1;
2; jidctflt.asm - floating-point IDCT (3DNow! & MMX)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2016, 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 floating-point implementation of the inverse DCT
18; (Discrete Cosine Transform). The following code is based directly on
19; the IJG's original jidctflt.c; see the jidctflt.c for more details.
20
21%include "jsimdext.inc"
22%include "jdct.inc"
23
24; --------------------------------------------------------------------------
25    SECTION     SEG_CONST
26
27    alignz      32
28    GLOBAL_DATA(jconst_idct_float_3dnow)
29
30EXTN(jconst_idct_float_3dnow):
31
32PD_1_414        times 2 dd 1.414213562373095048801689
33PD_1_847        times 2 dd 1.847759065022573512256366
34PD_1_082        times 2 dd 1.082392200292393968799446
35PD_2_613        times 2 dd 2.613125929752753055713286
36PD_RNDINT_MAGIC times 2 dd 100663296.0  ; (float)(0x00C00000 << 3)
37PB_CENTERJSAMP  times 8 db CENTERJSAMPLE
38
39    alignz      32
40
41; --------------------------------------------------------------------------
42    SECTION     SEG_TEXT
43    BITS        32
44;
45; Perform dequantization and inverse DCT on one block of coefficients.
46;
47; GLOBAL(void)
48; jsimd_idct_float_3dnow(void *dct_table, JCOEFPTR coef_block,
49;                        JSAMPARRAY output_buf, JDIMENSION output_col)
50;
51
52%define dct_table(b)   (b) + 8          ; void *dct_table
53%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
54%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
55%define output_col(b)  (b) + 20         ; JDIMENSION output_col
56
57%define original_ebp   ebp + 0
58%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_MMWORD
59                                        ; mmword wk[WK_NUM]
60%define WK_NUM         2
61%define workspace      wk(0) - DCTSIZE2 * SIZEOF_FAST_FLOAT
62                                        ; FAST_FLOAT workspace[DCTSIZE2]
63
64    align       32
65    GLOBAL_FUNCTION(jsimd_idct_float_3dnow)
66
67EXTN(jsimd_idct_float_3dnow):
68    push        ebp
69    mov         eax, esp                    ; eax = original ebp
70    sub         esp, byte 4
71    and         esp, byte (-SIZEOF_MMWORD)  ; align to 64 bits
72    mov         [esp], eax
73    mov         ebp, esp                    ; ebp = aligned ebp
74    lea         esp, [workspace]
75    push        ebx
76;   push        ecx                     ; need not be preserved
77;   push        edx                     ; need not be preserved
78    push        esi
79    push        edi
80
81    get_GOT     ebx                     ; get GOT address
82
83    ; ---- Pass 1: process columns from input, store into work array.
84
85;   mov         eax, [original_ebp]
86    mov         edx, POINTER [dct_table(eax)]    ; quantptr
87    mov         esi, JCOEFPTR [coef_block(eax)]  ; inptr
88    lea         edi, [workspace]                 ; FAST_FLOAT *wsptr
89    mov         ecx, DCTSIZE/2                   ; ctr
90    alignx      16, 7
91.columnloop:
92%ifndef NO_ZERO_COLUMN_TEST_FLOAT_3DNOW
93    mov         eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
94    or          eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
95    jnz         short .columnDCT
96
97    pushpic     ebx                     ; save GOT address
98    mov         ebx, dword [DWBLOCK(3,0,esi,SIZEOF_JCOEF)]
99    mov         eax, dword [DWBLOCK(4,0,esi,SIZEOF_JCOEF)]
100    or          ebx, dword [DWBLOCK(5,0,esi,SIZEOF_JCOEF)]
101    or          eax, dword [DWBLOCK(6,0,esi,SIZEOF_JCOEF)]
102    or          ebx, dword [DWBLOCK(7,0,esi,SIZEOF_JCOEF)]
103    or          eax, ebx
104    poppic      ebx                     ; restore GOT address
105    jnz         short .columnDCT
106
107    ; -- AC terms all zero
108
109    movd        mm0, dword [DWBLOCK(0,0,esi,SIZEOF_JCOEF)]
