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, D. R. Commander.
6;
7; Based on the x86 SIMD extension for IJG JPEG library
8; Copyright (C) 1999-2006, MIYASAKA Masaru.
9; For conditions of distribution and use, see copyright notice in jsimdext.inc
10;
11; This file should be assembled with NASM (Netwide Assembler),
12; can *not* be assembled with Microsoft's MASM or any compatible
13; assembler (including Borland's Turbo Assembler).
14; NASM is available from http://nasm.sourceforge.net/ or
15; http://sourceforge.net/project/showfiles.php?group_id=6208
16;
17; This file contains a slow-but-accurate integer implementation of the
18; inverse DCT (Discrete Cosine Transform). The following code is based
19; directly on the IJG's original jidctint.c; see the jidctint.c for
20; more details.
21;
22; [TAB8]
23
24%include "jsimdext.inc"
25%include "jdct.inc"
26
27; --------------------------------------------------------------------------
28
29%define CONST_BITS  13
30%define PASS1_BITS  2
31
32%define DESCALE_P1  (CONST_BITS - PASS1_BITS)
33%define DESCALE_P2  (CONST_BITS + PASS1_BITS + 3)
34
35%if CONST_BITS == 13
36F_0_298 equ  2446  ; FIX(0.298631336)
37F_0_390 equ  3196  ; FIX(0.390180644)
38F_0_541 equ  4433  ; FIX(0.541196100)
39F_0_765 equ  6270  ; FIX(0.765366865)
40F_0_899 equ  7373  ; FIX(0.899976223)
41F_1_175 equ  9633  ; FIX(1.175875602)
42F_1_501 equ 12299  ; FIX(1.501321110)
43F_1_847 equ 15137  ; FIX(1.847759065)
44F_1_961 equ 16069  ; FIX(1.961570560)
45F_2_053 equ 16819  ; FIX(2.053119869)
46F_2_562 equ 20995  ; FIX(2.562915447)
47F_3_072 equ 25172  ; FIX(3.072711026)
48%else
49; NASM cannot do compile-time arithmetic on floating-point constants.
50%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
51F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS)  ; FIX(0.298631336)
52F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS)  ; FIX(0.390180644)
53F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS)  ; FIX(0.541196100)
54F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
55F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
56F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS)  ; FIX(1.175875602)
57F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS)  ; FIX(1.501321110)
58F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
59F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS)  ; FIX(1.961570560)
60F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS)  ; FIX(2.053119869)
61F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
62F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS)  ; FIX(3.072711026)
63%endif
64
65; --------------------------------------------------------------------------
66; In-place 8x8x16-bit inverse matrix transpose using AVX2 instructions
67; %1-%4: Input/output registers
68; %5-%8: Temp registers
69
70%macro dotranspose 8
71    ; %5=(00 10 20 30 40 50 60 70  01 11 21 31 41 51 61 71)
72    ; %6=(03 13 23 33 43 53 63 73  02 12 22 32 42 52 62 72)
73    ; %7=(04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75)
74    ; %8=(07 17 27 37 47 57 67 77  06 16 26 36 46 56 66 76)
75
76    vpermq      %5, %1, 0xD8
77    vpermq      %6, %2, 0x72
78    vpermq      %7, %3, 0xD8
79    vpermq      %8, %4, 0x72
80    ; transpose coefficients(phase 1)
81    ; %5=(00 10 20 30 01 11 21 31  40 50 60 70 41 51 61 71)
82    ; %6=(02 12 22 32 03 13 23 33  42 52 62 72 43 53 63 73)
83    ; %7=(04 14 24 34 05 15 25 35  44 54 64 74 45 55 65 75)
84    ; %8=(06 16 26 36 07 17 27 37  46 56 66 76 47 57 67 77)
85
86    vpunpcklwd  %1, %5, %6
87    vpunpckhwd  %2, %5, %6
88    vpunpcklwd  %3, %7, %8
89    vpunpckhwd  %4, %7, %8
90    ; transpose coefficients(phase 2)
91    ; %1=(00 02 10 12 20 22 30 32  40 42 50 52 60 62 70 72)
92    ; %2=(01 03 11 13 21 23 31 33  41 43 51 53 61 63 71 73)
93    ; %3=(04 06 14 16 24 26 34 36  44 46 54 56 64 66 74 76)
94    ; %4=(05 07 15 17 25 27 35 37  45 47 55 57 65 67 75 77)
