• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jidctint.asm - accurate integer IDCT (64-bit AVX2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
6; Copyright (C) 2018, Matthias Räncker.
7;
8; Based on the x86 SIMD extension for IJG JPEG library
9; Copyright (C) 1999-2006, MIYASAKA Masaru.
10; For conditions of distribution and use, see copyright notice in jsimdext.inc
11;
12; This file should be assembled with NASM (Netwide Assembler),
13; can *not* be assembled with Microsoft's MASM or any compatible
14; assembler (including Borland's Turbo Assembler).
15; NASM is available from http://nasm.sourceforge.net/ or
16; http://sourceforge.net/project/showfiles.php?group_id=6208
17;
18; This file contains a slower but more accurate integer implementation of the
19; inverse DCT (Discrete Cosine Transform). The following code is based
20; directly on the IJG's original jidctint.c; see the jidctint.c for
21; more details.
22
23%include "jsimdext.inc"
24%include "jdct.inc"
25
26; --------------------------------------------------------------------------
27
28%define CONST_BITS  13
29%define PASS1_BITS  2
30
31%define DESCALE_P1  (CONST_BITS - PASS1_BITS)
32%define DESCALE_P2  (CONST_BITS + PASS1_BITS + 3)
33
34%if CONST_BITS == 13
35F_0_298 equ  2446  ; FIX(0.298631336)
36F_0_390 equ  3196  ; FIX(0.390180644)
37F_0_541 equ  4433  ; FIX(0.541196100)
38F_0_765 equ  6270  ; FIX(0.765366865)
39F_0_899 equ  7373  ; FIX(0.899976223)
40F_1_175 equ  9633  ; FIX(1.175875602)
41F_1_501 equ 12299  ; FIX(1.501321110)
42F_1_847 equ 15137  ; FIX(1.847759065)
43F_1_961 equ 16069  ; FIX(1.961570560)
44F_2_053 equ 16819  ; FIX(2.053119869)
45F_2_562 equ 20995  ; FIX(2.562915447)
46F_3_072 equ 25172  ; FIX(3.072711026)
47%else
48; NASM cannot do compile-time arithmetic on floating-point constants.
49%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
50F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS)  ; FIX(0.298631336)
51F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS)  ; FIX(0.390180644)
52F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS)  ; FIX(0.541196100)
53F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
54F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
55F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS)  ; FIX(1.175875602)
56F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS)  ; FIX(1.501321110)
57F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
58F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS)  ; FIX(1.961570560)
59F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS)  ; FIX(2.053119869)
60F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
61F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS)  ; FIX(3.072711026)
62%endif
63
64; --------------------------------------------------------------------------
65; In-place 8x8x16-bit inverse matrix transpose using AVX2 instructions
66; %1-%4: Input/output registers
67; %5-%8: Temp registers
68
69%macro dotranspose 8
70    ; %5=(00 10 20 30 40 50 60 70  01 11 21 31 41 51 61 71)
71    ; %6=(03 13 23 33 43 53 63 73  02 12 22 32 42 52 62 72)
72    ; %7=(04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75)
73    ; %8=(07 17 27 37 47 57 67 77  06 16 26 36 46 56 66 76)
74
75    vpermq      %5, %1, 0xD8
76    vpermq      %6, %2, 0x72
77    vpermq      %7, %3, 0xD8
78    vpermq      %8, %4, 0x72
79    ; transpose coefficients(phase 1)
80    ; %5=(00 10 20 30 01 11 21 31  40 50 60 70 41 51 61 71)
81    ; %6=(02 12 22 32 03 13 23 33  42 52 62 72 43 53 63 73)
82    ; %7=(04 14 24 34 05 15 25 35  44 54 64 74 45 55 65 75)
83    ; %8=(06 16 26 36 07 17 27 37  46 56 66 76 47 57 67 77)
84
85    vpunpcklwd  %1, %5, %6
86    vpunpckhwd  %2, %5, %6
87    vpunpcklwd  %3, %7, %8
88    vpunpckhwd  %4, %7, %8
