conv_acc_neon_impl.h File Reference

Accelerated Viterbi decoder implementation: straight port of SSE to NEON based on Tom Tsous work. More...

Go to the source code of this file.

Macros
#define	__always_inline   inline __attribute__((always_inline))
#define	NEON_BUTTERFLY(M0, M1, M2, M3, M4)
#define	NEON_DEINTERLEAVE_K5(M0, M1, M2, M3)
#define	NEON_DEINTERLEAVE_K7(M0, M1, M2, M3, M4, M5, M6, M7, M8, M9, M10, M11, M12, M13, M14, M15)
#define	NEON_BRANCH_METRIC_N2(M0, M1, M2, M3, M4, M6, M7)
#define	NEON_BRANCH_METRIC_N4(M0, M1, M2, M3, M4, M5)
#define	NEON_NORMALIZE_K5(M0, M1, M2, M3)
#define	NEON_NORMALIZE_K7(M0, M1, M2, M3, M4, M5, M6, M7, M8, M9, M10, M11)

Functions
__always_inline void	_neon_metrics_k5_n2 (const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, int norm)
__always_inline void	_neon_metrics_k5_n4 (const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, int norm)
static __always_inline void	_neon_metrics_k7_n2 (const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, int norm)
static __always_inline void	_neon_metrics_k7_n4 (const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, int norm)

Detailed Description

Accelerated Viterbi decoder implementation: straight port of SSE to NEON based on Tom Tsous work.

Macro Definition Documentation

__always_inline

#define __always_inline   inline __attribute__((always_inline))

NEON_BRANCH_METRIC_N2

#define NEON_BRANCH_METRIC_N2	(		M0,
			M1,
			M2,
			M3,
			M4,
			M6,
			M7
	)

Value:

{ \
        M0 = vmulq_s16(M4, M0); \
        M1 = vmulq_s16(M4, M1); \
        M2 = vmulq_s16(M4, M2); \
        M3 = vmulq_s16(M4, M3); \
        M6 = vcombine_s16(vpadd_s16(vget_low_s16(M0), vget_high_s16(M0)), vpadd_s16(vget_low_s16(M1), vget_high_s16(M1))); \
        M7 = vcombine_s16(vpadd_s16(vget_low_s16(M2), vget_high_s16(M2)), vpadd_s16(vget_low_s16(M3), vget_high_s16(M3))); \
}

NEON_BRANCH_METRIC_N4

#define NEON_BRANCH_METRIC_N4	(		M0,
			M1,
			M2,
			M3,
			M4,
			M5
	)

Value:

{ \
        M0 = vmulq_s16(M4, M0); \
        M1 = vmulq_s16(M4, M1); \
        M2 = vmulq_s16(M4, M2); \
        M3 = vmulq_s16(M4, M3); \
        int16x4_t t1 = vpadd_s16(vpadd_s16(vget_low_s16(M0), vget_high_s16(M0)), vpadd_s16(vget_low_s16(M1), vget_high_s16(M1))); \
        int16x4_t t2 = vpadd_s16(vpadd_s16(vget_low_s16(M2), vget_high_s16(M2)), vpadd_s16(vget_low_s16(M3), vget_high_s16(M3))); \
        M5 = vcombine_s16(t1, t2); \
}

NEON_BUTTERFLY

#define NEON_BUTTERFLY	(		M0,
			M1,
			M2,
			M3,
			M4
	)

Value:

{ \
        M3 = vqaddq_s16(M0, M2); \
        M4 = vqsubq_s16(M1, M2); \
        M0 = vqsubq_s16(M0, M2); \
        M1 = vqaddq_s16(M1, M2); \
        M2 = vmaxq_s16(M3, M4); \
        M3 = vreinterpretq_s16_u16(vcgtq_s16(M3, M4)); \
        M4 = vmaxq_s16(M0, M1); \
        M1 = vreinterpretq_s16_u16(vcgtq_s16(M0, M1)); \
}

NEON_DEINTERLEAVE_K5

#define NEON_DEINTERLEAVE_K5	(		M0,
			M1,
			M2,
			M3
	)

Value:

{ \
        int16x8x2_t tmp; \
        tmp = vuzpq_s16(M0, M1); \
        M2 = tmp.val[0]; \
        M3 = tmp.val[1]; \
}