95
96    vpunpcklwd  %5, %1, %2
97    vpunpcklwd  %6, %3, %4
98    vpunpckhwd  %7, %1, %2
99    vpunpckhwd  %8, %3, %4
100    ; transpose coefficients(phase 3)
101    ; %5=(00 01 02 03 10 11 12 13  40 41 42 43 50 51 52 53)
102    ; %6=(04 05 06 07 14 15 16 17  44 45 46 47 54 55 56 57)
103    ; %7=(20 21 22 23 30 31 32 33  60 61 62 63 70 71 72 73)
104    ; %8=(24 25 26 27 34 35 36 37  64 65 66 67 74 75 76 77)
105
106    vpunpcklqdq %1, %5, %6
107    vpunpckhqdq %2, %5, %6
108    vpunpcklqdq %3, %7, %8
109    vpunpckhqdq %4, %7, %8
110    ; transpose coefficients(phase 4)
111    ; %1=(00 01 02 03 04 05 06 07  40 41 42 43 44 45 46 47)
112    ; %2=(10 11 12 13 14 15 16 17  50 51 52 53 54 55 56 57)
113    ; %3=(20 21 22 23 24 25 26 27  60 61 62 63 64 65 66 67)
114    ; %4=(30 31 32 33 34 35 36 37  70 71 72 73 74 75 76 77)
115%endmacro
116
117; --------------------------------------------------------------------------
118; In-place 8x8x16-bit slow integer inverse DCT using AVX2 instructions
119; %1-%4:  Input/output registers
120; %5-%12: Temp registers
121; %9:     Pass (1 or 2)
122
123%macro dodct 13
124    ; -- Even part
125
126    ; (Original)
127    ; z1 = (z2 + z3) * 0.541196100;
128    ; tmp2 = z1 + z3 * -1.847759065;
129    ; tmp3 = z1 + z2 * 0.765366865;
130    ;
131    ; (This implementation)
132    ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
133    ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
134
135    vperm2i128  %6, %3, %3, 0x01        ; %6=in6_2
136    vpunpcklwd  %5, %3, %6              ; %5=in26_62L
137    vpunpckhwd  %6, %3, %6              ; %6=in26_62H
138    vpmaddwd    %5, %5, [rel PW_F130_F054_MF130_F054]  ; %5=tmp3_2L
139    vpmaddwd    %6, %6, [rel PW_F130_F054_MF130_F054]  ; %6=tmp3_2H
140
141    vperm2i128  %7, %1, %1, 0x01        ; %7=in4_0
142    vpsignw     %1, %1, [rel PW_1_NEG1]
143    vpaddw      %7, %7, %1              ; %7=(in0+in4)_(in0-in4)
144
145    vpxor       %1, %1, %1
146    vpunpcklwd  %8, %1, %7              ; %8=tmp0_1L
147    vpunpckhwd  %1, %1, %7              ; %1=tmp0_1H
148    vpsrad      %8, %8, (16-CONST_BITS)  ; vpsrad %8,16 & vpslld %8,CONST_BITS
149    vpsrad      %1, %1, (16-CONST_BITS)  ; vpsrad %1,16 & vpslld %1,CONST_BITS
150
151    vpsubd      %11, %8, %5             ; %11=tmp0_1L-tmp3_2L=tmp13_12L
152    vpaddd      %9, %8, %5              ; %9=tmp0_1L+tmp3_2L=tmp10_11L
153    vpsubd      %12, %1, %6             ; %12=tmp0_1H-tmp3_2H=tmp13_12H
154    vpaddd      %10, %1, %6             ; %10=tmp0_1H+tmp3_2H=tmp10_11H
155
156    ; -- Odd part
157
158    vpaddw      %1, %4, %2              ; %1=in7_5+in3_1=z3_4
159
160    ; (Original)
161    ; z5 = (z3 + z4) * 1.175875602;
162    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
163    ; z3 += z5;  z4 += z5;
164    ;
165    ; (This implementation)
166    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
167    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
168
169    vperm2i128  %8, %1, %1, 0x01        ; %8=z4_3
170    vpunpcklwd  %7, %1, %8              ; %7=z34_43L
171    vpunpckhwd  %8, %1, %8              ; %8=z34_43H
172    vpmaddwd    %7, %7, [rel PW_MF078_F117_F078_F117]  ; %7=z3_4L
173    vpmaddwd    %8, %8, [rel PW_MF078_F117_F078_F117]  ; %8=z3_4H
174
175    ; (Original)
176    ; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
177    ; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
178    ; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
179    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
180    ; tmp0 += z1 + z3;  tmp1 += z2 + z4;
181    ; tmp2 += z2 + z3;  tmp3 += z1 + z4;
182    ;
183    ; (This implementation)
184    ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
185    ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
186    ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
187    ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
