123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593 |
- ;
- ; jdmrgext.asm - merged upsampling/color conversion (64-bit AVX2)
- ;
- ; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
- ; Copyright (C) 2009, 2012, 2016, D. R. Commander.
- ; Copyright (C) 2015, Intel Corporation.
- ;
- ; Based on the x86 SIMD extension for IJG JPEG library
- ; Copyright (C) 1999-2006, MIYASAKA Masaru.
- ; For conditions of distribution and use, see copyright notice in jsimdext.inc
- ;
- ; This file should be assembled with NASM (Netwide Assembler),
- ; can *not* be assembled with Microsoft's MASM or any compatible
- ; assembler (including Borland's Turbo Assembler).
- ; NASM is available from http://nasm.sourceforge.net/ or
- ; http://sourceforge.net/project/showfiles.php?group_id=6208
- %include "jcolsamp.inc"
- ; --------------------------------------------------------------------------
- ;
- ; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
- ;
- ; GLOBAL(void)
- ; jsimd_h2v1_merged_upsample_avx2(JDIMENSION output_width,
- ; JSAMPIMAGE input_buf,
- ; JDIMENSION in_row_group_ctr,
- ; JSAMPARRAY output_buf);
- ;
- ; r10d = JDIMENSION output_width
- ; r11 = JSAMPIMAGE input_buf
- ; r12d = JDIMENSION in_row_group_ctr
- ; r13 = JSAMPARRAY output_buf
- %define wk(i) rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD ; ymmword wk[WK_NUM]
- %define WK_NUM 3
- align 32
- GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_avx2)
- EXTN(jsimd_h2v1_merged_upsample_avx2):
- push rbp
- mov rax, rsp ; rax = original rbp
- sub rsp, byte 4
- and rsp, byte (-SIZEOF_YMMWORD) ; align to 256 bits
- mov [rsp], rax
- mov rbp, rsp ; rbp = aligned rbp
- lea rsp, [wk(0)]
- collect_args 4
- push rbx
- mov ecx, r10d ; col
- test rcx, rcx
- jz near .return
- push rcx
- mov rdi, r11
- mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
- mov rdi, r13
- mov rsi, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW] ; inptr0
- mov rbx, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW] ; inptr1
- mov rdx, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW] ; inptr2
- mov rdi, JSAMPROW [rdi] ; outptr
- pop rcx ; col
- .columnloop:
- vmovdqu ymm6, YMMWORD [rbx] ; ymm6=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV)
- vmovdqu ymm7, YMMWORD [rdx] ; ymm7=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV)
- vpxor ymm1, ymm1, ymm1 ; ymm1=(all 0's)
- vpcmpeqw ymm3, ymm3, ymm3
- vpsllw ymm3, ymm3, 7 ; ymm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
- vpermq ymm6, ymm6, 0xd8 ; ymm6=Cb(01234567GHIJKLMN89ABCDEFOPQRSTUV)
- vpermq ymm7, ymm7, 0xd8 ; ymm7=Cr(01234567GHIJKLMN89ABCDEFOPQRSTUV)
- vpunpcklbw ymm4, ymm6, ymm1 ; ymm4=Cb(0123456789ABCDEF)=CbL
- vpunpckhbw ymm6, ymm6, ymm1 ; ymm6=Cb(GHIJKLMNOPQRSTUV)=CbH
- vpunpcklbw ymm0, ymm7, ymm1 ; ymm0=Cr(0123456789ABCDEF)=CrL
- vpunpckhbw ymm7, ymm7, ymm1 ; ymm7=Cr(GHIJKLMNOPQRSTUV)=CrH
- vpaddw ymm5, ymm6, ymm3
- vpaddw ymm2, ymm4, ymm3
- vpaddw ymm1, ymm7, ymm3
- vpaddw ymm3, ymm0, ymm3
- ; (Original)
- ; R = Y + 1.40200 * Cr
- ; G = Y - 0.34414 * Cb - 0.71414 * Cr
- ; B = Y + 1.77200 * Cb
- ;
- ; (This implementation)
- ; R = Y + 0.40200 * Cr + Cr
- ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
- ; B = Y - 0.22800 * Cb + Cb + Cb
- vpaddw ymm6, ymm5, ymm5 ; ymm6=2*CbH
- vpaddw ymm4, ymm2, ymm2 ; ymm4=2*CbL
- vpaddw ymm7, ymm1, ymm1 ; ymm7=2*CrH
- vpaddw ymm0, ymm3, ymm3 ; ymm0=2*CrL
- vpmulhw ymm6, ymm6, [rel PW_MF0228] ; ymm6=(2*CbH * -FIX(0.22800))
- vpmulhw ymm4, ymm4, [rel PW_MF0228] ; ymm4=(2*CbL * -FIX(0.22800))
- vpmulhw ymm7, ymm7, [rel PW_F0402] ; ymm7=(2*CrH * FIX(0.40200))
- vpmulhw ymm0, ymm0, [rel PW_F0402] ; ymm0=(2*CrL * FIX(0.40200))
- vpaddw ymm6, ymm6, [rel PW_ONE]
- vpaddw ymm4, ymm4, [rel PW_ONE]
- vpsraw ymm6, ymm6, 1 ; ymm6=(CbH * -FIX(0.22800))
- vpsraw ymm4, ymm4, 1 ; ymm4=(CbL * -FIX(0.22800))
- vpaddw ymm7, ymm7, [rel PW_ONE]
- vpaddw ymm0, ymm0, [rel PW_ONE]
- vpsraw ymm7, ymm7, 1 ; ymm7=(CrH * FIX(0.40200))
- vpsraw ymm0, ymm0, 1 ; ymm0=(CrL * FIX(0.40200))
- vpaddw ymm6, ymm6, ymm5
- vpaddw ymm4, ymm4, ymm2
- vpaddw ymm6, ymm6, ymm5 ; ymm6=(CbH * FIX(1.77200))=(B-Y)H
- vpaddw ymm4, ymm4, ymm2 ; ymm4=(CbL * FIX(1.77200))=(B-Y)L
- vpaddw ymm7, ymm7, ymm1 ; ymm7=(CrH * FIX(1.40200))=(R-Y)H
