Simd Library Release Notes (2026).

April X, 2026 (version 6.2.160)

Algorithms

New features

• Possibility to use non constant B matrix in framework SynetInnerProduct32f.
• Function SimdSynetInnerProduct32fExternalBufferSize.
• Parameter 'activation' to function SimdSynetInnerProduct16bInit.
• Parameter 'params' to function SimdSynetInnerProduct16bSetParams.

Improving

• AMX-BF16 optimizations of class SynetMergedConvolution16bCdc.
• AMX-BF16 optimizations of class SynetMergedConvolution16bCd.
• AMX-BF16 optimizations of class SynetMergedConvolution16bDc.

Bug fixing

• Error in SSE4.1, AVX2, AVX-512BW optimizations of function SynetQuantizedPreluLayerForward (possible aligned read of unaligned memory).
• Error in SSE4.1, AVX2, AVX-512BW optimizations of function SynetQuantizedScaleLayerForward (possible aligned read of unaligned memory).
• Error in SSE4.1 optimizations of class ResizerFloatBilinear (possible aligned read of unaligned memory).
• Error in SSE4.1 optimizations of class ResizerBf16Bilinear (possible aligned read of unaligned memory).
• Error in SSE4.1, AVX2 optimizations of class ResizerByteBilinear (possible aligned read of unaligned memory).
• Error in SSE4.1 optimizations of class ResizerByteBilinear (possible aligned write to unaligned memory).
• Error in SSE4.1 optimizations of class ResizerFloatBilinear (possible aligned write to unaligned memory).
• Memory leak in function Simd::Detection::LoadStringXml.
• Possible crash in function Simd::ImageLoadFromFile.
• Possible crash in function Simd::Base64Decode.
• Possible crash in AVX-512BW optimization of function TransformImageRotate270.

March 3, 2026 (version 6.2.159)

Algorithms

New features

• Base implementation, AMX-BF16 optimizations of class SynetConvolution16bNhwcSpecV2.
• Support of SimdSynetUnaryOperation32fRound in function SynetUnaryOperation32f.
• Support of SimdSynetUnaryOperation32fSign in function SynetUnaryOperation32f.

Bug fixing

• Error in AMX-BF16 optimizations of class SynetConvolution16bNhwcGemmV1 (kernel Convolution16bNhwcGemm_Macro32x32).
• Error in Base implementation of class SynetQuantizedConvolutionNhwcDepthwiseV2 (multithread using of SimdSynetQuantizedConvolutionForward).
• Error in Base implementation of class SynetQuantizedConvolutionNhwcDepthwiseV3 (multithread using of SimdSynetQuantizedConvolutionForward).
• Error in AMX-BF16 optimizations of class SynetConvolution16bNhwcGemmV1 (Inv2x2, kernel Convolution16bNhwcGemm_MacroNx32, unaligned dstH*dstW).
• Error in AMX-BF16 optimizations of class SynetConvolution16bNhwcGemmV1 (Inv2x2, kernel Convolution16bNhwcGemm_MacroNx32, unaligned dstC).
• Error in SSE4.1, AVX2, AVX-512BW, NEON optimizations of function AbsDifference (wrong alignment checking).
• Error in AVX-512BW optimizations of class SynetConvolution32fGemmNN (case of extra large padding).
• Error in AVX-512BW optimizations of class ResizerNearest.
• Error in AMX-BF16 optimizations of class SynetMergedConvolution16bCdc (batch > 1, small input size).
• Error in functions BodyH, BodyW (file SimdSynetConvParam.h).

February 3, 2026 (version 6.2.158)

Algorithms

New features

• Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function MidpointFilterSquare3x3.
• Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function MidpointFilterSquare5x5.
• Base implementation of class SynetConvolution16bNhwcSpecV2.
• Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function MinFilterSquare3x3.
• Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function MinFilterSquare5x5.
• Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function MaxFilterSquare3x3.
• Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of function MaxFilterSquare5x5.

Improving

• AMX-BF16 optimizations of class SynetConvolution16bNhwcGemmV1.

Test framework

New features

• Tests for verifying functionality of function MidpointFilterSquare3x3.
• Tests for verifying functionality of function MidpointFilterSquare5x5.
• Tests for verifying functionality of function MinFilterSquare3x3.
• Tests for verifying functionality of function MinFilterSquare5x5.
• Tests for verifying functionality of function MaxFilterSquare3x3.
• Tests for verifying functionality of function MaxFilterSquare5x5.

January 2, 2026 (version 6.2.157)

Algorithms

New features

• Function Simd::Resize for Simd::Frame.
• Base implementation of function DrawLine.
• Base implementation of function DrawRectangle.
• Base implementation of function FontInit.
• Base implementation of function FontResize.
• Base implementation of function FontHeight.
• Base implementation of function FontMeasure.
• Base implementation of function FontDraw.

Improving

• Base implementation, AMX-BF16 optimizations of class SynetConvolution16bNhwcGemmV1.
• AVX-512BW optimizations of function SynetPoolingMax32f (case of SynetPoolingMax32f2DNhwcSolid2x2).
• AVX-512BW optimizations of function SynetMergedConvolution32f (InputConvolution1x1).
• AVX-512BW optimizations of function SynetMergedConvolution32f (DepthwiseConvolution_k3p1d1s1w6).
• Simd::DrawLine uses SimdDrawLine instead its own implementation.
• Simd::DrawRectangle uses SimdDrawRectangle instead its own implementation.
• Simd::Font uses functions SimdFontInit, SimdFontResize, SimdFontHeight, SimdFontMeasure, SimdFontDraw instead of its own implementation.

Python wrapper

New features

• Function Simd.ResizeFrame.
• Function Simd.ResizedFrame.
• Yuv444p member to Simd.FrameFormat enumeration.
• Method Simd.ImageFrame.Save.
• Method Simd.ImageFrame.Load.
• Function Simd.Lib.StretchGray2x2.
• Function Simd.StretchGray2x2.
• Function Simd.Lib.BgraToYuv444p.
• Function Simd.Lib.Yuv444pToRgb.
• Function Simd.Lib.ReduceGray2x2.
• Function Simd.ReduceGray2x2.
• Function Simd.Lib.BgrToYuv444p.
• Function Simd.Lib.BgraToYuv444p.
• Function Simd.Lib.Yuv444pToBgr.
• Function Simd.Lib.Yuv444pToRgba.
• Function Simd.Lib.DrawLine.
• Method Simd.Image.DrawLine.
• Function Simd.Lib.DrawRectangle.
• Method Simd.Image.DrawRectangle.
• Function Simd.Lib.FontInit.
• Function Simd.Lib.FontResize.
• Function Simd.Lib.FontHeight.
• Function Simd.Lib.FontMeasure.
• Function Simd.Lib.FontDraw.
• Class Simd.TextFont.
• Method Simd.Image.DrawFilledRectangle.

Improving

• Support of Simd.FrameFormat.Yuv444p in method Simd.ImageFrame.Recreate.
• Support of Simd.FrameFormat.Yuv444p in method Simd.ImageFrame.Convert.

Bug fixing

• Error in method Simd.Frame.Convert.

Renaming

• Function Simd.Resize to Simd.ResizeImage.
• Function Simd.Resized to Simd.ResizedImage.

Test framework

New features

• Tests for verifying functionality of function DrawLine.
• Tests for verifying functionality of function DrawRectangle.

Bug fixing

• Error in method Test::PerformanceMeasurerStorage::Clear.