2 ; jidctfst.asm - fast integer IDCT (64-bit SSE2)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright (C) 2009, 2016, D. R. Commander.
6 ; Copyright (C) 2018, Matthias Räncker.
7 ; Copyright (C) 2023, Aliaksiej Kandracienka.
9 ; Based on the x86 SIMD extension for IJG JPEG library
10 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
11 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
13 ; This file should be assembled with NASM (Netwide Assembler),
14 ; can *not* be assembled with Microsoft's MASM or any compatible
15 ; assembler (including Borland's Turbo Assembler).
16 ; NASM is available from http://nasm.sourceforge.net/ or
17 ; http://sourceforge.net/project/showfiles.php?group_id=6208
19 ; This file contains a fast, not so accurate integer implementation of
20 ; the inverse DCT (Discrete Cosine Transform). The following code is
21 ; based directly on the IJG's original jidctfst.c; see the jidctfst.c
24 %include "jsimdext.inc"
27 ; --------------------------------------------------------------------------
29 %define CONST_BITS 8 ; 14 is also OK.
32 %if IFAST_SCALE_BITS != PASS1_BITS
33 %error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
37 F_1_082 equ 277 ; FIX(1.082392200)
38 F_1_414 equ 362 ; FIX(1.414213562)
39 F_1_847 equ 473 ; FIX(1.847759065)
40 F_2_613 equ 669 ; FIX(2.613125930)
41 F_1_613 equ (F_2_613 - 256) ; FIX(2.613125930) - FIX(1)
43 ; NASM cannot do compile-time arithmetic on floating-point constants.
44 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n))
45 F_1_082 equ DESCALE(1162209775, 30 - CONST_BITS) ; FIX(1.082392200)
46 F_1_414 equ DESCALE(1518500249, 30 - CONST_BITS) ; FIX(1.414213562)
47 F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065)
48 F_2_613 equ DESCALE(2805822602, 30 - CONST_BITS) ; FIX(2.613125930)
49 F_1_613 equ (F_2_613 - (1 << CONST_BITS)) ; FIX(2.613125930) - FIX(1)
52 ; --------------------------------------------------------------------------
55 ; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
56 ; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
58 %define PRE_MULTIPLY_SCALE_BITS 2
59 %define CONST_SHIFT (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
62 GLOBAL_DATA(jconst_idct_ifast_sse2)
64 EXTN(jconst_idct_ifast_sse2):
66 PW_F1414 times 8 dw F_1_414 << CONST_SHIFT
67 PW_F1847 times 8 dw F_1_847 << CONST_SHIFT
68 PW_MF1613 times 8 dw -F_1_613 << CONST_SHIFT
69 PW_F1082 times 8 dw F_1_082 << CONST_SHIFT
70 PB_CENTERJSAMP times 16 db CENTERJSAMPLE
74 ; --------------------------------------------------------------------------
78 ; Perform dequantization and inverse DCT on one block of coefficients.
81 ; jsimd_idct_ifast_sse2(void *dct_table, JCOEFPTR coef_block,
82 ; JSAMPARRAY output_buf, JDIMENSION output_col)
85 ; r10 = jpeg_component_info *compptr
86 ; r11 = JCOEFPTR coef_block
87 ; r12 = JSAMPARRAY output_buf
88 ; r13d = JDIMENSION output_col
90 %define wk(i) r15 - (WK_NUM - (i)) * SIZEOF_XMMWORD
95 GLOBAL_FUNCTION(jsimd_idct_ifast_sse2)
97 EXTN(jsimd_idct_ifast_sse2):
101 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
102 ; Allocate stack space for wk array. r15 is used to access it.
104 sub rsp, byte (SIZEOF_XMMWORD * WK_NUM)
107 ; ---- Pass 1: process columns from input.
