2 ; jidctint.asm - accurate integer IDCT (64-bit SSE2)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright (C) 2009, 2016, 2020, 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 slower but more accurate integer implementation of the
20 ; inverse DCT (Discrete Cosine Transform). The following code is based
21 ; directly on the IJG's original jidctint.c; see the jidctint.c for
24 %include "jsimdext.inc"
27 ; --------------------------------------------------------------------------
32 %define DESCALE_P1 (CONST_BITS - PASS1_BITS)
33 %define DESCALE_P2 (CONST_BITS + PASS1_BITS + 3)
36 F_0_298 equ 2446 ; FIX(0.298631336)
37 F_0_390 equ 3196 ; FIX(0.390180644)
38 F_0_541 equ 4433 ; FIX(0.541196100)
39 F_0_765 equ 6270 ; FIX(0.765366865)
40 F_0_899 equ 7373 ; FIX(0.899976223)
41 F_1_175 equ 9633 ; FIX(1.175875602)
42 F_1_501 equ 12299 ; FIX(1.501321110)
43 F_1_847 equ 15137 ; FIX(1.847759065)
44 F_1_961 equ 16069 ; FIX(1.961570560)
45 F_2_053 equ 16819 ; FIX(2.053119869)
46 F_2_562 equ 20995 ; FIX(2.562915447)
47 F_3_072 equ 25172 ; FIX(3.072711026)
49 ; NASM cannot do compile-time arithmetic on floating-point constants.
50 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n))
51 F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS) ; FIX(0.298631336)
52 F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS) ; FIX(0.390180644)
53 F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS) ; FIX(0.541196100)
54 F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS) ; FIX(0.765366865)
55 F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS) ; FIX(0.899976223)
56 F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS) ; FIX(1.175875602)
57 F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS) ; FIX(1.501321110)
58 F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065)
59 F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS) ; FIX(1.961570560)
60 F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS) ; FIX(2.053119869)
61 F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS) ; FIX(2.562915447)
62 F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS) ; FIX(3.072711026)
65 ; --------------------------------------------------------------------------
69 GLOBAL_DATA(jconst_idct_islow_sse2)
71 EXTN(jconst_idct_islow_sse2):
73 PW_F130_F054 times 4 dw (F_0_541 + F_0_765), F_0_541
74 PW_F054_MF130 times 4 dw F_0_541, (F_0_541 - F_1_847)
75 PW_MF078_F117 times 4 dw (F_1_175 - F_1_961), F_1_175
76 PW_F117_F078 times 4 dw F_1_175, (F_1_175 - F_0_390)
77 PW_MF060_MF089 times 4 dw (F_0_298 - F_0_899), -F_0_899
78 PW_MF089_F060 times 4 dw -F_0_899, (F_1_501 - F_0_899)
79 PW_MF050_MF256 times 4 dw (F_2_053 - F_2_562), -F_2_562
80 PW_MF256_F050 times 4 dw -F_2_562, (F_3_072 - F_2_562)
81 PD_DESCALE_P1 times 4 dd 1 << (DESCALE_P1 - 1)
82 PD_DESCALE_P2 times 4 dd 1 << (DESCALE_P2 - 1)
83 PB_CENTERJSAMP times 16 db CENTERJSAMPLE
87 ; --------------------------------------------------------------------------
91 ; Perform dequantization and inverse DCT on one block of coefficients.
94 ; jsimd_idct_islow_sse2(void *dct_table, JCOEFPTR coef_block,
95 ; JSAMPARRAY output_buf, JDIMENSION output_col)
98 ; r10 = jpeg_component_info *compptr
99 ; r11 = JCOEFPTR coef_block
100 ; r12 = JSAMPARRAY output_buf
101 ; r13d = JDIMENSION output_col
103 %define wk(i) r15 - (WK_NUM - (i)) * SIZEOF_XMMWORD
108 GLOBAL_FUNCTION(jsimd_idct_islow_sse2)
110 EXTN(jsimd_idct_islow_sse2):
114 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
115 ; Allocate stack space for wk array. r15 is used to access it.
117 sub rsp, (SIZEOF_XMMWORD * WK_NUM)
120 ; ---- Pass 1: process columns from input.
