• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jdsample.asm - upsampling (64-bit AVX2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2009, 2016, D. R. Commander.
6; Copyright (C) 2015, Intel Corporation.
7;
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
11;
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
17
18%include "jsimdext.inc"
19
20; --------------------------------------------------------------------------
21    SECTION     SEG_CONST
22
23    alignz      32
24    GLOBAL_DATA(jconst_fancy_upsample_avx2)
25
26EXTN(jconst_fancy_upsample_avx2):
27
28PW_ONE   times 16 dw 1
29PW_TWO   times 16 dw 2
30PW_THREE times 16 dw 3
31PW_SEVEN times 16 dw 7
32PW_EIGHT times 16 dw 8
33
34    alignz      32
35
36; --------------------------------------------------------------------------
37    SECTION     SEG_TEXT
38    BITS        64
39;
40; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
41;
42; The upsampling algorithm is linear interpolation between pixel centers,
43; also known as a "triangle filter".  This is a good compromise between
44; speed and visual quality.  The centers of the output pixels are 1/4 and 3/4
45; of the way between input pixel centers.
46;
47; GLOBAL(void)
48; jsimd_h2v1_fancy_upsample_avx2(int max_v_samp_factor,
49;                                JDIMENSION downsampled_width,
50;                                JSAMPARRAY input_data,
51;                                JSAMPARRAY *output_data_ptr);
52;
53
54; r10 = int max_v_samp_factor
55; r11d = JDIMENSION downsampled_width
56; r12 = JSAMPARRAY input_data
57; r13 = JSAMPARRAY *output_data_ptr
58
59    align       32
60    GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_avx2)
61
62EXTN(jsimd_h2v1_fancy_upsample_avx2):
63    push        rbp
64    mov         rax, rsp
65    mov         rbp, rsp
66    push_xmm    3
67    collect_args 4
68
69    mov         eax, r11d               ; colctr
70    test        rax, rax
71    jz          near .return
72
73    mov         rcx, r10                ; rowctr
74    test        rcx, rcx
75    jz          near .return
76
77    mov         rsi, r12                ; input_data
78    mov         rdi, r13
79    mov         rdi, JSAMPARRAY [rdi]   ; output_data
80
81    vpxor       ymm0, ymm0, ymm0                 ; ymm0=(all 0's)
82    vpcmpeqb    xmm9, xmm9, xmm9
83    vpsrldq     xmm10, xmm9, (SIZEOF_XMMWORD-1)  ; (ff -- -- -- ... -- --) LSB is ff
84
85    vpslldq     xmm9, xmm9, (SIZEOF_XMMWORD-1)
86    vperm2i128  ymm9, ymm9, ymm9, 1              ; (---- ---- ... ---- ---- ff) MSB is ff
87
88.rowloop:
89    push        rax                     ; colctr
90    push        rdi
91    push        rsi
92
93    mov         rsi, JSAMPROW [rsi]     ; inptr
94    mov         rdi, JSAMPROW [rdi]     ; outptr
95
96    test        rax, SIZEOF_YMMWORD-1
97    jz          short .skip
98    mov         dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
99    mov         JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
100.skip:
101    vpand       ymm7, ymm10, YMMWORD [rsi+0*SIZEOF_YMMWORD]
102
103    add         rax, byte SIZEOF_YMMWORD-1
104    and         rax, byte -SIZEOF_YMMWORD
105    cmp         rax, byte SIZEOF_YMMWORD
106    ja          short .columnloop
107
108.columnloop_last:
109    vpand       ymm6, ymm9, YMMWORD [rsi+0*SIZEOF_YMMWORD]
110    jmp         short .upsample
111
112.columnloop:
113    vmovdqu     ymm6, YMMWORD [rsi+1*SIZEOF_YMMWORD]
114    vperm2i128  ymm6, ymm0, ymm6, 0x20
115    vpslldq     ymm6, ymm6, 15
116
117.upsample:
118    vmovdqu     ymm1, YMMWORD [rsi+0*SIZEOF_YMMWORD]  ; ymm1=( 0  1  2 ... 29 30 31)
119
120    vperm2i128  ymm2, ymm0, ymm1, 0x20
121    vpalignr    ymm2, ymm1, ymm2, 15            ; ymm2=(--  0  1 ... 28 29 30)
122    vperm2i128  ymm4, ymm0, ymm1, 0x03
123    vpalignr    ymm3, ymm4, ymm1, 1             ; ymm3=( 1  2  3 ... 30 31 --)
