shithub: openh264

ref: f8ae7fcc5f1a9fce5df39f752a78fe644d5e05bd
dir: /codec/encoder/core/arm/reconstruct_neon.S/

View raw version
/*!
 * \copy
 *     Copyright (c)  2013, Cisco Systems
 *     All rights reserved.
 *
 *     Redistribution and use in source and binary forms, with or without
 *     modification, are permitted provided that the following conditions
 *     are met:
 *
 *        * Redistributions of source code must retain the above copyright
 *          notice, this list of conditions and the following disclaimer.
 *
 *        * Redistributions in binary form must reproduce the above copyright
 *          notice, this list of conditions and the following disclaimer in
 *          the documentation and/or other materials provided with the
 *          distribution.
 *
 *     THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 *     "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 *     LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
 *     FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
 *     COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
 *     INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
 *     BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 *     LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
 *     CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 *     LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
 *     ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 *     POSSIBILITY OF SUCH DAMAGE.
 *
 */

#ifdef  HAVE_NEON
.text
#include "arm_arch_common_macro.S"

#ifdef __APPLE__
.macro	LOAD_4x4_DATA_FOR_DCT
//	{	//	input: $0~$3, src1*, src1_stride, src2*, src2_stride
    vld2.16	{$0[0],$1[0]}, [$4], $5
    vld2.16	{$2[0],$3[0]}, [$6], $7
    vld2.16	{$0[1],$1[1]}, [$4], $5
    vld2.16	{$2[1],$3[1]}, [$6], $7

    vld2.16	{$0[2],$1[2]}, [$4], $5
    vld2.16	{$2[2],$3[2]}, [$6], $7
    vld2.16	{$0[3],$1[3]}, [$4], $5
    vld2.16	{$2[3],$3[3]}, [$6], $7
//	}
.endm

.macro	LOAD_8x8_DATA_FOR_DCT
//	{	//	input: $0~$3, src1*, src2*; untouched r2:src1_stride &r4:src2_stride
    vld1.64	{$0}, [$8], r2
    vld1.64	{$4}, [$9], r4
    vld1.64	{$1}, [$8], r2
    vld1.64	{$5}, [$9], r4

    vld1.64	{$2}, [$8], r2
    vld1.64	{$6}, [$9], r4
    vld1.64	{$3}, [$8], r2
    vld1.64	{$7}, [$9], r4
//	}
.endm

.macro	DCT_ROW_TRANSFORM_TOTAL_16BITS
//	{	//	input: src_d[0]~[3], working: [4]~[7]
    vadd.s16		$4, $0, $3			//int16 s[0] = data[i] + data[i3];
    vsub.s16		$7, $0, $3			//int16 s[3] = data[i] - data[i3];
    vadd.s16		$5, $1, $2			//int16 s[1] = data[i1] + data[i2];
    vsub.s16		$6, $1, $2			//int16 s[2] = data[i1] - data[i2];

    vadd.s16		$0, $4, $5			//int16 dct[i ] = s[0] + s[1];
    vsub.s16		$2, $4, $5			//int16 dct[i2] = s[0] - s[1];
    vshl.s16		$1, $7, #1
    vshl.s16		$3, $6, #1
    vadd.s16		$1, $1, $6			//int16 dct[i1] = (s[3] << 1) + s[2];
    vsub.s16		$3, $7, $3			//int16 dct[i3] = s[3] - (s[2] << 1);
//	}
.endm

.macro	MATRIX_TRANSFORM_EACH_16BITS
//	{	//	input & output: src_d[0]~[3];[0 1 2 3]+[4 5 6 7]+[8 9 10 11]+[12 13 14 15]
    vtrn.s16		$0, $1				//[0 1 2 3]+[4 5 6 7]-->[0 4 2 6]+[1 5 3 7]
    vtrn.s16		$2, $3				//[8 9 10 11]+[12 13 14 15]-->[8 12 10 14]+[9 13 11 15]
    vtrn.32		$0, $2				//[0 4 2 6]+[8 12 10 14]-->[0 4 8 12]+[2 6 10 14]
    vtrn.32		$1, $3				//[1 5 3 7]+[9 13 11 15]-->[1 5 9 13]+[3 7 11 15]
//	}
.endm

.macro	NEWQUANT_COEF_EACH_16BITS	// if coef <= 0, - coef; else , coef;
//	{	//	input:	coef, ff (dst), ff_d0, ff_d1, mf_d0, md_d1
    veor.s16		$6, $6			// init 0 , and keep 0;
    vaba.s16		$1, $0, $6		// f + abs(coef - 0)
    vmull.s16		$7, $2, $4
    vmull.s16		$8, $3, $5
    vshr.s32		$7, #16
    vshr.s32		$8, #16
    vmovn.s32		$2, $7
    vmovn.s32		$3, $8

    vcgt.s16		$7, $0, #0		// if true, location of coef == 11111111
    vbif.s16		$6, $1, $7		// if (x<0) reserved part; else keep 0 untouched
    vshl.s16		$6, #1
    vsub.s16		$1, $1, $6		// if x > 0, -= 0; else x-= 2x
//	}
.endm

.macro	NEWQUANT_COEF_EACH_16BITS_MAX	// if coef <= 0, - coef; else , coef;
//	{	//	input:	coef, ff (dst), ff_d0, ff_d1, mf_d0(max), md_d1
    veor.s16		$6, $6			// init 0 , and keep 0;
    vaba.s16		$1, $0, $6		// f + abs(coef - 0)
    vmull.s16		$7, $2, $4
    vmull.s16		$8, $3, $5
    vshr.s32		$7, #16
    vshr.s32		$8, #16
    vmovn.s32		$2, $7
    vmovn.s32		$3, $8

    vcgt.s16		$7, $0, #0		// if true, location of coef == 11111111
    vbif.s16		$6, $1, $7		// if (x<0) reserved part; else keep 0 untouched
    vshl.s16		$6, #1
    vmax.s16		$9, $2, $3
    vsub.s16		$1, $1, $6		// if x > 0, -= 0; else x-= 2x
//	}
.endm

