• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jdsample.asm - upsampling (SSE2)
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%include "jsimdext.inc"
18
19; --------------------------------------------------------------------------
20    SECTION     SEG_CONST
21
22    alignz      32
23    GLOBAL_DATA(jconst_fancy_upsample_sse2)
24
25EXTN(jconst_fancy_upsample_sse2):
26
27PW_ONE   times 8 dw 1
28PW_TWO   times 8 dw 2
29PW_THREE times 8 dw 3
30PW_SEVEN times 8 dw 7
31PW_EIGHT times 8 dw 8
32
33    alignz      32
34
35; --------------------------------------------------------------------------
36    SECTION     SEG_TEXT
37    BITS        32
38;
39; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
40;
41; The upsampling algorithm is linear interpolation between pixel centers,
42; also known as a "triangle filter".  This is a good compromise between
43; speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
44; of the way between input pixel centers.
45;
46; GLOBAL(void)
47; jsimd_h2v1_fancy_upsample_sse2(int max_v_samp_factor,
48;                                JDIMENSION downsampled_width,
49;                                JSAMPARRAY input_data,
50;                                JSAMPARRAY *output_data_ptr);
51;
52
53%define max_v_samp(b)       (b) + 8     ; int max_v_samp_factor
54%define downsamp_width(b)   (b) + 12    ; JDIMENSION downsampled_width
55%define input_data(b)       (b) + 16    ; JSAMPARRAY input_data
56%define output_data_ptr(b)  (b) + 20    ; JSAMPARRAY *output_data_ptr
57
58    align       32
59    GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_sse2)
60
61EXTN(jsimd_h2v1_fancy_upsample_sse2):
62    push        ebp
63    mov         ebp, esp
64    pushpic     ebx
65;   push        ecx                     ; need not be preserved
66;   push        edx                     ; need not be preserved
67    push        esi
68    push        edi
69
70    get_GOT     ebx                     ; get GOT address
71
72    mov         eax, JDIMENSION [downsamp_width(ebp)]  ; colctr
73    test        eax, eax
74    jz          near .return
75
76    mov         ecx, INT [max_v_samp(ebp)]  ; rowctr
77    test        ecx, ecx
78    jz          near .return
79
80    mov         esi, JSAMPARRAY [input_data(ebp)]    ; input_data
81    mov         edi, POINTER [output_data_ptr(ebp)]
82    mov         edi, JSAMPARRAY [edi]                ; output_data
83    alignx      16, 7
84.rowloop:
85    push        eax                     ; colctr
86    push        edi
87    push        esi
88
89    mov         esi, JSAMPROW [esi]     ; inptr
90    mov         edi, JSAMPROW [edi]     ; outptr
91
92    test        eax, SIZEOF_XMMWORD-1
93    jz          short .skip
94    mov         dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
95    mov         JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
96.skip:
97    pxor        xmm0, xmm0              ; xmm0=(all 0's)
98    pcmpeqb     xmm7, xmm7
99    psrldq      xmm7, (SIZEOF_XMMWORD-1)
100    pand        xmm7, XMMWORD [esi+0*SIZEOF_XMMWORD]
101
102    add         eax, byte SIZEOF_XMMWORD-1
103    and         eax, byte -SIZEOF_XMMWORD
104    cmp         eax, byte SIZEOF_XMMWORD
105    ja          short .columnloop
106    alignx      16, 7
107
108.columnloop_last:
109    pcmpeqb     xmm6, xmm6
110    pslldq      xmm6, (SIZEOF_XMMWORD-1)
111    pand        xmm6, XMMWORD [esi+0*SIZEOF_XMMWORD]
112    jmp         short .upsample
