summaryrefslogtreecommitdiffstats
path: root/media/libjpeg/simd/arm/aarch64/jchuff-neon.c
blob: 607a116070c9883a818f7faaf4425c8b7978fb82 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
/*
 * jchuff-neon.c - Huffman entropy encoding (64-bit Arm Neon)
 *
 * Copyright (C) 2020-2021, Arm Limited.  All Rights Reserved.
 * Copyright (C) 2020, 2022, D. R. Commander.  All Rights Reserved.
 *
 * This software is provided 'as-is', without any express or implied
 * warranty.  In no event will the authors be held liable for any damages
 * arising from the use of this software.
 *
 * Permission is granted to anyone to use this software for any purpose,
 * including commercial applications, and to alter it and redistribute it
 * freely, subject to the following restrictions:
 *
 * 1. The origin of this software must not be misrepresented; you must not
 *    claim that you wrote the original software. If you use this software
 *    in a product, an acknowledgment in the product documentation would be
 *    appreciated but is not required.
 * 2. Altered source versions must be plainly marked as such, and must not be
 *    misrepresented as being the original software.
 * 3. This notice may not be removed or altered from any source distribution.
 *
 * NOTE: All referenced figures are from
 * Recommendation ITU-T T.81 (1992) | ISO/IEC 10918-1:1994.
 */

#define JPEG_INTERNALS
#include "../../../jinclude.h"
#include "../../../jpeglib.h"
#include "../../../jsimd.h"
#include "../../../jdct.h"
#include "../../../jsimddct.h"
#include "../../jsimd.h"
#include "../align.h"
#include "../jchuff.h"
#include "neon-compat.h"

#include <limits.h>

#include <arm_neon.h>


ALIGN(16) static const uint8_t jsimd_huff_encode_one_block_consts[] = {
    0,   1,   2,   3,  16,  17,  32,  33,
   18,  19,   4,   5,   6,   7,  20,  21,
   34,  35,  48,  49, 255, 255,  50,  51,
   36,  37,  22,  23,   8,   9,  10,  11,
  255, 255,   6,   7,  20,  21,  34,  35,
   48,  49, 255, 255,  50,  51,  36,  37,
   54,  55,  40,  41,  26,  27,  12,  13,
   14,  15,  28,  29,  42,  43,  56,  57,
    6,   7,  20,  21,  34,  35,  48,  49,
   50,  51,  36,  37,  22,  23,   8,   9,
   26,  27,  12,  13, 255, 255,  14,  15,
   28,  29,  42,  43,  56,  57, 255, 255,
   52,  53,  54,  55,  40,  41,  26,  27,
   12,  13, 255, 255,  14,  15,  28,  29,
   26,  27,  40,  41,  42,  43,  28,  29,
   14,  15,  30,  31,  44,  45,  46,  47
};

/* The AArch64 implementation of the FLUSH() macro triggers a UBSan misaligned
 * address warning because the macro sometimes writes a 64-bit value to a
 * non-64-bit-aligned address.  That behavior is technically undefined per
 * the C specification, but it is supported by the AArch64 architecture and
 * compilers.
 */
#if defined(__has_feature)
#if __has_feature(undefined_behavior_sanitizer)
__attribute__((no_sanitize("alignment")))
#endif
#endif
JOCTET *jsimd_huff_encode_one_block_neon(void *state, JOCTET *buffer,
                                         JCOEFPTR block, int last_dc_val,
                                         c_derived_tbl *dctbl,
                                         c_derived_tbl *actbl)
{
  uint16_t block_diff[DCTSIZE2];

  /* Load lookup table indices for rows of zig-zag ordering. */
#ifdef HAVE_VLD1Q_U8_X4
  const uint8x16x4_t idx_rows_0123 =
    vld1q_u8_x4(jsimd_huff_encode_one_block_consts + 0 * DCTSIZE);
  const uint8x16x4_t idx_rows_4567 =
    vld1q_u8_x4(jsimd_huff_encode_one_block_consts + 8 * DCTSIZE);
#else
  /* GCC does not currently support intrinsics vl1dq_<type>_x4(). */
  const uint8x16x4_t idx_rows_0123 = { {
    vld1q_u8(jsimd_huff_encode_one_block_consts + 0 * DCTSIZE),
    vld1q_u8(jsimd_huff_encode_one_block_consts + 2 * DCTSIZE),
    vld1q_u8(jsimd_huff_encode_one_block_consts + 4 * DCTSIZE),
    vld1q_u8(jsimd_huff_encode_one_block_consts + 6 * DCTSIZE)
  } };
  const uint8x16x4_t idx_rows_4567 = { {
    vld1q_u8(jsimd_huff_encode_one_block_consts + 8 * DCTSIZE),
    vld1q_u8(jsimd_huff_encode_one_block_consts + 10 * DCTSIZE),
    vld1q_u8(jsimd_huff_encode_one_block_consts + 12 * DCTSIZE),
    vld1q_u8(jsimd_huff_encode_one_block_consts + 14 * DCTSIZE)
  } };
#endif