- vpaddw ymm0, ymm0, ymm3 ; ymm0=(CrL * FIX(1.40200))=(R-Y)L
- vmovdqa YMMWORD [wk(0)], ymm6 ; wk(0)=(B-Y)H
- vmovdqa YMMWORD [wk(1)], ymm7 ; wk(1)=(R-Y)H
- vpunpckhwd ymm6, ymm5, ymm1
- vpunpcklwd ymm5, ymm5, ymm1
- vpmaddwd ymm5, ymm5, [rel PW_MF0344_F0285]
- vpmaddwd ymm6, ymm6, [rel PW_MF0344_F0285]
- vpunpckhwd ymm7, ymm2, ymm3
- vpunpcklwd ymm2, ymm2, ymm3
- vpmaddwd ymm2, ymm2, [rel PW_MF0344_F0285]
- vpmaddwd ymm7, ymm7, [rel PW_MF0344_F0285]
- vpaddd ymm5, ymm5, [rel PD_ONEHALF]
- vpaddd ymm6, ymm6, [rel PD_ONEHALF]
- vpsrad ymm5, ymm5, SCALEBITS
- vpsrad ymm6, ymm6, SCALEBITS
- vpaddd ymm2, ymm2, [rel PD_ONEHALF]
- vpaddd ymm7, ymm7, [rel PD_ONEHALF]
- vpsrad ymm2, ymm2, SCALEBITS
- vpsrad ymm7, ymm7, SCALEBITS
- vpackssdw ymm5, ymm5, ymm6 ; ymm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
- vpackssdw ymm2, ymm2, ymm7 ; ymm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
- vpsubw ymm5, ymm5, ymm1 ; ymm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
- vpsubw ymm2, ymm2, ymm3 ; ymm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
- vmovdqa YMMWORD [wk(2)], ymm5 ; wk(2)=(G-Y)H
- mov al, 2 ; Yctr
- jmp short .Yloop_1st
- .Yloop_2nd:
- vmovdqa ymm0, YMMWORD [wk(1)] ; ymm0=(R-Y)H
- vmovdqa ymm2, YMMWORD [wk(2)] ; ymm2=(G-Y)H
- vmovdqa ymm4, YMMWORD [wk(0)] ; ymm4=(B-Y)H
- .Yloop_1st:
- vmovdqu ymm7, YMMWORD [rsi] ; ymm7=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV)
- vpcmpeqw ymm6, ymm6, ymm6
- vpsrlw ymm6, ymm6, BYTE_BIT ; ymm6={0xFF 0x00 0xFF 0x00 ..}
- vpand ymm6, ymm6, ymm7 ; ymm6=Y(02468ACEGIKMOQSU)=YE
- vpsrlw ymm7, ymm7, BYTE_BIT ; ymm7=Y(13579BDFHJLNPRTV)=YO
- vmovdqa ymm1, ymm0 ; ymm1=ymm0=(R-Y)(L/H)
- vmovdqa ymm3, ymm2 ; ymm3=ymm2=(G-Y)(L/H)
- vmovdqa ymm5, ymm4 ; ymm5=ymm4=(B-Y)(L/H)
- vpaddw ymm0, ymm0, ymm6 ; ymm0=((R-Y)+YE)=RE=R(02468ACEGIKMOQSU)
- vpaddw ymm1, ymm1, ymm7 ; ymm1=((R-Y)+YO)=RO=R(13579BDFHJLNPRTV)
- vpackuswb ymm0, ymm0, ymm0 ; ymm0=R(02468ACE********GIKMOQSU********)
- vpackuswb ymm1, ymm1, ymm1 ; ymm1=R(13579BDF********HJLNPRTV********)
- vpaddw ymm2, ymm2, ymm6 ; ymm2=((G-Y)+YE)=GE=G(02468ACEGIKMOQSU)
- vpaddw ymm3, ymm3, ymm7 ; ymm3=((G-Y)+YO)=GO=G(13579BDFHJLNPRTV)
- vpackuswb ymm2, ymm2, ymm2 ; ymm2=G(02468ACE********GIKMOQSU********)
- vpackuswb ymm3, ymm3, ymm3 ; ymm3=G(13579BDF********HJLNPRTV********)
- vpaddw ymm4, ymm4, ymm6 ; ymm4=((B-Y)+YE)=BE=B(02468ACEGIKMOQSU)
- vpaddw ymm5, ymm5, ymm7 ; ymm5=((B-Y)+YO)=BO=B(13579BDFHJLNPRTV)
- vpackuswb ymm4, ymm4, ymm4 ; ymm4=B(02468ACE********GIKMOQSU********)
- vpackuswb ymm5, ymm5, ymm5 ; ymm5=B(13579BDF********HJLNPRTV********)
- %if RGB_PIXELSIZE == 3 ; ---------------
- ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
- ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
- ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
- ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
- ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
- ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
- ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
- ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
- vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
- ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
- vpunpcklbw ymmE, ymmE, ymmB ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F
- ; 2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V)
- vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F
- ; 1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V)
- vpsrldq ymmH, ymmA, 2 ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G
- ; 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --)
- vpunpckhwd ymmG, ymmA, ymmE ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F
- ; 0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V)
- vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07
- ; 0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N)
- vpsrldq ymmE, ymmE, 2 ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H
- ; 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --)
- vpsrldq ymmB, ymmD, 2 ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H
- ; 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --)
- vpunpckhwd ymmC, ymmD, ymmH ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G
