• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * jidctfst-neon.c - fast integer IDCT (Arm Neon)
3  *
4  * Copyright (C) 2020, Arm Limited.  All Rights Reserved.
5  *
6  * This software is provided 'as-is', without any express or implied
7  * warranty.  In no event will the authors be held liable for any damages
8  * arising from the use of this software.
9  *
10  * Permission is granted to anyone to use this software for any purpose,
11  * including commercial applications, and to alter it and redistribute it
12  * freely, subject to the following restrictions:
13  *
14  * 1. The origin of this software must not be misrepresented; you must not
15  *    claim that you wrote the original software. If you use this software
16  *    in a product, an acknowledgment in the product documentation would be
17  *    appreciated but is not required.
18  * 2. Altered source versions must be plainly marked as such, and must not be
19  *    misrepresented as being the original software.
20  * 3. This notice may not be removed or altered from any source distribution.
21  */
22 
23 #define JPEG_INTERNALS
24 #include "../../jinclude.h"
25 #include "../../jpeglib.h"
26 #include "../../jsimd.h"
27 #include "../../jdct.h"
28 #include "../../jsimddct.h"
29 #include "../jsimd.h"
30 #include "align.h"
31 
32 #include <arm_neon.h>
33 
34 
35 /* jsimd_idct_ifast_neon() performs dequantization and a fast, not so accurate
36  * inverse DCT (Discrete Cosine Transform) on one block of coefficients.  It
37  * uses the same calculations and produces exactly the same output as IJG's
38  * original jpeg_idct_ifast() function, which can be found in jidctfst.c.
39  *
40  * Scaled integer constants are used to avoid floating-point arithmetic:
41  *    0.082392200 =  2688 * 2^-15
42  *    0.414213562 = 13568 * 2^-15
43  *    0.847759065 = 27776 * 2^-15
44  *    0.613125930 = 20096 * 2^-15
45  *
46  * See jidctfst.c for further details of the IDCT algorithm.  Where possible,
47  * the variable names and comments here in jsimd_idct_ifast_neon() match up
48  * with those in jpeg_idct_ifast().
49  */
50 
51 #define PASS1_BITS  2
52 
53 #define F_0_082  2688
54 #define F_0_414  13568
55 #define F_0_847  27776
56 #define F_0_613  20096
57 
58 
59 ALIGN(16) static const int16_t jsimd_idct_ifast_neon_consts[] = {
60   F_0_082, F_0_414, F_0_847, F_0_613
61 };
62 
jsimd_idct_ifast_neon(void * dct_table,JCOEFPTR coef_block,JSAMPARRAY output_buf,JDIMENSION output_col)63 void jsimd_idct_ifast_neon(void *dct_table, JCOEFPTR coef_block,
64                            JSAMPARRAY output_buf, JDIMENSION output_col)
65 {
66   IFAST_MULT_TYPE *quantptr = dct_table;
67 
68   /* Load DCT coefficients. */
69   int16x8_t row0 = vld1q_s16(coef_block + 0 * DCTSIZE);
70   int16x8_t row1 = vld1q_s16(coef_block + 1 * DCTSIZE);
71   int16x8_t row2 = vld1q_s16(coef_block + 2 * DCTSIZE);
72   int16x8_t row3 = vld1q_s16(coef_block + 3 * DCTSIZE);
73   int16x8_t row4 = vld1q_s16(coef_block + 4 * DCTSIZE);
74   int16x8_t row5 = vld1q_s16(coef_block + 5 * DCTSIZE);
75   int16x8_t row6 = vld1q_s16(coef_block + 6 * DCTSIZE);
76   int16x8_t row7 = vld1q_s16(coef_block + 7 * DCTSIZE);
77 
78   /* Load quantization table values for DC coefficients. */
79   int16x8_t quant_row0 = vld1q_s16(quantptr + 0 * DCTSIZE);
80   /* Dequantize DC coefficients. */
81   row0 = vmulq_s16(row0, quant_row0);
82 
83   /* Construct bitmap to test if all AC coefficients are 0. */
84   int16x8_t bitmap = vorrq_s16(row1, row2);
85   bitmap = vorrq_s16(bitmap, row3);
86   bitmap = vorrq_s16(bitmap, row4);
87   bitmap = vorrq_s16(bitmap, row5);
88   bitmap = vorrq_s16(bitmap, row6);
89   bitmap = vorrq_s16(bitmap, row7);
90 
91   int64_t left_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 0);
92   int64_t right_ac_bitmap = vgetq_lane_s64(vreinterpretq_s64_s16(bitmap), 1);
93 
94   /* Load IDCT conversion constants. */
95   const int16x4_t consts = vld1_s16(jsimd_idct_ifast_neon_consts);
96 
97   if (left_ac_bitmap == 0 && right_ac_bitmap == 0) {
98     /* All AC coefficients are zero.