.macro	QUANT_DUALWORD_COEF_EACH_16BITS	// if coef <= 0, - coef; else , coef;
//	{	//	input:	coef, ff (dst), mf , working_d (all 0), working_q
    vaba.s16		$1, $0, $3		// f + abs(coef - 0)
    vmull.s16		$4, $1, $2		// *= mf
    vshr.s32		$4, #16
    vmovn.s32		$1, $4			// >> 16

    vcgt.s16		$2, $0, #0		// if true, location of coef == 11111111
    vbif.s16		$3, $1, $2		// if (x<0) reserved part; else keep 0 untouched
    vshl.s16		$3, #1
    vsub.s16		$1, $1, $3		// if x > 0, -= 0; else x-= 2x
//	}
.endm

.macro	DC_ZERO_COUNT_IN_DUALWORD
//	{	//	input:	coef, dst_d, working_d (all 0x01)
    vceq.s16	$1, $0, #0
    vand.s16	$1, $2
    vpadd.s16	$1, $1, $1
    vpadd.s16	$1, $1, $1
//	}
.endm

.macro	SELECT_MAX_IN_ABS_COEF
//	{	//	input:	coef_0, coef_1, max_q (identy to follow two)
    vmax.s16		$2, $0, $1		// max 1st in $3 & max 2nd in $4
    vpmax.s16		$3, $3, $4		// max 1st in $3[0][1] & max 2nd in $3[2][3]
    vpmax.s16		$3, $3, $4		// max 1st in $3[0][1]
//	}
.endm

.macro	ZERO_COUNT_IN_2_QUARWORD
//	{	//	input:	coef_0 (identy to $3 $4), coef_1(identy to $5 $6), mask_q
    vceq.s16	$0, #0
    vceq.s16	$1, #0
    vand.s16	$0, $2
    vand.s16	$1, $2

    vpadd.s16	$3, $3, $5
    vpadd.s16	$4, $4, $6
    vpadd.s16	$3, $3, $4		// 8-->4
    vpadd.s16	$3, $3, $3
    vpadd.s16	$3, $3, $3
//	}
.endm

.macro	HDM_QUANT_2x2_TOTAL_16BITS
//	{	//	input: src_d[0]~[3], working_d, dst_d
    vshr.s64	$1, $0, #32
    vadd.s16	$2, $0, $1		// [0] = rs[0] + rs[32];[1] = rs[16] + rs[48];
    vsub.s16	$1, $0, $1		// [0] = rs[0] - rs[32];[1] = rs[16] - rs[48];
    vtrn.s16	$2, $1
    vtrn.s32	$2, $1
//	}
.endm

.macro	IHDM_4x4_TOTAL_16BITS
//	{	//	input: each src_d[0]~[3](dst), working_q0, working_q1, working_q2
    vshr.s64	$1, $0, #32
    vadd.s16	$2, $0, $1		// [0] = rs[0] + rs[2];[1] = rs[1] + rs[3];
    vsub.s16	$1, $0, $1		// [0] = rs[0] - rs[2];[1] = rs[1] - rs[3];
    vtrn.s16	$2, $1
    vrev32.16	$1, $1
    vtrn.s32	$2, $1			// [0] = rs[0] + rs[2];[1] = rs[0] - rs[2];[2] = rs[1] - rs[3];[3] = rs[1] + rs[3];

    vrev64.16	$1, $2
    vadd.s16	$0, $2, $1		// [0] = rs[0] + rs[3];[1] = rs[1] + rs[2];
    vsub.s16	$1, $2, $1
    vrev32.16	$1, $1			// [0] = rs[1] - rs[2];[1] = rs[0] - rs[3];
    vtrn.s32	$0, $1			// [0] = rs[0] + rs[3];[1] = rs[1] + rs[2];[2] = rs[1] - rs[2];[3] = rs[0] - rs[3];
//	}
.endm

.macro	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP
//	{	//	input: pred_d[0]/[1](output), dct_q0/1, working_q0/1;
    vmovl.u8		$4,$0
    vmovl.u8		$5,$1
    vadd.s16		$4,$2
    vadd.s16		$5,$3
    vqmovun.s16	$0,$4
    vqmovun.s16	$1,$5
//	}
.endm

.macro	ROW_TRANSFORM_1_STEP_TOTAL_16BITS
//	{	//	input: src_d[0]~[3], output: e_d[0]~[3];
    vadd.s16		$4, $0, $2			//int16 e[i][0] = src[0] + src[2];
    vsub.s16		$5, $0, $2			//int16 e[i][1] = src[0] - src[2];
    vshr.s16		$6, $1, #1
    vshr.s16		$7, $3, #1
    vsub.s16		$6, $6, $3			//int16 e[i][2] = (src[1]>>1)-src[3];
    vadd.s16		$7, $1, $7			//int16 e[i][3] = src[1] + (src[3]>>1);
//	}
.endm

.macro	TRANSFORM_TOTAL_16BITS	// both row & col transform used
//	{	//	output: f_q[0]~[3], input: e_q[0]~[3];
    vadd.s16		$0, $4, $7			//int16 f[i][0] = e[i][0] + e[i][3];
    vadd.s16		$1, $5, $6			//int16 f[i][1] = e[i][1] + e[i][2];
    vsub.s16		$2, $5, $6			//int16 f[i][2] = e[i][1] - e[i][2];
    vsub.s16		$3, $4, $7			//int16 f[i][3] = e[i][0] - e[i][3];
//	}
.endm


.macro	ROW_TRANSFORM_0_STEP
//	{	//	input: src_d[0]~[3], output: e_q[0]~[3];
    vaddl.s16		$4, $0, $2			//int32 e[i][0] = src[0] + src[2];
    vsubl.s16		$5, $0, $2			//int32 e[i][1] = src[0] - src[2];
    vsubl.s16		$6, $1, $3			//int32 e[i][2] = src[1] - src[3];
    vaddl.s16		$7, $1, $3			//int32 e[i][3] = src[1] + src[3];
//	}
.endm

.macro	ROW_TRANSFORM_1_STEP
//	{	//	input: src_d[0]~[3], output: e_q[0]~[3]; working: $8 $9
    vaddl.s16		$4, $0, $2			//int32 e[i][0] = src[0] + src[2];
    vsubl.s16		$5, $0, $2			//int32 e[i][1] = src[0] - src[2];
    vshr.s16		$8, $1, #1
    vshr.s16		$9, $3, #1
    vsubl.s16		$6, $8, $3			//int32 e[i][2] = (src[1]>>1)-src[3];
    vaddl.s16		$7, $1, $9			//int32 e[i][3] = src[1] + (src[3]>>1);
//	}
.endm