89    ; transpose coefficients(phase 2)
90    ; %1=(00 02 10 12 20 22 30 32  40 42 50 52 60 62 70 72)
91    ; %2=(01 03 11 13 21 23 31 33  41 43 51 53 61 63 71 73)
92    ; %3=(04 06 14 16 24 26 34 36  44 46 54 56 64 66 74 76)
93    ; %4=(05 07 15 17 25 27 35 37  45 47 55 57 65 67 75 77)
94
95    vpunpcklwd  %5, %1, %2
96    vpunpcklwd  %6, %3, %4
97    vpunpckhwd  %7, %1, %2
98    vpunpckhwd  %8, %3, %4
99    ; transpose coefficients(phase 3)
100    ; %5=(00 01 02 03 10 11 12 13  40 41 42 43 50 51 52 53)
101    ; %6=(04 05 06 07 14 15 16 17  44 45 46 47 54 55 56 57)
102    ; %7=(20 21 22 23 30 31 32 33  60 61 62 63 70 71 72 73)
103    ; %8=(24 25 26 27 34 35 36 37  64 65 66 67 74 75 76 77)
104
105    vpunpcklqdq %1, %5, %6
106    vpunpckhqdq %2, %5, %6
107    vpunpcklqdq %3, %7, %8
108    vpunpckhqdq %4, %7, %8
109    ; transpose coefficients(phase 4)
110    ; %1=(00 01 02 03 04 05 06 07  40 41 42 43 44 45 46 47)
111    ; %2=(10 11 12 13 14 15 16 17  50 51 52 53 54 55 56 57)
112    ; %3=(20 21 22 23 24 25 26 27  60 61 62 63 64 65 66 67)
113    ; %4=(30 31 32 33 34 35 36 37  70 71 72 73 74 75 76 77)
114%endmacro
115
116; --------------------------------------------------------------------------
117; In-place 8x8x16-bit accurate integer inverse DCT using AVX2 instructions
118; %1-%4:  Input/output registers
119; %5-%12: Temp registers
120; %9:     Pass (1 or 2)
121
122%macro dodct 13
123    ; -- Even part
124
125    ; (Original)
126    ; z1 = (z2 + z3) * 0.541196100;
127    ; tmp2 = z1 + z3 * -1.847759065;
128    ; tmp3 = z1 + z2 * 0.765366865;
129    ;
130    ; (This implementation)
131    ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
132    ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
133
134    vperm2i128  %6, %3, %3, 0x01        ; %6=in6_2
135    vpunpcklwd  %5, %3, %6              ; %5=in26_62L
136    vpunpckhwd  %6, %3, %6              ; %6=in26_62H
137    vpmaddwd    %5, %5, [rel PW_F130_F054_MF130_F054]  ; %5=tmp3_2L
138    vpmaddwd    %6, %6, [rel PW_F130_F054_MF130_F054]  ; %6=tmp3_2H
139
140    vperm2i128  %7, %1, %1, 0x01        ; %7=in4_0
141    vpsignw     %1, %1, [rel PW_1_NEG1]
142    vpaddw      %7, %7, %1              ; %7=(in0+in4)_(in0-in4)
143
144    vpxor       %1, %1, %1
145    vpunpcklwd  %8, %1, %7              ; %8=tmp0_1L
146    vpunpckhwd  %1, %1, %7              ; %1=tmp0_1H
147    vpsrad      %8, %8, (16-CONST_BITS)  ; vpsrad %8,16 & vpslld %8,CONST_BITS
148    vpsrad      %1, %1, (16-CONST_BITS)  ; vpsrad %1,16 & vpslld %1,CONST_BITS
149
150    vpsubd      %11, %8, %5             ; %11=tmp0_1L-tmp3_2L=tmp13_12L
151    vpaddd      %9, %8, %5              ; %9=tmp0_1L+tmp3_2L=tmp10_11L
152    vpsubd      %12, %1, %6             ; %12=tmp0_1H-tmp3_2H=tmp13_12H
153    vpaddd      %10, %1, %6             ; %10=tmp0_1H+tmp3_2H=tmp10_11H
154
155    ; -- Odd part
156
157    vpaddw      %1, %4, %2              ; %1=in7_5+in3_1=z3_4
158
159    ; (Original)
160    ; z5 = (z3 + z4) * 1.175875602;
161    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
162    ; z3 += z5;  z4 += z5;
163    ;
164    ; (This implementation)
165    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
166    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
167
168    vperm2i128  %8, %1, %1, 0x01        ; %8=z4_3
169    vpunpcklwd  %7, %1, %8              ; %7=z34_43L
170    vpunpckhwd  %8, %1, %8              ; %8=z34_43H
171    vpmaddwd    %7, %7, [rel PW_MF078_F117_F078_F117]  ; %7=z3_4L
172    vpmaddwd    %8, %8, [rel PW_MF078_F117_F078_F117]  ; %8=z3_4H
173
174    ; (Original)
175    ; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
176    ; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
177    ; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
178    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