99      * Compute DC values and duplicate into vectors.
100      */
101     int16x8_t dcval = row0;
102     row1 = dcval;
103     row2 = dcval;
104     row3 = dcval;
105     row4 = dcval;
106     row5 = dcval;
107     row6 = dcval;
108     row7 = dcval;
109   } else if (left_ac_bitmap == 0) {
110     /* AC coefficients are zero for columns 0, 1, 2, and 3.
111      * Use DC values for these columns.
112      */
113     int16x4_t dcval = vget_low_s16(row0);
114 
115     /* Commence regular fast IDCT computation for columns 4, 5, 6, and 7. */
116 
117     /* Load quantization table. */
118     int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE + 4);
119     int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE + 4);
120     int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE + 4);
121     int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE + 4);
122     int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE + 4);
123     int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE + 4);
124     int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE + 4);
125 
126     /* Even part: dequantize DCT coefficients. */
127     int16x4_t tmp0 = vget_high_s16(row0);
128     int16x4_t tmp1 = vmul_s16(vget_high_s16(row2), quant_row2);
129     int16x4_t tmp2 = vmul_s16(vget_high_s16(row4), quant_row4);
130     int16x4_t tmp3 = vmul_s16(vget_high_s16(row6), quant_row6);
131 
132     int16x4_t tmp10 = vadd_s16(tmp0, tmp2);   /* phase 3 */
133     int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
134 
135     int16x4_t tmp13 = vadd_s16(tmp1, tmp3);   /* phases 5-3 */
136     int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
137     int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
138     tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
139     tmp12 = vsub_s16(tmp12, tmp13);
140 
141     tmp0 = vadd_s16(tmp10, tmp13);            /* phase 2 */
142     tmp3 = vsub_s16(tmp10, tmp13);
143     tmp1 = vadd_s16(tmp11, tmp12);
144     tmp2 = vsub_s16(tmp11, tmp12);
145 
146     /* Odd part: dequantize DCT coefficients. */
147     int16x4_t tmp4 = vmul_s16(vget_high_s16(row1), quant_row1);
148     int16x4_t tmp5 = vmul_s16(vget_high_s16(row3), quant_row3);
149     int16x4_t tmp6 = vmul_s16(vget_high_s16(row5), quant_row5);
150     int16x4_t tmp7 = vmul_s16(vget_high_s16(row7), quant_row7);
151 
152     int16x4_t z13 = vadd_s16(tmp6, tmp5);     /* phase 6 */
153     int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
154     int16x4_t z11 = vadd_s16(tmp4, tmp7);
155     int16x4_t z12 = vsub_s16(tmp4, tmp7);
156 
157     tmp7 = vadd_s16(z11, z13);                /* phase 5 */
158     int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
159     tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
160     tmp11 = vadd_s16(tmp11, z11_sub_z13);
161 
162     int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
163     int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
164     z5 = vadd_s16(z5, z10_add_z12);
165     tmp10 = vqdmulh_lane_s16(z12, consts, 0);
166     tmp10 = vadd_s16(tmp10, z12);
167     tmp10 = vsub_s16(tmp10, z5);
168     tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
169     tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
170     tmp12 = vadd_s16(tmp12, z5);
171 
172     tmp6 = vsub_s16(tmp12, tmp7);             /* phase 2 */
173     tmp5 = vsub_s16(tmp11, tmp6);
174     tmp4 = vadd_s16(tmp10, tmp5);
175 
176     row0 = vcombine_s16(dcval, vadd_s16(tmp0, tmp7));
177     row7 = vcombine_s16(dcval, vsub_s16(tmp0, tmp7));
178     row1 = vcombine_s16(dcval, vadd_s16(tmp1, tmp6));
179     row6 = vcombine_s16(dcval, vsub_s16(tmp1, tmp6));
180     row2 = vcombine_s16(dcval, vadd_s16(tmp2, tmp5));
181     row5 = vcombine_s16(dcval, vsub_s16(tmp2, tmp5));
182     row4 = vcombine_s16(dcval, vadd_s16(tmp3, tmp4));
183     row3 = vcombine_s16(dcval, vsub_s16(tmp3, tmp4));
184   } else if (right_ac_bitmap == 0) {
185     /* AC coefficients are zero for columns 4, 5, 6, and 7.
