Types used in Synet Framework. More...

Data Structures
struct	SimdConvolutionParameters

Typedefs
typedef struct SimdConvolutionParameters	SimdConvolutionParameters

Enumerations
enum	SimdConvolutionActivationType { SimdConvolutionActivationIdentity = 0 , SimdConvolutionActivationRelu , SimdConvolutionActivationLeakyRelu , SimdConvolutionActivationRestrictRange , SimdConvolutionActivationPrelu , SimdConvolutionActivationElu , SimdConvolutionActivationHswish , SimdConvolutionActivationMish , SimdConvolutionActivationHardSigmoid , SimdConvolutionActivationSwish , SimdConvolutionActivationGelu }

enum	SimdSynetCompatibilityType { SimdSynetCompatibilityDefault = 0 , SimdSynetCompatibilityFmaUse = 0 , SimdSynetCompatibilityFmaNoTail = 1 , SimdSynetCompatibilityFmaAvoid = 2 , SimdSynetCompatibilityFmaMask = 3 , SimdSynetCompatibility8iPrecise = 0 , SimdSynetCompatibility8iOverflow = 4 , SimdSynetCompatibility8iNarrowed = 8 , SimdSynetCompatibility8iMask = 12 , SimdSynetCompatibility16bfAvoid = 0 , SimdSynetCompatibility16bfHard = 16 , SimdSynetCompatibility16bfSoft = 32 , SimdSynetCompatibility16bfMask = 48 , SimdSynetCompatibility16fpAvoid = 0 , SimdSynetCompatibility16fpHard = 64 , SimdSynetCompatibility16fpSoft = 128 , SimdSynetCompatibility16fpMask = 192 }

enum	SimdSynetEltwiseOperationType { SimdSynetEltwiseOperationProduct , SimdSynetEltwiseOperationSum , SimdSynetEltwiseOperationMax , SimdSynetEltwiseOperationMin }

enum	SimdSynetUnaryOperation32fType { SimdSynetUnaryOperation32fAbs , SimdSynetUnaryOperation32fCeil , SimdSynetUnaryOperation32fCos , SimdSynetUnaryOperation32fErf , SimdSynetUnaryOperation32fExp , SimdSynetUnaryOperation32fFloor , SimdSynetUnaryOperation32fLog , SimdSynetUnaryOperation32fNeg , SimdSynetUnaryOperation32fNot , SimdSynetUnaryOperation32fRcp , SimdSynetUnaryOperation32fRound , SimdSynetUnaryOperation32fRsqrt , SimdSynetUnaryOperation32fSign , SimdSynetUnaryOperation32fSin , SimdSynetUnaryOperation32fSqrt , SimdSynetUnaryOperation32fTanh , SimdSynetUnaryOperation32fZero }

enum	SimdTensorFormatType { SimdTensorFormatUnknown = -1 , SimdTensorFormatNchw , SimdTensorFormatNhwc }

enum	SimdTensorDataType { SimdTensorDataUnknown = -1 , SimdTensorData32f , SimdTensorData32i , SimdTensorData8i , SimdTensorData8u , SimdTensorData64i , SimdTensorData64u , SimdTensorDataBool , SimdTensorData16b , SimdTensorData16f }

Detailed Description

Types used in Synet Framework.

Typedef Documentation

◆ SimdConvolutionParameters

typedef struct SimdConvolutionParameters SimdConvolutionParameters

Describes convolution and deconvolution geometry, tensor types and activation.

This structure is passed to convolution, deconvolution and merged-convolution initialization functions. For convolution, destination spatial size must satisfy:

dstH = (srcH + padY + padH - (dilationY*(kernelY - 1) + 1)) / strideY + 1
dstW = (srcW + padX + padW - (dilationX*(kernelX - 1) + 1)) / strideX + 1

For deconvolution, destination spatial size must satisfy:

dstH = strideY*(srcH - 1) + dilationY*(kernelY - 1) + 1 - padY - padH
dstW = strideX*(srcW - 1) + dilationX*(kernelX - 1) + 1 - padX - padW

The weight tensor contains kernelY*kernelX*srcC*dstC/group elements. A depthwise convolution is represented by group == srcC == dstC.

Enumeration Type Documentation

◆ SimdConvolutionActivationType

enum SimdConvolutionActivationType

Describes activation functions used by Synet convolution, deconvolution, inner product, merged convolution, quantized convolution and quantized add APIs.

Activations are applied after bias/normalization/addition as documented by the corresponding function. Parameters are passed through the params or actParams arguments of those functions. A NULL parameter pointer is valid only for activations that do not use parameters.