.macro	TRANSFORM_4BYTES	// both row & col transform used
//	{	//	output: f_q[0]~[3], input: e_q[0]~[3];
    vadd.s32		$0, $4, $7			//int16 f[i][0] = e[i][0] + e[i][3];
    vadd.s32		$1, $5, $6			//int16 f[i][1] = e[i][1] + e[i][2];
    vsub.s32		$2, $5, $6			//int16 f[i][2] = e[i][1] - e[i][2];
    vsub.s32		$3, $4, $7			//int16 f[i][3] = e[i][0] - e[i][3];
//	}
.endm

.macro	COL_TRANSFORM_0_STEP
//	{	//	input: src_q[0]~[3], output: e_q[0]~[3];
    vadd.s32		$4, $0, $2			//int32 e[0][j] = f[0][j] + f[2][j];
    vsub.s32		$5, $0, $2			//int32 e[1][j] = f[0][j] - f[2][j];
    vsub.s32		$6, $1, $3			//int32 e[2][j] = (f[1][j]>>1) - f[3][j];
    vadd.s32		$7, $1, $3			//int32 e[3][j] = f[1][j] + (f[3][j]>>1);
//	}
.endm

.macro	COL_TRANSFORM_1_STEP
//	{	//	input: src_q[0]~[3], output: e_q[0]~[3];
    vadd.s32		$4, $0, $2			//int32 e[0][j] = f[0][j] + f[2][j];
    vsub.s32		$5, $0, $2			//int32 e[1][j] = f[0][j] - f[2][j];
    vshr.s32		$6, $1, #1
    vshr.s32		$7, $3, #1
    vsub.s32		$6, $6, $3			//int32 e[2][j] = (f[1][j]>>1) - f[3][j];
    vadd.s32		$7, $1, $7			//int32 e[3][j] = f[1][j] + (f[3][j]>>1);
//	}
.endm
#else
.macro	LOAD_4x4_DATA_FOR_DCT arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
//	{	//	input: \arg0~\arg3, src1*, src1_stride, src2*, src2_stride
    vld2.16	{\arg0[0],\arg1[0]}, [\arg4], \arg5
    vld2.16	{\arg2[0],\arg3[0]}, [\arg6], \arg7
    vld2.16	{\arg0[1],\arg1[1]}, [\arg4], \arg5
    vld2.16	{\arg2[1],\arg3[1]}, [\arg6], \arg7

    vld2.16	{\arg0[2],\arg1[2]}, [\arg4], \arg5
    vld2.16	{\arg2[2],\arg3[2]}, [\arg6], \arg7
    vld2.16	{\arg0[3],\arg1[3]}, [\arg4], \arg5
    vld2.16	{\arg2[3],\arg3[3]}, [\arg6], \arg7
//	}
.endm

.macro	LOAD_8x8_DATA_FOR_DCT arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9
//	{	//	input: \arg0~\arg3, src1*, src2*; untouched r2:src1_stride &r4:src2_stride
    vld1.64	{\arg0}, [\arg8], r2
    vld1.64	{\arg4}, [\arg9], r4
    vld1.64	{\arg1}, [\arg8], r2
    vld1.64	{\arg5}, [\arg9], r4

    vld1.64	{\arg2}, [\arg8], r2
    vld1.64	{\arg6}, [\arg9], r4
    vld1.64	{\arg3}, [\arg8], r2
    vld1.64	{\arg7}, [\arg9], r4
//	}
.endm

.macro	DCT_ROW_TRANSFORM_TOTAL_16BITS arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
//	{	//	input: src_d[0]~[3], working: [4]~[7]
    vadd.s16		\arg4, \arg0, \arg3			//int16 s[0] = data[i] + data[i3];
    vsub.s16		\arg7, \arg0, \arg3			//int16 s[3] = data[i] - data[i3];
    vadd.s16		\arg5, \arg1, \arg2			//int16 s[1] = data[i1] + data[i2];
    vsub.s16		\arg6, \arg1, \arg2			//int16 s[2] = data[i1] - data[i2];

    vadd.s16		\arg0, \arg4, \arg5			//int16 dct[i ] = s[0] + s[1];
    vsub.s16		\arg2, \arg4, \arg5			//int16 dct[i2] = s[0] - s[1];
    vshl.s16		\arg1, \arg7, #1
    vshl.s16		\arg3, \arg6, #1
    vadd.s16		\arg1, \arg1, \arg6			//int16 dct[i1] = (s[3] << 1) + s[2];
    vsub.s16		\arg3, \arg7, \arg3			//int16 dct[i3] = s[3] - (s[2] << 1);
//	}
.endm

.macro	MATRIX_TRANSFORM_EACH_16BITS arg0, arg1, arg2, arg3
//	{	//	input & output: src_d[0]~[3];[0 1 2 3]+[4 5 6 7]+[8 9 10 11]+[12 13 14 15]
    vtrn.s16		\arg0, \arg1				//[0 1 2 3]+[4 5 6 7]-->[0 4 2 6]+[1 5 3 7]
    vtrn.s16		\arg2, \arg3				//[8 9 10 11]+[12 13 14 15]-->[8 12 10 14]+[9 13 11 15]
    vtrn.32		\arg0, \arg2				//[0 4 2 6]+[8 12 10 14]-->[0 4 8 12]+[2 6 10 14]
    vtrn.32		\arg1, \arg3				//[1 5 3 7]+[9 13 11 15]-->[1 5 9 13]+[3 7 11 15]
//	}
.endm