186      * Use DC values for these columns.
187      */
188     int16x4_t dcval = vget_high_s16(row0);
189 
190     /* Commence regular fast IDCT computation for columns 0, 1, 2, and 3. */
191 
192     /* Load quantization table. */
193     int16x4_t quant_row1 = vld1_s16(quantptr + 1 * DCTSIZE);
194     int16x4_t quant_row2 = vld1_s16(quantptr + 2 * DCTSIZE);
195     int16x4_t quant_row3 = vld1_s16(quantptr + 3 * DCTSIZE);
196     int16x4_t quant_row4 = vld1_s16(quantptr + 4 * DCTSIZE);
197     int16x4_t quant_row5 = vld1_s16(quantptr + 5 * DCTSIZE);
198     int16x4_t quant_row6 = vld1_s16(quantptr + 6 * DCTSIZE);
199     int16x4_t quant_row7 = vld1_s16(quantptr + 7 * DCTSIZE);
200 
201     /* Even part: dequantize DCT coefficients. */
202     int16x4_t tmp0 = vget_low_s16(row0);
203     int16x4_t tmp1 = vmul_s16(vget_low_s16(row2), quant_row2);
204     int16x4_t tmp2 = vmul_s16(vget_low_s16(row4), quant_row4);
205     int16x4_t tmp3 = vmul_s16(vget_low_s16(row6), quant_row6);
206 
207     int16x4_t tmp10 = vadd_s16(tmp0, tmp2);   /* phase 3 */
208     int16x4_t tmp11 = vsub_s16(tmp0, tmp2);
209 
210     int16x4_t tmp13 = vadd_s16(tmp1, tmp3);   /* phases 5-3 */
211     int16x4_t tmp1_sub_tmp3 = vsub_s16(tmp1, tmp3);
212     int16x4_t tmp12 = vqdmulh_lane_s16(tmp1_sub_tmp3, consts, 1);
213     tmp12 = vadd_s16(tmp12, tmp1_sub_tmp3);
214     tmp12 = vsub_s16(tmp12, tmp13);
215 
216     tmp0 = vadd_s16(tmp10, tmp13);            /* phase 2 */
217     tmp3 = vsub_s16(tmp10, tmp13);
218     tmp1 = vadd_s16(tmp11, tmp12);
219     tmp2 = vsub_s16(tmp11, tmp12);
220 
221     /* Odd part: dequantize DCT coefficients. */
222     int16x4_t tmp4 = vmul_s16(vget_low_s16(row1), quant_row1);
223     int16x4_t tmp5 = vmul_s16(vget_low_s16(row3), quant_row3);
224     int16x4_t tmp6 = vmul_s16(vget_low_s16(row5), quant_row5);
225     int16x4_t tmp7 = vmul_s16(vget_low_s16(row7), quant_row7);
226 
227     int16x4_t z13 = vadd_s16(tmp6, tmp5);     /* phase 6 */
228     int16x4_t neg_z10 = vsub_s16(tmp5, tmp6);
229     int16x4_t z11 = vadd_s16(tmp4, tmp7);
230     int16x4_t z12 = vsub_s16(tmp4, tmp7);
231 
232     tmp7 = vadd_s16(z11, z13);                /* phase 5 */
233     int16x4_t z11_sub_z13 = vsub_s16(z11, z13);
234     tmp11 = vqdmulh_lane_s16(z11_sub_z13, consts, 1);
235     tmp11 = vadd_s16(tmp11, z11_sub_z13);
236 
237     int16x4_t z10_add_z12 = vsub_s16(z12, neg_z10);
238     int16x4_t z5 = vqdmulh_lane_s16(z10_add_z12, consts, 2);
239     z5 = vadd_s16(z5, z10_add_z12);
240     tmp10 = vqdmulh_lane_s16(z12, consts, 0);
241     tmp10 = vadd_s16(tmp10, z12);
242     tmp10 = vsub_s16(tmp10, z5);
243     tmp12 = vqdmulh_lane_s16(neg_z10, consts, 3);
244     tmp12 = vadd_s16(tmp12, vadd_s16(neg_z10, neg_z10));
245     tmp12 = vadd_s16(tmp12, z5);