  /* Load 8x8 block of DCT coefficients. */
#ifdef HAVE_VLD1Q_U8_X4
  const int8x16x4_t tbl_rows_0123 =
    vld1q_s8_x4((int8_t *)(block + 0 * DCTSIZE));
  const int8x16x4_t tbl_rows_4567 =
    vld1q_s8_x4((int8_t *)(block + 4 * DCTSIZE));
#else
  const int8x16x4_t tbl_rows_0123 = { {
    vld1q_s8((int8_t *)(block + 0 * DCTSIZE)),
    vld1q_s8((int8_t *)(block + 1 * DCTSIZE)),
    vld1q_s8((int8_t *)(block + 2 * DCTSIZE)),
    vld1q_s8((int8_t *)(block + 3 * DCTSIZE))
  } };
  const int8x16x4_t tbl_rows_4567 = { {
    vld1q_s8((int8_t *)(block + 4 * DCTSIZE)),
    vld1q_s8((int8_t *)(block + 5 * DCTSIZE)),
    vld1q_s8((int8_t *)(block + 6 * DCTSIZE)),
    vld1q_s8((int8_t *)(block + 7 * DCTSIZE))
  } };
#endif

  /* Initialise extra lookup tables. */
  const int8x16x4_t tbl_rows_2345 = { {
    tbl_rows_0123.val[2], tbl_rows_0123.val[3],
    tbl_rows_4567.val[0], tbl_rows_4567.val[1]
  } };
  const int8x16x3_t tbl_rows_567 =
    { { tbl_rows_4567.val[1], tbl_rows_4567.val[2], tbl_rows_4567.val[3] } };

  /* Shuffle coefficients into zig-zag order. */
  int16x8_t row0 =
    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[0]));
  int16x8_t row1 =
    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[1]));
  int16x8_t row2 =
    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_2345, idx_rows_0123.val[2]));
  int16x8_t row3 =
    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_0123, idx_rows_0123.val[3]));
  int16x8_t row4 =
    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_4567, idx_rows_4567.val[0]));
  int16x8_t row5 =
    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_2345, idx_rows_4567.val[1]));
  int16x8_t row6 =
    vreinterpretq_s16_s8(vqtbl4q_s8(tbl_rows_4567, idx_rows_4567.val[2]));
  int16x8_t row7 =
    vreinterpretq_s16_s8(vqtbl3q_s8(tbl_rows_567, idx_rows_4567.val[3]));

  /* Compute DC coefficient difference value (F.1.1.5.1). */
  row0 = vsetq_lane_s16(block[0] - last_dc_val, row0, 0);
  /* Initialize AC coefficient lanes not reachable by lookup tables. */
  row1 =
    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[0]),
                                  0), row1, 2);
  row2 =
    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[1]),
                                  4), row2, 0);
  row2 =
    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[2]),
                                  0), row2, 5);
  row5 =
    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[1]),
                                  7), row5, 2);
  row5 =
    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_4567.val[2]),
                                  3), row5, 7);
  row6 =
    vsetq_lane_s16(vgetq_lane_s16(vreinterpretq_s16_s8(tbl_rows_0123.val[3]),
                                  7), row6, 5);

  /* DCT block is now in zig-zag order; start Huffman encoding process. */

  /* Construct bitmap to accelerate encoding of AC coefficients.  A set bit
   * means that the corresponding coefficient != 0.
   */
  uint16x8_t row0_ne_0 = vtstq_s16(row0, row0);
  uint16x8_t row1_ne_0 = vtstq_s16(row1, row1);
  uint16x8_t row2_ne_0 = vtstq_s16(row2, row2);
  uint16x8_t row3_ne_0 = vtstq_s16(row3, row3);
  uint16x8_t row4_ne_0 = vtstq_s16(row4, row4);
  uint16x8_t row5_ne_0 = vtstq_s16(row5, row5);
  uint16x8_t row6_ne_0 = vtstq_s16(row6, row6);
  uint16x8_t row7_ne_0 = vtstq_s16(row7, row7);

  uint8x16_t row10_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row1_ne_0),
                                    vreinterpretq_u8_u16(row0_ne_0));
  uint8x16_t row32_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row3_ne_0),
                                    vreinterpretq_u8_u16(row2_ne_0));
  uint8x16_t row54_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row5_ne_0),
                                    vreinterpretq_u8_u16(row4_ne_0));
  uint8x16_t row76_ne_0 = vuzp1q_u8(vreinterpretq_u8_u16(row7_ne_0),
                                    vreinterpretq_u8_u16(row6_ne_0));

