diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Argb32.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Argb32.PixelOperations.Generated.cs index e6b8922e9..6449351cc 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Argb32.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Argb32.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; @@ -45,25 +46,25 @@ namespace SixLabors.ImageSharp.PixelFormats /// internal override void FromVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, false); } /// internal override void ToVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, false); } /// internal override void FromScaledVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, true); } /// internal override void ToScaledVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, true); } /// diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgr24.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgr24.PixelOperations.Generated.cs index 9b1740013..9232cf454 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgr24.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgr24.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; @@ -45,25 +46,25 @@ namespace SixLabors.ImageSharp.PixelFormats /// internal override void FromVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, false); } /// internal override void ToVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, false); } /// internal override void FromScaledVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, true); } /// internal override void ToScaledVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, true); } diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgra32.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgra32.PixelOperations.Generated.cs index 37d6b72d7..4f56b75c5 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgra32.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Bgra32.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; @@ -45,25 +46,25 @@ namespace SixLabors.ImageSharp.PixelFormats /// internal override void FromVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, false); } /// internal override void ToVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, false); } /// internal override void FromScaledVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, true); } /// internal override void ToScaledVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, true); } /// diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray16.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray16.PixelOperations.Generated.cs index 638db1d0d..81882185d 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray16.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray16.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray8.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray8.PixelOperations.Generated.cs index 6bf0693c2..f6678a4f8 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray8.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Gray8.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb24.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb24.PixelOperations.Generated.cs index 6ff87eb38..aae8b2f63 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb24.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb24.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; @@ -45,25 +46,25 @@ namespace SixLabors.ImageSharp.PixelFormats /// internal override void FromVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, false); } /// internal override void ToVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, false); } /// internal override void FromScaledVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, true); } /// internal override void ToScaledVector4(Configuration configuration, ReadOnlySpan sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, true); } diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb48.