110
111    punpcklwd   mm0, mm0
112    psrad       mm0, (DWORD_BIT-WORD_BIT)
113    pi2fd       mm0, mm0
114
115    pfmul       mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
116
117    movq        mm1, mm0
118    punpckldq   mm0, mm0
119    punpckhdq   mm1, mm1
120
121    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm0
122    movq        MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm0
123    movq        MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm0
124    movq        MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0
125    movq        MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm1
126    movq        MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm1
127    movq        MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm1
128    movq        MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm1
129    jmp         near .nextcolumn
130    alignx      16, 7
131%endif
132.columnDCT:
133
134    ; -- Even part
135
136    movd        mm0, dword [DWBLOCK(0,0,esi,SIZEOF_JCOEF)]
137    movd        mm1, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
138    movd        mm2, dword [DWBLOCK(4,0,esi,SIZEOF_JCOEF)]
139    movd        mm3, dword [DWBLOCK(6,0,esi,SIZEOF_JCOEF)]
140
141    punpcklwd   mm0, mm0
142    punpcklwd   mm1, mm1
143    psrad       mm0, (DWORD_BIT-WORD_BIT)
144    psrad       mm1, (DWORD_BIT-WORD_BIT)
145    pi2fd       mm0, mm0
146    pi2fd       mm1, mm1
147
148    pfmul       mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
149    pfmul       mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
150
151    punpcklwd   mm2, mm2
152    punpcklwd   mm3, mm3
153    psrad       mm2, (DWORD_BIT-WORD_BIT)
154    psrad       mm3, (DWORD_BIT-WORD_BIT)
155    pi2fd       mm2, mm2
156    pi2fd       mm3, mm3
157
158    pfmul       mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
159    pfmul       mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
160
161    movq        mm4, mm0
162    movq        mm5, mm1
163    pfsub       mm0, mm2                ; mm0=tmp11
164    pfsub       mm1, mm3
165    pfadd       mm4, mm2                ; mm4=tmp10
166    pfadd       mm5, mm3                ; mm5=tmp13
167
168    pfmul       mm1, [GOTOFF(ebx,PD_1_414)]
169    pfsub       mm1, mm5                ; mm1=tmp12
170
171    movq        mm6, mm4
172    movq        mm7, mm0
173    pfsub       mm4, mm5                ; mm4=tmp3
174    pfsub       mm0, mm1                ; mm0=tmp2
175    pfadd       mm6, mm5                ; mm6=tmp0
176    pfadd       mm7, mm1                ; mm7=tmp1
177
178    movq        MMWORD [wk(1)], mm4     ; tmp3
179    movq        MMWORD [wk(0)], mm0     ; tmp2
180
181    ; -- Odd part
182
183    movd        mm2, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
184    movd        mm3, dword [DWBLOCK(3,0,esi,SIZEOF_JCOEF)]
185    movd        mm5, dword [DWBLOCK(5,0,esi,SIZEOF_JCOEF)]
186    movd        mm1, dword [DWBLOCK(7,0,esi,SIZEOF_JCOEF)]
187
188    punpcklwd   mm2, mm2
189    punpcklwd   mm3, mm3
190    psrad       mm2, (DWORD_BIT-WORD_BIT)
191    psrad       mm3, (DWORD_BIT-WORD_BIT)
192    pi2fd       mm2, mm2
193    pi2fd       mm3, mm3
194
195    pfmul       mm2, MMWORD [MMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
196    pfmul       mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
197
198    punpcklwd   mm5, mm5
199    punpcklwd   mm1, mm1
200    psrad       mm5, (DWORD_BIT-WORD_BIT)
201    psrad       mm1, (DWORD_BIT-WORD_BIT)
202    pi2fd       mm5, mm5
203    pi2fd       mm1, mm1
204
205    pfmul       mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
206    pfmul       mm1, MMWORD [MMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
207
208    movq        mm4, mm2
209    movq        mm0, mm5