122 mov rdx, r10 ; quantptr
125 %ifndef NO_ZERO_COLUMN_TEST_ISLOW_SSE2
126 mov eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
127 or eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
130 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
131 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
132 por xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
133 por xmm1, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
134 por xmm0, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
135 por xmm1, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
136 por xmm0, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
144 ; -- AC terms all zero
146 movdqa xmm5, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
147 pmullw xmm5, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
149 psllw xmm5, PASS1_BITS
151 movdqa xmm4, xmm5 ; xmm5=in0=(00 01 02 03 04 05 06 07)
152 punpcklwd xmm5, xmm5 ; xmm5=(00 00 01 01 02 02 03 03)
153 punpckhwd xmm4, xmm4 ; xmm4=(04 04 05 05 06 06 07 07)
155 pshufd xmm7, xmm5, 0x00 ; xmm7=col0=(00 00 00 00 00 00 00 00)
156 pshufd xmm6, xmm5, 0x55 ; xmm6=col1=(01 01 01 01 01 01 01 01)
157 pshufd xmm1, xmm5, 0xAA ; xmm1=col2=(02 02 02 02 02 02 02 02)
158 pshufd xmm5, xmm5, 0xFF ; xmm5=col3=(03 03 03 03 03 03 03 03)
159 pshufd xmm0, xmm4, 0x00 ; xmm0=col4=(04 04 04 04 04 04 04 04)
160 pshufd xmm3, xmm4, 0x55 ; xmm3=col5=(05 05 05 05 05 05 05 05)
161 pshufd xmm2, xmm4, 0xAA ; xmm2=col6=(06 06 06 06 06 06 06 06)
162 pshufd xmm4, xmm4, 0xFF ; xmm4=col7=(07 07 07 07 07 07 07 07)
164 movdqa XMMWORD [wk(8)], xmm6 ; wk(8)=col1
165 movdqa XMMWORD [wk(9)], xmm5 ; wk(9)=col3
166 movdqa XMMWORD [wk(10)], xmm3 ; wk(10)=col5
167 movdqa XMMWORD [wk(11)], xmm4 ; wk(11)=col7
174 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
175 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
176 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
177 pmullw xmm1, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
178 movdqa xmm2, XMMWORD [XMMBLOCK(4,0,rsi,SIZEOF_JCOEF)]
179 movdqa xmm3, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
180 pmullw xmm2, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
181 pmullw xmm3, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
184 ; z1 = (z2 + z3) * 0.541196100;
185 ; tmp2 = z1 + z3 * -1.847759065;
186 ; tmp3 = z1 + z2 * 0.765366865;
188 ; (This implementation)
189 ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
190 ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
192 movdqa xmm4, xmm1 ; xmm1=in2=z2
194 punpcklwd xmm4, xmm3 ; xmm3=in6=z3
198 pmaddwd xmm4, [rel PW_F130_F054] ; xmm4=tmp3L
199 pmaddwd xmm5, [rel PW_F130_F054] ; xmm5=tmp3H
200 pmaddwd xmm1, [rel PW_F054_MF130] ; xmm1=tmp2L
201 pmaddwd xmm3, [rel PW_F054_MF130] ; xmm3=tmp2H
204 paddw xmm0, xmm2 ; xmm0=in0+in4
205 psubw xmm6, xmm2 ; xmm6=in0-in4
209 punpcklwd xmm7, xmm0 ; xmm7=tmp0L
210 punpckhwd xmm2, xmm0 ; xmm2=tmp0H
211 psrad xmm7, (16-CONST_BITS) ; psrad xmm7,16 & pslld xmm7,CONST_BITS
212 psrad xmm2, (16-CONST_BITS) ; psrad xmm2,16 & pslld xmm2,CONST_BITS
215 paddd xmm7, xmm4 ; xmm7=tmp10L
216 psubd xmm0, xmm4 ; xmm0=tmp13L
