.Add(ref blockRef, 6);
tmp6 = Unsafe.Add(ref blockRef, 1) - Unsafe.Add(ref blockRef, 6);
tmp2 = Unsafe.Add(ref blockRef, 2) + Unsafe.Add(ref blockRef, 5);
tmp5 = Unsafe.Add(ref blockRef, 2) - Unsafe.Add(ref blockRef, 5);
tmp3 = Unsafe.Add(ref blockRef, 3) + Unsafe.Add(ref blockRef, 4);
tmp4 = Unsafe.Add(ref blockRef, 3) - Unsafe.Add(ref blockRef, 4);

// Even part
tmp10 = tmp0 + tmp3;
tmp13 = tmp0 - tmp3;
tmp11 = tmp1 + tmp2;
tmp12 = tmp1 - tmp2;

Unsafe.Add(ref blockRef, 0) = tmp10 + tmp11;
Unsafe.Add(ref blockRef, 4) = tmp10 - tmp11;

z1 = (tmp12 + tmp13) * 0.707106781f;
Unsafe.Add(ref blockRef, 2) = tmp13 + z1;
Unsafe.Add(ref blockRef, 6) = tmp13 - z1;

// Odd part
tmp10 = tmp4 + tmp5;
tmp11 = tmp5 + tmp6;
tmp12 = tmp6 + tmp7;

z5 = (tmp10 - tmp12) * 0.382683433f;
z2 = (0.541196100f * tmp10) + z5;
z4 = (1.306562965f * tmp12) + z5;
z3 = tmp11 * 0.707106781f;

z11 = tmp7 + z3;
z13 = tmp7 - z3;

Unsafe.Add(ref blockRef, 5) = z13 + z2;
Unsafe.Add(ref blockRef, 3) = z13 - z2;
Unsafe.Add(ref blockRef, 1) = z11 + z4;
Unsafe.Add(ref blockRef, 7) = z11 - z4;

blockRef = ref Unsafe.Add(ref blockRef, dctSize);
}

// Second pass - process columns
blockRef = ref Unsafe.As<Block8x8F, float>(ref block);
for (int ctr = 7; ctr >= 0; ctr--)
{
tmp0 = Unsafe.Add(ref blockRef, dctSize * 0) + Unsafe.Add(ref blockRef, dctSize * 7);
tmp7 = Unsafe.Add(ref blockRef, dctSize * 0) - Unsafe.Add(ref blockRef, dctSize * 7);
tmp1 = Unsafe.Add(ref blockRef, dctSize * 1) + Unsafe.Add(ref blockRef, dctSize * 6);
tmp6 = Unsafe.Add(ref blockRef, dctSize * 1) - Unsafe.Add(ref blockRef, dctSize * 6);
tmp2 = Unsafe.Add(ref blockRef, dctSize * 2) + Unsafe.Add(ref blockRef, dctSize * 5);
tmp5 = Unsafe.Add(ref blockRef, dctSize * 2) - Unsafe.Add(ref blockRef, dctSize * 5);
tmp3 = Unsafe.Add(ref blockRef, dctSize * 3) + Unsafe.Add(ref blockRef, dctSize * 4);
tmp4 = Unsafe.Add(ref blockRef, dctSize * 3) - Unsafe.Add(ref blockRef, dctSize * 4);

// Even part
tmp10 = tmp0 + tmp3;
tmp13 = tmp0 - tmp3;
tmp11 = tmp1 + tmp2;
tmp12 = tmp1 - tmp2;

Unsafe.Add(ref blockRef, dctSize * 0) = tmp10 + tmp11;
Unsafe.Add(ref blockRef, dctSize * 4) = tmp10 - tmp11;

z1 = (tmp12 + tmp13) * 0.707106781f;
Unsafe.Add(ref blockRef, dctSize * 2) = tmp13 + z1;
Unsafe.Add(ref blockRef, dctSize * 6) = tmp13 - z1;

// Odd part
tmp10 = tmp4 + tmp5;
tmp11 = tmp5 + tmp6;
tmp12 = tmp6 + tmp7;

z5 = (tmp10 - tmp12) * 0.382683433f;
z2 = (0.541196100f * tmp10) + z5;
z4 = (1.306562965f * tmp12) + z5;
z3 = tmp11 * 0.707106781f;

z11 = tmp7 + z3;
z13 = tmp7 - z3;

Unsafe.Add(ref blockRef, dctSize * 5) = z13 + z2;
Unsafe.Add(ref blockRef, dctSize * 3) = z13 - z2;
Unsafe.Add(ref blockRef, dctSize * 1) = z11 + z4;
Unsafe.Add(ref blockRef, dctSize * 7) = z11 - z4;

blockRef = ref Unsafe.Add(ref blockRef, 1);
}
}

/// <summary>
/// Apply floating point FDCT inplace using <see cref="Vector4"/> API.
/// </summary>
/// <remarks>
/// This implementation must be called only if hardware supports 4
/// floating point numbers vector. Otherwise explicit scalar
/// implementation <see cref="FDCT_Scalar"/> is faster
/// because it does not rely on block transposition.
/// </remarks>
/// <param name="block">Input block.</param>
public static void FDCT_Vector4(ref Block8x8F block)
{
DebugGuard.IsTrue(Vector.IsHardwareAccelerated, "Scalar implementation should be called for non-accelerated hardware.");

// First pass - process rows
block.TransposeInplace();
// First pass - process columns
FDCT8x4_Vector4(ref block.V0L);
FDCT8x4_Vector4(ref block.V0R);

// Second pass - process columns
// Second pass - process rows
block.TransposeInplace();
FDCT8x4_Vector4(ref block.V0L);
FDCT8x4_Vector4(ref block.V0R);
Expand Down
Loading