1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jiss2red-64.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,576 @@
1.4 +;
1.5 +; jiss2red-64.asm - reduced-size IDCT (64-bit SSE2)
1.6 +;
1.7 +; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
1.8 +; Copyright 2009 D. R. Commander
1.9 +;
1.10 +; Based on
1.11 +; x86 SIMD extension for IJG JPEG library
1.12 +; Copyright (C) 1999-2006, MIYASAKA Masaru.
1.13 +; For conditions of distribution and use, see copyright notice in jsimdext.inc
1.14 +;
1.15 +; This file should be assembled with NASM (Netwide Assembler),
1.16 +; can *not* be assembled with Microsoft's MASM or any compatible
1.17 +; assembler (including Borland's Turbo Assembler).
1.18 +; NASM is available from http://nasm.sourceforge.net/ or
1.19 +; http://sourceforge.net/project/showfiles.php?group_id=6208
1.20 +;
1.21 +; This file contains inverse-DCT routines that produce reduced-size
1.22 +; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
1.23 +; The following code is based directly on the IJG's original jidctred.c;
1.24 +; see the jidctred.c for more details.
1.25 +;
1.26 +; [TAB8]
1.27 +
1.28 +%include "jsimdext.inc"
1.29 +%include "jdct.inc"
1.30 +
1.31 +; --------------------------------------------------------------------------
1.32 +
1.33 +%define CONST_BITS 13
1.34 +%define PASS1_BITS 2
1.35 +
1.36 +%define DESCALE_P1_4 (CONST_BITS-PASS1_BITS+1)
1.37 +%define DESCALE_P2_4 (CONST_BITS+PASS1_BITS+3+1)
1.38 +%define DESCALE_P1_2 (CONST_BITS-PASS1_BITS+2)
1.39 +%define DESCALE_P2_2 (CONST_BITS+PASS1_BITS+3+2)
1.40 +
1.41 +%if CONST_BITS == 13
1.42 +F_0_211 equ 1730 ; FIX(0.211164243)
1.43 +F_0_509 equ 4176 ; FIX(0.509795579)
1.44 +F_0_601 equ 4926 ; FIX(0.601344887)
1.45 +F_0_720 equ 5906 ; FIX(0.720959822)
1.46 +F_0_765 equ 6270 ; FIX(0.765366865)
1.47 +F_0_850 equ 6967 ; FIX(0.850430095)
1.48 +F_0_899 equ 7373 ; FIX(0.899976223)
1.49 +F_1_061 equ 8697 ; FIX(1.061594337)
1.50 +F_1_272 equ 10426 ; FIX(1.272758580)
1.51 +F_1_451 equ 11893 ; FIX(1.451774981)
1.52 +F_1_847 equ 15137 ; FIX(1.847759065)
1.53 +F_2_172 equ 17799 ; FIX(2.172734803)
1.54 +F_2_562 equ 20995 ; FIX(2.562915447)
1.55 +F_3_624 equ 29692 ; FIX(3.624509785)
1.56 +%else
1.57 +; NASM cannot do compile-time arithmetic on floating-point constants.
1.58 +%define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
1.59 +F_0_211 equ DESCALE( 226735879,30-CONST_BITS) ; FIX(0.211164243)
1.60 +F_0_509 equ DESCALE( 547388834,30-CONST_BITS) ; FIX(0.509795579)
1.61 +F_0_601 equ DESCALE( 645689155,30-CONST_BITS) ; FIX(0.601344887)
1.62 +F_0_720 equ DESCALE( 774124714,30-CONST_BITS) ; FIX(0.720959822)
1.63 +F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
1.64 +F_0_850 equ DESCALE( 913142361,30-CONST_BITS) ; FIX(0.850430095)
1.65 +F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
1.66 +F_1_061 equ DESCALE(1139878239,30-CONST_BITS) ; FIX(1.061594337)
1.67 +F_1_272 equ DESCALE(1366614119,30-CONST_BITS) ; FIX(1.272758580)
1.68 +F_1_451 equ DESCALE(1558831516,30-CONST_BITS) ; FIX(1.451774981)
1.69 +F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
1.70 +F_2_172 equ DESCALE(2332956230,30-CONST_BITS) ; FIX(2.172734803)
1.71 +F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
1.72 +F_3_624 equ DESCALE(3891787747,30-CONST_BITS) ; FIX(3.624509785)
1.73 +%endif
1.74 +
1.75 +; --------------------------------------------------------------------------
1.76 + SECTION SEG_CONST
1.77 +
1.78 + alignz 16
