Skip to content

Releases: ermig1979/Simd

Simd v6.1.143

04 Nov 15:26
@ermig1979 ermig1979
v6.1.143
fe88018
Compare
Choose a tag to compare
Simd v6.1.143 Latest
Latest

Algorithms

New features
  • Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of class SynetConvolution16bNhwcDepthwise.
  • AVX-512BW kernel Convolution32fNhwcDepthwise_k7p3d1s1w4 for class SynetConvolution32fNhwcDepthwise.
  • AMX-BF16 kernel DepthwiseConvolution_k7p3d1s1w4 for class SynetMergedConvolution16b.
  • AVX-512BW kernel Convolution32fNhwcDepthwise_k7p3d1s1w6 for class SynetConvolution32fNhwcDepthwise.
  • AVX-512BW kernel Convolution32fNhwcDepthwise_k7p3d1s1w8 for class SynetConvolution32fNhwcDepthwise.
  • AMX-BF16 kernel DepthwiseConvolution_k7p3d1s1w6 for class SynetMergedConvolution16b.
  • AMX-BF16 kernel DepthwiseConvolution_k7p3d1s1w8 for class SynetMergedConvolution16b.
  • AVX-512BW kernel Convolution32fNhwcDepthwise_k7p3d1s1w4 for framework SynetMergedConvolution32f.
  • AVX-512BW kernel Convolution32fNhwcDepthwise_k7p3d1s1w6 for framework SynetMergedConvolution32f.
  • AVX-512BW kernel Convolution32fNhwcDepthwise_k7p3d1s1w8 for framework SynetMergedConvolution32f.
  • AMX-BF16 kernel DepthwiseConvolution_k5p2d1s1w8 for class SynetMergedConvolution16b.
  • Base implementation of function SimdYuv444pToRgbaV2.
Improving
  • AVX-512BW optimizations of function Convolution32fNhwcDepthwiseDefault.
  • AMX-BF16 optimizations of function DepthwiseConvolutionLargePad.
Bug fixing
  • Error in Base implementation of class SynetDeconvolution16bNhwcGemm.

Test framework

New features
  • Tests for verifying functionality of function SimdYuv444pToRgbaV2.
Assets 2
Loading

Simd v6.1.142

01 Oct 07:21
@ermig1979 ermig1979
v6.1.142
76fcac5
Compare
Choose a tag to compare
Simd v6.1.142

Algorithms

New features
  • Base implementation of class SynetDeconvolution16bGemm.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetDeconvolution16bNhwcGemm.
  • AMX-BF16 (AVX-512VBMI) optimizations of function DeinterleaveUv.
  • AMX-BF16 (AVX-512VBMI) optimizations of function DeinterleaveBgr.
  • AMX-BF16 (AVX-512VBMI) optimizations of function DeinterleaveBgra.
Improving
  • AVX-512BW optimizations of function ConvolutionDirectNhwcConvolutionBiasActivationDepthwise.
Removing
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetConvolution32fBf16NhwcGemm.
  • Base implementation of class SynetConvolution32fBf16Gemm.
  • Parameter 'compatibility' from function SynetConvolution32fInit.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetMergedConvolution32fBf16Cdc.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetMergedConvolution32fBf16Cd.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetMergedConvolution32fBf16Dc.
  • Base implementation of class SynetMergedConvolution32fBf16.
  • Parameter 'compatibility' from function SynetMergedConvolution32fInit.

Test framework

New features
  • Tests for verifying functionality of SynetDeconvolution16b framework.
Assets 2
Loading

Simd v6.1.141

02 Sep 08:07
@ermig1979 ermig1979
v6.1.141
30bcb02
Compare
Choose a tag to compare
Simd v6.1.141

Algorithms

New features
  • Support of BFloat16 in Base implementation, SSE4.1, AVX2, AVX-512BW, NEON optimizations of class ResizerNearest.
Bug fixing
  • Compiler warning in function Simd::LitterCpuCache.
  • Error in AVX-512BW optimizations of class SynetInnerProduct16bGemmNN.
Assets 2
Loading

