2 ; jfdctflt.asm - floating-point FDCT (64-bit SSE)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright (C) 2009, 2016, D. R. Commander.
6 ; Copyright (C) 2023, Aliaksiej Kandracienka.
8 ; Based on the x86 SIMD extension for IJG JPEG library
9 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
10 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
12 ; This file should be assembled with NASM (Netwide Assembler),
13 ; can *not* be assembled with Microsoft's MASM or any compatible
14 ; assembler (including Borland's Turbo Assembler).
15 ; NASM is available from http://nasm.sourceforge.net/ or
16 ; http://sourceforge.net/project/showfiles.php?group_id=6208
18 ; This file contains a floating-point implementation of the forward DCT
19 ; (Discrete Cosine Transform). The following code is based directly on
20 ; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
22 %include "jsimdext.inc"
25 ; --------------------------------------------------------------------------
27 %macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
31 %macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
35 ; --------------------------------------------------------------------------
39 GLOBAL_DATA(jconst_fdct_float_sse)
41 EXTN(jconst_fdct_float_sse):
43 PD_0_382 times 4 dd 0.382683432365089771728460
44 PD_0_707 times 4 dd 0.707106781186547524400844
45 PD_0_541 times 4 dd 0.541196100146196984399723
46 PD_1_306 times 4 dd 1.306562964876376527856643
50 ; --------------------------------------------------------------------------
54 ; Perform the forward DCT on one block of samples.
57 ; jsimd_fdct_float_sse(FAST_FLOAT *data)
60 ; r10 = FAST_FLOAT *data
62 %define wk(i) r15 - (WK_NUM - (i)) * SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
66 GLOBAL_FUNCTION(jsimd_fdct_float_sse)
68 EXTN(jsimd_fdct_float_sse):
72 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
73 ; Allocate stack space for wk array. r15 is used to access it.
75 sub rsp, byte (SIZEOF_XMMWORD * WK_NUM)
78 ; ---- Pass 1: process rows.
80 mov rdx, r10 ; (FAST_FLOAT *)
84 movaps xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
85 movaps xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
86 movaps xmm2, XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)]
87 movaps xmm3, XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)]
89 ; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
90 ; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
92 movaps xmm4, xmm0 ; transpose coefficients(phase 1)
93 unpcklps xmm0, xmm1 ; xmm0=(20 30 21 31)
94 unpckhps xmm4, xmm1 ; xmm4=(22 32 23 33)
95 movaps xmm5, xmm2 ; transpose coefficients(phase 1)
96 unpcklps xmm2, xmm3 ; xmm2=(24 34 25 35)
97 unpckhps xmm5, xmm3 ; xmm5=(26 36 27 37)
99 movaps xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
100 movaps xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
101 movaps xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
102 movaps xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
104 ; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
105 ; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
107 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 32 23 33)
108 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(24 34 25 35)
