2 ; jfdctint.asm - accurate integer FDCT (64-bit AVX2)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright (C) 2009, 2016, 2018, 2020, D. R. Commander.
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
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
17 ; This file contains a slower but more 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
22 %include "jsimdext.inc"
25 ; --------------------------------------------------------------------------
30 %define DESCALE_P1 (CONST_BITS - PASS1_BITS)
31 %define DESCALE_P2 (CONST_BITS + PASS1_BITS)
34 F_0_298 equ 2446 ; FIX(0.298631336)
35 F_0_390 equ 3196 ; FIX(0.390180644)
36 F_0_541 equ 4433 ; FIX(0.541196100)
37 F_0_765 equ 6270 ; FIX(0.765366865)
38 F_0_899 equ 7373 ; FIX(0.899976223)
39 F_1_175 equ 9633 ; FIX(1.175875602)
40 F_1_501 equ 12299 ; FIX(1.501321110)
41 F_1_847 equ 15137 ; FIX(1.847759065)
42 F_1_961 equ 16069 ; FIX(1.961570560)
43 F_2_053 equ 16819 ; FIX(2.053119869)
44 F_2_562 equ 20995 ; FIX(2.562915447)
45 F_3_072 equ 25172 ; FIX(3.072711026)
47 ; NASM cannot do compile-time arithmetic on floating-point constants.
48 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n))
49 F_0_298 equ DESCALE( 320652955, 30 - CONST_BITS) ; FIX(0.298631336)
50 F_0_390 equ DESCALE( 418953276, 30 - CONST_BITS) ; FIX(0.390180644)
51 F_0_541 equ DESCALE( 581104887, 30 - CONST_BITS) ; FIX(0.541196100)
52 F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS) ; FIX(0.765366865)
53 F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS) ; FIX(0.899976223)
54 F_1_175 equ DESCALE(1262586813, 30 - CONST_BITS) ; FIX(1.175875602)
55 F_1_501 equ DESCALE(1612031267, 30 - CONST_BITS) ; FIX(1.501321110)
56 F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065)
57 F_1_961 equ DESCALE(2106220350, 30 - CONST_BITS) ; FIX(1.961570560)
58 F_2_053 equ DESCALE(2204520673, 30 - CONST_BITS) ; FIX(2.053119869)
59 F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS) ; FIX(2.562915447)
60 F_3_072 equ DESCALE(3299298341, 30 - CONST_BITS) ; FIX(3.072711026)
63 ; --------------------------------------------------------------------------
64 ; In-place 8x8x16-bit matrix transpose using AVX2 instructions
65 ; %1-%4: Input/output registers
66 ; %5-%8: Temp registers
69 ; %1=(00 01 02 03 04 05 06 07 40 41 42 43 44 45 46 47)
70 ; %2=(10 11 12 13 14 15 16 17 50 51 52 53 54 55 56 57)
71 ; %3=(20 21 22 23 24 25 26 27 60 61 62 63 64 65 66 67)
72 ; %4=(30 31 32 33 34 35 36 37 70 71 72 73 74 75 76 77)
78 ; transpose coefficients(phase 1)
79 ; %5=(00 10 01 11 02 12 03 13 40 50 41 51 42 52 43 53)
80 ; %6=(04 14 05 15 06 16 07 17 44 54 45 55 46 56 47 57)
81 ; %7=(20 30 21 31 22 32 23 33 60 70 61 71 62 72 63 73)
82 ; %8=(24 34 25 35 26 36 27 37 64 74 65 75 66 76 67 77)
88 ; transpose coefficients(phase 2)
89 ; %1=(00 10 20 30 01 11 21 31 40 50 60 70 41 51 61 71)
90 ; %2=(02 12 22 32 03 13 23 33 42 52 62 72 43 53 63 73)
91 ; %3=(04 14 24 34 05 15 25 35 44 54 64 74 45 55 65 75)
92 ; %4=(06 16 26 36 07 17 27 37 46 56 66 76 47 57 67 77)
98 ; transpose coefficients(phase 3)
99 ; %1=(01 11 21 31 41 51 61 71 00 10 20 30 40 50 60 70)
100 ; %2=(03 13 23 33 43 53 63 73 02 12 22 32 42 52 62 72)
101 ; %3=(04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75)
102 ; %4=(06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77)
105 ; --------------------------------------------------------------------------
