ScaleLayer functions

Functions to accelerate layer scale in Synet Framework. More...

Functions
SIMD_API void *	SimdSynetScale16bInit (size_t channels, size_t spatial, SimdTensorDataType srcType, SimdTensorDataType dstType, SimdTensorFormatType format, SimdBool norm, SimdBool bias)
	Initializes FP32/BF16 scale and bias algorithm. More...
SIMD_API void	SimdSynetScale16bForward (void *context, const uint8_t *src, const float *norm, const float *bias, uint8_t *dst)
	Performs forward propagation of FP32/BF16 scale and bias algorithm. More...
SIMD_API void	SimdSynetScaleLayerForward (const float *src, const float *scale, const float *bias, size_t channels, size_t height, size_t width, float *dst, SimdTensorFormatType format, SimdSynetCompatibilityType compatibility)
	Performs forward propagation of FP32 ScaleLayer. More...

Detailed Description

Functions to accelerate layer scale in Synet Framework.

Function Documentation

SimdSynetScale16bInit()

void * SimdSynetScale16bInit	(	size_t	channels,
		size_t	spatial,
		SimdTensorDataType	srcType,
		SimdTensorDataType	dstType,
		SimdTensorFormatType	format,
		SimdBool	norm,
		SimdBool	bias
	)

Initializes FP32/BF16 scale and bias algorithm.

The context applies per-channel operation dst = src*norm + bias, dst = src*norm or dst = src + bias depending on norm and bias flags. Source and destination tensors can be FP32 or BF16.

Parameters

[in]	channels	- a number of channels in the (input/output) image tensor.
[in]	spatial	- a spatial size (height*width) of (input/output) image tensor.
[in]	srcType	- a type of input tensor. It can be SimdTensorData32f or SimdTensorData16b.
[in]	dstType	- a type of output tensor. It can be SimdTensorData32f or SimdTensorData16b.
[in]	format	- a format of input/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.

Returns

a pointer to scale context. On error it returns NULL. It must be released with using of function SimdRelease. This pointer is used in function SimdSynetScale16bForward.

SimdSynetScale16bForward()

void SimdSynetScale16bForward	(	void *	context,
		const uint8_t *	src,
		const float *	norm,
		const float *	bias,
		uint8_t *	dst
	)

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

Algorithm's details:

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

Parameters

[in]	context	- a pointer to scale context. It must be created by function SimdSynetScale16bInit and released by function SimdRelease.
[in]	src	- a pointer to input tensor data. Its type is defined by parameter srcType of SimdSynetScale16bInit.
[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 SimdSynetScale16bInit.

SimdSynetScaleLayerForward()

void SimdSynetScaleLayerForward	(	const float *	src,
		const float *	scale,
		const float *	bias,
		size_t	channels,
		size_t	height,
		size_t	width,
		float *	dst,
		SimdTensorFormatType	format,
		SimdSynetCompatibilityType	compatibility
	)

Performs forward propagation of FP32 ScaleLayer.

Algorithm's details (example for NCHW tensor format):

for(c = 0; c < channels; ++c)
    for(h = 0; h < height; ++h)
        for(w = 0; w < width; ++w)
            dst[(c*height + h)*width + w] = src[(c*height + h)*width + w]*scale[c] + (bias ? bias[c] : 0);

Note

This function is used in Synet Framework.

Parameters

[in]	src	- a pointer to the 32-bit float array with input image tensor. The size of the array is equal to channels*height*width.
[in]	scale	- a pointer to the 32-bit float array with scale coefficients. The size of the array is equal to channels.
[in]	bias	- a pointer to the 32-bit float array with bias coefficients. The size of the array is equal to channels. Can be NULL.
[in]	channels	- a number of channels in the (input/output) image tensor.
[in]	height	- a height of (input/output) image tensor.
[in]	width	- a width of (input/output) image tensor.
[out]	dst	- a pointer to the 32-bit float array with output image tensor. The size of the array is equal to channels*height*width.
[in]	format	- a format of input and output image tensors. It can be SimdTensorFormatNchw or SimdTensorFormatNhwc.
[in]	compatibility	- reserved compatibility flags. Current implementation does not use this parameter.