218 paddd xmm2, xmm5 ; xmm2=tmp10H
219 psubd xmm4, xmm5 ; xmm4=tmp13H
221 movdqa XMMWORD [wk(0)], xmm7 ; wk(0)=tmp10L
222 movdqa XMMWORD [wk(1)], xmm2 ; wk(1)=tmp10H
223 movdqa XMMWORD [wk(2)], xmm0 ; wk(2)=tmp13L
224 movdqa XMMWORD [wk(3)], xmm4 ; wk(3)=tmp13H
228 punpcklwd xmm5, xmm6 ; xmm5=tmp1L
229 punpckhwd xmm7, xmm6 ; xmm7=tmp1H
230 psrad xmm5, (16-CONST_BITS) ; psrad xmm5,16 & pslld xmm5,CONST_BITS
231 psrad xmm7, (16-CONST_BITS) ; psrad xmm7,16 & pslld xmm7,CONST_BITS
234 paddd xmm5, xmm1 ; xmm5=tmp11L
235 psubd xmm2, xmm1 ; xmm2=tmp12L
237 paddd xmm7, xmm3 ; xmm7=tmp11H
238 psubd xmm0, xmm3 ; xmm0=tmp12H
240 movdqa XMMWORD [wk(4)], xmm5 ; wk(4)=tmp11L
241 movdqa XMMWORD [wk(5)], xmm7 ; wk(5)=tmp11H
242 movdqa XMMWORD [wk(6)], xmm2 ; wk(6)=tmp12L
243 movdqa XMMWORD [wk(7)], xmm0 ; wk(7)=tmp12H
247 movdqa xmm4, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
248 movdqa xmm6, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
249 pmullw xmm4, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
250 pmullw xmm6, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
251 movdqa xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
252 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
253 pmullw xmm1, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
254 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
258 paddw xmm5, xmm3 ; xmm5=z3
259 paddw xmm7, xmm1 ; xmm7=z4
262 ; z5 = (z3 + z4) * 1.175875602;
263 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
264 ; z3 += z5; z4 += z5;
266 ; (This implementation)
267 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
268 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
276 pmaddwd xmm2, [rel PW_MF078_F117] ; xmm2=z3L
277 pmaddwd xmm0, [rel PW_MF078_F117] ; xmm0=z3H
278 pmaddwd xmm5, [rel PW_F117_F078] ; xmm5=z4L
279 pmaddwd xmm7, [rel PW_F117_F078] ; xmm7=z4H
281 movdqa XMMWORD [wk(10)], xmm2 ; wk(10)=z3L
282 movdqa XMMWORD [wk(11)], xmm0 ; wk(11)=z3H
285 ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
286 ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
287 ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
288 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
289 ; tmp0 += z1 + z3; tmp1 += z2 + z4;
290 ; tmp2 += z2 + z3; tmp3 += z1 + z4;
292 ; (This implementation)
293 ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
294 ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
295 ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
296 ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
297 ; tmp0 += z3; tmp1 += z4;
298 ; tmp2 += z3; tmp3 += z4;
306 pmaddwd xmm2, [rel PW_MF060_MF089] ; xmm2=tmp0L
307 pmaddwd xmm0, [rel PW_MF060_MF089] ; xmm0=tmp0H
308 pmaddwd xmm3, [rel PW_MF089_F060] ; xmm3=tmp3L
309 pmaddwd xmm4, [rel PW_MF089_F060] ; xmm4=tmp3H
311 paddd xmm2, XMMWORD [wk(10)] ; xmm2=tmp0L
312 paddd xmm0, XMMWORD [wk(11)] ; xmm0=tmp0H
313 paddd xmm3, xmm5 ; xmm3=tmp3L
314 paddd xmm4, xmm7 ; xmm4=tmp3H
316 movdqa XMMWORD [wk(8)], xmm2 ; wk(8)=tmp0L
317 movdqa XMMWORD [wk(9)], xmm0 ; wk(9)=tmp0H
325 pmaddwd xmm2, [rel PW_MF050_MF256] ; xmm2=tmp1L
