1;
2; jfdctint.asm - accurate integer FDCT (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; forward DCT (Discrete Cosine Transform). The following code is based
19; directly on the IJG's original jfdctint.c; see the jfdctint.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)
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 matrix transpose using AVX2 instructions
67; %1-%4: Input/output registers
68; %5-%8: Temp registers
69
70%macro dotranspose 8
71    ; %1=(00 01 02 03 04 05 06 07  40 41 42 43 44 45 46 47)
72    ; %2=(10 11 12 13 14 15 16 17  50 51 52 53 54 55 56 57)
73    ; %3=(20 21 22 23 24 25 26 27  60 61 62 63 64 65 66 67)
74    ; %4=(30 31 32 33 34 35 36 37  70 71 72 73 74 75 76 77)
75
76    vpunpcklwd  %5, %1, %2
77    vpunpckhwd  %6, %1, %2
78    vpunpcklwd  %7, %3, %4
79    vpunpckhwd  %8, %3, %4
80    ; transpose coefficients(phase 1)
81    ; %5=(00 10 01 11 02 12 03 13  40 50 41 51 42 52 43 53)
82    ; %6=(04 14 05 15 06 16 07 17  44 54 45 55 46 56 47 57)
83    ; %7=(20 30 21 31 22 32 23 33  60 70 61 71 62 72 63 73)
84    ; %8=(24 34 25 35 26 36 27 37  64 74 65 75 66 76 67 77)
85
86    vpunpckldq  %1, %5, %7
87    vpunpckhdq  %2, %5, %7
88    vpunpckldq  %3, %6, %8
89    vpunpckhdq  %4, %6, %8
90    ; transpose coefficients(phase 2)
91    ; %1=(00 10 20 30 01 11 21 31  40 50 60 70 41 51 61 71)
92    ; %2=(02 12 22 32 03 13 23 33  42 52 62 72 43 53 63 73)
93    ; %3=(04 14 24 34 05 15 25 35  44 54 64 74 45 55 65 75)
94    ; %4=(06 16 26 36 07 17 27 37  46 56 66 76 47 57 67 77)
95
96    vpermq      %1, %1, 0x8D
97    vpermq      %2, %2, 0x8D
98    vpermq      %3, %3, 0xD8
99    vpermq      %4, %4, 0xD8
100    ; transpose coefficients(phase 3)
101    ; %1=(01 11 21 31 41 51 61 71  00 10 20 30 40 50 60 70)
102    ; %2=(03 13 23 33 43 53 63 73  02 12 22 32 42 52 62 72)
103    ; %3=(04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75)
104    ; %4=(06 16 26 36 46 56 66 76  07 17 27 37 47 57 67 77)
105%endmacro
106
107; --------------------------------------------------------------------------
108; In-place 8x8x16-bit slow integer forward DCT using AVX2 instructions
109; %1-%4: Input/output registers
110; %5-%8: Temp registers
111; %9:    Pass (1 or 2)
112
113%macro dodct 9
114    vpsubw      %5, %1, %4              ; %5=data1_0-data6_7=tmp6_7
115    vpaddw      %6, %1, %4              ; %6=data1_0+data6_7=tmp1_0
116    vpaddw      %7, %2, %3              ; %7=data3_2+data4_5=tmp3_2
117    vpsubw      %8, %2, %3              ; %8=data3_2-data4_5=tmp4_5
118
119    ; -- Even part
120
121    vperm2i128  %6, %6, %6, 0x01        ; %6=tmp0_1
122    vpaddw      %1, %6, %7              ; %1=tmp0_1+tmp3_2=tmp10_11
123    vpsubw      %6, %6, %7              ; %6=tmp0_1-tmp3_2=tmp13_12
124
125    vperm2i128  %7, %1, %1, 0x01        ; %7=tmp11_10
126    vpsignw     %1, %1, [rel PW_1_NEG1]  ; %1=tmp10_neg11
127    vpaddw      %7, %7, %1              ; %7=(tmp10+tmp11)_(tmp10-tmp11)
128%if %9 == 1
129    vpsllw      %1, %7, PASS1_BITS      ; %1=data0_4
130%else
131    vpaddw      %7, %7, [rel PW_DESCALE_P2X]
132    vpsraw      %1, %7, PASS1_BITS      ; %1=data0_4
133%endif
134
135    ; (Original)
136    ; z1 = (tmp12 + tmp13) * 0.541196100;
137    ; data2 = z1 + tmp13 * 0.765366865;
138    ; data6 = z1 + tmp12 * -1.847759065;
139    ;
140    ; (This implementation)
141    ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
142    ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
143
144    vperm2i128  %7, %6, %6, 0x01        ; %7=tmp12_13
145    vpunpcklwd  %2, %6, %7
146    vpunpckhwd  %6, %6, %7
147    vpmaddwd    %2, %2, [rel PW_F130_F054_MF130_F054]  ; %2=data2_6L
148    vpmaddwd    %6, %6, [rel PW_F130_F054_MF130_F054]  ; %6=data2_6H
149
150    vpaddd      %2, %2, [rel PD_DESCALE_P %+ %9]
151    vpaddd      %6, %6, [rel PD_DESCALE_P %+ %9]
152    vpsrad      %2, %2, DESCALE_P %+ %9
153    vpsrad      %6, %6, DESCALE_P %+ %9
154
155    vpackssdw   %3, %2, %6              ; %6=data2_6
156
157    ; -- Odd part
158
