GridSampleLayer functions

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

Enumerations
enum	SimdGridSampleInterpType {   SimdGridSampleInterpBilinear = 0 ,   SimdGridSampleInterpNearest ,   SimdGridSampleInterpBicubic }
enum	SimdGridSamplePaddingType {   SimdGridSamplePaddingZeros = 0 ,   SimdGridSamplePaddingBorder ,   SimdGridSamplePaddingReflect }

Functions
SIMD_API void *	SimdSynetGridSample2dInit (size_t batch, size_t channels, size_t srcH, size_t srcW, size_t dstH, size_t dstW, SimdTensorDataType type, SimdGridSampleInterpType interp, SimdGridSamplePaddingType padding, SimdBool align)
	Initializes an ONNX-compatible GridSample-2D context. More...
SIMD_API size_t	SimdSynetGridSample2dInternalBufferSize (const void *context)
	Gets the size in bytes of internal storage used by a GridSample-2D context. More...
SIMD_API void	SimdSynetGridSample2dForward (void *context, const uint8_t *src, const uint8_t *grd, uint8_t *dst)
	Performs GridSample-2D forward propagation. More...

Detailed Description

Functions to accelerate GridSampleLayer in Synet Framework.

Enumeration Type Documentation

SimdGridSampleInterpType

enum SimdGridSampleInterpType

Describes interpolation modes used by SimdSynetGridSample2dInit.

The grid tensor stores normalized (x, y) coordinates. During SimdSynetGridSample2dForward these modes define how source pixels around each denormalized coordinate are combined.

Enumerator
SimdGridSampleInterpBilinear	Bilinear interpolation from the four neighboring source pixels.
SimdGridSampleInterpNearest	Nearest-neighbor interpolation after rounding the denormalized coordinate.
SimdGridSampleInterpBicubic	Bicubic interpolation from a 4x4 neighborhood.

SimdGridSamplePaddingType

enum SimdGridSamplePaddingType

Describes padding modes used by SimdSynetGridSample2dInit.

Padding is applied when a denormalized grid coordinate falls outside the source image. The border used for this test depends on the align parameter of SimdSynetGridSample2dInit.

Enumerator
SimdGridSamplePaddingZeros	Use zero for out-of-bound source samples.
SimdGridSamplePaddingBorder	Clamp out-of-bound source samples to the nearest border pixel.
SimdGridSamplePaddingReflect	Reflect out-of-bound source samples across the border before reading the source pixel.

Function Documentation

SimdSynetGridSample2dInit()

void * SimdSynetGridSample2dInit	(	size_t	batch,
		size_t	channels,
		size_t	srcH,
		size_t	srcW,
		size_t	dstH,
		size_t	dstW,
		SimdTensorDataType	type,
		SimdGridSampleInterpType	interp,
		SimdGridSamplePaddingType	padding,
		SimdBool	align
	)

Initializes an ONNX-compatible GridSample-2D context.

The function creates a context for NCHW tensors. The input tensor shape is batch*channels*srcH*srcW, the grid tensor shape is batch*dstH*dstW*2 and the output tensor shape is batch*channels*dstH*dstW. The grid stores normalized coordinates in the range [-1, 1] as pairs (x, y). The current implementation supports FP32 source, grid and destination tensors.

Parameters

[in]	batch	- a batch size.
[in]	channels	- a number of channels in the input and output tensors.
[in]	srcH	- a height of input tensor.
[in]	srcW	- a width of input tensor.
[in]	dstH	- a height of output tensor.
[in]	dstW	- a width of output tensor.
[in]	type	- a type of input, grid and output tensor. Currently only FP32 is supported.
[in]	interp	- an interpolation type: bilinear, nearest or bicubic.
[in]	padding	- a padding type for out-of-bound grid coordinates.
[in]	align	- a flag corresponding to ONNX/PyTorch align_corners.

Returns

a pointer to grid sample 2D context. On error it returns NULL. It must be released with using of function SimdRelease. This pointer is used in functions SimdSynetGridSample2dInternalBufferSize, and SimdSynetGridSample2dForward.

SimdSynetGridSample2dInternalBufferSize()

size_t SimdSynetGridSample2dInternalBufferSize

(

const void *

context

)

Gets the size in bytes of internal storage used by a GridSample-2D context.

The optimized bilinear/zero-padding FP32 implementation stores padded source rows, precomputed source indexes and interpolation coefficients internally. The reference implementation returns 0.

Parameters

[in]

context

- a pointer to grid sample 2D context. It must be created by function SimdSynetGridSample2dInit and released by function SimdRelease.

Returns

a number of bytes used by internal buffers.

SimdSynetGridSample2dForward()

void SimdSynetGridSample2dForward	(	void *	context,
		const uint8_t *	src,
		const uint8_t *	grd,
		uint8_t *	dst
	)

Performs GridSample-2D forward propagation.

For every output pixel the function denormalizes grid coordinates to input image coordinates, applies the selected padding rule for out-of-bound coordinates and samples the input with nearest, bilinear or bicubic interpolation:

if(align)
    x = (gridX + 1) * (srcW - 1) / 2;
else
    x = ((gridX + 1) * srcW - 1) / 2;
y is computed in the same way from gridY and srcH.
dst[b, c, dy, dx] = Interpolate(src[b, c], x, y, interp, padding);

Parameters

[in]	context	- a pointer to grid sample 2D context. It must be created by function SimdSynetGridSample2dInit and released by function SimdRelease.
[in]	src	- a pointer to input tensor. Its shape is batch * channels * srcH * srcW.
[in]	grd	- a pointer to grid tensor. Its shape is batch * dstH * dstW * 2.
[out]	dst	- a pointer to output tensor. Its shape is batch * channels * dstH * dstW.