188    ; tmp0 += z3;  tmp1 += z4;
189    ; tmp2 += z3;  tmp3 += z4;
190
191    vperm2i128  %2, %2, %2, 0x01        ; %2=in1_3
192    vpunpcklwd  %3, %4, %2              ; %3=in71_53L
193    vpunpckhwd  %4, %4, %2              ; %4=in71_53H
194
195    vpmaddwd    %5, %3, [rel PW_MF060_MF089_MF050_MF256]  ; %5=tmp0_1L
196    vpmaddwd    %6, %4, [rel PW_MF060_MF089_MF050_MF256]  ; %6=tmp0_1H
197    vpaddd      %5, %5, %7              ; %5=tmp0_1L+z3_4L=tmp0_1L
198    vpaddd      %6, %6, %8              ; %6=tmp0_1H+z3_4H=tmp0_1H
199
200    vpmaddwd    %3, %3, [rel PW_MF089_F060_MF256_F050]  ; %3=tmp3_2L
201    vpmaddwd    %4, %4, [rel PW_MF089_F060_MF256_F050]  ; %4=tmp3_2H
202    vperm2i128  %7, %7, %7, 0x01        ; %7=z4_3L
203    vperm2i128  %8, %8, %8, 0x01        ; %8=z4_3H
204    vpaddd      %7, %3, %7              ; %7=tmp3_2L+z4_3L=tmp3_2L
205    vpaddd      %8, %4, %8              ; %8=tmp3_2H+z4_3H=tmp3_2H
206
207    ; -- Final output stage
208
209    vpaddd      %1, %9, %7              ; %1=tmp10_11L+tmp3_2L=data0_1L
210    vpaddd      %2, %10, %8             ; %2=tmp10_11H+tmp3_2H=data0_1H
211    vpaddd      %1, %1, [rel PD_DESCALE_P %+ %13]
212    vpaddd      %2, %2, [rel PD_DESCALE_P %+ %13]
213    vpsrad      %1, %1, DESCALE_P %+ %13
214    vpsrad      %2, %2, DESCALE_P %+ %13
215    vpackssdw   %1, %1, %2              ; %1=data0_1
216
217    vpsubd      %3, %9, %7              ; %3=tmp10_11L-tmp3_2L=data7_6L
218    vpsubd      %4, %10, %8             ; %4=tmp10_11H-tmp3_2H=data7_6H
219    vpaddd      %3, %3, [rel PD_DESCALE_P %+ %13]
220    vpaddd      %4, %4, [rel PD_DESCALE_P %+ %13]
221    vpsrad      %3, %3, DESCALE_P %+ %13
222    vpsrad      %4, %4, DESCALE_P %+ %13
223    vpackssdw   %4, %3, %4              ; %4=data7_6
224
225    vpaddd      %7, %11, %5             ; %7=tmp13_12L+tmp0_1L=data3_2L
226    vpaddd      %8, %12, %6             ; %8=tmp13_12H+tmp0_1H=data3_2H
227    vpaddd      %7, %7, [rel PD_DESCALE_P %+ %13]
228    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %13]
229    vpsrad      %7, %7, DESCALE_P %+ %13
230    vpsrad      %8, %8, DESCALE_P %+ %13
231    vpackssdw   %2, %7, %8              ; %2=data3_2
232
233    vpsubd      %7, %11, %5             ; %7=tmp13_12L-tmp0_1L=data4_5L
234    vpsubd      %8, %12, %6             ; %8=tmp13_12H-tmp0_1H=data4_5H
235    vpaddd      %7, %7, [rel PD_DESCALE_P %+ %13]
236    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %13]
237    vpsrad      %7, %7, DESCALE_P %+ %13
238    vpsrad      %8, %8, DESCALE_P %+ %13
239    vpackssdw   %3, %7, %8              ; %3=data4_5
240%endmacro
241
242; --------------------------------------------------------------------------
243    SECTION     SEG_CONST
244
245    alignz      32
246    GLOBAL_DATA(jconst_idct_islow_avx2)
247
248EXTN(jconst_idct_islow_avx2):
249
250PW_F130_F054_MF130_F054    times 4  dw  (F_0_541 + F_0_765),  F_0_541
251                           times 4  dw  (F_0_541 - F_1_847),  F_0_541
252PW_MF078_F117_F078_F117    times 4  dw  (F_1_175 - F_1_961),  F_1_175
253                           times 4  dw  (F_1_175 - F_0_390),  F_1_175
254PW_MF060_MF089_MF050_MF256 times 4  dw  (F_0_298 - F_0_899), -F_0_899
255                           times 4  dw  (F_2_053 - F_2_562), -F_2_562
256PW_MF089_F060_MF256_F050   times 4  dw -F_0_899, (F_1_501 - F_0_899)
257                           times 4  dw -F_2_562, (F_3_072 - F_2_562)
258PD_DESCALE_P1              times 8  dd  1 << (DESCALE_P1 - 1)
259PD_DESCALE_P2              times 8  dd  1 << (DESCALE_P2 - 1)
260PB_CENTERJSAMP             times 32 db  CENTERJSAMPLE
261PW_1_NEG1                  times 8  dw  1
262                           times 8  dw -1
263
264    alignz      32
265
266; --------------------------------------------------------------------------
267    SECTION     SEG_TEXT
268    BITS        64
269;
270; Perform dequantization and inverse DCT on one block of coefficients.
