Simd v5.0.115

Algorithms

New features

• Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16, AMX optimizations of class SynetMergedConvolution32fBf16Cdc.
• Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16, AMX optimizations of class SynetMergedConvolution32fBf16Cd.
• Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16, AMX optimizations of class SynetMergedConvolution32fBf16Dc.
• AVX-512BF16 extension support.
• AVX-512BF16 optimizations of function Float32ToBFloat16.
• AVX-512BF16, AMX optimizations of class SynetConvolution32fBf16Nhwc.
• AMX extension support.
• Support of 3D pooling in Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function SynetPoolingMax32f.

Improving

• AVX-512BW optimizations of function Fill32f.

Renaming

• Rename function SynetPoolingForwardAverage to SynetPoolingAverage.
• Rename function SynetPoolingForwardMax32f to SynetPoolingMax32f.
• Rename function SynetPoolingForwardMax8u to SynetPoolingMax8u.

Replacing

• Replace AVX-512F optimizations to AVX-512BW for function SvmSumLinear.
• Replace AVX-512F optimizations to AVX-512BW for function Fill32f.
• Replace AVX-512F optimizations to AVX-512BW for class ResizerNearest.
• Replace AVX-512F optimizations to AVX-512BW for class ResizerFloatBilinear.
• Replace AVX-512F optimizations to AVX-512BW for function SquaredDifferenceSum32f.
• Replace AVX-512F optimizations to AVX-512BW for function SquaredDifferenceKahanSum32f.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralConvolutionForward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution2x2Forward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution2x2Backward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution2x2Sum.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution3x3Forward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution3x3Backward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution3x3Sum.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution4x4Forward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution4x4Backward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution4x4Sum.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution5x5Forward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution5x5Backward.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddConvolution5x5Sum.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralProductSum.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAdaptiveGradientUpdate.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralPooling1x1Max3x3.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralPooling2x2Max2x2.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralPooling2x2Max3x3.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralUpdateWeights.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddValue.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddVector.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralAddVectorMultipliedByValue.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralRoughSigmoid.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralRoughSigmoid2.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralDerivativeSigmoid.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralRoughTanh.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralDerivativeTanh.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralDerivativeRelu.
• Replace AVX-512F optimizations to AVX-512BW for function NeuralPow.
• Replace AVX-512F optimizations to AVX-512BW for class SynetConvolution32fGemmNN.
• Replace AVX-512F optimizations to AVX-512BW for class SynetConvolution32fGemmNT.
• Replace AVX-512F optimizations to AVX-512BW for class SynetConvolution32fWinograd.
• Replace AVX-512F optimizations to AVX-512BW for class SynetDeconvolution32fGemmNN.
• Replace AVX-512F optimizations to AVX-512BW for class SynetDeconvolution32fNhwcDirect2x2.
• Replace AVX-512F optimizations to AVX-512BW for function SynetDeconvolution32fInit.
• Replace AVX-512F optimizations to AVX-512BW for class SynetInnerProduct32fGemm.
• Replace AVX-512F optimizations to AVX-512BW for class SynetInnerProduct32fProd.
• Replace AVX-512F optimizations to AVX-512BW for function SynetInnerProduct32fInit.
• Replace AVX-512F optimizations to AVX-512BW for function ConvolutionBiasAndActivation.
• Replace AVX-512F optimizations to AVX-512BW for function SynetReorderImage.
• Replace AVX-512F optimizations to AVX-512BW for function SynetReorderFilter.
• Replace AVX-512F optimizations to AVX-512BW for function Gemm32fNN.
• Replace AVX-512F optimizations to AVX-512BW for function Gemm32fNT.
• Replace AVX-512F optimizations to AVX-512BW for function SynetFusedLayerForward0.
• Replace AVX-512F optimizations to AVX-512BW for function SynetFusedLayerForward1.
• Replace AVX-512F optimizations to AVX-512BW for function SynetFusedLayerForward2.
• Replace AVX-512F optimizations to AVX-512BW for function SynetFusedLayerForward3.
• Replace AVX-512F optimizations to AVX-512BW for function SynetFusedLayerForward4.
• Replace AVX-512F optimizations to AVX-512BW for function SynetFusedLayerForward8.
• Replace AVX-512F optimizations to AVX-512BW for function SynetFusedLayerForward9.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel1x3Block1x4SetFilter.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel1x3Block1x4SetInput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel1x3Block1x4SetOutput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel1x5Block1x4SetFilter.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel1x5Block1x4SetInput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel1x5Block1x4SetOutput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel2x2Block2x2SetFilter.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel2x2Block2x2SetInput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel2x2Block2x2SetOutput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel2x2Block4x4SetFilter.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel2x2Block4x4SetInput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel2x2Block4x4SetOutput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block2x2SetFilter.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block2x2SetInput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block2x2SetOutput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block3x3SetFilter.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block3x3SetInput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block3x3SetOutput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block4x4SetFilter.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block4x4SetInput.
• Replace AVX-512F optimizations to AVX-512BW for function WinogradKernel3x3Block4x4SetOutput.
• Replace AVX-512F optimizations to AVX-512BW for function SynetElu32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetHardSigmoid32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetHswish32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetMish32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetPreluLayerForward.
• Replace AVX-512F optimizations to AVX-512BW for function SynetRelu32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetRestrictRange32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetSigmoid32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetSoftplus32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetSwish32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetTanh32f.
• Replace AVX-512F optimizations to AVX-512BW for function SynetScaleLayerForward.
• Replace AVX-512F optimizations to AVX-512BW for function SynetPoolingAverage.
• Replace AVX-512F optimizations to AVX-512BW for function SynetAddBias.
• Replace AVX-512F optimizations to AVX-512BW for function SynetEltwiseLayerForward.
• Replace AVX-512F optimizations to AVX-512BW for function SynetInnerProductLayerForward.
• Replace AVX-512F optimizations to AVX-512BW for function SynetLrnLayerCrossChannels.

