Simd v4.6.95

Algorithms

New features

• AVX2, AVX-512BW and AVX-512VNNI optimizations of SynetMergedConvolution8iCdc class.
• AVX2, AVX-512BW and AVX-512VNNI optimizations of SynetMergedConvolution8iCd class.
• AVX2, AVX-512BW and AVX-512VNNI optimizations of SynetMergedConvolution8iDc class.
• SSE4.1, AVX2, AVX-512BW optimizations of function SynetConvert8uTo32f.
• Base implementation, SSE2, SSSE3 AVX2, AVX-512BW optimizations of function AlphaPremultiply.
• Base implementation of function AlphaUnpremultiply.

Bug fixing

• GCC v10 compilation error in file SimdGemm.h.
• Error in IECompatible method of SynetMergedConvolution8i.

Test framework

New features

• Tests for verifying functionality of function AlphaPremultiply.
• Tests for verifying functionality of function AlphaUnpremultiply.

Documentation

Bug fixing

• There are no references to C++ wrappers in description of API functions.

Simd v4.6.94

Algorithms

New features

• Base implementation of SynetMergedConvolution8i class.
• Base implementation of function SynetConvert8uTo32f.
• Base implementation and SSE4.1 optimizations of SynetMergedConvolution8iCdc class.
• Base implementation and SSE4.1 optimizations of SynetMergedConvolution8iCd class.
• Base implementation and SSE4.1 optimizations of SynetMergedConvolution8iDc class.

Bug fixing

• Performance degradation in class Convolution32fNhwcDirect (weights size >> L3 cache).
• Performance degradation in class Convolution32fGemmNN (weights size >> L3 cache).

Test framework

New features

• Tests for verifying functionality of SynetMergedConvolution8i class.
• Tests for verifying functionality of function SynetConvert8uTo32f.

Documentation

Improving

• Improve structuring of Synet documentation.

Simd v4.6.93

Algorithms

New features

• Full support of SimdConvolutionActivationType in SynetConvolution8i class.
• Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512VNNI optimizations of SynetConvolution8iNhwcDepthwise class.
• Extend class MergedConvolution32f (2 merged convolutions).
• Base implementation, SSE2, AVX, AVX2, AVX-512F optimizations of MergedConvolution32fCd class.
• Base implementation, SSE2, AVX, AVX2, AVX-512F optimizations of MergedConvolution32fDc class.

Improving

• Reducing of compilation time and assembled size of Simd Library.

Renaming

• Class MergedConvolution32f to MergedConvolution32fCdc.

• Performance degradation in class Convolution32fNhwcDirect (dilation != 1).

Test framework

New features

• Tests for verifying functionality of class MergedConvolution32f (2 merged convolutions).

Simd v4.6.92

Algorithms

New features

• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function SynetAdd8i.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of function SynetInnerProduct8i.

Improving

• Reducing of compilation time and assembled size of Simd Library.

Bug fixing

• Error in SSE4.1, AVX2, AVX-512BW optimizations of SynetScale8i class (wrong alignment check).
• Error in performance annotation of SynetConvolution8i class.
• Compiler error in file SimdBaseSynetConvolution8i.cpp (for old compilers).
• Compiler errors in files SimdAvx2Synet.cpp, SimdAvx2SynetScale.cpp (WIN32, MSVS).

Test framework

New features

• Tests for verifying functionality of function SynetAdd8i.
• Tests for verifying functionality of function SynetInnerProduct8i.

Simd v4.6.91

Algorithms

New features

• Extend SimdSynetCompatibilityType enumeration.
• Add support of SimdSynetCompatibility8iNarrowed to Base implementation, SSE2, AVX2, AVX-512BW and NEON optimizations of function SynetConvert32fTo8u.
• Add support of SimdSynetCompatibility8iNarrowed to Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512VNNI and NEON optimizations of SynetConvolution8iNhwcDirect class.
• Add support of SimdConvolutionActivationPrelu to Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512VNNI and NEON optimizations of SynetConvolution8iNhwcDirect class.
• Base implementation, SSE4.1, AVX2, AVX-512BW optimizations of SynetScale8i class.

Improving

• Reducing of size of applications or shared libraries which use Simd as static library.

Bug fixing

• Error in class SynetConvolution8i (batch > 1).

Test framework

New features

• Tests for verifying functionality of SynetScale8i framework.

Simd v4.6.90

Algorithms

New features

• Rgb24 format in Frame structure.
• Rgb24 format in Convert function.
• Base implementation, SSS3, AVX2, AVX-512BW and NEON optimizations of function RgbToGray.
• Base implementation, SSS3, AVX2, AVX-512BW and NEON optimizations of function RgbToBgra.
• Base implementation, SSS3, AVX2, AVX-512BW and NEON optimizations of function BgraToRgb.
• AVX2 optimization of function BgraToBgr.
• Function LitterCpuCache.
• Base implementation, SSSE3, AVX2, AVX-512BW and NEON optimizations of function Yuv444pToRgb.
• Base implementation, SSSE3, AVX2, AVX-512BW and NEON optimizations of function Yuv422pToRgb.
• Base implementation, SSSE3, AVX2, AVX-512BW and NEON optimizations of function Yuv420pToRgb.

Improving

• NEON optimization of function BgrToGray.

Bug fixing

• Error in class SynetConvolution8i (group != 1).
• Wrong assert condition in SSE2, AVX, AVX2, AVX-512F and NEON optimization of class Convolution32fNhwcDirect.
• Compiler error when SIMD_AVX2_DISABLE macro is uncommented.
• Int32 overflow in function SynetConvolution8i::SetParams.