113    alignx      16, 7
114
115.columnloop:
116    movdqa      xmm6, XMMWORD [esi+1*SIZEOF_XMMWORD]
117    pslldq      xmm6, (SIZEOF_XMMWORD-1)
118
119.upsample:
120    movdqa      xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
121    movdqa      xmm2, xmm1
122    movdqa      xmm3, xmm1                ; xmm1=( 0  1  2 ... 13 14 15)
123    pslldq      xmm2, 1                   ; xmm2=(--  0  1 ... 12 13 14)
124    psrldq      xmm3, 1                   ; xmm3=( 1  2  3 ... 14 15 --)
125
126    por         xmm2, xmm7                ; xmm2=(-1  0  1 ... 12 13 14)
127    por         xmm3, xmm6                ; xmm3=( 1  2  3 ... 14 15 16)
128
129    movdqa      xmm7, xmm1
130    psrldq      xmm7, (SIZEOF_XMMWORD-1)  ; xmm7=(15 -- -- ... -- -- --)
131
132    movdqa      xmm4, xmm1
133    punpcklbw   xmm1, xmm0                ; xmm1=( 0  1  2  3  4  5  6  7)
134    punpckhbw   xmm4, xmm0                ; xmm4=( 8  9 10 11 12 13 14 15)
135    movdqa      xmm5, xmm2
136    punpcklbw   xmm2, xmm0                ; xmm2=(-1  0  1  2  3  4  5  6)
137    punpckhbw   xmm5, xmm0                ; xmm5=( 7  8  9 10 11 12 13 14)
138    movdqa      xmm6, xmm3
139    punpcklbw   xmm3, xmm0                ; xmm3=( 1  2  3  4  5  6  7  8)
140    punpckhbw   xmm6, xmm0                ; xmm6=( 9 10 11 12 13 14 15 16)
141
142    pmullw      xmm1, [GOTOFF(ebx,PW_THREE)]
143    pmullw      xmm4, [GOTOFF(ebx,PW_THREE)]
144    paddw       xmm2, [GOTOFF(ebx,PW_ONE)]
145    paddw       xmm5, [GOTOFF(ebx,PW_ONE)]
146    paddw       xmm3, [GOTOFF(ebx,PW_TWO)]
147    paddw       xmm6, [GOTOFF(ebx,PW_TWO)]
148
149    paddw       xmm2, xmm1
150    paddw       xmm5, xmm4
151    psrlw       xmm2, 2                 ; xmm2=OutLE=( 0  2  4  6  8 10 12 14)
152    psrlw       xmm5, 2                 ; xmm5=OutHE=(16 18 20 22 24 26 28 30)
153    paddw       xmm3, xmm1
154    paddw       xmm6, xmm4
155    psrlw       xmm3, 2                 ; xmm3=OutLO=( 1  3  5  7  9 11 13 15)
156    psrlw       xmm6, 2                 ; xmm6=OutHO=(17 19 21 23 25 27 29 31)
157
158    psllw       xmm3, BYTE_BIT
159    psllw       xmm6, BYTE_BIT
160    por         xmm2, xmm3              ; xmm2=OutL=( 0  1  2 ... 13 14 15)
161    por         xmm5, xmm6              ; xmm5=OutH=(16 17 18 ... 29 30 31)
162
163    movdqa      XMMWORD [edi+0*SIZEOF_XMMWORD], xmm2
164    movdqa      XMMWORD [edi+1*SIZEOF_XMMWORD], xmm5
165
166    sub         eax, byte SIZEOF_XMMWORD
167    add         esi, byte 1*SIZEOF_XMMWORD  ; inptr
168    add         edi, byte 2*SIZEOF_XMMWORD  ; outptr
169    cmp         eax, byte SIZEOF_XMMWORD
170    ja          near .columnloop
171    test        eax, eax
172    jnz         near .columnloop_last
173
174    pop         esi
175    pop         edi
176    pop         eax
177
178    add         esi, byte SIZEOF_JSAMPROW  ; input_data
179    add         edi, byte SIZEOF_JSAMPROW  ; output_data
180    dec         ecx                        ; rowctr
181    jg          near .rowloop
182
183.return:
184    pop         edi
185    pop         esi
186;   pop         edx                     ; need not be preserved
187;   pop         ecx                     ; need not be preserved
188    poppic      ebx
189    pop         ebp
190    ret
191
192; --------------------------------------------------------------------------
193;
194; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
195; Again a triangle filter; see comments for h2v1 case, above.