110 movaps xmm4, xmm6 ; transpose coefficients(phase 1)
111 unpcklps xmm6, xmm7 ; xmm6=(00 10 01 11)
112 unpckhps xmm4, xmm7 ; xmm4=(02 12 03 13)
113 movaps xmm2, xmm1 ; transpose coefficients(phase 1)
114 unpcklps xmm1, xmm3 ; xmm1=(04 14 05 15)
115 unpckhps xmm2, xmm3 ; xmm2=(06 16 07 17)
117 movaps xmm7, xmm6 ; transpose coefficients(phase 2)
118 unpcklps2 xmm6, xmm0 ; xmm6=(00 10 20 30)=data0
119 unpckhps2 xmm7, xmm0 ; xmm7=(01 11 21 31)=data1
120 movaps xmm3, xmm2 ; transpose coefficients(phase 2)
121 unpcklps2 xmm2, xmm5 ; xmm2=(06 16 26 36)=data6
122 unpckhps2 xmm3, xmm5 ; xmm3=(07 17 27 37)=data7
126 subps xmm7, xmm2 ; xmm7=data1-data6=tmp6
127 subps xmm6, xmm3 ; xmm6=data0-data7=tmp7
128 addps xmm0, xmm2 ; xmm0=data1+data6=tmp1
129 addps xmm5, xmm3 ; xmm5=data0+data7=tmp0
131 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 32 23 33)
132 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(24 34 25 35)
133 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6
134 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
136 movaps xmm7, xmm4 ; transpose coefficients(phase 2)
137 unpcklps2 xmm4, xmm2 ; xmm4=(02 12 22 32)=data2
138 unpckhps2 xmm7, xmm2 ; xmm7=(03 13 23 33)=data3
139 movaps xmm6, xmm1 ; transpose coefficients(phase 2)
140 unpcklps2 xmm1, xmm3 ; xmm1=(04 14 24 34)=data4
141 unpckhps2 xmm6, xmm3 ; xmm6=(05 15 25 35)=data5
145 addps xmm7, xmm1 ; xmm7=data3+data4=tmp3
146 addps xmm4, xmm6 ; xmm4=data2+data5=tmp2
147 subps xmm2, xmm1 ; xmm2=data3-data4=tmp4
148 subps xmm3, xmm6 ; xmm3=data2-data5=tmp5
154 subps xmm5, xmm7 ; xmm5=tmp13
155 subps xmm0, xmm4 ; xmm0=tmp12
156 addps xmm1, xmm7 ; xmm1=tmp10
157 addps xmm6, xmm4 ; xmm6=tmp11
160 mulps xmm0, [rel PD_0_707] ; xmm0=z1
164 subps xmm1, xmm6 ; xmm1=data4
165 subps xmm5, xmm0 ; xmm5=data6
166 addps xmm7, xmm6 ; xmm7=data0
167 addps xmm4, xmm0 ; xmm4=data2
169 movaps XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)], xmm1
170 movaps XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
171 movaps XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
172 movaps XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
176 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6
177 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7
179 addps xmm2, xmm3 ; xmm2=tmp10
180 addps xmm3, xmm6 ; xmm3=tmp11
181 addps xmm6, xmm0 ; xmm6=tmp12, xmm0=tmp7
183 mulps xmm3, [rel PD_0_707] ; xmm3=z3
185 movaps xmm1, xmm2 ; xmm1=tmp10
187 mulps xmm2, [rel PD_0_382] ; xmm2=z5
188 mulps xmm1, [rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
189 mulps xmm6, [rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
190 addps xmm1, xmm2 ; xmm1=z2
191 addps xmm6, xmm2 ; xmm6=z4
194 subps xmm0, xmm3 ; xmm0=z13
195 addps xmm5, xmm3 ; xmm5=z11
199 subps xmm0, xmm1 ; xmm0=data3
200 subps xmm5, xmm6 ; xmm5=data7
201 addps xmm7, xmm1 ; xmm7=data5
202 addps xmm4, xmm6 ; xmm4=data1
204 movaps XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
205 movaps XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
206 movaps XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)], xmm7
207 movaps XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
209 add rdx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
213 ; ---- Pass 2: process columns.