.macro	NEWQUANT_COEF_EACH_16BITS arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8
//	{	//	input:	coef, ff (dst), ff_d0, ff_d1, mf_d0, md_d1
    veor.s16		\arg6, \arg6			// init 0 , and keep 0;
    vaba.s16		\arg1, \arg0, \arg6		// f + abs(coef - 0)
    vmull.s16		\arg7, \arg2, \arg4
    vmull.s16		\arg8, \arg3, \arg5
    vshr.s32		\arg7, #16
    vshr.s32		\arg8, #16
    vmovn.s32		\arg2, \arg7
    vmovn.s32		\arg3, \arg8

    vcgt.s16		\arg7, \arg0, #0		// if true, location of coef == 11111111
    vbif.s16		\arg6, \arg1, \arg7		// if (x<0) reserved part; else keep 0 untouched
    vshl.s16		\arg6, #1
    vsub.s16		\arg1, \arg1, \arg6		// if x > 0, -= 0; else x-= 2x
//	}
.endm

.macro	NEWQUANT_COEF_EACH_16BITS_MAX arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9
//	{	//	input:	coef, ff (dst), ff_d0, ff_d1, mf_d0(max), md_d1
    veor.s16		\arg6, \arg6			// init 0 , and keep 0;
    vaba.s16		\arg1, \arg0, \arg6		// f + abs(coef - 0)
    vmull.s16		\arg7, \arg2, \arg4
    vmull.s16		\arg8, \arg3, \arg5
    vshr.s32		\arg7, #16
    vshr.s32		\arg8, #16
    vmovn.s32		\arg2, \arg7
    vmovn.s32		\arg3, \arg8

    vcgt.s16		\arg7, \arg0, #0		// if true, location of coef == 11111111
    vbif.s16		\arg6, \arg1, \arg7		// if (x<0) reserved part; else keep 0 untouched
    vshl.s16		\arg6, #1
    vmax.s16		\arg9, \arg2, \arg3
    vsub.s16		\arg1, \arg1, \arg6		// if x > 0, -= 0; else x-= 2x
//	}
.endm

.macro	QUANT_DUALWORD_COEF_EACH_16BITS arg0, arg1, arg2, arg3, arg4
//	{	//	input:	coef, ff (dst), mf , working_d (all 0), working_q
    vaba.s16		\arg1, \arg0, \arg3		// f + abs(coef - 0)
    vmull.s16		\arg4, \arg1, \arg2		// *= mf
    vshr.s32		\arg4, #16
    vmovn.s32		\arg1, \arg4			// >> 16

    vcgt.s16		\arg2, \arg0, #0		// if true, location of coef == 11111111
    vbif.s16		\arg3, \arg1, \arg2		// if (x<0) reserved part; else keep 0 untouched
    vshl.s16		\arg3, #1
    vsub.s16		\arg1, \arg1, \arg3		// if x > 0, -= 0; else x-= 2x
//	}
.endm

.macro	DC_ZERO_COUNT_IN_DUALWORD arg0, arg1, arg2
//	{	//	input:	coef, dst_d, working_d (all 0x01)
    vceq.s16	\arg1, \arg0, #0
    vand.s16	\arg1, \arg2
    vpadd.s16	\arg1, \arg1, \arg1
    vpadd.s16	\arg1, \arg1, \arg1
//	}
.endm

.macro	SELECT_MAX_IN_ABS_COEF arg0, arg1, arg2, arg3, arg4
//	{	//	input:	coef_0, coef_1, max_q (identy to follow two), output: max_d0, max_d1
    vmax.s16		\arg2, \arg0, \arg1		// max 1st in \arg3 & max 2nd in \arg4
    vpmax.s16		\arg3, \arg3, \arg4		// max 1st in \arg3[0][1] & max 2nd in \arg3[2][3]
    vpmax.s16		\arg3, \arg3, \arg4		// max 1st in \arg3[0][1]
//	}
.endm

.macro	ZERO_COUNT_IN_2_QUARWORD arg0, arg1, arg2, arg3, arg4, arg5, arg6
//	{	//	input:	coef_0 (identy to \arg3 \arg4), coef_1(identy to \arg5 \arg6), mask_q
    vceq.s16	\arg0, #0
    vceq.s16	\arg1, #0
    vand.s16	\arg0, \arg2
    vand.s16	\arg1, \arg2

    vpadd.s16	\arg3, \arg3, \arg5
    vpadd.s16	\arg4, \arg4, \arg6
    vpadd.s16	\arg3, \arg3, \arg4		// 8-->4
    vpadd.s16	\arg3, \arg3, \arg3
    vpadd.s16	\arg3, \arg3, \arg3
//	}
.endm

.macro	HDM_QUANT_2x2_TOTAL_16BITS arg0, arg1, arg2
//	{	//	input: src_d[0]~[3], working_d, dst_d
    vshr.s64	\arg1, \arg0, #32
    vadd.s16	\arg2, \arg0, \arg1		// [0] = rs[0] + rs[32];[1] = rs[16] + rs[48];
    vsub.s16	\arg1, \arg0, \arg1		// [0] = rs[0] - rs[32];[1] = rs[16] - rs[48];
    vtrn.s16	\arg2, \arg1
    vtrn.s32	\arg2, \arg1
//	}
.endm

.macro	IHDM_4x4_TOTAL_16BITS arg0, arg1, arg2
//	{	//	input: each src_d[0]~[3](dst), working_q0, working_q1, working_q2
    vshr.s64	\arg1, \arg0, #32
    vadd.s16	\arg2, \arg0, \arg1		// [0] = rs[0] + rs[2];[1] = rs[1] + rs[3];
    vsub.s16	\arg1, \arg0, \arg1		// [0] = rs[0] - rs[2];[1] = rs[1] - rs[3];
    vtrn.s16	\arg2, \arg1
    vrev32.16	\arg1, \arg1
    vtrn.s32	\arg2, \arg1			// [0] = rs[0] + rs[2];[1] = rs[0] - rs[2];[2] = rs[1] - rs[3];[3] = rs[1] + rs[3];

    vrev64.16	\arg1, \arg2
    vadd.s16	\arg0, \arg2, \arg1		// [0] = rs[0] + rs[3];[1] = rs[1] + rs[2];
    vsub.s16	\arg1, \arg2, \arg1
    vrev32.16	\arg1, \arg1			// [0] = rs[1] - rs[2];[1] = rs[0] - rs[3];
    vtrn.s32	\arg0, \arg1			// [0] = rs[0] + rs[3];[1] = rs[1] + rs[2];[2] = rs[1] - rs[2];[3] = rs[0] - rs[3];
//	}
.endm