124
125    vpor        ymm2, ymm2, ymm7                ; ymm2=(-1  0  1 ... 28 29 30)
126    vpor        ymm3, ymm3, ymm6                ; ymm3=( 1  2  3 ... 30 31 32)
127
128    vpsrldq     ymm7, ymm4, (SIZEOF_XMMWORD-1)  ; ymm7=(31 -- -- ... -- -- --)
129
130    vpunpckhbw  ymm4, ymm1, ymm0                ; ymm4=( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
131    vpunpcklbw  ymm5, ymm1, ymm0                ; ymm5=( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
132    vperm2i128  ymm1, ymm5, ymm4, 0x20          ; ymm1=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
133    vperm2i128  ymm4, ymm5, ymm4, 0x31          ; ymm4=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
134
135    vpunpckhbw  ymm5, ymm2, ymm0                ; ymm5=( 7  8  9 10 11 12 13 14 23 24 25 26 27 28 29 30)
136    vpunpcklbw  ymm6, ymm2, ymm0                ; ymm6=(-1  0  1  2  3  4  5  6 15 16 17 18 19 20 21 22)
137    vperm2i128  ymm2, ymm6, ymm5, 0x20          ; ymm2=(-1  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
138    vperm2i128  ymm5, ymm6, ymm5, 0x31          ; ymm5=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
139
140    vpunpckhbw  ymm6, ymm3, ymm0                ; ymm6=( 1  2  3  4  5  6  7  8 17 18 19 20 21 22 23 24)
141    vpunpcklbw  ymm8, ymm3, ymm0                ; ymm8=( 9 10 11 12 13 14 15 16 25 26 27 28 29 30 31 32)
142    vperm2i128  ymm3, ymm8, ymm6, 0x20          ; ymm3=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16)
143    vperm2i128  ymm6, ymm8, ymm6, 0x31          ; ymm6=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32)
144
145    vpmullw     ymm1, ymm1, [rel PW_THREE]
146    vpmullw     ymm4, ymm4, [rel PW_THREE]
147    vpaddw      ymm2, ymm2, [rel PW_ONE]
148    vpaddw      ymm5, ymm5, [rel PW_ONE]
149    vpaddw      ymm3, ymm3, [rel PW_TWO]
150    vpaddw      ymm6, ymm6, [rel PW_TWO]
151
152    vpaddw      ymm2, ymm2, ymm1
153    vpaddw      ymm5, ymm5, ymm4
154    vpsrlw      ymm2, ymm2, 2                   ; ymm2=OutLE=( 0  2  4  6  8 10 12 14 16 18 20 22 24 26 28 30)
155    vpsrlw      ymm5, ymm5, 2                   ; ymm5=OutHE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62)
156    vpaddw      ymm3, ymm3, ymm1
157    vpaddw      ymm6, ymm6, ymm4
158    vpsrlw      ymm3, ymm3, 2                   ; ymm3=OutLO=( 1  3  5  7  9 11 13 15 17 19 21 23 25 27 29 31)
159    vpsrlw      ymm6, ymm6, 2                   ; ymm6=OutHO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63)
160
161    vpsllw      ymm3, ymm3, BYTE_BIT
162    vpsllw      ymm6, ymm6, BYTE_BIT
163    vpor        ymm2, ymm2, ymm3                ; ymm2=OutL=( 0  1  2 ... 29 30 31)
164    vpor        ymm5, ymm5, ymm6                ; ymm5=OutH=(32 33 34 ... 61 62 63)
165
166    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm2
167    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm5
168
169    sub         rax, byte SIZEOF_YMMWORD
170    add         rsi, byte 1*SIZEOF_YMMWORD  ; inptr
171    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
172    cmp         rax, byte SIZEOF_YMMWORD
173    ja          near .columnloop
174    test        eax, eax
175    jnz         near .columnloop_last
176
177    pop         rsi
178    pop         rdi
179    pop         rax
180
181    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
182    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
183    dec         rcx                        ; rowctr
184    jg          near .rowloop
185
186.return:
187    vzeroupper
188    uncollect_args 4
189    pop_xmm     3
190    pop         rbp
191    ret
192
193; --------------------------------------------------------------------------
194;
195; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
196; Again a triangle filter; see comments for h2v1 case, above.