Enumerator
SimdConvolutionActivationIdentity	Identity activation: no transformation is applied. dst[i] = src[i];
SimdConvolutionActivationRelu	ReLU activation function. It does not use parameters. dst[i] = Max(0, src[i]);
SimdConvolutionActivationLeakyRelu	Leaky ReLU activation function. It has one parameter: negative slope (params[0]). dst[i] = src[i] > 0 ? src[i] : slope*src[i];
SimdConvolutionActivationRestrictRange	RestrictRange activation function. It clamps the value to the interval [lower, upper], where lower = params[0] and upper = params[1]. dst[i] = Min(Max(lower, src[i]), upper);
SimdConvolutionActivationPrelu	PReLU activation function. It uses one negative slope per destination channel. For a convolution output with dstC channels and spatial size n = dstHdstW, params[c] is used for channel c. dst[cn + j] = src[cn + j] > 0 ? src[cn + j] : params[c]src[cn + j];
SimdConvolutionActivationElu	ELU activation function. It has one parameter: alpha (params[0]). dst[i] = src[i] >= 0 ? src[i] : alpha*(Exp(src[i]) - 1);
SimdConvolutionActivationHswish	H-Swish (https://arxiv.org/pdf/1905.02244.pdf) activation function. It has two parameters: shift (params[0]) and scale (params[1]). Typical values are shift = 3 and scale = 1/6. dst[i] = Max(Min(src[i], shift) + shift, 0)scalesrc[i];
SimdConvolutionActivationMish	Mish (https://arxiv.org/abs/1908.08681) activation function. It has one parameter: threshold (params[0]). Values greater than threshold are returned unchanged. dst[i] = src[i] > threshold ? src[i] : src[i]*(1 - 2/(Square(Exp(src[i]) + 1) + 1));
SimdConvolutionActivationHardSigmoid	HardSigmoid (https://pytorch.org/docs/stable/generated/torch.nn.Hardsigmoid.html) activation function. It has two parameters: scale (params[0]) and shift (params[1]). dst[i] = Max(0, Min(src[i] * scale + shift, 1));
SimdConvolutionActivationSwish	Swish (https://en.wikipedia.org/wiki/Swish_function) activation function. It has one parameter: slope (params[0]). dst[i] = src[i]/(1 + Exp(-slope*src[i]));
SimdConvolutionActivationGelu	GELU (https://en.wikipedia.org/wiki/Activation_function) activation function. It does not use parameters. dst[i] = src[i] * (1 + erf(src[i]/sqrt(2))) / 2;

◆ SimdSynetCompatibilityType

enum SimdSynetCompatibilityType

Describes Synet calculation compatibility flags.

Values are grouped into independent bit fields and can be combined with bitwise OR. The FMA field controls use of fused multiply-add instructions, the 8-bit integer field controls quantized multiplication/range policy, and the BF16/FP16 fields request use of reduced-precision formats. Masks are provided to extract each field.

This type is used in functions such as SimdSynetAdd8i, SimdSynetScaleLayerForward, SimdSynetConvert32fTo8u, SimdSynetConvert8uTo32f, SimdSynetInnerProduct8i, SimdSynetScale8iInit, SimdSynetConvolution16bInit, SimdSynetConvolution8iInit, SimdSynetMergedConvolution16bInit and SimdSynetMergedConvolution8iInit.

Enumerator
SimdSynetCompatibilityDefault	Default compatibility value: use the fastest supported policy for each field.
SimdSynetCompatibilityFmaUse	Allow FMA instructions where implementations support them.
SimdSynetCompatibilityFmaNoTail	Avoid FMA instructions at row tails for stricter reproducibility.
SimdSynetCompatibilityFmaAvoid	Avoid FMA instructions completely.
SimdSynetCompatibilityFmaMask	Mask used to extract FMA policy bits.
SimdSynetCompatibility8iPrecise	Use precise 8-bit integer multiplication (VNNI or 16-bit emulation).
SimdSynetCompatibility8iOverflow	Allow 16-bit intermediate overflow in 8-bit integer multiplication paths.
SimdSynetCompatibility8iNarrowed	Use narrowed input ranges (signed [-90..90], unsigned [0..180]) to avoid 16-bit overflow.
SimdSynetCompatibility8iMask	Mask used to extract 8-bit integer multiplication policy bits.
SimdSynetCompatibility16bfAvoid	Do not request BFloat16 (Brain Floating Point) internal computation.
SimdSynetCompatibility16bfHard	Use BFloat16 internal computation only when hardware support exists.
SimdSynetCompatibility16bfSoft	Use BFloat16 internal computation with software emulation when hardware support is absent.
SimdSynetCompatibility16bfMask	Mask used to extract BFloat16 policy bits.
SimdSynetCompatibility16fpAvoid	Do not request 16-bit floating point (Half Precision) internal computation.
SimdSynetCompatibility16fpHard	Use 16-bit floating point internal computation only when hardware support exists.
SimdSynetCompatibility16fpSoft	Use 16-bit floating point internal computation with software emulation when hardware support is absent.
SimdSynetCompatibility16fpMask	Mask used to extract 16-bit floating point policy bits.