Test framework

New features

• Tests for verifying functionality of function RgbToGray.
• Tests for verifying functionality of function RgbToBgra.
• Tests for verifying functionality of function BgraToRgb.
• Tests for verifying functionality of function Yuv444pToRgb.
• Tests for verifying functionality of function Yuv422pToRgb.
• Tests for verifying functionality of function Yuv420pToRgb.

Simd v4.6.89

Algorithms

Bug fixing

• Microsoft Visual Studio 2013 compiler errors in files: SimdSynetConvolution8i.h, SimdSse2SynetConvolution32f.cpp, SimdAvx2Reduce.cpp.
• Buffer overrun in SSE4.1, AVX2, NEON optimizations of SynetConvolution8iNhwcDirect class.
• Visual Studio 2017 internal compiler error in function Avx512f::ConvolutionBiasAndActivation (Win32/Release).
• Compiler error in NEON optimization of class SynetConvolution8iNhwcDirect (ARM, 32-bit).
• Error in AVX2 optimization of function SynetScaleLayerForward.
• Error in base implementation of SquaredDifferenceKahanSum32f (Visual Studio 2017).
• Error in AVX-512BW optimization of class SynetConvolution8iNhwcDirect (Visual Studio 2017/2019, Release).
• Error in class SynetConvolution32fNhwcDirect (large parameters srcC and dstC).

Test framework

Bug fixing

• Microsoft Visual Studio 2013 compiler errors in files: TestTensor.h, TestSynetActivation.cpp.
• Test report is not generated if output directory is not exists.
• Error in test SynetConvert32fTo8uAutoTest.

Infrastructure

New features

• Script to test Simd compiled with different version of Microsoft Visual Studio.
• New structure of Microsoft Visual Studio 2019 project files.

Removing

• Remove project files of Microsoft Visual Studio 2012.

Simd v4.6.88

Algorithms

New features

• AVX-512VNNI extension support.
• AVX2, AVX-512BW, AVX-512VNNI and NEON optimizations of SynetConvolution8iNhwcDirect class.
• Base implementation and SSE4.1, AVX2 AVX-512BW and NEON optimizations of function SynetPoolingForwardMax8u.

Renaming

• SynetPoolingForwardMax to SynetPoolingForwardMax32f.

Improving

• SSE4.1 optimization of SynetConvolution8iNhwcDirect class.
• SSE2, AVX, AVX2, AVX-512F and NEON optimizations of SynetConvolution32fNhwcDirect class.

Bug fixing

• Microsoft Visual Studio 2015 compiler error in function SynetConvert32fTo8u.
• Degradation of performance of AVX2 code.
• Microsoft Visual Studio compiler error in function Extract64i (32-bit mode).

Test framework

New features

• Tests for verifying functionality of function SynetPoolingForwardMax8u.

Simd v4.5.87

Algorithms

New features

• Add parameter of bitwise compatibility of function SynetScaleLayerForward and Inference Engine.
• Add parameter 'type' to function SynetShuffleLayerForward.
• Base implementation, SSE2, AVX2, AVX-512BW amd NEON optimizations of function SynetConvert32fTo8u.
• SimdSynetCompatibilityType enumeration.
• Base implementation of SynetConvolution8iGemmNN class.
• Base implementation and SSE4.1 optimization of SynetConvolution8iNhwcDirect class.

Renaming

• SimdSynetConvertImage to SimdSynetReorderImage.
• SimdSynetConvertFilter to SimdSynetReorderFilter.

Test framework

New features

• A new commandline test parameter -c - a number of channels in test image for performance testing.
• A new commandline test parameter -mt - a minimal test execution time (in milliseconds).
• Tests for verifying functionality of SynetConvolution8i framework.
• Tests for verifying functionality of function SynetConvert32fTo8u.

Documentation

Bug fixing

• Error in description of method Detection::LoadStringXml.

Simd v4.5.86

Algorithms

New features

• SimdResizeMethodInferenceEngineInterp method in Resizer framework.

Improving

• Performance of Convolution32f framework (NHWC format, kernel=3x3, stride=1x1, large H and W).
• Performance of AVX-512F and NEON optimizations of function GemmPackA.
• Performance of Convolution32f framework (NHWC format, GemmNN method).
• Performance of SSE2, AVX, AVX2, AVX-512F and NEON optimizations of Convolution32f framework (NHWC format, NhwcDirect method, kernel=1x1).
• Performance of AVX-512F optimization of MergedConvolution32f framework (input convolution).
• Performance of AVX2 and AVX-512F optimizations of MergedConvolution32f framework (output convolution).
• Performance of Convolution32f framework (stride > 1).
• Performance of AVX-512F optimization of Gemm32fNN function (add 6x64 and 6x48 micro kernel).

Bug fixing

• Error in AVX-512F optimization of function WinogradKernel3x3Block2x2SetOutput (NCHW format).
• Error in SSE, AVX, AVX-512F and NEON optimizations of function SynetPoolingForwardAverage (NHWC format).
• Error in AVX-512F optimization of function SynetInnerProductLayerForward.
• Error in AVX, AVX2 and AVX-512F optimizations of function Gemm32fNT.
• Error in function WinogradKernel3x3Block4x4SetInput (padX != padY != padW != padH).
• Error in debug FLOPS annotation of Deconvolution32f framework.
• MergedConvolution32f framework doesn't work with stride == 3.