1;
2; jidctflt.asm - floating-point IDCT (SSE & 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; [TAB8]
22
23%include "jsimdext.inc"
24%include "jdct.inc"
25
26; --------------------------------------------------------------------------
27
28%macro unpcklps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
29    shufps      %1, %2, 0x44
30%endmacro
31
32%macro unpckhps2 2  ; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
33    shufps      %1, %2, 0xEE
34%endmacro
35
36; --------------------------------------------------------------------------
37    SECTION     SEG_CONST
38
39    alignz      32
40    GLOBAL_DATA(jconst_idct_float_sse)
41
42EXTN(jconst_idct_float_sse):
43
44PD_1_414       times 4 dd  1.414213562373095048801689
45PD_1_847       times 4 dd  1.847759065022573512256366
46PD_1_082       times 4 dd  1.082392200292393968799446
47PD_M2_613      times 4 dd -2.613125929752753055713286
48PD_0_125       times 4 dd  0.125        ; 1/8
49PB_CENTERJSAMP times 8 db  CENTERJSAMPLE
50
51    alignz      32
52
53; --------------------------------------------------------------------------
54    SECTION     SEG_TEXT
55    BITS        32
56;
57; Perform dequantization and inverse DCT on one block of coefficients.
58;
59; GLOBAL(void)
60; jsimd_idct_float_sse(void *dct_table, JCOEFPTR coef_block,
61;                      JSAMPARRAY output_buf, JDIMENSION output_col)
62;
63
64%define dct_table(b)   (b) + 8          ; void *dct_table
65%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
66%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
67%define output_col(b)  (b) + 20         ; JDIMENSION output_col
68
69%define original_ebp   ebp + 0
70%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
71                                        ; xmmword wk[WK_NUM]
72%define WK_NUM         2
73%define workspace      wk(0) - DCTSIZE2 * SIZEOF_FAST_FLOAT
74                                        ; FAST_FLOAT workspace[DCTSIZE2]
75
76    align       32
77    GLOBAL_FUNCTION(jsimd_idct_float_sse)
78
79EXTN(jsimd_idct_float_sse):
80    push        ebp
81    mov         eax, esp                     ; eax = original ebp
82    sub         esp, byte 4
83    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
84    mov         [esp], eax
85    mov         ebp, esp                     ; ebp = aligned ebp
86    lea         esp, [workspace]
87    push        ebx
88;   push        ecx                     ; need not be preserved
89;   push        edx                     ; need not be preserved
90    push        esi
91    push        edi
92
93    get_GOT     ebx                     ; get GOT address
94
95    ; ---- Pass 1: process columns from input, store into work array.
96
97;   mov         eax, [original_ebp]
98    mov         edx, POINTER [dct_table(eax)]    ; quantptr
99    mov         esi, JCOEFPTR [coef_block(eax)]  ; inptr
100    lea         edi, [workspace]                 ; FAST_FLOAT *wsptr
101    mov         ecx, DCTSIZE/4                   ; ctr
102    alignx      16, 7
103.columnloop:
104%ifndef NO_ZERO_COLUMN_TEST_FLOAT_SSE
105    mov         eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
106    or          eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
107    jnz         near .columnDCT
108
109    movq        mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
110    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
111    por         mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
112    por         mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
113    por         mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
114    por         mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
115    por         mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