326 pmaddwd xmm0, [rel PW_MF050_MF256] ; xmm0=tmp1H
327 pmaddwd xmm1, [rel PW_MF256_F050] ; xmm1=tmp2L
328 pmaddwd xmm6, [rel PW_MF256_F050] ; xmm6=tmp2H
330 paddd xmm2, xmm5 ; xmm2=tmp1L
331 paddd xmm0, xmm7 ; xmm0=tmp1H
332 paddd xmm1, XMMWORD [wk(10)] ; xmm1=tmp2L
333 paddd xmm6, XMMWORD [wk(11)] ; xmm6=tmp2H
335 movdqa XMMWORD [wk(10)], xmm2 ; wk(10)=tmp1L
336 movdqa XMMWORD [wk(11)], xmm0 ; wk(11)=tmp1H
338 ; -- Final output stage
340 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=tmp10L
341 movdqa xmm7, XMMWORD [wk(1)] ; xmm7=tmp10H
345 paddd xmm5, xmm3 ; xmm5=data0L
346 paddd xmm7, xmm4 ; xmm7=data0H
347 psubd xmm2, xmm3 ; xmm2=data7L
348 psubd xmm0, xmm4 ; xmm0=data7H
350 movdqa xmm3, [rel PD_DESCALE_P1] ; xmm3=[rel PD_DESCALE_P1]
354 psrad xmm5, DESCALE_P1
355 psrad xmm7, DESCALE_P1
358 psrad xmm2, DESCALE_P1
359 psrad xmm0, DESCALE_P1
361 packssdw xmm5, xmm7 ; xmm5=data0=(00 01 02 03 04 05 06 07)
362 packssdw xmm2, xmm0 ; xmm2=data7=(70 71 72 73 74 75 76 77)
364 movdqa xmm4, XMMWORD [wk(4)] ; xmm4=tmp11L
365 movdqa xmm3, XMMWORD [wk(5)] ; xmm3=tmp11H
369 paddd xmm4, xmm1 ; xmm4=data1L
370 paddd xmm3, xmm6 ; xmm3=data1H
371 psubd xmm7, xmm1 ; xmm7=data6L
372 psubd xmm0, xmm6 ; xmm0=data6H
374 movdqa xmm1, [rel PD_DESCALE_P1] ; xmm1=[rel PD_DESCALE_P1]
378 psrad xmm4, DESCALE_P1
379 psrad xmm3, DESCALE_P1
382 psrad xmm7, DESCALE_P1
383 psrad xmm0, DESCALE_P1
385 packssdw xmm4, xmm3 ; xmm4=data1=(10 11 12 13 14 15 16 17)
386 packssdw xmm7, xmm0 ; xmm7=data6=(60 61 62 63 64 65 66 67)
388 movdqa xmm6, xmm5 ; transpose coefficients(phase 1)
389 punpcklwd xmm5, xmm4 ; xmm5=(00 10 01 11 02 12 03 13)
390 punpckhwd xmm6, xmm4 ; xmm6=(04 14 05 15 06 16 07 17)
391 movdqa xmm1, xmm7 ; transpose coefficients(phase 1)
392 punpcklwd xmm7, xmm2 ; xmm7=(60 70 61 71 62 72 63 73)
393 punpckhwd xmm1, xmm2 ; xmm1=(64 74 65 75 66 76 67 77)
395 movdqa xmm3, XMMWORD [wk(6)] ; xmm3=tmp12L
396 movdqa xmm0, XMMWORD [wk(7)] ; xmm0=tmp12H
397 movdqa xmm4, XMMWORD [wk(10)] ; xmm4=tmp1L
398 movdqa xmm2, XMMWORD [wk(11)] ; xmm2=tmp1H
400 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(00 10 01 11 02 12 03 13)
401 movdqa XMMWORD [wk(1)], xmm6 ; wk(1)=(04 14 05 15 06 16 07 17)
402 movdqa XMMWORD [wk(4)], xmm7 ; wk(4)=(60 70 61 71 62 72 63 73)
403 movdqa XMMWORD [wk(5)], xmm1 ; wk(5)=(64 74 65 75 66 76 67 77)
407 paddd xmm3, xmm4 ; xmm3=data2L
408 paddd xmm0, xmm2 ; xmm0=data2H
409 psubd xmm5, xmm4 ; xmm5=data5L
410 psubd xmm6, xmm2 ; xmm6=data5H
412 movdqa xmm7, [rel PD_DESCALE_P1] ; xmm7=[rel PD_DESCALE_P1]
416 psrad xmm3, DESCALE_P1
417 psrad xmm0, DESCALE_P1
420 psrad xmm5, DESCALE_P1
421 psrad xmm6, DESCALE_P1
423 packssdw xmm3, xmm0 ; xmm3=data2=(20 21 22 23 24 25 26 27)
424 packssdw xmm5, xmm6 ; xmm5=data5=(50 51 52 53 54 55 56 57)
426 movdqa xmm1, XMMWORD [wk(2)] ; xmm1=tmp13L
427 movdqa xmm4, XMMWORD [wk(3)] ; xmm4=tmp13H
428 movdqa xmm2, XMMWORD [wk(8)] ; xmm2=tmp0L
429 movdqa xmm7, XMMWORD [wk(9)] ; xmm7=tmp0H
433 paddd xmm1, xmm2 ; xmm1=data3L
434 paddd xmm4, xmm7 ; xmm4=data3H
435 psubd xmm0, xmm2 ; xmm0=data4L
436 psubd xmm6, xmm7 ; xmm6=data4H