210    pfadd       mm2, mm1                ; mm2=z11
211    pfadd       mm5, mm3                ; mm5=z13
212    pfsub       mm4, mm1                ; mm4=z12
213    pfsub       mm0, mm3                ; mm0=z10
214
215    movq        mm1, mm2
216    pfsub       mm2, mm5
217    pfadd       mm1, mm5                ; mm1=tmp7
218
219    pfmul       mm2, [GOTOFF(ebx,PD_1_414)]  ; mm2=tmp11
220
221    movq        mm3, mm0
222    pfadd       mm0, mm4
223    pfmul       mm0, [GOTOFF(ebx,PD_1_847)]  ; mm0=z5
224    pfmul       mm3, [GOTOFF(ebx,PD_2_613)]  ; mm3=(z10 * 2.613125930)
225    pfmul       mm4, [GOTOFF(ebx,PD_1_082)]  ; mm4=(z12 * 1.082392200)
226    pfsubr      mm3, mm0                     ; mm3=tmp12
227    pfsub       mm4, mm0                     ; mm4=tmp10
228
229    ; -- Final output stage
230
231    pfsub       mm3, mm1                ; mm3=tmp6
232    movq        mm5, mm6
233    movq        mm0, mm7
234    pfadd       mm6, mm1                ; mm6=data0=(00 01)
235    pfadd       mm7, mm3                ; mm7=data1=(10 11)
236    pfsub       mm5, mm1                ; mm5=data7=(70 71)
237    pfsub       mm0, mm3                ; mm0=data6=(60 61)
238    pfsub       mm2, mm3                ; mm2=tmp5
239
240    movq        mm1, mm6                ; transpose coefficients
241    punpckldq   mm6, mm7                ; mm6=(00 10)
242    punpckhdq   mm1, mm7                ; mm1=(01 11)
243    movq        mm3, mm0                ; transpose coefficients
244    punpckldq   mm0, mm5                ; mm0=(60 70)
245    punpckhdq   mm3, mm5                ; mm3=(61 71)
246
247    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], mm6
248    movq        MMWORD [MMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], mm1
249    movq        MMWORD [MMBLOCK(0,3,edi,SIZEOF_FAST_FLOAT)], mm0
250    movq        MMWORD [MMBLOCK(1,3,edi,SIZEOF_FAST_FLOAT)], mm3
251
252    movq        mm7, MMWORD [wk(0)]     ; mm7=tmp2
253    movq        mm5, MMWORD [wk(1)]     ; mm5=tmp3
254
255    pfadd       mm4, mm2                ; mm4=tmp4
256    movq        mm6, mm7
257    movq        mm1, mm5
258    pfadd       mm7, mm2                ; mm7=data2=(20 21)
259    pfadd       mm5, mm4                ; mm5=data4=(40 41)
260    pfsub       mm6, mm2                ; mm6=data5=(50 51)
261    pfsub       mm1, mm4                ; mm1=data3=(30 31)
262
263    movq        mm0, mm7                ; transpose coefficients
264    punpckldq   mm7, mm1                ; mm7=(20 30)
265    punpckhdq   mm0, mm1                ; mm0=(21 31)
266    movq        mm3, mm5                ; transpose coefficients
267    punpckldq   mm5, mm6                ; mm5=(40 50)
268    punpckhdq   mm3, mm6                ; mm3=(41 51)
269
270    movq        MMWORD [MMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], mm7
271    movq        MMWORD [MMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], mm0
272    movq        MMWORD [MMBLOCK(0,2,edi,SIZEOF_FAST_FLOAT)], mm5
273    movq        MMWORD [MMBLOCK(1,2,edi,SIZEOF_FAST_FLOAT)], mm3
274
275.nextcolumn:
276    add         esi, byte 2*SIZEOF_JCOEF               ; coef_block
277    add         edx, byte 2*SIZEOF_FLOAT_MULT_TYPE     ; quantptr
278    add         edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT  ; wsptr
279    dec         ecx                                    ; ctr
280    jnz         near .columnloop
281
282    ; -- Prefetch the next coefficient block
283
284    prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
285    prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
286    prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
287    prefetch [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
288
289    ; ---- Pass 2: process rows from work array, store into output array.