116    por         mm1, mm0
117    packsswb    mm1, mm1
118    movd        eax, mm1
119    test        eax, eax
120    jnz         short .columnDCT
121
122    ; -- AC terms all zero
123
124    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
125
126    punpckhwd   mm1, mm0                   ; mm1=(** 02 ** 03)
127    punpcklwd   mm0, mm0                   ; mm0=(00 00 01 01)
128    psrad       mm1, (DWORD_BIT-WORD_BIT)  ; mm1=in0H=(02 03)
129    psrad       mm0, (DWORD_BIT-WORD_BIT)  ; mm0=in0L=(00 01)
130    cvtpi2ps    xmm3, mm1                  ; xmm3=(02 03 ** **)
131    cvtpi2ps    xmm0, mm0                  ; xmm0=(00 01 ** **)
132    movlhps     xmm0, xmm3                 ; xmm0=in0=(00 01 02 03)
133
134    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
135
136    movaps      xmm1, xmm0
137    movaps      xmm2, xmm0
138    movaps      xmm3, xmm0
139
140    shufps      xmm0, xmm0, 0x00        ; xmm0=(00 00 00 00)
141    shufps      xmm1, xmm1, 0x55        ; xmm1=(01 01 01 01)
142    shufps      xmm2, xmm2, 0xAA        ; xmm2=(02 02 02 02)
143    shufps      xmm3, xmm3, 0xFF        ; xmm3=(03 03 03 03)
144
145    movaps      XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm0
146    movaps      XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
147    movaps      XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm1
148    movaps      XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
149    movaps      XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm2
150    movaps      XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm2
151    movaps      XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm3
152    movaps      XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
153    jmp         near .nextcolumn
154    alignx      16, 7
155%endif
156.columnDCT:
157
158    ; -- Even part
159
160    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
161    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
162    movq        mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
163    movq        mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
164
165    punpckhwd   mm4, mm0                ; mm4=(** 02 ** 03)
166    punpcklwd   mm0, mm0                ; mm0=(00 00 01 01)
167    punpckhwd   mm5, mm1                ; mm5=(** 22 ** 23)
168    punpcklwd   mm1, mm1                ; mm1=(20 20 21 21)
169
170    psrad       mm4, (DWORD_BIT-WORD_BIT)  ; mm4=in0H=(02 03)
171    psrad       mm0, (DWORD_BIT-WORD_BIT)  ; mm0=in0L=(00 01)
172    cvtpi2ps    xmm4, mm4                  ; xmm4=(02 03 ** **)
173    cvtpi2ps    xmm0, mm0                  ; xmm0=(00 01 ** **)
174    psrad       mm5, (DWORD_BIT-WORD_BIT)  ; mm5=in2H=(22 23)
175    psrad       mm1, (DWORD_BIT-WORD_BIT)  ; mm1=in2L=(20 21)
176    cvtpi2ps    xmm5, mm5                  ; xmm5=(22 23 ** **)
177    cvtpi2ps    xmm1, mm1                  ; xmm1=(20 21 ** **)
178
179    punpckhwd   mm6, mm2                ; mm6=(** 42 ** 43)
180    punpcklwd   mm2, mm2                ; mm2=(40 40 41 41)
181    punpckhwd   mm7, mm3                ; mm7=(** 62 ** 63)
182    punpcklwd   mm3, mm3                ; mm3=(60 60 61 61)
183
184    psrad       mm6, (DWORD_BIT-WORD_BIT)  ; mm6=in4H=(42 43)
185    psrad       mm2, (DWORD_BIT-WORD_BIT)  ; mm2=in4L=(40 41)
186    cvtpi2ps    xmm6, mm6                  ; xmm6=(42 43 ** **)
187    cvtpi2ps    xmm2, mm2                  ; xmm2=(40 41 ** **)
188    psrad       mm7, (DWORD_BIT-WORD_BIT)  ; mm7=in6H=(62 63)
189    psrad       mm3, (DWORD_BIT-WORD_BIT)  ; mm3=in6L=(60 61)
190    cvtpi2ps    xmm7, mm7                  ; xmm7=(62 63 ** **)
191    cvtpi2ps    xmm3, mm3                  ; xmm3=(60 61 ** **)
192