159    vpaddw      %7, %8, %5              ; %7=tmp4_5+tmp6_7=z3_4
160
161    ; (Original)
162    ; z5 = (z3 + z4) * 1.175875602;
163    ; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
164    ; z3 += z5;  z4 += z5;
165    ;
166    ; (This implementation)
167    ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
168    ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
169
170    vperm2i128  %2, %7, %7, 0x01        ; %2=z4_3
171    vpunpcklwd  %6, %7, %2
172    vpunpckhwd  %7, %7, %2
173    vpmaddwd    %6, %6, [rel PW_MF078_F117_F078_F117]  ; %6=z3_4L
174    vpmaddwd    %7, %7, [rel PW_MF078_F117_F078_F117]  ; %7=z3_4H
175
176    ; (Original)
177    ; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
178    ; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
179    ; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
180    ; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
181    ; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
182    ; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
183    ;
184    ; (This implementation)
185    ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
186    ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
187    ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
188    ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
189    ; data7 = tmp4 + z3;  data5 = tmp5 + z4;
190    ; data3 = tmp6 + z3;  data1 = tmp7 + z4;
191
192    vperm2i128  %4, %5, %5, 0x01        ; %4=tmp7_6
193    vpunpcklwd  %2, %8, %4
194    vpunpckhwd  %4, %8, %4
195    vpmaddwd    %2, %2, [rel PW_MF060_MF089_MF050_MF256]  ; %2=tmp4_5L
196    vpmaddwd    %4, %4, [rel PW_MF060_MF089_MF050_MF256]  ; %4=tmp4_5H
197
198    vpaddd      %2, %2, %6              ; %2=data7_5L
199    vpaddd      %4, %4, %7              ; %4=data7_5H
200
201    vpaddd      %2, %2, [rel PD_DESCALE_P %+ %9]
202    vpaddd      %4, %4, [rel PD_DESCALE_P %+ %9]
203    vpsrad      %2, %2, DESCALE_P %+ %9
204    vpsrad      %4, %4, DESCALE_P %+ %9
205
206    vpackssdw   %4, %2, %4              ; %4=data7_5
207
208    vperm2i128  %2, %8, %8, 0x01        ; %2=tmp5_4
209    vpunpcklwd  %8, %5, %2
210    vpunpckhwd  %5, %5, %2
211    vpmaddwd    %8, %8, [rel PW_F050_MF256_F060_MF089]  ; %8=tmp6_7L
212    vpmaddwd    %5, %5, [rel PW_F050_MF256_F060_MF089]  ; %5=tmp6_7H
213
214    vpaddd      %8, %8, %6              ; %8=data3_1L
215    vpaddd      %5, %5, %7              ; %5=data3_1H
216
217    vpaddd      %8, %8, [rel PD_DESCALE_P %+ %9]
218    vpaddd      %5, %5, [rel PD_DESCALE_P %+ %9]
219    vpsrad      %8, %8, DESCALE_P %+ %9
220    vpsrad      %5, %5, DESCALE_P %+ %9
221
222    vpackssdw   %2, %8, %5              ; %2=data3_1
223%endmacro
224
225; --------------------------------------------------------------------------
226    SECTION     SEG_CONST
227
228    alignz      32
229    GLOBAL_DATA(jconst_fdct_islow_avx2)
230
231EXTN(jconst_fdct_islow_avx2):
232
233PW_F130_F054_MF130_F054    times 4  dw  (F_0_541 + F_0_765),  F_0_541
234                           times 4  dw  (F_0_541 - F_1_847),  F_0_541
235PW_MF078_F117_F078_F117    times 4  dw  (F_1_175 - F_1_961),  F_1_175
236                           times 4  dw  (F_1_175 - F_0_390),  F_1_175
237PW_MF060_MF089_MF050_MF256 times 4  dw  (F_0_298 - F_0_899), -F_0_899
238                           times 4  dw  (F_2_053 - F_2_562), -F_2_562
239PW_F050_MF256_F060_MF089   times 4  dw  (F_3_072 - F_2_562), -F_2_562
240                           times 4  dw  (F_1_501 - F_0_899), -F_0_899
241PD_DESCALE_P1              times 8  dd  1 << (DESCALE_P1 - 1)
242PD_DESCALE_P2              times 8  dd  1 << (DESCALE_P2 - 1)
243PW_DESCALE_P2X             times 16 dw  1 << (PASS1_BITS - 1)
244PW_1_NEG1                  times 8  dw  1
245                           times 8  dw -1
246
247    alignz      32
248
249; --------------------------------------------------------------------------
250    SECTION     SEG_TEXT
251    BITS        64
252;
253; Perform the forward DCT on one block of samples.