290
291    mov         eax, [original_ebp]
292    lea         esi, [workspace]                   ; FAST_FLOAT *wsptr
293    mov         edi, JSAMPARRAY [output_buf(eax)]  ; (JSAMPROW *)
294    mov         eax, JDIMENSION [output_col(eax)]
295    mov         ecx, DCTSIZE/2                     ; ctr
296    alignx      16, 7
297.rowloop:
298
299    ; -- Even part
300
301    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
302    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
303    movq        mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
304    movq        mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
305
306    movq        mm4, mm0
307    movq        mm5, mm1
308    pfsub       mm0, mm2                ; mm0=tmp11
309    pfsub       mm1, mm3
310    pfadd       mm4, mm2                ; mm4=tmp10
311    pfadd       mm5, mm3                ; mm5=tmp13
312
313    pfmul       mm1, [GOTOFF(ebx,PD_1_414)]
314    pfsub       mm1, mm5                ; mm1=tmp12
315
316    movq        mm6, mm4
317    movq        mm7, mm0
318    pfsub       mm4, mm5                ; mm4=tmp3
319    pfsub       mm0, mm1                ; mm0=tmp2
320    pfadd       mm6, mm5                ; mm6=tmp0
321    pfadd       mm7, mm1                ; mm7=tmp1
322
323    movq        MMWORD [wk(1)], mm4     ; tmp3
324    movq        MMWORD [wk(0)], mm0     ; tmp2
325
326    ; -- Odd part
327
328    movq        mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
329    movq        mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
330    movq        mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
331    movq        mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
332
333    movq        mm4, mm2
334    movq        mm0, mm5
335    pfadd       mm2, mm1                ; mm2=z11
336    pfadd       mm5, mm3                ; mm5=z13
337    pfsub       mm4, mm1                ; mm4=z12
338    pfsub       mm0, mm3                ; mm0=z10
339
340    movq        mm1, mm2
341    pfsub       mm2, mm5
342    pfadd       mm1, mm5                ; mm1=tmp7
343
344    pfmul       mm2, [GOTOFF(ebx,PD_1_414)]  ; mm2=tmp11
345
346    movq        mm3, mm0
347    pfadd       mm0, mm4
348    pfmul       mm0, [GOTOFF(ebx,PD_1_847)]  ; mm0=z5
349    pfmul       mm3, [GOTOFF(ebx,PD_2_613)]  ; mm3=(z10 * 2.613125930)
350    pfmul       mm4, [GOTOFF(ebx,PD_1_082)]  ; mm4=(z12 * 1.082392200)
351    pfsubr      mm3, mm0                     ; mm3=tmp12
352    pfsub       mm4, mm0                     ; mm4=tmp10
353
354    ; -- Final output stage
355
356    pfsub       mm3, mm1                ; mm3=tmp6
357    movq        mm5, mm6
358    movq        mm0, mm7
359    pfadd       mm6, mm1                ; mm6=data0=(00 10)
360    pfadd       mm7, mm3                ; mm7=data1=(01 11)
361    pfsub       mm5, mm1                ; mm5=data7=(07 17)
362    pfsub       mm0, mm3                ; mm0=data6=(06 16)
363    pfsub       mm2, mm3                ; mm2=tmp5
364
365    movq        mm1, [GOTOFF(ebx,PD_RNDINT_MAGIC)]  ; mm1=[PD_RNDINT_MAGIC]
366    pcmpeqd     mm3, mm3
367    psrld       mm3, WORD_BIT           ; mm3={0xFFFF 0x0000 0xFFFF 0x0000}
368
369    pfadd       mm6, mm1                ; mm6=roundint(data0/8)=(00 ** 10 **)
370    pfadd       mm7, mm1                ; mm7=roundint(data1/8)=(01 ** 11 **)
371    pfadd       mm0, mm1                ; mm0=roundint(data6/8)=(06 ** 16 **)
372    pfadd       mm5, mm1                ; mm5=roundint(data7/8)=(07 ** 17 **)
373
374    pand        mm6, mm3                ; mm6=(00 -- 10 --)