193    movlhps     xmm0, xmm4              ; xmm0=in0=(00 01 02 03)
194    movlhps     xmm1, xmm5              ; xmm1=in2=(20 21 22 23)
195    mulps       xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
196    mulps       xmm1, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
197
198    movlhps     xmm2, xmm6              ; xmm2=in4=(40 41 42 43)
199    movlhps     xmm3, xmm7              ; xmm3=in6=(60 61 62 63)
200    mulps       xmm2, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
201    mulps       xmm3, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
202
203    movaps      xmm4, xmm0
204    movaps      xmm5, xmm1
205    subps       xmm0, xmm2              ; xmm0=tmp11
206    subps       xmm1, xmm3
207    addps       xmm4, xmm2              ; xmm4=tmp10
208    addps       xmm5, xmm3              ; xmm5=tmp13
209
210    mulps       xmm1, [GOTOFF(ebx,PD_1_414)]
211    subps       xmm1, xmm5              ; xmm1=tmp12
212
213    movaps      xmm6, xmm4
214    movaps      xmm7, xmm0
215    subps       xmm4, xmm5              ; xmm4=tmp3
216    subps       xmm0, xmm1              ; xmm0=tmp2
217    addps       xmm6, xmm5              ; xmm6=tmp0
218    addps       xmm7, xmm1              ; xmm7=tmp1
219
220    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
221    movaps      XMMWORD [wk(0)], xmm0   ; tmp2
222
223    ; -- Odd part
224
225    movq        mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
226    movq        mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
227    movq        mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
228    movq        mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
229
230    punpckhwd   mm6, mm4                ; mm6=(** 12 ** 13)
231    punpcklwd   mm4, mm4                ; mm4=(10 10 11 11)
232    punpckhwd   mm2, mm0                ; mm2=(** 32 ** 33)
233    punpcklwd   mm0, mm0                ; mm0=(30 30 31 31)
234
235    psrad       mm6, (DWORD_BIT-WORD_BIT)  ; mm6=in1H=(12 13)
236    psrad       mm4, (DWORD_BIT-WORD_BIT)  ; mm4=in1L=(10 11)
237    cvtpi2ps    xmm4, mm6                  ; xmm4=(12 13 ** **)
238    cvtpi2ps    xmm2, mm4                  ; xmm2=(10 11 ** **)
239    psrad       mm2, (DWORD_BIT-WORD_BIT)  ; mm2=in3H=(32 33)
240    psrad       mm0, (DWORD_BIT-WORD_BIT)  ; mm0=in3L=(30 31)
241    cvtpi2ps    xmm0, mm2                  ; xmm0=(32 33 ** **)
242    cvtpi2ps    xmm3, mm0                  ; xmm3=(30 31 ** **)
243
244    punpckhwd   mm7, mm5                ; mm7=(** 52 ** 53)
245    punpcklwd   mm5, mm5                ; mm5=(50 50 51 51)
246    punpckhwd   mm3, mm1                ; mm3=(** 72 ** 73)
247    punpcklwd   mm1, mm1                ; mm1=(70 70 71 71)
248
249    movlhps     xmm2, xmm4              ; xmm2=in1=(10 11 12 13)
250    movlhps     xmm3, xmm0              ; xmm3=in3=(30 31 32 33)
251
252    psrad       mm7, (DWORD_BIT-WORD_BIT)  ; mm7=in5H=(52 53)
253    psrad       mm5, (DWORD_BIT-WORD_BIT)  ; mm5=in5L=(50 51)
254    cvtpi2ps    xmm4, mm7                  ; xmm4=(52 53 ** **)
255    cvtpi2ps    xmm5, mm5                  ; xmm5=(50 51 ** **)
256    psrad       mm3, (DWORD_BIT-WORD_BIT)  ; mm3=in7H=(72 73)
257    psrad       mm1, (DWORD_BIT-WORD_BIT)  ; mm1=in7L=(70 71)
258    cvtpi2ps    xmm0, mm3                  ; xmm0=(72 73 ** **)
259    cvtpi2ps    xmm1, mm1                  ; xmm1=(70 71 ** **)
260
261    mulps       xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
262    mulps       xmm3, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
263
264    movlhps     xmm5, xmm4              ; xmm5=in5=(50 51 52 53)
265    movlhps     xmm1, xmm0              ; xmm1=in7=(70 71 72 73)
266    mulps       xmm5, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