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb48.PixelOperations.Generated.cs index 9b4584d76..c828053a4 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb48.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgb48.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba32.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba32.PixelOperations.Generated.cs index 8e15ca1f4..9c29bd044 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba32.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba32.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba64.PixelOperations.Generated.cs b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba64.PixelOperations.Generated.cs index caaba7809..db9cb84be 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba64.PixelOperations.Generated.cs +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/Rgba64.PixelOperations.Generated.cs @@ -3,6 +3,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; diff --git a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/_Common.ttinclude b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/_Common.ttinclude index 4501f4972..cc8cb0e2f 100644 --- a/src/ImageSharp/PixelFormats/PixelImplementations/Generated/_Common.ttinclude +++ b/src/ImageSharp/PixelFormats/PixelImplementations/Generated/_Common.ttinclude @@ -8,6 +8,7 @@ // +using SixLabors.ImageSharp.PixelFormats.Utils; using System; using System.Buffers; using System.Numerics; @@ -114,25 +115,25 @@ using System.Runtime.InteropServices; /// internal override void FromVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span<<#=pixelType#>> destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, false); } /// internal override void ToVector4(Configuration configuration, ReadOnlySpan<<#=pixelType#>> sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, false); } /// internal override void FromScaledVector4(Configuration configuration, ReadOnlySpan sourceVectors, Span<<#=pixelType#>> destPixels) { - this.RunRgba32CompatibleFromVector4Conversion(configuration, sourceVectors, destPixels); + Vector4Converters.RgbaCompatible.FromVector4(configuration, this, sourceVectors, destPixels, true); } /// internal override void ToScaledVector4(Configuration configuration, ReadOnlySpan<<#=pixelType#>> sourcePixels, Span destVectors) { - this.RunRgba32CompatibleToVector4Conversion(configuration, sourcePixels, destVectors); + Vector4Converters.RgbaCompatible.ToVector4(configuration, this, sourcePixels, destVectors, true); } <#+ diff --git a/src/ImageSharp/PixelFormats/PixelOperations{TPixel}.cs b/src/ImageSharp/PixelFormats/PixelOperations{TPixel}.cs index 9238e1b47..f4eb19be3 100644 --- a/src/ImageSharp/PixelFormats/PixelOperations{TPixel}.cs +++ b/src/ImageSharp/PixelFormats/PixelOperations{TPixel}.cs @@ -17,12 +17,6 @@ namespace SixLabors.ImageSharp.PixelFormats public partial class PixelOperations where TPixel : struct, IPixel { - /// - /// It's not worth to bother the transitive pixel conversion method below this limit. - /// The value depends on the actual gain brought by the SIMD characteristics of the executing CPU and JIT. - /// - private static readonly int Vector4ConversionThreshold = CalculateVector4ConversionThreshold(); - /// /// Gets the global instance for the pixel type /// @@ -42,7 +36,7 @@ namespace SixLabors.ImageSharp.PixelFormats Guard.NotNull(configuration, nameof(configuration)); Guard.DestinationShouldNotBeTooShort(sourceVectors, destPixels, nameof(destPixels)); - FromVector4DefaultImpl(sourceVectors, destPixels); + Utils.Vector4Converters.Default.DangerousFromVector4(sourceVectors, destPixels); } /// @@ -59,7 +53,7 @@ namespace SixLabors.ImageSharp.PixelFormats Guard.NotNull(configuration, nameof(configuration)); Guard.DestinationShouldNotBeTooShort(sourcePixels, destVectors, nameof(destVectors)); - ToVector4DefaultImpl(sourcePixels, destVectors); + Utils.Vector4Converters.Default.DangerousToVector4(sourcePixels, destVectors); } /// @@ -76,15 +70,7 @@ namespace SixLabors.ImageSharp.PixelFormats Guard.NotNull(configuration, nameof(configuration)); Guard.DestinationShouldNotBeTooShort(sourceVectors, destinationColors, nameof(destinationColors)); - ref Vector4 sourceRef = ref MemoryMarshal.GetReference(sourceVectors); - ref TPixel destRef = ref MemoryMarshal.GetReference(destinationColors); - - for (int i = 0; i < sourceVectors.Length; i++) - { - ref Vector4 sp = ref Unsafe.Add(ref sourceRef, i); - ref TPixel dp = ref Unsafe.Add(ref destRef, i); - dp.FromScaledVector4(sp); - } + Utils.Vector4Converters.Default.DangerousFromScaledVector4(sourceVectors, destinationColors); } /// @@ -101,15 +87,7 @@ namespace SixLabors.ImageSharp.PixelFormats Guard.NotNull(configuration, nameof(configuration)); Guard.DestinationShouldNotBeTooShort(sourceColors, destinationVectors, nameof(destinationVectors)); - ref TPixel sourceRef = ref MemoryMarshal.GetReference(sourceColors); - ref Vector4 destRef = ref MemoryMarshal.GetReference(destinationVectors); - - for (int i = 0; i < sourceColors.Length; i++) - { - ref TPixel sp = ref Unsafe.Add(ref sourceRef, i); - ref Vector4 dp = ref Unsafe.Add(ref destRef, i); - dp = sp.ToScaledVector4(); - } + Utils.Vector4Converters.Default.DangerousToScaledVector4(sourceColors, destinationVectors); } /// @@ -170,122 +148,5 @@ namespace SixLabors.ImageSharp.PixelFormats dp.FromScaledVector4(sp.ToScaledVector4()); } } - - // TODO: The Vector4 helpers should be moved to a utility class. - - /// - /// Provides an efficient default implementation for and - /// which is applicable for -compatible pixel types where - /// returns the same scaled result as . - /// The method is works by internally converting to a therefore it's not applicable for that type! - /// - [MethodImpl(InliningOptions.ShortMethod)] - internal void RunRgba32CompatibleToVector4Conversion( - Configuration configuration, - ReadOnlySpan sourcePixels, - Span destVectors) - { - Guard.NotNull(configuration, nameof(configuration)); - Guard.DestinationShouldNotBeTooShort(sourcePixels, destVectors, nameof(destVectors)); - - int count = sourcePixels.Length; - - // Not worth for small buffers: - if (count < Vector4ConversionThreshold) - { - ToVector4DefaultImpl(sourcePixels, destVectors); - return; - } - - // Using the last quarter of 'destVectors' as a temporary buffer to avoid allocation: - int countWithoutLastItem = count - 1; - ReadOnlySpan reducedSource = sourcePixels.Slice(0, countWithoutLastItem); - Span lastQuarterOfDestBuffer = MemoryMarshal.Cast(destVectors).Slice((3 * count) + 1, countWithoutLastItem); - this.ToRgba32(configuration, reducedSource, lastQuarterOfDestBuffer); - - // 'destVectors' and 'lastQuarterOfDestBuffer' are ovelapping buffers, - // but we are always reading/writing at different positions: - SimdUtils.BulkConvertByteToNormalizedFloat( - MemoryMarshal.Cast(lastQuarterOfDestBuffer), - MemoryMarshal.Cast(destVectors.Slice(0, countWithoutLastItem))); - - destVectors[countWithoutLastItem] = sourcePixels[countWithoutLastItem].ToVector4(); - } - - /// - /// Provides an efficient default implementation for and - /// which is applicable for -compatible pixel types where - /// returns the same scaled result as . - /// The method is works by internally converting to a therefore it's not applicable for that type! - /// - [MethodImpl(InliningOptions.ShortMethod)] - internal void