Functions
PolarisStatus	polaris_gemm (PolarisContext ctxt, PolarisTransType trans_a, PolarisTransType trans_b, int m, int n, int k, float alpha, const void a, PolarisDataType type_a, int lda, const void b, PolarisDataType type_b, int ldb, float beta, void c, PolarisDataType type_c, int ldc, const void *bias, PolarisActivationType activation)
	Perform matrix multiplication with bias and activation support. More...

PolarisStatus	polaris_activation_2d (PolarisContext ctxt, PolarisActivationType type, int m, int n, float alpha, const void x, PolarisDataType type_x, int stride_x, float beta, void *y, PolarisDataType type_y, int stride_y)
	Compute activation functions. More...

PolarisStatus	polaris_deactivation_2d (PolarisContext ctxt, PolarisActivationType type, int m, int n, float alpha, const void y, PolarisDataType type_y, int stride_y, const void dy, PolarisDataType type_dy, int stride_dy, const void x, PolarisDataType type_x, int stride_x, float beta, void *dx, PolarisDataType type_dx, int stride_dx)
	Compute derived activation functions. More...

PolarisStatus	polaris_elementwise_2d (PolarisContext ctxt, PolarisElementWiseType type, int m, int n, float alpha0, const void x0, PolarisDataType type_x0, int stride_x0, float alpha1, const void x1, PolarisDataType type_x1, int stride_x1, float beta, void y, PolarisDataType type_y, int stride_y)
	Element wise functions. More...

PolarisStatus	polaris_transpose (PolarisContext ctxt, int m, int n, const void x, PolarisDataType type_x, int stride_x, void *y, PolarisDataType type_y, int stride_y)
	Perform matrix transpose on FPGA. More...

Detailed Description

Functions that involve computation. This group provides the core computation ability of Polaris library.

Function Documentation

◆ polaris_gemm()

PolarisStatus polaris_gemm	(	PolarisContext *	ctxt,
		PolarisTransType	trans_a,
		PolarisTransType	trans_b,
		int	m,
		int	n,
		int	k,
		float	alpha,
		const void *	a,
		PolarisDataType	type_a,
		int	lda,
		const void *	b,
		PolarisDataType	type_b,
		int	ldb,
		float	beta,
		void *	c,
		PolarisDataType	type_c,
		int	ldc,
		const void *	bias,
		PolarisActivationType	activation
	)

Perform matrix multiplication with bias and activation support.

This function perform the calculation as the formula:

c = activation( alpha * op(a) * op(b) + beta * c + bias )

This function behaves similar with SGEMM function in CBLAS, with the extended ability of adding bias vector and performing activation functions, which match better for the need of deep neural network. In another word, this function provides the ability of the forward process of a full-connected layer.

Parameters like trans_a, trans_b, m, n, k, alpha, a, lda b, ldb, beta, c, ldc all have the same meaning with that in CBLAS SGEMM.

bias is a vector with dimension 1 * n, which will be added to each row of output matrix c.

activation indicates the activation function type to be performed after the bias procession.

Caller could specify the data type of input matrix a, b, c with type_a, type_b, type_c , data type of bias is the same with c.

Note: this function best performed with trans_a equals to POLARIS_NO_TRANS and trans_b equals to POLARIS_TRANS. (Currently supports calling with this mode only)

Warning: lda, ldb and ldc is currently not supported in this version of Polaris, this functionality will be released later in December.
For data type, only POLARIS_FP32 is available for now, support for other data type will be released later.

Parameters

[in]	ctxt	pointer to PolarisContext object
[in]	trans_a	if POLARIS_NO_TRANS, `a` is `m * k`, otherwise `k * m`
[in]	trans_b	if POLARIS_NO_TRANS, `b` is `n * k`, otherwise `k * n`
[in]	m	dimension m
[in]	n	dimension n
[in]	k	dimension k
[in]	alpha	scalar parameter
[in]	a	FPGA address of matrix a
[in]	type_a	data type of matrix a
[in]	lda	same as LDA in BLAS
[in]	b	FPGA address of matrix b
[in]	type_b	data type of matrix b
[in]	ldb	same as LDB in BLAS
[in]	beta	scalar parameter
[in,out]	c	FPGA address of matrix c
[in]	type_c	data type of matrix c and bias
[in]	ldc	same as LDC in BLAS
[in]	bias	FPGA address of bias
[in]	activation	activaion type

Returns: Execution status, POLARIS_OK on success.

◆ polaris_activation_2d()

PolarisStatus polaris_activation_2d	(	PolarisContext *	ctxt,
		PolarisActivationType	type,
		int	m,
		int	n,
		float	alpha,
		const void *	x,
		PolarisDataType	type_x,
		int	stride_x,
		float	beta,
		void *	y,
		PolarisDataType	type_y,
		int	stride_y
	)

Compute activation functions.