Simd v6.1.140

19 Aug 10:38
@ermig1979 ermig1979
v6.1.140
ae3262a
Compare
Choose a tag to compare
Simd v6.1.140

Algorithms

New features
  • Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function SynetRelu16b.
  • API of SynetAdd16b framework.
  • Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of class SynetAdd16bUniform.
  • Base implementation, SSE4.1, AVX2, AVX-512BW optimizations, AMX-BF16 of class SynetConvolution16bNchwGemm.
Improving
  • AMX-BF16 optimizations of class SynetInnerProduct16bGemmNN.
Bug fixing
  • Error in Base implementation of class SynetMergedConvolution16bCdc.
  • Error in Base implementation of class SynetMergedConvolution16bDc.
  • Error in Base implementation of class SynetInnerProduct16bGemmNN.
  • Error in Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function Float32ToBFloat16.

Test framework

New features
  • Tests for verifying functionality of function SynetRelu16b.
  • Tests for verifying functionality of SynetAdd16b framework.
Assets 2
Loading

Simd v6.1.139

01 Jul 14:09
@ermig1979 ermig1979
v6.1.139
ea84623
Compare
Choose a tag to compare
Simd v6.1.139

Algorithms

New features
  • API of SynetInnerProduct16b framework.
  • Base implementation of class SynetInnerProduct16bRef.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetInnerProduct16bGemmNN.
Bug fixing
  • Error in AVX-512BF16 optimizations of class SynetConvolution16bNhwcDirect.
  • Error in Base implementation of class SynetConvolution16bNhwcGemm.
  • Error in SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of function Convert16bNhwcDirect.
  • Error in SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of function Reorder16bNhwcDirect.
  • Error in Base implementation of class SynetMergedConvolution16bCdc.
  • Error in Base implementation of class SynetMergedConvolution16bDc.
  • Error in Base implementation of class SynetMergedConvolution16bCd.
  • Error in AMX-BF16 optimizations of class SynetMergedConvolution16bDc.

Test framework

New features
  • Tests for verifying functionality of SynetInnerProduct16b framework.
Assets 2
Loading

Simd v6.1.138

03 Jun 06:53
@ermig1979 ermig1979
v6.1.138
1c9c068
Compare
Choose a tag to compare
Simd v6.1.138

Algorithms

New features
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16 optimizations of class SynetConvolution16bNhwcDirect.
  • SimdCpuInfoCurrentFrequency in SimdCpuInfoType enumeration.
  • API of SynetMergedConvolution16b framework.
  • Base implementation of class SynetMergedConvolution16b.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16 optimizations of class SynetMergedConvolution16bDc.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16 optimizations of class SynetMergedConvolution16bCd.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16 optimizations of class SynetMergedConvolution16bCdc.
  • Support of YUV420P format to Simd::Frame.
Improving
  • AVX-512BF16 optimizations of class SynetConvolution16bNhwcGemm.
Bug fixing
  • Errors in Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16 optimizations of class SynetConvolution16bNhwcGemm.
  • Error in Base implementation of class SynetMergedConvolution8i.

Test framework

New features
  • -wu command line option to set CPU warm up time in milliseconds.
  • Tests for verifying functionality of SynetMergedConvolution16b framework.

Infrastructure

Bug fixing
  • Errors in build_and_test_gcc section in Github actions script for CMake.
Assets 2
Loading

Simd v6.1.137

02 May 07:26
@ermig1979 ermig1979
v6.1.137
6b090e8
Compare
Choose a tag to compare
Simd v6.1.137

Algorithms

New features
  • AMX-BF16 (AVX-512VBMI) optimizations of function DescrIntCosineDistance.
  • AMX-BF16 (AVX-512VBMI, AMX-INT8) optimizations of function DescrIntCosineDistancesMxNa.
  • AMX-BF16 (AVX-512VBMI, AMX-INT8) optimizations of function DescrIntCosineDistancesMxNp.
  • API of SynetConvolution16b framework.
  • Base implementation of class SynetConvolution16bGemm.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512BF16 optimizations of class SynetConvolution16bNhwcGemm.
Improving
  • AVX-512VNNI optimizations of function DescrIntCosineDistance.
  • AVX-512VNNI optimizations of function DescrIntCosineDistancesMxNa.
  • AVX-512VNNI optimizations of function DescrIntCosineDistancesMxNp.