RunRgba32CompatibleFromVector4Conversion( - Configuration configuration, - ReadOnlySpan sourceVectors, - Span destPixels) - { - Guard.NotNull(configuration, nameof(configuration)); - Guard.DestinationShouldNotBeTooShort(sourceVectors, destPixels, nameof(destPixels)); - - int count = sourceVectors.Length; - - // Not worth for small buffers: - if (count < Vector4ConversionThreshold) - { - FromVector4DefaultImpl(sourceVectors, destPixels); - return; - } - - // For the opposite direction it's not easy to implement the trick used in RunRgba32CompatibleToVector4Conversion, - // so let's allocate a temporary buffer as usually: - using (IMemoryOwner tempBuffer = configuration.MemoryAllocator.Allocate(count)) - { - Span tempSpan = tempBuffer.Memory.Span; - - SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows( - MemoryMarshal.Cast(sourceVectors), - MemoryMarshal.Cast(tempSpan)); - - this.FromRgba32(configuration, tempSpan, destPixels); - } - } - - [MethodImpl(InliningOptions.ShortMethod)] - private static void FromVector4DefaultImpl(ReadOnlySpan sourceVectors, Span destPixels) - { - ref Vector4 sourceRef = ref MemoryMarshal.GetReference(sourceVectors); - ref TPixel destRef = ref MemoryMarshal.GetReference(destPixels); - - for (int i = 0; i < sourceVectors.Length; i++) - { - ref Vector4 sp = ref Unsafe.Add(ref sourceRef, i); - ref TPixel dp = ref Unsafe.Add(ref destRef, i); - dp.FromVector4(sp); - } - } - - [MethodImpl(InliningOptions.ShortMethod)] - private static void ToVector4DefaultImpl(ReadOnlySpan sourcePixels, Span destVectors) - { - ref TPixel sourceRef = ref MemoryMarshal.GetReference(sourcePixels); - ref Vector4 destRef = ref MemoryMarshal.GetReference(destVectors); - - for (int i = 0; i < sourcePixels.Length; i++) - { - ref TPixel sp = ref Unsafe.Add(ref sourceRef, i); - ref Vector4 dp = ref Unsafe.Add(ref destRef, i); - dp = sp.ToVector4(); - } - } - - private static int CalculateVector4ConversionThreshold() - { - if (!Vector.IsHardwareAccelerated) - { - return int.MaxValue; - } - - return SimdUtils.ExtendedIntrinsics.IsAvailable && SimdUtils.IsAvx2CompatibleArchitecture ? 256 : 128; - } } } \ No newline at end of file diff --git a/src/ImageSharp/PixelFormats/PixelConverter.cs b/src/ImageSharp/PixelFormats/Utils/PixelConverter.cs similarity index 98% rename from src/ImageSharp/PixelFormats/PixelConverter.cs rename to src/ImageSharp/PixelFormats/Utils/PixelConverter.cs index 8fde490fd..2336dbee7 100644 --- a/src/ImageSharp/PixelFormats/PixelConverter.cs +++ b/src/ImageSharp/PixelFormats/Utils/PixelConverter.cs @@ -1,11 +1,10 @@ // Copyright (c) Six Labors and contributors. // Licensed under the Apache License, Version 2.0. -using System; using System.Buffers.Binary; using System.Runtime.CompilerServices; -namespace SixLabors.ImageSharp.PixelFormats +namespace SixLabors.ImageSharp.PixelFormats.Utils { /// /// Contains optimized implementations for conversion between pixel formats. diff --git a/src/ImageSharp/PixelFormats/PixelExtensions.cs b/src/ImageSharp/PixelFormats/Utils/PixelExtensions.cs similarity index 93% rename from src/ImageSharp/PixelFormats/PixelExtensions.cs rename to src/ImageSharp/PixelFormats/Utils/PixelExtensions.cs index 175696ab6..2284b8fc2 100644 --- a/src/ImageSharp/PixelFormats/PixelExtensions.cs +++ b/src/ImageSharp/PixelFormats/Utils/PixelExtensions.cs @@ -1,7 +1,7 @@ // Copyright (c) Six Labors and contributors. // Licensed under the Apache License, Version 2.0. -namespace SixLabors.ImageSharp.PixelFormats +namespace SixLabors.ImageSharp.PixelFormats.Utils { /// /// Low-performance extension methods to help conversion syntax, suitable for testing purposes. diff --git a/src/ImageSharp/PixelFormats/Utils/Vector4Converters.Default.cs