196;
197; GLOBAL(void)
198; jsimd_h2v2_fancy_upsample_sse2(int max_v_samp_factor,
199;                                JDIMENSION downsampled_width,
200;                                JSAMPARRAY input_data,
201;                                JSAMPARRAY *output_data_ptr);
202;
203
204%define max_v_samp(b)       (b) + 8     ; int max_v_samp_factor
205%define downsamp_width(b)   (b) + 12    ; JDIMENSION downsampled_width
206%define input_data(b)       (b) + 16    ; JSAMPARRAY input_data
207%define output_data_ptr(b)  (b) + 20    ; JSAMPARRAY *output_data_ptr
208
209%define original_ebp  ebp + 0
210%define wk(i)         ebp - (WK_NUM - (i)) * SIZEOF_XMMWORD
211                                        ; xmmword wk[WK_NUM]
212%define WK_NUM        4
213%define gotptr        wk(0) - SIZEOF_POINTER  ; void *gotptr
214
215    align       32
216    GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_sse2)
217
218EXTN(jsimd_h2v2_fancy_upsample_sse2):
219    push        ebp
220    mov         eax, esp                     ; eax = original ebp
221    sub         esp, byte 4
222    and         esp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
223    mov         [esp], eax
224    mov         ebp, esp                     ; ebp = aligned ebp
225    lea         esp, [wk(0)]
226    pushpic     eax                     ; make a room for GOT address
227    push        ebx
228;   push        ecx                     ; need not be preserved
229;   push        edx                     ; need not be preserved
230    push        esi
231    push        edi
232
233    get_GOT     ebx                     ; get GOT address
234    movpic      POINTER [gotptr], ebx   ; save GOT address
235
236    mov         edx, eax                ; edx = original ebp
237    mov         eax, JDIMENSION [downsamp_width(edx)]  ; colctr
238    test        eax, eax
239    jz          near .return
240
241    mov         ecx, INT [max_v_samp(edx)]  ; rowctr
242    test        ecx, ecx
243    jz          near .return
244
245    mov         esi, JSAMPARRAY [input_data(edx)]    ; input_data
246    mov         edi, POINTER [output_data_ptr(edx)]
247    mov         edi, JSAMPARRAY [edi]                ; output_data
248    alignx      16, 7
249.rowloop:
250    push        eax                     ; colctr
251    push        ecx
252    push        edi
253    push        esi
254
255    mov         ecx, JSAMPROW [esi-1*SIZEOF_JSAMPROW]  ; inptr1(above)
256    mov         ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]  ; inptr0
257    mov         esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW]  ; inptr1(below)
258    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]  ; outptr0
259    mov         edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]  ; outptr1
260
261    test        eax, SIZEOF_XMMWORD-1
262    jz          short .skip
263    push        edx
264    mov         dl, JSAMPLE [ecx+(eax-1)*SIZEOF_JSAMPLE]
265    mov         JSAMPLE [ecx+eax*SIZEOF_JSAMPLE], dl
266    mov         dl, JSAMPLE [ebx+(eax-1)*SIZEOF_JSAMPLE]
267    mov         JSAMPLE [ebx+eax*SIZEOF_JSAMPLE], dl
268    mov         dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
269    mov         JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
270    pop         edx
271.skip:
272    ; -- process the first column block
273
274    movdqa      xmm0, XMMWORD [ebx+0*SIZEOF_XMMWORD]  ; xmm0=row[ 0][0]
275    movdqa      xmm1, XMMWORD [ecx+0*SIZEOF_XMMWORD]  ; xmm1=row[-1][0]
276    movdqa      xmm2, XMMWORD [esi+0*SIZEOF_XMMWORD]  ; xmm2=row[+1][0]
277
278    pushpic     ebx
279    movpic      ebx, POINTER [gotptr]   ; load GOT address
280
