Functions to accelerate conversion in Synet Framework. More...

Functions
SIMD_API void	SimdSynetConvert32fTo8u (const float src, size_t batch, size_t channels, size_t height, size_t width, SimdTensorFormatType format, const float scale, const float shift, uint8_t dst, SimdSynetCompatibilityType compatibility)
	Converts an FP32 tensor to a UINT8 tensor using per-channel scale and shift. More...

SIMD_API void	SimdSynetConvert8uTo32f (const uint8_t src, size_t batch, size_t channels, size_t height, size_t width, SimdTensorFormatType format, const float scale, const float shift, float dst, SimdSynetCompatibilityType compatibility)
	Converts a UINT8 tensor to an FP32 tensor using per-channel scale and shift. More...

SIMD_API void	SimdSynetSetInput (const uint8_t src, size_t width, size_t height, size_t stride, SimdPixelFormatType srcFormat, const float lower, const float upper, float dst, size_t channels, SimdTensorFormatType dstFormat)
	Converts an 8-bit image to normalized FP32 neural-network input tensor. More...

template<template< class > class A>
SIMD_INLINE void	SynetSetInput (const View< A > &src, const float lower, const float upper, float *dst, size_t channels, SimdTensorFormatType format, bool isRgb=false)
	Sets image to the input of neural network of Synet Framework. More...

Detailed Description

Functions to accelerate conversion in Synet Framework.

Function Documentation

◆ SimdSynetConvert32fTo8u()

void SimdSynetConvert32fTo8u	(	const float *	src,
		size_t	batch,
		size_t	channels,
		size_t	height,
		size_t	width,
		SimdTensorFormatType	format,
		const float *	scale,
		const float *	shift,
		uint8_t *	dst,
		SimdSynetCompatibilityType	compatibility
	)

Converts an FP32 tensor to a UINT8 tensor using per-channel scale and shift.

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

upper = isNarrowed(compatibility) ? 180 : 255;
for(b = 0; b < batch; ++b)
    for(c = 0; c < channels; ++c)
        for(h = 0; h < height; ++h)
            for(w = 0; w < width; ++w)
            {
                offs = ((b*channels + c)*height + h)*width + w;
                dst[offs] = restrict(round(src[offs]*scale[c] + shift[c]), 0, upper);
            }

For NHWC tensor format the same calculation uses offset ((b*height + h)*width + w)*channels + c.

Note: This function is used in Synet Framework.

Parameters

[in]	src	- a pointer to the FP32 input tensor.
[in]	batch	- a batch size of input and output tensors.
[in]	channels	- a number of channels in input and output tensors.
[in]	height	- a height of input and output tensors.
[in]	width	- a width of input and output tensors.
[in]	format	- a format of input and output tensors. Can be NCHW or NHWC.
[in]	scale	- a pointer to the 32-bit float array with per-channel scale coefficients.
[in]	shift	- a pointer to the 32-bit float array with per-channel shift coefficients.
[out]	dst	- a pointer to the UINT8 output tensor.
[in]	compatibility	- calculation compatibility flags. When narrowed 8-bit mode is active, output is limited to [0, 180], otherwise to [0, 255].

◆ SimdSynetConvert8uTo32f()

void SimdSynetConvert8uTo32f	(	const uint8_t *	src,
		size_t	batch,
		size_t	channels,
		size_t	height,
		size_t	width,
		SimdTensorFormatType	format,
		const float *	scale,
		const float *	shift,
		float *	dst,
		SimdSynetCompatibilityType	compatibility
	)

Converts a UINT8 tensor to an FP32 tensor using per-channel scale and shift.

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

for(b = 0; b < batch; ++b)
    for(c = 0; c < channels; ++c)
        for(h = 0; h < height; ++h)
            for(w = 0; w < width; ++w)
            {
                offs = ((b*channels + c)*height + h)*width + w;
                dst[offs] = src[offs]*scale[c] + shift[c];
            }

For NHWC tensor format the same calculation uses offset ((b*height + h)*width + w)*channels + c.

Note: This function is used in Synet Framework.

Parameters

[in]	src	- a pointer to the UINT8 input tensor.
[in]	batch	- a batch size of input and output tensors.
[in]	channels	- a number of channels in input and output tensors.
[in]	height	- a height of input and output tensors.
[in]	width	- a width of input and output tensors.
[in]	format	- a format of input and output tensors. Can be NCHW or NHWC.
[in]	scale	- a pointer to the 32-bit float array with per-channel scale coefficients.
[in]	shift	- a pointer to the 32-bit float array with per-channel shift coefficients.
[out]	dst	- a pointer to the FP32 output tensor.
[in]	compatibility	- calculation compatibility flags.