Test framework

New features
  • Tests for verifying functionality of SynetConvolution16b framework.
Assets 2
Loading

Simd v6.1.136

02 Apr 10:40
@ermig1979 ermig1979
v6.1.136
70225b5
Compare
Choose a tag to compare
Simd v6.1.136

Algorithms

New features
  • AMX-BF16 (AVX-512VBMI) optimizations of function ChangeColors.
  • AMX-BF16 (AVX-512VBMI) optimizations of function NormalizeHistogram.
Improving
  • AMX-BF16 optimizations of class SynetConvolution32fBf16NhwcGemm.
Bug fixing
  • Error in Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetConvolution32fBf16NhwcGemm.

Test framework

New features
  • Command line parameter to disable testing of some SIMD extensions.
Bug fixing
  • Error in test of function Nv12SaveAsJpegToMemory.
Assets 2
Loading

Simd v6.1.135

01 Mar 14:54
@ermig1979 ermig1979
v6.1.135
37ed82a
Compare
Choose a tag to compare
Simd v6.1.135

Algorithms

New features
  • Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetConvolution32fBf16NhwcGemm.
  • AMX-BF16 optimizations of function Float32ToBFloat16.
  • Support of SimdSynetUnaryOperation32fCos in function SynetUnaryOperation32f.
  • Support of SimdSynetUnaryOperation32fSin in function SynetUnaryOperation32f.
Bug fixing
  • Error in function SimdCpuInfo (wrong AMX-BF16 detection).
  • Error in AVX-512BF16 optimization of function Float32ToBFloat16.
  • Error in AMX initialization in function AmxBf16::SupportedByOS.
  • Crash in function AmxBf16::ConvolutionBf16NhwcConv_2.
  • Error in Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetMergedConvolution32fBf16Cdc.
  • Error in Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetMergedConvolution32fBf16Cd.
  • Error in Base implementation, SSE4.1, AVX2, AVX-512BW, AMX-BF16 optimizations of class SynetMergedConvolution32fBf16Dc.
Removing
  • AVX-512BF16 optimizations of function Float32ToBFloat16.
  • AVX-512BF16 optimizations of SynetConvolution32fBf16Nhwc.
  • AVX-512BF16 optimizations of class SynetMergedConvolution32fBf16Cdc.
  • AVX-512BF16 optimizations of class SynetMergedConvolution32fBf16Cd.
  • AVX-512BF16 optimizations of class SynetMergedConvolution32fBf16Dc.
  • Stopping of separate support of AVX-512BF16 extension (only together with AMX-BF16).

Test framework

Bug fixing
  • Error in test of SynetMergedConvolution32f framework.

Infrastructure

Removing
  • Avx512Bf16 project for MSVS-2022.
  • Avx512Bf16 project for MSVS-2019.
  • Avx512Bf16 project for MSVS-2015.
  • Avx512Bf16 project for MSVS-2017.
  • Avx512Bf16 project for CMake.
Assets 2
Loading

