• file
• revisions
• annotate
• diff
• comparison
• raw

media/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2.c@ac0c01689b40 (annotated)

media/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2.c

Thu, 15 Jan 2015 15:59:08 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Thu, 15 Jan 2015 15:59:08 +0100

branch

TOR_BUG_9701

changeset 10

ac0c01689b40

permissions

-rw-r--r--

Implement a real Private Browsing Mode condition by changing the API/ABI;
This solves Tor bug #9701, complying with disk avoidance documented in
https://www.torproject.org/projects/torbrowser/design/#disk-avoidance.

michael@0	1	/*
michael@0	2	* Copyright (c) 2012 The WebM project authors. All Rights Reserved.
michael@0	3	*
michael@0	4	* Use of this source code is governed by a BSD-style license
michael@0	5	* that can be found in the LICENSE file in the root of the source
michael@0	6	* tree. An additional intellectual property rights grant can be found
michael@0	7	* in the file PATENTS. All contributing project authors may
michael@0	8	* be found in the AUTHORS file in the root of the source tree.
michael@0	9	*/
michael@0	10
michael@0	11	#include <emmintrin.h> // SSE2
michael@0	12	#include "vp9/common/vp9_idct.h" // for cospi constants
michael@0	13	#include "vpx_ports/mem.h"
michael@0	14
michael@0	15	#if FDCT32x32_HIGH_PRECISION
michael@0	16	static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
michael@0	17	__m128i buf0, buf1;
michael@0	18	buf0 = _mm_mul_epu32(a, b);
michael@0	19	a = _mm_srli_epi64(a, 32);
michael@0	20	b = _mm_srli_epi64(b, 32);
michael@0	21	buf1 = _mm_mul_epu32(a, b);
michael@0	22	return _mm_add_epi64(buf0, buf1);
michael@0	23	}
michael@0	24
michael@0	25	static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) {
michael@0	26	__m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
michael@0	27	__m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
michael@0	28	return _mm_unpacklo_epi64(buf0, buf1);
michael@0	29	}
michael@0	30	#endif
michael@0	31
michael@0	32	void FDCT32x32_2D(const int16_t *input,
michael@0	33	int16_t *output_org, int stride) {
michael@0	34	// Calculate pre-multiplied strides
michael@0	35	const int str1 = stride;
michael@0	36	const int str2 = 2 * stride;
michael@0	37	const int str3 = 2 * stride + str1;
michael@0	38	// We need an intermediate buffer between passes.
michael@0	39	DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]);
michael@0	40	// Constants
michael@0	41	// When we use them, in one case, they are all the same. In all others
michael@0	42	// it's a pair of them that we need to repeat four times. This is done
michael@0	43	// by constructing the 32 bit constant corresponding to that pair.
michael@0	44	const __m128i k__cospi_p16_p16 = _mm_set1_epi16(+cospi_16_64);
michael@0	45	const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
michael@0	46	const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
michael@0	47	const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
michael@0	48	const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64, cospi_8_64);
michael@0	49	const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64, cospi_20_64);
michael@0	50	const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
michael@0	51	const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
michael@0	52	const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64, cospi_4_64);
michael@0	53	const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
michael@0	54	const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
michael@0	55	const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64, cospi_2_64);
michael@0	56	const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64, cospi_18_64);
michael@0	57	const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64, cospi_10_64);
michael@0	58	const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64, cospi_26_64);
michael@0	59	const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
michael@0	60	const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
michael@0	61	const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
michael@0	62	const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
michael@0	63	const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64, cospi_1_64);
michael@0	64	const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64, cospi_17_64);
michael@0	65	const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64, cospi_9_64);
michael@0	66	const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64, cospi_25_64);
michael@0	67	const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64, cospi_7_64);
michael@0	68	const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64, cospi_23_64);
michael@0	69	const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64, cospi_15_64);
michael@0	70	const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64, cospi_31_64);
michael@0	71	const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64, cospi_5_64);
michael@0	72	const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64, cospi_21_64);
michael@0	73	const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64, cospi_13_64);
michael@0	74	const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64, cospi_29_64);
michael@0	75	const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64, cospi_3_64);
michael@0	76	const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64, cospi_19_64);
michael@0	77	const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64, cospi_11_64);
michael@0	78	const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64, cospi_27_64);
michael@0	79	const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
michael@0	80	const __m128i kZero = _mm_set1_epi16(0);
michael@0	81	const __m128i kOne = _mm_set1_epi16(1);
michael@0	82	// Do the two transform/transpose passes
michael@0	83	int pass;
michael@0	84	for (pass = 0; pass < 2; ++pass) {
michael@0	85	// We process eight columns (transposed rows in second pass) at a time.
michael@0	86	int column_start;
michael@0	87	for (column_start = 0; column_start < 32; column_start += 8) {
michael@0	88	__m128i step1[32];
michael@0	89	__m128i step2[32];
michael@0	90	__m128i step3[32];
michael@0	91	__m128i out[32];
michael@0	92	// Stage 1
michael@0	93	// Note: even though all the loads below are aligned, using the aligned
michael@0	94	// intrinsic make the code slightly slower.
michael@0	95	if (0 == pass) {
michael@0	96	const int16_t *in = &input[column_start];
michael@0	97	// step1[i] = (in[ 0 * stride] + in[(32 - 1) * stride]) << 2;
michael@0	98	// Note: the next four blocks could be in a loop. That would help the
michael@0	99	// instruction cache but is actually slower.
michael@0	100	{
michael@0	101	const int16_t *ina = in + 0 * str1;
michael@0	102	const int16_t *inb = in + 31 * str1;
michael@0	103	__m128i *step1a = &step1[ 0];
michael@0	104	__m128i *step1b = &step1[31];
michael@0	105	const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
michael@0	106	const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
michael@0	107	const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
michael@0	108	const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
michael@0	109	const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
michael@0	110	const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
michael@0	111	const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
michael@0	112	const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
michael@0	113	step1a[ 0] = _mm_add_epi16(ina0, inb0);
michael@0	114	step1a[ 1] = _mm_add_epi16(ina1, inb1);
michael@0	115	step1a[ 2] = _mm_add_epi16(ina2, inb2);
michael@0	116	step1a[ 3] = _mm_add_epi16(ina3, inb3);
michael@0	117	step1b[-3] = _mm_sub_epi16(ina3, inb3);
michael@0	118	step1b[-2] = _mm_sub_epi16(ina2, inb2);
michael@0	119	step1b[-1] = _mm_sub_epi16(ina1, inb1);
michael@0	120	step1b[-0] = _mm_sub_epi16(ina0, inb0);
michael@0	121	step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
michael@0	122	step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
michael@0	123	step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
michael@0	124	step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
michael@0	125	step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
michael@0	126	step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
michael@0	127	step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
michael@0	128	step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
michael@0	129	}
michael@0	130	{
michael@0	131	const int16_t *ina = in + 4 * str1;
michael@0	132	const int16_t *inb = in + 27 * str1;
michael@0	133	__m128i *step1a = &step1[ 4];
michael@0	134	__m128i *step1b = &step1[27];
michael@0	135	const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
michael@0	136	const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
michael@0	137	const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
michael@0	138	const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
michael@0	139	const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
michael@0	140	const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
michael@0	141	const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
michael@0	142	const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
michael@0	143	step1a[ 0] = _mm_add_epi16(ina0, inb0);
michael@0	144	step1a[ 1] = _mm_add_epi16(ina1, inb1);
michael@0	145	step1a[ 2] = _mm_add_epi16(ina2, inb2);
michael@0	146	step1a[ 3] = _mm_add_epi16(ina3, inb3);
michael@0	147	step1b[-3] = _mm_sub_epi16(ina3, inb3);
michael@0	148	step1b[-2] = _mm_sub_epi16(ina2, inb2);
michael@0	149	step1b[-1] = _mm_sub_epi16(ina1, inb1);
michael@0	150	step1b[-0] = _mm_sub_epi16(ina0, inb0);
michael@0	151	step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
michael@0	152	step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
michael@0	153	step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
michael@0	154	step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
michael@0	155	step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
michael@0	156	step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
michael@0	157	step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
michael@0	158	step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
michael@0	159	}
michael@0	160	{
michael@0	161	const int16_t *ina = in + 8 * str1;
michael@0	162	const int16_t *inb = in + 23 * str1;
michael@0	163	__m128i *step1a = &step1[ 8];
michael@0	164	__m128i *step1b = &step1[23];
michael@0	165	const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
michael@0	166	const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
michael@0	167	const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
michael@0	168	const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
michael@0	169	const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
michael@0	170	const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
michael@0	171	const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
michael@0	172	const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
michael@0	173	step1a[ 0] = _mm_add_epi16(ina0, inb0);
michael@0	174	step1a[ 1] = _mm_add_epi16(ina1, inb1);
michael@0	175	step1a[ 2] = _mm_add_epi16(ina2, inb2);
michael@0	176	step1a[ 3] = _mm_add_epi16(ina3, inb3);
michael@0	177	step1b[-3] = _mm_sub_epi16(ina3, inb3);
michael@0	178	step1b[-2] = _mm_sub_epi16(ina2, inb2);
michael@0	179	step1b[-1] = _mm_sub_epi16(ina1, inb1);
michael@0	180	step1b[-0] = _mm_sub_epi16(ina0, inb0);
michael@0	181	step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
michael@0	182	step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
michael@0	183	step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
michael@0	184	step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
michael@0	185	step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
michael@0	186	step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
michael@0	187	step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
michael@0	188	step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
michael@0	189	}
michael@0	190	{
michael@0	191	const int16_t *ina = in + 12 * str1;
michael@0	192	const int16_t *inb = in + 19 * str1;
michael@0	193	__m128i *step1a = &step1[12];
michael@0	194	__m128i *step1b = &step1[19];
michael@0	195	const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
michael@0	196	const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
michael@0	197	const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
michael@0	198	const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
michael@0	199	const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
michael@0	200	const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
michael@0	201	const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
michael@0	202	const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
michael@0	203	step1a[ 0] = _mm_add_epi16(ina0, inb0);
michael@0	204	step1a[ 1] = _mm_add_epi16(ina1, inb1);
michael@0	205	step1a[ 2] = _mm_add_epi16(ina2, inb2);
michael@0	206	step1a[ 3] = _mm_add_epi16(ina3, inb3);
michael@0	207	step1b[-3] = _mm_sub_epi16(ina3, inb3);
michael@0	208	step1b[-2] = _mm_sub_epi16(ina2, inb2);
michael@0	209	step1b[-1] = _mm_sub_epi16(ina1, inb1);
michael@0	210	step1b[-0] = _mm_sub_epi16(ina0, inb0);
michael@0	211	step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
michael@0	212	step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
michael@0	213	step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
michael@0	214	step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
michael@0	215	step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
michael@0	216	step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
michael@0	217	step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
michael@0	218	step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
michael@0	219	}
michael@0	220	} else {
michael@0	221	int16_t *in = &intermediate[column_start];
michael@0	222	// step1[i] = in[ 0 * 32] + in[(32 - 1) * 32];
michael@0	223	// Note: using the same approach as above to have common offset is
michael@0	224	// counter-productive as all offsets can be calculated at compile
michael@0	225	// time.
michael@0	226	// Note: the next four blocks could be in a loop. That would help the
michael@0	227	// instruction cache but is actually slower.
michael@0	228	{
michael@0	229	__m128i in00 = _mm_loadu_si128((const __m128i *)(in + 0 * 32));
michael@0	230	__m128i in01 = _mm_loadu_si128((const __m128i *)(in + 1 * 32));
michael@0	231	__m128i in02 = _mm_loadu_si128((const __m128i *)(in + 2 * 32));
michael@0	232	__m128i in03 = _mm_loadu_si128((const __m128i *)(in + 3 * 32));
michael@0	233	__m128i in28 = _mm_loadu_si128((const __m128i *)(in + 28 * 32));
michael@0	234	__m128i in29 = _mm_loadu_si128((const __m128i *)(in + 29 * 32));
michael@0	235	__m128i in30 = _mm_loadu_si128((const __m128i *)(in + 30 * 32));
michael@0	236	__m128i in31 = _mm_loadu_si128((const __m128i *)(in + 31 * 32));
michael@0	237	step1[ 0] = _mm_add_epi16(in00, in31);
michael@0	238	step1[ 1] = _mm_add_epi16(in01, in30);
michael@0	239	step1[ 2] = _mm_add_epi16(in02, in29);
michael@0	240	step1[ 3] = _mm_add_epi16(in03, in28);
michael@0	241	step1[28] = _mm_sub_epi16(in03, in28);
michael@0	242	step1[29] = _mm_sub_epi16(in02, in29);
michael@0	243	step1[30] = _mm_sub_epi16(in01, in30);
michael@0	244	step1[31] = _mm_sub_epi16(in00, in31);
michael@0	245	}
michael@0	246	{
michael@0	247	__m128i in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 32));
michael@0	248	__m128i in05 = _mm_loadu_si128((const __m128i *)(in + 5 * 32));
michael@0	249	__m128i in06 = _mm_loadu_si128((const __m128i *)(in + 6 * 32));
michael@0	250	__m128i in07 = _mm_loadu_si128((const __m128i *)(in + 7 * 32));
michael@0	251	__m128i in24 = _mm_loadu_si128((const __m128i *)(in + 24 * 32));
michael@0	252	__m128i in25 = _mm_loadu_si128((const __m128i *)(in + 25 * 32));
michael@0	253	__m128i in26 = _mm_loadu_si128((const __m128i *)(in + 26 * 32));
michael@0	254	__m128i in27 = _mm_loadu_si128((const __m128i *)(in + 27 * 32));
michael@0	255	step1[ 4] = _mm_add_epi16(in04, in27);
michael@0	256	step1[ 5] = _mm_add_epi16(in05, in26);
michael@0	257	step1[ 6] = _mm_add_epi16(in06, in25);
michael@0	258	step1[ 7] = _mm_add_epi16(in07, in24);
michael@0	259	step1[24] = _mm_sub_epi16(in07, in24);
michael@0	260	step1[25] = _mm_sub_epi16(in06, in25);
michael@0	261	step1[26] = _mm_sub_epi16(in05, in26);
michael@0	262	step1[27] = _mm_sub_epi16(in04, in27);
michael@0	263	}
michael@0	264	{
michael@0	265	__m128i in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 32));
michael@0	266	__m128i in09 = _mm_loadu_si128((const __m128i *)(in + 9 * 32));
michael@0	267	__m128i in10 = _mm_loadu_si128((const __m128i *)(in + 10 * 32));
michael@0	268	__m128i in11 = _mm_loadu_si128((const __m128i *)(in + 11 * 32));
michael@0	269	__m128i in20 = _mm_loadu_si128((const __m128i *)(in + 20 * 32));
michael@0	270	__m128i in21 = _mm_loadu_si128((const __m128i *)(in + 21 * 32));
michael@0	271	__m128i in22 = _mm_loadu_si128((const __m128i *)(in + 22 * 32));
michael@0	272	__m128i in23 = _mm_loadu_si128((const __m128i *)(in + 23 * 32));
michael@0	273	step1[ 8] = _mm_add_epi16(in08, in23);
michael@0	274	step1[ 9] = _mm_add_epi16(in09, in22);
michael@0	275	step1[10] = _mm_add_epi16(in10, in21);
michael@0	276	step1[11] = _mm_add_epi16(in11, in20);
michael@0	277	step1[20] = _mm_sub_epi16(in11, in20);
michael@0	278	step1[21] = _mm_sub_epi16(in10, in21);
michael@0	279	step1[22] = _mm_sub_epi16(in09, in22);
michael@0	280	step1[23] = _mm_sub_epi16(in08, in23);
michael@0	281	}
michael@0	282	{
michael@0	283	__m128i in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 32));
michael@0	284	__m128i in13 = _mm_loadu_si128((const __m128i *)(in + 13 * 32));
michael@0	285	__m128i in14 = _mm_loadu_si128((const __m128i *)(in + 14 * 32));
michael@0	286	__m128i in15 = _mm_loadu_si128((const __m128i *)(in + 15 * 32));
michael@0	287	__m128i in16 = _mm_loadu_si128((const __m128i *)(in + 16 * 32));
michael@0	288	__m128i in17 = _mm_loadu_si128((const __m128i *)(in + 17 * 32));
michael@0	289	__m128i in18 = _mm_loadu_si128((const __m128i *)(in + 18 * 32));
michael@0	290	__m128i in19 = _mm_loadu_si128((const __m128i *)(in + 19 * 32));
michael@0	291	step1[12] = _mm_add_epi16(in12, in19);
michael@0	292	step1[13] = _mm_add_epi16(in13, in18);
michael@0	293	step1[14] = _mm_add_epi16(in14, in17);
michael@0	294	step1[15] = _mm_add_epi16(in15, in16);
michael@0	295	step1[16] = _mm_sub_epi16(in15, in16);
michael@0	296	step1[17] = _mm_sub_epi16(in14, in17);
michael@0	297	step1[18] = _mm_sub_epi16(in13, in18);
michael@0	298	step1[19] = _mm_sub_epi16(in12, in19);
michael@0	299	}
michael@0	300	}
michael@0	301	// Stage 2
michael@0	302	{
michael@0	303	step2[ 0] = _mm_add_epi16(step1[0], step1[15]);
michael@0	304	step2[ 1] = _mm_add_epi16(step1[1], step1[14]);
michael@0	305	step2[ 2] = _mm_add_epi16(step1[2], step1[13]);
michael@0	306	step2[ 3] = _mm_add_epi16(step1[3], step1[12]);
michael@0	307	step2[ 4] = _mm_add_epi16(step1[4], step1[11]);
michael@0	308	step2[ 5] = _mm_add_epi16(step1[5], step1[10]);
michael@0	309	step2[ 6] = _mm_add_epi16(step1[6], step1[ 9]);
michael@0	310	step2[ 7] = _mm_add_epi16(step1[7], step1[ 8]);
michael@0	311	step2[ 8] = _mm_sub_epi16(step1[7], step1[ 8]);
michael@0	312	step2[ 9] = _mm_sub_epi16(step1[6], step1[ 9]);
michael@0	313	step2[10] = _mm_sub_epi16(step1[5], step1[10]);
michael@0	314	step2[11] = _mm_sub_epi16(step1[4], step1[11]);
michael@0	315	step2[12] = _mm_sub_epi16(step1[3], step1[12]);
michael@0	316	step2[13] = _mm_sub_epi16(step1[2], step1[13]);
michael@0	317	step2[14] = _mm_sub_epi16(step1[1], step1[14]);
michael@0	318	step2[15] = _mm_sub_epi16(step1[0], step1[15]);
michael@0	319	}
michael@0	320	{
michael@0	321	const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
michael@0	322	const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
michael@0	323	const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
michael@0	324	const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
michael@0	325	const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
michael@0	326	const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
michael@0	327	const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
michael@0	328	const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
michael@0	329	const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
michael@0	330	const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
michael@0	331	const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
michael@0	332	const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
michael@0	333	const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
michael@0	334	const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
michael@0	335	const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
michael@0	336	const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
michael@0	337	const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
michael@0	338	const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
michael@0	339	const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
michael@0	340	const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
michael@0	341	const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
michael@0	342	const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
michael@0	343	const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
michael@0	344	const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
michael@0	345	// dct_const_round_shift
michael@0	346	const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
michael@0	347	const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
michael@0	348	const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
michael@0	349	const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
michael@0	350	const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
michael@0	351	const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
michael@0	352	const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
michael@0	353	const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
michael@0	354	const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
michael@0	355	const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
michael@0	356	const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
michael@0	357	const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
michael@0	358	const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
michael@0	359	const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
michael@0	360	const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
michael@0	361	const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
michael@0	362	const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
michael@0	363	const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
michael@0	364	const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
michael@0	365	const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
michael@0	366	const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
michael@0	367	const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
michael@0	368	const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
michael@0	369	const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
michael@0	370	const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
michael@0	371	const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
michael@0	372	const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
michael@0	373	const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
michael@0	374	const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
michael@0	375	const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
michael@0	376	const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
michael@0	377	const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
michael@0	378	// Combine
michael@0	379	step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
michael@0	380	step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
michael@0	381	step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
michael@0	382	step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
michael@0	383	step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
michael@0	384	step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
michael@0	385	step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
michael@0	386	step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
michael@0	387	}
michael@0	388
michael@0	389	#if !FDCT32x32_HIGH_PRECISION
michael@0	390	// dump the magnitude by half, hence the intermediate values are within
michael@0	391	// the range of 16 bits.