246 
247     tmp6 = vsub_s16(tmp12, tmp7);             /* phase 2 */
248     tmp5 = vsub_s16(tmp11, tmp6);
249     tmp4 = vadd_s16(tmp10, tmp5);
250 
251     row0 = vcombine_s16(vadd_s16(tmp0, tmp7), dcval);
252     row7 = vcombine_s16(vsub_s16(tmp0, tmp7), dcval);
253     row1 = vcombine_s16(vadd_s16(tmp1, tmp6), dcval);
254     row6 = vcombine_s16(vsub_s16(tmp1, tmp6), dcval);
255     row2 = vcombine_s16(vadd_s16(tmp2, tmp5), dcval);
256     row5 = vcombine_s16(vsub_s16(tmp2, tmp5), dcval);
257     row4 = vcombine_s16(vadd_s16(tmp3, tmp4), dcval);
258     row3 = vcombine_s16(vsub_s16(tmp3, tmp4), dcval);
259   } else {
260     /* Some AC coefficients are non-zero; full IDCT calculation required. */
261 
262     /* Load quantization table. */
263     int16x8_t quant_row1 = vld1q_s16(quantptr + 1 * DCTSIZE);
264     int16x8_t quant_row2 = vld1q_s16(quantptr + 2 * DCTSIZE);
265     int16x8_t quant_row3 = vld1q_s16(quantptr + 3 * DCTSIZE);
266     int16x8_t quant_row4 = vld1q_s16(quantptr + 4 * DCTSIZE);
267     int16x8_t quant_row5 = vld1q_s16(quantptr + 5 * DCTSIZE);
268     int16x8_t quant_row6 = vld1q_s16(quantptr + 6 * DCTSIZE);
269     int16x8_t quant_row7 = vld1q_s16(quantptr + 7 * DCTSIZE);
270 
271     /* Even part: dequantize DCT coefficients. */
272     int16x8_t tmp0 = row0;
273     int16x8_t tmp1 = vmulq_s16(row2, quant_row2);
274     int16x8_t tmp2 = vmulq_s16(row4, quant_row4);
275     int16x8_t tmp3 = vmulq_s16(row6, quant_row6);
276 
277     int16x8_t tmp10 = vaddq_s16(tmp0, tmp2);   /* phase 3 */
278     int16x8_t tmp11 = vsubq_s16(tmp0, tmp2);
279 
280     int16x8_t tmp13 = vaddq_s16(tmp1, tmp3);   /* phases 5-3 */
281     int16x8_t tmp1_sub_tmp3 = vsubq_s16(tmp1, tmp3);
282     int16x8_t tmp12 = vqdmulhq_lane_s16(tmp1_sub_tmp3, consts, 1);
283     tmp12 = vaddq_s16(tmp12, tmp1_sub_tmp3);
284     tmp12 = vsubq_s16(tmp12, tmp13);
285 
286     tmp0 = vaddq_s16(tmp10, tmp13);            /* phase 2 */
287     tmp3 = vsubq_s16(tmp10, tmp13);
288     tmp1 = vaddq_s16(tmp11, tmp12);
289     tmp2 = vsubq_s16(tmp11, tmp12);
290 
291     /* Odd part: dequantize DCT coefficients. */
292     int16x8_t tmp4 = vmulq_s16(row1, quant_row1);
293     int16x8_t tmp5 = vmulq_s16(row3, quant_row3);
294     int16x8_t tmp6 = vmulq_s16(row5, quant_row5);
295     int16x8_t tmp7 = vmulq_s16(row7, quant_row7);
296 
297     int16x8_t z13 = vaddq_s16(tmp6, tmp5);     /* phase 6 */
298     int16x8_t neg_z10 = vsubq_s16(tmp5, tmp6);
299     int16x8_t z11 = vaddq_s16(tmp4, tmp7);
300     int16x8_t z12 = vsubq_s16(tmp4, tmp7);
301 
302     tmp7 = vaddq_s16(z11, z13);                /* phase 5 */
303     int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
304     tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