106 ; In-place 8x8x16-bit accurate integer forward DCT using AVX2 instructions
107 ; %1-%4: Input/output registers
108 ; %5-%8: Temp registers
112 vpsubw %5, %1, %4 ; %5=data1_0-data6_7=tmp6_7
113 vpaddw %6, %1, %4 ; %6=data1_0+data6_7=tmp1_0
114 vpaddw %7, %2, %3 ; %7=data3_2+data4_5=tmp3_2
115 vpsubw %8, %2, %3 ; %8=data3_2-data4_5=tmp4_5
119 vperm2i128 %6, %6, %6, 0x01 ; %6=tmp0_1
120 vpaddw %1, %6, %7 ; %1=tmp0_1+tmp3_2=tmp10_11
121 vpsubw %6, %6, %7 ; %6=tmp0_1-tmp3_2=tmp13_12
123 vperm2i128 %7, %1, %1, 0x01 ; %7=tmp11_10
124 vpsignw %1, %1, [rel PW_1_NEG1] ; %1=tmp10_neg11
125 vpaddw %7, %7, %1 ; %7=(tmp10+tmp11)_(tmp10-tmp11)
127 vpsllw %1, %7, PASS1_BITS ; %1=data0_4
129 vpaddw %7, %7, [rel PW_DESCALE_P2X]
130 vpsraw %1, %7, PASS1_BITS ; %1=data0_4
134 ; z1 = (tmp12 + tmp13) * 0.541196100;
135 ; data2 = z1 + tmp13 * 0.765366865;
136 ; data6 = z1 + tmp12 * -1.847759065;
138 ; (This implementation)
139 ; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
140 ; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
142 vperm2i128 %7, %6, %6, 0x01 ; %7=tmp12_13
143 vpunpcklwd %2, %6, %7
144 vpunpckhwd %6, %6, %7
145 vpmaddwd %2, %2, [rel PW_F130_F054_MF130_F054] ; %2=data2_6L
146 vpmaddwd %6, %6, [rel PW_F130_F054_MF130_F054] ; %6=data2_6H
148 vpaddd %2, %2, [rel PD_DESCALE_P %+ %9]
149 vpaddd %6, %6, [rel PD_DESCALE_P %+ %9]
150 vpsrad %2, %2, DESCALE_P %+ %9
151 vpsrad %6, %6, DESCALE_P %+ %9
153 vpackssdw %3, %2, %6 ; %6=data2_6
157 vpaddw %7, %8, %5 ; %7=tmp4_5+tmp6_7=z3_4
160 ; z5 = (z3 + z4) * 1.175875602;
161 ; z3 = z3 * -1.961570560; z4 = z4 * -0.390180644;
162 ; z3 += z5; z4 += z5;
164 ; (This implementation)
165 ; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
166 ; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
168 vperm2i128 %2, %7, %7, 0x01 ; %2=z4_3
169 vpunpcklwd %6, %7, %2
170 vpunpckhwd %7, %7, %2
171 vpmaddwd %6, %6, [rel PW_MF078_F117_F078_F117] ; %6=z3_4L
172 vpmaddwd %7, %7, [rel PW_MF078_F117_F078_F117] ; %7=z3_4H
175 ; z1 = tmp4 + tmp7; z2 = tmp5 + tmp6;
176 ; tmp4 = tmp4 * 0.298631336; tmp5 = tmp5 * 2.053119869;
177 ; tmp6 = tmp6 * 3.072711026; tmp7 = tmp7 * 1.501321110;
178 ; z1 = z1 * -0.899976223; z2 = z2 * -2.562915447;
179 ; data7 = tmp4 + z1 + z3; data5 = tmp5 + z2 + z4;
180 ; data3 = tmp6 + z2 + z3; data1 = tmp7 + z1 + z4;
182 ; (This implementation)
183 ; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
184 ; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
185 ; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
186 ; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
187 ; data7 = tmp4 + z3; data5 = tmp5 + z4;
188 ; data3 = tmp6 + z3; data1 = tmp7 + z4;
190 vperm2i128 %4, %5, %5, 0x01 ; %4=tmp7_6
191 vpunpcklwd %2, %8, %4
192 vpunpckhwd %4, %8, %4
193 vpmaddwd %2, %2, [rel PW_MF060_MF089_MF050_MF256] ; %2=tmp4_5L
194 vpmaddwd %4, %4, [rel PW_MF060_MF089_MF050_MF256] ; %4=tmp4_5H
196 vpaddd %2, %2, %6 ; %2=data7_5L
197 vpaddd %4, %4, %7 ; %4=data7_5H
199 vpaddd %2, %2, [rel PD_DESCALE_P %+ %9]
200 vpaddd %4, %4, [rel PD_DESCALE_P %+ %9]
201 vpsrad %2, %2, DESCALE_P %+ %9
202 vpsrad %4, %4, DESCALE_P %+ %9
204 vpackssdw %4, %2, %4 ; %4=data7_5
206 vperm2i128 %2, %8, %8, 0x01 ; %2=tmp5_4
207 vpunpcklwd %8, %5, %2
208 vpunpckhwd %5, %5, %2
209 vpmaddwd %8, %8, [rel PW_F050_MF256_F060_MF089] ; %8=tmp6_7L
210 vpmaddwd %5, %5, [rel PW_F050_MF256_F060_MF089] ; %5=tmp6_7H