Simd v4.10.114

Algorithms

New features

• Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function Yuv420pToUyvy422.
• AVX-512BW, NEON optimizations of function Uyvy422ToYuv420p.
• AVX-512BW, NEON optimizations of function Uyvy422ToBgr.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function Float32ToBFloat16.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function BFloat16ToFloat32.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function BFloat16ToFloat32.
• Base implementation of class SynetConvolution32fBf16Gemm.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of class SynetConvolution32fBf16Nhwc.
• Base implementation of class SynetMergedConvolution32fBf16.

Removing

• Remove external GEMM function parameter from function SynetConvolution32fInit.
• Remove external GEMM function parameter from function SynetDeconvolution32fInit.

Test framework

New features

• Tests for verifying functionality of function Yuv420pToUyvy422.
• Tests for verifying functionality of function Float32ToBFloat16.
• Tests for verifying functionality of function BFloat16ToFloat32.

Infrastructure

New features

• Project files for Microsoft Visual Studio 2022.

Simd v4.9.113

Algorithms

New features

• SSE4.1, AVX2, AVX-512BW optimizations of class ResizerByteArea2x2.

Improving

• Base implementation of class ResizerByteArea1x1.

Bug fixing

• Error in Base implementation of class ResizerByteArea2x2.
• Error in AVX optimizations of class SynetConvolution32fDirectNchw.

Removing

• SimdSynetCompatibilityFloatZero flag.

Infrastructure

New features

• Git commit ID info in function SimdVersion.
• Git branch name in function SimdVersion.

Simd v4.9.112

Algorithms

New features

• NEON optimizations of function Base64Encode.
• NEON optimizations of ImageJpegSaver class.
• NEON optimizations of function Yuv420pSaveAsJpegToMemory.
• NEON optimizations of function Nv12SaveAsJpegToMemory.
• Owner method in View structure.
• Owner method in Frame structure.
• Capture method in View structure.
• Capture method in Frame structure.
• Base implementation of class ResizerByteAreaReduced2x2.

Bug fixing

• MSVS compiler error in AVX-512BW optimizations of function Yuv420pToBgraV2.
• Error in AVX2 optimizations of function BgraToRgb.
• Error (aligned reading of unaligned memory) in SSE4.1, AVX2, AVX-512BW optimizations of function InterleaveBgra.
• Error in function View::ToOcv.
• Error in View copy constructor (from OpenCV Mat).

Test framework

Bug fixing

• Wrong default ROOT_PATH for Linux.
• Error in test SynetConvert32fTo8uAutoTest.
• Special test ResizeYuv420pSpecialTest.

Simd v4.9.111

Algorithms

New features

• AVX2, AVX-512BW optimizations of ResizerByteBicubic class.
• SSE4.1, AVX2, AVX-512BW, NEON optimizations of function Base64Decode.
• NEON optimizations of function SynetSwish32f.
• Swish activation function to NEON optimizations of SynetConvolution32f framework.
• Swish activation function to NEON optimizations of SynetDeconvolution32f framework.
• Swish activation function to NEON optimizations of SynetMergedConvolution32f framework.
• Swish activation function to NEON optimizations of SynetConvolution8i framework.
• Swish activation function to NEON optimizations of SynetMergedConvolution8i framework.
• NEON optimizations of function Yuv444pToBgraV2.
• SSE2, AVX2, AVX-512BW, NEON optimizations of function Yuv420pToBgraV2.

Improving

• SSE4.1 optimizations of ResizerByteBicubic class.

Bug fixing

• Compiler error in NEON optimizations of function AlphaUnpremultiply.
• MSVS Compiler warnings in SSE4.1, AVX2, AVX-512BW optimizations of function TransformImage.

Simd v4.9.110

Algorithms

New features

• Base implementation, SSE4.1 optimizations of ResizerByteBicubic class.
• Base implementation of function BgraToYuv444pV2.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function Nv12SaveAsJpegToMemory.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function Yuv420pSaveAsJpegToMemory.
• Base implementation of function BgraToYuv420pV2.

Bug fixing

• Error in SSE4.1, AVX2, AVX-512BW optimizations of function BgraToRgba.
• Error in SSE4.1, AVX2 optimizations of function BgraToBgr.
• Error in SSE4.1, AVX2 optimizations of function BgraToRgb.
• Error in Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function AlphaUnpremultiply.