michael@0	392	if (1 == pass) {
michael@0	393	__m128i s3_00_0 = _mm_cmplt_epi16(step2[ 0], kZero);
michael@0	394	__m128i s3_01_0 = _mm_cmplt_epi16(step2[ 1], kZero);
michael@0	395	__m128i s3_02_0 = _mm_cmplt_epi16(step2[ 2], kZero);
michael@0	396	__m128i s3_03_0 = _mm_cmplt_epi16(step2[ 3], kZero);
michael@0	397	__m128i s3_04_0 = _mm_cmplt_epi16(step2[ 4], kZero);
michael@0	398	__m128i s3_05_0 = _mm_cmplt_epi16(step2[ 5], kZero);
michael@0	399	__m128i s3_06_0 = _mm_cmplt_epi16(step2[ 6], kZero);
michael@0	400	__m128i s3_07_0 = _mm_cmplt_epi16(step2[ 7], kZero);
michael@0	401	__m128i s2_08_0 = _mm_cmplt_epi16(step2[ 8], kZero);
michael@0	402	__m128i s2_09_0 = _mm_cmplt_epi16(step2[ 9], kZero);
michael@0	403	__m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero);
michael@0	404	__m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero);
michael@0	405	__m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero);
michael@0	406	__m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero);
michael@0	407	__m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero);
michael@0	408	__m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero);
michael@0	409	__m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero);
michael@0	410	__m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero);
michael@0	411	__m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero);
michael@0	412	__m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero);
michael@0	413	__m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero);
michael@0	414	__m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero);
michael@0	415	__m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero);
michael@0	416	__m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero);
michael@0	417	__m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero);
michael@0	418	__m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero);
michael@0	419	__m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero);
michael@0	420	__m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero);
michael@0	421	__m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero);
michael@0	422	__m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero);
michael@0	423	__m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero);
michael@0	424	__m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero);
michael@0	425
michael@0	426	step2[ 0] = _mm_sub_epi16(step2[ 0], s3_00_0);
michael@0	427	step2[ 1] = _mm_sub_epi16(step2[ 1], s3_01_0);
michael@0	428	step2[ 2] = _mm_sub_epi16(step2[ 2], s3_02_0);
michael@0	429	step2[ 3] = _mm_sub_epi16(step2[ 3], s3_03_0);
michael@0	430	step2[ 4] = _mm_sub_epi16(step2[ 4], s3_04_0);
michael@0	431	step2[ 5] = _mm_sub_epi16(step2[ 5], s3_05_0);
michael@0	432	step2[ 6] = _mm_sub_epi16(step2[ 6], s3_06_0);
michael@0	433	step2[ 7] = _mm_sub_epi16(step2[ 7], s3_07_0);
michael@0	434	step2[ 8] = _mm_sub_epi16(step2[ 8], s2_08_0);
michael@0	435	step2[ 9] = _mm_sub_epi16(step2[ 9], s2_09_0);
michael@0	436	step2[10] = _mm_sub_epi16(step2[10], s3_10_0);
michael@0	437	step2[11] = _mm_sub_epi16(step2[11], s3_11_0);
michael@0	438	step2[12] = _mm_sub_epi16(step2[12], s3_12_0);
michael@0	439	step2[13] = _mm_sub_epi16(step2[13], s3_13_0);
michael@0	440	step2[14] = _mm_sub_epi16(step2[14], s2_14_0);
michael@0	441	step2[15] = _mm_sub_epi16(step2[15], s2_15_0);
michael@0	442	step1[16] = _mm_sub_epi16(step1[16], s3_16_0);
michael@0	443	step1[17] = _mm_sub_epi16(step1[17], s3_17_0);
michael@0	444	step1[18] = _mm_sub_epi16(step1[18], s3_18_0);
michael@0	445	step1[19] = _mm_sub_epi16(step1[19], s3_19_0);
michael@0	446	step2[20] = _mm_sub_epi16(step2[20], s3_20_0);
michael@0	447	step2[21] = _mm_sub_epi16(step2[21], s3_21_0);
michael@0	448	step2[22] = _mm_sub_epi16(step2[22], s3_22_0);
michael@0	449	step2[23] = _mm_sub_epi16(step2[23], s3_23_0);
michael@0	450	step2[24] = _mm_sub_epi16(step2[24], s3_24_0);
michael@0	451	step2[25] = _mm_sub_epi16(step2[25], s3_25_0);
michael@0	452	step2[26] = _mm_sub_epi16(step2[26], s3_26_0);
michael@0	453	step2[27] = _mm_sub_epi16(step2[27], s3_27_0);
michael@0	454	step1[28] = _mm_sub_epi16(step1[28], s3_28_0);
michael@0	455	step1[29] = _mm_sub_epi16(step1[29], s3_29_0);
michael@0	456	step1[30] = _mm_sub_epi16(step1[30], s3_30_0);
michael@0	457	step1[31] = _mm_sub_epi16(step1[31], s3_31_0);
michael@0	458
michael@0	459	step2[ 0] = _mm_add_epi16(step2[ 0], kOne);
michael@0	460	step2[ 1] = _mm_add_epi16(step2[ 1], kOne);
michael@0	461	step2[ 2] = _mm_add_epi16(step2[ 2], kOne);
michael@0	462	step2[ 3] = _mm_add_epi16(step2[ 3], kOne);
michael@0	463	step2[ 4] = _mm_add_epi16(step2[ 4], kOne);
michael@0	464	step2[ 5] = _mm_add_epi16(step2[ 5], kOne);
michael@0	465	step2[ 6] = _mm_add_epi16(step2[ 6], kOne);
michael@0	466	step2[ 7] = _mm_add_epi16(step2[ 7], kOne);
michael@0	467	step2[ 8] = _mm_add_epi16(step2[ 8], kOne);
michael@0	468	step2[ 9] = _mm_add_epi16(step2[ 9], kOne);
michael@0	469	step2[10] = _mm_add_epi16(step2[10], kOne);
michael@0	470	step2[11] = _mm_add_epi16(step2[11], kOne);
michael@0	471	step2[12] = _mm_add_epi16(step2[12], kOne);
michael@0	472	step2[13] = _mm_add_epi16(step2[13], kOne);
michael@0	473	step2[14] = _mm_add_epi16(step2[14], kOne);
michael@0	474	step2[15] = _mm_add_epi16(step2[15], kOne);
michael@0	475	step1[16] = _mm_add_epi16(step1[16], kOne);
michael@0	476	step1[17] = _mm_add_epi16(step1[17], kOne);
michael@0	477	step1[18] = _mm_add_epi16(step1[18], kOne);
michael@0	478	step1[19] = _mm_add_epi16(step1[19], kOne);
michael@0	479	step2[20] = _mm_add_epi16(step2[20], kOne);
michael@0	480	step2[21] = _mm_add_epi16(step2[21], kOne);
michael@0	481	step2[22] = _mm_add_epi16(step2[22], kOne);
michael@0	482	step2[23] = _mm_add_epi16(step2[23], kOne);
michael@0	483	step2[24] = _mm_add_epi16(step2[24], kOne);
michael@0	484	step2[25] = _mm_add_epi16(step2[25], kOne);
michael@0	485	step2[26] = _mm_add_epi16(step2[26], kOne);
michael@0	486	step2[27] = _mm_add_epi16(step2[27], kOne);
michael@0	487	step1[28] = _mm_add_epi16(step1[28], kOne);
michael@0	488	step1[29] = _mm_add_epi16(step1[29], kOne);
michael@0	489	step1[30] = _mm_add_epi16(step1[30], kOne);
michael@0	490	step1[31] = _mm_add_epi16(step1[31], kOne);
michael@0	491
michael@0	492	step2[ 0] = _mm_srai_epi16(step2[ 0], 2);
michael@0	493	step2[ 1] = _mm_srai_epi16(step2[ 1], 2);
michael@0	494	step2[ 2] = _mm_srai_epi16(step2[ 2], 2);
michael@0	495	step2[ 3] = _mm_srai_epi16(step2[ 3], 2);
michael@0	496	step2[ 4] = _mm_srai_epi16(step2[ 4], 2);
michael@0	497	step2[ 5] = _mm_srai_epi16(step2[ 5], 2);
michael@0	498	step2[ 6] = _mm_srai_epi16(step2[ 6], 2);
michael@0	499	step2[ 7] = _mm_srai_epi16(step2[ 7], 2);
michael@0	500	step2[ 8] = _mm_srai_epi16(step2[ 8], 2);
michael@0	501	step2[ 9] = _mm_srai_epi16(step2[ 9], 2);
michael@0	502	step2[10] = _mm_srai_epi16(step2[10], 2);
michael@0	503	step2[11] = _mm_srai_epi16(step2[11], 2);
michael@0	504	step2[12] = _mm_srai_epi16(step2[12], 2);
michael@0	505	step2[13] = _mm_srai_epi16(step2[13], 2);
michael@0	506	step2[14] = _mm_srai_epi16(step2[14], 2);
michael@0	507	step2[15] = _mm_srai_epi16(step2[15], 2);
michael@0	508	step1[16] = _mm_srai_epi16(step1[16], 2);
michael@0	509	step1[17] = _mm_srai_epi16(step1[17], 2);
michael@0	510	step1[18] = _mm_srai_epi16(step1[18], 2);
michael@0	511	step1[19] = _mm_srai_epi16(step1[19], 2);
michael@0	512	step2[20] = _mm_srai_epi16(step2[20], 2);
michael@0	513	step2[21] = _mm_srai_epi16(step2[21], 2);
michael@0	514	step2[22] = _mm_srai_epi16(step2[22], 2);
michael@0	515	step2[23] = _mm_srai_epi16(step2[23], 2);
michael@0	516	step2[24] = _mm_srai_epi16(step2[24], 2);
michael@0	517	step2[25] = _mm_srai_epi16(step2[25], 2);
michael@0	518	step2[26] = _mm_srai_epi16(step2[26], 2);
michael@0	519	step2[27] = _mm_srai_epi16(step2[27], 2);
michael@0	520	step1[28] = _mm_srai_epi16(step1[28], 2);
michael@0	521	step1[29] = _mm_srai_epi16(step1[29], 2);
michael@0	522	step1[30] = _mm_srai_epi16(step1[30], 2);
michael@0	523	step1[31] = _mm_srai_epi16(step1[31], 2);
michael@0	524	}
michael@0	525	#endif
michael@0	526
michael@0	527	#if FDCT32x32_HIGH_PRECISION
michael@0	528	if (pass == 0) {
michael@0	529	#endif
michael@0	530	// Stage 3
michael@0	531	{
michael@0	532	step3[0] = _mm_add_epi16(step2[(8 - 1)], step2[0]);
michael@0	533	step3[1] = _mm_add_epi16(step2[(8 - 2)], step2[1]);
michael@0	534	step3[2] = _mm_add_epi16(step2[(8 - 3)], step2[2]);
michael@0	535	step3[3] = _mm_add_epi16(step2[(8 - 4)], step2[3]);
michael@0	536	step3[4] = _mm_sub_epi16(step2[(8 - 5)], step2[4]);
michael@0	537	step3[5] = _mm_sub_epi16(step2[(8 - 6)], step2[5]);
michael@0	538	step3[6] = _mm_sub_epi16(step2[(8 - 7)], step2[6]);
michael@0	539	step3[7] = _mm_sub_epi16(step2[(8 - 8)], step2[7]);
michael@0	540	}
michael@0	541	{
michael@0	542	const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
michael@0	543	const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
michael@0	544	const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
michael@0	545	const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
michael@0	546	const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
michael@0	547	const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
michael@0	548	const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
michael@0	549	const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
michael@0	550	const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
michael@0	551	const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
michael@0	552	const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
michael@0	553	const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
michael@0	554	// dct_const_round_shift
michael@0	555	const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
michael@0	556	const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
michael@0	557	const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
michael@0	558	const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
michael@0	559	const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
michael@0	560	const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
michael@0	561	const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
michael@0	562	const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
michael@0	563	const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
michael@0	564	const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
michael@0	565	const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
michael@0	566	const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
michael@0	567	const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
michael@0	568	const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
michael@0	569	const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
michael@0	570	const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
michael@0	571	// Combine
michael@0	572	step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
michael@0	573	step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
michael@0	574	step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
michael@0	575	step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
michael@0	576	}
michael@0	577	{
michael@0	578	step3[16] = _mm_add_epi16(step2[23], step1[16]);
michael@0	579	step3[17] = _mm_add_epi16(step2[22], step1[17]);
michael@0	580	step3[18] = _mm_add_epi16(step2[21], step1[18]);
michael@0	581	step3[19] = _mm_add_epi16(step2[20], step1[19]);
michael@0	582	step3[20] = _mm_sub_epi16(step1[19], step2[20]);
michael@0	583	step3[21] = _mm_sub_epi16(step1[18], step2[21]);
michael@0	584	step3[22] = _mm_sub_epi16(step1[17], step2[22]);
michael@0	585	step3[23] = _mm_sub_epi16(step1[16], step2[23]);
michael@0	586	step3[24] = _mm_sub_epi16(step1[31], step2[24]);
michael@0	587	step3[25] = _mm_sub_epi16(step1[30], step2[25]);
michael@0	588	step3[26] = _mm_sub_epi16(step1[29], step2[26]);
michael@0	589	step3[27] = _mm_sub_epi16(step1[28], step2[27]);
michael@0	590	step3[28] = _mm_add_epi16(step2[27], step1[28]);
michael@0	591	step3[29] = _mm_add_epi16(step2[26], step1[29]);
michael@0	592	step3[30] = _mm_add_epi16(step2[25], step1[30]);
michael@0	593	step3[31] = _mm_add_epi16(step2[24], step1[31]);
michael@0	594	}
michael@0	595
michael@0	596	// Stage 4
michael@0	597	{
michael@0	598	step1[ 0] = _mm_add_epi16(step3[ 3], step3[ 0]);
michael@0	599	step1[ 1] = _mm_add_epi16(step3[ 2], step3[ 1]);
michael@0	600	step1[ 2] = _mm_sub_epi16(step3[ 1], step3[ 2]);
michael@0	601	step1[ 3] = _mm_sub_epi16(step3[ 0], step3[ 3]);
michael@0	602	step1[ 8] = _mm_add_epi16(step3[11], step2[ 8]);
michael@0	603	step1[ 9] = _mm_add_epi16(step3[10], step2[ 9]);
michael@0	604	step1[10] = _mm_sub_epi16(step2[ 9], step3[10]);
michael@0	605	step1[11] = _mm_sub_epi16(step2[ 8], step3[11]);
michael@0	606	step1[12] = _mm_sub_epi16(step2[15], step3[12]);
michael@0	607	step1[13] = _mm_sub_epi16(step2[14], step3[13]);
michael@0	608	step1[14] = _mm_add_epi16(step3[13], step2[14]);
michael@0	609	step1[15] = _mm_add_epi16(step3[12], step2[15]);
michael@0	610	}
michael@0	611	{
michael@0	612	const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
michael@0	613	const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
michael@0	614	const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
michael@0	615	const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
michael@0	616	const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
michael@0	617	const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
michael@0	618	// dct_const_round_shift
michael@0	619	const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
michael@0	620	const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
michael@0	621	const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
michael@0	622	const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
michael@0	623	const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
michael@0	624	const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
michael@0	625	const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
michael@0	626	const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
michael@0	627	// Combine
michael@0	628	step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
michael@0	629	step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
michael@0	630	}
michael@0	631	{
michael@0	632	const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
michael@0	633	const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
michael@0	634	const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
michael@0	635	const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
michael@0	636	const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
michael@0	637	const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
michael@0	638	const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
michael@0	639	const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
michael@0	640	const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
michael@0	641	const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
michael@0	642	const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
michael@0	643	const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
michael@0	644	const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
michael@0	645	const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
michael@0	646	const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
michael@0	647	const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
michael@0	648	const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
michael@0	649	const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
michael@0	650	const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
michael@0	651	const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
michael@0	652	const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
michael@0	653	const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
michael@0	654	const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
michael@0	655	const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
michael@0	656	// dct_const_round_shift
michael@0	657	const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
michael@0	658	const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
michael@0	659	const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
michael@0	660	const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
michael@0	661	const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
michael@0	662	const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
michael@0	663	const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
michael@0	664	const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
michael@0	665	const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
michael@0	666	const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
michael@0	667	const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
michael@0	668	const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
michael@0	669	const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
michael@0	670	const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
michael@0	671	const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
michael@0	672	const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
michael@0	673	const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
michael@0	674	const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
michael@0	675	const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
michael@0	676	const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
michael@0	677	const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
michael@0	678	const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
michael@0	679	const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
michael@0	680	const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
michael@0	681	const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
michael@0	682	const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
michael@0	683	const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
michael@0	684	const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
michael@0	685	const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
michael@0	686	const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
michael@0	687	const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
michael@0	688	const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
michael@0	689	// Combine
michael@0	690	step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
michael@0	691	step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
michael@0	692	step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
michael@0	693	step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
michael@0	694	step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
michael@0	695	step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
michael@0	696	step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
michael@0	697	step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
michael@0	698	}
michael@0	699	// Stage 5
michael@0	700	{
michael@0	701	step2[4] = _mm_add_epi16(step1[5], step3[4]);
michael@0	702	step2[5] = _mm_sub_epi16(step3[4], step1[5]);
michael@0	703	step2[6] = _mm_sub_epi16(step3[7], step1[6]);
michael@0	704	step2[7] = _mm_add_epi16(step1[6], step3[7]);
michael@0	705	}
michael@0	706	{
michael@0	707	const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
michael@0	708	const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
michael@0	709	const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
michael@0	710	const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
michael@0	711	const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
michael@0	712	const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
michael@0	713	const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
michael@0	714	const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
michael@0	715	const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
michael@0	716	const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
michael@0	717	const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
michael@0	718	const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
michael@0	719	// dct_const_round_shift
michael@0	720	const __m128i out_00_4 = _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
michael@0	721	const __m128i out_00_5 = _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
michael@0	722	const __m128i out_16_4 = _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
michael@0	723	const __m128i out_16_5 = _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
michael@0	724	const __m128i out_08_4 = _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
michael@0	725	const __m128i out_08_5 = _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
michael@0	726	const __m128i out_24_4 = _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
michael@0	727	const __m128i out_24_5 = _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
michael@0	728	const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
michael@0	729	const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
michael@0	730	const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
michael@0	731	const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
michael@0	732	const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
michael@0	733	const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
michael@0	734	const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
michael@0	735	const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
michael@0	736	// Combine
michael@0	737	out[ 0] = _mm_packs_epi32(out_00_6, out_00_7);
michael@0	738	out[16] = _mm_packs_epi32(out_16_6, out_16_7);
michael@0	739	out[ 8] = _mm_packs_epi32(out_08_6, out_08_7);
michael@0	740	out[24] = _mm_packs_epi32(out_24_6, out_24_7);
michael@0	741	}
michael@0	742	{
michael@0	743	const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[ 9], step1[14]);
michael@0	744	const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[ 9], step1[14]);
michael@0	745	const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
michael@0	746	const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
michael@0	747	const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
michael@0	748	const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
michael@0	749	const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
michael@0	750	const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
michael@0	751	const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
michael@0	752	const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
michael@0	753	const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
michael@0	754	const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
michael@0	755	// dct_const_round_shift
michael@0	756	const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
michael@0	757	const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
michael@0	758	const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
michael@0	759	const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
michael@0	760	const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
michael@0	761	const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
michael@0	762	const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
michael@0	763	const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
michael@0	764	const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
michael@0	765	const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
michael@0	766	const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
michael@0	767	const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
michael@0	768	const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
michael@0	769	const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
michael@0	770	const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
michael@0	771	const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
michael@0	772	// Combine
michael@0	773	step2[ 9] = _mm_packs_epi32(s2_09_6, s2_09_7);
michael@0	774	step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
michael@0	775	step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
michael@0	776	step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
michael@0	777	}
michael@0	778	{
michael@0	779	step2[16] = _mm_add_epi16(step1[19], step3[16]);
michael@0	780	step2[17] = _mm_add_epi16(step1[18], step3[17]);
michael@0	781	step2[18] = _mm_sub_epi16(step3[17], step1[18]);
michael@0	782	step2[19] = _mm_sub_epi16(step3[16], step1[19]);
michael@0	783	step2[20] = _mm_sub_epi16(step3[23], step1[20]);
michael@0	784	step2[21] = _mm_sub_epi16(step3[22], step1[21]);
michael@0	785	step2[22] = _mm_add_epi16(step1[21], step3[22]);
michael@0	786	step2[23] = _mm_add_epi16(step1[20], step3[23]);
michael@0	787	step2[24] = _mm_add_epi16(step1[27], step3[24]);
michael@0	788	step2[25] = _mm_add_epi16(step1[26], step3[25]);
michael@0	789	step2[26] = _mm_sub_epi16(step3[25], step1[26]);
michael@0	790	step2[27] = _mm_sub_epi16(step3[24], step1[27]);
