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.
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 floating-point implementation of the forward DCT
18 ; (Discrete Cosine Transform). The following code is based directly on
19 ; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
21 %include "jsimdext.inc"
24 ; --------------------------------------------------------------------------
26 %macro unpcklps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
30 %macro unpckhps2 2 ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
34 ; --------------------------------------------------------------------------
38 GLOBAL_DATA(jconst_fdct_float_sse)
40 EXTN(jconst_fdct_float_sse):
42 PD_0_382 times 4 dd 0.382683432365089771728460
43 PD_0_707 times 4 dd 0.707106781186547524400844
44 PD_0_541 times 4 dd 0.541196100146196984399723
45 PD_1_306 times 4 dd 1.306562964876376527856643
49 ; --------------------------------------------------------------------------
53 ; Perform the forward DCT on one block of samples.
56 ; jsimd_fdct_float_sse(FAST_FLOAT *data)
59 ; r10 = FAST_FLOAT *data
61 %define wk(i) rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
65 GLOBAL_FUNCTION(jsimd_fdct_float_sse)
67 EXTN(jsimd_fdct_float_sse):
69 mov rax, rsp ; rax = original rbp
71 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
73 mov rbp, rsp ; rbp = aligned rbp
77 ; ---- Pass 1: process rows.
79 mov rdx, r10 ; (FAST_FLOAT *)
83 movaps xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
84 movaps xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
85 movaps xmm2, XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)]
86 movaps xmm3, XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)]
88 ; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
89 ; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
91 movaps xmm4, xmm0 ; transpose coefficients(phase 1)
92 unpcklps xmm0, xmm1 ; xmm0=(20 30 21 31)
93 unpckhps xmm4, xmm1 ; xmm4=(22 32 23 33)
94 movaps xmm5, xmm2 ; transpose coefficients(phase 1)
95 unpcklps xmm2, xmm3 ; xmm2=(24 34 25 35)
96 unpckhps xmm5, xmm3 ; xmm5=(26 36 27 37)
98 movaps xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
99 movaps xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
100 movaps xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
101 movaps xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
103 ; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
104 ; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
106 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 32 23 33)
107 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(24 34 25 35)
109 movaps xmm4, xmm6 ; transpose coefficients(phase 1)
110 unpcklps xmm6, xmm7 ; xmm6=(00 10 01 11)
111 unpckhps xmm4, xmm7 ; xmm4=(02 12 03 13)
112 movaps xmm2, xmm1 ; transpose coefficients(phase 1)
113 unpcklps xmm1, xmm3 ; xmm1=(04 14 05 15)
114 unpckhps xmm2, xmm3 ; xmm2=(06 16 07 17)
116 movaps xmm7, xmm6 ; transpose coefficients(phase 2)
117 unpcklps2 xmm6, xmm0 ; xmm6=(00 10 20 30)=data0
118 unpckhps2 xmm7, xmm0 ; xmm7=(01 11 21 31)=data1
119 movaps xmm3, xmm2 ; transpose coefficients(phase 2)
120 unpcklps2 xmm2, xmm5 ; xmm2=(06 16 26 36)=data6
121 unpckhps2 xmm3, xmm5 ; xmm3=(07 17 27 37)=data7
125 subps xmm7, xmm2 ; xmm7=data1-data6=tmp6
126 subps xmm6, xmm3 ; xmm6=data0-data7=tmp7
127 addps xmm0, xmm2 ; xmm0=data1+data6=tmp1
128 addps xmm5, xmm3 ; xmm5=data0+data7=tmp0
130 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 32 23 33)