197;
198; GLOBAL(void)
199; jsimd_h2v2_fancy_upsample_avx2(int max_v_samp_factor,
200;                                JDIMENSION downsampled_width,
201;                                JSAMPARRAY input_data,
202;                                JSAMPARRAY *output_data_ptr);
203;
204
205; r10 = int max_v_samp_factor
206; r11d = JDIMENSION downsampled_width
207; r12 = JSAMPARRAY input_data
208; r13 = JSAMPARRAY *output_data_ptr
209
210%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD  ; ymmword wk[WK_NUM]
211%define WK_NUM  4
212
213    align       32
214    GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_avx2)
215
216EXTN(jsimd_h2v2_fancy_upsample_avx2):
217    push        rbp
218    mov         rax, rsp                     ; rax = original rbp
219    sub         rsp, byte 4
220    and         rsp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
221    mov         [rsp], rax
222    mov         rbp, rsp                     ; rbp = aligned rbp
223    lea         rsp, [wk(0)]
224    push_xmm    3
225    collect_args 4
226    push        rbx
227
228    mov         eax, r11d               ; colctr
229    test        rax, rax
230    jz          near .return
231
232    mov         rcx, r10                ; rowctr
233    test        rcx, rcx
234    jz          near .return
235
236    mov         rsi, r12                ; input_data
237    mov         rdi, r13
238    mov         rdi, JSAMPARRAY [rdi]   ; output_data
239.rowloop:
240    push        rax                     ; colctr
241    push        rcx
242    push        rdi
243    push        rsi
244
245    mov         rcx, JSAMPROW [rsi-1*SIZEOF_JSAMPROW]  ; inptr1(above)
246    mov         rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW]  ; inptr0
247    mov         rsi, JSAMPROW [rsi+1*SIZEOF_JSAMPROW]  ; inptr1(below)
248    mov         rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]  ; outptr0
249    mov         rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]  ; outptr1
250
251    vpxor       ymm8, ymm8, ymm8                 ; ymm8=(all 0's)
252    vpcmpeqb    xmm9, xmm9, xmm9
253    vpsrldq     xmm10, xmm9, (SIZEOF_XMMWORD-2)  ; (ffff ---- ---- ... ---- ----) LSB is ffff
254    vpslldq     xmm9, xmm9, (SIZEOF_XMMWORD-2)
255    vperm2i128  ymm9, ymm9, ymm9, 1              ; (---- ---- ... ---- ---- ffff) MSB is ffff
256
257    test        rax, SIZEOF_YMMWORD-1
258    jz          short .skip
259    push        rdx
260    mov         dl, JSAMPLE [rcx+(rax-1)*SIZEOF_JSAMPLE]
261    mov         JSAMPLE [rcx+rax*SIZEOF_JSAMPLE], dl
262    mov         dl, JSAMPLE [rbx+(rax-1)*SIZEOF_JSAMPLE]
263    mov         JSAMPLE [rbx+rax*SIZEOF_JSAMPLE], dl
264    mov         dl, JSAMPLE [rsi+(rax-1)*SIZEOF_JSAMPLE]
265    mov         JSAMPLE [rsi+rax*SIZEOF_JSAMPLE], dl    ; insert a dummy sample
266    pop         rdx
267.skip:
268    ; -- process the first column block
269
270    vmovdqu     ymm0, YMMWORD [rbx+0*SIZEOF_YMMWORD]  ; ymm0=row[ 0][0]
271    vmovdqu     ymm1, YMMWORD [rcx+0*SIZEOF_YMMWORD]  ; ymm1=row[-1][0]
272    vmovdqu     ymm2, YMMWORD [rsi+0*SIZEOF_YMMWORD]  ; ymm2=row[+1][0]
273
274    vpunpckhbw  ymm4, ymm0, ymm8        ; ymm4=row[ 0]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
275    vpunpcklbw  ymm5, ymm0, ymm8        ; ymm5=row[ 0]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
276    vperm2i128  ymm0, ymm5, ymm4, 0x20  ; ymm0=row[ 0]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
277    vperm2i128  ymm4, ymm5, ymm4, 0x31  ; ymm4=row[ 0](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
278
279    vpunpckhbw  ymm5, ymm1, ymm8        ; ymm5=row[-1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
280    vpunpcklbw  ymm6, ymm1, ymm8        ; ymm6=row[-1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