michael@0	791	step2[28] = _mm_sub_epi16(step3[31], step1[28]);
michael@0	792	step2[29] = _mm_sub_epi16(step3[30], step1[29]);
michael@0	793	step2[30] = _mm_add_epi16(step1[29], step3[30]);
michael@0	794	step2[31] = _mm_add_epi16(step1[28], step3[31]);
michael@0	795	}
michael@0	796	// Stage 6
michael@0	797	{
michael@0	798	const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
michael@0	799	const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
michael@0	800	const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
michael@0	801	const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
michael@0	802	const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
michael@0	803	const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
michael@0	804	const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
michael@0	805	const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
michael@0	806	const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
michael@0	807	const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
michael@0	808	const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
michael@0	809	const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
michael@0	810	const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
michael@0	811	const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
michael@0	812	const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
michael@0	813	const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
michael@0	814	// dct_const_round_shift
michael@0	815	const __m128i out_04_4 = _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
michael@0	816	const __m128i out_04_5 = _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
michael@0	817	const __m128i out_20_4 = _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
michael@0	818	const __m128i out_20_5 = _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
michael@0	819	const __m128i out_12_4 = _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
michael@0	820	const __m128i out_12_5 = _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
michael@0	821	const __m128i out_28_4 = _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
michael@0	822	const __m128i out_28_5 = _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
michael@0	823	const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
michael@0	824	const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
michael@0	825	const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
michael@0	826	const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
michael@0	827	const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
michael@0	828	const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
michael@0	829	const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
michael@0	830	const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
michael@0	831	// Combine
michael@0	832	out[ 4] = _mm_packs_epi32(out_04_6, out_04_7);
michael@0	833	out[20] = _mm_packs_epi32(out_20_6, out_20_7);
michael@0	834	out[12] = _mm_packs_epi32(out_12_6, out_12_7);
michael@0	835	out[28] = _mm_packs_epi32(out_28_6, out_28_7);
michael@0	836	}
michael@0	837	{
michael@0	838	step3[ 8] = _mm_add_epi16(step2[ 9], step1[ 8]);
michael@0	839	step3[ 9] = _mm_sub_epi16(step1[ 8], step2[ 9]);
michael@0	840	step3[10] = _mm_sub_epi16(step1[11], step2[10]);
michael@0	841	step3[11] = _mm_add_epi16(step2[10], step1[11]);
michael@0	842	step3[12] = _mm_add_epi16(step2[13], step1[12]);
michael@0	843	step3[13] = _mm_sub_epi16(step1[12], step2[13]);
michael@0	844	step3[14] = _mm_sub_epi16(step1[15], step2[14]);
michael@0	845	step3[15] = _mm_add_epi16(step2[14], step1[15]);
michael@0	846	}
michael@0	847	{
michael@0	848	const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
michael@0	849	const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
michael@0	850	const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
michael@0	851	const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
michael@0	852	const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
michael@0	853	const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
michael@0	854	const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
michael@0	855	const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
michael@0	856	const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
michael@0	857	const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
michael@0	858	const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
michael@0	859	const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
michael@0	860	const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
michael@0	861	const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
michael@0	862	const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
michael@0	863	const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
michael@0	864	const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
michael@0	865	const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
michael@0	866	const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
michael@0	867	const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
michael@0	868	const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
michael@0	869	const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
michael@0	870	const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
michael@0	871	const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
michael@0	872	// dct_const_round_shift
michael@0	873	const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
michael@0	874	const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
michael@0	875	const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
michael@0	876	const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
michael@0	877	const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
michael@0	878	const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
michael@0	879	const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
michael@0	880	const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
michael@0	881	const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
michael@0	882	const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
michael@0	883	const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
michael@0	884	const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
michael@0	885	const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
michael@0	886	const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
michael@0	887	const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
michael@0	888	const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
michael@0	889	const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
michael@0	890	const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
michael@0	891	const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
michael@0	892	const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
michael@0	893	const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
michael@0	894	const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
michael@0	895	const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
michael@0	896	const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
michael@0	897	const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
michael@0	898	const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
michael@0	899	const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
michael@0	900	const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
michael@0	901	const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
michael@0	902	const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
michael@0	903	const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
michael@0	904	const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
michael@0	905	// Combine
michael@0	906	step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
michael@0	907	step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
michael@0	908	step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
michael@0	909	step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
michael@0	910	// Combine
michael@0	911	step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
michael@0	912	step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
michael@0	913	step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
michael@0	914	step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
michael@0	915	}
michael@0	916	// Stage 7
michael@0	917	{
michael@0	918	const __m128i out_02_0 = _mm_unpacklo_epi16(step3[ 8], step3[15]);
michael@0	919	const __m128i out_02_1 = _mm_unpackhi_epi16(step3[ 8], step3[15]);
michael@0	920	const __m128i out_18_0 = _mm_unpacklo_epi16(step3[ 9], step3[14]);
michael@0	921	const __m128i out_18_1 = _mm_unpackhi_epi16(step3[ 9], step3[14]);
michael@0	922	const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
michael@0	923	const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
michael@0	924	const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
michael@0	925	const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
michael@0	926	const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
michael@0	927	const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
michael@0	928	const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
michael@0	929	const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
michael@0	930	const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
michael@0	931	const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
michael@0	932	const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
michael@0	933	const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
michael@0	934	const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
michael@0	935	const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
michael@0	936	const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
michael@0	937	const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
michael@0	938	const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
michael@0	939	const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
michael@0	940	const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
michael@0	941	const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
michael@0	942	// dct_const_round_shift
michael@0	943	const __m128i out_02_4 = _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
michael@0	944	const __m128i out_02_5 = _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
michael@0	945	const __m128i out_18_4 = _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
michael@0	946	const __m128i out_18_5 = _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
michael@0	947	const __m128i out_10_4 = _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
michael@0	948	const __m128i out_10_5 = _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
michael@0	949	const __m128i out_26_4 = _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
michael@0	950	const __m128i out_26_5 = _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
michael@0	951	const __m128i out_06_4 = _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
michael@0	952	const __m128i out_06_5 = _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
michael@0	953	const __m128i out_22_4 = _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
michael@0	954	const __m128i out_22_5 = _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
michael@0	955	const __m128i out_14_4 = _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
michael@0	956	const __m128i out_14_5 = _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
michael@0	957	const __m128i out_30_4 = _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
michael@0	958	const __m128i out_30_5 = _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
michael@0	959	const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
michael@0	960	const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
michael@0	961	const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
michael@0	962	const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
michael@0	963	const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
michael@0	964	const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
michael@0	965	const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
michael@0	966	const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
michael@0	967	const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
michael@0	968	const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
michael@0	969	const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
michael@0	970	const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
michael@0	971	const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
michael@0	972	const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
michael@0	973	const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
michael@0	974	const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
michael@0	975	// Combine
michael@0	976	out[ 2] = _mm_packs_epi32(out_02_6, out_02_7);
michael@0	977	out[18] = _mm_packs_epi32(out_18_6, out_18_7);
michael@0	978	out[10] = _mm_packs_epi32(out_10_6, out_10_7);
michael@0	979	out[26] = _mm_packs_epi32(out_26_6, out_26_7);
michael@0	980	out[ 6] = _mm_packs_epi32(out_06_6, out_06_7);
michael@0	981	out[22] = _mm_packs_epi32(out_22_6, out_22_7);
michael@0	982	out[14] = _mm_packs_epi32(out_14_6, out_14_7);
michael@0	983	out[30] = _mm_packs_epi32(out_30_6, out_30_7);
michael@0	984	}
michael@0	985	{
michael@0	986	step1[16] = _mm_add_epi16(step3[17], step2[16]);
michael@0	987	step1[17] = _mm_sub_epi16(step2[16], step3[17]);
michael@0	988	step1[18] = _mm_sub_epi16(step2[19], step3[18]);
michael@0	989	step1[19] = _mm_add_epi16(step3[18], step2[19]);
michael@0	990	step1[20] = _mm_add_epi16(step3[21], step2[20]);
michael@0	991	step1[21] = _mm_sub_epi16(step2[20], step3[21]);
michael@0	992	step1[22] = _mm_sub_epi16(step2[23], step3[22]);
michael@0	993	step1[23] = _mm_add_epi16(step3[22], step2[23]);
michael@0	994	step1[24] = _mm_add_epi16(step3[25], step2[24]);
michael@0	995	step1[25] = _mm_sub_epi16(step2[24], step3[25]);
michael@0	996	step1[26] = _mm_sub_epi16(step2[27], step3[26]);
michael@0	997	step1[27] = _mm_add_epi16(step3[26], step2[27]);
michael@0	998	step1[28] = _mm_add_epi16(step3[29], step2[28]);
michael@0	999	step1[29] = _mm_sub_epi16(step2[28], step3[29]);
michael@0	1000	step1[30] = _mm_sub_epi16(step2[31], step3[30]);
michael@0	1001	step1[31] = _mm_add_epi16(step3[30], step2[31]);
michael@0	1002	}
michael@0	1003	// Final stage --- outputs indices are bit-reversed.
michael@0	1004	{
michael@0	1005	const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
michael@0	1006	const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
michael@0	1007	const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
michael@0	1008	const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
michael@0	1009	const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
michael@0	1010	const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
michael@0	1011	const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
michael@0	1012	const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
michael@0	1013	const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
michael@0	1014	const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
michael@0	1015	const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
michael@0	1016	const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
michael@0	1017	const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
michael@0	1018	const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
michael@0	1019	const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
michael@0	1020	const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
michael@0	1021	const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
michael@0	1022	const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
michael@0	1023	const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
michael@0	1024	const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
michael@0	1025	const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
michael@0	1026	const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
michael@0	1027	const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
michael@0	1028	const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
michael@0	1029	// dct_const_round_shift
michael@0	1030	const __m128i out_01_4 = _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
michael@0	1031	const __m128i out_01_5 = _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
michael@0	1032	const __m128i out_17_4 = _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
michael@0	1033	const __m128i out_17_5 = _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
michael@0	1034	const __m128i out_09_4 = _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
michael@0	1035	const __m128i out_09_5 = _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
michael@0	1036	const __m128i out_25_4 = _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
michael@0	1037	const __m128i out_25_5 = _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
michael@0	1038	const __m128i out_07_4 = _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
michael@0	1039	const __m128i out_07_5 = _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
michael@0	1040	const __m128i out_23_4 = _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
michael@0	1041	const __m128i out_23_5 = _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
michael@0	1042	const __m128i out_15_4 = _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
michael@0	1043	const __m128i out_15_5 = _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
michael@0	1044	const __m128i out_31_4 = _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
michael@0	1045	const __m128i out_31_5 = _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
michael@0	1046	const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
michael@0	1047	const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
michael@0	1048	const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
michael@0	1049	const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
michael@0	1050	const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
michael@0	1051	const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
michael@0	1052	const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
michael@0	1053	const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
michael@0	1054	const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
michael@0	1055	const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
michael@0	1056	const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
michael@0	1057	const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
michael@0	1058	const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
michael@0	1059	const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
michael@0	1060	const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
michael@0	1061	const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
michael@0	1062	// Combine
michael@0	1063	out[ 1] = _mm_packs_epi32(out_01_6, out_01_7);
michael@0	1064	out[17] = _mm_packs_epi32(out_17_6, out_17_7);
michael@0	1065	out[ 9] = _mm_packs_epi32(out_09_6, out_09_7);
michael@0	1066	out[25] = _mm_packs_epi32(out_25_6, out_25_7);
michael@0	1067	out[ 7] = _mm_packs_epi32(out_07_6, out_07_7);
michael@0	1068	out[23] = _mm_packs_epi32(out_23_6, out_23_7);
michael@0	1069	out[15] = _mm_packs_epi32(out_15_6, out_15_7);
michael@0	1070	out[31] = _mm_packs_epi32(out_31_6, out_31_7);
michael@0	1071	}
michael@0	1072	{
michael@0	1073	const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
michael@0	1074	const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
michael@0	1075	const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
michael@0	1076	const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
michael@0	1077	const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
michael@0	1078	const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
michael@0	1079	const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
michael@0	1080	const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
michael@0	1081	const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
michael@0	1082	const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
michael@0	1083	const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
michael@0	1084	const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
michael@0	1085	const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
michael@0	1086	const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
michael@0	1087	const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
michael@0	1088	const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
michael@0	1089	const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
michael@0	1090	const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
michael@0	1091	const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
michael@0	1092	const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
michael@0	1093	const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
michael@0	1094	const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
michael@0	1095	const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
michael@0	1096	const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
michael@0	1097	// dct_const_round_shift
michael@0	1098	const __m128i out_05_4 = _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
michael@0	1099	const __m128i out_05_5 = _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
michael@0	1100	const __m128i out_21_4 = _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
michael@0	1101	const __m128i out_21_5 = _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
michael@0	1102	const __m128i out_13_4 = _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
michael@0	1103	const __m128i out_13_5 = _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
michael@0	1104	const __m128i out_29_4 = _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
michael@0	1105	const __m128i out_29_5 = _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
michael@0	1106	const __m128i out_03_4 = _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
michael@0	1107	const __m128i out_03_5 = _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
michael@0	1108	const __m128i out_19_4 = _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
michael@0	1109	const __m128i out_19_5 = _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
michael@0	1110	const __m128i out_11_4 = _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
michael@0	1111	const __m128i out_11_5 = _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
michael@0	1112	const __m128i out_27_4 = _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
michael@0	1113	const __m128i out_27_5 = _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
michael@0	1114	const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
michael@0	1115	const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
michael@0	1116	const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
michael@0	1117	const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
michael@0	1118	const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
michael@0	1119	const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
michael@0	1120	const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
michael@0	1121	const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
michael@0	1122	const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
michael@0	1123	const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
michael@0	1124	const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
michael@0	1125	const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
michael@0	1126	const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
michael@0	1127	const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
michael@0	1128	const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
michael@0	1129	const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
michael@0	1130	// Combine
michael@0	1131	out[ 5] = _mm_packs_epi32(out_05_6, out_05_7);
michael@0	1132	out[21] = _mm_packs_epi32(out_21_6, out_21_7);
michael@0	1133	out[13] = _mm_packs_epi32(out_13_6, out_13_7);
michael@0	1134	out[29] = _mm_packs_epi32(out_29_6, out_29_7);
michael@0	1135	out[ 3] = _mm_packs_epi32(out_03_6, out_03_7);
michael@0	1136	out[19] = _mm_packs_epi32(out_19_6, out_19_7);
michael@0	1137	out[11] = _mm_packs_epi32(out_11_6, out_11_7);
michael@0	1138	out[27] = _mm_packs_epi32(out_27_6, out_27_7);
michael@0	1139	}
michael@0	1140	#if FDCT32x32_HIGH_PRECISION
michael@0	1141	} else {
michael@0	1142	__m128i lstep1[64], lstep2[64], lstep3[64];
michael@0	1143	__m128i u[32], v[32], sign[16];
michael@0	1144	const __m128i K32One = _mm_set_epi32(1, 1, 1, 1);
michael@0	1145	// start using 32-bit operations
michael@0	1146	// stage 3
michael@0	1147	{
michael@0	1148	// expanding to 32-bit length priori to addition operations