375    pslld       mm7, WORD_BIT           ; mm7=(-- 01 -- 11)
376    pand        mm0, mm3                ; mm0=(06 -- 16 --)
377    pslld       mm5, WORD_BIT           ; mm5=(-- 07 -- 17)
378    por         mm6, mm7                ; mm6=(00 01 10 11)
379    por         mm0, mm5                ; mm0=(06 07 16 17)
380
381    movq        mm1, MMWORD [wk(0)]     ; mm1=tmp2
382    movq        mm3, MMWORD [wk(1)]     ; mm3=tmp3
383
384    pfadd       mm4, mm2                ; mm4=tmp4
385    movq        mm7, mm1
386    movq        mm5, mm3
387    pfadd       mm1, mm2                ; mm1=data2=(02 12)
388    pfadd       mm3, mm4                ; mm3=data4=(04 14)
389    pfsub       mm7, mm2                ; mm7=data5=(05 15)
390    pfsub       mm5, mm4                ; mm5=data3=(03 13)
391
392    movq        mm2, [GOTOFF(ebx,PD_RNDINT_MAGIC)]  ; mm2=[PD_RNDINT_MAGIC]
393    pcmpeqd     mm4, mm4
394    psrld       mm4, WORD_BIT           ; mm4={0xFFFF 0x0000 0xFFFF 0x0000}
395
396    pfadd       mm3, mm2                ; mm3=roundint(data4/8)=(04 ** 14 **)
397    pfadd       mm7, mm2                ; mm7=roundint(data5/8)=(05 ** 15 **)
398    pfadd       mm1, mm2                ; mm1=roundint(data2/8)=(02 ** 12 **)
399    pfadd       mm5, mm2                ; mm5=roundint(data3/8)=(03 ** 13 **)
400
401    pand        mm3, mm4                ; mm3=(04 -- 14 --)
402    pslld       mm7, WORD_BIT           ; mm7=(-- 05 -- 15)
403    pand        mm1, mm4                ; mm1=(02 -- 12 --)
404    pslld       mm5, WORD_BIT           ; mm5=(-- 03 -- 13)
405    por         mm3, mm7                ; mm3=(04 05 14 15)
406    por         mm1, mm5                ; mm1=(02 03 12 13)
407
408    movq        mm2, [GOTOFF(ebx,PB_CENTERJSAMP)]  ; mm2=[PB_CENTERJSAMP]
409
410    packsswb    mm6, mm3                ; mm6=(00 01 10 11 04 05 14 15)
411    packsswb    mm1, mm0                ; mm1=(02 03 12 13 06 07 16 17)
412    paddb       mm6, mm2
413    paddb       mm1, mm2
414
415    movq        mm4, mm6                ; transpose coefficients(phase 2)
416    punpcklwd   mm6, mm1                ; mm6=(00 01 02 03 10 11 12 13)
417    punpckhwd   mm4, mm1                ; mm4=(04 05 06 07 14 15 16 17)
418
419    movq        mm7, mm6                ; transpose coefficients(phase 3)
420    punpckldq   mm6, mm4                ; mm6=(00 01 02 03 04 05 06 07)
421    punpckhdq   mm7, mm4                ; mm7=(10 11 12 13 14 15 16 17)
422
423    pushpic     ebx                     ; save GOT address
424
425    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
426    mov         ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
427    movq        MMWORD [edx+eax*SIZEOF_JSAMPLE], mm6
428    movq        MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm7
429
430    poppic      ebx                     ; restore GOT address
431
432    add         esi, byte 2*SIZEOF_FAST_FLOAT  ; wsptr
433    add         edi, byte 2*SIZEOF_JSAMPROW
434    dec         ecx                            ; ctr
435    jnz         near .rowloop
436
437    femms                               ; empty MMX/3DNow! state
438
439    pop         edi
440    pop         esi
441;   pop         edx                     ; need not be preserved
442;   pop         ecx                     ; need not be preserved
443    pop         ebx
444    mov         esp, ebp                ; esp <- aligned ebp
445    pop         esp                     ; esp <- original ebp
446    pop         ebp
447    ret
448
449; For some reason, the OS X linker does not honor the request to align the
450; segment unless we do this.
451    align       32
452