109 mov rdx, r10 ; quantptr
112 %ifndef NO_ZERO_COLUMN_TEST_IFAST_SSE2
113 mov eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
114 or eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
117 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
118 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
119 por xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
120 por xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
121 por xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
122 por xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
123 por xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
131 ; -- AC terms all zero
133 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
134 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
136 movdqa xmm7, xmm0 ; xmm0=in0=(00 01 02 03 04 05 06 07)
137 punpcklwd xmm0, xmm0 ; xmm0=(00 00 01 01 02 02 03 03)
138 punpckhwd xmm7, xmm7 ; xmm7=(04 04 05 05 06 06 07 07)
140 pshufd xmm6, xmm0, 0x00 ; xmm6=col0=(00 00 00 00 00 00 00 00)
141 pshufd xmm2, xmm0, 0x55 ; xmm2=col1=(01 01 01 01 01 01 01 01)
142 pshufd xmm5, xmm0, 0xAA ; xmm5=col2=(02 02 02 02 02 02 02 02)
143 pshufd xmm0, xmm0, 0xFF ; xmm0=col3=(03 03 03 03 03 03 03 03)
144 pshufd xmm1, xmm7, 0x00 ; xmm1=col4=(04 04 04 04 04 04 04 04)
145 pshufd xmm4, xmm7, 0x55 ; xmm4=col5=(05 05 05 05 05 05 05 05)
146 pshufd xmm3, xmm7, 0xAA ; xmm3=col6=(06 06 06 06 06 06 06 06)
147 pshufd xmm7, xmm7, 0xFF ; xmm7=col7=(07 07 07 07 07 07 07 07)
149 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=col1
150 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=col3
157 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
158 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
159 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
160 pmullw xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
161 movdqa xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
162 movdqa xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
163 pmullw xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
164 pmullw xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
168 psubw xmm0, xmm2 ; xmm0=tmp11
170 paddw xmm4, xmm2 ; xmm4=tmp10
171 paddw xmm5, xmm3 ; xmm5=tmp13
173 psllw xmm1, PRE_MULTIPLY_SCALE_BITS
174 pmulhw xmm1, [rel PW_F1414]
175 psubw xmm1, xmm5 ; xmm1=tmp12
179 psubw xmm4, xmm5 ; xmm4=tmp3
180 psubw xmm0, xmm1 ; xmm0=tmp2
181 paddw xmm6, xmm5 ; xmm6=tmp0
182 paddw xmm7, xmm1 ; xmm7=tmp1
184 movdqa XMMWORD [wk(1)], xmm4 ; wk(1)=tmp3
185 movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=tmp2
189 movdqa xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
190 movdqa xmm3, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
191 pmullw xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
192 pmullw xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
193 movdqa xmm5, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
194 movdqa xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
195 pmullw xmm5, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
196 pmullw xmm1, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_IFAST_MULT_TYPE)]
200 psubw xmm2, xmm1 ; xmm2=z12
201 psubw xmm5, xmm3 ; xmm5=z10
202 paddw xmm4, xmm1 ; xmm4=z11
203 paddw xmm0, xmm3 ; xmm0=z13
205 movdqa xmm1, xmm5 ; xmm1=z10(unscaled)
206 psllw xmm2, PRE_MULTIPLY_SCALE_BITS
207 psllw xmm5, PRE_MULTIPLY_SCALE_BITS
211 paddw xmm3, xmm0 ; xmm3=tmp7
213 psllw xmm4, PRE_MULTIPLY_SCALE_BITS
214 pmulhw xmm4, [rel PW_F1414] ; xmm4=tmp11
216 ; To avoid overflow...