281    pxor        xmm3, xmm3              ; xmm3=(all 0's)
282    movdqa      xmm4, xmm0
283    punpcklbw   xmm0, xmm3              ; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
284    punpckhbw   xmm4, xmm3              ; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
285    movdqa      xmm5, xmm1
286    punpcklbw   xmm1, xmm3              ; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
287    punpckhbw   xmm5, xmm3              ; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
288    movdqa      xmm6, xmm2
289    punpcklbw   xmm2, xmm3              ; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
290    punpckhbw   xmm6, xmm3              ; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
291
292    pmullw      xmm0, [GOTOFF(ebx,PW_THREE)]
293    pmullw      xmm4, [GOTOFF(ebx,PW_THREE)]
294
295    pcmpeqb     xmm7, xmm7
296    psrldq      xmm7, (SIZEOF_XMMWORD-2)
297
298    paddw       xmm1, xmm0              ; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
299    paddw       xmm5, xmm4              ; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
300    paddw       xmm2, xmm0              ; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
301    paddw       xmm6, xmm4              ; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
302
303    movdqa      XMMWORD [edx+0*SIZEOF_XMMWORD], xmm1  ; temporarily save
304    movdqa      XMMWORD [edx+1*SIZEOF_XMMWORD], xmm5  ; the intermediate data
305    movdqa      XMMWORD [edi+0*SIZEOF_XMMWORD], xmm2
306    movdqa      XMMWORD [edi+1*SIZEOF_XMMWORD], xmm6
307
308    pand        xmm1, xmm7              ; xmm1=( 0 -- -- -- -- -- -- --)
309    pand        xmm2, xmm7              ; xmm2=( 0 -- -- -- -- -- -- --)
310
311    movdqa      XMMWORD [wk(0)], xmm1
312    movdqa      XMMWORD [wk(1)], xmm2
313
314    poppic      ebx
315
316    add         eax, byte SIZEOF_XMMWORD-1
317    and         eax, byte -SIZEOF_XMMWORD
318    cmp         eax, byte SIZEOF_XMMWORD
319    ja          short .columnloop
320    alignx      16, 7
321
322.columnloop_last:
323    ; -- process the last column block
324
325    pushpic     ebx
326    movpic      ebx, POINTER [gotptr]   ; load GOT address
327
328    pcmpeqb     xmm1, xmm1
329    pslldq      xmm1, (SIZEOF_XMMWORD-2)
330    movdqa      xmm2, xmm1
331
332    pand        xmm1, XMMWORD [edx+1*SIZEOF_XMMWORD]
333    pand        xmm2, XMMWORD [edi+1*SIZEOF_XMMWORD]
334
335    movdqa      XMMWORD [wk(2)], xmm1   ; xmm1=(-- -- -- -- -- -- -- 15)
336    movdqa      XMMWORD [wk(3)], xmm2   ; xmm2=(-- -- -- -- -- -- -- 15)
337
338    jmp         near .upsample
339    alignx      16, 7
340
341.columnloop:
342    ; -- process the next column block
343
344    movdqa      xmm0, XMMWORD [ebx+1*SIZEOF_XMMWORD]  ; xmm0=row[ 0][1]
345    movdqa      xmm1, XMMWORD [ecx+1*SIZEOF_XMMWORD]  ; xmm1=row[-1][1]
346    movdqa      xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]  ; xmm2=row[+1][1]
347
348    pushpic     ebx
349    movpic      ebx, POINTER [gotptr]   ; load GOT address
350
351    pxor        xmm3, xmm3              ; xmm3=(all 0's)
352    movdqa      xmm4, xmm0
353    punpcklbw   xmm0, xmm3              ; xmm0=row[ 0]( 0  1  2  3  4  5  6  7)
354    punpckhbw   xmm4, xmm3              ; xmm4=row[ 0]( 8  9 10 11 12 13 14 15)
355    movdqa      xmm5, xmm1
356    punpcklbw   xmm1, xmm3              ; xmm1=row[-1]( 0  1  2  3  4  5  6  7)
357    punpckhbw   xmm5, xmm3              ; xmm5=row[-1]( 8  9 10 11 12 13 14 15)
358    movdqa      xmm6, xmm2
359    punpcklbw   xmm2, xmm3              ; xmm2=row[+1]( 0  1  2  3  4  5  6  7)
360    punpckhbw   xmm6, xmm3              ; xmm6=row[+1]( 8  9 10 11 12 13 14 15)