305     tmp11 = vaddq_s16(tmp11, z11_sub_z13);
306 
307     int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
308     int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
309     z5 = vaddq_s16(z5, z10_add_z12);
310     tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
311     tmp10 = vaddq_s16(tmp10, z12);
312     tmp10 = vsubq_s16(tmp10, z5);
313     tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
314     tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
315     tmp12 = vaddq_s16(tmp12, z5);
316 
317     tmp6 = vsubq_s16(tmp12, tmp7);             /* phase 2 */
318     tmp5 = vsubq_s16(tmp11, tmp6);
319     tmp4 = vaddq_s16(tmp10, tmp5);
320 
321     row0 = vaddq_s16(tmp0, tmp7);
322     row7 = vsubq_s16(tmp0, tmp7);
323     row1 = vaddq_s16(tmp1, tmp6);
324     row6 = vsubq_s16(tmp1, tmp6);
325     row2 = vaddq_s16(tmp2, tmp5);
326     row5 = vsubq_s16(tmp2, tmp5);
327     row4 = vaddq_s16(tmp3, tmp4);
328     row3 = vsubq_s16(tmp3, tmp4);
329   }
330 
331   /* Transpose rows to work on columns in pass 2. */
332   int16x8x2_t rows_01 = vtrnq_s16(row0, row1);
333   int16x8x2_t rows_23 = vtrnq_s16(row2, row3);
334   int16x8x2_t rows_45 = vtrnq_s16(row4, row5);
335   int16x8x2_t rows_67 = vtrnq_s16(row6, row7);
336 
337   int32x4x2_t rows_0145_l = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[0]),
338                                       vreinterpretq_s32_s16(rows_45.val[0]));
339   int32x4x2_t rows_0145_h = vtrnq_s32(vreinterpretq_s32_s16(rows_01.val[1]),
340                                       vreinterpretq_s32_s16(rows_45.val[1]));
341   int32x4x2_t rows_2367_l = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[0]),
342                                       vreinterpretq_s32_s16(rows_67.val[0]));
343   int32x4x2_t rows_2367_h = vtrnq_s32(vreinterpretq_s32_s16(rows_23.val[1]),
344                                       vreinterpretq_s32_s16(rows_67.val[1]));
345 
346   int32x4x2_t cols_04 = vzipq_s32(rows_0145_l.val[0], rows_2367_l.val[0]);
347   int32x4x2_t cols_15 = vzipq_s32(rows_0145_h.val[0], rows_2367_h.val[0]);
348   int32x4x2_t cols_26 = vzipq_s32(rows_0145_l.val[1], rows_2367_l.val[1]);
349   int32x4x2_t cols_37 = vzipq_s32(rows_0145_h.val[1], rows_2367_h.val[1]);
350 
351   int16x8_t col0 = vreinterpretq_s16_s32(cols_04.val[0]);
352   int16x8_t col1 = vreinterpretq_s16_s32(cols_15.val[0]);
353   int16x8_t col2 = vreinterpretq_s16_s32(cols_26.val[0]);
354   int16x8_t col3 = vreinterpretq_s16_s32(cols_37.val[0]);
355   int16x8_t col4 = vreinterpretq_s16_s32(cols_04.val[1]);
356   int16x8_t col5 = vreinterpretq_s16_s32(cols_15.val[1]);
357   int16x8_t col6 = vreinterpretq_s16_s32(cols_26.val[1]);
358   int16x8_t col7 = vreinterpretq_s16_s32(cols_37.val[1]);
359 
360   /* 1-D IDCT, pass 2 */