281    vperm2i128  ymm1, ymm6, ymm5, 0x20  ; ymm1=row[-1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
282    vperm2i128  ymm5, ymm6, ymm5, 0x31  ; ymm5=row[-1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
283
284    vpunpckhbw  ymm6, ymm2, ymm8        ; ymm6=row[+1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
285    vpunpcklbw  ymm3, ymm2, ymm8        ; ymm3=row[+1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
286    vperm2i128  ymm2, ymm3, ymm6, 0x20  ; ymm2=row[+1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
287    vperm2i128  ymm6, ymm3, ymm6, 0x31  ; ymm6=row[+1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
288
289    vpmullw     ymm0, ymm0, [rel PW_THREE]
290    vpmullw     ymm4, ymm4, [rel PW_THREE]
291
292    vpaddw      ymm1, ymm1, ymm0        ; ymm1=Int0L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
293    vpaddw      ymm5, ymm5, ymm4        ; ymm5=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
294    vpaddw      ymm2, ymm2, ymm0        ; ymm2=Int1L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
295    vpaddw      ymm6, ymm6, ymm4        ; ymm6=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
296
297    vmovdqu     YMMWORD [rdx+0*SIZEOF_YMMWORD], ymm1  ; temporarily save
298    vmovdqu     YMMWORD [rdx+1*SIZEOF_YMMWORD], ymm5  ; the intermediate data
299    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm2
300    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm6
301
302    vpand       ymm1, ymm1, ymm10       ; ymm1=( 0 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
303    vpand       ymm2, ymm2, ymm10       ; ymm2=( 0 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
304
305    vmovdqa     YMMWORD [wk(0)], ymm1
306    vmovdqa     YMMWORD [wk(1)], ymm2
307
308    add         rax, byte SIZEOF_YMMWORD-1
309    and         rax, byte -SIZEOF_YMMWORD
310    cmp         rax, byte SIZEOF_YMMWORD
311    ja          short .columnloop
312
313.columnloop_last:
314    ; -- process the last column block
315
316    vpand       ymm1, ymm9, YMMWORD [rdx+1*SIZEOF_YMMWORD]
317    vpand       ymm2, ymm9, YMMWORD [rdi+1*SIZEOF_YMMWORD]
318
319    vmovdqa     YMMWORD [wk(2)], ymm1   ; ymm1=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 31)
320    vmovdqa     YMMWORD [wk(3)], ymm2   ; ymm2=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 31)
321
322    jmp         near .upsample
323
324.columnloop:
325    ; -- process the next column block
326
327    vmovdqu     ymm0, YMMWORD [rbx+1*SIZEOF_YMMWORD]  ; ymm0=row[ 0][1]
328    vmovdqu     ymm1, YMMWORD [rcx+1*SIZEOF_YMMWORD]  ; ymm1=row[-1][1]
329    vmovdqu     ymm2, YMMWORD [rsi+1*SIZEOF_YMMWORD]  ; ymm2=row[+1][1]
330
331    vpunpckhbw  ymm4, ymm0, ymm8        ; ymm4=row[ 0]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
332    vpunpcklbw  ymm5, ymm0, ymm8        ; ymm5=row[ 0]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
333    vperm2i128  ymm0, ymm5, ymm4, 0x20  ; ymm0=row[ 0]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
334    vperm2i128  ymm4, ymm5, ymm4, 0x31  ; ymm4=row[ 0](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
335
336    vpunpckhbw  ymm5, ymm1, ymm8        ; ymm5=row[-1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
337    vpunpcklbw  ymm6, ymm1, ymm8        ; ymm6=row[-1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
338    vperm2i128  ymm1, ymm6, ymm5, 0x20  ; ymm1=row[-1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
339    vperm2i128  ymm5, ymm6, ymm5, 0x31  ; ymm5=row[-1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
340
341    vpunpckhbw  ymm6, ymm2, ymm8        ; ymm6=row[+1]( 8  9 10 11 12 13 14 15 24 25 26 27 28 29 30 31)
342    vpunpcklbw  ymm7, ymm2, ymm8        ; ymm7=row[+1]( 0  1  2  3  4  5  6  7 16 17 18 19 20 21 22 23)