  /* { 0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01 } */
  const uint8x16_t bitmap_mask =
    vreinterpretq_u8_u64(vdupq_n_u64(0x0102040810204080));

  uint8x16_t bitmap_rows_10 = vandq_u8(row10_ne_0, bitmap_mask);
  uint8x16_t bitmap_rows_32 = vandq_u8(row32_ne_0, bitmap_mask);
  uint8x16_t bitmap_rows_54 = vandq_u8(row54_ne_0, bitmap_mask);
  uint8x16_t bitmap_rows_76 = vandq_u8(row76_ne_0, bitmap_mask);

  uint8x16_t bitmap_rows_3210 = vpaddq_u8(bitmap_rows_32, bitmap_rows_10);
  uint8x16_t bitmap_rows_7654 = vpaddq_u8(bitmap_rows_76, bitmap_rows_54);
  uint8x16_t bitmap_rows_76543210 = vpaddq_u8(bitmap_rows_7654,
                                              bitmap_rows_3210);
  uint8x8_t bitmap_all = vpadd_u8(vget_low_u8(bitmap_rows_76543210),
                                  vget_high_u8(bitmap_rows_76543210));

  /* Shift left to remove DC bit. */
  bitmap_all =
    vreinterpret_u8_u64(vshl_n_u64(vreinterpret_u64_u8(bitmap_all), 1));
  /* Count bits set (number of non-zero coefficients) in bitmap. */
  unsigned int non_zero_coefficients = vaddv_u8(vcnt_u8(bitmap_all));
  /* Move bitmap to 64-bit scalar register. */
  uint64_t bitmap = vget_lane_u64(vreinterpret_u64_u8(bitmap_all), 0);

  /* Set up state and bit buffer for output bitstream. */
  working_state *state_ptr = (working_state *)state;
  int free_bits = state_ptr->cur.free_bits;
  size_t put_buffer = state_ptr->cur.put_buffer;

  /* Encode DC coefficient. */

  /* For negative coeffs: diff = abs(coeff) -1 = ~abs(coeff) */
  int16x8_t abs_row0 = vabsq_s16(row0);
  int16x8_t row0_lz = vclzq_s16(abs_row0);
  uint16x8_t row0_mask = vshlq_u16(vcltzq_s16(row0), vnegq_s16(row0_lz));
  uint16x8_t row0_diff = veorq_u16(vreinterpretq_u16_s16(abs_row0), row0_mask);
  /* Find nbits required to specify sign and amplitude of coefficient. */
  unsigned int lz = vgetq_lane_u16(vreinterpretq_u16_s16(row0_lz), 0);
  unsigned int nbits = 16 - lz;
  /* Emit Huffman-coded symbol and additional diff bits. */
  unsigned int diff = vgetq_lane_u16(row0_diff, 0);
  PUT_CODE(dctbl->ehufco[nbits], dctbl->ehufsi[nbits], diff)

  /* Encode AC coefficients. */

  unsigned int r = 0;  /* r = run length of zeros */
  unsigned int i = 1;  /* i = number of coefficients encoded */
  /* Code and size information for a run length of 16 zero coefficients */
  const unsigned int code_0xf0 = actbl->ehufco[0xf0];
  const unsigned int size_0xf0 = actbl->ehufsi[0xf0];

  /* The most efficient method of computing nbits and diff depends on the
   * number of non-zero coefficients.  If the bitmap is not too sparse (> 8
   * non-zero AC coefficients), it is beneficial to do all of the work using
   * Neon; else we do some of the work using Neon and the rest on demand using
   * scalar code.
   */
  if (non_zero_coefficients > 8) {
    uint8_t block_nbits[DCTSIZE2];