.macro	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP arg0, arg1, arg2, arg3, arg4, arg5
//	{	//	input: pred_d[0]/[1](output), dct_q0/1, working_q0/1;
    vmovl.u8		\arg4,\arg0
    vmovl.u8		\arg5,\arg1
    vadd.s16		\arg4,\arg2
    vadd.s16		\arg5,\arg3
    vqmovun.s16	\arg0,\arg4
    vqmovun.s16	\arg1,\arg5
//	}
.endm

.macro	ROW_TRANSFORM_1_STEP_TOTAL_16BITS arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
//	{	//	input: src_d[0]~[3], output: e_d[0]~[3];
    vadd.s16		\arg4, \arg0, \arg2			//int16 e[i][0] = src[0] + src[2];
    vsub.s16		\arg5, \arg0, \arg2			//int16 e[i][1] = src[0] - src[2];
    vshr.s16		\arg6, \arg1, #1
    vshr.s16		\arg7, \arg3, #1
    vsub.s16		\arg6, \arg6, \arg3			//int16 e[i][2] = (src[1]>>1)-src[3];
    vadd.s16		\arg7, \arg1, \arg7			//int16 e[i][3] = src[1] + (src[3]>>1);
//	}
.endm

.macro	TRANSFORM_TOTAL_16BITS arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7	// both row & col transform used
//	{	//	output: f_q[0]~[3], input: e_q[0]~[3];
    vadd.s16		\arg0, \arg4, \arg7			//int16 f[i][0] = e[i][0] + e[i][3];
    vadd.s16		\arg1, \arg5, \arg6			//int16 f[i][1] = e[i][1] + e[i][2];
    vsub.s16		\arg2, \arg5, \arg6			//int16 f[i][2] = e[i][1] - e[i][2];
    vsub.s16		\arg3, \arg4, \arg7			//int16 f[i][3] = e[i][0] - e[i][3];
//	}
.endm


.macro	ROW_TRANSFORM_0_STEP arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
//	{	//	input: src_d[0]~[3], output: e_q[0]~[3];
    vaddl.s16		\arg4, \arg0, \arg2			//int32 e[i][0] = src[0] + src[2];
    vsubl.s16		\arg5, \arg0, \arg2			//int32 e[i][1] = src[0] - src[2];
    vsubl.s16		\arg6, \arg1, \arg3			//int32 e[i][2] = src[1] - src[3];
    vaddl.s16		\arg7, \arg1, \arg3			//int32 e[i][3] = src[1] + src[3];
//	}
.endm

.macro	ROW_TRANSFORM_1_STEP arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7, arg8, arg9
//	{	//	input: src_d[0]~[3], output: e_q[0]~[3]; working: \arg8 \arg9
    vaddl.s16		\arg4, \arg0, \arg2			//int32 e[i][0] = src[0] + src[2];
    vsubl.s16		\arg5, \arg0, \arg2			//int32 e[i][1] = src[0] - src[2];
    vshr.s16		\arg8, \arg1, #1
    vshr.s16		\arg9, \arg3, #1
    vsubl.s16		\arg6, \arg8, \arg3			//int32 e[i][2] = (src[1]>>1)-src[3];
    vaddl.s16		\arg7, \arg1, \arg9			//int32 e[i][3] = src[1] + (src[3]>>1);
//	}
.endm

.macro	TRANSFORM_4BYTES arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7	// both row & col transform used
//	{	//	output: f_q[0]~[3], input: e_q[0]~[3];
    vadd.s32		\arg0, \arg4, \arg7			//int16 f[i][0] = e[i][0] + e[i][3];
    vadd.s32		\arg1, \arg5, \arg6			//int16 f[i][1] = e[i][1] + e[i][2];
    vsub.s32		\arg2, \arg5, \arg6			//int16 f[i][2] = e[i][1] - e[i][2];
    vsub.s32		\arg3, \arg4, \arg7			//int16 f[i][3] = e[i][0] - e[i][3];
//	}
.endm

.macro	COL_TRANSFORM_0_STEP arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
//	{	//	input: src_q[0]~[3], output: e_q[0]~[3];
    vadd.s32		\arg4, \arg0, \arg2			//int32 e[0][j] = f[0][j] + f[2][j];
    vsub.s32		\arg5, \arg0, \arg2			//int32 e[1][j] = f[0][j] - f[2][j];
    vsub.s32		\arg6, \arg1, \arg3			//int32 e[2][j] = (f[1][j]>>1) - f[3][j];
    vadd.s32		\arg7, \arg1, \arg3			//int32 e[3][j] = f[1][j] + (f[3][j]>>1);
//	}
.endm

.macro	COL_TRANSFORM_1_STEP arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7
//	{	//	input: src_q[0]~[3], output: e_q[0]~[3];
    vadd.s32		\arg4, \arg0, \arg2			//int32 e[0][j] = f[0][j] + f[2][j];
    vsub.s32		\arg5, \arg0, \arg2			//int32 e[1][j] = f[0][j] - f[2][j];
    vshr.s32		\arg6, \arg1, #1
    vshr.s32		\arg7, \arg3, #1
    vsub.s32		\arg6, \arg6, \arg3			//int32 e[2][j] = (f[1][j]>>1) - f[3][j];
    vadd.s32		\arg7, \arg1, \arg7			//int32 e[3][j] = f[1][j] + (f[3][j]>>1);
//	}
.endm
#endif


WELS_ASM_FUNC_BEGIN WelsDctT4_neon
	push		{r4}
	ldr			r4, [sp, #4]

	LOAD_4x4_DATA_FOR_DCT	d4, d5, d6, d7, r1, r2, r3, r4

	vsubl.u8	q0, d4, d6
	vsubl.u8	q1, d5, d7
	vtrn.s32	q0, q1
	vswp		d1, d2

	// horizontal transform
	DCT_ROW_TRANSFORM_TOTAL_16BITS		d0, d1, d2, d3, d4, d5, d6, d7

	// transform element
	MATRIX_TRANSFORM_EACH_16BITS	d0, d1, d2, d3

	//	vertical transform
	DCT_ROW_TRANSFORM_TOTAL_16BITS		d0, d1, d2, d3, d4, d5, d6, d7

	// transform element
	MATRIX_TRANSFORM_EACH_16BITS	d0, d1, d2, d3

	vst1.s16		{q0, q1}, [r0]!