b/src/ImageSharp/PixelFormats/Utils/Vector4Converters.Default.cs new file mode 100644 index 000000000..139dbfa10 --- /dev/null +++ b/src/ImageSharp/PixelFormats/Utils/Vector4Converters.Default.cs @@ -0,0 +1,89 @@ +using System; +using System.Numerics; +using System.Runtime.CompilerServices; +using System.Runtime.InteropServices; + +namespace SixLabors.ImageSharp.PixelFormats.Utils +{ + /// + /// Helper class for (bulk) conversion of buffers to/from other buffer types. + /// + internal static partial class Vector4Converters + { + /// + /// Provides default implementations for batched to/from conversion. + /// WARNING: The methods are operating without bounds checking and input validation! + /// Input validation is the responsibility of the caller! + /// + public static class Default + { + [MethodImpl(InliningOptions.ShortMethod)] + internal static void DangerousFromVector4( + ReadOnlySpan sourceVectors, + Span destPixels) + where TPixel : struct, IPixel + { + ref Vector4 sourceRef = ref MemoryMarshal.GetReference(sourceVectors); + ref TPixel destRef = ref MemoryMarshal.GetReference(destPixels); + + for (int i = 0; i < sourceVectors.Length; i++) + { + ref Vector4 sp = ref Unsafe.Add(ref sourceRef, i); + ref TPixel dp = ref Unsafe.Add(ref destRef, i); + dp.FromVector4(sp); + } + } + + [MethodImpl(InliningOptions.ShortMethod)] + internal static void DangerousToVector4( + ReadOnlySpan sourcePixels, + Span destVectors) + where TPixel : struct, IPixel + { + ref TPixel sourceRef = ref MemoryMarshal.GetReference(sourcePixels); + ref Vector4 destRef = ref MemoryMarshal.GetReference(destVectors); + + for (int i = 0; i < sourcePixels.Length; i++) + { + ref TPixel sp = ref Unsafe.Add(ref sourceRef, i); + ref Vector4 dp = ref Unsafe.Add(ref destRef, i); + dp = sp.ToVector4(); + } + } + + [MethodImpl(InliningOptions.ShortMethod)] + internal static void DangerousFromScaledVector4( + ReadOnlySpan sourceVectors, + Span destinationColors) + where TPixel : struct, IPixel + { + ref Vector4 sourceRef = ref MemoryMarshal.GetReference(sourceVectors); + ref TPixel destRef = ref MemoryMarshal.GetReference(destinationColors); + + for (int i = 0; i < sourceVectors.Length; i++) + { + ref Vector4 sp = ref Unsafe.Add(ref sourceRef, i); + ref TPixel dp = ref Unsafe.Add(ref destRef, i); + dp.FromScaledVector4(sp); + } + } + + [MethodImpl(InliningOptions.ShortMethod)] + internal static void DangerousToScaledVector4( + ReadOnlySpan sourceColors, + Span destinationVectors) + where TPixel : struct, IPixel + { + ref TPixel sourceRef = ref MemoryMarshal.GetReference(sourceColors); + ref Vector4 destRef = ref MemoryMarshal.GetReference(destinationVectors); + + for (int i = 0; i < sourceColors.Length; i++) + { + ref TPixel sp = ref Unsafe.Add(ref sourceRef, i); + ref Vector4 dp = ref Unsafe.Add(ref destRef, i); + dp = sp.ToScaledVector4(); + } + } + } + } +} \ No newline at end of file diff --git a/src/ImageSharp/PixelFormats/Utils/Vector4Converters.RgbaCompatible.cs b/src/ImageSharp/PixelFormats/Utils/Vector4Converters.RgbaCompatible.cs new file mode 100644 index 000000000..7c57fe4fb --- /dev/null +++ b/src/ImageSharp/PixelFormats/Utils/Vector4Converters.RgbaCompatible.cs @@ -0,0 +1,154 @@ +using System; +using System.Buffers; +using System.Numerics; +using System.Runtime.CompilerServices; +using System.Runtime.InteropServices; + +namespace SixLabors.ImageSharp.PixelFormats.Utils +{ + /// + /// Contains + /// + internal static partial class Vector4Converters + { + /// + /// Provides efficient implementations for batched to/from conversion. + /// which is applicable for -compatible pixel types where + /// returns the same scaled result as . + /// The method is works by internally converting to a therefore it's not applicable for that type! + /// + public