This function calculate the following formula:

y = alpha * ActivationType(x) + beta * y

where both x and y have length elements. Activation type is indicated by type.

Parameters

[in]	ctxt	pointer to PolarisContext object
[in]	type	activation type
[in]	m	number of rows
[in]	n	number of cols
[in]	alpha	alpha value
[in]	x	input matrix
[in]	type_x	data type of `x`
[in]	stride_x	stride of `x`
[in]	beta	beta value
[out]	y	output matrix
[in]	type_y	data type of `y`
[in]	stride_y	stride of `y`

Note: Not fully supported by FPGA hardware yet.
Currently supports using with m = 1.

Returns: Execution status, POLARIS_OK on success.

◆ polaris_deactivation_2d()

PolarisStatus polaris_deactivation_2d	(	PolarisContext *	ctxt,
		PolarisActivationType	type,
		int	m,
		int	n,
		float	alpha,
		const void *	y,
		PolarisDataType	type_y,
		int	stride_y,
		const void *	dy,
		PolarisDataType	type_dy,
		int	stride_dy,
		const void *	x,
		PolarisDataType	type_x,
		int	stride_x,
		float	beta,
		void *	dx,
		PolarisDataType	type_dx,
		int	stride_dx
	)

Compute derived activation functions.

This function computes the following formula:

dx = DerivedActivation(y, dy, x)

It is used in the backward phase, and computes dx according to y, dy and x, where x and y is actually the input and output data separately of the forward phase of the same operation, while dy is the gradient of y.

Parameters

[in]	ctxt	pointer to PolarisContext
[in]	type	activation type
[in]	m	number of rows
[in]	n	number of cols
[in]	alpha	alpha value
[in]	y	backward src data (output data of the forward phase)
[in]	type_y	data type of `y`
[in]	stride_y	stride of `y`
[in]	dy	gradient of `y`
[in]	type_dy	data type of `dy`
[in]	stride_dy	stride of `dy`
[in]	x	backward destination data (input data of the forward phase)
[in]	type_x	data type of `x`
[in]	stride_x	stride of `x`
[in]	beta	beta value
[out]	dx	gradient of `x`
[in]	type_dx	data type of `dx`
[in]	stride_dx	stride of `dx`

Note: Not supported by FPGA hardware yet.

Returns: Execution status, POLARIS_OK on success.

◆ polaris_elementwise_2d()

PolarisStatus polaris_elementwise_2d	(	PolarisContext *	ctxt,
		PolarisElementWiseType	type,
		int	m,
		int	n,
		float	alpha0,
		const void *	x0,
		PolarisDataType	type_x0,
		int	stride_x0,
		float	alpha1,
		const void *	x1,
		PolarisDataType	type_x1,
		int	stride_x1,
		float	beta,
		void *	y,
		PolarisDataType	type_y,
		int	stride_y
	)

Element wise functions.

Perform element-wise operations indicated by the following fomula:

y = ElementWiseOperation(alpha0 * x0, alpha1 * x1) + beta * y

where type indicates the exact operation to be performed.

POLARIS_ADD := v0 + v1
POLARIS_MUL := v0 * v1
POLARIS_MIN := min(v0, v1)
POLARIS_MAX := max(v0, v1)

where v0 is alpha0 * x0 and v1 is alpha1 * x1.

Parameters

[in]	ctxt	pointer to PolarisContext
[in]	type	element wise operation type
[in]	m	number of rows
[in]	n	number of cols
[in]	alpha0	scalar parameter
[in]	x0	input matrix
[in]	type_x0	data type of `x0`
[in]	stride_x0	stride of `x0`
[in]	alpha1	scalar parameter
[in]	x1	input matrix
[in]	type_x1	data type of `x1`
[in]	stride_x1	stride of `x1`
[in]	beta	scalar parameter
[out]	y	output matrix
[in]	type_y	data type of `y`
[in]	stride_y	stride of `y`

Note: Not fully supported by FPGA hardware yet.
Currently supports using with m = 1.

Returns: Execution status, POLARIS_OK on success.

◆ polaris_transpose()

PolarisStatus polaris_transpose	(	PolarisContext *	ctxt,
		int	m,
		int	n,
		const void *	x,
		PolarisDataType	type_x,
		int	stride_x,
		void *	y,
		PolarisDataType	type_y,
		int	stride_y
	)

Perform matrix transpose on FPGA.

Dimension of matrix x is m * n, and y is n * m.

Warning: address of y and x should NOT be the same.

Parameters

[in]	ctxt	pointer to PolarisContext object
[in]	m	dimension m
[in]	n	dimension n
[in]	x	FPGA address of input matrix
[in]	type_x	data type of `x`
[in]	stride_x	stride of `x`
[out]	y	FPGA address of output matrix
[in]	type_y	data type of `y`
[in]	stride_y	stride of `y`

Note: Not supported by FPGA hardware yet.

Returns: Execution status, POLARIS_OK on success.