◆ SimdSynetEltwiseOperationType

enum SimdSynetEltwiseOperationType

Describes operation type used by SimdSynetEltwiseLayerForward.

The function combines at least two equally sized FP32 arrays element by element. The weight array is used only by SimdSynetEltwiseOperationSum.

Enumerator
SimdSynetEltwiseOperationProduct	Product of corresponding elements from all input arrays.
SimdSynetEltwiseOperationSum	Weighted sum of corresponding elements from all input arrays.
SimdSynetEltwiseOperationMax	Maximum of corresponding elements from all input arrays.
SimdSynetEltwiseOperationMin	Minimum of corresponding elements from all input arrays.

◆ SimdSynetUnaryOperation32fType

enum SimdSynetUnaryOperation32fType

Describes unary operation type used by SimdSynetUnaryOperation32f.

Each operation is applied independently to every 32-bit floating point element of the input array.

Enumerator
SimdSynetUnaryOperation32fAbs	Absolute value: `dst = Abs(src)`.
SimdSynetUnaryOperation32fCeil	Ceiling: `dst = Ceil(src)`.
SimdSynetUnaryOperation32fCos	Cosine: `dst = Cos(src)`.
SimdSynetUnaryOperation32fErf	Error function: `dst = Erf(src)`.
SimdSynetUnaryOperation32fExp	Exponent: `dst = Exp(src)`.
SimdSynetUnaryOperation32fFloor	Floor: `dst = Floor(src)`.
SimdSynetUnaryOperation32fLog	Natural logarithm: `dst = Log(src)`.
SimdSynetUnaryOperation32fNeg	Negation: `dst = -src`.
SimdSynetUnaryOperation32fNot	Bitwise NOT applied to the IEEE-754 representation of the source value.
SimdSynetUnaryOperation32fRcp	Reciprocal: `dst = 1/src`.
SimdSynetUnaryOperation32fRound	Round to nearest integer value in FP32 representation.
SimdSynetUnaryOperation32fRsqrt	Reciprocal square root: `dst = 1/Sqrt(src)`.
SimdSynetUnaryOperation32fSign	Sign function: -1 for negative values, 0 for zero and 1 for positive values.
SimdSynetUnaryOperation32fSin	Sine: `dst = Sin(src)`.
SimdSynetUnaryOperation32fSqrt	Square root: `dst = Sqrt(src)`.
SimdSynetUnaryOperation32fTanh	Hyperbolic tangent: `dst = Tanh(src)`.
SimdSynetUnaryOperation32fZero	Zeroing: `dst = 0`.

◆ SimdTensorFormatType

enum SimdTensorFormatType

Describes tensor memory layout used by Synet functions.

Most functions use 4D tensors with dimensions batch (N), channels (C), height (H) and width (W). Some shape-based helper functions accept SimdTensorFormatUnknown when layout is irrelevant.

Enumerator
SimdTensorFormatUnknown	Unknown or layout-independent tensor format.
SimdTensorFormatNchw	NCHW layout: offset = ((nC + c)H + h)*W + w.
SimdTensorFormatNhwc	NHWC layout: offset = ((nH + h)W + w)*C + c.

◆ SimdTensorDataType

enum SimdTensorDataType

Describes tensor element data type used by Synet functions.

The value defines interpretation and size of each tensor element. Reduced precision values SimdTensorData16b and SimdTensorData16f are stored in 16-bit containers.

Enumerator
SimdTensorDataUnknown	Unknown tensor data type.
SimdTensorData32f	32-bit floating point (single precision).
SimdTensorData32i	32-bit signed integer.
SimdTensorData8i	8-bit signed integer.
SimdTensorData8u	8-bit unsigned integer.
SimdTensorData64i	64-bit signed integer.
SimdTensorData64u	64-bit unsigned integer.
SimdTensorDataBool	Boolean value stored in one byte.
SimdTensorData16b	16-bit BFloat16 (Brain Floating Point) stored in uint16_t.
SimdTensorData16f	16-bit floating point (Half Precision) stored in uint16_t/int16_t.

Simd Library Documentation.

Data Structures

Typedefs

Enumerations