343    vperm2i128  ymm2, ymm7, ymm6, 0x20  ; ymm2=row[+1]( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
344    vperm2i128  ymm6, ymm7, ymm6, 0x31  ; ymm6=row[+1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
345
346    vpmullw     ymm0, ymm0, [rel PW_THREE]
347    vpmullw     ymm4, ymm4, [rel PW_THREE]
348
349    vpaddw      ymm1, ymm1, ymm0        ; ymm1=Int0L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
350    vpaddw      ymm5, ymm5, ymm4        ; ymm5=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
351    vpaddw      ymm2, ymm2, ymm0        ; ymm2=Int1L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
352    vpaddw      ymm6, ymm6, ymm4        ; ymm6=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
353
354    vmovdqu     YMMWORD [rdx+2*SIZEOF_YMMWORD], ymm1  ; temporarily save
355    vmovdqu     YMMWORD [rdx+3*SIZEOF_YMMWORD], ymm5  ; the intermediate data
356    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymm2
357    vmovdqu     YMMWORD [rdi+3*SIZEOF_YMMWORD], ymm6
358
359    vperm2i128  ymm1, ymm8, ymm1, 0x20
360    vpslldq     ymm1, ymm1, 14          ; ymm1=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- --  0)
361    vperm2i128  ymm2, ymm8, ymm2, 0x20
362    vpslldq     ymm2, ymm2, 14          ; ymm2=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- --  0)
363
364    vmovdqa     YMMWORD [wk(2)], ymm1
365    vmovdqa     YMMWORD [wk(3)], ymm2
366
367.upsample:
368    ; -- process the upper row
369
370    vmovdqu     ymm7, YMMWORD [rdx+0*SIZEOF_YMMWORD]  ; ymm7=Int0L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
371    vmovdqu     ymm3, YMMWORD [rdx+1*SIZEOF_YMMWORD]  ; ymm3=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
372
373    vperm2i128  ymm0, ymm8, ymm7, 0x03
374    vpalignr    ymm0, ymm0, ymm7, 2     ; ymm0=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 --)
375    vperm2i128  ymm4, ymm8, ymm3, 0x20
376    vpslldq     ymm4, ymm4, 14          ; ymm4=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 16)
377
378    vperm2i128  ymm5, ymm8, ymm7, 0x03
379    vpsrldq     ymm5, ymm5, 14          ; ymm5=(15 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
380    vperm2i128  ymm6, ymm8, ymm3, 0x20
381    vpalignr    ymm6, ymm3, ymm6, 14    ; ymm6=(-- 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
382
383    vpor        ymm0, ymm0, ymm4        ; ymm0=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16)
384    vpor        ymm5, ymm5, ymm6        ; ymm5=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
385
386    vperm2i128  ymm2, ymm8, ymm3, 0x03
387    vpalignr    ymm2, ymm2, ymm3, 2     ; ymm2=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 --)
388    vperm2i128  ymm4, ymm8, ymm3, 0x03
389    vpsrldq     ymm4, ymm4, 14          ; ymm4=(31 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
390    vperm2i128  ymm1, ymm8, ymm7, 0x20
391    vpalignr    ymm1, ymm7, ymm1, 14    ; ymm1=(--  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
392
393    vpor        ymm1, ymm1, YMMWORD [wk(0)]  ; ymm1=(-1  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
394    vpor        ymm2, ymm2, YMMWORD [wk(2)]  ; ymm2=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32)
395
396    vmovdqa     YMMWORD [wk(0)], ymm4
397
398    vpmullw     ymm7, ymm7, [rel PW_THREE]
399    vpmullw     ymm3, ymm3, [rel PW_THREE]
400    vpaddw      ymm1, ymm1, [rel PW_EIGHT]
401    vpaddw      ymm5, ymm5, [rel PW_EIGHT]
402    vpaddw      ymm0, ymm0, [rel PW_SEVEN]
403    vpaddw      ymm2, [rel PW_SEVEN]
404
405    vpaddw      ymm1, ymm1, ymm7
406    vpaddw      ymm5, ymm5, ymm3
407    vpsrlw      ymm1, ymm1, 4           ; ymm1=Out0LE=( 0  2  4  6  8 10 12 14 16 18 20 22 24 26 28 30)
408    vpsrlw      ymm5, ymm5, 4           ; ymm5=Out0HE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62)