    int16x8_t abs_row1 = vabsq_s16(row1);
    int16x8_t abs_row2 = vabsq_s16(row2);
    int16x8_t abs_row3 = vabsq_s16(row3);
    int16x8_t abs_row4 = vabsq_s16(row4);
    int16x8_t abs_row5 = vabsq_s16(row5);
    int16x8_t abs_row6 = vabsq_s16(row6);
    int16x8_t abs_row7 = vabsq_s16(row7);
    int16x8_t row1_lz = vclzq_s16(abs_row1);
    int16x8_t row2_lz = vclzq_s16(abs_row2);
    int16x8_t row3_lz = vclzq_s16(abs_row3);
    int16x8_t row4_lz = vclzq_s16(abs_row4);
    int16x8_t row5_lz = vclzq_s16(abs_row5);
    int16x8_t row6_lz = vclzq_s16(abs_row6);
    int16x8_t row7_lz = vclzq_s16(abs_row7);
    /* Narrow leading zero count to 8 bits. */
    uint8x16_t row01_lz = vuzp1q_u8(vreinterpretq_u8_s16(row0_lz),
                                    vreinterpretq_u8_s16(row1_lz));
    uint8x16_t row23_lz = vuzp1q_u8(vreinterpretq_u8_s16(row2_lz),
                                    vreinterpretq_u8_s16(row3_lz));
    uint8x16_t row45_lz = vuzp1q_u8(vreinterpretq_u8_s16(row4_lz),
                                    vreinterpretq_u8_s16(row5_lz));
    uint8x16_t row67_lz = vuzp1q_u8(vreinterpretq_u8_s16(row6_lz),
                                    vreinterpretq_u8_s16(row7_lz));
    /* Compute nbits needed to specify magnitude of each coefficient. */
    uint8x16_t row01_nbits = vsubq_u8(vdupq_n_u8(16), row01_lz);
    uint8x16_t row23_nbits = vsubq_u8(vdupq_n_u8(16), row23_lz);
    uint8x16_t row45_nbits = vsubq_u8(vdupq_n_u8(16), row45_lz);
    uint8x16_t row67_nbits = vsubq_u8(vdupq_n_u8(16), row67_lz);
    /* Store nbits. */
    vst1q_u8(block_nbits + 0 * DCTSIZE, row01_nbits);
    vst1q_u8(block_nbits + 2 * DCTSIZE, row23_nbits);
    vst1q_u8(block_nbits + 4 * DCTSIZE, row45_nbits);
    vst1q_u8(block_nbits + 6 * DCTSIZE, row67_nbits);
    /* Mask bits not required to specify sign and amplitude of diff. */
    uint16x8_t row1_mask = vshlq_u16(vcltzq_s16(row1), vnegq_s16(row1_lz));
    uint16x8_t row2_mask = vshlq_u16(vcltzq_s16(row2), vnegq_s16(row2_lz));
    uint16x8_t row3_mask = vshlq_u16(vcltzq_s16(row3), vnegq_s16(row3_lz));
    uint16x8_t row4_mask = vshlq_u16(vcltzq_s16(row4), vnegq_s16(row4_lz));
    uint16x8_t row5_mask = vshlq_u16(vcltzq_s16(row5), vnegq_s16(row5_lz));
    uint16x8_t row6_mask = vshlq_u16(vcltzq_s16(row6), vnegq_s16(row6_lz));
    uint16x8_t row7_mask = vshlq_u16(vcltzq_s16(row7), vnegq_s16(row7_lz));
    /* diff = abs(coeff) ^ sign(coeff) [no-op for positive coefficients] */
    uint16x8_t row1_diff = veorq_u16(vreinterpretq_u16_s16(abs_row1),
                                     row1_mask);
    uint16x8_t row2_diff = veorq_u16(vreinterpretq_u16_s16(abs_row2),
                                     row2_mask);
    uint16x8_t row3_diff = veorq_u16(vreinterpretq_u16_s16(abs_row3),
                                     row3_mask);
    uint16x8_t row4_diff = veorq_u16(vreinterpretq_u16_s16(abs_row4),
                                     row4_mask);
    uint16x8_t row5_diff = veorq_u16(vreinterpretq_u16_s16(abs_row5),
                                     row5_mask);
    uint16x8_t row6_diff = veorq_u16(vreinterpretq_u16_s16(abs_row6),
                                     row6_mask);
    uint16x8_t row7_diff = veorq_u16(vreinterpretq_u16_s16(abs_row7),
                                     row7_mask);
    /* Store diff bits. */
    vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
    vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
    vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
    vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
    vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
    vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
    vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
    vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);