michael@0	1149	lstep2[ 0] = _mm_unpacklo_epi16(step2[ 0], kZero);
michael@0	1150	lstep2[ 1] = _mm_unpackhi_epi16(step2[ 0], kZero);
michael@0	1151	lstep2[ 2] = _mm_unpacklo_epi16(step2[ 1], kZero);
michael@0	1152	lstep2[ 3] = _mm_unpackhi_epi16(step2[ 1], kZero);
michael@0	1153	lstep2[ 4] = _mm_unpacklo_epi16(step2[ 2], kZero);
michael@0	1154	lstep2[ 5] = _mm_unpackhi_epi16(step2[ 2], kZero);
michael@0	1155	lstep2[ 6] = _mm_unpacklo_epi16(step2[ 3], kZero);
michael@0	1156	lstep2[ 7] = _mm_unpackhi_epi16(step2[ 3], kZero);
michael@0	1157	lstep2[ 8] = _mm_unpacklo_epi16(step2[ 4], kZero);
michael@0	1158	lstep2[ 9] = _mm_unpackhi_epi16(step2[ 4], kZero);
michael@0	1159	lstep2[10] = _mm_unpacklo_epi16(step2[ 5], kZero);
michael@0	1160	lstep2[11] = _mm_unpackhi_epi16(step2[ 5], kZero);
michael@0	1161	lstep2[12] = _mm_unpacklo_epi16(step2[ 6], kZero);
michael@0	1162	lstep2[13] = _mm_unpackhi_epi16(step2[ 6], kZero);
michael@0	1163	lstep2[14] = _mm_unpacklo_epi16(step2[ 7], kZero);
michael@0	1164	lstep2[15] = _mm_unpackhi_epi16(step2[ 7], kZero);
michael@0	1165	lstep2[ 0] = _mm_madd_epi16(lstep2[ 0], kOne);
michael@0	1166	lstep2[ 1] = _mm_madd_epi16(lstep2[ 1], kOne);
michael@0	1167	lstep2[ 2] = _mm_madd_epi16(lstep2[ 2], kOne);
michael@0	1168	lstep2[ 3] = _mm_madd_epi16(lstep2[ 3], kOne);
michael@0	1169	lstep2[ 4] = _mm_madd_epi16(lstep2[ 4], kOne);
michael@0	1170	lstep2[ 5] = _mm_madd_epi16(lstep2[ 5], kOne);
michael@0	1171	lstep2[ 6] = _mm_madd_epi16(lstep2[ 6], kOne);
michael@0	1172	lstep2[ 7] = _mm_madd_epi16(lstep2[ 7], kOne);
michael@0	1173	lstep2[ 8] = _mm_madd_epi16(lstep2[ 8], kOne);
michael@0	1174	lstep2[ 9] = _mm_madd_epi16(lstep2[ 9], kOne);
michael@0	1175	lstep2[10] = _mm_madd_epi16(lstep2[10], kOne);
michael@0	1176	lstep2[11] = _mm_madd_epi16(lstep2[11], kOne);
michael@0	1177	lstep2[12] = _mm_madd_epi16(lstep2[12], kOne);
michael@0	1178	lstep2[13] = _mm_madd_epi16(lstep2[13], kOne);
michael@0	1179	lstep2[14] = _mm_madd_epi16(lstep2[14], kOne);
michael@0	1180	lstep2[15] = _mm_madd_epi16(lstep2[15], kOne);
michael@0	1181
michael@0	1182	lstep3[ 0] = _mm_add_epi32(lstep2[14], lstep2[ 0]);
michael@0	1183	lstep3[ 1] = _mm_add_epi32(lstep2[15], lstep2[ 1]);
michael@0	1184	lstep3[ 2] = _mm_add_epi32(lstep2[12], lstep2[ 2]);
michael@0	1185	lstep3[ 3] = _mm_add_epi32(lstep2[13], lstep2[ 3]);
michael@0	1186	lstep3[ 4] = _mm_add_epi32(lstep2[10], lstep2[ 4]);
michael@0	1187	lstep3[ 5] = _mm_add_epi32(lstep2[11], lstep2[ 5]);
michael@0	1188	lstep3[ 6] = _mm_add_epi32(lstep2[ 8], lstep2[ 6]);
michael@0	1189	lstep3[ 7] = _mm_add_epi32(lstep2[ 9], lstep2[ 7]);
michael@0	1190	lstep3[ 8] = _mm_sub_epi32(lstep2[ 6], lstep2[ 8]);
michael@0	1191	lstep3[ 9] = _mm_sub_epi32(lstep2[ 7], lstep2[ 9]);
michael@0	1192	lstep3[10] = _mm_sub_epi32(lstep2[ 4], lstep2[10]);
michael@0	1193	lstep3[11] = _mm_sub_epi32(lstep2[ 5], lstep2[11]);
michael@0	1194	lstep3[12] = _mm_sub_epi32(lstep2[ 2], lstep2[12]);
michael@0	1195	lstep3[13] = _mm_sub_epi32(lstep2[ 3], lstep2[13]);
michael@0	1196	lstep3[14] = _mm_sub_epi32(lstep2[ 0], lstep2[14]);
michael@0	1197	lstep3[15] = _mm_sub_epi32(lstep2[ 1], lstep2[15]);
michael@0	1198	}
michael@0	1199	{
michael@0	1200	const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
michael@0	1201	const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
michael@0	1202	const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
michael@0	1203	const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
michael@0	1204	const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
michael@0	1205	const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
michael@0	1206	const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
michael@0	1207	const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
michael@0	1208	const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
michael@0	1209	const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
michael@0	1210	const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
michael@0	1211	const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
michael@0	1212	// dct_const_round_shift
michael@0	1213	const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
michael@0	1214	const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
michael@0	1215	const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
michael@0	1216	const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
michael@0	1217	const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
michael@0	1218	const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
michael@0	1219	const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
michael@0	1220	const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
michael@0	1221	lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
michael@0	1222	lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
michael@0	1223	lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
michael@0	1224	lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
michael@0	1225	lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
michael@0	1226	lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
michael@0	1227	lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
michael@0	1228	lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
michael@0	1229	}
michael@0	1230	{
michael@0	1231	lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero);
michael@0	1232	lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero);
michael@0	1233	lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero);
michael@0	1234	lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero);
michael@0	1235	lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero);
michael@0	1236	lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero);
michael@0	1237	lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero);
michael@0	1238	lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero);
michael@0	1239	lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero);
michael@0	1240	lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero);
michael@0	1241	lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero);
michael@0	1242	lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero);
michael@0	1243	lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero);
michael@0	1244	lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero);
michael@0	1245	lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero);
michael@0	1246	lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero);
michael@0	1247	lstep2[40] = _mm_madd_epi16(lstep2[40], kOne);
michael@0	1248	lstep2[41] = _mm_madd_epi16(lstep2[41], kOne);
michael@0	1249	lstep2[42] = _mm_madd_epi16(lstep2[42], kOne);
michael@0	1250	lstep2[43] = _mm_madd_epi16(lstep2[43], kOne);
michael@0	1251	lstep2[44] = _mm_madd_epi16(lstep2[44], kOne);
michael@0	1252	lstep2[45] = _mm_madd_epi16(lstep2[45], kOne);
michael@0	1253	lstep2[46] = _mm_madd_epi16(lstep2[46], kOne);
michael@0	1254	lstep2[47] = _mm_madd_epi16(lstep2[47], kOne);
michael@0	1255	lstep2[48] = _mm_madd_epi16(lstep2[48], kOne);
michael@0	1256	lstep2[49] = _mm_madd_epi16(lstep2[49], kOne);
michael@0	1257	lstep2[50] = _mm_madd_epi16(lstep2[50], kOne);
michael@0	1258	lstep2[51] = _mm_madd_epi16(lstep2[51], kOne);
michael@0	1259	lstep2[52] = _mm_madd_epi16(lstep2[52], kOne);
michael@0	1260	lstep2[53] = _mm_madd_epi16(lstep2[53], kOne);
michael@0	1261	lstep2[54] = _mm_madd_epi16(lstep2[54], kOne);
michael@0	1262	lstep2[55] = _mm_madd_epi16(lstep2[55], kOne);
michael@0	1263
michael@0	1264	lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero);
michael@0	1265	lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero);
michael@0	1266	lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero);
michael@0	1267	lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero);
michael@0	1268	lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero);
michael@0	1269	lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero);
michael@0	1270	lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero);
michael@0	1271	lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero);
michael@0	1272	lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero);
michael@0	1273	lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero);
michael@0	1274	lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero);
michael@0	1275	lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero);
michael@0	1276	lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero);
michael@0	1277	lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero);
michael@0	1278	lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero);
michael@0	1279	lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero);
michael@0	1280	lstep1[32] = _mm_madd_epi16(lstep1[32], kOne);
michael@0	1281	lstep1[33] = _mm_madd_epi16(lstep1[33], kOne);
michael@0	1282	lstep1[34] = _mm_madd_epi16(lstep1[34], kOne);
michael@0	1283	lstep1[35] = _mm_madd_epi16(lstep1[35], kOne);
michael@0	1284	lstep1[36] = _mm_madd_epi16(lstep1[36], kOne);
michael@0	1285	lstep1[37] = _mm_madd_epi16(lstep1[37], kOne);
michael@0	1286	lstep1[38] = _mm_madd_epi16(lstep1[38], kOne);
michael@0	1287	lstep1[39] = _mm_madd_epi16(lstep1[39], kOne);
michael@0	1288	lstep1[56] = _mm_madd_epi16(lstep1[56], kOne);
michael@0	1289	lstep1[57] = _mm_madd_epi16(lstep1[57], kOne);
michael@0	1290	lstep1[58] = _mm_madd_epi16(lstep1[58], kOne);
michael@0	1291	lstep1[59] = _mm_madd_epi16(lstep1[59], kOne);
michael@0	1292	lstep1[60] = _mm_madd_epi16(lstep1[60], kOne);
michael@0	1293	lstep1[61] = _mm_madd_epi16(lstep1[61], kOne);
michael@0	1294	lstep1[62] = _mm_madd_epi16(lstep1[62], kOne);
michael@0	1295	lstep1[63] = _mm_madd_epi16(lstep1[63], kOne);
michael@0	1296
michael@0	1297	lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]);
michael@0	1298	lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]);
michael@0	1299
michael@0	1300	lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]);
michael@0	1301	lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]);
michael@0	1302	lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]);
michael@0	1303	lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]);
michael@0	1304	lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]);
michael@0	1305	lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]);
michael@0	1306	lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]);
michael@0	1307	lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]);
michael@0	1308	lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]);
michael@0	1309	lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]);
michael@0	1310	lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]);
michael@0	1311	lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]);
michael@0	1312	lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]);
michael@0	1313	lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]);
michael@0	1314	lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]);
michael@0	1315	lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]);
michael@0	1316	lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]);
michael@0	1317	lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]);
michael@0	1318	lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]);
michael@0	1319	lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]);
michael@0	1320	lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]);
michael@0	1321	lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]);
michael@0	1322	lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]);
michael@0	1323	lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]);
michael@0	1324	lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]);
michael@0	1325	lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]);
michael@0	1326	lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]);
michael@0	1327	lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]);
michael@0	1328	lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]);
michael@0	1329	lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]);
michael@0	1330	}
michael@0	1331
michael@0	1332	// stage 4
michael@0	1333	{
michael@0	1334	// expanding to 32-bit length priori to addition operations
michael@0	1335	lstep2[16] = _mm_unpacklo_epi16(step2[ 8], kZero);
michael@0	1336	lstep2[17] = _mm_unpackhi_epi16(step2[ 8], kZero);
michael@0	1337	lstep2[18] = _mm_unpacklo_epi16(step2[ 9], kZero);
michael@0	1338	lstep2[19] = _mm_unpackhi_epi16(step2[ 9], kZero);
michael@0	1339	lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero);
michael@0	1340	lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero);
michael@0	1341	lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero);
michael@0	1342	lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero);
michael@0	1343	lstep2[16] = _mm_madd_epi16(lstep2[16], kOne);
michael@0	1344	lstep2[17] = _mm_madd_epi16(lstep2[17], kOne);
michael@0	1345	lstep2[18] = _mm_madd_epi16(lstep2[18], kOne);
michael@0	1346	lstep2[19] = _mm_madd_epi16(lstep2[19], kOne);
michael@0	1347	lstep2[28] = _mm_madd_epi16(lstep2[28], kOne);
michael@0	1348	lstep2[29] = _mm_madd_epi16(lstep2[29], kOne);
michael@0	1349	lstep2[30] = _mm_madd_epi16(lstep2[30], kOne);
michael@0	1350	lstep2[31] = _mm_madd_epi16(lstep2[31], kOne);
michael@0	1351
michael@0	1352	lstep1[ 0] = _mm_add_epi32(lstep3[ 6], lstep3[ 0]);
michael@0	1353	lstep1[ 1] = _mm_add_epi32(lstep3[ 7], lstep3[ 1]);
michael@0	1354	lstep1[ 2] = _mm_add_epi32(lstep3[ 4], lstep3[ 2]);
michael@0	1355	lstep1[ 3] = _mm_add_epi32(lstep3[ 5], lstep3[ 3]);
michael@0	1356	lstep1[ 4] = _mm_sub_epi32(lstep3[ 2], lstep3[ 4]);
michael@0	1357	lstep1[ 5] = _mm_sub_epi32(lstep3[ 3], lstep3[ 5]);
michael@0	1358	lstep1[ 6] = _mm_sub_epi32(lstep3[ 0], lstep3[ 6]);
michael@0	1359	lstep1[ 7] = _mm_sub_epi32(lstep3[ 1], lstep3[ 7]);
michael@0	1360	lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]);
michael@0	1361	lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]);
michael@0	1362	lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]);
michael@0	1363	lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]);
michael@0	1364	lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]);
michael@0	1365	lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]);
michael@0	1366	lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]);
michael@0	1367	lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]);
michael@0	1368	lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]);
michael@0	1369	lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]);
michael@0	1370	lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]);
michael@0	1371	lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]);
michael@0	1372	lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]);
michael@0	1373	lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]);
michael@0	1374	lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]);
michael@0	1375	lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]);
michael@0	1376	}
michael@0	1377	{
michael@0	1378	// to be continued...
michael@0	1379	//
michael@0	1380	const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
michael@0	1381	const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
michael@0	1382
michael@0	1383	u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]);
michael@0	1384	u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]);
michael@0	1385	u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]);
michael@0	1386	u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]);
michael@0	1387
michael@0	1388	// TODO(jingning): manually inline k_madd_epi32_ to further hide
michael@0	1389	// instruction latency.
michael@0	1390	v[ 0] = k_madd_epi32(u[0], k32_p16_m16);
michael@0	1391	v[ 1] = k_madd_epi32(u[1], k32_p16_m16);
michael@0	1392	v[ 2] = k_madd_epi32(u[2], k32_p16_m16);
michael@0	1393	v[ 3] = k_madd_epi32(u[3], k32_p16_m16);
michael@0	1394	v[ 4] = k_madd_epi32(u[0], k32_p16_p16);
michael@0	1395	v[ 5] = k_madd_epi32(u[1], k32_p16_p16);
michael@0	1396	v[ 6] = k_madd_epi32(u[2], k32_p16_p16);
michael@0	1397	v[ 7] = k_madd_epi32(u[3], k32_p16_p16);
michael@0	1398
michael@0	1399	u[0] = k_packs_epi64(v[0], v[1]);
michael@0	1400	u[1] = k_packs_epi64(v[2], v[3]);
michael@0	1401	u[2] = k_packs_epi64(v[4], v[5]);
michael@0	1402	u[3] = k_packs_epi64(v[6], v[7]);
michael@0	1403
michael@0	1404	v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
michael@0	1405	v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
michael@0	1406	v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
michael@0	1407	v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
michael@0	1408
michael@0	1409	lstep1[10] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
michael@0	1410	lstep1[11] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
michael@0	1411	lstep1[12] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
michael@0	1412	lstep1[13] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
michael@0	1413	}
michael@0	1414	{
michael@0	1415	const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
michael@0	1416	const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
michael@0	1417	const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
michael@0	1418
michael@0	1419	u[ 0] = _mm_unpacklo_epi32(lstep3[36], lstep3[58]);
michael@0	1420	u[ 1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]);
michael@0	1421	u[ 2] = _mm_unpacklo_epi32(lstep3[37], lstep3[59]);
michael@0	1422	u[ 3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]);
michael@0	1423	u[ 4] = _mm_unpacklo_epi32(lstep3[38], lstep3[56]);
michael@0	1424	u[ 5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]);
michael@0	1425	u[ 6] = _mm_unpacklo_epi32(lstep3[39], lstep3[57]);
michael@0	1426	u[ 7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]);
michael@0	1427	u[ 8] = _mm_unpacklo_epi32(lstep3[40], lstep3[54]);
michael@0	1428	u[ 9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]);
michael@0	1429	u[10] = _mm_unpacklo_epi32(lstep3[41], lstep3[55]);
michael@0	1430	u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]);
michael@0	1431	u[12] = _mm_unpacklo_epi32(lstep3[42], lstep3[52]);
michael@0	1432	u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]);
michael@0	1433	u[14] = _mm_unpacklo_epi32(lstep3[43], lstep3[53]);
michael@0	1434	u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]);
michael@0	1435
michael@0	1436	v[ 0] = k_madd_epi32(u[ 0], k32_m08_p24);
michael@0	1437	v[ 1] = k_madd_epi32(u[ 1], k32_m08_p24);
michael@0	1438	v[ 2] = k_madd_epi32(u[ 2], k32_m08_p24);
michael@0	1439	v[ 3] = k_madd_epi32(u[ 3], k32_m08_p24);
michael@0	1440	v[ 4] = k_madd_epi32(u[ 4], k32_m08_p24);
michael@0	1441	v[ 5] = k_madd_epi32(u[ 5], k32_m08_p24);
michael@0	1442	v[ 6] = k_madd_epi32(u[ 6], k32_m08_p24);
michael@0	1443	v[ 7] = k_madd_epi32(u[ 7], k32_m08_p24);
michael@0	1444	v[ 8] = k_madd_epi32(u[ 8], k32_m24_m08);
michael@0	1445	v[ 9] = k_madd_epi32(u[ 9], k32_m24_m08);
michael@0	1446	v[10] = k_madd_epi32(u[10], k32_m24_m08);
michael@0	1447	v[11] = k_madd_epi32(u[11], k32_m24_m08);
michael@0	1448	v[12] = k_madd_epi32(u[12], k32_m24_m08);
michael@0	1449	v[13] = k_madd_epi32(u[13], k32_m24_m08);
michael@0	1450	v[14] = k_madd_epi32(u[14], k32_m24_m08);
michael@0	1451	v[15] = k_madd_epi32(u[15], k32_m24_m08);
michael@0	1452	v[16] = k_madd_epi32(u[12], k32_m08_p24);
michael@0	1453	v[17] = k_madd_epi32(u[13], k32_m08_p24);
michael@0	1454	v[18] = k_madd_epi32(u[14], k32_m08_p24);
michael@0	1455	v[19] = k_madd_epi32(u[15], k32_m08_p24);
michael@0	1456	v[20] = k_madd_epi32(u[ 8], k32_m08_p24);
michael@0	1457	v[21] = k_madd_epi32(u[ 9], k32_m08_p24);
michael@0	1458	v[22] = k_madd_epi32(u[10], k32_m08_p24);
michael@0	1459	v[23] = k_madd_epi32(u[11], k32_m08_p24);
michael@0	1460	v[24] = k_madd_epi32(u[ 4], k32_p24_p08);
michael@0	1461	v[25] = k_madd_epi32(u[ 5], k32_p24_p08);
michael@0	1462	v[26] = k_madd_epi32(u[ 6], k32_p24_p08);
michael@0	1463	v[27] = k_madd_epi32(u[ 7], k32_p24_p08);
michael@0	1464	v[28] = k_madd_epi32(u[ 0], k32_p24_p08);
michael@0	1465	v[29] = k_madd_epi32(u[ 1], k32_p24_p08);
michael@0	1466	v[30] = k_madd_epi32(u[ 2], k32_p24_p08);
michael@0	1467	v[31] = k_madd_epi32(u[ 3], k32_p24_p08);
michael@0	1468
michael@0	1469	u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
michael@0	1470	u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
michael@0	1471	u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
michael@0	1472	u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
michael@0	1473	u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
michael@0	1474	u[ 5] = k_packs_epi64(v[10], v[11]);
michael@0	1475	u[ 6] = k_packs_epi64(v[12], v[13]);
michael@0	1476	u[ 7] = k_packs_epi64(v[14], v[15]);
michael@0	1477	u[ 8] = k_packs_epi64(v[16], v[17]);
michael@0	1478	u[ 9] = k_packs_epi64(v[18], v[19]);
michael@0	1479	u[10] = k_packs_epi64(v[20], v[21]);
michael@0	1480	u[11] = k_packs_epi64(v[22], v[23]);
michael@0	1481	u[12] = k_packs_epi64(v[24], v[25]);
michael@0	1482	u[13] = k_packs_epi64(v[26], v[27]);
michael@0	1483	u[14] = k_packs_epi64(v[28], v[29]);
michael@0	1484	u[15] = k_packs_epi64(v[30], v[31]);
michael@0	1485
michael@0	1486	v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
michael@0	1487	v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
michael@0	1488	v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
michael@0	1489	v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
michael@0	1490	v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
michael@0	1491	v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
michael@0	1492	v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
michael@0	1493	v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
michael@0	1494	v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
michael@0	1495	v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
michael@0	1496	v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
michael@0	1497	v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
michael@0	1498	v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
michael@0	1499	v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
michael@0	1500	v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
michael@0	1501	v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
michael@0	1502
michael@0	1503	lstep1[36] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
michael@0	1504	lstep1[37] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
michael@0	1505	lstep1[38] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
michael@0	1506	lstep1[39] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
michael@0	1507	lstep1[40] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
michael@0	1508	lstep1[41] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
michael@0	1509	lstep1[42] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
michael@0	1510	lstep1[43] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
michael@0	1511	lstep1[52] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
michael@0	1512	lstep1[53] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
michael@0	1513	lstep1[54] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
michael@0	1514	lstep1[55] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
michael@0	1515	lstep1[56] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
michael@0	1516	lstep1[57] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
michael@0	1517	lstep1[58] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
michael@0	1518	lstep1[59] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
michael@0	1519	}
michael@0	1520	// stage 5
michael@0	1521	{
michael@0	1522	lstep2[ 8] = _mm_add_epi32(lstep1[10], lstep3[ 8]);
michael@0	1523	lstep2[ 9] = _mm_add_epi32(lstep1[11], lstep3[ 9]);
michael@0	1524	lstep2[10] = _mm_sub_epi32(lstep3[ 8], lstep1[10]);
michael@0	1525	lstep2[11] = _mm_sub_epi32(lstep3[ 9], lstep1[11]);
michael@0	1526	lstep2[12] = _mm_sub_epi32(lstep3[14], lstep1[12]);
michael@0	1527	lstep2[13] = _mm_sub_epi32(lstep3[15], lstep1[13]);
michael@0	1528	lstep2[14] = _mm_add_epi32(lstep1[12], lstep3[14]);
michael@0	1529	lstep2[15] = _mm_add_epi32(lstep1[13], lstep3[15]);
michael@0	1530	}
michael@0	1531	{
michael@0	1532	const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
michael@0	1533	const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
michael@0	1534	const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
michael@0	1535	const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
michael@0	1536
michael@0	1537	u[0] = _mm_unpacklo_epi32(lstep1[0], lstep1[2]);
michael@0	1538	u[1] = _mm_unpackhi_epi32(lstep1[0], lstep1[2]);
michael@0	1539	u[2] = _mm_unpacklo_epi32(lstep1[1], lstep1[3]);
michael@0	1540	u[3] = _mm_unpackhi_epi32(lstep1[1], lstep1[3]);
michael@0	1541	u[4] = _mm_unpacklo_epi32(lstep1[4], lstep1[6]);
michael@0	1542	u[5] = _mm_unpackhi_epi32(lstep1[4], lstep1[6]);
michael@0	1543	u[6] = _mm_unpacklo_epi32(lstep1[5], lstep1[7]);
michael@0	1544	u[7] = _mm_unpackhi_epi32(lstep1[5], lstep1[7]);
michael@0	1545
michael@0	1546	// TODO(jingning): manually inline k_madd_epi32_ to further hide
michael@0	1547	// instruction latency.