◆ SimdSynetSetInput()

void SimdSynetSetInput	(	const uint8_t *	src,
		size_t	width,
		size_t	height,
		size_t	stride,
		SimdPixelFormatType	srcFormat,
		const float *	lower,
		const float *	upper,
		float *	dst,
		size_t	channels,
		SimdTensorFormatType	dstFormat
	)

Converts an 8-bit image to normalized FP32 neural-network input tensor.

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

for(c = 0; c < channels; ++c)
    for(y = 0; y < height; ++y)
        for(x = 0; x < width; ++x)
            dst[(c*height + y)*width + x] = src[stride*y + x*4 + c]*(upper[c] - lower[c])/255 + lower[c];

Each output value is mapped from [0, 255] to [lower[c], upper[c]]. Note that there are following relationships:

upper[c] = (1 - mean[c]) / std[c];
lower[c] = - mean[c] / std[c];

Also this algorithm assumes that channel order of output tensor is BGR. In case of RGB channel order you need to change parameter srcFormat: SimdPixelFormatBgr24 <-> SimdPixelFormatRgb24, SimdPixelFormatBgra32 <-> SimdPixelFormatRgba32. The actual pixel data of the input image does not need to be changed.

Note: This function has a C++ wrappers: Simd::SynetSetInput(const View & src, const float * lower, const float * upper, float * dst, size_t channels, SimdTensorFormatType format, bool isRgb = false).

Parameters

[in]	src	- a pointer to pixels data of input image.
[in]	width	- a width of input image and output image tensor.
[in]	height	- a height of input image and output image tensor.
[in]	stride	- a row size of input image.
[in]	srcFormat	- a pixel format of input image. There are supported following pixel formats: SimdPixelFormatGray8, SimdPixelFormatBgr24, SimdPixelFormatBgra32, SimdPixelFormatRgb24, SimdPixelFormatRgba32.
[in]	lower	- a pointer to lower bounds of output tensor values. The size of the array must be equal to channels.
[in]	upper	- a pointer to upper bounds of output tensor values. The size of the array must be equal to channels.
[out]	dst	- a pointer to the output 32-bit float image tensor.
[in]	channels	- a number of channels in the output image tensor. It can be 1 or 3.
[in]	dstFormat	- a format of output image tensor. There are supported following tensor formats: SimdTensorFormatNchw, SimdTensorFormatNhwc.

◆ SynetSetInput()

void SynetSetInput	(	const View< A > &	src,
		const float *	lower,
		const float *	upper,
		float *	dst,
		size_t	channels,
		SimdTensorFormatType	format,
		bool	isRgb = `false`
	)

Sets image to the input of neural network of Synet Framework.

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

for(c = 0; c < channels; ++c)
    for(y = 0; y < src.height; ++y)
        for(x = 0; x < src.width; ++x)
            dst[(c*height + y)*width + x] = src.data[src.stride*y + src.width*4 + c]*(upper[c] - lower[c])/255 + lower[c];

Note that there are following relationships:

upper[c] = (1 - mean[c]) / std[c];
lower[c] = - mean[c] / std[c];

Note: This function is a C++ wrapper for function SimdSynetSetInput.

Parameters

[in]	src	- an input image.There are supported following image formats: View<A>::Gray8, View<A>::Bgr24, View<A>::Bgra32, View<A>::Rgb24, View<A>::Rgba32.
[in]	lower	- a pointer to the array with lower bound of values of the output tensor. The size of the array have to correspond number of channels in the output image tensor.
[in]	upper	- a pointer to the array with upper bound of values of the output tensor. The size of the array have to correspond number of channels in the output image tensor.
[out]	dst	- a pointer to the output 32-bit float image tensor.
[in]	channels	- a number of channels in the output image tensor. It can be 1 or 3.
[in]	format	- a format of output image tensor. There are supported following tensor formats: SimdTensorFormatNchw, SimdTensorFormatNhwc.
[in]	isRgb	- is channel order of output tensor is RGB or BGR. Its default value is false.

Simd Library Documentation.

Functions

Detailed Description

Function Documentation

◆ SimdSynetConvert32fTo8u()

◆ SimdSynetConvert8uTo32f()

◆ SimdSynetSetInput()

◆ SynetSetInput()