361
362    pmullw      xmm0, [GOTOFF(ebx,PW_THREE)]
363    pmullw      xmm4, [GOTOFF(ebx,PW_THREE)]
364
365    paddw       xmm1, xmm0              ; xmm1=Int0L=( 0  1  2  3  4  5  6  7)
366    paddw       xmm5, xmm4              ; xmm5=Int0H=( 8  9 10 11 12 13 14 15)
367    paddw       xmm2, xmm0              ; xmm2=Int1L=( 0  1  2  3  4  5  6  7)
368    paddw       xmm6, xmm4              ; xmm6=Int1H=( 8  9 10 11 12 13 14 15)
369
370    movdqa      XMMWORD [edx+2*SIZEOF_XMMWORD], xmm1  ; temporarily save
371    movdqa      XMMWORD [edx+3*SIZEOF_XMMWORD], xmm5  ; the intermediate data
372    movdqa      XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
373    movdqa      XMMWORD [edi+3*SIZEOF_XMMWORD], xmm6
374
375    pslldq      xmm1, (SIZEOF_XMMWORD-2)  ; xmm1=(-- -- -- -- -- -- --  0)
376    pslldq      xmm2, (SIZEOF_XMMWORD-2)  ; xmm2=(-- -- -- -- -- -- --  0)
377
378    movdqa      XMMWORD [wk(2)], xmm1
379    movdqa      XMMWORD [wk(3)], xmm2
380
381.upsample:
382    ; -- process the upper row
383
384    movdqa      xmm7, XMMWORD [edx+0*SIZEOF_XMMWORD]
385    movdqa      xmm3, XMMWORD [edx+1*SIZEOF_XMMWORD]
386
387    movdqa      xmm0, xmm7                ; xmm7=Int0L=( 0  1  2  3  4  5  6  7)
388    movdqa      xmm4, xmm3                ; xmm3=Int0H=( 8  9 10 11 12 13 14 15)
389    psrldq      xmm0, 2                   ; xmm0=( 1  2  3  4  5  6  7 --)
390    pslldq      xmm4, (SIZEOF_XMMWORD-2)  ; xmm4=(-- -- -- -- -- -- --  8)
391    movdqa      xmm5, xmm7
392    movdqa      xmm6, xmm3
393    psrldq      xmm5, (SIZEOF_XMMWORD-2)  ; xmm5=( 7 -- -- -- -- -- -- --)
394    pslldq      xmm6, 2                   ; xmm6=(--  8  9 10 11 12 13 14)
395
396    por         xmm0, xmm4                ; xmm0=( 1  2  3  4  5  6  7  8)
397    por         xmm5, xmm6                ; xmm5=( 7  8  9 10 11 12 13 14)
398
399    movdqa      xmm1, xmm7
400    movdqa      xmm2, xmm3
401    pslldq      xmm1, 2                   ; xmm1=(--  0  1  2  3  4  5  6)
402    psrldq      xmm2, 2                   ; xmm2=( 9 10 11 12 13 14 15 --)
403    movdqa      xmm4, xmm3
404    psrldq      xmm4, (SIZEOF_XMMWORD-2)  ; xmm4=(15 -- -- -- -- -- -- --)
405
406    por         xmm1, XMMWORD [wk(0)]     ; xmm1=(-1  0  1  2  3  4  5  6)
407    por         xmm2, XMMWORD [wk(2)]     ; xmm2=( 9 10 11 12 13 14 15 16)
408
409    movdqa      XMMWORD [wk(0)], xmm4
410
411    pmullw      xmm7, [GOTOFF(ebx,PW_THREE)]
412    pmullw      xmm3, [GOTOFF(ebx,PW_THREE)]
413    paddw       xmm1, [GOTOFF(ebx,PW_EIGHT)]
414    paddw       xmm5, [GOTOFF(ebx,PW_EIGHT)]
415    paddw       xmm0, [GOTOFF(ebx,PW_SEVEN)]
416    paddw       xmm2, [GOTOFF(ebx,PW_SEVEN)]
417
418    paddw       xmm1, xmm7
419    paddw       xmm5, xmm3
420    psrlw       xmm1, 4                 ; xmm1=Out0LE=( 0  2  4  6  8 10 12 14)
421    psrlw       xmm5, 4                 ; xmm5=Out0HE=(16 18 20 22 24 26 28 30)
422    paddw       xmm0, xmm7
423    paddw       xmm2, xmm3
424    psrlw       xmm0, 4                 ; xmm0=Out0LO=( 1  3  5  7  9 11 13 15)
425    psrlw       xmm2, 4                 ; xmm2=Out0HO=(17 19 21 23 25 27 29 31)
426
427    psllw       xmm0, BYTE_BIT
428    psllw       xmm2, BYTE_BIT
429    por         xmm1, xmm0              ; xmm1=Out0L=( 0  1  2 ... 13 14 15)
430    por         xmm5, xmm2              ; xmm5=Out0H=(16 17 18 ... 29 30 31)
431
432    movdqa      XMMWORD [edx+0*SIZEOF_XMMWORD], xmm1
433    movdqa      XMMWORD [edx+1*SIZEOF_XMMWORD], xmm5
434
435    ; -- process the lower row
436
437    movdqa      xmm6, XMMWORD [edi+0*SIZEOF_XMMWORD]
438    movdqa      xmm4, XMMWORD [edi+1*SIZEOF_XMMWORD]
439
440    movdqa      xmm7, xmm6                ; xmm6=Int1L=( 0  1  2  3  4  5  6  7)