267    mulps       xmm1, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FLOAT_MULT_TYPE)]
268
269    movaps      xmm4, xmm2
270    movaps      xmm0, xmm5
271    addps       xmm2, xmm1              ; xmm2=z11
272    addps       xmm5, xmm3              ; xmm5=z13
273    subps       xmm4, xmm1              ; xmm4=z12
274    subps       xmm0, xmm3              ; xmm0=z10
275
276    movaps      xmm1, xmm2
277    subps       xmm2, xmm5
278    addps       xmm1, xmm5              ; xmm1=tmp7
279
280    mulps       xmm2, [GOTOFF(ebx,PD_1_414)]  ; xmm2=tmp11
281
282    movaps      xmm3, xmm0
283    addps       xmm0, xmm4
284    mulps       xmm0, [GOTOFF(ebx,PD_1_847)]   ; xmm0=z5
285    mulps       xmm3, [GOTOFF(ebx,PD_M2_613)]  ; xmm3=(z10 * -2.613125930)
286    mulps       xmm4, [GOTOFF(ebx,PD_1_082)]   ; xmm4=(z12 * 1.082392200)
287    addps       xmm3, xmm0                     ; xmm3=tmp12
288    subps       xmm4, xmm0                     ; xmm4=tmp10
289
290    ; -- Final output stage
291
292    subps       xmm3, xmm1              ; xmm3=tmp6
293    movaps      xmm5, xmm6
294    movaps      xmm0, xmm7
295    addps       xmm6, xmm1              ; xmm6=data0=(00 01 02 03)
296    addps       xmm7, xmm3              ; xmm7=data1=(10 11 12 13)
297    subps       xmm5, xmm1              ; xmm5=data7=(70 71 72 73)
298    subps       xmm0, xmm3              ; xmm0=data6=(60 61 62 63)
299    subps       xmm2, xmm3              ; xmm2=tmp5
300
301    movaps      xmm1, xmm6              ; transpose coefficients(phase 1)
302    unpcklps    xmm6, xmm7              ; xmm6=(00 10 01 11)
303    unpckhps    xmm1, xmm7              ; xmm1=(02 12 03 13)
304    movaps      xmm3, xmm0              ; transpose coefficients(phase 1)
305    unpcklps    xmm0, xmm5              ; xmm0=(60 70 61 71)
306    unpckhps    xmm3, xmm5              ; xmm3=(62 72 63 73)
307
308    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp2
309    movaps      xmm5, XMMWORD [wk(1)]   ; xmm5=tmp3
310
311    movaps      XMMWORD [wk(0)], xmm0   ; wk(0)=(60 70 61 71)
312    movaps      XMMWORD [wk(1)], xmm3   ; wk(1)=(62 72 63 73)
313
314    addps       xmm4, xmm2              ; xmm4=tmp4
315    movaps      xmm0, xmm7
316    movaps      xmm3, xmm5
317    addps       xmm7, xmm2              ; xmm7=data2=(20 21 22 23)
318    addps       xmm5, xmm4              ; xmm5=data4=(40 41 42 43)
319    subps       xmm0, xmm2              ; xmm0=data5=(50 51 52 53)
320    subps       xmm3, xmm4              ; xmm3=data3=(30 31 32 33)
321
322    movaps      xmm2, xmm7              ; transpose coefficients(phase 1)
323    unpcklps    xmm7, xmm3              ; xmm7=(20 30 21 31)
324    unpckhps    xmm2, xmm3              ; xmm2=(22 32 23 33)
325    movaps      xmm4, xmm5              ; transpose coefficients(phase 1)
326    unpcklps    xmm5, xmm0              ; xmm5=(40 50 41 51)
327    unpckhps    xmm4, xmm0              ; xmm4=(42 52 43 53)
328
329    movaps      xmm3, xmm6              ; transpose coefficients(phase 2)
330    unpcklps2   xmm6, xmm7              ; xmm6=(00 10 20 30)
331    unpckhps2   xmm3, xmm7              ; xmm3=(01 11 21 31)
332    movaps      xmm0, xmm1              ; transpose coefficients(phase 2)
333    unpcklps2   xmm1, xmm2              ; xmm1=(02 12 22 32)
334    unpckhps2   xmm0, xmm2              ; xmm0=(03 13 23 33)
335
336    movaps      xmm7, XMMWORD [wk(0)]   ; xmm7=(60 70 61 71)
337    movaps      xmm2, XMMWORD [wk(1)]   ; xmm2=(62 72 63 73)
338
339    movaps      XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm6
340    movaps      XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
341    movaps      XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_FAST_FLOAT)], xmm1