Test framework

New features

• Tests for verifying functionality of function BgraToYuv444pV2.
• Tests for verifying functionality of function Nv12SaveAsJpegToMemory.
• Tests for verifying functionality of function Yuv420pSaveAsJpegToMemory.
• Tests for verifying functionality of function BgraToYuv420pV2.

Simd v4.9.109

Algorithms

New features

• Parameter Uyvy422ToBgr to function.
• SSE4.1, AVX2 optimizations of function Uyvy422ToBgr.
• Base implementation, SSE4.1, AVX2 optimizations of function Uyvy422ToYuv420p.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function Base64Encode.
• Base implementation of function Base64Decode.

Improving

• AVX2 optimizations of class ResizerNearest for Bgr24, Uv16.

Renaming

• Function UyvyToBgr to Uyvy422ToBgr.

Test framework

New features

• Tests for verifying functionality of function Uyvy422ToYuv420p.
• Tests for verifying functionality of function Base64Encode.
• Tests for verifying functionality of function Base64Decode.

Documentation

Changes

• Update developers list.

Simd v4.9.108

Algorithms

New features

• SSE4.1, AVX2, AVX-512F, AVX-512BW optimizations of class ResizerNearest.
• Add SimdResizeMethodNearestPytorch to SimdResizeMethodType enumeration.
• Add parameter BackgroundStatUpdateTime to Motion Detector.
• MotionDetector performance optimization (case of falling star).
• 16-bit UYVY image format in View.
• Base implementation of function UyvyToBgr.
• Base implementation, SSE2, AVX2, AVX-512F optimizations of function SynetSwish32f.
• SimdConvolutionActivationSwish item of SimdConvolutionActivationType enumeration.
• Swish activation function to Base implementation, SSE2, AVX2, AVX-512F optimizations of SynetConvolution32f framework.
• Swish activation function to Base implementation, SSE2, AVX2, AVX-512F optimizations of SynetDeconvolution32f framework.
• Swish activation function to Base implementation, SSE2, AVX2, AVX-512F optimizations of SynetMergedConvolution32f framework.
• Swish activation function to Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512VNNI optimizations of SynetConvolution8i framework.
• Swish activation function to Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512VNNI optimizations of SynetMergedConvolution8i framework.
• SimdYuvType enumeration.
• Base implementation, SSE2, AVX2, AVX-512BW optimizations of function Yuv444pToBgraV2.
• Function Simd::Resize supports images with 16-bit channel size.
• Base implementation function Yuv420pToBgraV2.

Improving

• Refactoring of SimdResizeMethodType enumeration.

Bug fixing

• Stack corruption in function Simd::Avx2::JpegWriteBlockSubs.

Test framework

New features

• Tests for verifying functionality of function UyvyToBgr.
• Tests for verifying functionality of function SynetSwish32f.
• Tests for verifying functionality of function Yuv444pToBgraV2.
• Tests for verifying functionality of function Yuv420pToBgraV2.

Infrastructure

Bug fixing

• Wrong compiler options correction in Cmake.

Simd v4.9.107

Algorithms

New features

• Internal class Holder to replace std::unique_ptr for old compilers without support of C++11 standard.
• SimdBayerLayoutType enumeration.
• Base implementation of class ResizerNearest.

Bug fixing

• Compiler error when defined macro SIMD_SSE2_DISABLE.
• Compiler error when defined macro SIMD_NEON_DISABLE.

Infrastructure

New features

• SIMD_ROOT Cmake parameter.

Simd v4.9.106

Algorithms

New features

• Base implementation, SSE2, AVX, AVX-512F, NEON optimizations of function SynetHardSigmoid32f.
• SimdConvolutionActivationHardSigmoid item of SimdConvolutionActivationType enumeration.
• HardSigmoid activation function to Base implementation, SSE2, AVX, AVX2, AVX-512F, NEON optimizations of SynetConvolution32f framework.
• HardSigmoid activation function to Base implementation, SSE2, AVX, AVX2, AVX-512F, NEON optimizations of SynetDeconvolution32f framework.
• HardSigmoid activation function to Base implementation, SSE2, AVX, AVX2, AVX-512F, NEON optimizations of SynetMergedConvolution32f framework.
• NEON optimizations of SynetMergedConvolution32fDc class.
• NEON optimizations of SynetMergedConvolution32fCd class.
• NEON optimizations of SynetInnerProduct32fGemm class.
• NEON optimizations of SynetInnerProduct32fProd class.
• HardSigmoid activation function to Base implementation, SSE41, AVX2, AVX-512BW, AVX-512VNNI, NEON optimizations of SynetConvolution8i framework.
• HardSigmoid activation function to Base implementation, SSE41, AVX2, AVX-512BW, AVX-512VNNI optimizations of SynetMergedConvolution8i framework.

Bug fixing

• Compiler error in file SimdInit.h (CLang, Windows).

Removing

• Remove including SimdConfig.h in SimdLib.h.

Test framework

New features

• Tests for verifying functionality of function SynetHardSigmoid32f.
• '-pi' test parameter (to print internal performance statistics of Simd Library to console).