	pop		{r4}
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsDctFourT4_neon
	push		{r4}
	ldr			r4, [sp, #4]

	LOAD_8x8_DATA_FOR_DCT	d16, d17, d18, d19, d20, d21, d22, d23, r1, r3

	vsubl.u8	q0, d16, d20
	vsubl.u8	q1, d17, d21
	vsubl.u8	q2, d18, d22
	vsubl.u8	q3, d19, d23
	MATRIX_TRANSFORM_EACH_16BITS	q0, q1, q2, q3

	// horizontal transform
	DCT_ROW_TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	// transform element
	MATRIX_TRANSFORM_EACH_16BITS	q0, q1, q2, q3

	//	vertical transform
	DCT_ROW_TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	vswp		d1, d2
	vswp		d5, d6
	vswp		q1, q2
	vst1.s16		{q0, q1}, [r0]!
	vst1.s16		{q2, q3}, [r0]!

	////////////////
	LOAD_8x8_DATA_FOR_DCT	d16, d17, d18, d19, d20, d21, d22, d23, r1, r3

	vsubl.u8	q0, d16, d20
	vsubl.u8	q1, d17, d21
	vsubl.u8	q2, d18, d22
	vsubl.u8	q3, d19, d23
	MATRIX_TRANSFORM_EACH_16BITS	q0, q1, q2, q3

	// horizontal transform
	DCT_ROW_TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	// transform element
	MATRIX_TRANSFORM_EACH_16BITS	q0, q1, q2, q3

	//	vertical transform
	DCT_ROW_TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	vswp		d1, d2
	vswp		d5, d6
	vswp		q1, q2
	vst1.s16		{q0, q1}, [r0]!
	vst1.s16		{q2, q3}, [r0]!

	pop		{r4}
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsQuant4x4_neon
	vld1.s16		{q2}, [r1]
	vld1.s16		{q0, q1}, [r0]
	vld1.s16		{q3}, [r2]

	vmov			q8, q2

	NEWQUANT_COEF_EACH_16BITS	q0, q2, d4, d5, d6, d7, q9, q10, q11
	vst1.s16		{q2}, [r0]!

	NEWQUANT_COEF_EACH_16BITS	q1, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r0]!

WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsQuant4x4Dc_neon

	vld1.s16		{q0, q1}, [r0]
	vdup.s16		q2, r1		// even ff range [0, 768]
	vdup.s16		q3, r2

	vmov			q8, q2

	NEWQUANT_COEF_EACH_16BITS	q0, q2, d4, d5, d6, d7, q9, q10, q11
	vst1.s16		{q2}, [r0]!

	NEWQUANT_COEF_EACH_16BITS	q1, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r0]!

WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsQuantFour4x4_neon
	vld1.s16		{q2}, [r1]
	vld1.s16		{q3}, [r2]
	mov				r1, r0

	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q0, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q1, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!

	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q0, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q1, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!

	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q0, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q1, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!

	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q0, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS	q1, q8, d16, d17, d6, d7, q9, q10, q11
	vst1.s16		{q8}, [r1]!

WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsQuantFour4x4Max_neon
	vld1.s16		{q2}, [r1]
	vld1.s16		{q3}, [r2]
	mov				r1, r0

	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q0, q8, d16, d17, d6, d7, q9, q10, q11, d26
	vst1.s16		{q8}, [r1]!
	vmov			q12, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q1, q12, d24, d25, d6, d7, q9, q10, q11, d28
	vst1.s16		{q12}, [r1]!		// then 1st 16 elem in d26 & d28

	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q0, q8, d16, d17, d6, d7, q9, q10, q11, d27
	vst1.s16		{q8}, [r1]!
	vmov			q12, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q1, q12, d24, d25, d6, d7, q9, q10, q11, d29
	vst1.s16		{q12}, [r1]!	// then 2nd 16 elem in d27 & d29

	SELECT_MAX_IN_ABS_COEF	q13, q14, q0, d0, d1
	vst1.s32		{d0[0]}, [r3]!

	///////////
	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q0, q8, d16, d17, d6, d7, q9, q10, q11, d26
	vst1.s16		{q8}, [r1]!
	vmov			q12, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q1, q12, d24, d25, d6, d7, q9, q10, q11, d28
	vst1.s16		{q12}, [r1]!		// then 3rd 16 elem in d26 & d28

	vld1.s16		{q0, q1}, [r0]!
	vmov			q8, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q0, q8, d16, d17, d6, d7, q9, q10, q11, d27
	vst1.s16		{q8}, [r1]!
	vmov			q12, q2
	NEWQUANT_COEF_EACH_16BITS_MAX	q1, q12, d24, d25, d6, d7, q9, q10, q11, d29
	vst1.s16		{q12}, [r1]!	// then 4th 16 elem in d27 & d29

	SELECT_MAX_IN_ABS_COEF	q13, q14, q0, d0, d1
	vst1.s32		{d0[0]}, [r3]!

WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsHadamardT4Dc_neon
	push	{r2,r3}
	mov		r2, #64	// 2*16*sizeof(int16_t)
	add		r3, r1, #32

	vld1.s16		{d0}, [r1], r2
	vld1.s16		{d1}, [r3], r2
	vld1.s16		{d4}, [r1], r2
	vld1.s16		{d5}, [r3], r2
	vld1.s16		{d2}, [r1], r2
	vld1.s16		{d3}, [r3], r2
	vld1.s16		{d6}, [r1], r2
	vld1.s16		{d7}, [r3], r2
	vtrn.16		q0, q2		// d0[0 4], d1[1 5]
	vtrn.16		q1, q3		// d2[2 6], d3[3 7]

	vld1.s16		{d16}, [r1], r2
	vld1.s16		{d17}, [r3], r2
	vld1.s16		{d20}, [r1], r2
	vld1.s16		{d21}, [r3], r2
	vld1.s16		{d18}, [r1], r2
	vld1.s16		{d19}, [r3], r2
	vld1.s16		{d22}, [r1], r2
	vld1.s16		{d23}, [r3], r2
	vtrn.16		q8, q10		//d16[08 12],d17[09 13]
	vtrn.16		q9, q11		//d18[10 14],d19[11 15]

	vtrn.32		q0, q8		// d0 [0 4 08 12] = dct[idx],		d1[1 5 09 13] = dct[idx+16]
	vtrn.32		q1, q9		// d2 [2 6 10 14] = dct[idx+64],	d3[3 7 11 15] = dct[idx+80]