438 movdqa xmm2, [rel PD_DESCALE_P1] ; xmm2=[rel PD_DESCALE_P1]
442 psrad xmm1, DESCALE_P1
443 psrad xmm4, DESCALE_P1
446 psrad xmm0, DESCALE_P1
447 psrad xmm6, DESCALE_P1
449 packssdw xmm1, xmm4 ; xmm1=data3=(30 31 32 33 34 35 36 37)
450 packssdw xmm0, xmm6 ; xmm0=data4=(40 41 42 43 44 45 46 47)
452 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(00 10 01 11 02 12 03 13)
453 movdqa xmm2, XMMWORD [wk(1)] ; xmm2=(04 14 05 15 06 16 07 17)
455 movdqa xmm4, xmm3 ; transpose coefficients(phase 1)
456 punpcklwd xmm3, xmm1 ; xmm3=(20 30 21 31 22 32 23 33)
457 punpckhwd xmm4, xmm1 ; xmm4=(24 34 25 35 26 36 27 37)
458 movdqa xmm6, xmm0 ; transpose coefficients(phase 1)
459 punpcklwd xmm0, xmm5 ; xmm0=(40 50 41 51 42 52 43 53)
460 punpckhwd xmm6, xmm5 ; xmm6=(44 54 45 55 46 56 47 57)
462 movdqa xmm1, xmm7 ; transpose coefficients(phase 2)
463 punpckldq xmm7, xmm3 ; xmm7=(00 10 20 30 01 11 21 31)
464 punpckhdq xmm1, xmm3 ; xmm1=(02 12 22 32 03 13 23 33)
465 movdqa xmm5, xmm2 ; transpose coefficients(phase 2)
466 punpckldq xmm2, xmm4 ; xmm2=(04 14 24 34 05 15 25 35)
467 punpckhdq xmm5, xmm4 ; xmm5=(06 16 26 36 07 17 27 37)
469 movdqa xmm3, XMMWORD [wk(4)] ; xmm3=(60 70 61 71 62 72 63 73)
470 movdqa xmm4, XMMWORD [wk(5)] ; xmm4=(64 74 65 75 66 76 67 77)
472 movdqa XMMWORD [wk(6)], xmm2 ; wk(6)=(04 14 24 34 05 15 25 35)
473 movdqa XMMWORD [wk(7)], xmm5 ; wk(7)=(06 16 26 36 07 17 27 37)
475 movdqa xmm2, xmm0 ; transpose coefficients(phase 2)
476 punpckldq xmm0, xmm3 ; xmm0=(40 50 60 70 41 51 61 71)
477 punpckhdq xmm2, xmm3 ; xmm2=(42 52 62 72 43 53 63 73)
478 movdqa xmm5, xmm6 ; transpose coefficients(phase 2)
479 punpckldq xmm6, xmm4 ; xmm6=(44 54 64 74 45 55 65 75)
480 punpckhdq xmm5, xmm4 ; xmm5=(46 56 66 76 47 57 67 77)
482 movdqa xmm3, xmm7 ; transpose coefficients(phase 3)
483 punpcklqdq xmm7, xmm0 ; xmm7=col0=(00 10 20 30 40 50 60 70)
484 punpckhqdq xmm3, xmm0 ; xmm3=col1=(01 11 21 31 41 51 61 71)
485 movdqa xmm4, xmm1 ; transpose coefficients(phase 3)
486 punpcklqdq xmm1, xmm2 ; xmm1=col2=(02 12 22 32 42 52 62 72)
487 punpckhqdq xmm4, xmm2 ; xmm4=col3=(03 13 23 33 43 53 63 73)
489 movdqa xmm0, XMMWORD [wk(6)] ; xmm0=(04 14 24 34 05 15 25 35)
490 movdqa xmm2, XMMWORD [wk(7)] ; xmm2=(06 16 26 36 07 17 27 37)
492 movdqa XMMWORD [wk(8)], xmm3 ; wk(8)=col1
493 movdqa XMMWORD [wk(9)], xmm4 ; wk(9)=col3
495 movdqa xmm3, xmm0 ; transpose coefficients(phase 3)
496 punpcklqdq xmm0, xmm6 ; xmm0=col4=(04 14 24 34 44 54 64 74)
497 punpckhqdq xmm3, xmm6 ; xmm3=col5=(05 15 25 35 45 55 65 75)
498 movdqa xmm4, xmm2 ; transpose coefficients(phase 3)
499 punpcklqdq xmm2, xmm5 ; xmm2=col6=(06 16 26 36 46 56 66 76)
500 punpckhqdq xmm4, xmm5 ; xmm4=col7=(07 17 27 37 47 57 67 77)
502 movdqa XMMWORD [wk(10)], xmm3 ; wk(10)=col5
503 movdqa XMMWORD [wk(11)], xmm4 ; wk(11)=col7
506 ; -- Prefetch the next coefficient block
508 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
509 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
510 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
511 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
513 ; ---- Pass 2: process rows from work array, store into output array.