michael@0	1548	v[ 0] = k_madd_epi32(u[0], k32_p16_p16);
michael@0	1549	v[ 1] = k_madd_epi32(u[1], k32_p16_p16);
michael@0	1550	v[ 2] = k_madd_epi32(u[2], k32_p16_p16);
michael@0	1551	v[ 3] = k_madd_epi32(u[3], k32_p16_p16);
michael@0	1552	v[ 4] = k_madd_epi32(u[0], k32_p16_m16);
michael@0	1553	v[ 5] = k_madd_epi32(u[1], k32_p16_m16);
michael@0	1554	v[ 6] = k_madd_epi32(u[2], k32_p16_m16);
michael@0	1555	v[ 7] = k_madd_epi32(u[3], k32_p16_m16);
michael@0	1556	v[ 8] = k_madd_epi32(u[4], k32_p24_p08);
michael@0	1557	v[ 9] = k_madd_epi32(u[5], k32_p24_p08);
michael@0	1558	v[10] = k_madd_epi32(u[6], k32_p24_p08);
michael@0	1559	v[11] = k_madd_epi32(u[7], k32_p24_p08);
michael@0	1560	v[12] = k_madd_epi32(u[4], k32_m08_p24);
michael@0	1561	v[13] = k_madd_epi32(u[5], k32_m08_p24);
michael@0	1562	v[14] = k_madd_epi32(u[6], k32_m08_p24);
michael@0	1563	v[15] = k_madd_epi32(u[7], k32_m08_p24);
michael@0	1564
michael@0	1565	u[0] = k_packs_epi64(v[0], v[1]);
michael@0	1566	u[1] = k_packs_epi64(v[2], v[3]);
michael@0	1567	u[2] = k_packs_epi64(v[4], v[5]);
michael@0	1568	u[3] = k_packs_epi64(v[6], v[7]);
michael@0	1569	u[4] = k_packs_epi64(v[8], v[9]);
michael@0	1570	u[5] = k_packs_epi64(v[10], v[11]);
michael@0	1571	u[6] = k_packs_epi64(v[12], v[13]);
michael@0	1572	u[7] = k_packs_epi64(v[14], v[15]);
michael@0	1573
michael@0	1574	v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
michael@0	1575	v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
michael@0	1576	v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
michael@0	1577	v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
michael@0	1578	v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
michael@0	1579	v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
michael@0	1580	v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
michael@0	1581	v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
michael@0	1582
michael@0	1583	u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
michael@0	1584	u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
michael@0	1585	u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
michael@0	1586	u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
michael@0	1587	u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
michael@0	1588	u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
michael@0	1589	u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
michael@0	1590	u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
michael@0	1591
michael@0	1592	sign[0] = _mm_cmplt_epi32(u[0], kZero);
michael@0	1593	sign[1] = _mm_cmplt_epi32(u[1], kZero);
michael@0	1594	sign[2] = _mm_cmplt_epi32(u[2], kZero);
michael@0	1595	sign[3] = _mm_cmplt_epi32(u[3], kZero);
michael@0	1596	sign[4] = _mm_cmplt_epi32(u[4], kZero);
michael@0	1597	sign[5] = _mm_cmplt_epi32(u[5], kZero);
michael@0	1598	sign[6] = _mm_cmplt_epi32(u[6], kZero);
michael@0	1599	sign[7] = _mm_cmplt_epi32(u[7], kZero);
michael@0	1600
michael@0	1601	u[0] = _mm_sub_epi32(u[0], sign[0]);
michael@0	1602	u[1] = _mm_sub_epi32(u[1], sign[1]);
michael@0	1603	u[2] = _mm_sub_epi32(u[2], sign[2]);
michael@0	1604	u[3] = _mm_sub_epi32(u[3], sign[3]);
michael@0	1605	u[4] = _mm_sub_epi32(u[4], sign[4]);
michael@0	1606	u[5] = _mm_sub_epi32(u[5], sign[5]);
michael@0	1607	u[6] = _mm_sub_epi32(u[6], sign[6]);
michael@0	1608	u[7] = _mm_sub_epi32(u[7], sign[7]);
michael@0	1609
michael@0	1610	u[0] = _mm_add_epi32(u[0], K32One);
michael@0	1611	u[1] = _mm_add_epi32(u[1], K32One);
michael@0	1612	u[2] = _mm_add_epi32(u[2], K32One);
michael@0	1613	u[3] = _mm_add_epi32(u[3], K32One);
michael@0	1614	u[4] = _mm_add_epi32(u[4], K32One);
michael@0	1615	u[5] = _mm_add_epi32(u[5], K32One);
michael@0	1616	u[6] = _mm_add_epi32(u[6], K32One);
michael@0	1617	u[7] = _mm_add_epi32(u[7], K32One);
michael@0	1618
michael@0	1619	u[0] = _mm_srai_epi32(u[0], 2);
michael@0	1620	u[1] = _mm_srai_epi32(u[1], 2);
michael@0	1621	u[2] = _mm_srai_epi32(u[2], 2);
michael@0	1622	u[3] = _mm_srai_epi32(u[3], 2);
michael@0	1623	u[4] = _mm_srai_epi32(u[4], 2);
michael@0	1624	u[5] = _mm_srai_epi32(u[5], 2);
michael@0	1625	u[6] = _mm_srai_epi32(u[6], 2);
michael@0	1626	u[7] = _mm_srai_epi32(u[7], 2);
michael@0	1627
michael@0	1628	// Combine
michael@0	1629	out[ 0] = _mm_packs_epi32(u[0], u[1]);
michael@0	1630	out[16] = _mm_packs_epi32(u[2], u[3]);
michael@0	1631	out[ 8] = _mm_packs_epi32(u[4], u[5]);
michael@0	1632	out[24] = _mm_packs_epi32(u[6], u[7]);
michael@0	1633	}
michael@0	1634	{
michael@0	1635	const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
michael@0	1636	const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
michael@0	1637	const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
michael@0	1638
michael@0	1639	u[0] = _mm_unpacklo_epi32(lstep1[18], lstep1[28]);
michael@0	1640	u[1] = _mm_unpackhi_epi32(lstep1[18], lstep1[28]);
michael@0	1641	u[2] = _mm_unpacklo_epi32(lstep1[19], lstep1[29]);
michael@0	1642	u[3] = _mm_unpackhi_epi32(lstep1[19], lstep1[29]);
michael@0	1643	u[4] = _mm_unpacklo_epi32(lstep1[20], lstep1[26]);
michael@0	1644	u[5] = _mm_unpackhi_epi32(lstep1[20], lstep1[26]);
michael@0	1645	u[6] = _mm_unpacklo_epi32(lstep1[21], lstep1[27]);
michael@0	1646	u[7] = _mm_unpackhi_epi32(lstep1[21], lstep1[27]);
michael@0	1647
michael@0	1648	v[0] = k_madd_epi32(u[0], k32_m08_p24);
michael@0	1649	v[1] = k_madd_epi32(u[1], k32_m08_p24);
michael@0	1650	v[2] = k_madd_epi32(u[2], k32_m08_p24);
michael@0	1651	v[3] = k_madd_epi32(u[3], k32_m08_p24);
michael@0	1652	v[4] = k_madd_epi32(u[4], k32_m24_m08);
michael@0	1653	v[5] = k_madd_epi32(u[5], k32_m24_m08);
michael@0	1654	v[6] = k_madd_epi32(u[6], k32_m24_m08);
michael@0	1655	v[7] = k_madd_epi32(u[7], k32_m24_m08);
michael@0	1656	v[ 8] = k_madd_epi32(u[4], k32_m08_p24);
michael@0	1657	v[ 9] = k_madd_epi32(u[5], k32_m08_p24);
michael@0	1658	v[10] = k_madd_epi32(u[6], k32_m08_p24);
michael@0	1659	v[11] = k_madd_epi32(u[7], k32_m08_p24);
michael@0	1660	v[12] = k_madd_epi32(u[0], k32_p24_p08);
michael@0	1661	v[13] = k_madd_epi32(u[1], k32_p24_p08);
michael@0	1662	v[14] = k_madd_epi32(u[2], k32_p24_p08);
michael@0	1663	v[15] = k_madd_epi32(u[3], k32_p24_p08);
michael@0	1664
michael@0	1665	u[0] = k_packs_epi64(v[0], v[1]);
michael@0	1666	u[1] = k_packs_epi64(v[2], v[3]);
michael@0	1667	u[2] = k_packs_epi64(v[4], v[5]);
michael@0	1668	u[3] = k_packs_epi64(v[6], v[7]);
michael@0	1669	u[4] = k_packs_epi64(v[8], v[9]);
michael@0	1670	u[5] = k_packs_epi64(v[10], v[11]);
michael@0	1671	u[6] = k_packs_epi64(v[12], v[13]);
michael@0	1672	u[7] = k_packs_epi64(v[14], v[15]);
michael@0	1673
michael@0	1674	u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
michael@0	1675	u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
michael@0	1676	u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
michael@0	1677	u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
michael@0	1678	u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
michael@0	1679	u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
michael@0	1680	u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
michael@0	1681	u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
michael@0	1682
michael@0	1683	lstep2[18] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
michael@0	1684	lstep2[19] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
michael@0	1685	lstep2[20] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
michael@0	1686	lstep2[21] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
michael@0	1687	lstep2[26] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
michael@0	1688	lstep2[27] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
michael@0	1689	lstep2[28] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
michael@0	1690	lstep2[29] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
michael@0	1691	}
michael@0	1692	{
michael@0	1693	lstep2[32] = _mm_add_epi32(lstep1[38], lstep3[32]);
michael@0	1694	lstep2[33] = _mm_add_epi32(lstep1[39], lstep3[33]);
michael@0	1695	lstep2[34] = _mm_add_epi32(lstep1[36], lstep3[34]);
michael@0	1696	lstep2[35] = _mm_add_epi32(lstep1[37], lstep3[35]);
michael@0	1697	lstep2[36] = _mm_sub_epi32(lstep3[34], lstep1[36]);
michael@0	1698	lstep2[37] = _mm_sub_epi32(lstep3[35], lstep1[37]);
michael@0	1699	lstep2[38] = _mm_sub_epi32(lstep3[32], lstep1[38]);
michael@0	1700	lstep2[39] = _mm_sub_epi32(lstep3[33], lstep1[39]);
michael@0	1701	lstep2[40] = _mm_sub_epi32(lstep3[46], lstep1[40]);
michael@0	1702	lstep2[41] = _mm_sub_epi32(lstep3[47], lstep1[41]);
michael@0	1703	lstep2[42] = _mm_sub_epi32(lstep3[44], lstep1[42]);
michael@0	1704	lstep2[43] = _mm_sub_epi32(lstep3[45], lstep1[43]);
michael@0	1705	lstep2[44] = _mm_add_epi32(lstep1[42], lstep3[44]);
michael@0	1706	lstep2[45] = _mm_add_epi32(lstep1[43], lstep3[45]);
michael@0	1707	lstep2[46] = _mm_add_epi32(lstep1[40], lstep3[46]);
michael@0	1708	lstep2[47] = _mm_add_epi32(lstep1[41], lstep3[47]);
michael@0	1709	lstep2[48] = _mm_add_epi32(lstep1[54], lstep3[48]);
michael@0	1710	lstep2[49] = _mm_add_epi32(lstep1[55], lstep3[49]);
michael@0	1711	lstep2[50] = _mm_add_epi32(lstep1[52], lstep3[50]);
michael@0	1712	lstep2[51] = _mm_add_epi32(lstep1[53], lstep3[51]);
michael@0	1713	lstep2[52] = _mm_sub_epi32(lstep3[50], lstep1[52]);
michael@0	1714	lstep2[53] = _mm_sub_epi32(lstep3[51], lstep1[53]);
michael@0	1715	lstep2[54] = _mm_sub_epi32(lstep3[48], lstep1[54]);
michael@0	1716	lstep2[55] = _mm_sub_epi32(lstep3[49], lstep1[55]);
michael@0	1717	lstep2[56] = _mm_sub_epi32(lstep3[62], lstep1[56]);
michael@0	1718	lstep2[57] = _mm_sub_epi32(lstep3[63], lstep1[57]);
michael@0	1719	lstep2[58] = _mm_sub_epi32(lstep3[60], lstep1[58]);
michael@0	1720	lstep2[59] = _mm_sub_epi32(lstep3[61], lstep1[59]);
michael@0	1721	lstep2[60] = _mm_add_epi32(lstep1[58], lstep3[60]);
michael@0	1722	lstep2[61] = _mm_add_epi32(lstep1[59], lstep3[61]);
michael@0	1723	lstep2[62] = _mm_add_epi32(lstep1[56], lstep3[62]);
michael@0	1724	lstep2[63] = _mm_add_epi32(lstep1[57], lstep3[63]);
michael@0	1725	}
michael@0	1726	// stage 6
michael@0	1727	{
michael@0	1728	const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
michael@0	1729	const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
michael@0	1730	const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
michael@0	1731	const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
michael@0	1732
michael@0	1733	u[0] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
michael@0	1734	u[1] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
michael@0	1735	u[2] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
michael@0	1736	u[3] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
michael@0	1737	u[4] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
michael@0	1738	u[5] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
michael@0	1739	u[6] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
michael@0	1740	u[7] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
michael@0	1741	u[8] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
michael@0	1742	u[9] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
michael@0	1743	u[10] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
michael@0	1744	u[11] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
michael@0	1745	u[12] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
michael@0	1746	u[13] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
michael@0	1747	u[14] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
michael@0	1748	u[15] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
michael@0	1749
michael@0	1750	v[0] = k_madd_epi32(u[0], k32_p28_p04);
michael@0	1751	v[1] = k_madd_epi32(u[1], k32_p28_p04);
michael@0	1752	v[2] = k_madd_epi32(u[2], k32_p28_p04);
michael@0	1753	v[3] = k_madd_epi32(u[3], k32_p28_p04);
michael@0	1754	v[4] = k_madd_epi32(u[4], k32_p12_p20);
michael@0	1755	v[5] = k_madd_epi32(u[5], k32_p12_p20);
michael@0	1756	v[6] = k_madd_epi32(u[6], k32_p12_p20);
michael@0	1757	v[7] = k_madd_epi32(u[7], k32_p12_p20);
michael@0	1758	v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
michael@0	1759	v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
michael@0	1760	v[10] = k_madd_epi32(u[10], k32_m20_p12);
michael@0	1761	v[11] = k_madd_epi32(u[11], k32_m20_p12);
michael@0	1762	v[12] = k_madd_epi32(u[12], k32_m04_p28);
michael@0	1763	v[13] = k_madd_epi32(u[13], k32_m04_p28);
michael@0	1764	v[14] = k_madd_epi32(u[14], k32_m04_p28);
michael@0	1765	v[15] = k_madd_epi32(u[15], k32_m04_p28);
michael@0	1766
michael@0	1767	u[0] = k_packs_epi64(v[0], v[1]);
michael@0	1768	u[1] = k_packs_epi64(v[2], v[3]);
michael@0	1769	u[2] = k_packs_epi64(v[4], v[5]);
michael@0	1770	u[3] = k_packs_epi64(v[6], v[7]);
michael@0	1771	u[4] = k_packs_epi64(v[8], v[9]);
michael@0	1772	u[5] = k_packs_epi64(v[10], v[11]);
michael@0	1773	u[6] = k_packs_epi64(v[12], v[13]);
michael@0	1774	u[7] = k_packs_epi64(v[14], v[15]);
michael@0	1775
michael@0	1776	v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
michael@0	1777	v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
michael@0	1778	v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
michael@0	1779	v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
michael@0	1780	v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
michael@0	1781	v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
michael@0	1782	v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
michael@0	1783	v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
michael@0	1784
michael@0	1785	u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
michael@0	1786	u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
michael@0	1787	u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
michael@0	1788	u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
michael@0	1789	u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
michael@0	1790	u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
michael@0	1791	u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
michael@0	1792	u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
michael@0	1793
michael@0	1794	sign[0] = _mm_cmplt_epi32(u[0], kZero);
michael@0	1795	sign[1] = _mm_cmplt_epi32(u[1], kZero);
michael@0	1796	sign[2] = _mm_cmplt_epi32(u[2], kZero);
michael@0	1797	sign[3] = _mm_cmplt_epi32(u[3], kZero);
michael@0	1798	sign[4] = _mm_cmplt_epi32(u[4], kZero);
michael@0	1799	sign[5] = _mm_cmplt_epi32(u[5], kZero);
michael@0	1800	sign[6] = _mm_cmplt_epi32(u[6], kZero);
michael@0	1801	sign[7] = _mm_cmplt_epi32(u[7], kZero);
michael@0	1802
michael@0	1803	u[0] = _mm_sub_epi32(u[0], sign[0]);
michael@0	1804	u[1] = _mm_sub_epi32(u[1], sign[1]);
michael@0	1805	u[2] = _mm_sub_epi32(u[2], sign[2]);
michael@0	1806	u[3] = _mm_sub_epi32(u[3], sign[3]);
michael@0	1807	u[4] = _mm_sub_epi32(u[4], sign[4]);
michael@0	1808	u[5] = _mm_sub_epi32(u[5], sign[5]);
michael@0	1809	u[6] = _mm_sub_epi32(u[6], sign[6]);
michael@0	1810	u[7] = _mm_sub_epi32(u[7], sign[7]);
michael@0	1811
michael@0	1812	u[0] = _mm_add_epi32(u[0], K32One);
michael@0	1813	u[1] = _mm_add_epi32(u[1], K32One);
michael@0	1814	u[2] = _mm_add_epi32(u[2], K32One);
michael@0	1815	u[3] = _mm_add_epi32(u[3], K32One);
michael@0	1816	u[4] = _mm_add_epi32(u[4], K32One);
michael@0	1817	u[5] = _mm_add_epi32(u[5], K32One);
michael@0	1818	u[6] = _mm_add_epi32(u[6], K32One);
michael@0	1819	u[7] = _mm_add_epi32(u[7], K32One);
michael@0	1820
michael@0	1821	u[0] = _mm_srai_epi32(u[0], 2);
michael@0	1822	u[1] = _mm_srai_epi32(u[1], 2);
michael@0	1823	u[2] = _mm_srai_epi32(u[2], 2);
michael@0	1824	u[3] = _mm_srai_epi32(u[3], 2);
michael@0	1825	u[4] = _mm_srai_epi32(u[4], 2);
michael@0	1826	u[5] = _mm_srai_epi32(u[5], 2);
michael@0	1827	u[6] = _mm_srai_epi32(u[6], 2);
michael@0	1828	u[7] = _mm_srai_epi32(u[7], 2);
michael@0	1829
michael@0	1830	out[ 4] = _mm_packs_epi32(u[0], u[1]);
michael@0	1831	out[20] = _mm_packs_epi32(u[2], u[3]);
michael@0	1832	out[12] = _mm_packs_epi32(u[4], u[5]);
michael@0	1833	out[28] = _mm_packs_epi32(u[6], u[7]);
michael@0	1834	}
michael@0	1835	{
michael@0	1836	lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]);
michael@0	1837	lstep3[17] = _mm_add_epi32(lstep2[19], lstep1[17]);
michael@0	1838	lstep3[18] = _mm_sub_epi32(lstep1[16], lstep2[18]);
michael@0	1839	lstep3[19] = _mm_sub_epi32(lstep1[17], lstep2[19]);
michael@0	1840	lstep3[20] = _mm_sub_epi32(lstep1[22], lstep2[20]);
michael@0	1841	lstep3[21] = _mm_sub_epi32(lstep1[23], lstep2[21]);
michael@0	1842	lstep3[22] = _mm_add_epi32(lstep2[20], lstep1[22]);
