Quantized addition framework

A framework to accelerate Quantized addition in Synet Framework. More...

Functions
SIMD_API void *	SimdSynetQuantizedAddInit (const size_t *aShape, size_t aCount, SimdTensorDataType aType, const float *aScale, int32_t aZero, const size_t *bShape, size_t bCount, SimdTensorDataType bType, const float *bScale, int32_t bZero, SimdConvolutionActivationType actType, const float *actParams, SimdTensorDataType dstType, const float *dstScale, int32_t dstZero)
	Initializes element-wise quantized addition of two tensors with optional activation. More...
SIMD_API void	SimdSynetQuantizedAddForward (void *context, const uint8_t *a, const uint8_t *b, uint8_t *dst)
	Performs element-wise quantized addition. More...

Detailed Description

A framework to accelerate Quantized addition in Synet Framework.

Function Documentation

SimdSynetQuantizedAddInit()

void * SimdSynetQuantizedAddInit	(	const size_t *	aShape,
		size_t	aCount,
		SimdTensorDataType	aType,
		const float *	aScale,
		int32_t	aZero,
		const size_t *	bShape,
		size_t	bCount,
		SimdTensorDataType	bType,
		const float *	bScale,
		int32_t	bZero,
		SimdConvolutionActivationType	actType,
		const float *	actParams,
		SimdTensorDataType	dstType,
		const float *	dstScale,
		int32_t	dstZero
	)

Initializes element-wise quantized addition of two tensors with optional activation.

The current implementation supports equal input shapes. For each element it dequantizes UINT8 inputs as (value - zero)*scale, adds the two values, applies activation if it is specified and converts the result to FP32 or UINT8 output. FP32 inputs and outputs ignore the corresponding quantization zero.

Note

This function has a C++ wrapper: Simd::SynetQuantizedAdd.

Parameters

[in]	aShape	- a pointer to shape of input A tensor.
[in]	aCount	- a count of dimensions of input A tensor.
[in]	aType	- a type of input A tensor. It can be SimdTensorData32f or SimdTensorData8u.
[in]	aScale	- a pointer to quantization scale of input A tensor. Can be NULL (scale is 1.0).
[in]	aZero	- a quantization zero of input A tensor.
[in]	bShape	- a pointer to shape of input B tensor.
[in]	bCount	- a count of dimensions of input B tensor.
[in]	bType	- a type of input B tensor. It can be SimdTensorData32f or SimdTensorData8u.
[in]	bScale	- a pointer to quantization scale of input B tensor. Can be NULL (scale is 1.0).
[in]	bZero	- a quantization zero of input B tensor.
[in]	actType	- an activation function type applied after addition. Supported optimized path uses SimdConvolutionActivationIdentity or SimdConvolutionActivationRelu.
[in]	actParams	- a pointer to activation function parameters. Can be NULL.
[in]	dstType	- a type of output tensor. It can be SimdTensorData32f or SimdTensorData8u.
[in]	dstScale	- a pointer to output quantization scale. Can be NULL (scale is 1.0).
[in]	dstZero	- an output quantization zero.

Returns

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

SimdSynetQuantizedAddForward()

void SimdSynetQuantizedAddForward	(	void *	context,
		const uint8_t *	a,
		const uint8_t *	b,
		uint8_t *	dst
	)

Performs element-wise quantized addition.

Algorithm's details for UINT8 output:

for(i = 0; i < size; ++i)
{
    value = Activate((a[i] - aZero)*aScale + (b[i] - bZero)*bScale, actType, actParams);
    dst[i] = RestrictRange(Round(value/dstScale) + dstZero, 0, 255);
}

Note

This function has a C++ wrapper: Simd::SynetQuantizedAdd.

Parameters

[in]	context	- a pointer to quantized addition context. It must be created by function SimdSynetQuantizedAddInit and released by function SimdRelease.
[in]	a	- a pointer to input A tensor data. Its type is defined by parameter aType of SimdSynetQuantizedAddInit.
[in]	b	- a pointer to input B tensor data. Its type is defined by parameter bType of SimdSynetQuantizedAddInit.
[out]	dst	- a pointer to output tensor data. Its type is defined by parameter dstType of SimdSynetQuantizedAddInit.