- ; 1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --)
- vpunpcklwd ymmD, ymmD, ymmH ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18
- ; 1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O)
- vpunpckhwd ymmF, ymmE, ymmB ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H
- ; 2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --)
- vpunpcklwd ymmE, ymmE, ymmB ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29
- ; 2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P)
- vpshufd ymmH, ymmA, 0x4E ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03
- ; 0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J)
- vpunpckldq ymmA, ymmA, ymmD ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14
- ; 0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K)
- vpunpckhdq ymmD, ymmD, ymmE ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29
- ; 1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P)
- vpunpckldq ymmE, ymmE, ymmH ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07
- ; 2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N)
- vpshufd ymmH, ymmG, 0x4E ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B
- ; 0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R)
- vpunpckldq ymmG, ymmG, ymmC ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C
- ; 0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S)
- vpunpckhdq ymmC, ymmC, ymmF ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H
- ; 1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --)
- vpunpckldq ymmF, ymmF, ymmH ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F
- ; 2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V)
- vpunpcklqdq ymmH, ymmA, ymmE ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
- ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
- vpunpcklqdq ymmG, ymmD, ymmG ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
- ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
- vpunpcklqdq ymmC, ymmF, ymmC ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
- ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
- vperm2i128 ymmA, ymmH, ymmG, 0x20 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
- ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
- vperm2i128 ymmD, ymmC, ymmH, 0x30 ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
- ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
- vperm2i128 ymmF, ymmG, ymmC, 0x31 ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
- ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
- cmp rcx, byte SIZEOF_YMMWORD
- jb short .column_st64
- test rdi, SIZEOF_YMMWORD-1
- jnz short .out1
- ; --(aligned)-------------------
- vmovntdq YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- vmovntdq YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
- vmovntdq YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
- jmp short .out0
- .out1: ; --(unaligned)-----------------
- vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
- vmovdqu YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
- .out0:
- add rdi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr
- sub rcx, byte SIZEOF_YMMWORD
- jz near .endcolumn
- add rsi, byte SIZEOF_YMMWORD ; inptr0
- dec al ; Yctr
- jnz near .Yloop_2nd
- add rbx, byte SIZEOF_YMMWORD ; inptr1
- add rdx, byte SIZEOF_YMMWORD ; inptr2
- jmp near .columnloop
- .column_st64:
- lea rcx, [rcx+rcx*2] ; imul ecx, RGB_PIXELSIZE
- cmp rcx, byte 2*SIZEOF_YMMWORD
- jb short .column_st32
- vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
- add rdi, byte 2*SIZEOF_YMMWORD ; outptr
- vmovdqa ymmA, ymmF
- sub rcx, byte 2*SIZEOF_YMMWORD
- jmp short .column_st31
- .column_st32:
- cmp rcx, byte SIZEOF_YMMWORD
- jb short .column_st31
- vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- add rdi, byte SIZEOF_YMMWORD ; outptr
- vmovdqa ymmA, ymmD
- sub rcx, byte SIZEOF_YMMWORD
- jmp short .column_st31
- .column_st31:
- cmp rcx, byte SIZEOF_XMMWORD
- jb short .column_st15
- vmovdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
- add rdi, byte SIZEOF_XMMWORD ; outptr
- vperm2i128 ymmA, ymmA, ymmA, 1
- sub rcx, byte SIZEOF_XMMWORD
- .column_st15:
- ; Store the lower 8 bytes of xmmA to the output when it has enough
- ; space.