static class RgbaCompatible + { + /// + /// It's not worth to bother the transitive pixel conversion method below this limit. + /// The value depends on the actual gain brought by the SIMD characteristics of the executing CPU and JIT. + /// + private static readonly int Vector4ConversionThreshold = CalculateVector4ConversionThreshold(); + + /// + /// Provides an efficient default implementation for + /// and + /// which is applicable for -compatible pixel types where + /// returns the same scaled result as . + /// The method is works by internally converting to a therefore it's not applicable for that type! + /// + [MethodImpl(InliningOptions.ShortMethod)] + internal static void ToVector4( + Configuration configuration, + PixelOperations pixelOperations, + ReadOnlySpan sourcePixels, + Span destVectors, + bool scaled) + where TPixel : struct, IPixel + { + Guard.NotNull(configuration, nameof(configuration)); + Guard.DestinationShouldNotBeTooShort(sourcePixels, destVectors, nameof(destVectors)); + + int count = sourcePixels.Length; + + // Not worth for small buffers: + if (count < Vector4ConversionThreshold) + { + ToVector4Fallback(sourcePixels, destVectors, scaled); + + return; + } + + // Using the last quarter of 'destVectors' as a temporary buffer to avoid allocation: + int countWithoutLastItem = count - 1; + ReadOnlySpan reducedSource = sourcePixels.Slice(0, countWithoutLastItem); + Span lastQuarterOfDestBuffer = MemoryMarshal.Cast(destVectors).Slice((3 * count) + 1, countWithoutLastItem); + pixelOperations.ToRgba32(configuration, reducedSource, lastQuarterOfDestBuffer); + + // 'destVectors' and 'lastQuarterOfDestBuffer' are ovelapping buffers, + // but we are always reading/writing at different positions: + SimdUtils.BulkConvertByteToNormalizedFloat( + MemoryMarshal.Cast(lastQuarterOfDestBuffer), + MemoryMarshal.Cast(destVectors.Slice(0, countWithoutLastItem))); + + destVectors[countWithoutLastItem] = sourcePixels[countWithoutLastItem].ToVector4(); + } + + /// + /// Provides an efficient default implementation for + /// and + /// which is applicable for -compatible pixel types where + /// returns the same scaled result as . + /// The method is works by internally converting to a therefore it's not applicable for that type! + /// + [MethodImpl(InliningOptions.ShortMethod)] + internal static void FromVector4( + Configuration configuration, + PixelOperations pixelOperations, + ReadOnlySpan sourceVectors, + Span destPixels, + bool scaled) + where TPixel : struct, IPixel + { + Guard.NotNull(configuration, nameof(configuration)); + Guard.DestinationShouldNotBeTooShort(sourceVectors, destPixels, nameof(destPixels)); + + int count = sourceVectors.Length; + + // Not worth for small buffers: + if (count < Vector4ConversionThreshold) + { + FromVector4Fallback(sourceVectors, destPixels, scaled); + + return; + } + + // For the opposite direction it's not easy to implement the trick used in RunRgba32CompatibleToVector4Conversion, + // so let's allocate a temporary buffer as usually: + using (IMemoryOwner tempBuffer = configuration.MemoryAllocator.Allocate(count)) + { + Span tempSpan = tempBuffer.Memory.Span; + + SimdUtils.BulkConvertNormalizedFloatToByteClampOverflows( + MemoryMarshal.Cast(sourceVectors), + MemoryMarshal.Cast(tempSpan)); + + pixelOperations.FromRgba32(configuration, tempSpan, destPixels); + } + } + + [MethodImpl(InliningOptions.ColdPath)] + private static void ToVector4Fallback(ReadOnlySpan sourcePixels, Span destVectors, bool scaled) + where TPixel : struct, IPixel + { + if (scaled) + { + Default.DangerousToScaledVector4(sourcePixels, destVectors); + } + else + { + Default.DangerousToVector4(sourcePixels, destVectors); + } + } + + [MethodImpl(InliningOptions.ColdPath)] + private static void FromVector4Fallback(ReadOnlySpan sourceVectors, Span destPixels, bool scaled) + where TPixel : struct, IPixel + { + if (scaled) + { + Default.DangerousFromScaledVector4(sourceVectors, destPixels); + } + else + { + Default.DangerousFromVector4(sourceVectors, destPixels); + } + } + + private static int CalculateVector4ConversionThreshold() + { + if (!Vector.IsHardwareAccelerated) + { + return int.MaxValue; + } + + return SimdUtils.ExtendedIntrinsics.IsAvailable && SimdUtils.IsAvx2CompatibleArchitecture ? 