131 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(24 34 25 35)
132 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6
133 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
135 movaps xmm7, xmm4 ; transpose coefficients(phase 2)
136 unpcklps2 xmm4, xmm2 ; xmm4=(02 12 22 32)=data2
137 unpckhps2 xmm7, xmm2 ; xmm7=(03 13 23 33)=data3
138 movaps xmm6, xmm1 ; transpose coefficients(phase 2)
139 unpcklps2 xmm1, xmm3 ; xmm1=(04 14 24 34)=data4
140 unpckhps2 xmm6, xmm3 ; xmm6=(05 15 25 35)=data5
144 addps xmm7, xmm1 ; xmm7=data3+data4=tmp3
145 addps xmm4, xmm6 ; xmm4=data2+data5=tmp2
146 subps xmm2, xmm1 ; xmm2=data3-data4=tmp4
147 subps xmm3, xmm6 ; xmm3=data2-data5=tmp5
153 subps xmm5, xmm7 ; xmm5=tmp13
154 subps xmm0, xmm4 ; xmm0=tmp12
155 addps xmm1, xmm7 ; xmm1=tmp10
156 addps xmm6, xmm4 ; xmm6=tmp11
159 mulps xmm0, [rel PD_0_707] ; xmm0=z1
163 subps xmm1, xmm6 ; xmm1=data4
164 subps xmm5, xmm0 ; xmm5=data6
165 addps xmm7, xmm6 ; xmm7=data0
166 addps xmm4, xmm0 ; xmm4=data2
168 movaps XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)], xmm1
169 movaps XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
170 movaps XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
171 movaps XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
175 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6
176 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7
178 addps xmm2, xmm3 ; xmm2=tmp10
179 addps xmm3, xmm6 ; xmm3=tmp11
180 addps xmm6, xmm0 ; xmm6=tmp12, xmm0=tmp7
182 mulps xmm3, [rel PD_0_707] ; xmm3=z3
184 movaps xmm1, xmm2 ; xmm1=tmp10
186 mulps xmm2, [rel PD_0_382] ; xmm2=z5
187 mulps xmm1, [rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
188 mulps xmm6, [rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
189 addps xmm1, xmm2 ; xmm1=z2
190 addps xmm6, xmm2 ; xmm6=z4
193 subps xmm0, xmm3 ; xmm0=z13
194 addps xmm5, xmm3 ; xmm5=z11
198 subps xmm0, xmm1 ; xmm0=data3
199 subps xmm5, xmm6 ; xmm5=data7
200 addps xmm7, xmm1 ; xmm7=data5
201 addps xmm4, xmm6 ; xmm4=data1
203 movaps XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
204 movaps XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
205 movaps XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)], xmm7
206 movaps XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
208 add rdx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
212 ; ---- Pass 2: process columns.
214 mov rdx, r10 ; (FAST_FLOAT *)
218 movaps xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
219 movaps xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
220 movaps xmm2, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)]
221 movaps xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)]
223 ; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
224 ; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
226 movaps xmm4, xmm0 ; transpose coefficients(phase 1)
227 unpcklps xmm0, xmm1 ; xmm0=(02 03 12 13)
228 unpckhps xmm4, xmm1 ; xmm4=(22 23 32 33)
229 movaps xmm5, xmm2 ; transpose coefficients(phase 1)
230 unpcklps xmm2, xmm3 ; xmm2=(42 43 52 53)
231 unpckhps xmm5, xmm3 ; xmm5=(62 63 72 73)
233 movaps xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
234 movaps xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
235 movaps xmm1, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)]
236 movaps xmm3, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)]
238 ; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