361 
362   /* Even part */
363   int16x8_t tmp10 = vaddq_s16(col0, col4);
364   int16x8_t tmp11 = vsubq_s16(col0, col4);
365 
366   int16x8_t tmp13 = vaddq_s16(col2, col6);
367   int16x8_t col2_sub_col6 = vsubq_s16(col2, col6);
368   int16x8_t tmp12 = vqdmulhq_lane_s16(col2_sub_col6, consts, 1);
369   tmp12 = vaddq_s16(tmp12, col2_sub_col6);
370   tmp12 = vsubq_s16(tmp12, tmp13);
371 
372   int16x8_t tmp0 = vaddq_s16(tmp10, tmp13);
373   int16x8_t tmp3 = vsubq_s16(tmp10, tmp13);
374   int16x8_t tmp1 = vaddq_s16(tmp11, tmp12);
375   int16x8_t tmp2 = vsubq_s16(tmp11, tmp12);
376 
377   /* Odd part */
378   int16x8_t z13 = vaddq_s16(col5, col3);
379   int16x8_t neg_z10 = vsubq_s16(col3, col5);
380   int16x8_t z11 = vaddq_s16(col1, col7);
381   int16x8_t z12 = vsubq_s16(col1, col7);
382 
383   int16x8_t tmp7 = vaddq_s16(z11, z13);      /* phase 5 */
384   int16x8_t z11_sub_z13 = vsubq_s16(z11, z13);
385   tmp11 = vqdmulhq_lane_s16(z11_sub_z13, consts, 1);
386   tmp11 = vaddq_s16(tmp11, z11_sub_z13);
387 
388   int16x8_t z10_add_z12 = vsubq_s16(z12, neg_z10);
389   int16x8_t z5 = vqdmulhq_lane_s16(z10_add_z12, consts, 2);
390   z5 = vaddq_s16(z5, z10_add_z12);
391   tmp10 = vqdmulhq_lane_s16(z12, consts, 0);
392   tmp10 = vaddq_s16(tmp10, z12);
393   tmp10 = vsubq_s16(tmp10, z5);
394   tmp12 = vqdmulhq_lane_s16(neg_z10, consts, 3);
395   tmp12 = vaddq_s16(tmp12, vaddq_s16(neg_z10, neg_z10));
396   tmp12 = vaddq_s16(tmp12, z5);
397 
398   int16x8_t tmp6 = vsubq_s16(tmp12, tmp7);   /* phase 2 */
399   int16x8_t tmp5 = vsubq_s16(tmp11, tmp6);
400   int16x8_t tmp4 = vaddq_s16(tmp10, tmp5);
401 
402   col0 = vaddq_s16(tmp0, tmp7);
403   col7 = vsubq_s16(tmp0, tmp7);
404   col1 = vaddq_s16(tmp1, tmp6);
405   col6 = vsubq_s16(tmp1, tmp6);
406   col2 = vaddq_s16(tmp2, tmp5);
407   col5 = vsubq_s16(tmp2, tmp5);
408   col4 = vaddq_s16(tmp3, tmp4);
409   col3 = vsubq_s16(tmp3, tmp4);
410 
411   /* Scale down by a factor of 8, narrowing to 8-bit. */
412   int8x16_t cols_01_s8 = vcombine_s8(vqshrn_n_s16(col0, PASS1_BITS + 3),
413                                      vqshrn_n_s16(col1, PASS1_BITS + 3));
414   int8x16_t cols_45_s8 = vcombine_s8(vqshrn_n_s16(col4, PASS1_BITS + 3),
415                                      vqshrn_n_s16(col5, PASS1_BITS + 3));
416   int8x16_t cols_23_s8 = vcombine_s8(vqshrn_n_s16(col2, PASS1_BITS + 3),
417                                      vqshrn_n_s16(col3, PASS1_BITS + 3));
418   int8x16_t cols_67_s8 = vcombine_s8(vqshrn_n_s16(col6, PASS1_BITS + 3),
419                                      vqshrn_n_s16(col7, PASS1_BITS + 3));
420   /* Clamp to range [0-255]. */
421   uint8x16_t cols_01 =
422     vreinterpretq_u8_s8
423       (vaddq_s8(cols_01_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