179    ; tmp0 += z1 + z3;  tmp1 += z2 + z4;
180    ; tmp2 += z2 + z3;  tmp3 += z1 + z4;
181    ;
182    ; (This implementation)
183    ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
184    ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
185    ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
186    ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
187    ; tmp0 += z3;  tmp1 += z4;
188    ; tmp2 += z3;  tmp3 += z4;
189
190    vperm2i128  %2, %2, %2, 0x01        ; %2=in1_3
191    vpunpcklwd  %3, %4, %2              ; %3=in71_53L
192    vpunpckhwd  %4, %4, %2              ; %4=in71_53H
193
194    vpmaddwd    %5, %3, [rel PW_MF060_MF089_MF050_MF256]  ; %5=tmp0_1L
195    vpmaddwd    %6, %4, [rel PW_MF060_MF089_MF050_MF256]  ; %6=tmp0_1H
196    vpaddd      %5, %5, %7              ; %5=tmp0_1L+z3_4L=tmp0_1L
197    vpaddd      %6, %6, %8              ; %6=tmp0_1H+z3_4H=tmp0_1H
198
199    vpmaddwd    %3, %3, [rel PW_MF089_F060_MF256_F050]  ; %3=tmp3_2L
200    vpmaddwd    %4, %4, [rel PW_MF089_F060_MF256_F050]  ; %4=tmp3_2H
201    vperm2i128  %7, %7, %7, 0x01        ; %7=z4_3L
202    vperm2i128  %8, %8, %8, 0x01        ; %8=z4_3H
203    vpaddd      %7, %3, %7              ; %7=tmp3_2L+z4_3L=tmp3_2L
204    vpaddd      %8, %4, %8              ; %8=tmp3_2H+z4_3H=tmp3_2H
205
206    ; -- Final output stage
207
208    vpaddd      %1, %9, %7              ; %1=tmp10_11L+tmp3_2L=data0_1L
209    vpaddd      %2, %10, %8             ; %2=tmp10_11H+tmp3_2H=data0_1H
210    vpaddd      %1, %1, [rel PD_DESCALE_P %+ %13]
211    vpaddd      %2, %2, [rel PD_DESCALE_P %+ %13]
212    vpsrad      %1, %1, DESCALE_P %+ %13
213    vpsrad      %2, %2, DESCALE_P %+ %13
214    vpackssdw   %1, %1, %2              ; %1=data0_1
215
216    vpsubd      %3, %9, %7              ; %3=tmp10_11L-tmp3_2L=data7_6L
217    vpsubd      %4, %10, %8             ; %4=tmp10_11H-tmp3_2H=data7_6H
218    vpaddd      %3, %3, [rel PD_DESCALE_P %+ %13]
219    vpaddd      %4, %4, [rel PD_DESCALE_P %+ %13]
220    vpsrad      %3, %3, DESCALE_P %+ %13
221    vpsrad      %4, %4, DESCALE_P %+ %13
222    vpackssdw   %4, %3, %4              ; %4=data7_6
223
224    vpaddd      %7, %11, %5             ; %7=tmp13_12L+tmp0_1L=data3_2L
225    vpaddd      %8, %12, %6             ; %8=tmp13_12H+tmp0_1H=data3_2H
226    vpaddd      %7, %7, [rel PD_DESCALE_P %+ %13]
227    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %13]
228    vpsrad      %7, %7, DESCALE_P %+ %13
229    vpsrad      %8, %8, DESCALE_P %+ %13
230    vpackssdw   %2, %7, %8              ; %2=data3_2
231
232    vpsubd      %7, %11, %5             ; %7=tmp13_12L-tmp0_1L=data4_5L
233    vpsubd      %8, %12, %6             ; %8=tmp13_12H-tmp0_1H=data4_5H
234    vpaddd      %7, %7, [rel PD_DESCALE_P %+ %13]
235    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %13]
236    vpsrad      %7, %7, DESCALE_P %+ %13
237    vpsrad      %8, %8, DESCALE_P %+ %13
238    vpackssdw   %3, %7, %8              ; %3=data4_5
239%endmacro
240
241; --------------------------------------------------------------------------
242    SECTION     SEG_CONST
243
244    alignz      32
245    GLOBAL_DATA(jconst_idct_islow_avx2)
246
247EXTN(jconst_idct_islow_avx2):
248
249PW_F130_F054_MF130_F054    times 4  dw  (F_0_541 + F_0_765),  F_0_541
250                           times 4  dw  (F_0_541 - F_1_847),  F_0_541
251PW_MF078_F117_F078_F117    times 4  dw  (F_1_175 - F_1_961),  F_1_175
252                           times 4  dw  (F_1_175 - F_0_390),  F_1_175
253PW_MF060_MF089_MF050_MF256 times 4  dw  (F_0_298 - F_0_899), -F_0_899
254                           times 4  dw  (F_2_053 - F_2_562), -F_2_562
255PW_MF089_F060_MF256_F050   times 4  dw -F_0_899, (F_1_501 - F_0_899)