212 vpaddd %8, %8, %6 ; %8=data3_1L
213 vpaddd %5, %5, %7 ; %5=data3_1H
215 vpaddd %8, %8, [rel PD_DESCALE_P %+ %9]
216 vpaddd %5, %5, [rel PD_DESCALE_P %+ %9]
217 vpsrad %8, %8, DESCALE_P %+ %9
218 vpsrad %5, %5, DESCALE_P %+ %9
220 vpackssdw %2, %8, %5 ; %2=data3_1
223 ; --------------------------------------------------------------------------
227 GLOBAL_DATA(jconst_fdct_islow_avx2)
229 EXTN(jconst_fdct_islow_avx2):
231 PW_F130_F054_MF130_F054 times 4 dw (F_0_541 + F_0_765), F_0_541
232 times 4 dw (F_0_541 - F_1_847), F_0_541
233 PW_MF078_F117_F078_F117 times 4 dw (F_1_175 - F_1_961), F_1_175
234 times 4 dw (F_1_175 - F_0_390), F_1_175
235 PW_MF060_MF089_MF050_MF256 times 4 dw (F_0_298 - F_0_899), -F_0_899
236 times 4 dw (F_2_053 - F_2_562), -F_2_562
237 PW_F050_MF256_F060_MF089 times 4 dw (F_3_072 - F_2_562), -F_2_562
238 times 4 dw (F_1_501 - F_0_899), -F_0_899
239 PD_DESCALE_P1 times 8 dd 1 << (DESCALE_P1 - 1)
240 PD_DESCALE_P2 times 8 dd 1 << (DESCALE_P2 - 1)
241 PW_DESCALE_P2X times 16 dw 1 << (PASS1_BITS - 1)
242 PW_1_NEG1 times 8 dw 1
247 ; --------------------------------------------------------------------------
251 ; Perform the forward DCT on one block of samples.
254 ; jsimd_fdct_islow_avx2(DCTELEM *data)
257 ; r10 = DCTELEM *data
260 GLOBAL_FUNCTION(jsimd_fdct_islow_avx2)
262 EXTN(jsimd_fdct_islow_avx2):
267 ; ---- Pass 1: process rows.
269 vmovdqu ymm4, YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)]
270 vmovdqu ymm5, YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)]
271 vmovdqu ymm6, YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)]
272 vmovdqu ymm7, YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)]
273 ; ymm4=(00 01 02 03 04 05 06 07 10 11 12 13 14 15 16 17)
274 ; ymm5=(20 21 22 23 24 25 26 27 30 31 32 33 34 35 36 37)
275 ; ymm6=(40 41 42 43 44 45 46 47 50 51 52 53 54 55 56 57)
276 ; ymm7=(60 61 62 63 64 65 66 67 70 71 72 73 74 75 76 77)
278 vperm2i128 ymm0, ymm4, ymm6, 0x20
279 vperm2i128 ymm1, ymm4, ymm6, 0x31
280 vperm2i128 ymm2, ymm5, ymm7, 0x20
281 vperm2i128 ymm3, ymm5, ymm7, 0x31
282 ; ymm0=(00 01 02 03 04 05 06 07 40 41 42 43 44 45 46 47)
283 ; ymm1=(10 11 12 13 14 15 16 17 50 51 52 53 54 55 56 57)
284 ; ymm2=(20 21 22 23 24 25 26 27 60 61 62 63 64 65 66 67)
285 ; ymm3=(30 31 32 33 34 35 36 37 70 71 72 73 74 75 76 77)
287 dotranspose ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7
289 dodct ymm0, ymm1, ymm2, ymm3, ymm4, ymm5, ymm6, ymm7, 1
290 ; ymm0=data0_4, ymm1=data3_1, ymm2=data2_6, ymm3=data7_5
292 ; ---- Pass 2: process columns.
294 vperm2i128 ymm4, ymm1, ymm3, 0x20 ; ymm4=data3_7
295 vperm2i128 ymm1, ymm1, ymm3, 0x31 ; ymm1=data1_5
297 dotranspose ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7
299 dodct ymm0, ymm1, ymm2, ymm4, ymm3, ymm5, ymm6, ymm7, 2
300 ; ymm0=data0_4, ymm1=data3_1, ymm2=data2_6, ymm4=data7_5
302 vperm2i128 ymm3, ymm0, ymm1, 0x30 ; ymm3=data0_1
303 vperm2i128 ymm5, ymm2, ymm1, 0x20 ; ymm5=data2_3
304 vperm2i128 ymm6, ymm0, ymm4, 0x31 ; ymm6=data4_5
305 vperm2i128 ymm7, ymm2, ymm4, 0x21 ; ymm7=data6_7
307 vmovdqu YMMWORD [YMMBLOCK(0,0,r10,SIZEOF_DCTELEM)], ymm3
308 vmovdqu YMMWORD [YMMBLOCK(2,0,r10,SIZEOF_DCTELEM)], ymm5
309 vmovdqu YMMWORD [YMMBLOCK(4,0,r10,SIZEOF_DCTELEM)], ymm6
310 vmovdqu YMMWORD [YMMBLOCK(6,0,r10,SIZEOF_DCTELEM)], ymm7
317 ; For some reason, the OS X linker does not honor the request to align the
318 ; segment unless we do this.