342    movaps      XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_FAST_FLOAT)], xmm0
343
344    movaps      xmm6, xmm5              ; transpose coefficients(phase 2)
345    unpcklps2   xmm5, xmm7              ; xmm5=(40 50 60 70)
346    unpckhps2   xmm6, xmm7              ; xmm6=(41 51 61 71)
347    movaps      xmm3, xmm4              ; transpose coefficients(phase 2)
348    unpcklps2   xmm4, xmm2              ; xmm4=(42 52 62 72)
349    unpckhps2   xmm3, xmm2              ; xmm3=(43 53 63 73)
350
351    movaps      XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm5
352    movaps      XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm6
353    movaps      XMMWORD [XMMBLOCK(2,1,edi,SIZEOF_FAST_FLOAT)], xmm4
354    movaps      XMMWORD [XMMBLOCK(3,1,edi,SIZEOF_FAST_FLOAT)], xmm3
355
356.nextcolumn:
357    add         esi, byte 4*SIZEOF_JCOEF               ; coef_block
358    add         edx, byte 4*SIZEOF_FLOAT_MULT_TYPE     ; quantptr
359    add         edi,      4*DCTSIZE*SIZEOF_FAST_FLOAT  ; wsptr
360    dec         ecx                                    ; ctr
361    jnz         near .columnloop
362
363    ; -- Prefetch the next coefficient block
364
365    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 0*32]
366    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 1*32]
367    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 2*32]
368    prefetchnta [esi + (DCTSIZE2-8)*SIZEOF_JCOEF + 3*32]
369
370    ; ---- Pass 2: process rows from work array, store into output array.
371
372    mov         eax, [original_ebp]
373    lea         esi, [workspace]                   ; FAST_FLOAT *wsptr
374    mov         edi, JSAMPARRAY [output_buf(eax)]  ; (JSAMPROW *)
375    mov         eax, JDIMENSION [output_col(eax)]
376    mov         ecx, DCTSIZE/4                     ; ctr
377    alignx      16, 7
378.rowloop:
379
380    ; -- Even part
381
382    movaps      xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
383    movaps      xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_FAST_FLOAT)]
384    movaps      xmm2, XMMWORD [XMMBLOCK(4,0,esi,SIZEOF_FAST_FLOAT)]
385    movaps      xmm3, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_FAST_FLOAT)]
386
387    movaps      xmm4, xmm0
388    movaps      xmm5, xmm1
389    subps       xmm0, xmm2              ; xmm0=tmp11
390    subps       xmm1, xmm3
391    addps       xmm4, xmm2              ; xmm4=tmp10
392    addps       xmm5, xmm3              ; xmm5=tmp13
393
394    mulps       xmm1, [GOTOFF(ebx,PD_1_414)]
395    subps       xmm1, xmm5              ; xmm1=tmp12
396
397    movaps      xmm6, xmm4
398    movaps      xmm7, xmm0
399    subps       xmm4, xmm5              ; xmm4=tmp3
400    subps       xmm0, xmm1              ; xmm0=tmp2
401    addps       xmm6, xmm5              ; xmm6=tmp0
402    addps       xmm7, xmm1              ; xmm7=tmp1
403
404    movaps      XMMWORD [wk(1)], xmm4   ; tmp3
405    movaps      XMMWORD [wk(0)], xmm0   ; tmp2
406
407    ; -- Odd part
408
409    movaps      xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
410    movaps      xmm3, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_FAST_FLOAT)]
411    movaps      xmm5, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_FAST_FLOAT)]
412    movaps      xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_FAST_FLOAT)]
413
414    movaps      xmm4, xmm2
415    movaps      xmm0, xmm5
416    addps       xmm2, xmm1              ; xmm2=z11
417    addps       xmm5, xmm3              ; xmm5=z13
418    subps       xmm4, xmm1              ; xmm4=z12
419    subps       xmm0, xmm3              ; xmm0=z10
420
421    movaps      xmm1, xmm2
422    subps       xmm2, xmm5
423    addps       xmm1, xmm5              ; xmm1=tmp7
424
425    mulps       xmm2, [GOTOFF(ebx,PD_1_414)]  ; xmm2=tmp11