- cmp rcx, byte SIZEOF_MMWORD
- jb short .column_st7
- vmovq XMM_MMWORD [rdi], xmmA
- add rdi, byte SIZEOF_MMWORD
- sub rcx, byte SIZEOF_MMWORD
- vpsrldq xmmA, xmmA, SIZEOF_MMWORD
- .column_st7:
- ; Store the lower 4 bytes of xmmA to the output when it has enough
- ; space.
- cmp rcx, byte SIZEOF_DWORD
- jb short .column_st3
- vmovd XMM_DWORD [rdi], xmmA
- add rdi, byte SIZEOF_DWORD
- sub rcx, byte SIZEOF_DWORD
- vpsrldq xmmA, xmmA, SIZEOF_DWORD
- .column_st3:
- ; Store the lower 2 bytes of rax to the output when it has enough
- ; space.
- vmovd eax, xmmA
- cmp rcx, byte SIZEOF_WORD
- jb short .column_st1
- mov word [rdi], ax
- add rdi, byte SIZEOF_WORD
- sub rcx, byte SIZEOF_WORD
- shr rax, 16
- .column_st1:
- ; Store the lower 1 byte of rax to the output when it has enough
- ; space.
- test rcx, rcx
- jz short .endcolumn
- mov byte [rdi], al
- %else ; RGB_PIXELSIZE == 4 ; -----------
- %ifdef RGBX_FILLER_0XFF
- vpcmpeqb ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********)
- vpcmpeqb ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********)
- %else
- vpxor ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********)
- vpxor ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********)
- %endif
- ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
- ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
- ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
- ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
- ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
- ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
- ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **)
- ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **)
- vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
- ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
- vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E
- ; 2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U)
- vpunpcklbw ymmB, ymmB, ymmD ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F
- ; 0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V)
- vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F
- ; 2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V)
- vpunpckhwd ymmC, ymmA, ymmE ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E
- ; 0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U)
- vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36
- ; 0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M)
- vpunpckhwd ymmG, ymmB, ymmF ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F
- ; 0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V)
- vpunpcklwd ymmB, ymmB, ymmF ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37
- ; 0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N)
- vpunpckhdq ymmE, ymmA, ymmB ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
- ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
- vpunpckldq ymmB, ymmA, ymmB ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
- ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
- vpunpckhdq ymmF, ymmC, ymmG ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
- ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
- vpunpckldq ymmG, ymmC, ymmG ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
- ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
- vperm2i128 ymmA, ymmB, ymmE, 0x20 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
- ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
- vperm2i128 ymmD, ymmG, ymmF, 0x20 ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
- ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
- vperm2i128 ymmC, ymmB, ymmE, 0x31 ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
- ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
- vperm2i128 ymmH, ymmG, ymmF, 0x31 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
- ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
- cmp rcx, byte SIZEOF_YMMWORD
- jb short .column_st64
- test rdi, SIZEOF_YMMWORD-1
- jnz short .out1
- ; --(aligned)-------------------
- vmovntdq YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- vmovntdq YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
- vmovntdq YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
- vmovntdq YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
- jmp short .out0
- .out1: ; --(unaligned)-----------------
- vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
- vmovdqu YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
- vmovdqu YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
- .out0:
- add rdi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr
- sub rcx, byte SIZEOF_YMMWORD
- jz near .endcolumn
- add rsi, byte SIZEOF_YMMWORD ; inptr0
- dec al
- jnz near .Yloop_2nd
- add rbx, byte SIZEOF_YMMWORD ; inptr1
- add rdx, byte SIZEOF_YMMWORD ; inptr2
- jmp near .columnloop
- .column_st64:
- cmp rcx, byte SIZEOF_YMMWORD/2
- jb short .column_st32
- vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
- add rdi, byte 2*SIZEOF_YMMWORD ; outptr
- vmovdqa ymmA, ymmC
- vmovdqa ymmD, ymmH
- sub rcx, byte SIZEOF_YMMWORD/2
- .column_st32:
- cmp rcx, byte SIZEOF_YMMWORD/4
- jb short .column_st16
- vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
- add rdi, byte SIZEOF_YMMWORD ; outptr
- vmovdqa ymmA, ymmD
- sub rcx, byte SIZEOF_YMMWORD/4
- .column_st16:
- cmp rcx, byte SIZEOF_YMMWORD/8
- jb short .column_st15
- vmovdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
- add rdi, byte SIZEOF_XMMWORD ; outptr
- vperm2i128 ymmA, ymmA, ymmA, 1
- sub rcx, byte SIZEOF_YMMWORD/8
- .column_st15:
- ; Store two pixels (8 bytes) of ymmA to the output when it has enough
- ; space.