441    movdqa      xmm3, xmm4                ; xmm4=Int1H=( 8  9 10 11 12 13 14 15)
442    psrldq      xmm7, 2                   ; xmm7=( 1  2  3  4  5  6  7 --)
443    pslldq      xmm3, (SIZEOF_XMMWORD-2)  ; xmm3=(-- -- -- -- -- -- --  8)
444    movdqa      xmm0, xmm6
445    movdqa      xmm2, xmm4
446    psrldq      xmm0, (SIZEOF_XMMWORD-2)  ; xmm0=( 7 -- -- -- -- -- -- --)
447    pslldq      xmm2, 2                   ; xmm2=(--  8  9 10 11 12 13 14)
448
449    por         xmm7, xmm3                ; xmm7=( 1  2  3  4  5  6  7  8)
450    por         xmm0, xmm2                ; xmm0=( 7  8  9 10 11 12 13 14)
451
452    movdqa      xmm1, xmm6
453    movdqa      xmm5, xmm4
454    pslldq      xmm1, 2                   ; xmm1=(--  0  1  2  3  4  5  6)
455    psrldq      xmm5, 2                   ; xmm5=( 9 10 11 12 13 14 15 --)
456    movdqa      xmm3, xmm4
457    psrldq      xmm3, (SIZEOF_XMMWORD-2)  ; xmm3=(15 -- -- -- -- -- -- --)
458
459    por         xmm1, XMMWORD [wk(1)]     ; xmm1=(-1  0  1  2  3  4  5  6)
460    por         xmm5, XMMWORD [wk(3)]     ; xmm5=( 9 10 11 12 13 14 15 16)
461
462    movdqa      XMMWORD [wk(1)], xmm3
463
464    pmullw      xmm6, [GOTOFF(ebx,PW_THREE)]
465    pmullw      xmm4, [GOTOFF(ebx,PW_THREE)]
466    paddw       xmm1, [GOTOFF(ebx,PW_EIGHT)]
467    paddw       xmm0, [GOTOFF(ebx,PW_EIGHT)]
468    paddw       xmm7, [GOTOFF(ebx,PW_SEVEN)]
469    paddw       xmm5, [GOTOFF(ebx,PW_SEVEN)]
470
471    paddw       xmm1, xmm6
472    paddw       xmm0, xmm4
473    psrlw       xmm1, 4                 ; xmm1=Out1LE=( 0  2  4  6  8 10 12 14)
474    psrlw       xmm0, 4                 ; xmm0=Out1HE=(16 18 20 22 24 26 28 30)
475    paddw       xmm7, xmm6
476    paddw       xmm5, xmm4
477    psrlw       xmm7, 4                 ; xmm7=Out1LO=( 1  3  5  7  9 11 13 15)
478    psrlw       xmm5, 4                 ; xmm5=Out1HO=(17 19 21 23 25 27 29 31)
479
480    psllw       xmm7, BYTE_BIT
481    psllw       xmm5, BYTE_BIT
482    por         xmm1, xmm7              ; xmm1=Out1L=( 0  1  2 ... 13 14 15)
483    por         xmm0, xmm5              ; xmm0=Out1H=(16 17 18 ... 29 30 31)
484
485    movdqa      XMMWORD [edi+0*SIZEOF_XMMWORD], xmm1
486    movdqa      XMMWORD [edi+1*SIZEOF_XMMWORD], xmm0
487
488    poppic      ebx
489
490    sub         eax, byte SIZEOF_XMMWORD
491    add         ecx, byte 1*SIZEOF_XMMWORD  ; inptr1(above)
492    add         ebx, byte 1*SIZEOF_XMMWORD  ; inptr0
493    add         esi, byte 1*SIZEOF_XMMWORD  ; inptr1(below)
494    add         edx, byte 2*SIZEOF_XMMWORD  ; outptr0
495    add         edi, byte 2*SIZEOF_XMMWORD  ; outptr1
496    cmp         eax, byte SIZEOF_XMMWORD
497    ja          near .columnloop
498    test        eax, eax
499    jnz         near .columnloop_last
500
501    pop         esi
502    pop         edi
503    pop         ecx
504    pop         eax
505
506    add         esi, byte 1*SIZEOF_JSAMPROW  ; input_data
507    add         edi, byte 2*SIZEOF_JSAMPROW  ; output_data
508    sub         ecx, byte 2                  ; rowctr
509    jg          near .rowloop
510
511.return:
512    pop         edi
513    pop         esi
514;   pop         edx                     ; need not be preserved
515;   pop         ecx                     ; need not be preserved
516    pop         ebx
517    mov         esp, ebp                ; esp <- aligned ebp
518    pop         esp                     ; esp <- original ebp
519    pop         ebp
520    ret
521
522; --------------------------------------------------------------------------
523;
524; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
525; It's still a box filter.