426
427    movaps      xmm3, xmm0
428    addps       xmm0, xmm4
429    mulps       xmm0, [GOTOFF(ebx,PD_1_847)]   ; xmm0=z5
430    mulps       xmm3, [GOTOFF(ebx,PD_M2_613)]  ; xmm3=(z10 * -2.613125930)
431    mulps       xmm4, [GOTOFF(ebx,PD_1_082)]   ; xmm4=(z12 * 1.082392200)
432    addps       xmm3, xmm0                     ; xmm3=tmp12
433    subps       xmm4, xmm0                     ; xmm4=tmp10
434
435    ; -- Final output stage
436
437    subps       xmm3, xmm1              ; xmm3=tmp6
438    movaps      xmm5, xmm6
439    movaps      xmm0, xmm7
440    addps       xmm6, xmm1              ; xmm6=data0=(00 10 20 30)
441    addps       xmm7, xmm3              ; xmm7=data1=(01 11 21 31)
442    subps       xmm5, xmm1              ; xmm5=data7=(07 17 27 37)
443    subps       xmm0, xmm3              ; xmm0=data6=(06 16 26 36)
444    subps       xmm2, xmm3              ; xmm2=tmp5
445
446    movaps      xmm1, [GOTOFF(ebx,PD_0_125)]  ; xmm1=[PD_0_125]
447
448    mulps       xmm6, xmm1              ; descale(1/8)
449    mulps       xmm7, xmm1              ; descale(1/8)
450    mulps       xmm5, xmm1              ; descale(1/8)
451    mulps       xmm0, xmm1              ; descale(1/8)
452
453    movhlps     xmm3, xmm6
454    movhlps     xmm1, xmm7
455    cvtps2pi    mm0, xmm6               ; round to int32, mm0=data0L=(00 10)
456    cvtps2pi    mm1, xmm7               ; round to int32, mm1=data1L=(01 11)
457    cvtps2pi    mm2, xmm3               ; round to int32, mm2=data0H=(20 30)
458    cvtps2pi    mm3, xmm1               ; round to int32, mm3=data1H=(21 31)
459    packssdw    mm0, mm2                ; mm0=data0=(00 10 20 30)
460    packssdw    mm1, mm3                ; mm1=data1=(01 11 21 31)
461
462    movhlps     xmm6, xmm5
463    movhlps     xmm7, xmm0
464    cvtps2pi    mm4, xmm5               ; round to int32, mm4=data7L=(07 17)
465    cvtps2pi    mm5, xmm0               ; round to int32, mm5=data6L=(06 16)
466    cvtps2pi    mm6, xmm6               ; round to int32, mm6=data7H=(27 37)
467    cvtps2pi    mm7, xmm7               ; round to int32, mm7=data6H=(26 36)
468    packssdw    mm4, mm6                ; mm4=data7=(07 17 27 37)
469    packssdw    mm5, mm7                ; mm5=data6=(06 16 26 36)
470
471    packsswb    mm0, mm5                ; mm0=(00 10 20 30 06 16 26 36)
472    packsswb    mm1, mm4                ; mm1=(01 11 21 31 07 17 27 37)
473
474    movaps      xmm3, XMMWORD [wk(0)]   ; xmm3=tmp2
475    movaps      xmm1, XMMWORD [wk(1)]   ; xmm1=tmp3
476
477    movaps      xmm6, [GOTOFF(ebx,PD_0_125)]  ; xmm6=[PD_0_125]
478
479    addps       xmm4, xmm2              ; xmm4=tmp4
480    movaps      xmm5, xmm3
481    movaps      xmm0, xmm1
482    addps       xmm3, xmm2              ; xmm3=data2=(02 12 22 32)
483    addps       xmm1, xmm4              ; xmm1=data4=(04 14 24 34)
484    subps       xmm5, xmm2              ; xmm5=data5=(05 15 25 35)
485    subps       xmm0, xmm4              ; xmm0=data3=(03 13 23 33)
486
487    mulps       xmm3, xmm6              ; descale(1/8)
488    mulps       xmm1, xmm6              ; descale(1/8)
489    mulps       xmm5, xmm6              ; descale(1/8)
490    mulps       xmm0, xmm6              ; descale(1/8)
491
492    movhlps     xmm7, xmm3
493    movhlps     xmm2, xmm1
494    cvtps2pi    mm2, xmm3               ; round to int32, mm2=data2L=(02 12)
495    cvtps2pi    mm3, xmm1               ; round to int32, mm3=data4L=(04 14)
496    cvtps2pi    mm6, xmm7               ; round to int32, mm6=data2H=(22 32)
497    cvtps2pi    mm7, xmm2               ; round to int32, mm7=data4H=(24 34)
498    packssdw    mm2, mm6                ; mm2=data2=(02 12 22 32)
499    packssdw    mm3, mm7                ; mm3=data4=(04 14 24 34)
