Simd Library Release Notes (2025).

February X, 2025 (version X.X.146)

Algorithms

New features

• AVX2, AVX-512BW optimizations of class ResizerBf16Bilinear.

Improving

• SSE4.1 optimizations of class ResizerBf16Bilinear.

January 1, 2025 (version 6.1.145)

Algorithms

New features

• Parameter add in function SimdSynetMergedConvolution16bInit.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function SynetTiledScale2D32f.
• AMX-BF16 kernel DepthwiseConvolution_k5p2d1s1w6 for class SynetMergedConvolution16b.
• AMX-BF16 kernel DepthwiseConvolution_k5p2d1s1w4 for class SynetMergedConvolution16b.
• AMX-BF16 kernel DepthwiseConvolution_k3p1d1s1w8 for class SynetMergedConvolution16b.
• AMX-BF16 kernel DepthwiseConvolution_k3p1d1s1w6 for class SynetMergedConvolution16b.
• Base implementation, SSE4.1 optimizations of class ResizerBf16Bilinear.

Improving

• Extend using of AMX-BF16 optimization of function DepthwiseConvolution_k7p3d1s1w4.
• Extend using of AMX-BF16 optimization of function DepthwiseConvolution_k7p3d1s1w6.
• Extend using of AMX-BF16 optimization of function DepthwiseConvolution_k7p3d1s1w8.
• Extend using of AVX-512BW optimization of function Convolution32fNhwcDepthwise_k7p3d1s1w4.
• Extend using of AMX-BF16 optimization of function DepthwiseConvolution_k5p2d1s1w8.
• Performance of SynetConvolution32f (NHWC, srcC=1, dstС=1).

Bug fixing

• Error in AMX-BF16 optimizations of class SynetInnerProduct16bGemmNN.
• Error in AVX-512BW optimizations of class SynetAdd16bUniform.
• Error in AMX-BF16 optimizations of function DepthwiseConvolutionDefault.
• Error in AMX-BF16 optimizations of function DepthwiseConvolutionLargePad.
• Error in Base implementation of class SynetMergedConvolution16bCdc.
• Error in Base implementation of class SynetMergedConvolution16bCd.
• Error in class InputMemoryStream.

Removing

• Parameter compatibility in function SimdSynetMergedConvolution16bInit.
• Parameter internal in function SimdSynetMergedConvolution16bSetParams.

Test framework

New features

• Tests for verifying functionality of function SynetTiledScale2D32f.