254;
255; GLOBAL(void)
256; jsimd_fdct_islow_avx2(DCTELEM *data)
257;
258
259; r10 = DCTELEM *data
260
261    align       32
262    GLOBAL_FUNCTION(jsimd_fdct_islow_avx2)
263
264EXTN(jsimd_fdct_islow_avx2):
265    push        rbp
266    mov         rax, rsp
267    mov         rbp, rsp
268    collect_args 1
269
270    ; ---- Pass 1: process rows.
271
272    vmovdqu     ymm4, YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)]
273    vmovdqu     ymm5, YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)]
274    vmovdqu     ymm6, YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)]
275    vmovdqu     ymm7, YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)]
276    ; ymm4=(00 01 02 03 04 05 06 07  10 11 12 13 14 15 16 17)
277    ; ymm5=(20 21 22 23 24 25 26 27  30 31 32 33 34 35 36 37)
278    ; ymm6=(40 41 42 43 44 45 46 47  50 51 52 53 54 55 56 57)
279    ; ymm7=(60 61 62 63 64 65 66 67  70 71 72 73 74 75 76 77)
280
281    vperm2i128  ymm0, ymm4, ymm6, 0x20
282    vperm2i128  ymm1, ymm4, ymm6, 0x31
283    vperm2i128  ymm2, ymm5, ymm7, 0x20
284    vperm2i128  ymm3, ymm5, ymm7, 0x31
285    ; ymm0=(00 01 02 03 04 05 06 07  40 41 42 43 44 45 46 47)
286    ; ymm1=(10 11 12 13 14 15 16 17  50 51 52 53 54 55 56 57)
287    ; ymm2=(20 21 22 23 24 25 26 27  60 61 62 63 64 65 66 67)
288    ; ymm3=(30 31 32 33 34 35 36 37  70 71 72 73 74 75 76 77)
289
290    dotranspose ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
291
292    dodct       ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, 1
293    ; ymm0=data0_4, ymm1=data3_1, ymm2=data2_6, ymm3=data7_5
294
295    ; ---- Pass 2: process columns.
296
297    vperm2i128  ymm4, ymm1, ymm3, 0x20  ; ymm4=data3_7
298    vperm2i128  ymm1, ymm1, ymm3, 0x31  ; ymm1=data1_5
299
300    dotranspose ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7
301
302    dodct       ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7, 2
303    ; ymm0=data0_4, ymm1=data3_1, ymm2=data2_6, ymm4=data7_5
304
305    vperm2i128 ymm3, ymm0, ymm1, 0x30   ; ymm3=data0_1
306    vperm2i128 ymm5, ymm2, ymm1, 0x20   ; ymm5=data2_3
307    vperm2i128 ymm6, ymm0, ymm4, 0x31   ; ymm6=data4_5
308    vperm2i128 ymm7, ymm2, ymm4, 0x21   ; ymm7=data6_7
309
310    vmovdqu     YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)], ymm3
311    vmovdqu     YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)], ymm5
312    vmovdqu     YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)], ymm6
313    vmovdqu     YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)], ymm7
314
315    vzeroupper
316    uncollect_args 1
317    pop         rbp
318    ret
319
320; For some reason, the OS X linker does not honor the request to align the
321; segment unless we do this.
322    align       32
323