219 ; tmp12 = -2.613125930 * z10 + z5;
221 ; (This implementation)
222 ; tmp12 = (-1.613125930 - 1) * z10 + z5;
223 ; = -1.613125930 * z10 - z10 + z5;
227 pmulhw xmm5, [rel PW_F1847] ; xmm5=z5
228 pmulhw xmm0, [rel PW_MF1613]
229 pmulhw xmm2, [rel PW_F1082]
231 psubw xmm2, xmm5 ; xmm2=tmp10
232 paddw xmm0, xmm5 ; xmm0=tmp12
234 ; -- Final output stage
236 psubw xmm0, xmm3 ; xmm0=tmp6
239 paddw xmm6, xmm3 ; xmm6=data0=(00 01 02 03 04 05 06 07)
240 paddw xmm7, xmm0 ; xmm7=data1=(10 11 12 13 14 15 16 17)
241 psubw xmm1, xmm3 ; xmm1=data7=(70 71 72 73 74 75 76 77)
242 psubw xmm5, xmm0 ; xmm5=data6=(60 61 62 63 64 65 66 67)
243 psubw xmm4, xmm0 ; xmm4=tmp5
245 movdqa xmm3, xmm6 ; transpose coefficients(phase 1)
246 punpcklwd xmm6, xmm7 ; xmm6=(00 10 01 11 02 12 03 13)
247 punpckhwd xmm3, xmm7 ; xmm3=(04 14 05 15 06 16 07 17)
248 movdqa xmm0, xmm5 ; transpose coefficients(phase 1)
249 punpcklwd xmm5, xmm1 ; xmm5=(60 70 61 71 62 72 63 73)
250 punpckhwd xmm0, xmm1 ; xmm0=(64 74 65 75 66 76 67 77)
252 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp2
253 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=tmp3
255 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(60 70 61 71 62 72 63 73)
256 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(64 74 65 75 66 76 67 77)
258 paddw xmm2, xmm4 ; xmm2=tmp4
261 paddw xmm7, xmm4 ; xmm7=data2=(20 21 22 23 24 25 26 27)
262 paddw xmm1, xmm2 ; xmm1=data4=(40 41 42 43 44 45 46 47)
263 psubw xmm5, xmm4 ; xmm5=data5=(50 51 52 53 54 55 56 57)
264 psubw xmm0, xmm2 ; xmm0=data3=(30 31 32 33 34 35 36 37)
266 movdqa xmm4, xmm7 ; transpose coefficients(phase 1)
267 punpcklwd xmm7, xmm0 ; xmm7=(20 30 21 31 22 32 23 33)
268 punpckhwd xmm4, xmm0 ; xmm4=(24 34 25 35 26 36 27 37)
269 movdqa xmm2, xmm1 ; transpose coefficients(phase 1)
270 punpcklwd xmm1, xmm5 ; xmm1=(40 50 41 51 42 52 43 53)
271 punpckhwd xmm2, xmm5 ; xmm2=(44 54 45 55 46 56 47 57)
273 movdqa xmm0, xmm3 ; transpose coefficients(phase 2)
274 punpckldq xmm3, xmm4 ; xmm3=(04 14 24 34 05 15 25 35)
275 punpckhdq xmm0, xmm4 ; xmm0=(06 16 26 36 07 17 27 37)
276 movdqa xmm5, xmm6 ; transpose coefficients(phase 2)
277 punpckldq xmm6, xmm7 ; xmm6=(00 10 20 30 01 11 21 31)
278 punpckhdq xmm5, xmm7 ; xmm5=(02 12 22 32 03 13 23 33)
280 movdqa xmm4, XMMWORD [wk(0)] ; xmm4=(60 70 61 71 62 72 63 73)
281 movdqa xmm7, XMMWORD [wk(1)] ; xmm7=(64 74 65 75 66 76 67 77)
283 movdqa XMMWORD [wk(0)], xmm3 ; wk(0)=(04 14 24 34 05 15 25 35)
284 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=(06 16 26 36 07 17 27 37)
286 movdqa xmm3, xmm1 ; transpose coefficients(phase 2)
287 punpckldq xmm1, xmm4 ; xmm1=(40 50 60 70 41 51 61 71)