515 mov rdi, r12 ; (JSAMPROW *)
520 ; xmm7=col0, xmm1=col2, xmm0=col4, xmm2=col6
523 ; z1 = (z2 + z3) * 0.541196100;
524 ; tmp2 = z1 + z3 * -1.847759065;
525 ; tmp3 = z1 + z2 * 0.765366865;
527 ; (This implementation)
528 ; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
529 ; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
531 movdqa xmm6, xmm1 ; xmm1=in2=z2
533 punpcklwd xmm6, xmm2 ; xmm2=in6=z3
537 pmaddwd xmm6, [rel PW_F130_F054] ; xmm6=tmp3L
538 pmaddwd xmm5, [rel PW_F130_F054] ; xmm5=tmp3H
539 pmaddwd xmm1, [rel PW_F054_MF130] ; xmm1=tmp2L
540 pmaddwd xmm2, [rel PW_F054_MF130] ; xmm2=tmp2H
543 paddw xmm7, xmm0 ; xmm7=in0+in4
544 psubw xmm3, xmm0 ; xmm3=in0-in4
548 punpcklwd xmm4, xmm7 ; xmm4=tmp0L
549 punpckhwd xmm0, xmm7 ; xmm0=tmp0H
550 psrad xmm4, (16-CONST_BITS) ; psrad xmm4,16 & pslld xmm4,CONST_BITS
551 psrad xmm0, (16-CONST_BITS) ; psrad xmm0,16 & pslld xmm0,CONST_BITS
554 paddd xmm4, xmm6 ; xmm4=tmp10L
555 psubd xmm7, xmm6 ; xmm7=tmp13L
557 paddd xmm0, xmm5 ; xmm0=tmp10H
558 psubd xmm6, xmm5 ; xmm6=tmp13H
560 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=tmp10L
561 movdqa XMMWORD [wk(1)], xmm0 ; wk(1)=tmp10H
562 movdqa XMMWORD [wk(2)], xmm7 ; wk(2)=tmp13L
563 movdqa XMMWORD [wk(3)], xmm6 ; wk(3)=tmp13H
567 punpcklwd xmm5, xmm3 ; xmm5=tmp1L
568 punpckhwd xmm4, xmm3 ; xmm4=tmp1H
569 psrad xmm5, (16-CONST_BITS) ; psrad xmm5,16 & pslld xmm5,CONST_BITS
570 psrad xmm4, (16-CONST_BITS) ; psrad xmm4,16 & pslld xmm4,CONST_BITS
573 paddd xmm5, xmm1 ; xmm5=tmp11L
574 psubd xmm0, xmm1 ; xmm0=tmp12L
576 paddd xmm4, xmm2 ; xmm4=tmp11H
577 psubd xmm7, xmm2 ; xmm7=tmp12H
579 movdqa XMMWORD [wk(4)], xmm5 ; wk(4)=tmp11L
580 movdqa XMMWORD [wk(5)], xmm4 ; wk(5)=tmp11H
581 movdqa XMMWORD [wk(6)], xmm0 ; wk(6)=tmp12L
582 movdqa XMMWORD [wk(7)], xmm7 ; wk(7)=tmp12H
586 movdqa xmm6, XMMWORD [wk(9)] ; xmm6=col3
587 movdqa xmm3, XMMWORD [wk(8)] ; xmm3=col1
588 movdqa xmm1, XMMWORD [wk(11)] ; xmm1=col7
589 movdqa xmm2, XMMWORD [wk(10)] ; xmm2=col5
593 paddw xmm5, xmm1 ; xmm5=z3
594 paddw xmm4, xmm2 ; xmm4=z4
597 ; z5 = (z3 + z4) * 1.175875602;
598 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
599 ; z3 += z5; z4 += z5;
601 ; (This implementation)
602 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
603 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
611 pmaddwd xmm0, [rel PW_MF078_F117] ; xmm0=z3L
612 pmaddwd xmm7, [rel PW_MF078_F117] ; xmm7=z3H
613 pmaddwd xmm5, [rel PW_F117_F078] ; xmm5=z4L
614 pmaddwd xmm4, [rel PW_F117_F078] ; xmm4=z4H
616 movdqa XMMWORD [wk(10)], xmm0 ; wk(10)=z3L
617 movdqa XMMWORD [wk(11)], xmm7 ; wk(11)=z3H
620 ; z1 = tmp0 + tmp3; z2 = tmp1 + tmp2;
621 ; tmp0 = tmp0 * 0.298631336; tmp1 = tmp1 * 2.053119869;
622 ; tmp2 = tmp2 * 3.072711026; tmp3 = tmp3 * 1.501321110;
623 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
624 ; tmp0 += z1 + z3; tmp1 += z2 + z4;
625 ; tmp2 += z2 + z3; tmp3 += z1 + z4;
627 ; (This implementation)
628 ; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
629 ; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
630 ; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
631 ; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
632 ; tmp0 += z3; tmp1 += z4;
633 ; tmp2 += z3; tmp3 += z4;
641 pmaddwd xmm0, [rel PW_MF060_MF089] ; xmm0=tmp0L
642 pmaddwd xmm7, [rel PW_MF060_MF089] ; xmm7=tmp0H