409    vpaddw      ymm0, ymm0, ymm7
410    vpaddw      ymm2, ymm2, ymm3
411    vpsrlw      ymm0, ymm0, 4           ; ymm0=Out0LO=( 1  3  5  7  9 11 13 15 17 19 21 23 25 27 29 31)
412    vpsrlw      ymm2, ymm2, 4           ; ymm2=Out0HO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63)
413
414    vpsllw      ymm0, ymm0, BYTE_BIT
415    vpsllw      ymm2, ymm2, BYTE_BIT
416    vpor        ymm1, ymm1, ymm0        ; ymm1=Out0L=( 0  1  2 ... 29 30 31)
417    vpor        ymm5, ymm5, ymm2        ; ymm5=Out0H=(32 33 34 ... 61 62 63)
418
419    vmovdqu     YMMWORD [rdx+0*SIZEOF_YMMWORD], ymm1
420    vmovdqu     YMMWORD [rdx+1*SIZEOF_YMMWORD], ymm5
421
422    ; -- process the lower row
423
424    vmovdqu     ymm6, YMMWORD [rdi+0*SIZEOF_YMMWORD]  ; ymm6=Int1L=( 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15)
425    vmovdqu     ymm4, YMMWORD [rdi+1*SIZEOF_YMMWORD]  ; ymm4=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
426
427    vperm2i128  ymm7, ymm8, ymm6, 0x03
428    vpalignr    ymm7, ymm7, ymm6, 2     ; ymm7=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 --)
429    vperm2i128  ymm3, ymm8, ymm4, 0x20
430    vpslldq     ymm3, ymm3, 14          ; ymm3=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 16)
431
432    vperm2i128  ymm0, ymm8, ymm6, 0x03
433    vpsrldq     ymm0, ymm0, 14          ; ymm0=(15 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
434    vperm2i128  ymm2, ymm8, ymm4, 0x20
435    vpalignr    ymm2, ymm4, ymm2, 14    ; ymm2=(-- 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
436
437    vpor        ymm7, ymm7, ymm3        ; ymm7=( 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16)
438    vpor        ymm0, ymm0, ymm2        ; ymm0=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30)
439
440    vperm2i128  ymm5, ymm8, ymm4, 0x03
441    vpalignr    ymm5, ymm5, ymm4, 2     ; ymm5=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 --)
442    vperm2i128  ymm3, ymm8, ymm4, 0x03
443    vpsrldq     ymm3, ymm3, 14          ; ymm3=(31 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --)
444    vperm2i128  ymm1, ymm8, ymm6, 0x20
445    vpalignr    ymm1, ymm6, ymm1, 14    ; ymm1=(--  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
446
447    vpor        ymm1, ymm1, YMMWORD [wk(1)]  ; ymm1=(-1  0  1  2  3  4  5  6  7  8  9 10 11 12 13 14)
448    vpor        ymm5, ymm5, YMMWORD [wk(3)]  ; ymm5=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32)
449
450    vmovdqa     YMMWORD [wk(1)], ymm3
451
452    vpmullw     ymm6, ymm6, [rel PW_THREE]
453    vpmullw     ymm4, ymm4, [rel PW_THREE]
454    vpaddw      ymm1, ymm1, [rel PW_EIGHT]
455    vpaddw      ymm0, ymm0, [rel PW_EIGHT]
456    vpaddw      ymm7, ymm7, [rel PW_SEVEN]
457    vpaddw      ymm5, ymm5, [rel PW_SEVEN]
458
459    vpaddw      ymm1, ymm1, ymm6
460    vpaddw      ymm0, ymm0, ymm4
461    vpsrlw      ymm1, ymm1, 4           ; ymm1=Out1LE=( 0  2  4  6  8 10 12 14 16 18 20 22 24 26 28 30)
462    vpsrlw      ymm0, ymm0, 4           ; ymm0=Out1HE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62)
463    vpaddw      ymm7, ymm7, ymm6
464    vpaddw      ymm5, ymm5, ymm4
465    vpsrlw      ymm7, ymm7, 4           ; ymm7=Out1LO=( 1  3  5  7  9 11 13 15 17 19 21 23 25 27 29 31)
466    vpsrlw      ymm5, ymm5, 4           ; ymm5=Out1HO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63)
467
468    vpsllw      ymm7, ymm7, BYTE_BIT
469    vpsllw      ymm5, ymm5, BYTE_BIT
470    vpor        ymm1, ymm1, ymm7        ; ymm1=Out1L=( 0  1  2 ... 29 30 31)
471    vpor        ymm0, ymm0, ymm5        ; ymm0=Out1H=(32 33 34 ... 61 62 63)
472
473    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm1
474    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm0
475
476    sub         rax, byte SIZEOF_YMMWORD
477    add         rcx, byte 1*SIZEOF_YMMWORD  ; inptr1(above)