256                           times 4  dw -F_2_562, (F_3_072 - F_2_562)
257PD_DESCALE_P1              times 8  dd  1 << (DESCALE_P1 - 1)
258PD_DESCALE_P2              times 8  dd  1 << (DESCALE_P2 - 1)
259PB_CENTERJSAMP             times 32 db  CENTERJSAMPLE
260PW_1_NEG1                  times 8  dw  1
261                           times 8  dw -1
262
263    alignz      32
264
265; --------------------------------------------------------------------------
266    SECTION     SEG_TEXT
267    BITS        64
268;
269; Perform dequantization and inverse DCT on one block of coefficients.
270;
271; GLOBAL(void)
272; jsimd_idct_islow_avx2(void *dct_table, JCOEFPTR coef_block,
273;                       JSAMPARRAY output_buf, JDIMENSION output_col)
274;
275
276; r10 = jpeg_component_info *compptr
277; r11 = JCOEFPTR coef_block
278; r12 = JSAMPARRAY output_buf
279; r13d = JDIMENSION output_col
280
281    align       32
282    GLOBAL_FUNCTION(jsimd_idct_islow_avx2)
283
284EXTN(jsimd_idct_islow_avx2):
285    push        rbp
286    mov         rax, rsp                     ; rax = original rbp
287    mov         rbp, rsp                     ; rbp = aligned rbp
288    push_xmm    4
289    collect_args 4
290
291    ; ---- Pass 1: process columns.
292
293%ifndef NO_ZERO_COLUMN_TEST_ISLOW_AVX2
294    mov         eax, dword [DWBLOCK(1,0,r11,SIZEOF_JCOEF)]
295    or          eax, dword [DWBLOCK(2,0,r11,SIZEOF_JCOEF)]
296    jnz         near .columnDCT
297
298    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,r11,SIZEOF_JCOEF)]
299    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,r11,SIZEOF_JCOEF)]
300    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(3,0,r11,SIZEOF_JCOEF)]
301    vpor        xmm1, xmm1, XMMWORD [XMMBLOCK(4,0,r11,SIZEOF_JCOEF)]
302    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(5,0,r11,SIZEOF_JCOEF)]
303    vpor        xmm1, xmm1, XMMWORD [XMMBLOCK(6,0,r11,SIZEOF_JCOEF)]
304    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(7,0,r11,SIZEOF_JCOEF)]
305    vpor        xmm1, xmm1, xmm0
306    vpacksswb   xmm1, xmm1, xmm1
307    vpacksswb   xmm1, xmm1, xmm1
308    movd        eax, xmm1
309    test        rax, rax
310    jnz         short .columnDCT
311
312    ; -- AC terms all zero
313
314    movdqa      xmm5, XMMWORD [XMMBLOCK(0,0,r11,SIZEOF_JCOEF)]
315    vpmullw     xmm5, xmm5, XMMWORD [XMMBLOCK(0,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
316
317    vpsllw      xmm5, xmm5, PASS1_BITS
318
319    vpunpcklwd  xmm4, xmm5, xmm5        ; xmm4=(00 00 01 01 02 02 03 03)
320    vpunpckhwd  xmm5, xmm5, xmm5        ; xmm5=(04 04 05 05 06 06 07 07)
321    vinserti128 ymm4, ymm4, xmm5, 1
322
323    vpshufd     ymm0, ymm4, 0x00        ; ymm0=col0_4=(00 00 00 00 00 00 00 00  04 04 04 04 04 04 04 04)
324    vpshufd     ymm1, ymm4, 0x55        ; ymm1=col1_5=(01 01 01 01 01 01 01 01  05 05 05 05 05 05 05 05)
325    vpshufd     ymm2, ymm4, 0xAA        ; ymm2=col2_6=(02 02 02 02 02 02 02 02  06 06 06 06 06 06 06 06)
326    vpshufd     ymm3, ymm4, 0xFF        ; ymm3=col3_7=(03 03 03 03 03 03 03 03  07 07 07 07 07 07 07 07)
327
328    jmp         near .column_end
329%endif
330.columnDCT:
331
332    vmovdqu     ymm4, YMMWORD [YMMBLOCK(0,0,r11,SIZEOF_JCOEF)]  ; ymm4=in0_1
333    vmovdqu     ymm5, YMMWORD [YMMBLOCK(2,0,r11,SIZEOF_JCOEF)]  ; ymm5=in2_3
334    vmovdqu     ymm6, YMMWORD [YMMBLOCK(4,0,r11,SIZEOF_JCOEF)]  ; ymm6=in4_5
335    vmovdqu     ymm7, YMMWORD [YMMBLOCK(6,0,r11,SIZEOF_JCOEF)]  ; ymm7=in6_7
336    vpmullw     ymm4, ymm4, YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
337    vpmullw     ymm5, ymm5, YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
338    vpmullw     ymm6, ymm6, YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
339    vpmullw     ymm7, ymm7, YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