271;
272; GLOBAL(void)
273; jsimd_idct_islow_avx2(void *dct_table, JCOEFPTR coef_block,
274;                       JSAMPARRAY output_buf, JDIMENSION output_col)
275;
276
277; r10 = jpeg_component_info *compptr
278; r11 = JCOEFPTR coef_block
279; r12 = JSAMPARRAY output_buf
280; r13d = JDIMENSION output_col
281
282    align       32
283    GLOBAL_FUNCTION(jsimd_idct_islow_avx2)
284
285EXTN(jsimd_idct_islow_avx2):
286    push        rbp
287    mov         rax, rsp                     ; rax = original rbp
288    mov         rbp, rsp                     ; rbp = aligned rbp
289    push_xmm    4
290    collect_args 4
291
292    ; ---- Pass 1: process columns.
293
294%ifndef NO_ZERO_COLUMN_TEST_ISLOW_AVX2
295    mov         eax, DWORD [DWBLOCK(1,0,r11,SIZEOF_JCOEF)]
296    or          eax, DWORD [DWBLOCK(2,0,r11,SIZEOF_JCOEF)]
297    jnz         near .columnDCT
298
299    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,r11,SIZEOF_JCOEF)]
300    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,r11,SIZEOF_JCOEF)]
301    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(3,0,r11,SIZEOF_JCOEF)]
302    vpor        xmm1, xmm1, XMMWORD [XMMBLOCK(4,0,r11,SIZEOF_JCOEF)]
303    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(5,0,r11,SIZEOF_JCOEF)]
304    vpor        xmm1, xmm1, XMMWORD [XMMBLOCK(6,0,r11,SIZEOF_JCOEF)]
305    vpor        xmm0, xmm0, XMMWORD [XMMBLOCK(7,0,r11,SIZEOF_JCOEF)]
306    vpor        xmm1, xmm1, xmm0
307    vpacksswb   xmm1, xmm1, xmm1
308    vpacksswb   xmm1, xmm1, xmm1
309    movd        eax, xmm1
310    test        rax, rax
311    jnz         short .columnDCT
312
313    ; -- AC terms all zero
314
315    movdqa      xmm5, XMMWORD [XMMBLOCK(0,0,r11,SIZEOF_JCOEF)]
316    vpmullw     xmm5, xmm5, XMMWORD [XMMBLOCK(0,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
317
318    vpsllw      xmm5, xmm5, PASS1_BITS
319
320    vpunpcklwd  xmm4, xmm5, xmm5        ; xmm4=(00 00 01 01 02 02 03 03)
321    vpunpckhwd  xmm5, xmm5, xmm5        ; xmm5=(04 04 05 05 06 06 07 07)
322    vinserti128 ymm4, ymm4, xmm5, 1
323
324    vpshufd     ymm0, ymm4, 0x00        ; ymm0=col0_4=(00 00 00 00 00 00 00 00  04 04 04 04 04 04 04 04)
325    vpshufd     ymm1, ymm4, 0x55        ; ymm1=col1_5=(01 01 01 01 01 01 01 01  05 05 05 05 05 05 05 05)
326    vpshufd     ymm2, ymm4, 0xAA        ; ymm2=col2_6=(02 02 02 02 02 02 02 02  06 06 06 06 06 06 06 06)
327    vpshufd     ymm3, ymm4, 0xFF        ; ymm3=col3_7=(03 03 03 03 03 03 03 03  07 07 07 07 07 07 07 07)
328
329    jmp         near .column_end
330%endif
331.columnDCT:
332
333    vmovdqu     ymm4, YMMWORD [YMMBLOCK(0,0,r11,SIZEOF_JCOEF)]  ; ymm4=in0_1
334    vmovdqu     ymm5, YMMWORD [YMMBLOCK(2,0,r11,SIZEOF_JCOEF)]  ; ymm5=in2_3
335    vmovdqu     ymm6, YMMWORD [YMMBLOCK(4,0,r11,SIZEOF_JCOEF)]  ; ymm6=in4_5
336    vmovdqu     ymm7, YMMWORD [YMMBLOCK(6,0,r11,SIZEOF_JCOEF)]  ; ymm7=in6_7
337    vpmullw     ymm4, ymm4, YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
338    vpmullw     ymm5, ymm5, YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
339    vpmullw     ymm6, ymm6, YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
340    vpmullw     ymm7, ymm7, YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_ISLOW_MULT_TYPE)]
341
342    vperm2i128  ymm0, ymm4, ymm6, 0x20  ; ymm0=in0_4
343    vperm2i128  ymm1, ymm5, ymm4, 0x31  ; ymm1=in3_1
344    vperm2i128  ymm2, ymm5, ymm7, 0x20  ; ymm2=in2_6
345    vperm2i128  ymm3, ymm7, ymm6, 0x31  ; ymm3=in7_5
346