    while (bitmap != 0) {
      r = BUILTIN_CLZLL(bitmap);
      i += r;
      bitmap <<= r;
      nbits = block_nbits[i];
      diff = block_diff[i];
      while (r > 15) {
        /* If run length > 15, emit special run-length-16 codes. */
        PUT_BITS(code_0xf0, size_0xf0)
        r -= 16;
      }
      /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
      unsigned int rs = (r << 4) + nbits;
      PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
      i++;
      bitmap <<= 1;
    }
  } else if (bitmap != 0) {
    uint16_t block_abs[DCTSIZE2];
    /* Compute and store absolute value of coefficients. */
    int16x8_t abs_row1 = vabsq_s16(row1);
    int16x8_t abs_row2 = vabsq_s16(row2);
    int16x8_t abs_row3 = vabsq_s16(row3);
    int16x8_t abs_row4 = vabsq_s16(row4);
    int16x8_t abs_row5 = vabsq_s16(row5);
    int16x8_t abs_row6 = vabsq_s16(row6);
    int16x8_t abs_row7 = vabsq_s16(row7);
    vst1q_u16(block_abs + 0 * DCTSIZE, vreinterpretq_u16_s16(abs_row0));
    vst1q_u16(block_abs + 1 * DCTSIZE, vreinterpretq_u16_s16(abs_row1));
    vst1q_u16(block_abs + 2 * DCTSIZE, vreinterpretq_u16_s16(abs_row2));
    vst1q_u16(block_abs + 3 * DCTSIZE, vreinterpretq_u16_s16(abs_row3));
    vst1q_u16(block_abs + 4 * DCTSIZE, vreinterpretq_u16_s16(abs_row4));
    vst1q_u16(block_abs + 5 * DCTSIZE, vreinterpretq_u16_s16(abs_row5));
    vst1q_u16(block_abs + 6 * DCTSIZE, vreinterpretq_u16_s16(abs_row6));
    vst1q_u16(block_abs + 7 * DCTSIZE, vreinterpretq_u16_s16(abs_row7));
    /* Compute diff bits (without nbits mask) and store. */
    uint16x8_t row1_diff = veorq_u16(vreinterpretq_u16_s16(abs_row1),
                                     vcltzq_s16(row1));
    uint16x8_t row2_diff = veorq_u16(vreinterpretq_u16_s16(abs_row2),
                                     vcltzq_s16(row2));
    uint16x8_t row3_diff = veorq_u16(vreinterpretq_u16_s16(abs_row3),
                                     vcltzq_s16(row3));
    uint16x8_t row4_diff = veorq_u16(vreinterpretq_u16_s16(abs_row4),
                                     vcltzq_s16(row4));
    uint16x8_t row5_diff = veorq_u16(vreinterpretq_u16_s16(abs_row5),
                                     vcltzq_s16(row5));
    uint16x8_t row6_diff = veorq_u16(vreinterpretq_u16_s16(abs_row6),
                                     vcltzq_s16(row6));
    uint16x8_t row7_diff = veorq_u16(vreinterpretq_u16_s16(abs_row7),
                                     vcltzq_s16(row7));
    vst1q_u16(block_diff + 0 * DCTSIZE, row0_diff);
    vst1q_u16(block_diff + 1 * DCTSIZE, row1_diff);
    vst1q_u16(block_diff + 2 * DCTSIZE, row2_diff);
    vst1q_u16(block_diff + 3 * DCTSIZE, row3_diff);
    vst1q_u16(block_diff + 4 * DCTSIZE, row4_diff);
    vst1q_u16(block_diff + 5 * DCTSIZE, row5_diff);
    vst1q_u16(block_diff + 6 * DCTSIZE, row6_diff);
    vst1q_u16(block_diff + 7 * DCTSIZE, row7_diff);

    /* Same as above but must mask diff bits and compute nbits on demand. */
    while (bitmap != 0) {
      r = BUILTIN_CLZLL(bitmap);
      i += r;
      bitmap <<= r;
      lz = BUILTIN_CLZ(block_abs[i]);
      nbits = 32 - lz;
      diff = ((unsigned int)block_diff[i] << lz) >> lz;
      while (r > 15) {
        /* If run length > 15, emit special run-length-16 codes. */
        PUT_BITS(code_0xf0, size_0xf0)
        r -= 16;
      }
      /* Emit Huffman symbol for run length / number of bits. (F.1.2.2.1) */
      unsigned int rs = (r << 4) + nbits;
      PUT_CODE(actbl->ehufco[rs], actbl->ehufsi[rs], diff)
      i++;
      bitmap <<= 1;
    }
  }

  /* If the last coefficient(s) were zero, emit an end-of-block (EOB) code.
   * The value of RS for the EOB code is 0.
   */
  if (i != 64) {
    PUT_BITS(actbl->ehufco[0], actbl->ehufsi[0])
  }

  state_ptr->cur.put_buffer = put_buffer;
  state_ptr->cur.free_bits = free_bits;

  return buffer;
}