1.79 + global EXTN(jconst_idct_red_sse2)
1.80 +
1.81 +EXTN(jconst_idct_red_sse2):
1.82 +
1.83 +PW_F184_MF076 times 4 dw F_1_847,-F_0_765
1.84 +PW_F256_F089 times 4 dw F_2_562, F_0_899
1.85 +PW_F106_MF217 times 4 dw F_1_061,-F_2_172
1.86 +PW_MF060_MF050 times 4 dw -F_0_601,-F_0_509
1.87 +PW_F145_MF021 times 4 dw F_1_451,-F_0_211
1.88 +PW_F362_MF127 times 4 dw F_3_624,-F_1_272
1.89 +PW_F085_MF072 times 4 dw F_0_850,-F_0_720
1.90 +PD_DESCALE_P1_4 times 4 dd 1 << (DESCALE_P1_4-1)
1.91 +PD_DESCALE_P2_4 times 4 dd 1 << (DESCALE_P2_4-1)
1.92 +PD_DESCALE_P1_2 times 4 dd 1 << (DESCALE_P1_2-1)
1.93 +PD_DESCALE_P2_2 times 4 dd 1 << (DESCALE_P2_2-1)
1.94 +PB_CENTERJSAMP times 16 db CENTERJSAMPLE
1.95 +
1.96 + alignz 16
1.97 +
1.98 +; --------------------------------------------------------------------------
1.99 + SECTION SEG_TEXT
1.100 + BITS 64
1.101 +;
1.102 +; Perform dequantization and inverse DCT on one block of coefficients,
1.103 +; producing a reduced-size 4x4 output block.
1.104 +;
1.105 +; GLOBAL(void)
1.106 +; jsimd_idct_4x4_sse2 (void * dct_table, JCOEFPTR coef_block,
1.107 +; JSAMPARRAY output_buf, JDIMENSION output_col)
1.108 +;
1.109 +
1.110 +; r10 = void * dct_table
1.111 +; r11 = JCOEFPTR coef_block
1.112 +; r12 = JSAMPARRAY output_buf
1.113 +; r13 = JDIMENSION output_col
1.114 +
1.115 +%define original_rbp rbp+0
1.116 +%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
1.117 +%define WK_NUM 2
1.118 +
1.119 + align 16
1.120 + global EXTN(jsimd_idct_4x4_sse2)
1.121 +
1.122 +EXTN(jsimd_idct_4x4_sse2):
1.123 + push rbp
1.124 + mov rax,rsp ; rax = original rbp
1.125 + sub rsp, byte 4
1.126 + and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
1.127 + mov [rsp],rax
1.128 + mov rbp,rsp ; rbp = aligned rbp
1.129 + lea rsp, [wk(0)]
1.130 + collect_args
1.131 +
1.132 + ; ---- Pass 1: process columns from input.
1.133 +
1.134 + mov rdx, r10 ; quantptr
1.135 + mov rsi, r11 ; inptr
1.136 +
1.137 +%ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
1.138 + mov eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
1.139 + or eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
1.140 + jnz short .columnDCT
1.141 +
1.142 + movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
1.143 + movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
1.144 + por xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
1.145 + por xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
1.146 + por xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
1.147 + por xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
1.148 + por xmm0,xmm1
1.149 + packsswb xmm0,xmm0
1.150 + packsswb xmm0,xmm0
1.151 + movd eax,xmm0
1.152 + test rax,rax
1.153 + jnz short .columnDCT
1.154 +
1.155 + ; -- AC terms all zero
1.156 +
1.157 + movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
1.158 + pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.159 +
1.160 + psllw xmm0,PASS1_BITS
1.161 +
1.162 + movdqa xmm3,xmm0 ; xmm0=in0=(00 01 02 03 04 05 06 07)
1.163 + punpcklwd xmm0,xmm0 ; xmm0=(00 00 01 01 02 02 03 03)
1.164 + punpckhwd xmm3,xmm3 ; xmm3=(04 04 05 05 06 06 07 07)
1.165 +
1.166 + pshufd xmm1,xmm0,0x50 ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
1.167 + pshufd xmm0,xmm0,0xFA ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
1.168 + pshufd xmm6,xmm3,0x50 ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
1.169 + pshufd xmm3,xmm3,0xFA ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
1.170 +