347    dodct ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, ymm8, ymm9, ymm10, ymm11, 1
348    ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm3=data7_6
349
350    dotranspose ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
351    ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm3=data3_7
352
353.column_end:
354
355    ; -- Prefetch the next coefficient block
356
357    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 0*32]
358    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 1*32]
359    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 2*32]
360    prefetchnta [r11 + DCTSIZE2*SIZEOF_JCOEF + 3*32]
361
362    ; ---- Pass 2: process rows.
363
364    vperm2i128  ymm4, ymm3, ymm1, 0x31  ; ymm3=in7_5
365    vperm2i128  ymm1, ymm3, ymm1, 0x20  ; ymm1=in3_1
366
367    dodct ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7, ymm8, ymm9, ymm10, ymm11, 2
368    ; ymm0=data0_1, ymm1=data3_2, ymm2=data4_5, ymm4=data7_6
369
370    dotranspose ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7
371    ; ymm0=data0_4, ymm1=data1_5, ymm2=data2_6, ymm4=data3_7
372
373    vpacksswb   ymm0, ymm0, ymm1        ; ymm0=data01_45
374    vpacksswb   ymm1, ymm2, ymm4        ; ymm1=data23_67
375    vpaddb      ymm0, ymm0, [rel PB_CENTERJSAMP]
376    vpaddb      ymm1, ymm1, [rel PB_CENTERJSAMP]
377
378    vextracti128 xmm6, ymm1, 1          ; xmm3=data67
379    vextracti128 xmm4, ymm0, 1          ; xmm2=data45
380    vextracti128 xmm2, ymm1, 0          ; xmm1=data23
381    vextracti128 xmm0, ymm0, 0          ; xmm0=data01
382
383    vpshufd     xmm1, xmm0, 0x4E  ; xmm1=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
384    vpshufd     xmm3, xmm2, 0x4E  ; xmm3=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
385    vpshufd     xmm5, xmm4, 0x4E  ; xmm5=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
386    vpshufd     xmm7, xmm6, 0x4E  ; xmm7=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
387
388    vzeroupper
389
390    mov         eax, r13d
391
392    mov         rdx, JSAMPROW [r12+0*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
393    mov         rsi, JSAMPROW [r12+1*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
394    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm0
395    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
396
397    mov         rdx, JSAMPROW [r12+2*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
398    mov         rsi, JSAMPROW [r12+3*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
399    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
400    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
401
402    mov         rdx, JSAMPROW [r12+4*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
403    mov         rsi, JSAMPROW [r12+5*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
404    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
405    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
406
407    mov         rdx, JSAMPROW [r12+6*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
408    mov         rsi, JSAMPROW [r12+7*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
409    movq        XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
410    movq        XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
411
412    uncollect_args 4
413    pop_xmm     4
414    pop         rbp
415    ret
416
417; For some reason, the OS X linker does not honor the request to align the
418; segment unless we do this.
419    align       32
420