239 ; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
241 movaps XMMWORD [wk(0)], xmm4 ; wk(0)=(22 23 32 33)
242 movaps XMMWORD [wk(1)], xmm2 ; wk(1)=(42 43 52 53)
244 movaps xmm4, xmm6 ; transpose coefficients(phase 1)
245 unpcklps xmm6, xmm7 ; xmm6=(00 01 10 11)
246 unpckhps xmm4, xmm7 ; xmm4=(20 21 30 31)
247 movaps xmm2, xmm1 ; transpose coefficients(phase 1)
248 unpcklps xmm1, xmm3 ; xmm1=(40 41 50 51)
249 unpckhps xmm2, xmm3 ; xmm2=(60 61 70 71)
251 movaps xmm7, xmm6 ; transpose coefficients(phase 2)
252 unpcklps2 xmm6, xmm0 ; xmm6=(00 01 02 03)=data0
253 unpckhps2 xmm7, xmm0 ; xmm7=(10 11 12 13)=data1
254 movaps xmm3, xmm2 ; transpose coefficients(phase 2)
255 unpcklps2 xmm2, xmm5 ; xmm2=(60 61 62 63)=data6
256 unpckhps2 xmm3, xmm5 ; xmm3=(70 71 72 73)=data7
260 subps xmm7, xmm2 ; xmm7=data1-data6=tmp6
261 subps xmm6, xmm3 ; xmm6=data0-data7=tmp7
262 addps xmm0, xmm2 ; xmm0=data1+data6=tmp1
263 addps xmm5, xmm3 ; xmm5=data0+data7=tmp0
265 movaps xmm2, XMMWORD [wk(0)] ; xmm2=(22 23 32 33)
266 movaps xmm3, XMMWORD [wk(1)] ; xmm3=(42 43 52 53)
267 movaps XMMWORD [wk(0)], xmm7 ; wk(0)=tmp6
268 movaps XMMWORD [wk(1)], xmm6 ; wk(1)=tmp7
270 movaps xmm7, xmm4 ; transpose coefficients(phase 2)
271 unpcklps2 xmm4, xmm2 ; xmm4=(20 21 22 23)=data2
272 unpckhps2 xmm7, xmm2 ; xmm7=(30 31 32 33)=data3
273 movaps xmm6, xmm1 ; transpose coefficients(phase 2)
274 unpcklps2 xmm1, xmm3 ; xmm1=(40 41 42 43)=data4
275 unpckhps2 xmm6, xmm3 ; xmm6=(50 51 52 53)=data5
279 addps xmm7, xmm1 ; xmm7=data3+data4=tmp3
280 addps xmm4, xmm6 ; xmm4=data2+data5=tmp2
281 subps xmm2, xmm1 ; xmm2=data3-data4=tmp4
282 subps xmm3, xmm6 ; xmm3=data2-data5=tmp5
288 subps xmm5, xmm7 ; xmm5=tmp13
289 subps xmm0, xmm4 ; xmm0=tmp12
290 addps xmm1, xmm7 ; xmm1=tmp10
291 addps xmm6, xmm4 ; xmm6=tmp11
294 mulps xmm0, [rel PD_0_707] ; xmm0=z1
298 subps xmm1, xmm6 ; xmm1=data4
299 subps xmm5, xmm0 ; xmm5=data6
300 addps xmm7, xmm6 ; xmm7=data0
301 addps xmm4, xmm0 ; xmm4=data2
303 movaps XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)], xmm1
304 movaps XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
305 movaps XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
306 movaps XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
310 movaps xmm6, XMMWORD [wk(0)] ; xmm6=tmp6
311 movaps xmm0, XMMWORD [wk(1)] ; xmm0=tmp7
313 addps xmm2, xmm3 ; xmm2=tmp10
314 addps xmm3, xmm6 ; xmm3=tmp11
315 addps xmm6, xmm0 ; xmm6=tmp12, xmm0=tmp7
317 mulps xmm3, [rel PD_0_707] ; xmm3=z3
319 movaps xmm1, xmm2 ; xmm1=tmp10
321 mulps xmm2, [rel PD_0_382] ; xmm2=z5
322 mulps xmm1, [rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
323 mulps xmm6, [rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
324 addps xmm1, xmm2 ; xmm1=z2
325 addps xmm6, xmm2 ; xmm6=z4
328 subps xmm0, xmm3 ; xmm0=z13
329 addps xmm5, xmm3 ; xmm5=z11
333 subps xmm0, xmm1 ; xmm0=data3
334 subps xmm5, xmm6 ; xmm5=data7
335 addps xmm7, xmm1 ; xmm7=data5
336 addps xmm4, xmm6 ; xmm4=data1
338 movaps XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
339 movaps XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
340 movaps XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
341 movaps XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
343 add rdx, byte 4*SIZEOF_FAST_FLOAT
348 mov rsp, rbp ; rsp <- aligned rbp
349 pop rsp ; rsp <- original rbp
353 ; For some reason, the OS X linker does not honor the request to align the
354 ; segment unless we do this.