1.171 + jmp near .column_end
1.172 +%endif
1.173 +.columnDCT:
1.174 +
1.175 + ; -- Odd part
1.176 +
1.177 + movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
1.178 + movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
1.179 + pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.180 + pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.181 + movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
1.182 + movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
1.183 + pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.184 + pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.185 +
1.186 + movdqa xmm4,xmm0
1.187 + movdqa xmm5,xmm0
1.188 + punpcklwd xmm4,xmm1
1.189 + punpckhwd xmm5,xmm1
1.190 + movdqa xmm0,xmm4
1.191 + movdqa xmm1,xmm5
1.192 + pmaddwd xmm4,[rel PW_F256_F089] ; xmm4=(tmp2L)
1.193 + pmaddwd xmm5,[rel PW_F256_F089] ; xmm5=(tmp2H)
1.194 + pmaddwd xmm0,[rel PW_F106_MF217] ; xmm0=(tmp0L)
1.195 + pmaddwd xmm1,[rel PW_F106_MF217] ; xmm1=(tmp0H)
1.196 +
1.197 + movdqa xmm6,xmm2
1.198 + movdqa xmm7,xmm2
1.199 + punpcklwd xmm6,xmm3
1.200 + punpckhwd xmm7,xmm3
1.201 + movdqa xmm2,xmm6
1.202 + movdqa xmm3,xmm7
1.203 + pmaddwd xmm6,[rel PW_MF060_MF050] ; xmm6=(tmp2L)
1.204 + pmaddwd xmm7,[rel PW_MF060_MF050] ; xmm7=(tmp2H)
1.205 + pmaddwd xmm2,[rel PW_F145_MF021] ; xmm2=(tmp0L)
1.206 + pmaddwd xmm3,[rel PW_F145_MF021] ; xmm3=(tmp0H)
1.207 +
1.208 + paddd xmm6,xmm4 ; xmm6=tmp2L
1.209 + paddd xmm7,xmm5 ; xmm7=tmp2H
1.210 + paddd xmm2,xmm0 ; xmm2=tmp0L
1.211 + paddd xmm3,xmm1 ; xmm3=tmp0H
1.212 +
1.213 + movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=tmp0L
1.214 + movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=tmp0H
1.215 +
1.216 + ; -- Even part
1.217 +
1.218 + movdqa xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
1.219 + movdqa xmm5, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
1.220 + movdqa xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
1.221 + pmullw xmm4, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.222 + pmullw xmm5, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.223 + pmullw xmm0, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.224 +
1.225 + pxor xmm1,xmm1
1.226 + pxor xmm2,xmm2
1.227 + punpcklwd xmm1,xmm4 ; xmm1=tmp0L
1.228 + punpckhwd xmm2,xmm4 ; xmm2=tmp0H
1.229 + psrad xmm1,(16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
1.230 + psrad xmm2,(16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
1.231 +
1.232 + movdqa xmm3,xmm5 ; xmm5=in2=z2
1.233 + punpcklwd xmm5,xmm0 ; xmm0=in6=z3
1.234 + punpckhwd xmm3,xmm0
1.235 + pmaddwd xmm5,[rel PW_F184_MF076] ; xmm5=tmp2L
1.236 + pmaddwd xmm3,[rel PW_F184_MF076] ; xmm3=tmp2H
1.237 +
1.238 + movdqa xmm4,xmm1
1.239 + movdqa xmm0,xmm2
1.240 + paddd xmm1,xmm5 ; xmm1=tmp10L
1.241 + paddd xmm2,xmm3 ; xmm2=tmp10H
1.242 + psubd xmm4,xmm5 ; xmm4=tmp12L
1.243 + psubd xmm0,xmm3 ; xmm0=tmp12H
1.244 +
1.245 + ; -- Final output stage
1.246 +
1.247 + movdqa xmm5,xmm1
1.248 + movdqa xmm3,xmm2
1.249 + paddd xmm1,xmm6 ; xmm1=data0L
1.250 + paddd xmm2,xmm7 ; xmm2=data0H
1.251 + psubd xmm5,xmm6 ; xmm5=data3L
1.252 + psubd xmm3,xmm7 ; xmm3=data3H
1.253 +
1.254 + movdqa xmm6,[rel PD_DESCALE_P1_4] ; xmm6=[rel PD_DESCALE_P1_4]
1.255 +
1.256 + paddd xmm1,xmm6
1.257 + paddd xmm2,xmm6
1.258 + psrad xmm1,DESCALE_P1_4
1.259 + psrad xmm2,DESCALE_P1_4
1.260 + paddd xmm5,xmm6
1.261 + paddd xmm3,xmm6
1.262 + psrad xmm5,DESCALE_P1_4
1.263 + psrad xmm3,DESCALE_P1_4
1.264 +