340
341    vperm2i128  ymm0, ymm4, ymm6, 0x20  ; ymm0=in0_4
342    vperm2i128  ymm1, ymm5, ymm4, 0x31  ; ymm1=in3_1
343    vperm2i128  ymm2, ymm5, ymm7, 0x20  ; ymm2=in2_6
344    vperm2i128  ymm3, ymm7, ymm6, 0x31  ; ymm3=in7_5
345
346    dodct ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, ymm8, ymm9, ymm10, ymm11, 1
347    ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm3=data7_6
348
349    dotranspose ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
350    ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm3=data3_7
351
352.column_end:
353
354    ; -- Prefetch the next coefficient block
355
356    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 0*32]
357    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 1*32]
358    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 2*32]
359    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 3*32]
360
361    ; ---- Pass 2: process rows.
362
363    vperm2i128  ymm4, ymm3, ymm1, 0x31  ; ymm3=in7_5
364    vperm2i128  ymm1, ymm3, ymm1, 0x20  ; ymm1=in3_1
365
366    dodct ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7, ymm8, ymm9, ymm10, ymm11, 2
367    ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm4=data7_6
368
369    dotranspose ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7
370    ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm4=data3_7
371
372    vpacksswb   ymm0, ymm0, ymm1        ; ymm0=data01_45
373    vpacksswb   ymm1, ymm2, ymm4        ; ymm1=data23_67
374    vpaddb      ymm0, ymm0, [rel PB_CENTERJSAMP]
375    vpaddb      ymm1, ymm1, [rel PB_CENTERJSAMP]
376
377    vextracti128 xmm6, ymm1, 1          ; xmm3=data67
378    vextracti128 xmm4, ymm0, 1          ; xmm2=data45
379    vextracti128 xmm2, ymm1, 0          ; xmm1=data23
380    vextracti128 xmm0, ymm0, 0          ; xmm0=data01
381
382    vpshufd     xmm1, xmm0, 0x4E  ; xmm1=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
383    vpshufd     xmm3, xmm2, 0x4E  ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
384    vpshufd     xmm5, xmm4, 0x4E  ; xmm5=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
385    vpshufd     xmm7, xmm6, 0x4E  ; xmm7=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
386
387    vzeroupper
388
389    mov         eax, r13d
390
391    mov         rdxp, JSAMPROW [r12+0*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
392    mov         rsip, JSAMPROW [r12+1*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
393    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm0
394    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
395
396    mov         rdxp, JSAMPROW [r12+2*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
397    mov         rsip, JSAMPROW [r12+3*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
398    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
399    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
400
401    mov         rdxp, JSAMPROW [r12+4*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
402    mov         rsip, JSAMPROW [r12+5*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
403    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
404    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
405
406    mov         rdxp, JSAMPROW [r12+6*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
407    mov         rsip, JSAMPROW [r12+7*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
408    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
409    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
410
411    uncollect_args 4
412    pop_xmm     4
413    pop         rbp
414    ret
415
416; For some reason, the OS X linker does not honor the request to align the
417; segment unless we do this.
418    align       32
419