Simd v6.0.134

01 Feb 07:03
@ermig1979 ermig1979
v6.0.134
808e448
Compare
Choose a tag to compare
Simd v6.0.134

Algorithms

New features
  • SSE4.1 optimizations of ResizerFloatBilinear class.
Improving
  • Improve AVX2 optimizations of ResizerFloatBilinear class (AMD CPU).
  • Improve AVX2 optimizations of ResizerShortBilinear class (AMD CPU).
Bug fixing
  • MSVS compiler bug in file SimdAvx512bwYuvToBgraV2.
  • Linux, GCC-13 - crash in function SimdSynetInnerProduct32fForward.
  • MSVS compiler bug (Cmake, Windows for ARM64) with functions Extract4Sums.
  • GCC-9/10 - compiler error in AVX-512BW optimization of function YToGray.
  • GCC-9/10 - compiler error in AVX-512BW optimization of function GrayToY.
Replacing
  • Replace AVX optimizations to AVX2 for function CosineDistance32f.
  • Replace AVX optimizations to AVX2 for function Fill32f.
  • Replace AVX optimizations to AVX2 for ResizerFloatBilinear class.
  • Replace AVX optimizations to AVX2 for function SquaredDifferenceSum32f.
  • Replace AVX optimizations to AVX2 for function SquaredDifferenceKahanSum32f.
  • Replace AVX optimizations to AVX2 for function HogLiteFilterFeatures.
  • Replace AVX optimizations to AVX2 for function HogLiteResizeFeatures.
  • Replace AVX optimizations to AVX2 for function HogLiteCompressFeatures.
  • Replace AVX optimizations to AVX2 for function HogLiteFilterSeparable.
  • Replace AVX optimizations to AVX2 for function NeuralPooling2x2Max2x2.
  • Replace AVX optimizations to AVX2 for function NeuralProductSum.
  • Replace AVX optimizations to AVX2 for function NeuralAddVectorMultipliedByValue.
  • Replace AVX optimizations to AVX2 for function NeuralAddVector.
  • Replace AVX optimizations to AVX2 for function NeuralAddValue.
  • Replace AVX optimizations to AVX2 for function NeuralRoughSigmoid.
  • Replace AVX optimizations to AVX2 for function NeuralRoughSigmoid2.
  • Replace AVX optimizations to AVX2 for function NeuralRoughTanh.
  • Replace AVX optimizations to AVX2 for function NeuralDerivativeRelu.
  • Replace AVX optimizations to AVX2 for function NeuralDerivativeTanh.
  • Replace AVX optimizations to AVX2 for function NeuralDerivativeSigmoid.
  • Replace AVX optimizations to AVX2 for function NeuralUpdateWeights.
  • Replace AVX optimizations to AVX2 for function NeuralAdaptiveGradientUpdate.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution2x2Forward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution3x3Forward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution4x4Forward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution5x5Forward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution2x2Backward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution3x3Backward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution4x4Backward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution5x5Backward.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution2x2Sum.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution3x3Sum.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution4x4Sum.
  • Replace AVX optimizations to AVX2 for function NeuralAddConvolution5x5Sum.
  • Replace AVX optimizations to AVX2 for function NeuralConvolutionForward.
  • Replace AVX optimizations to AVX2 for function SynetAddBias.
  • Replace AVX optimizations to AVX2 for function SynetFusedLayerForward0.
  • Replace AVX optimizations to AVX2 for function SynetFusedLayerForward1.
  • Replace AVX optimizations to AVX2 for function SynetFusedLayerForward2.
  • Replace AVX optimizations to AVX2 for function SynetFusedLayerForward3.
  • Replace AVX optimizations to AVX2 for function SynetFusedLayerForward4.
  • Replace AVX optimizations to AVX2 for function SynetFusedLayerForward8.
  • Replace AVX optimizations to AVX2 for function SynetFusedLayerForward9.
  • Replace AVX optimizations to AVX2 for function SynetPoolingAverage.
  • Replace AVX optimizations to AVX2 for function SynetShuffleLayerForward.
  • Replace AVX optimizations to AVX2 for function SynetHardSigmoid32f.
  • Replace AVX optimizations to AVX2 for function SynetHswish32f.
  • Replace AVX optimizations to AVX2 for function SynetPreluLayerForward.
  • Replace AVX optimizations to AVX2 for function SynetRelu32f.
  • Replace AVX optimizations to AVX2 for function SynetRestrictRange32f.
  • Replace AVX optimizations to AVX2 for function WinogradKernel1x3Block1x4SetFilter.