478    add         rbx, byte 1*SIZEOF_YMMWORD  ; inptr0
479    add         rsi, byte 1*SIZEOF_YMMWORD  ; inptr1(below)
480    add         rdx, byte 2*SIZEOF_YMMWORD  ; outptr0
481    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr1
482    cmp         rax, byte SIZEOF_YMMWORD
483    ja          near .columnloop
484    test        rax, rax
485    jnz         near .columnloop_last
486
487    pop         rsi
488    pop         rdi
489    pop         rcx
490    pop         rax
491
492    add         rsi, byte 1*SIZEOF_JSAMPROW  ; input_data
493    add         rdi, byte 2*SIZEOF_JSAMPROW  ; output_data
494    sub         rcx, byte 2                  ; rowctr
495    jg          near .rowloop
496
497.return:
498    pop         rbx
499    vzeroupper
500    uncollect_args 4
501    pop_xmm     3
502    mov         rsp, rbp                ; rsp <- aligned rbp
503    pop         rsp                     ; rsp <- original rbp
504    pop         rbp
505    ret
506
507; --------------------------------------------------------------------------
508;
509; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
510; It's still a box filter.
511;
512; GLOBAL(void)
513; jsimd_h2v1_upsample_avx2(int max_v_samp_factor, JDIMENSION output_width,
514;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
515;
516
517; r10 = int max_v_samp_factor
518; r11d = JDIMENSION output_width
519; r12 = JSAMPARRAY input_data
520; r13 = JSAMPARRAY *output_data_ptr
521
522    align       32
523    GLOBAL_FUNCTION(jsimd_h2v1_upsample_avx2)
524
525EXTN(jsimd_h2v1_upsample_avx2):
526    push        rbp
527    mov         rax, rsp
528    mov         rbp, rsp
529    collect_args 4
530
531    mov         edx, r11d
532    add         rdx, byte (SIZEOF_YMMWORD-1)
533    and         rdx, -SIZEOF_YMMWORD
534    jz          near .return
535
536    mov         rcx, r10                ; rowctr
537    test        rcx, rcx
538    jz          short .return
539
540    mov         rsi, r12                ; input_data
541    mov         rdi, r13
542    mov         rdi, JSAMPARRAY [rdi]   ; output_data
543.rowloop:
544    push        rdi
545    push        rsi
546
547    mov         rsi, JSAMPROW [rsi]     ; inptr
548    mov         rdi, JSAMPROW [rdi]     ; outptr
549    mov         rax, rdx                ; colctr
550.columnloop:
551
552    cmp         rax, byte SIZEOF_YMMWORD
553    ja          near .above_16
554
555    vmovdqu     xmm0, XMMWORD [rsi+0*SIZEOF_YMMWORD]
556    vpunpckhbw  xmm1, xmm0, xmm0
557    vpunpcklbw  xmm0, xmm0, xmm0
558
559    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
560    vmovdqu     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
561
562    jmp         short .nextrow
563
564.above_16:
565    vmovdqu     ymm0, YMMWORD [rsi+0*SIZEOF_YMMWORD]
566
567    vpermq      ymm0, ymm0, 0xd8
568    vpunpckhbw  ymm1, ymm0, ymm0
569    vpunpcklbw  ymm0, ymm0, ymm0
570
571    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm0
572    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm1
573
574    sub         rax, byte 2*SIZEOF_YMMWORD
575    jz          short .nextrow
576
577    add         rsi, byte SIZEOF_YMMWORD    ; inptr
578    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
579    jmp         short .columnloop
580
581.nextrow:
582    pop         rsi
583    pop         rdi
584
585    add         rsi, byte SIZEOF_JSAMPROW  ; input_data
586    add         rdi, byte SIZEOF_JSAMPROW  ; output_data
587    dec         rcx                        ; rowctr
588    jg          short .rowloop
589
590.return:
591    vzeroupper
592    uncollect_args 4
593    pop         rbp
594    ret
595
596; --------------------------------------------------------------------------
597;
598; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
599; It's still a box filter.