michael@0	1843	lstep3[23] = _mm_add_epi32(lstep2[21], lstep1[23]);
michael@0	1844	lstep3[24] = _mm_add_epi32(lstep2[26], lstep1[24]);
michael@0	1845	lstep3[25] = _mm_add_epi32(lstep2[27], lstep1[25]);
michael@0	1846	lstep3[26] = _mm_sub_epi32(lstep1[24], lstep2[26]);
michael@0	1847	lstep3[27] = _mm_sub_epi32(lstep1[25], lstep2[27]);
michael@0	1848	lstep3[28] = _mm_sub_epi32(lstep1[30], lstep2[28]);
michael@0	1849	lstep3[29] = _mm_sub_epi32(lstep1[31], lstep2[29]);
michael@0	1850	lstep3[30] = _mm_add_epi32(lstep2[28], lstep1[30]);
michael@0	1851	lstep3[31] = _mm_add_epi32(lstep2[29], lstep1[31]);
michael@0	1852	}
michael@0	1853	{
michael@0	1854	const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
michael@0	1855	const __m128i k32_m28_m04 = pair_set_epi32(-cospi_28_64, -cospi_4_64);
michael@0	1856	const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
michael@0	1857	const __m128i k32_m12_m20 = pair_set_epi32(-cospi_12_64,
michael@0	1858	-cospi_20_64);
michael@0	1859	const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
michael@0	1860	const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
michael@0	1861
michael@0	1862	u[ 0] = _mm_unpacklo_epi32(lstep2[34], lstep2[60]);
michael@0	1863	u[ 1] = _mm_unpackhi_epi32(lstep2[34], lstep2[60]);
michael@0	1864	u[ 2] = _mm_unpacklo_epi32(lstep2[35], lstep2[61]);
michael@0	1865	u[ 3] = _mm_unpackhi_epi32(lstep2[35], lstep2[61]);
michael@0	1866	u[ 4] = _mm_unpacklo_epi32(lstep2[36], lstep2[58]);
michael@0	1867	u[ 5] = _mm_unpackhi_epi32(lstep2[36], lstep2[58]);
michael@0	1868	u[ 6] = _mm_unpacklo_epi32(lstep2[37], lstep2[59]);
michael@0	1869	u[ 7] = _mm_unpackhi_epi32(lstep2[37], lstep2[59]);
michael@0	1870	u[ 8] = _mm_unpacklo_epi32(lstep2[42], lstep2[52]);
michael@0	1871	u[ 9] = _mm_unpackhi_epi32(lstep2[42], lstep2[52]);
michael@0	1872	u[10] = _mm_unpacklo_epi32(lstep2[43], lstep2[53]);
michael@0	1873	u[11] = _mm_unpackhi_epi32(lstep2[43], lstep2[53]);
michael@0	1874	u[12] = _mm_unpacklo_epi32(lstep2[44], lstep2[50]);
michael@0	1875	u[13] = _mm_unpackhi_epi32(lstep2[44], lstep2[50]);
michael@0	1876	u[14] = _mm_unpacklo_epi32(lstep2[45], lstep2[51]);
michael@0	1877	u[15] = _mm_unpackhi_epi32(lstep2[45], lstep2[51]);
michael@0	1878
michael@0	1879	v[ 0] = k_madd_epi32(u[ 0], k32_m04_p28);
michael@0	1880	v[ 1] = k_madd_epi32(u[ 1], k32_m04_p28);
michael@0	1881	v[ 2] = k_madd_epi32(u[ 2], k32_m04_p28);
michael@0	1882	v[ 3] = k_madd_epi32(u[ 3], k32_m04_p28);
michael@0	1883	v[ 4] = k_madd_epi32(u[ 4], k32_m28_m04);
michael@0	1884	v[ 5] = k_madd_epi32(u[ 5], k32_m28_m04);
michael@0	1885	v[ 6] = k_madd_epi32(u[ 6], k32_m28_m04);
michael@0	1886	v[ 7] = k_madd_epi32(u[ 7], k32_m28_m04);
michael@0	1887	v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
michael@0	1888	v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
michael@0	1889	v[10] = k_madd_epi32(u[10], k32_m20_p12);
michael@0	1890	v[11] = k_madd_epi32(u[11], k32_m20_p12);
michael@0	1891	v[12] = k_madd_epi32(u[12], k32_m12_m20);
michael@0	1892	v[13] = k_madd_epi32(u[13], k32_m12_m20);
michael@0	1893	v[14] = k_madd_epi32(u[14], k32_m12_m20);
michael@0	1894	v[15] = k_madd_epi32(u[15], k32_m12_m20);
michael@0	1895	v[16] = k_madd_epi32(u[12], k32_m20_p12);
michael@0	1896	v[17] = k_madd_epi32(u[13], k32_m20_p12);
michael@0	1897	v[18] = k_madd_epi32(u[14], k32_m20_p12);
michael@0	1898	v[19] = k_madd_epi32(u[15], k32_m20_p12);
michael@0	1899	v[20] = k_madd_epi32(u[ 8], k32_p12_p20);
michael@0	1900	v[21] = k_madd_epi32(u[ 9], k32_p12_p20);
michael@0	1901	v[22] = k_madd_epi32(u[10], k32_p12_p20);
michael@0	1902	v[23] = k_madd_epi32(u[11], k32_p12_p20);
michael@0	1903	v[24] = k_madd_epi32(u[ 4], k32_m04_p28);
michael@0	1904	v[25] = k_madd_epi32(u[ 5], k32_m04_p28);
michael@0	1905	v[26] = k_madd_epi32(u[ 6], k32_m04_p28);
michael@0	1906	v[27] = k_madd_epi32(u[ 7], k32_m04_p28);
michael@0	1907	v[28] = k_madd_epi32(u[ 0], k32_p28_p04);
michael@0	1908	v[29] = k_madd_epi32(u[ 1], k32_p28_p04);
michael@0	1909	v[30] = k_madd_epi32(u[ 2], k32_p28_p04);
michael@0	1910	v[31] = k_madd_epi32(u[ 3], k32_p28_p04);
michael@0	1911
michael@0	1912	u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
michael@0	1913	u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
michael@0	1914	u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
michael@0	1915	u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
michael@0	1916	u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
michael@0	1917	u[ 5] = k_packs_epi64(v[10], v[11]);
michael@0	1918	u[ 6] = k_packs_epi64(v[12], v[13]);
michael@0	1919	u[ 7] = k_packs_epi64(v[14], v[15]);
michael@0	1920	u[ 8] = k_packs_epi64(v[16], v[17]);
michael@0	1921	u[ 9] = k_packs_epi64(v[18], v[19]);
michael@0	1922	u[10] = k_packs_epi64(v[20], v[21]);
michael@0	1923	u[11] = k_packs_epi64(v[22], v[23]);
michael@0	1924	u[12] = k_packs_epi64(v[24], v[25]);
michael@0	1925	u[13] = k_packs_epi64(v[26], v[27]);
michael@0	1926	u[14] = k_packs_epi64(v[28], v[29]);
michael@0	1927	u[15] = k_packs_epi64(v[30], v[31]);
michael@0	1928
michael@0	1929	v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
michael@0	1930	v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
michael@0	1931	v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
michael@0	1932	v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
michael@0	1933	v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
michael@0	1934	v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
michael@0	1935	v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
michael@0	1936	v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
michael@0	1937	v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
michael@0	1938	v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
michael@0	1939	v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
michael@0	1940	v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
michael@0	1941	v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
michael@0	1942	v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
michael@0	1943	v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
michael@0	1944	v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
michael@0	1945
michael@0	1946	lstep3[34] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
michael@0	1947	lstep3[35] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
michael@0	1948	lstep3[36] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
michael@0	1949	lstep3[37] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
michael@0	1950	lstep3[42] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
michael@0	1951	lstep3[43] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
michael@0	1952	lstep3[44] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
michael@0	1953	lstep3[45] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
michael@0	1954	lstep3[50] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
michael@0	1955	lstep3[51] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
michael@0	1956	lstep3[52] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
michael@0	1957	lstep3[53] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
michael@0	1958	lstep3[58] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
michael@0	1959	lstep3[59] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
michael@0	1960	lstep3[60] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
michael@0	1961	lstep3[61] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
michael@0	1962	}
michael@0	1963	// stage 7
michael@0	1964	{
michael@0	1965	const __m128i k32_p30_p02 = pair_set_epi32(cospi_30_64, cospi_2_64);
michael@0	1966	const __m128i k32_p14_p18 = pair_set_epi32(cospi_14_64, cospi_18_64);
michael@0	1967	const __m128i k32_p22_p10 = pair_set_epi32(cospi_22_64, cospi_10_64);
michael@0	1968	const __m128i k32_p06_p26 = pair_set_epi32(cospi_6_64, cospi_26_64);
michael@0	1969	const __m128i k32_m26_p06 = pair_set_epi32(-cospi_26_64, cospi_6_64);
michael@0	1970	const __m128i k32_m10_p22 = pair_set_epi32(-cospi_10_64, cospi_22_64);
michael@0	1971	const __m128i k32_m18_p14 = pair_set_epi32(-cospi_18_64, cospi_14_64);
michael@0	1972	const __m128i k32_m02_p30 = pair_set_epi32(-cospi_2_64, cospi_30_64);
michael@0	1973
michael@0	1974	u[ 0] = _mm_unpacklo_epi32(lstep3[16], lstep3[30]);
michael@0	1975	u[ 1] = _mm_unpackhi_epi32(lstep3[16], lstep3[30]);
michael@0	1976	u[ 2] = _mm_unpacklo_epi32(lstep3[17], lstep3[31]);
michael@0	1977	u[ 3] = _mm_unpackhi_epi32(lstep3[17], lstep3[31]);
michael@0	1978	u[ 4] = _mm_unpacklo_epi32(lstep3[18], lstep3[28]);
michael@0	1979	u[ 5] = _mm_unpackhi_epi32(lstep3[18], lstep3[28]);
michael@0	1980	u[ 6] = _mm_unpacklo_epi32(lstep3[19], lstep3[29]);
michael@0	1981	u[ 7] = _mm_unpackhi_epi32(lstep3[19], lstep3[29]);
michael@0	1982	u[ 8] = _mm_unpacklo_epi32(lstep3[20], lstep3[26]);
michael@0	1983	u[ 9] = _mm_unpackhi_epi32(lstep3[20], lstep3[26]);
michael@0	1984	u[10] = _mm_unpacklo_epi32(lstep3[21], lstep3[27]);
michael@0	1985	u[11] = _mm_unpackhi_epi32(lstep3[21], lstep3[27]);
michael@0	1986	u[12] = _mm_unpacklo_epi32(lstep3[22], lstep3[24]);
michael@0	1987	u[13] = _mm_unpackhi_epi32(lstep3[22], lstep3[24]);
michael@0	1988	u[14] = _mm_unpacklo_epi32(lstep3[23], lstep3[25]);
michael@0	1989	u[15] = _mm_unpackhi_epi32(lstep3[23], lstep3[25]);
michael@0	1990
michael@0	1991	v[ 0] = k_madd_epi32(u[ 0], k32_p30_p02);
michael@0	1992	v[ 1] = k_madd_epi32(u[ 1], k32_p30_p02);
michael@0	1993	v[ 2] = k_madd_epi32(u[ 2], k32_p30_p02);
michael@0	1994	v[ 3] = k_madd_epi32(u[ 3], k32_p30_p02);
michael@0	1995	v[ 4] = k_madd_epi32(u[ 4], k32_p14_p18);
michael@0	1996	v[ 5] = k_madd_epi32(u[ 5], k32_p14_p18);
michael@0	1997	v[ 6] = k_madd_epi32(u[ 6], k32_p14_p18);
michael@0	1998	v[ 7] = k_madd_epi32(u[ 7], k32_p14_p18);
michael@0	1999	v[ 8] = k_madd_epi32(u[ 8], k32_p22_p10);
michael@0	2000	v[ 9] = k_madd_epi32(u[ 9], k32_p22_p10);
michael@0	2001	v[10] = k_madd_epi32(u[10], k32_p22_p10);
michael@0	2002	v[11] = k_madd_epi32(u[11], k32_p22_p10);
michael@0	2003	v[12] = k_madd_epi32(u[12], k32_p06_p26);
michael@0	2004	v[13] = k_madd_epi32(u[13], k32_p06_p26);
michael@0	2005	v[14] = k_madd_epi32(u[14], k32_p06_p26);
michael@0	2006	v[15] = k_madd_epi32(u[15], k32_p06_p26);
michael@0	2007	v[16] = k_madd_epi32(u[12], k32_m26_p06);
michael@0	2008	v[17] = k_madd_epi32(u[13], k32_m26_p06);
michael@0	2009	v[18] = k_madd_epi32(u[14], k32_m26_p06);
michael@0	2010	v[19] = k_madd_epi32(u[15], k32_m26_p06);
michael@0	2011	v[20] = k_madd_epi32(u[ 8], k32_m10_p22);
michael@0	2012	v[21] = k_madd_epi32(u[ 9], k32_m10_p22);
michael@0	2013	v[22] = k_madd_epi32(u[10], k32_m10_p22);
michael@0	2014	v[23] = k_madd_epi32(u[11], k32_m10_p22);
michael@0	2015	v[24] = k_madd_epi32(u[ 4], k32_m18_p14);
michael@0	2016	v[25] = k_madd_epi32(u[ 5], k32_m18_p14);
michael@0	2017	v[26] = k_madd_epi32(u[ 6], k32_m18_p14);
michael@0	2018	v[27] = k_madd_epi32(u[ 7], k32_m18_p14);
michael@0	2019	v[28] = k_madd_epi32(u[ 0], k32_m02_p30);
michael@0	2020	v[29] = k_madd_epi32(u[ 1], k32_m02_p30);
michael@0	2021	v[30] = k_madd_epi32(u[ 2], k32_m02_p30);
michael@0	2022	v[31] = k_madd_epi32(u[ 3], k32_m02_p30);
michael@0	2023
michael@0	2024	u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
michael@0	2025	u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
michael@0	2026	u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
michael@0	2027	u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
michael@0	2028	u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
michael@0	2029	u[ 5] = k_packs_epi64(v[10], v[11]);
michael@0	2030	u[ 6] = k_packs_epi64(v[12], v[13]);
michael@0	2031	u[ 7] = k_packs_epi64(v[14], v[15]);
michael@0	2032	u[ 8] = k_packs_epi64(v[16], v[17]);
michael@0	2033	u[ 9] = k_packs_epi64(v[18], v[19]);
michael@0	2034	u[10] = k_packs_epi64(v[20], v[21]);
michael@0	2035	u[11] = k_packs_epi64(v[22], v[23]);
michael@0	2036	u[12] = k_packs_epi64(v[24], v[25]);
michael@0	2037	u[13] = k_packs_epi64(v[26], v[27]);
michael@0	2038	u[14] = k_packs_epi64(v[28], v[29]);
michael@0	2039	u[15] = k_packs_epi64(v[30], v[31]);
michael@0	2040
michael@0	2041	v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
michael@0	2042	v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
michael@0	2043	v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
michael@0	2044	v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
michael@0	2045	v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
michael@0	2046	v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
michael@0	2047	v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
michael@0	2048	v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
michael@0	2049	v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
michael@0	2050	v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
michael@0	2051	v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
michael@0	2052	v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
michael@0	2053	v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
michael@0	2054	v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
michael@0	2055	v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
michael@0	2056	v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
michael@0	2057
michael@0	2058	u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
michael@0	2059	u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
michael@0	2060	u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
michael@0	2061	u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
michael@0	2062	u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
michael@0	2063	u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
michael@0	2064	u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
michael@0	2065	u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
michael@0	2066	u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
michael@0	2067	u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
michael@0	2068	u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
michael@0	2069	u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
michael@0	2070	u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
michael@0	2071	u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
michael@0	2072	u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
michael@0	2073	u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
michael@0	2074
michael@0	2075	v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
michael@0	2076	v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
michael@0	2077	v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
michael@0	2078	v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
michael@0	2079	v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
michael@0	2080	v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
michael@0	2081	v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
michael@0	2082	v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
michael@0	2083	v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
michael@0	2084	v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
michael@0	2085	v[10] = _mm_cmplt_epi32(u[10], kZero);
michael@0	2086	v[11] = _mm_cmplt_epi32(u[11], kZero);
michael@0	2087	v[12] = _mm_cmplt_epi32(u[12], kZero);
michael@0	2088	v[13] = _mm_cmplt_epi32(u[13], kZero);
michael@0	2089	v[14] = _mm_cmplt_epi32(u[14], kZero);
michael@0	2090	v[15] = _mm_cmplt_epi32(u[15], kZero);
michael@0	2091
michael@0	2092	u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
michael@0	2093	u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
michael@0	2094	u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
michael@0	2095	u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
michael@0	2096	u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
michael@0	2097	u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
michael@0	2098	u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
michael@0	2099	u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
michael@0	2100	u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
michael@0	2101	u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
michael@0	2102	u[10] = _mm_sub_epi32(u[10], v[10]);
michael@0	2103	u[11] = _mm_sub_epi32(u[11], v[11]);
michael@0	2104	u[12] = _mm_sub_epi32(u[12], v[12]);
michael@0	2105	u[13] = _mm_sub_epi32(u[13], v[13]);
michael@0	2106	u[14] = _mm_sub_epi32(u[14], v[14]);
michael@0	2107	u[15] = _mm_sub_epi32(u[15], v[15]);
michael@0	2108
michael@0	2109	v[ 0] = _mm_add_epi32(u[ 0], K32One);
michael@0	2110	v[ 1] = _mm_add_epi32(u[ 1], K32One);