	ROW_TRANSFORM_0_STEP	d0, d1, d3, d2, q8, q11, q10, q9

	TRANSFORM_4BYTES		q0, q1, q3, q2, q8, q11, q10, q9

	// transform element 32bits
	vtrn.s32		q0, q1				//[0 1 2 3]+[4 5 6 7]-->[0 4 2 6]+[1 5 3 7]
	vtrn.s32		q2, q3				//[8 9 10 11]+[12 13 14 15]-->[8 12 10 14]+[9 13 11 15]
	vswp			d1, d4				//[0 4 2 6]+[8 12 10 14]-->[0 4 8 12]+[2 6 10 14]
	vswp			d3, d6				//[1 5 3 7]+[9 13 11 15]-->[1 5 9 13]+[3 7 11 15]

	COL_TRANSFORM_0_STEP	q0, q1, q3, q2, q8, q11, q10, q9

	TRANSFORM_4BYTES		q0, q1, q3, q2, q8, q11, q10, q9

	vrshrn.s32		d16, q0, #1
	vrshrn.s32		d17, q1, #1
	vrshrn.s32		d18, q2, #1
	vrshrn.s32		d19, q3, #1
	vst1.16	{q8, q9}, [r0]	//store

	pop		{r2,r3}
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsHadamardQuant2x2_neon

	vdup.s16	d1, r1				//ff
	vdup.s16	d2, r2				//mf
	veor		d3, d3

	mov			r1, #32
	mov			r2, r0

	vld1.s16	{d0[0]}, [r0], r1		//rs[00]
	vst1.s16	{d3[0]}, [r2], r1		//rs[00]=0
	vld1.s16	{d0[1]}, [r0], r1		//rs[16]
	vst1.s16	{d3[0]}, [r2], r1		//rs[16]=0
	vld1.s16	{d0[2]}, [r0], r1		//rs[32]
	vst1.s16	{d3[0]}, [r2], r1		//rs[32]=0
	vld1.s16	{d0[3]}, [r0], r1		//rs[48]
	vst1.s16	{d3[0]}, [r2], r1		//rs[48]=0

	HDM_QUANT_2x2_TOTAL_16BITS	d0, d4, d5		// output d5

	HDM_QUANT_2x2_TOTAL_16BITS	d5, d4, d0		// output d0

	QUANT_DUALWORD_COEF_EACH_16BITS	d0, d1, d2, d3, q2

	vst1.s16	d1, [r3]		// store to dct
	ldr			r2, [sp, #0]
	vst1.s16	d1, [r2]		// store to block

	mov			r1, #1
	vdup.s16	d3, r1
	DC_ZERO_COUNT_IN_DUALWORD	d1, d0, d3

	vmov	r0, r1, d0
	and		r0, #0x07		// range [0~4]
	rsb		r0, #4
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsHadamardQuant2x2SkipKernel_neon

	vdup.s16	d3, r1
	mov			r1, #32
	vld1.s16	{d0[0]}, [r0], r1		//rs[00]
	vld1.s16	{d0[1]}, [r0], r1		//rs[16]
	vld1.s16	{d0[2]}, [r0], r1		//rs[32]
	vld1.s16	{d0[3]}, [r0], r1		//rs[48]

	HDM_QUANT_2x2_TOTAL_16BITS	d0, d1, d2		// output d2

	HDM_QUANT_2x2_TOTAL_16BITS	d2, d1, d0		// output d0

	vabs.s16	d1, d0
	vcgt.s16	d1, d1, d3		// abs(dct[i])>threshold;
	vmov	r0, r1, d1
	orr		r0, r1
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsGetNoneZeroCount_neon
	push	{r1}
	vld1.s16	{q0, q1}, [r0]
	vmov.s16	q8, #1

	ZERO_COUNT_IN_2_QUARWORD	q0, q1, q8, d0, d1, d2, d3
	vmov	r0, r1, d0
	and		r0, #0x1F	// range [0~16]
	rsb		r0, #16
	pop		{r1}
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsDequant4x4_neon
	vld1.s16	{q0, q1}, [r0]
	vld1.u16	{q2}, [r1]

	vmul.s16	q8, q0, q2
	vmul.s16	q9, q1, q2

	vst1.s16	{q8, q9}, [r0]
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsDequantFour4x4_neon
	vld1.u16	{q12}, [r1]
	mov		r1, r0
	vld1.s16	{q0, q1}, [r0]!
	vld1.s16	{q2, q3}, [r0]!
	vmul.s16	q0, q0, q12
	vld1.s16	{q8, q9}, [r0]!
	vmul.s16	q1, q1, q12
	vld1.s16	{q10, q11}, [r0]!

	vst1.s16	{q0, q1}, [r1]!

	vmul.s16	q2, q2, q12
	vmul.s16	q3, q3, q12
	vmul.s16	q8, q8, q12
	vst1.s16	{q2, q3}, [r1]!

	vmul.s16	q9, q9, q12
	vmul.s16	q10, q10, q12
	vmul.s16	q11, q11, q12
	vst1.s16	{q8, q9}, [r1]!
	vst1.s16	{q10, q11}, [r1]!

WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsDequantIHadamard4x4_neon

	vld1.s16	{q0, q1}, [r0]
	vdup.s16	q8, r1

	IHDM_4x4_TOTAL_16BITS	q0, q2, q3
	IHDM_4x4_TOTAL_16BITS	q1, q2, q3

	MATRIX_TRANSFORM_EACH_16BITS	d0, d1, d2, d3

	IHDM_4x4_TOTAL_16BITS	q0, q2, q3
	vmul.s16	q0, q8

	IHDM_4x4_TOTAL_16BITS	q1, q2, q3
	vmul.s16	q1, q8

	MATRIX_TRANSFORM_EACH_16BITS	d0, d1, d2, d3
	vst1.s16	{q0, q1}, [r0]
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsIDctT4Rec_neon
	vld1.u32		{d16[0]}, [r2], r3
	push			{r4}
	ldr				r4, [sp, #4]
	vld1.u32		{d16[1]}, [r2], r3

	vld4.s16		{d0, d1, d2, d3}, [r4]		// cost 3 cycles!
	vld1.u32		{d17[0]}, [r2], r3
	vld1.u32		{d17[1]}, [r2], r3			// q7 is pred

	ROW_TRANSFORM_1_STEP_TOTAL_16BITS		d0, d1, d2, d3, d4, d5, d6, d7

	TRANSFORM_TOTAL_16BITS		d0, d1, d2, d3, d4, d5, d6, d7

	MATRIX_TRANSFORM_EACH_16BITS	d0, d1, d2, d3

	ROW_TRANSFORM_1_STEP_TOTAL_16BITS		d0, d1, d2, d3, d4, d5, d6, d7

	TRANSFORM_TOTAL_16BITS		d0, d1, d2, d3, d4, d5, d6, d7
	vrshr.s16		d0, d0, #6
	vrshr.s16		d1, d1, #6
	vrshr.s16		d2, d2, #6
	vrshr.s16		d3, d3, #6

	//after rounding 6, clip into [0, 255]
	vmovl.u8		q2,d16
	vadd.s16		q0,q2
	vqmovun.s16	d16,q0
	vst1.32		{d16[0]},[r0],r1
	vst1.32		{d16[1]},[r0],r1

	vmovl.u8		q2,d17
	vadd.s16		q1,q2
	vqmovun.s16	d17,q1
	vst1.32		{d17[0]},[r0],r1
	vst1.32		{d17[1]},[r0]

	pop			{r4}
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsIDctFourT4Rec_neon

	vld1.u64		{d24}, [r2], r3
	push			{r4}
	ldr				r4, [sp, #4]
	vld1.u64		{d25}, [r2], r3

	vld4.s16		{d0, d1, d2, d3}, [r4]!		// cost 3 cycles!
	vld1.u64		{d26}, [r2], r3
	vld1.u64		{d27}, [r2], r3
	vld4.s16		{d4, d5, d6, d7}, [r4]!		// cost 3 cycles!
	vswp			d1, d4
	vswp			d3, d6
	vswp			q1, q2						// q0~q3

	ROW_TRANSFORM_1_STEP_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	MATRIX_TRANSFORM_EACH_16BITS	q0, q1, q2, q3

	ROW_TRANSFORM_1_STEP_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11
	vrshr.s16		q0, q0, #6
	vrshr.s16		q1, q1, #6
	vrshr.s16		q2, q2, #6
	vrshr.s16		q3, q3, #6

	//after rounding 6, clip into [0, 255]
	vmovl.u8		q8,d24
	vadd.s16		q0,q8
	vqmovun.s16	d24,q0
	vst1.u8		{d24},[r0],r1

	vmovl.u8		q8,d25
	vadd.s16		q1,q8
	vqmovun.s16	d25,q1
	vst1.u8		{d25},[r0],r1

	vmovl.u8		q8,d26
	vadd.s16		q2,q8
	vqmovun.s16	d26,q2
	vst1.u8		{d26},[r0],r1

	vmovl.u8		q8,d27
	vadd.s16		q3,q8
	vqmovun.s16	d27,q3
	vst1.u8		{d27},[r0],r1

	vld1.u64		{d24}, [r2], r3
	vld1.u64		{d25}, [r2], r3

	vld4.s16		{d0, d1, d2, d3}, [r4]!		// cost 3 cycles!
	vld1.u64		{d26}, [r2], r3
	vld1.u64		{d27}, [r2], r3
	vld4.s16		{d4, d5, d6, d7}, [r4]!		// cost 3 cycles!
	vswp			d1, d4
	vswp			d3, d6
	vswp			q1, q2						// q0~q3

	ROW_TRANSFORM_1_STEP_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	MATRIX_TRANSFORM_EACH_16BITS	q0, q1, q2, q3

	ROW_TRANSFORM_1_STEP_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11

	TRANSFORM_TOTAL_16BITS		q0, q1, q2, q3, q8, q9, q10, q11
	vrshr.s16		q0, q0, #6
	vrshr.s16		q1, q1, #6
	vrshr.s16		q2, q2, #6
	vrshr.s16		q3, q3, #6

	//after rounding 6, clip into [0, 255]
	vmovl.u8		q8,d24
	vadd.s16		q0,q8
	vqmovun.s16	d24,q0
	vst1.u8		{d24},[r0],r1

	vmovl.u8		q8,d25
	vadd.s16		q1,q8
	vqmovun.s16	d25,q1
	vst1.u8		{d25},[r0],r1

	vmovl.u8		q8,d26
	vadd.s16		q2,q8
	vqmovun.s16	d26,q2
	vst1.u8		{d26},[r0],r1

	vmovl.u8		q8,d27
	vadd.s16		q3,q8
	vqmovun.s16	d27,q3
	vst1.u8		{d27},[r0],r1

	pop			{r4}
WELS_ASM_FUNC_END


WELS_ASM_FUNC_BEGIN WelsIDctRecI16x16Dc_neon
	push		{r4}
	ldr			r4, [sp, #4]

	vld1.s16	{q8,q9}, [r4]
	vrshr.s16		q8, q8, #6
	vrshr.s16		q9, q9, #6

	vdup.s16	d20, d16[0]
	vdup.s16	d21, d16[1]
	vdup.s16	d22, d16[2]
	vdup.s16	d23, d16[3]

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vdup.s16	d20, d17[0]
	vdup.s16	d21, d17[1]
	vdup.s16	d22, d17[2]
	vdup.s16	d23, d17[3]

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vdup.s16	d20, d18[0]
	vdup.s16	d21, d18[1]
	vdup.s16	d22, d18[2]
	vdup.s16	d23, d18[3]

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vdup.s16	d20, d19[0]
	vdup.s16	d21, d19[1]
	vdup.s16	d22, d19[2]
	vdup.s16	d23, d19[3]

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	vld1.u8	{q0}, [r2], r3
	MB_PRED_8BITS_ADD_DCT_16BITS_CLIP	d0, d1, q10, q11, q12, q13
	vst1.u8	{q0}, [r0], r1

	pop			{r4}
WELS_ASM_FUNC_END
#endif