- cmp rcx, byte SIZEOF_YMMWORD/16
- jb short .column_st7
- vmovq MMWORD [rdi], xmmA
- add rdi, byte SIZEOF_YMMWORD/16*4
- sub rcx, byte SIZEOF_YMMWORD/16
- vpsrldq xmmA, SIZEOF_YMMWORD/16*4
- .column_st7:
- ; Store one pixel (4 bytes) of ymmA to the output when it has enough
- ; space.
- test rcx, rcx
- jz short .endcolumn
- vmovd XMM_DWORD [rdi], xmmA
- %endif ; RGB_PIXELSIZE ; ---------------
- .endcolumn:
- sfence ; flush the write buffer
- .return:
- pop rbx
- vzeroupper
- uncollect_args 4
- mov rsp, rbp ; rsp <- aligned rbp
- pop rsp ; rsp <- original rbp
- pop rbp
- ret
- ; --------------------------------------------------------------------------
- ;
- ; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
- ;
- ; GLOBAL(void)
- ; jsimd_h2v2_merged_upsample_avx2(JDIMENSION output_width,
- ; JSAMPIMAGE input_buf,
- ; JDIMENSION in_row_group_ctr,
- ; JSAMPARRAY output_buf);
- ;
- ; r10d = JDIMENSION output_width
- ; r11 = JSAMPIMAGE input_buf
- ; r12d = JDIMENSION in_row_group_ctr
- ; r13 = JSAMPARRAY output_buf
- align 32
- GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_avx2)
- EXTN(jsimd_h2v2_merged_upsample_avx2):
- push rbp
- mov rax, rsp
- mov rbp, rsp
- collect_args 4
- push rbx
- mov eax, r10d
- mov rdi, r11
- mov ecx, r12d
- mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
- mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
- mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
- mov rdi, r13
- lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
- push rdx ; inptr2
- push rbx ; inptr1
- push rsi ; inptr00
- mov rbx, rsp
- push rdi
- push rcx
- push rax
- %ifdef WIN64
- mov r8, rcx
- mov r9, rdi
- mov rcx, rax
- mov rdx, rbx
- %else
- mov rdx, rcx
- mov rcx, rdi
- mov rdi, rax
- mov rsi, rbx
- %endif
- call EXTN(jsimd_h2v1_merged_upsample_avx2)
- pop rax
- pop rcx
- pop rdi
- pop rsi
- pop rbx
- pop rdx
- add rdi, byte SIZEOF_JSAMPROW ; outptr1
- add rsi, byte SIZEOF_JSAMPROW ; inptr01
- push rdx ; inptr2
- push rbx ; inptr1
- push rsi ; inptr00
- mov rbx, rsp
- push rdi
- push rcx
- push rax
- %ifdef WIN64
- mov r8, rcx
- mov r9, rdi
- mov rcx, rax
- mov rdx, rbx
- %else
- mov rdx, rcx
- mov rcx, rdi
- mov rdi, rax
- mov rsi, rbx
- %endif
- call EXTN(jsimd_h2v1_merged_upsample_avx2)
- pop rax
- pop rcx
- pop rdi
- pop rsi
- pop rbx
- pop rdx
- pop rbx
- uncollect_args 4
- pop rbp
- ret
- ; For some reason, the OS X linker does not honor the request to align the
- ; segment unless we do this.
- align 32
|