michael@0	2111	v[ 2] = _mm_add_epi32(u[ 2], K32One);
michael@0	2112	v[ 3] = _mm_add_epi32(u[ 3], K32One);
michael@0	2113	v[ 4] = _mm_add_epi32(u[ 4], K32One);
michael@0	2114	v[ 5] = _mm_add_epi32(u[ 5], K32One);
michael@0	2115	v[ 6] = _mm_add_epi32(u[ 6], K32One);
michael@0	2116	v[ 7] = _mm_add_epi32(u[ 7], K32One);
michael@0	2117	v[ 8] = _mm_add_epi32(u[ 8], K32One);
michael@0	2118	v[ 9] = _mm_add_epi32(u[ 9], K32One);
michael@0	2119	v[10] = _mm_add_epi32(u[10], K32One);
michael@0	2120	v[11] = _mm_add_epi32(u[11], K32One);
michael@0	2121	v[12] = _mm_add_epi32(u[12], K32One);
michael@0	2122	v[13] = _mm_add_epi32(u[13], K32One);
michael@0	2123	v[14] = _mm_add_epi32(u[14], K32One);
michael@0	2124	v[15] = _mm_add_epi32(u[15], K32One);
michael@0	2125
michael@0	2126	u[ 0] = _mm_srai_epi32(v[ 0], 2);
michael@0	2127	u[ 1] = _mm_srai_epi32(v[ 1], 2);
michael@0	2128	u[ 2] = _mm_srai_epi32(v[ 2], 2);
michael@0	2129	u[ 3] = _mm_srai_epi32(v[ 3], 2);
michael@0	2130	u[ 4] = _mm_srai_epi32(v[ 4], 2);
michael@0	2131	u[ 5] = _mm_srai_epi32(v[ 5], 2);
michael@0	2132	u[ 6] = _mm_srai_epi32(v[ 6], 2);
michael@0	2133	u[ 7] = _mm_srai_epi32(v[ 7], 2);
michael@0	2134	u[ 8] = _mm_srai_epi32(v[ 8], 2);
michael@0	2135	u[ 9] = _mm_srai_epi32(v[ 9], 2);
michael@0	2136	u[10] = _mm_srai_epi32(v[10], 2);
michael@0	2137	u[11] = _mm_srai_epi32(v[11], 2);
michael@0	2138	u[12] = _mm_srai_epi32(v[12], 2);
michael@0	2139	u[13] = _mm_srai_epi32(v[13], 2);
michael@0	2140	u[14] = _mm_srai_epi32(v[14], 2);
michael@0	2141	u[15] = _mm_srai_epi32(v[15], 2);
michael@0	2142
michael@0	2143	out[ 2] = _mm_packs_epi32(u[0], u[1]);
michael@0	2144	out[18] = _mm_packs_epi32(u[2], u[3]);
michael@0	2145	out[10] = _mm_packs_epi32(u[4], u[5]);
michael@0	2146	out[26] = _mm_packs_epi32(u[6], u[7]);
michael@0	2147	out[ 6] = _mm_packs_epi32(u[8], u[9]);
michael@0	2148	out[22] = _mm_packs_epi32(u[10], u[11]);
michael@0	2149	out[14] = _mm_packs_epi32(u[12], u[13]);
michael@0	2150	out[30] = _mm_packs_epi32(u[14], u[15]);
michael@0	2151	}
michael@0	2152	{
michael@0	2153	lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]);
michael@0	2154	lstep1[33] = _mm_add_epi32(lstep3[35], lstep2[33]);
michael@0	2155	lstep1[34] = _mm_sub_epi32(lstep2[32], lstep3[34]);
michael@0	2156	lstep1[35] = _mm_sub_epi32(lstep2[33], lstep3[35]);
michael@0	2157	lstep1[36] = _mm_sub_epi32(lstep2[38], lstep3[36]);
michael@0	2158	lstep1[37] = _mm_sub_epi32(lstep2[39], lstep3[37]);
michael@0	2159	lstep1[38] = _mm_add_epi32(lstep3[36], lstep2[38]);
michael@0	2160	lstep1[39] = _mm_add_epi32(lstep3[37], lstep2[39]);
michael@0	2161	lstep1[40] = _mm_add_epi32(lstep3[42], lstep2[40]);
michael@0	2162	lstep1[41] = _mm_add_epi32(lstep3[43], lstep2[41]);
michael@0	2163	lstep1[42] = _mm_sub_epi32(lstep2[40], lstep3[42]);
michael@0	2164	lstep1[43] = _mm_sub_epi32(lstep2[41], lstep3[43]);
michael@0	2165	lstep1[44] = _mm_sub_epi32(lstep2[46], lstep3[44]);
michael@0	2166	lstep1[45] = _mm_sub_epi32(lstep2[47], lstep3[45]);
michael@0	2167	lstep1[46] = _mm_add_epi32(lstep3[44], lstep2[46]);
michael@0	2168	lstep1[47] = _mm_add_epi32(lstep3[45], lstep2[47]);
michael@0	2169	lstep1[48] = _mm_add_epi32(lstep3[50], lstep2[48]);
michael@0	2170	lstep1[49] = _mm_add_epi32(lstep3[51], lstep2[49]);
michael@0	2171	lstep1[50] = _mm_sub_epi32(lstep2[48], lstep3[50]);
michael@0	2172	lstep1[51] = _mm_sub_epi32(lstep2[49], lstep3[51]);
michael@0	2173	lstep1[52] = _mm_sub_epi32(lstep2[54], lstep3[52]);
michael@0	2174	lstep1[53] = _mm_sub_epi32(lstep2[55], lstep3[53]);
michael@0	2175	lstep1[54] = _mm_add_epi32(lstep3[52], lstep2[54]);
michael@0	2176	lstep1[55] = _mm_add_epi32(lstep3[53], lstep2[55]);
michael@0	2177	lstep1[56] = _mm_add_epi32(lstep3[58], lstep2[56]);
michael@0	2178	lstep1[57] = _mm_add_epi32(lstep3[59], lstep2[57]);
michael@0	2179	lstep1[58] = _mm_sub_epi32(lstep2[56], lstep3[58]);
michael@0	2180	lstep1[59] = _mm_sub_epi32(lstep2[57], lstep3[59]);
michael@0	2181	lstep1[60] = _mm_sub_epi32(lstep2[62], lstep3[60]);
michael@0	2182	lstep1[61] = _mm_sub_epi32(lstep2[63], lstep3[61]);
michael@0	2183	lstep1[62] = _mm_add_epi32(lstep3[60], lstep2[62]);
michael@0	2184	lstep1[63] = _mm_add_epi32(lstep3[61], lstep2[63]);
michael@0	2185	}
michael@0	2186	// stage 8
michael@0	2187	{
michael@0	2188	const __m128i k32_p31_p01 = pair_set_epi32(cospi_31_64, cospi_1_64);
michael@0	2189	const __m128i k32_p15_p17 = pair_set_epi32(cospi_15_64, cospi_17_64);
michael@0	2190	const __m128i k32_p23_p09 = pair_set_epi32(cospi_23_64, cospi_9_64);
michael@0	2191	const __m128i k32_p07_p25 = pair_set_epi32(cospi_7_64, cospi_25_64);
michael@0	2192	const __m128i k32_m25_p07 = pair_set_epi32(-cospi_25_64, cospi_7_64);
michael@0	2193	const __m128i k32_m09_p23 = pair_set_epi32(-cospi_9_64, cospi_23_64);
michael@0	2194	const __m128i k32_m17_p15 = pair_set_epi32(-cospi_17_64, cospi_15_64);
michael@0	2195	const __m128i k32_m01_p31 = pair_set_epi32(-cospi_1_64, cospi_31_64);
michael@0	2196
michael@0	2197	u[ 0] = _mm_unpacklo_epi32(lstep1[32], lstep1[62]);
michael@0	2198	u[ 1] = _mm_unpackhi_epi32(lstep1[32], lstep1[62]);
michael@0	2199	u[ 2] = _mm_unpacklo_epi32(lstep1[33], lstep1[63]);
michael@0	2200	u[ 3] = _mm_unpackhi_epi32(lstep1[33], lstep1[63]);
michael@0	2201	u[ 4] = _mm_unpacklo_epi32(lstep1[34], lstep1[60]);
michael@0	2202	u[ 5] = _mm_unpackhi_epi32(lstep1[34], lstep1[60]);
michael@0	2203	u[ 6] = _mm_unpacklo_epi32(lstep1[35], lstep1[61]);
michael@0	2204	u[ 7] = _mm_unpackhi_epi32(lstep1[35], lstep1[61]);
michael@0	2205	u[ 8] = _mm_unpacklo_epi32(lstep1[36], lstep1[58]);
michael@0	2206	u[ 9] = _mm_unpackhi_epi32(lstep1[36], lstep1[58]);
michael@0	2207	u[10] = _mm_unpacklo_epi32(lstep1[37], lstep1[59]);
michael@0	2208	u[11] = _mm_unpackhi_epi32(lstep1[37], lstep1[59]);
michael@0	2209	u[12] = _mm_unpacklo_epi32(lstep1[38], lstep1[56]);
michael@0	2210	u[13] = _mm_unpackhi_epi32(lstep1[38], lstep1[56]);
michael@0	2211	u[14] = _mm_unpacklo_epi32(lstep1[39], lstep1[57]);
michael@0	2212	u[15] = _mm_unpackhi_epi32(lstep1[39], lstep1[57]);
michael@0	2213
michael@0	2214	v[ 0] = k_madd_epi32(u[ 0], k32_p31_p01);
michael@0	2215	v[ 1] = k_madd_epi32(u[ 1], k32_p31_p01);
michael@0	2216	v[ 2] = k_madd_epi32(u[ 2], k32_p31_p01);
michael@0	2217	v[ 3] = k_madd_epi32(u[ 3], k32_p31_p01);
michael@0	2218	v[ 4] = k_madd_epi32(u[ 4], k32_p15_p17);
michael@0	2219	v[ 5] = k_madd_epi32(u[ 5], k32_p15_p17);
michael@0	2220	v[ 6] = k_madd_epi32(u[ 6], k32_p15_p17);
michael@0	2221	v[ 7] = k_madd_epi32(u[ 7], k32_p15_p17);
michael@0	2222	v[ 8] = k_madd_epi32(u[ 8], k32_p23_p09);
michael@0	2223	v[ 9] = k_madd_epi32(u[ 9], k32_p23_p09);
michael@0	2224	v[10] = k_madd_epi32(u[10], k32_p23_p09);
michael@0	2225	v[11] = k_madd_epi32(u[11], k32_p23_p09);
michael@0	2226	v[12] = k_madd_epi32(u[12], k32_p07_p25);
michael@0	2227	v[13] = k_madd_epi32(u[13], k32_p07_p25);
michael@0	2228	v[14] = k_madd_epi32(u[14], k32_p07_p25);
michael@0	2229	v[15] = k_madd_epi32(u[15], k32_p07_p25);
michael@0	2230	v[16] = k_madd_epi32(u[12], k32_m25_p07);
michael@0	2231	v[17] = k_madd_epi32(u[13], k32_m25_p07);
michael@0	2232	v[18] = k_madd_epi32(u[14], k32_m25_p07);
michael@0	2233	v[19] = k_madd_epi32(u[15], k32_m25_p07);
michael@0	2234	v[20] = k_madd_epi32(u[ 8], k32_m09_p23);
michael@0	2235	v[21] = k_madd_epi32(u[ 9], k32_m09_p23);
michael@0	2236	v[22] = k_madd_epi32(u[10], k32_m09_p23);
michael@0	2237	v[23] = k_madd_epi32(u[11], k32_m09_p23);
michael@0	2238	v[24] = k_madd_epi32(u[ 4], k32_m17_p15);
michael@0	2239	v[25] = k_madd_epi32(u[ 5], k32_m17_p15);
michael@0	2240	v[26] = k_madd_epi32(u[ 6], k32_m17_p15);
michael@0	2241	v[27] = k_madd_epi32(u[ 7], k32_m17_p15);
michael@0	2242	v[28] = k_madd_epi32(u[ 0], k32_m01_p31);
michael@0	2243	v[29] = k_madd_epi32(u[ 1], k32_m01_p31);
michael@0	2244	v[30] = k_madd_epi32(u[ 2], k32_m01_p31);
michael@0	2245	v[31] = k_madd_epi32(u[ 3], k32_m01_p31);
michael@0	2246
michael@0	2247	u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
michael@0	2248	u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
michael@0	2249	u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
michael@0	2250	u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
michael@0	2251	u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
michael@0	2252	u[ 5] = k_packs_epi64(v[10], v[11]);
michael@0	2253	u[ 6] = k_packs_epi64(v[12], v[13]);
michael@0	2254	u[ 7] = k_packs_epi64(v[14], v[15]);
michael@0	2255	u[ 8] = k_packs_epi64(v[16], v[17]);
michael@0	2256	u[ 9] = k_packs_epi64(v[18], v[19]);
michael@0	2257	u[10] = k_packs_epi64(v[20], v[21]);
michael@0	2258	u[11] = k_packs_epi64(v[22], v[23]);
michael@0	2259	u[12] = k_packs_epi64(v[24], v[25]);
michael@0	2260	u[13] = k_packs_epi64(v[26], v[27]);
michael@0	2261	u[14] = k_packs_epi64(v[28], v[29]);
michael@0	2262	u[15] = k_packs_epi64(v[30], v[31]);
michael@0	2263
michael@0	2264	v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
michael@0	2265	v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
michael@0	2266	v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
michael@0	2267	v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
michael@0	2268	v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
michael@0	2269	v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
michael@0	2270	v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
michael@0	2271	v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
michael@0	2272	v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
michael@0	2273	v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
michael@0	2274	v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
michael@0	2275	v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
michael@0	2276	v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
michael@0	2277	v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
michael@0	2278	v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
michael@0	2279	v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
michael@0	2280
michael@0	2281	u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
michael@0	2282	u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
michael@0	2283	u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
michael@0	2284	u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
michael@0	2285	u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
michael@0	2286	u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
michael@0	2287	u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
michael@0	2288	u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
michael@0	2289	u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
michael@0	2290	u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
michael@0	2291	u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
michael@0	2292	u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
michael@0	2293	u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
michael@0	2294	u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
michael@0	2295	u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
michael@0	2296	u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
michael@0	2297
michael@0	2298	v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
michael@0	2299	v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
michael@0	2300	v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
michael@0	2301	v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
michael@0	2302	v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
michael@0	2303	v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
michael@0	2304	v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
michael@0	2305	v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
michael@0	2306	v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
michael@0	2307	v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
michael@0	2308	v[10] = _mm_cmplt_epi32(u[10], kZero);
michael@0	2309	v[11] = _mm_cmplt_epi32(u[11], kZero);
michael@0	2310	v[12] = _mm_cmplt_epi32(u[12], kZero);
michael@0	2311	v[13] = _mm_cmplt_epi32(u[13], kZero);
michael@0	2312	v[14] = _mm_cmplt_epi32(u[14], kZero);
michael@0	2313	v[15] = _mm_cmplt_epi32(u[15], kZero);
michael@0	2314
michael@0	2315	u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
michael@0	2316	u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
michael@0	2317	u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
michael@0	2318	u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
michael@0	2319	u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
michael@0	2320	u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
michael@0	2321	u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
michael@0	2322	u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
michael@0	2323	u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
michael@0	2324	u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
michael@0	2325	u[10] = _mm_sub_epi32(u[10], v[10]);
michael@0	2326	u[11] = _mm_sub_epi32(u[11], v[11]);
michael@0	2327	u[12] = _mm_sub_epi32(u[12], v[12]);
michael@0	2328	u[13] = _mm_sub_epi32(u[13], v[13]);
michael@0	2329	u[14] = _mm_sub_epi32(u[14], v[14]);
michael@0	2330	u[15] = _mm_sub_epi32(u[15], v[15]);
michael@0	2331
michael@0	2332	v[0] = _mm_add_epi32(u[0], K32One);
michael@0	2333	v[1] = _mm_add_epi32(u[1], K32One);
michael@0	2334	v[2] = _mm_add_epi32(u[2], K32One);
michael@0	2335	v[3] = _mm_add_epi32(u[3], K32One);
michael@0	2336	v[4] = _mm_add_epi32(u[4], K32One);
michael@0	2337	v[5] = _mm_add_epi32(u[5], K32One);
michael@0	2338	v[6] = _mm_add_epi32(u[6], K32One);
michael@0	2339	v[7] = _mm_add_epi32(u[7], K32One);
michael@0	2340	v[8] = _mm_add_epi32(u[8], K32One);
michael@0	2341	v[9] = _mm_add_epi32(u[9], K32One);
michael@0	2342	v[10] = _mm_add_epi32(u[10], K32One);
michael@0	2343	v[11] = _mm_add_epi32(u[11], K32One);
michael@0	2344	v[12] = _mm_add_epi32(u[12], K32One);
michael@0	2345	v[13] = _mm_add_epi32(u[13], K32One);
michael@0	2346	v[14] = _mm_add_epi32(u[14], K32One);
michael@0	2347	v[15] = _mm_add_epi32(u[15], K32One);
michael@0	2348
michael@0	2349	u[0] = _mm_srai_epi32(v[0], 2);
michael@0	2350	u[1] = _mm_srai_epi32(v[1], 2);
michael@0	2351	u[2] = _mm_srai_epi32(v[2], 2);
michael@0	2352	u[3] = _mm_srai_epi32(v[3], 2);
michael@0	2353	u[4] = _mm_srai_epi32(v[4], 2);
michael@0	2354	u[5] = _mm_srai_epi32(v[5], 2);
michael@0	2355	u[6] = _mm_srai_epi32(v[6], 2);
michael@0	2356	u[7] = _mm_srai_epi32(v[7], 2);
michael@0	2357	u[8] = _mm_srai_epi32(v[8], 2);
michael@0	2358	u[9] = _mm_srai_epi32(v[9], 2);
michael@0	2359	u[10] = _mm_srai_epi32(v[10], 2);
michael@0	2360	u[11] = _mm_srai_epi32(v[11], 2);
michael@0	2361	u[12] = _mm_srai_epi32(v[12], 2);
michael@0	2362	u[13] = _mm_srai_epi32(v[13], 2);
michael@0	2363	u[14] = _mm_srai_epi32(v[14], 2);
michael@0	2364	u[15] = _mm_srai_epi32(v[15], 2);
michael@0	2365
michael@0	2366	out[ 1] = _mm_packs_epi32(u[0], u[1]);
michael@0	2367	out[17] = _mm_packs_epi32(u[2], u[3]);
michael@0	2368	out[ 9] = _mm_packs_epi32(u[4], u[5]);
michael@0	2369	out[25] = _mm_packs_epi32(u[6], u[7]);
michael@0	2370	out[ 7] = _mm_packs_epi32(u[8], u[9]);
michael@0	2371	out[23] = _mm_packs_epi32(u[10], u[11]);
michael@0	2372	out[15] = _mm_packs_epi32(u[12], u[13]);
michael@0	2373	out[31] = _mm_packs_epi32(u[14], u[15]);
michael@0	2374	}
michael@0	2375	{
michael@0	2376	const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64);
michael@0	2377	const __m128i k32_p11_p21 = pair_set_epi32(cospi_11_64, cospi_21_64);
michael@0	2378	const __m128i k32_p19_p13 = pair_set_epi32(cospi_19_64, cospi_13_64);
michael@0	2379	const __m128i k32_p03_p29 = pair_set_epi32(cospi_3_64, cospi_29_64);
michael@0	2380	const __m128i k32_m29_p03 = pair_set_epi32(-cospi_29_64, cospi_3_64);
michael@0	2381	const __m128i k32_m13_p19 = pair_set_epi32(-cospi_13_64, cospi_19_64);
michael@0	2382	const __m128i k32_m21_p11 = pair_set_epi32(-cospi_21_64, cospi_11_64);
michael@0	2383	const __m128i k32_m05_p27 = pair_set_epi32(-cospi_5_64, cospi_27_64);
michael@0	2384
michael@0	2385	u[ 0] = _mm_unpacklo_epi32(lstep1[40], lstep1[54]);
michael@0	2386	u[ 1] = _mm_unpackhi_epi32(lstep1[40], lstep1[54]);
michael@0	2387	u[ 2] = _mm_unpacklo_epi32(lstep1[41], lstep1[55]);
michael@0	2388	u[ 3] = _mm_unpackhi_epi32(lstep1[41], lstep1[55]);
michael@0	2389	u[ 4] = _mm_unpacklo_epi32(lstep1[42], lstep1[52]);
michael@0	2390	u[ 5] = _mm_unpackhi_epi32(lstep1[42], lstep1[52]);
michael@0	2391	u[ 6] = _mm_unpacklo_epi32(lstep1[43], lstep1[53]);
michael@0	2392	u[ 7] = _mm_unpackhi_epi32(lstep1[43], lstep1[53]);