526;
527; GLOBAL(void)
528; jsimd_h2v1_upsample_sse2(int max_v_samp_factor, JDIMENSION output_width,
529;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
530;
531
532%define max_v_samp(b)       (b) + 8     ; int max_v_samp_factor
533%define output_width(b)     (b) + 12    ; JDIMENSION output_width
534%define input_data(b)       (b) + 16    ; JSAMPARRAY input_data
535%define output_data_ptr(b)  (b) + 20    ; JSAMPARRAY *output_data_ptr
536
537    align       32
538    GLOBAL_FUNCTION(jsimd_h2v1_upsample_sse2)
539
540EXTN(jsimd_h2v1_upsample_sse2):
541    push        ebp
542    mov         ebp, esp
543;   push        ebx                     ; unused
544;   push        ecx                     ; need not be preserved
545;   push        edx                     ; need not be preserved
546    push        esi
547    push        edi
548
549    mov         edx, JDIMENSION [output_width(ebp)]
550    add         edx, byte (2*SIZEOF_XMMWORD)-1
551    and         edx, byte -(2*SIZEOF_XMMWORD)
552    jz          short .return
553
554    mov         ecx, INT [max_v_samp(ebp)]  ; rowctr
555    test        ecx, ecx
556    jz          short .return
557
558    mov         esi, JSAMPARRAY [input_data(ebp)]    ; input_data
559    mov         edi, POINTER [output_data_ptr(ebp)]
560    mov         edi, JSAMPARRAY [edi]                ; output_data
561    alignx      16, 7
562.rowloop:
563    push        edi
564    push        esi
565
566    mov         esi, JSAMPROW [esi]     ; inptr
567    mov         edi, JSAMPROW [edi]     ; outptr
568    mov         eax, edx                ; colctr
569    alignx      16, 7
570.columnloop:
571
572    movdqa      xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
573
574    movdqa      xmm1, xmm0
575    punpcklbw   xmm0, xmm0
576    punpckhbw   xmm1, xmm1
577
578    movdqa      XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
579    movdqa      XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1
580
581    sub         eax, byte 2*SIZEOF_XMMWORD
582    jz          short .nextrow
583
584    movdqa      xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]
585
586    movdqa      xmm3, xmm2
587    punpcklbw   xmm2, xmm2
588    punpckhbw   xmm3, xmm3
589
590    movdqa      XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
591    movdqa      XMMWORD [edi+3*SIZEOF_XMMWORD], xmm3
592
593    sub         eax, byte 2*SIZEOF_XMMWORD
594    jz          short .nextrow
595
596    add         esi, byte 2*SIZEOF_XMMWORD  ; inptr
597    add         edi, byte 4*SIZEOF_XMMWORD  ; outptr
598    jmp         short .columnloop
599    alignx      16, 7
600
601.nextrow:
602    pop         esi
603    pop         edi
604
605    add         esi, byte SIZEOF_JSAMPROW  ; input_data
606    add         edi, byte SIZEOF_JSAMPROW  ; output_data
607    dec         ecx                        ; rowctr
608    jg          short .rowloop
609
610.return:
611    pop         edi
612    pop         esi
613;   pop         edx                     ; need not be preserved
614;   pop         ecx                     ; need not be preserved
615;   pop         ebx                     ; unused
616    pop         ebp
617    ret
618
619; --------------------------------------------------------------------------
620;
621; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
622; It's still a box filter.