NEON_DEINTERLEAVE_K7

#define NEON_DEINTERLEAVE_K7	(		M0,
			M1,
			M2,
			M3,
			M4,
			M5,
			M6,
			M7,
			M8,
			M9,
			M10,
			M11,
			M12,
			M13,
			M14,
			M15
	)

Value:

{ \
        int16x8x2_t tmp; \
        tmp = vuzpq_s16(M0, M1); \
        M8 = tmp.val[0]; M9 = tmp.val[1]; \
        tmp = vuzpq_s16(M2, M3); \
        M10 = tmp.val[0]; M11 = tmp.val[1]; \
        tmp = vuzpq_s16(M4, M5); \
        M12 = tmp.val[0]; M13 = tmp.val[1]; \
        tmp = vuzpq_s16(M6, M7); \
        M14 = tmp.val[0]; M15 = tmp.val[1]; \
}

NEON_NORMALIZE_K5

#define NEON_NORMALIZE_K5	(		M0,
			M1,
			M2,
			M3
	)

Value:

{ \
        M2 = vminq_s16(M0, M1); \
        int16x4_t t = vpmin_s16(vget_low_s16(M2), vget_high_s16(M2)); \
        t = vpmin_s16(t, t); \
        t = vpmin_s16(t, t); \
        M2 = vdupq_lane_s16(t, 0); \
        M0 = vqsubq_s16(M0, M2); \
        M1 = vqsubq_s16(M1, M2); \
}

NEON_NORMALIZE_K7

#define NEON_NORMALIZE_K7	(		M0,
			M1,
			M2,
			M3,
			M4,
			M5,
			M6,
			M7,
			M8,
			M9,
			M10,
			M11
	)

Value:

{ \
        M8 = vminq_s16(M0, M1); \
        M9 = vminq_s16(M2, M3); \
        M10 = vminq_s16(M4, M5); \
        M11 = vminq_s16(M6, M7); \
        M8 = vminq_s16(M8, M9); \
        M10 = vminq_s16(M10, M11); \
        M8 = vminq_s16(M8, M10); \
        int16x4_t t = vpmin_s16(vget_low_s16(M8), vget_high_s16(M8)); \
        t = vpmin_s16(t, t); \
        t = vpmin_s16(t, t); \
        M8 = vdupq_lane_s16(t, 0); \
        M0 = vqsubq_s16(M0, M8); \
        M1 = vqsubq_s16(M1, M8); \
        M2 = vqsubq_s16(M2, M8); \
        M3 = vqsubq_s16(M3, M8); \
        M4 = vqsubq_s16(M4, M8); \
        M5 = vqsubq_s16(M5, M8); \
        M6 = vqsubq_s16(M6, M8); \
        M7 = vqsubq_s16(M7, M8); \
}

Function Documentation

_neon_metrics_k5_n2()

__always_inline void _neon_metrics_k5_n2	(	const int16_t *	val,
		const int16_t *	outa,
		int16_t *	sumsa,
		int16_t *	paths,
		int	norm
	)

References NEON_BUTTERFLY, NEON_DEINTERLEAVE_K5, and NEON_NORMALIZE_K5.

_neon_metrics_k5_n4()

__always_inline void _neon_metrics_k5_n4	(	const int16_t *	val,
		const int16_t *	outa,
		int16_t *	sumsa,
		int16_t *	paths,
		int	norm
	)

References NEON_BRANCH_METRIC_N4, NEON_BUTTERFLY, NEON_DEINTERLEAVE_K5, and NEON_NORMALIZE_K5.

_neon_metrics_k7_n2()

static __always_inline void _neon_metrics_k7_n2 ( const int16_t *  val, const int16_t *  outa, int16_t *  sumsa, int16_t *  paths, int  norm  )

static

References NEON_BRANCH_METRIC_N2, NEON_BUTTERFLY, NEON_DEINTERLEAVE_K7, and NEON_NORMALIZE_K7.

_neon_metrics_k7_n4()

static __always_inline void _neon_metrics_k7_n4 ( const int16_t *  val, const int16_t *  outa, int16_t *  sumsa, int16_t *  paths, int  norm  )

static

References NEON_BRANCH_METRIC_N4, NEON_BUTTERFLY, NEON_DEINTERLEAVE_K7, and NEON_NORMALIZE_K7.