256 : 128; + } + } + } +} \ No newline at end of file diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4_Bgra32.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4_Bgra32.cs index 028bfe46f..39702d525 100644 --- a/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4_Bgra32.cs +++ b/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4_Bgra32.cs @@ -16,5 +16,26 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk this.source.GetSpan(), this.destination.GetSpan()); } + + // RESULTS: + // Method | Runtime | Count | Mean | Error | StdDev | Scaled | ScaledSD | Gen 0 | Allocated | + // ---------------------------- |-------- |------ |-----------:|------------:|-----------:|-------:|---------:|-------:|----------:| + // PixelOperations_Base | Clr | 64 | 339.9 ns | 138.30 ns | 7.8144 ns | 1.00 | 0.00 | 0.0072 | 24 B | + // PixelOperations_Specialized | Clr | 64 | 338.1 ns | 13.30 ns | 0.7515 ns | 0.99 | 0.02 | - | 0 B | + // | | | | | | | | | | + // PixelOperations_Base | Core | 64 | 245.6 ns | 29.05 ns | 1.6413 ns | 1.00 | 0.00 | 0.0072 | 24 B | + // PixelOperations_Specialized | Core | 64 | 257.1 ns | 37.89 ns | 2.1407 ns | 1.05 | 0.01 | - | 0 B | + // | | | | | | | | | | + // PixelOperations_Base | Clr | 256 | 972.7 ns | 61.98 ns | 3.5020 ns | 1.00 | 0.00 | 0.0057 | 24 B | + // PixelOperations_Specialized | Clr | 256 | 882.9 ns | 126.21 ns | 7.1312 ns | 0.91 | 0.01 | - | 0 B | + // | | | | | | | | | | + // PixelOperations_Base | Core | 256 | 910.0 ns | 90.87 ns | 5.1346 ns | 1.00 | 0.00 | 0.0067 | 24 B | + // PixelOperations_Specialized | Core | 256 | 448.4 ns | 15.77 ns | 0.8910 ns | 0.49 | 0.00 | - | 0 B | + // | | | | | | | | | | + // PixelOperations_Base | Clr | 2048 | 6,951.8 ns | 1,299.01 ns | 73.3963 ns | 1.00 | 0.00 | - | 24 B | + // PixelOperations_Specialized | Clr | 2048 | 5,852.3 ns | 630.56 ns | 35.6279 ns | 0.84 | 0.01 | - | 0 B | + // | | | | | | | | | | + // PixelOperations_Base | Core | 2048 | 6,937.5 ns | 1,692.19 ns | 95.6121 ns | 1.00 | 0.00 | - | 24 B | + // PixelOperations_Specialized | Core | 2048 | 2,994.5 ns | 1,126.65 ns | 63.6578 ns | 0.43 | 0.01 | - | 0 B | } } \ No newline at end of file diff --git a/tests/ImageSharp.Benchmarks/General/PixelConversion/PixelConversion_ConvertFromRgba32.cs b/tests/ImageSharp.Benchmarks/General/PixelConversion/PixelConversion_ConvertFromRgba32.cs index 424020e2f..9f1b2721b 100644 --- a/tests/ImageSharp.Benchmarks/General/PixelConversion/PixelConversion_ConvertFromRgba32.cs +++ b/tests/ImageSharp.Benchmarks/General/PixelConversion/PixelConversion_ConvertFromRgba32.cs @@ -7,6 +7,7 @@ using System.Runtime.InteropServices; using BenchmarkDotNet.Attributes; using SixLabors.ImageSharp.PixelFormats; +using SixLabors.ImageSharp.PixelFormats.Utils; namespace SixLabors.ImageSharp.Benchmarks.General.PixelConversion { diff --git a/tests/ImageSharp.Tests/PixelFormats/PixelConverterTests.cs b/tests/ImageSharp.Tests/PixelFormats/PixelConverterTests.cs index 9b32f7aee..c539e9dcf 100644 --- a/tests/ImageSharp.Tests/PixelFormats/PixelConverterTests.cs +++ b/tests/ImageSharp.Tests/PixelFormats/PixelConverterTests.cs @@ -1,4 +1,5 @@ using SixLabors.ImageSharp.PixelFormats; +using SixLabors.ImageSharp.PixelFormats.Utils; using Xunit;