643 pmaddwd xmm1, [rel PW_MF089_F060] ; xmm1=tmp3L
644 pmaddwd xmm3, [rel PW_MF089_F060] ; xmm3=tmp3H
646 paddd xmm0, XMMWORD [wk(10)] ; xmm0=tmp0L
647 paddd xmm7, XMMWORD [wk(11)] ; xmm7=tmp0H
648 paddd xmm1, xmm5 ; xmm1=tmp3L
649 paddd xmm3, xmm4 ; xmm3=tmp3H
651 movdqa XMMWORD [wk(8)], xmm0 ; wk(8)=tmp0L
652 movdqa XMMWORD [wk(9)], xmm7 ; wk(9)=tmp0H
660 pmaddwd xmm0, [rel PW_MF050_MF256] ; xmm0=tmp1L
661 pmaddwd xmm7, [rel PW_MF050_MF256] ; xmm7=tmp1H
662 pmaddwd xmm2, [rel PW_MF256_F050] ; xmm2=tmp2L
663 pmaddwd xmm6, [rel PW_MF256_F050] ; xmm6=tmp2H
665 paddd xmm0, xmm5 ; xmm0=tmp1L
666 paddd xmm7, xmm4 ; xmm7=tmp1H
667 paddd xmm2, XMMWORD [wk(10)] ; xmm2=tmp2L
668 paddd xmm6, XMMWORD [wk(11)] ; xmm6=tmp2H
670 movdqa XMMWORD [wk(10)], xmm0 ; wk(10)=tmp1L
671 movdqa XMMWORD [wk(11)], xmm7 ; wk(11)=tmp1H
673 ; -- Final output stage
675 movdqa xmm5, XMMWORD [wk(0)] ; xmm5=tmp10L
676 movdqa xmm4, XMMWORD [wk(1)] ; xmm4=tmp10H
680 paddd xmm5, xmm1 ; xmm5=data0L
681 paddd xmm4, xmm3 ; xmm4=data0H
682 psubd xmm0, xmm1 ; xmm0=data7L
683 psubd xmm7, xmm3 ; xmm7=data7H
685 movdqa xmm1, [rel PD_DESCALE_P2] ; xmm1=[rel PD_DESCALE_P2]
689 psrad xmm5, DESCALE_P2
690 psrad xmm4, DESCALE_P2
693 psrad xmm0, DESCALE_P2
694 psrad xmm7, DESCALE_P2
696 packssdw xmm5, xmm4 ; xmm5=data0=(00 10 20 30 40 50 60 70)
697 packssdw xmm0, xmm7 ; xmm0=data7=(07 17 27 37 47 57 67 77)
699 movdqa xmm3, XMMWORD [wk(4)] ; xmm3=tmp11L
700 movdqa xmm1, XMMWORD [wk(5)] ; xmm1=tmp11H
704 paddd xmm3, xmm2 ; xmm3=data1L
705 paddd xmm1, xmm6 ; xmm1=data1H
706 psubd xmm4, xmm2 ; xmm4=data6L
707 psubd xmm7, xmm6 ; xmm7=data6H
709 movdqa xmm2, [rel PD_DESCALE_P2] ; xmm2=[rel PD_DESCALE_P2]
713 psrad xmm3, DESCALE_P2
714 psrad xmm1, DESCALE_P2
717 psrad xmm4, DESCALE_P2
718 psrad xmm7, DESCALE_P2
720 packssdw xmm3, xmm1 ; xmm3=data1=(01 11 21 31 41 51 61 71)
721 packssdw xmm4, xmm7 ; xmm4=data6=(06 16 26 36 46 56 66 76)
723 packsswb xmm5, xmm4 ; xmm5=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
724 packsswb xmm3, xmm0 ; xmm3=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
726 movdqa xmm6, XMMWORD [wk(6)] ; xmm6=tmp12L
727 movdqa xmm2, XMMWORD [wk(7)] ; xmm2=tmp12H
728 movdqa xmm1, XMMWORD [wk(10)] ; xmm1=tmp1L
729 movdqa xmm7, XMMWORD [wk(11)] ; xmm7=tmp1H
731 movdqa XMMWORD [wk(0)], xmm5 ; wk(0)=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
732 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
736 paddd xmm6, xmm1 ; xmm6=data2L
737 paddd xmm2, xmm7 ; xmm2=data2H
738 psubd xmm4, xmm1 ; xmm4=data5L
739 psubd xmm0, xmm7 ; xmm0=data5H
741 movdqa xmm5, [rel PD_DESCALE_P2] ; xmm5=[rel PD_DESCALE_P2]
745 psrad xmm6, DESCALE_P2
746 psrad xmm2, DESCALE_P2
749 psrad xmm4, DESCALE_P2
750 psrad xmm0, DESCALE_P2
752 packssdw xmm6, xmm2 ; xmm6=data2=(02 12 22 32 42 52 62 72)
753 packssdw xmm4, xmm0 ; xmm4=data5=(05 15 25 35 45 55 65 75)
755 movdqa xmm3, XMMWORD [wk(2)] ; xmm3=tmp13L
756 movdqa xmm1, XMMWORD [wk(3)] ; xmm1=tmp13H
757 movdqa xmm7, XMMWORD [wk(8)] ; xmm7=tmp0L
758 movdqa xmm5, XMMWORD [wk(9)] ; xmm5=tmp0H
762 paddd xmm3, xmm7 ; xmm3=data3L
763 paddd xmm1, xmm5 ; xmm1=data3H
764 psubd xmm2, xmm7 ; xmm2=data4L
765 psubd xmm0, xmm5 ; xmm0=data4H
767 movdqa xmm7, [rel PD_DESCALE_P2] ; xmm7=[rel PD_DESCALE_P2]
771 psrad xmm3, DESCALE_P2
772 psrad xmm1, DESCALE_P2
