SynetScale16b Class Reference

The SynetScale16b class is a C++ wrapper of FP32/BF16 scale and bias. More...

#include <SimdSynet.hpp>

Public Member Functions
	SynetScale16b ()
virtual	~SynetScale16b ()
SIMD_INLINE void	Init (size_t channels, size_t spatial, SimdTensorDataType srcType, SimdTensorDataType dstType, SimdTensorFormatType format, SimdBool norm, SimdBool bias)
SIMD_INLINE bool	Enable () const
SIMD_INLINE void	Forward (const uint8_t *src, const float *norm, const float *bias, uint8_t *dst)
SIMD_INLINE void	Clear ()

Detailed Description

The SynetScale16b class is a C++ wrapper of FP32/BF16 scale and bias.

The class wraps C API functions SimdSynetScale16bInit and SimdSynetScale16bForward. It applies a per-channel affine transformation of FP32 or BF16 tensors. BF16 values are converted to FP32 before the operation and converted back after it when the corresponding tensor type is SimdTensorData16b. The context applies dst = src*norm + bias, dst = src*norm or dst = src + bias depending on the norm and bias flags:

value = ConvertToFloat(src);
if(norm) value *= norm[c];
if(bias) value += bias[c];
dst = ConvertFromFloat(value);

Algorithm's details (example for NCHW format):

for(c = 0; c < channels; ++c)
    for(s = 0; s < spatial; ++s)
    {
        value = ConvertToFloat(src[c*spatial + s]);
        if(norm) value *= norm[c];
        if(bias) value += bias[c];
        dst[c*spatial + s] = ConvertFromFloat(value);
    }

The current implementation creates a context for FP32 or BF16 input and output tensor types and SimdTensorFormatNchw or SimdTensorFormatNhwc tensor format. At least one of the norm and bias flags must be enabled. Call Init() before Forward(). Use Enable() to check that a context was created. The context is released by Clear() or by the destructor.

Using example:

#include "Simd/SimdSynet.hpp"

int main()
{
    const size_t channels = 4, spatial = 16;
    std::vector<float> src(channels * spatial, 1.0f), dst(channels * spatial, 0.0f);
    std::vector<float> norm(channels, 0.5f), bias(channels, 0.1f);

    Simd::SynetScale16b scale16b;
    scale16b.Init(channels, spatial, SimdTensorData32f, SimdTensorData32f, SimdTensorFormatNhwc, SimdTrue, SimdTrue);
    if (scale16b.Enable())
        scale16b.Forward((const uint8_t*)src.data(), norm.data(), bias.data(), (uint8_t*)dst.data());

    return 0;
}

Constructor & Destructor Documentation

SynetScale16b()

SynetScale16b

(

)

Creates a new empty SynetScale16b class.

~SynetScale16b()

virtual ~SynetScale16b ( )

virtual

SynetScale16b class destructor. Releases internal context.

Member Function Documentation

Init()

SIMD_INLINE void Init	(	size_t	channels,
		size_t	spatial,
		SimdTensorDataType	srcType,
		SimdTensorDataType	dstType,
		SimdTensorFormatType	format,
		SimdBool	norm,
		SimdBool	bias
	)

Initializes (or re-initializes) an FP32/BF16 scale and bias context.

Creates an internal context with using of function SimdSynetScale16bInit. The context is recreated only if channel count, spatial size, tensor types, tensor format or norm/bias flags were changed.

Note

This function is a C++ wrapper for function SimdSynetScale16bInit.

Parameters

[in]	channels	- a number of channels in input and output tensors.
[in]	spatial	- a spatial size (height*width) of input and output tensors.
[in]	srcType	- an input data type. It can be SimdTensorData32f or SimdTensorData16b.
[in]	dstType	- an output data type. It can be SimdTensorData32f or SimdTensorData16b.
[in]	format	- a format of input and output tensors. It can be SimdTensorFormatNchw or SimdTensorFormatNhwc.
[in]	norm	- a flag of presence of per-channel multiplication by norm.
[in]	bias	- a flag of presence of per-channel addition of bias.

Enable()

SIMD_INLINE bool Enable

(

)

const

Checks that the internal scale context was created.

Returns

true if the context exists and Forward() can be called.

Forward()

SIMD_INLINE void Forward	(	const uint8_t *	src,
		const float *	norm,
		const float *	bias,
		uint8_t *	dst
	)

Performs forward propagation of FP32/BF16 scale and bias algorithm.

The function applies per-channel scale and/or bias according to the flags stored in the context created by Init() and converts between FP32 and BF16 according to the types stored in that context.

Note

This function is a C++ wrapper for function SimdSynetScale16bForward.

Parameters

[in]	src	- a pointer to input tensor data. Its type is defined by parameter srcType of Init().
[in]	norm	- a pointer to FP32 array with per-channel scale coefficients. Can be NULL if norm flag is SimdFalse.
[in]	bias	- a pointer to FP32 array with per-channel bias coefficients. Can be NULL if bias flag is SimdFalse.
[out]	dst	- a pointer to output tensor data. Its type is defined by parameter dstType of Init().

Clear()

SIMD_INLINE void Clear

(

)

Releases internal context and clears stored scale parameters.