288 punpckhdq xmm3, xmm4 ; xmm3=(42 52 62 72 43 53 63 73)
289 movdqa xmm0, xmm2 ; transpose coefficients(phase 2)
290 punpckldq xmm2, xmm7 ; xmm2=(44 54 64 74 45 55 65 75)
291 punpckhdq xmm0, xmm7 ; xmm0=(46 56 66 76 47 57 67 77)
293 movdqa xmm4, xmm6 ; transpose coefficients(phase 3)
294 punpcklqdq xmm6, xmm1 ; xmm6=col0=(00 10 20 30 40 50 60 70)
295 punpckhqdq xmm4, xmm1 ; xmm4=col1=(01 11 21 31 41 51 61 71)
296 movdqa xmm7, xmm5 ; transpose coefficients(phase 3)
297 punpcklqdq xmm5, xmm3 ; xmm5=col2=(02 12 22 32 42 52 62 72)
298 punpckhqdq xmm7, xmm3 ; xmm7=col3=(03 13 23 33 43 53 63 73)
300 movdqa xmm1, XMMWORD [wk(0)] ; xmm1=(04 14 24 34 05 15 25 35)
301 movdqa xmm3, XMMWORD [wk(1)] ; xmm3=(06 16 26 36 07 17 27 37)
303 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=col1
304 movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=col3
306 movdqa xmm4, xmm1 ; transpose coefficients(phase 3)
307 punpcklqdq xmm1, xmm2 ; xmm1=col4=(04 14 24 34 44 54 64 74)
308 punpckhqdq xmm4, xmm2 ; xmm4=col5=(05 15 25 35 45 55 65 75)
309 movdqa xmm7, xmm3 ; transpose coefficients(phase 3)
310 punpcklqdq xmm3, xmm0 ; xmm3=col6=(06 16 26 36 46 56 66 76)
311 punpckhqdq xmm7, xmm0 ; xmm7=col7=(07 17 27 37 47 57 67 77)
314 ; -- Prefetch the next coefficient block
316 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
317 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
318 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
319 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
321 ; ---- Pass 2: process rows from work array, store into output array.
323 mov rdi, r12 ; (JSAMPROW *)
328 ; xmm6=col0, xmm5=col2, xmm1=col4, xmm3=col6
332 psubw xmm6, xmm1 ; xmm6=tmp11
334 paddw xmm2, xmm1 ; xmm2=tmp10
335 paddw xmm0, xmm3 ; xmm0=tmp13
337 psllw xmm5, PRE_MULTIPLY_SCALE_BITS
338 pmulhw xmm5, [rel PW_F1414]
339 psubw xmm5, xmm0 ; xmm5=tmp12
343 psubw xmm2, xmm0 ; xmm2=tmp3
344 psubw xmm6, xmm5 ; xmm6=tmp2
345 paddw xmm1, xmm0 ; xmm1=tmp0
346 paddw xmm3, xmm5 ; xmm3=tmp1
348 movdqa xmm0, XMMWORD [wk(0)] ; xmm0=col1
349 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=col3
351 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=tmp3
352 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=tmp2
356 ; xmm0=col1, xmm5=col3, xmm4=col5, xmm7=col7
360 psubw xmm0, xmm7 ; xmm0=z12
361 psubw xmm4, xmm5 ; xmm4=z10
362 paddw xmm2, xmm7 ; xmm2=z11
363 paddw xmm6, xmm5 ; xmm6=z13
365 movdqa xmm7, xmm4 ; xmm7=z10(unscaled)
366 psllw xmm0, PRE_MULTIPLY_SCALE_BITS
367 psllw xmm4, PRE_MULTIPLY_SCALE_BITS
371 paddw xmm5, xmm6 ; xmm5=tmp7
373 psllw xmm2, PRE_MULTIPLY_SCALE_BITS
374 pmulhw xmm2, [rel PW_F1414] ; xmm2=tmp11
376 ; To avoid overflow...