215 mov rdx, r10 ; (FAST_FLOAT *)
219 movaps xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
220 movaps xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
221 movaps xmm2, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)]
222 movaps xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)]
224 ; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
225 ; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
227 movaps xmm4, xmm0 ; transpose coefficients(phase 1)
228 unpcklps xmm0, xmm1 ; xmm0=(02 03 12 13)
229 unpckhps xmm4, xmm1 ; xmm4=(22 23 32 33)
230 movaps xmm5, xmm2 ; transpose coefficients(phase 1)
231 unpcklps xmm2, xmm3 ; xmm2=(42 43 52 53)
232 unpckhps xmm5, xmm3 ; xmm5=(62 63 72 73)
234 movaps xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
235 movaps xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
236 movaps xmm1, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)]
237 movaps xmm3, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)]
239 ; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
240 ; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
242 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 23 32 33)
243 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(42 43 52 53)
245 movaps xmm4, xmm6 ; transpose coefficients(phase 1)
246 unpcklps xmm6, xmm7 ; xmm6=(00 01 10 11)
247 unpckhps xmm4, xmm7 ; xmm4=(20 21 30 31)
248 movaps xmm2, xmm1 ; transpose coefficients(phase 1)
249 unpcklps xmm1, xmm3 ; xmm1=(40 41 50 51)
250 unpckhps xmm2, xmm3 ; xmm2=(60 61 70 71)
252 movaps xmm7, xmm6 ; transpose coefficients(phase 2)
253 unpcklps2 xmm6, xmm0 ; xmm6=(00 01 02 03)=data0
254 unpckhps2 xmm7, xmm0 ; xmm7=(10 11 12 13)=data1
255 movaps xmm3, xmm2 ; transpose coefficients(phase 2)
256 unpcklps2 xmm2, xmm5 ; xmm2=(60 61 62 63)=data6
257 unpckhps2 xmm3, xmm5 ; xmm3=(70 71 72 73)=data7
261 subps xmm7, xmm2 ; xmm7=data1-data6=tmp6
262 subps xmm6, xmm3 ; xmm6=data0-data7=tmp7
263 addps xmm0, xmm2 ; xmm0=data1+data6=tmp1
264 addps xmm5, xmm3 ; xmm5=data0+data7=tmp0
266 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 23 32 33)
267 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53)
268 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6
269 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
271 movaps xmm7, xmm4 ; transpose coefficients(phase 2)
272 unpcklps2 xmm4, xmm2 ; xmm4=(20 21 22 23)=data2
273 unpckhps2 xmm7, xmm2 ; xmm7=(30 31 32 33)=data3
274 movaps xmm6, xmm1 ; transpose coefficients(phase 2)
275 unpcklps2 xmm1, xmm3 ; xmm1=(40 41 42 43)=data4
276 unpckhps2 xmm6, xmm3 ; xmm6=(50 51 52 53)=data5
280 addps xmm7, xmm1 ; xmm7=data3+data4=tmp3
281 addps xmm4, xmm6 ; xmm4=data2+data5=tmp2
282 subps xmm2, xmm1 ; xmm2=data3-data4=tmp4
283 subps xmm3, xmm6 ; xmm3=data2-data5=tmp5
289 subps xmm5, xmm7 ; xmm5=tmp13
290 subps xmm0, xmm4 ; xmm0=tmp12
291 addps xmm1, xmm7 ; xmm1=tmp10
292 addps xmm6, xmm4 ; xmm6=tmp11
295 mulps xmm0, [rel PD_0_707] ; xmm0=z1
299 subps xmm1, xmm6 ; xmm1=data4
300 subps xmm5, xmm0 ; xmm5=data6
301 addps xmm7, xmm6 ; xmm7=data0
302 addps xmm4, xmm0 ; xmm4=data2
304 movaps XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)], xmm1
305 movaps XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
306 movaps XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
307 movaps XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
311 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6
312 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7
314 addps xmm2, xmm3 ; xmm2=tmp10
315 addps xmm3, xmm6 ; xmm3=tmp11
316 addps xmm6, xmm0 ; xmm6=tmp12, xmm0=tmp7
318 mulps xmm3, [rel PD_0_707] ; xmm3=z3
320 movaps xmm1, xmm2 ; xmm1=tmp10
322 mulps xmm2, [rel PD_0_382] ; xmm2=z5
323 mulps xmm1, [rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
324 mulps xmm6, [rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
325 addps xmm1, xmm2 ; xmm1=z2
326 addps xmm6, xmm2 ; xmm6=z4
329 subps xmm0, xmm3 ; xmm0=z13
330 addps xmm5, xmm3 ; xmm5=z11
334 subps xmm0, xmm1 ; xmm0=data3
335 subps xmm5, xmm6 ; xmm5=data7
336 addps xmm7, xmm1 ; xmm7=data5
337 addps xmm4, xmm6 ; xmm4=data1
339 movaps XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
340 movaps XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
341 movaps XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
342 movaps XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
344 add rdx, byte 4*SIZEOF_FAST_FLOAT
354 ; For some reason, the OS X linker does not honor the request to align the
355 ; segment unless we do this.