623;
624; GLOBAL(void)
625; jsimd_h2v2_upsample_sse2(int max_v_samp_factor, JDIMENSION output_width,
626;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
627;
628
629%define max_v_samp(b)       (b) + 8     ; int max_v_samp_factor
630%define output_width(b)     (b) + 12    ; JDIMENSION output_width
631%define input_data(b)       (b) + 16    ; JSAMPARRAY input_data
632%define output_data_ptr(b)  (b) + 20    ; JSAMPARRAY *output_data_ptr
633
634    align       32
635    GLOBAL_FUNCTION(jsimd_h2v2_upsample_sse2)
636
637EXTN(jsimd_h2v2_upsample_sse2):
638    push        ebp
639    mov         ebp, esp
640    push        ebx
641;   push        ecx                     ; need not be preserved
642;   push        edx                     ; need not be preserved
643    push        esi
644    push        edi
645
646    mov         edx, JDIMENSION [output_width(ebp)]
647    add         edx, byte (2*SIZEOF_XMMWORD)-1
648    and         edx, byte -(2*SIZEOF_XMMWORD)
649    jz          near .return
650
651    mov         ecx, INT [max_v_samp(ebp)]  ; rowctr
652    test        ecx, ecx
653    jz          near .return
654
655    mov         esi, JSAMPARRAY [input_data(ebp)]    ; input_data
656    mov         edi, POINTER [output_data_ptr(ebp)]
657    mov         edi, JSAMPARRAY [edi]                ; output_data
658    alignx      16, 7
659.rowloop:
660    push        edi
661    push        esi
662
663    mov         esi, JSAMPROW [esi]                    ; inptr
664    mov         ebx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]  ; outptr0
665    mov         edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]  ; outptr1
666    mov         eax, edx                               ; colctr
667    alignx      16, 7
668.columnloop:
669
670    movdqa      xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
671
672    movdqa      xmm1, xmm0
673    punpcklbw   xmm0, xmm0
674    punpckhbw   xmm1, xmm1
675
676    movdqa      XMMWORD [ebx+0*SIZEOF_XMMWORD], xmm0
677    movdqa      XMMWORD [ebx+1*SIZEOF_XMMWORD], xmm1
678    movdqa      XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
679    movdqa      XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1
680
681    sub         eax, byte 2*SIZEOF_XMMWORD
682    jz          short .nextrow
683
684    movdqa      xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]
685
686    movdqa      xmm3, xmm2
687    punpcklbw   xmm2, xmm2
688    punpckhbw   xmm3, xmm3
689
690    movdqa      XMMWORD [ebx+2*SIZEOF_XMMWORD], xmm2
691    movdqa      XMMWORD [ebx+3*SIZEOF_XMMWORD], xmm3
692    movdqa      XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
693    movdqa      XMMWORD [edi+3*SIZEOF_XMMWORD], xmm3
694
695    sub         eax, byte 2*SIZEOF_XMMWORD
696    jz          short .nextrow
697
698    add         esi, byte 2*SIZEOF_XMMWORD  ; inptr
699    add         ebx, byte 4*SIZEOF_XMMWORD  ; outptr0
700    add         edi, byte 4*SIZEOF_XMMWORD  ; outptr1
701    jmp         short .columnloop
702    alignx      16, 7
703
704.nextrow:
705    pop         esi
706    pop         edi
707
708    add         esi, byte 1*SIZEOF_JSAMPROW  ; input_data
709    add         edi, byte 2*SIZEOF_JSAMPROW  ; output_data
710    sub         ecx, byte 2                  ; rowctr
711    jg          short .rowloop
712
713.return:
714    pop         edi
715    pop         esi
716;   pop         edx                     ; need not be preserved
717;   pop         ecx                     ; need not be preserved
718    pop         ebx
719    pop         ebp
720    ret
721
722; For some reason, the OS X linker does not honor the request to align the
723; segment unless we do this.
724    align       32
725