424   uint8x16_t cols_45 =
425     vreinterpretq_u8_s8
426       (vaddq_s8(cols_45_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
427   uint8x16_t cols_23 =
428     vreinterpretq_u8_s8
429       (vaddq_s8(cols_23_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
430   uint8x16_t cols_67 =
431     vreinterpretq_u8_s8
432       (vaddq_s8(cols_67_s8, vreinterpretq_s8_u8(vdupq_n_u8(CENTERJSAMPLE))));
433 
434   /* Transpose block to prepare for store. */
435   uint32x4x2_t cols_0415 = vzipq_u32(vreinterpretq_u32_u8(cols_01),
436                                      vreinterpretq_u32_u8(cols_45));
437   uint32x4x2_t cols_2637 = vzipq_u32(vreinterpretq_u32_u8(cols_23),
438                                      vreinterpretq_u32_u8(cols_67));
439 
440   uint8x16x2_t cols_0145 = vtrnq_u8(vreinterpretq_u8_u32(cols_0415.val[0]),
441                                     vreinterpretq_u8_u32(cols_0415.val[1]));
442   uint8x16x2_t cols_2367 = vtrnq_u8(vreinterpretq_u8_u32(cols_2637.val[0]),
443                                     vreinterpretq_u8_u32(cols_2637.val[1]));
444   uint16x8x2_t rows_0426 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[0]),
445                                      vreinterpretq_u16_u8(cols_2367.val[0]));
446   uint16x8x2_t rows_1537 = vtrnq_u16(vreinterpretq_u16_u8(cols_0145.val[1]),
447                                      vreinterpretq_u16_u8(cols_2367.val[1]));
448 
449   uint8x16_t rows_04 = vreinterpretq_u8_u16(rows_0426.val[0]);
450   uint8x16_t rows_15 = vreinterpretq_u8_u16(rows_1537.val[0]);
451   uint8x16_t rows_26 = vreinterpretq_u8_u16(rows_0426.val[1]);
452   uint8x16_t rows_37 = vreinterpretq_u8_u16(rows_1537.val[1]);
453 
454   JSAMPROW outptr0 = output_buf[0] + output_col;
455   JSAMPROW outptr1 = output_buf[1] + output_col;
456   JSAMPROW outptr2 = output_buf[2] + output_col;
457   JSAMPROW outptr3 = output_buf[3] + output_col;
458   JSAMPROW outptr4 = output_buf[4] + output_col;
459   JSAMPROW outptr5 = output_buf[5] + output_col;
460   JSAMPROW outptr6 = output_buf[6] + output_col;
461   JSAMPROW outptr7 = output_buf[7] + output_col;
462 
463   /* Store DCT block to memory. */
464   vst1q_lane_u64((uint64_t *)outptr0, vreinterpretq_u64_u8(rows_04), 0);
465   vst1q_lane_u64((uint64_t *)outptr1, vreinterpretq_u64_u8(rows_15), 0);
466   vst1q_lane_u64((uint64_t *)outptr2, vreinterpretq_u64_u8(rows_26), 0);
467   vst1q_lane_u64((uint64_t *)outptr3, vreinterpretq_u64_u8(rows_37), 0);
468   vst1q_lane_u64((uint64_t *)outptr4, vreinterpretq_u64_u8(rows_04), 1);
469   vst1q_lane_u64((uint64_t *)outptr5, vreinterpretq_u64_u8(rows_15), 1);
470   vst1q_lane_u64((uint64_t *)outptr6, vreinterpretq_u64_u8(rows_26), 1);
471   vst1q_lane_u64((uint64_t *)outptr7, vreinterpretq_u64_u8(rows_37), 1);
472 }
473