michael@0	2393	u[ 8] = _mm_unpacklo_epi32(lstep1[44], lstep1[50]);
michael@0	2394	u[ 9] = _mm_unpackhi_epi32(lstep1[44], lstep1[50]);
michael@0	2395	u[10] = _mm_unpacklo_epi32(lstep1[45], lstep1[51]);
michael@0	2396	u[11] = _mm_unpackhi_epi32(lstep1[45], lstep1[51]);
michael@0	2397	u[12] = _mm_unpacklo_epi32(lstep1[46], lstep1[48]);
michael@0	2398	u[13] = _mm_unpackhi_epi32(lstep1[46], lstep1[48]);
michael@0	2399	u[14] = _mm_unpacklo_epi32(lstep1[47], lstep1[49]);
michael@0	2400	u[15] = _mm_unpackhi_epi32(lstep1[47], lstep1[49]);
michael@0	2401
michael@0	2402	v[ 0] = k_madd_epi32(u[ 0], k32_p27_p05);
michael@0	2403	v[ 1] = k_madd_epi32(u[ 1], k32_p27_p05);
michael@0	2404	v[ 2] = k_madd_epi32(u[ 2], k32_p27_p05);
michael@0	2405	v[ 3] = k_madd_epi32(u[ 3], k32_p27_p05);
michael@0	2406	v[ 4] = k_madd_epi32(u[ 4], k32_p11_p21);
michael@0	2407	v[ 5] = k_madd_epi32(u[ 5], k32_p11_p21);
michael@0	2408	v[ 6] = k_madd_epi32(u[ 6], k32_p11_p21);
michael@0	2409	v[ 7] = k_madd_epi32(u[ 7], k32_p11_p21);
michael@0	2410	v[ 8] = k_madd_epi32(u[ 8], k32_p19_p13);
michael@0	2411	v[ 9] = k_madd_epi32(u[ 9], k32_p19_p13);
michael@0	2412	v[10] = k_madd_epi32(u[10], k32_p19_p13);
michael@0	2413	v[11] = k_madd_epi32(u[11], k32_p19_p13);
michael@0	2414	v[12] = k_madd_epi32(u[12], k32_p03_p29);
michael@0	2415	v[13] = k_madd_epi32(u[13], k32_p03_p29);
michael@0	2416	v[14] = k_madd_epi32(u[14], k32_p03_p29);
michael@0	2417	v[15] = k_madd_epi32(u[15], k32_p03_p29);
michael@0	2418	v[16] = k_madd_epi32(u[12], k32_m29_p03);
michael@0	2419	v[17] = k_madd_epi32(u[13], k32_m29_p03);
michael@0	2420	v[18] = k_madd_epi32(u[14], k32_m29_p03);
michael@0	2421	v[19] = k_madd_epi32(u[15], k32_m29_p03);
michael@0	2422	v[20] = k_madd_epi32(u[ 8], k32_m13_p19);
michael@0	2423	v[21] = k_madd_epi32(u[ 9], k32_m13_p19);
michael@0	2424	v[22] = k_madd_epi32(u[10], k32_m13_p19);
michael@0	2425	v[23] = k_madd_epi32(u[11], k32_m13_p19);
michael@0	2426	v[24] = k_madd_epi32(u[ 4], k32_m21_p11);
michael@0	2427	v[25] = k_madd_epi32(u[ 5], k32_m21_p11);
michael@0	2428	v[26] = k_madd_epi32(u[ 6], k32_m21_p11);
michael@0	2429	v[27] = k_madd_epi32(u[ 7], k32_m21_p11);
michael@0	2430	v[28] = k_madd_epi32(u[ 0], k32_m05_p27);
michael@0	2431	v[29] = k_madd_epi32(u[ 1], k32_m05_p27);
michael@0	2432	v[30] = k_madd_epi32(u[ 2], k32_m05_p27);
michael@0	2433	v[31] = k_madd_epi32(u[ 3], k32_m05_p27);
michael@0	2434
michael@0	2435	u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
michael@0	2436	u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
michael@0	2437	u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
michael@0	2438	u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
michael@0	2439	u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
michael@0	2440	u[ 5] = k_packs_epi64(v[10], v[11]);
michael@0	2441	u[ 6] = k_packs_epi64(v[12], v[13]);
michael@0	2442	u[ 7] = k_packs_epi64(v[14], v[15]);
michael@0	2443	u[ 8] = k_packs_epi64(v[16], v[17]);
michael@0	2444	u[ 9] = k_packs_epi64(v[18], v[19]);
michael@0	2445	u[10] = k_packs_epi64(v[20], v[21]);
michael@0	2446	u[11] = k_packs_epi64(v[22], v[23]);
michael@0	2447	u[12] = k_packs_epi64(v[24], v[25]);
michael@0	2448	u[13] = k_packs_epi64(v[26], v[27]);
michael@0	2449	u[14] = k_packs_epi64(v[28], v[29]);
michael@0	2450	u[15] = k_packs_epi64(v[30], v[31]);
michael@0	2451
michael@0	2452	v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
michael@0	2453	v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
michael@0	2454	v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
michael@0	2455	v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
michael@0	2456	v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
michael@0	2457	v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
michael@0	2458	v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
michael@0	2459	v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
michael@0	2460	v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
michael@0	2461	v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
michael@0	2462	v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
michael@0	2463	v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
michael@0	2464	v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
michael@0	2465	v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
michael@0	2466	v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
michael@0	2467	v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
michael@0	2468
michael@0	2469	u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
michael@0	2470	u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
michael@0	2471	u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
michael@0	2472	u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
michael@0	2473	u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
michael@0	2474	u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
michael@0	2475	u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
michael@0	2476	u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
michael@0	2477	u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
michael@0	2478	u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
michael@0	2479	u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
michael@0	2480	u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
michael@0	2481	u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
michael@0	2482	u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
michael@0	2483	u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
michael@0	2484	u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
michael@0	2485
michael@0	2486	v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
michael@0	2487	v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
michael@0	2488	v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
michael@0	2489	v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
michael@0	2490	v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
michael@0	2491	v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
michael@0	2492	v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
michael@0	2493	v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
michael@0	2494	v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
michael@0	2495	v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
michael@0	2496	v[10] = _mm_cmplt_epi32(u[10], kZero);
michael@0	2497	v[11] = _mm_cmplt_epi32(u[11], kZero);
michael@0	2498	v[12] = _mm_cmplt_epi32(u[12], kZero);
michael@0	2499	v[13] = _mm_cmplt_epi32(u[13], kZero);
michael@0	2500	v[14] = _mm_cmplt_epi32(u[14], kZero);
michael@0	2501	v[15] = _mm_cmplt_epi32(u[15], kZero);
michael@0	2502
michael@0	2503	u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
michael@0	2504	u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
michael@0	2505	u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
michael@0	2506	u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
michael@0	2507	u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
michael@0	2508	u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
michael@0	2509	u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
michael@0	2510	u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
michael@0	2511	u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
michael@0	2512	u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
michael@0	2513	u[10] = _mm_sub_epi32(u[10], v[10]);
michael@0	2514	u[11] = _mm_sub_epi32(u[11], v[11]);
michael@0	2515	u[12] = _mm_sub_epi32(u[12], v[12]);
michael@0	2516	u[13] = _mm_sub_epi32(u[13], v[13]);
michael@0	2517	u[14] = _mm_sub_epi32(u[14], v[14]);
michael@0	2518	u[15] = _mm_sub_epi32(u[15], v[15]);
michael@0	2519
michael@0	2520	v[0] = _mm_add_epi32(u[0], K32One);
michael@0	2521	v[1] = _mm_add_epi32(u[1], K32One);
michael@0	2522	v[2] = _mm_add_epi32(u[2], K32One);
michael@0	2523	v[3] = _mm_add_epi32(u[3], K32One);
michael@0	2524	v[4] = _mm_add_epi32(u[4], K32One);
michael@0	2525	v[5] = _mm_add_epi32(u[5], K32One);
michael@0	2526	v[6] = _mm_add_epi32(u[6], K32One);
michael@0	2527	v[7] = _mm_add_epi32(u[7], K32One);
michael@0	2528	v[8] = _mm_add_epi32(u[8], K32One);
michael@0	2529	v[9] = _mm_add_epi32(u[9], K32One);
michael@0	2530	v[10] = _mm_add_epi32(u[10], K32One);
michael@0	2531	v[11] = _mm_add_epi32(u[11], K32One);
michael@0	2532	v[12] = _mm_add_epi32(u[12], K32One);
michael@0	2533	v[13] = _mm_add_epi32(u[13], K32One);
michael@0	2534	v[14] = _mm_add_epi32(u[14], K32One);
michael@0	2535	v[15] = _mm_add_epi32(u[15], K32One);
michael@0	2536
michael@0	2537	u[0] = _mm_srai_epi32(v[0], 2);
michael@0	2538	u[1] = _mm_srai_epi32(v[1], 2);
michael@0	2539	u[2] = _mm_srai_epi32(v[2], 2);
michael@0	2540	u[3] = _mm_srai_epi32(v[3], 2);
michael@0	2541	u[4] = _mm_srai_epi32(v[4], 2);
michael@0	2542	u[5] = _mm_srai_epi32(v[5], 2);
michael@0	2543	u[6] = _mm_srai_epi32(v[6], 2);
michael@0	2544	u[7] = _mm_srai_epi32(v[7], 2);
michael@0	2545	u[8] = _mm_srai_epi32(v[8], 2);
michael@0	2546	u[9] = _mm_srai_epi32(v[9], 2);
michael@0	2547	u[10] = _mm_srai_epi32(v[10], 2);
michael@0	2548	u[11] = _mm_srai_epi32(v[11], 2);
michael@0	2549	u[12] = _mm_srai_epi32(v[12], 2);
michael@0	2550	u[13] = _mm_srai_epi32(v[13], 2);
michael@0	2551	u[14] = _mm_srai_epi32(v[14], 2);
michael@0	2552	u[15] = _mm_srai_epi32(v[15], 2);
michael@0	2553
michael@0	2554	out[ 5] = _mm_packs_epi32(u[0], u[1]);
michael@0	2555	out[21] = _mm_packs_epi32(u[2], u[3]);
michael@0	2556	out[13] = _mm_packs_epi32(u[4], u[5]);
michael@0	2557	out[29] = _mm_packs_epi32(u[6], u[7]);
michael@0	2558	out[ 3] = _mm_packs_epi32(u[8], u[9]);
michael@0	2559	out[19] = _mm_packs_epi32(u[10], u[11]);
michael@0	2560	out[11] = _mm_packs_epi32(u[12], u[13]);
michael@0	2561	out[27] = _mm_packs_epi32(u[14], u[15]);
michael@0	2562	}
michael@0	2563	}
michael@0	2564	#endif
michael@0	2565	// Transpose the results, do it as four 8x8 transposes.
michael@0	2566	{
michael@0	2567	int transpose_block;
michael@0	2568	int16_t *output;
michael@0	2569	if (0 == pass) {
michael@0	2570	output = &intermediate[column_start * 32];
michael@0	2571	} else {
michael@0	2572	output = &output_org[column_start * 32];
michael@0	2573	}
michael@0	2574	for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
michael@0	2575	__m128i *this_out = &out[8 * transpose_block];
michael@0	2576	// 00 01 02 03 04 05 06 07
michael@0	2577	// 10 11 12 13 14 15 16 17
michael@0	2578	// 20 21 22 23 24 25 26 27
michael@0	2579	// 30 31 32 33 34 35 36 37
michael@0	2580	// 40 41 42 43 44 45 46 47
michael@0	2581	// 50 51 52 53 54 55 56 57
michael@0	2582	// 60 61 62 63 64 65 66 67
michael@0	2583	// 70 71 72 73 74 75 76 77
michael@0	2584	const __m128i tr0_0 = _mm_unpacklo_epi16(this_out[0], this_out[1]);
michael@0	2585	const __m128i tr0_1 = _mm_unpacklo_epi16(this_out[2], this_out[3]);
michael@0	2586	const __m128i tr0_2 = _mm_unpackhi_epi16(this_out[0], this_out[1]);
michael@0	2587	const __m128i tr0_3 = _mm_unpackhi_epi16(this_out[2], this_out[3]);
michael@0	2588	const __m128i tr0_4 = _mm_unpacklo_epi16(this_out[4], this_out[5]);
michael@0	2589	const __m128i tr0_5 = _mm_unpacklo_epi16(this_out[6], this_out[7]);
michael@0	2590	const __m128i tr0_6 = _mm_unpackhi_epi16(this_out[4], this_out[5]);
michael@0	2591	const __m128i tr0_7 = _mm_unpackhi_epi16(this_out[6], this_out[7]);
michael@0	2592	// 00 10 01 11 02 12 03 13
michael@0	2593	// 20 30 21 31 22 32 23 33
michael@0	2594	// 04 14 05 15 06 16 07 17
michael@0	2595	// 24 34 25 35 26 36 27 37
michael@0	2596	// 40 50 41 51 42 52 43 53
michael@0	2597	// 60 70 61 71 62 72 63 73
michael@0	2598	// 54 54 55 55 56 56 57 57
michael@0	2599	// 64 74 65 75 66 76 67 77
michael@0	2600	const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
michael@0	2601	const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
michael@0	2602	const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
michael@0	2603	const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
michael@0	2604	const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
michael@0	2605	const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
michael@0	2606	const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
michael@0	2607	const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
michael@0	2608	// 00 10 20 30 01 11 21 31
michael@0	2609	// 40 50 60 70 41 51 61 71
michael@0	2610	// 02 12 22 32 03 13 23 33
michael@0	2611	// 42 52 62 72 43 53 63 73
michael@0	2612	// 04 14 24 34 05 15 21 36
michael@0	2613	// 44 54 64 74 45 55 61 76
michael@0	2614	// 06 16 26 36 07 17 27 37
michael@0	2615	// 46 56 66 76 47 57 67 77
michael@0	2616	__m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
michael@0	2617	__m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
michael@0	2618	__m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
michael@0	2619	__m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
michael@0	2620	__m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
michael@0	2621	__m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
michael@0	2622	__m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
michael@0	2623	__m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
michael@0	2624	// 00 10 20 30 40 50 60 70
michael@0	2625	// 01 11 21 31 41 51 61 71
michael@0	2626	// 02 12 22 32 42 52 62 72
michael@0	2627	// 03 13 23 33 43 53 63 73
michael@0	2628	// 04 14 24 34 44 54 64 74
michael@0	2629	// 05 15 25 35 45 55 65 75
michael@0	2630	// 06 16 26 36 46 56 66 76
michael@0	2631	// 07 17 27 37 47 57 67 77
michael@0	2632	if (0 == pass) {
michael@0	2633	// output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
michael@0	2634	// TODO(cd): see quality impact of only doing
michael@0	2635	// output[j] = (output[j] + 1) >> 2;
michael@0	2636	// which would remove the code between here ...
michael@0	2637	__m128i tr2_0_0 = _mm_cmpgt_epi16(tr2_0, kZero);
michael@0	2638	__m128i tr2_1_0 = _mm_cmpgt_epi16(tr2_1, kZero);
michael@0	2639	__m128i tr2_2_0 = _mm_cmpgt_epi16(tr2_2, kZero);
michael@0	2640	__m128i tr2_3_0 = _mm_cmpgt_epi16(tr2_3, kZero);
michael@0	2641	__m128i tr2_4_0 = _mm_cmpgt_epi16(tr2_4, kZero);
michael@0	2642	__m128i tr2_5_0 = _mm_cmpgt_epi16(tr2_5, kZero);
michael@0	2643	__m128i tr2_6_0 = _mm_cmpgt_epi16(tr2_6, kZero);
michael@0	2644	__m128i tr2_7_0 = _mm_cmpgt_epi16(tr2_7, kZero);
michael@0	2645	tr2_0 = _mm_sub_epi16(tr2_0, tr2_0_0);
michael@0	2646	tr2_1 = _mm_sub_epi16(tr2_1, tr2_1_0);
michael@0	2647	tr2_2 = _mm_sub_epi16(tr2_2, tr2_2_0);
michael@0	2648	tr2_3 = _mm_sub_epi16(tr2_3, tr2_3_0);
michael@0	2649	tr2_4 = _mm_sub_epi16(tr2_4, tr2_4_0);
michael@0	2650	tr2_5 = _mm_sub_epi16(tr2_5, tr2_5_0);
michael@0	2651	tr2_6 = _mm_sub_epi16(tr2_6, tr2_6_0);
michael@0	2652	tr2_7 = _mm_sub_epi16(tr2_7, tr2_7_0);
michael@0	2653	// ... and here.
michael@0	2654	// PS: also change code in vp9/encoder/vp9_dct.c
michael@0	2655	tr2_0 = _mm_add_epi16(tr2_0, kOne);
michael@0	2656	tr2_1 = _mm_add_epi16(tr2_1, kOne);
michael@0	2657	tr2_2 = _mm_add_epi16(tr2_2, kOne);
michael@0	2658	tr2_3 = _mm_add_epi16(tr2_3, kOne);
michael@0	2659	tr2_4 = _mm_add_epi16(tr2_4, kOne);
michael@0	2660	tr2_5 = _mm_add_epi16(tr2_5, kOne);
michael@0	2661	tr2_6 = _mm_add_epi16(tr2_6, kOne);
michael@0	2662	tr2_7 = _mm_add_epi16(tr2_7, kOne);
michael@0	2663	tr2_0 = _mm_srai_epi16(tr2_0, 2);
michael@0	2664	tr2_1 = _mm_srai_epi16(tr2_1, 2);
michael@0	2665	tr2_2 = _mm_srai_epi16(tr2_2, 2);
michael@0	2666	tr2_3 = _mm_srai_epi16(tr2_3, 2);
michael@0	2667	tr2_4 = _mm_srai_epi16(tr2_4, 2);
michael@0	2668	tr2_5 = _mm_srai_epi16(tr2_5, 2);
michael@0	2669	tr2_6 = _mm_srai_epi16(tr2_6, 2);
michael@0	2670	tr2_7 = _mm_srai_epi16(tr2_7, 2);
michael@0	2671	}
michael@0	2672	// Note: even though all these stores are aligned, using the aligned
michael@0	2673	// intrinsic make the code slightly slower.
michael@0	2674	_mm_storeu_si128((__m128i *)(output + 0 * 32), tr2_0);
michael@0	2675	_mm_storeu_si128((__m128i *)(output + 1 * 32), tr2_1);
michael@0	2676	_mm_storeu_si128((__m128i *)(output + 2 * 32), tr2_2);
michael@0	2677	_mm_storeu_si128((__m128i *)(output + 3 * 32), tr2_3);
michael@0	2678	_mm_storeu_si128((__m128i *)(output + 4 * 32), tr2_4);
michael@0	2679	_mm_storeu_si128((__m128i *)(output + 5 * 32), tr2_5);
michael@0	2680	_mm_storeu_si128((__m128i *)(output + 6 * 32), tr2_6);
michael@0	2681	_mm_storeu_si128((__m128i *)(output + 7 * 32), tr2_7);
michael@0	2682	// Process next 8x8
michael@0	2683	output += 8;
michael@0	2684	}
michael@0	2685	}
michael@0	2686	}
michael@0	2687	}
michael@0	2688	} // NOLINT