500
501    movhlps     xmm4, xmm5
502    movhlps     xmm6, xmm0
503    cvtps2pi    mm5, xmm5               ; round to int32, mm5=data5L=(05 15)
504    cvtps2pi    mm4, xmm0               ; round to int32, mm4=data3L=(03 13)
505    cvtps2pi    mm6, xmm4               ; round to int32, mm6=data5H=(25 35)
506    cvtps2pi    mm7, xmm6               ; round to int32, mm7=data3H=(23 33)
507    packssdw    mm5, mm6                ; mm5=data5=(05 15 25 35)
508    packssdw    mm4, mm7                ; mm4=data3=(03 13 23 33)
509
510    movq        mm6, [GOTOFF(ebx,PB_CENTERJSAMP)]  ; mm6=[PB_CENTERJSAMP]
511
512    packsswb    mm2, mm3                ; mm2=(02 12 22 32 04 14 24 34)
513    packsswb    mm4, mm5                ; mm4=(03 13 23 33 05 15 25 35)
514
515    paddb       mm0, mm6
516    paddb       mm1, mm6
517    paddb       mm2, mm6
518    paddb       mm4, mm6
519
520    movq        mm7, mm0                ; transpose coefficients(phase 1)
521    punpcklbw   mm0, mm1                ; mm0=(00 01 10 11 20 21 30 31)
522    punpckhbw   mm7, mm1                ; mm7=(06 07 16 17 26 27 36 37)
523    movq        mm3, mm2                ; transpose coefficients(phase 1)
524    punpcklbw   mm2, mm4                ; mm2=(02 03 12 13 22 23 32 33)
525    punpckhbw   mm3, mm4                ; mm3=(04 05 14 15 24 25 34 35)
526
527    movq        mm5, mm0                ; transpose coefficients(phase 2)
528    punpcklwd   mm0, mm2                ; mm0=(00 01 02 03 10 11 12 13)
529    punpckhwd   mm5, mm2                ; mm5=(20 21 22 23 30 31 32 33)
530    movq        mm6, mm3                ; transpose coefficients(phase 2)
531    punpcklwd   mm3, mm7                ; mm3=(04 05 06 07 14 15 16 17)
532    punpckhwd   mm6, mm7                ; mm6=(24 25 26 27 34 35 36 37)
533
534    movq        mm1, mm0                ; transpose coefficients(phase 3)
535    punpckldq   mm0, mm3                ; mm0=(00 01 02 03 04 05 06 07)
536    punpckhdq   mm1, mm3                ; mm1=(10 11 12 13 14 15 16 17)
537    movq        mm4, mm5                ; transpose coefficients(phase 3)
538    punpckldq   mm5, mm6                ; mm5=(20 21 22 23 24 25 26 27)
539    punpckhdq   mm4, mm6                ; mm4=(30 31 32 33 34 35 36 37)
540
541    pushpic     ebx                     ; save GOT address
542
543    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
544    mov         ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
545    movq        MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
546    movq        MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
547    mov         edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
548    mov         ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
549    movq        MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
550    movq        MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
551
552    poppic      ebx                     ; restore GOT address
553
554    add         esi, byte 4*SIZEOF_FAST_FLOAT  ; wsptr
555    add         edi, byte 4*SIZEOF_JSAMPROW
556    dec         ecx                            ; ctr
557    jnz         near .rowloop
558
559    emms                                ; empty MMX state
560
561    pop         edi
562    pop         esi
563;   pop         edx                     ; need not be preserved
564;   pop         ecx                     ; need not be preserved
565    pop         ebx
566    mov         esp, ebp                ; esp <- aligned ebp
567    pop         esp                     ; esp <- original ebp
568    pop         ebp
569    ret
570
571; For some reason, the OS X linker does not honor the request to align the
572; segment unless we do this.
573    align       32
574