  • Replace AVX optimizations to AVX2 for function WinogradKernel1x3Block1x4SetInput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel1x3Block1x4SetOutput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel1x5Block1x4SetFilter.
  • Replace AVX optimizations to AVX2 for function WinogradKernel1x5Block1x4SetInput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel1x5Block1x4SetOutput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel2x2Block2x2SetFilter.
  • Replace AVX optimizations to AVX2 for function WinogradKernel2x2Block2x2SetInput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel2x2Block2x2SetOutput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel2x2Block4x4SetFilter.
  • Replace AVX optimizations to AVX2 for function WinogradKernel2x2Block4x4SetInput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel2x2Block4x4SetOutput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block2x2SetFilter.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block2x2SetInput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block2x2SetOutput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block3x3SetFilter.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block3x3SetInput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block3x3SetOutput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block4x4SetFilter.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block4x4SetInput.
  • Replace AVX optimizations to AVX2 for function WinogradKernel3x3Block4x4SetOutput.
  • Replace AVX optimizations to AVX2 for function GemmPackA.
  • Replace AVX optimizations to AVX2 for function GemmPackB.
  • Replace AVX optimizations to AVX2 for function GemmScaleC.
  • Replace AVX optimizations to AVX2 for function SynetScaleLayerForward.
  • Replace AVX optimizations to AVX2 for function SynetInnerProductLayerForward.
  • Replace AVX optimizations to AVX2 for function SynetInnerProduct32fInit.
  • Replace AVX optimizations to AVX2 for function SynetEltwiseLayerForward.
  • Replace AVX optimizations to AVX2 for function SynetDeconvolution32fInit.
  • Replace AVX optimizations to AVX2 for function SynetMergedConvolution32fInit.
  • Replace AVX optimizations to AVX2 for SynetInnerProduct32fGemm class.
  • Replace AVX optimizations to AVX2 for SynetInnerProduct32fProd class.
  • Replace AVX optimizations to AVX2 for SynetDeconvolution32fGemmNN class.
  • Replace AVX optimizations to AVX2 for SynetDeconvolution32fNhwcDirect2x2 class.
  • Replace AVX optimizations to AVX2 for SynetMergedConvolution32fCdc class.
  • Replace AVX optimizations to AVX2 for SynetMergedConvolution32fCd class.
  • Replace AVX optimizations to AVX2 for SynetMergedConvolution32fDc class.
  • Replace AVX optimizations to AVX2 for SynetConvolution32fDepthwiseDotProduct class.
  • Replace AVX optimizations to AVX2 for SynetConvolution32fDirectNchw class.
  • Replace AVX optimizations to AVX2 for SynetConvolution32fDirectNhwc class.
  • Replace AVX optimizations to AVX2 for SynetConvolution32fNhwcDirect class.
  • Replace AVX optimizations to AVX2 for SynetConvolution32fGemmNT class.
  • Replace AVX optimizations to AVX2 for SynetConvolution32fGemmNТ class.
  • Replace AVX optimizations to AVX2 for SynetConvolution32fWinograd class.
  • Replace AVX optimizations to AVX2 for function SynetConvolution32fInit.
  • Replace AVX optimizations to AVX2 for function SynetMergedConvolution32fInit.
  • Replace AVX optimizations to AVX2 for function SynetMergedConvolution32fInit.
Removing
  • Base implementation, SSE4.1, AVX, AVX-512BW, NEON, VSX optimizations of function SvmSumLinear.
  • Stopping of separate support of AVX extension (only together with AVX2).
  • Base implementation, SSE4.1, AVX2, AVX-512BW, NEON, VMX optimizations of function EdgeBackgroundGrowRangeSlow.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, NEON, VMX optimizations of function EdgeBackgroundGrowRangeFast.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, NEON, VMX optimizations of function EdgeBackgroundIncrementCount.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, NEON, VMX optimizations of function EdgeBackgroundAdjustRange.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, NEON, VMX optimizations of function EdgeBackgroundAdjustRangeMasked.
  • Base implementation, SSE4.1, AVX2, AVX-512BW, NEON, VMX optimizations of function EdgeBackgroundShiftRange.
  • Base implementatio...
Read more
Assets 2
Loading