379 ; tmp12 = -2.613125930 * z10 + z5;
381 ; (This implementation)
382 ; tmp12 = (-1.613125930 - 1) * z10 + z5;
383 ; = -1.613125930 * z10 - z10 + z5;
387 pmulhw xmm4, [rel PW_F1847] ; xmm4=z5
388 pmulhw xmm6, [rel PW_MF1613]
389 pmulhw xmm0, [rel PW_F1082]
391 psubw xmm0, xmm4 ; xmm0=tmp10
392 paddw xmm6, xmm4 ; xmm6=tmp12
394 ; -- Final output stage
396 psubw xmm6, xmm5 ; xmm6=tmp6
399 paddw xmm1, xmm5 ; xmm1=data0=(00 10 20 30 40 50 60 70)
400 paddw xmm3, xmm6 ; xmm3=data1=(01 11 21 31 41 51 61 71)
401 psraw xmm1, (PASS1_BITS+3) ; descale
402 psraw xmm3, (PASS1_BITS+3) ; descale
403 psubw xmm7, xmm5 ; xmm7=data7=(07 17 27 37 47 57 67 77)
404 psubw xmm4, xmm6 ; xmm4=data6=(06 16 26 36 46 56 66 76)
405 psraw xmm7, (PASS1_BITS+3) ; descale
406 psraw xmm4, (PASS1_BITS+3) ; descale
407 psubw xmm2, xmm6 ; xmm2=tmp5
409 packsswb xmm1, xmm4 ; xmm1=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
410 packsswb xmm3, xmm7 ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
412 movdqa xmm5, XMMWORD [wk(1)] ; xmm5=tmp2
413 movdqa xmm6, XMMWORD [wk(0)] ; xmm6=tmp3
415 paddw xmm0, xmm2 ; xmm0=tmp4
418 paddw xmm5, xmm2 ; xmm5=data2=(02 12 22 32 42 52 62 72)
419 paddw xmm6, xmm0 ; xmm6=data4=(04 14 24 34 44 54 64 74)
420 psraw xmm5, (PASS1_BITS+3) ; descale
421 psraw xmm6, (PASS1_BITS+3) ; descale
422 psubw xmm4, xmm2 ; xmm4=data5=(05 15 25 35 45 55 65 75)
423 psubw xmm7, xmm0 ; xmm7=data3=(03 13 23 33 43 53 63 73)
424 psraw xmm4, (PASS1_BITS+3) ; descale
425 psraw xmm7, (PASS1_BITS+3) ; descale
427 movdqa xmm2, [rel PB_CENTERJSAMP] ; xmm2=[rel PB_CENTERJSAMP]
429 packsswb xmm5, xmm6 ; xmm5=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
430 packsswb xmm7, xmm4 ; xmm7=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
437 movdqa xmm0, xmm1 ; transpose coefficients(phase 1)
438 punpcklbw xmm1, xmm3 ; xmm1=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
439 punpckhbw xmm0, xmm3 ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
440 movdqa xmm6, xmm5 ; transpose coefficients(phase 1)
441 punpcklbw xmm5, xmm7 ; xmm5=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
442 punpckhbw xmm6, xmm7 ; xmm6=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
444 movdqa xmm4, xmm1 ; transpose coefficients(phase 2)
445 punpcklwd xmm1, xmm5 ; xmm1=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
446 punpckhwd xmm4, xmm5 ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
447 movdqa xmm2, xmm6 ; transpose coefficients(phase 2)
448 punpcklwd xmm6, xmm0 ; xmm6=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
449 punpckhwd xmm2, xmm0 ; xmm2=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
451 movdqa xmm3, xmm1 ; transpose coefficients(phase 3)
452 punpckldq xmm1, xmm6 ; xmm1=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
453 punpckhdq xmm3, xmm6 ; xmm3=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
454 movdqa xmm7, xmm4 ; transpose coefficients(phase 3)
455 punpckldq xmm4, xmm2 ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
456 punpckhdq xmm7, xmm2 ; xmm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
458 pshufd xmm5, xmm1, 0x4E ; xmm5=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
459 pshufd xmm0, xmm3, 0x4E ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
460 pshufd xmm6, xmm4, 0x4E ; xmm6=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
461 pshufd xmm2, xmm7, 0x4E ; xmm2=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
463 mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
464 mov rsip, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
465 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
466 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
467 mov rdxp, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
468 mov rsip, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
469 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
470 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm7
472 mov rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
473 mov rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
474 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm5
475 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
476 mov rdxp, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
477 mov rsip, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
478 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
479 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
488 ; For some reason, the OS X linker does not honor the request to align the
489 ; segment unless we do this.