600;
601; GLOBAL(void)
602; jsimd_h2v2_upsample_avx2(int max_v_samp_factor, JDIMENSION output_width,
603;                          JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr);
604;
605
606; r10 = int max_v_samp_factor
607; r11d = JDIMENSION output_width
608; r12 = JSAMPARRAY input_data
609; r13 = JSAMPARRAY *output_data_ptr
610
611    align       32
612    GLOBAL_FUNCTION(jsimd_h2v2_upsample_avx2)
613
614EXTN(jsimd_h2v2_upsample_avx2):
615    push        rbp
616    mov         rax, rsp
617    mov         rbp, rsp
618    collect_args 4
619    push        rbx
620
621    mov         edx, r11d
622    add         rdx, byte (SIZEOF_YMMWORD-1)
623    and         rdx, -SIZEOF_YMMWORD
624    jz          near .return
625
626    mov         rcx, r10                ; rowctr
627    test        rcx, rcx
628    jz          near .return
629
630    mov         rsi, r12                ; input_data
631    mov         rdi, r13
632    mov         rdi, JSAMPARRAY [rdi]   ; output_data
633.rowloop:
634    push        rdi
635    push        rsi
636
637    mov         rsi, JSAMPROW [rsi]                    ; inptr
638    mov         rbx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]  ; outptr0
639    mov         rdi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]  ; outptr1
640    mov         rax, rdx                               ; colctr
641.columnloop:
642
643    cmp         rax, byte SIZEOF_YMMWORD
644    ja          short .above_16
645
646    vmovdqu     xmm0, XMMWORD [rsi+0*SIZEOF_XMMWORD]
647    vpunpckhbw  xmm1, xmm0, xmm0
648    vpunpcklbw  xmm0, xmm0, xmm0
649
650    vmovdqu     XMMWORD [rbx+0*SIZEOF_XMMWORD], xmm0
651    vmovdqu     XMMWORD [rbx+1*SIZEOF_XMMWORD], xmm1
652    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmm0
653    vmovdqu     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmm1
654
655    jmp         near .nextrow
656
657.above_16:
658    vmovdqu     ymm0, YMMWORD [rsi+0*SIZEOF_YMMWORD]
659
660    vpermq      ymm0, ymm0, 0xd8
661    vpunpckhbw  ymm1, ymm0, ymm0
662    vpunpcklbw  ymm0, ymm0, ymm0
663
664    vmovdqu     YMMWORD [rbx+0*SIZEOF_YMMWORD], ymm0
665    vmovdqu     YMMWORD [rbx+1*SIZEOF_YMMWORD], ymm1
666    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymm0
667    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymm1
668
669    sub         rax, byte 2*SIZEOF_YMMWORD
670    jz          short .nextrow
671
672    add         rsi, byte SIZEOF_YMMWORD  ; inptr
673    add         rbx, 2*SIZEOF_YMMWORD     ; outptr0
674    add         rdi, 2*SIZEOF_YMMWORD     ; outptr1
675    jmp         short .columnloop
676
677.nextrow:
678    pop         rsi
679    pop         rdi
680
681    add         rsi, byte 1*SIZEOF_JSAMPROW  ; input_data
682    add         rdi, byte 2*SIZEOF_JSAMPROW  ; output_data
683    sub         rcx, byte 2                  ; rowctr
684    jg          near .rowloop
685
686.return:
687    pop         rbx
688    vzeroupper
689    uncollect_args 4
690    pop         rbp
691    ret
692
693; For some reason, the OS X linker does not honor the request to align the
694; segment unless we do this.
695    align       32
696