775 psrad xmm2, DESCALE_P2
776 psrad xmm0, DESCALE_P2
778 movdqa xmm5, [rel PB_CENTERJSAMP] ; xmm5=[rel PB_CENTERJSAMP]
780 packssdw xmm3, xmm1 ; xmm3=data3=(03 13 23 33 43 53 63 73)
781 packssdw xmm2, xmm0 ; xmm2=data4=(04 14 24 34 44 54 64 74)
783 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=(00 10 20 30 40 50 60 70 06 16 26 36 46 56 66 76)
784 movdqa xmm1, XMMWORD [wk(1)] ; xmm1=(01 11 21 31 41 51 61 71 07 17 27 37 47 57 67 77)
786 packsswb xmm6, xmm2 ; xmm6=(02 12 22 32 42 52 62 72 04 14 24 34 44 54 64 74)
787 packsswb xmm3, xmm4 ; xmm3=(03 13 23 33 43 53 63 73 05 15 25 35 45 55 65 75)
794 movdqa xmm0, xmm7 ; transpose coefficients(phase 1)
795 punpcklbw xmm7, xmm1 ; xmm7=(00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71)
796 punpckhbw xmm0, xmm1 ; xmm0=(06 07 16 17 26 27 36 37 46 47 56 57 66 67 76 77)
797 movdqa xmm2, xmm6 ; transpose coefficients(phase 1)
798 punpcklbw xmm6, xmm3 ; xmm6=(02 03 12 13 22 23 32 33 42 43 52 53 62 63 72 73)
799 punpckhbw xmm2, xmm3 ; xmm2=(04 05 14 15 24 25 34 35 44 45 54 55 64 65 74 75)
801 movdqa xmm4, xmm7 ; transpose coefficients(phase 2)
802 punpcklwd xmm7, xmm6 ; xmm7=(00 01 02 03 10 11 12 13 20 21 22 23 30 31 32 33)
803 punpckhwd xmm4, xmm6 ; xmm4=(40 41 42 43 50 51 52 53 60 61 62 63 70 71 72 73)
804 movdqa xmm5, xmm2 ; transpose coefficients(phase 2)
805 punpcklwd xmm2, xmm0 ; xmm2=(04 05 06 07 14 15 16 17 24 25 26 27 34 35 36 37)
806 punpckhwd xmm5, xmm0 ; xmm5=(44 45 46 47 54 55 56 57 64 65 66 67 74 75 76 77)
808 movdqa xmm1, xmm7 ; transpose coefficients(phase 3)
809 punpckldq xmm7, xmm2 ; xmm7=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
810 punpckhdq xmm1, xmm2 ; xmm1=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
811 movdqa xmm3, xmm4 ; transpose coefficients(phase 3)
812 punpckldq xmm4, xmm5 ; xmm4=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
813 punpckhdq xmm3, xmm5 ; xmm3=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
815 pshufd xmm6, xmm7, 0x4E ; xmm6=(10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07)
816 pshufd xmm0, xmm1, 0x4E ; xmm0=(30 31 32 33 34 35 36 37 20 21 22 23 24 25 26 27)
817 pshufd xmm2, xmm4, 0x4E ; xmm2=(50 51 52 53 54 55 56 57 40 41 42 43 44 45 46 47)
818 pshufd xmm5, xmm3, 0x4E ; xmm5=(70 71 72 73 74 75 76 77 60 61 62 63 64 65 66 67)
820 mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
821 mov rsip, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
822 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm7
823 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm1
824 mov rdxp, JSAMPROW [rdi+4*SIZEOF_JSAMPROW]
825 mov rsip, JSAMPROW [rdi+6*SIZEOF_JSAMPROW]
826 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
827 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
829 mov rdxp, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
830 mov rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
831 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm6
832 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm0
833 mov rdxp, JSAMPROW [rdi+5*SIZEOF_JSAMPROW]
834 mov rsip, JSAMPROW [rdi+7*SIZEOF_JSAMPROW]
835 movq XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE], xmm2
836 movq XMM_MMWORD [rsi+rax*SIZEOF_JSAMPLE], xmm5
844 ; For some reason, the OS X linker does not honor the request to align the
845 ; segment unless we do this.