1.265 + packssdw xmm1,xmm2 ; xmm1=data0=(00 01 02 03 04 05 06 07)
1.266 + packssdw xmm5,xmm3 ; xmm5=data3=(30 31 32 33 34 35 36 37)
1.267 +
1.268 + movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp0L
1.269 + movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp0H
1.270 +
1.271 + movdqa xmm2,xmm4
1.272 + movdqa xmm3,xmm0
1.273 + paddd xmm4,xmm7 ; xmm4=data1L
1.274 + paddd xmm0,xmm6 ; xmm0=data1H
1.275 + psubd xmm2,xmm7 ; xmm2=data2L
1.276 + psubd xmm3,xmm6 ; xmm3=data2H
1.277 +
1.278 + movdqa xmm7,[rel PD_DESCALE_P1_4] ; xmm7=[rel PD_DESCALE_P1_4]
1.279 +
1.280 + paddd xmm4,xmm7
1.281 + paddd xmm0,xmm7
1.282 + psrad xmm4,DESCALE_P1_4
1.283 + psrad xmm0,DESCALE_P1_4
1.284 + paddd xmm2,xmm7
1.285 + paddd xmm3,xmm7
1.286 + psrad xmm2,DESCALE_P1_4
1.287 + psrad xmm3,DESCALE_P1_4
1.288 +
1.289 + packssdw xmm4,xmm0 ; xmm4=data1=(10 11 12 13 14 15 16 17)
1.290 + packssdw xmm2,xmm3 ; xmm2=data2=(20 21 22 23 24 25 26 27)
1.291 +
1.292 + movdqa xmm6,xmm1 ; transpose coefficients(phase 1)
1.293 + punpcklwd xmm1,xmm4 ; xmm1=(00 10 01 11 02 12 03 13)
1.294 + punpckhwd xmm6,xmm4 ; xmm6=(04 14 05 15 06 16 07 17)
1.295 + movdqa xmm7,xmm2 ; transpose coefficients(phase 1)
1.296 + punpcklwd xmm2,xmm5 ; xmm2=(20 30 21 31 22 32 23 33)
1.297 + punpckhwd xmm7,xmm5 ; xmm7=(24 34 25 35 26 36 27 37)
1.298 +
1.299 + movdqa xmm0,xmm1 ; transpose coefficients(phase 2)
1.300 + punpckldq xmm1,xmm2 ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
1.301 + punpckhdq xmm0,xmm2 ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
1.302 + movdqa xmm3,xmm6 ; transpose coefficients(phase 2)
1.303 + punpckldq xmm6,xmm7 ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
1.304 + punpckhdq xmm3,xmm7 ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
1.305 +.column_end:
1.306 +
1.307 + ; -- Prefetch the next coefficient block
1.308 +
1.309 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
1.310 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
1.311 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
1.312 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
1.313 +
1.314 + ; ---- Pass 2: process rows, store into output array.
1.315 +
1.316 + mov rax, [original_rbp]
1.317 + mov rdi, r12 ; (JSAMPROW *)
1.318 + mov rax, r13
1.319 +
1.320 + ; -- Even part
1.321 +
1.322 + pxor xmm4,xmm4
1.323 + punpcklwd xmm4,xmm1 ; xmm4=tmp0
1.324 + psrad xmm4,(16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
1.325 +
1.326 + ; -- Odd part
1.327 +
1.328 + punpckhwd xmm1,xmm0
1.329 + punpckhwd xmm6,xmm3
1.330 + movdqa xmm5,xmm1
1.331 + movdqa xmm2,xmm6
1.332 + pmaddwd xmm1,[rel PW_F256_F089] ; xmm1=(tmp2)
1.333 + pmaddwd xmm6,[rel PW_MF060_MF050] ; xmm6=(tmp2)
1.334 + pmaddwd xmm5,[rel PW_F106_MF217] ; xmm5=(tmp0)
1.335 + pmaddwd xmm2,[rel PW_F145_MF021] ; xmm2=(tmp0)
1.336 +
1.337 + paddd xmm6,xmm1 ; xmm6=tmp2
1.338 + paddd xmm2,xmm5 ; xmm2=tmp0
1.339 +
1.340 + ; -- Even part
1.341 +
1.342 + punpcklwd xmm0,xmm3
1.343 + pmaddwd xmm0,[rel PW_F184_MF076] ; xmm0=tmp2
1.344 +
1.345 + movdqa xmm7,xmm4
1.346 + paddd xmm4,xmm0 ; xmm4=tmp10
1.347 + psubd xmm7,xmm0 ; xmm7=tmp12
1.348 +
1.349 + ; -- Final output stage
1.350 +
1.351 + movdqa xmm1,[rel PD_DESCALE_P2_4] ; xmm1=[rel PD_DESCALE_P2_4]
1.352 +
1.353 + movdqa xmm5,xmm4
1.354 + movdqa xmm3,xmm7
1.355 + paddd xmm4,xmm6 ; xmm4=data0=(00 10 20 30)
1.356 + paddd xmm7,xmm2 ; xmm7=data1=(01 11 21 31)
1.357 + psubd xmm5,xmm6 ; xmm5=data3=(03 13 23 33)
1.358 + psubd xmm3,xmm2 ; xmm3=data2=(02 12 22 32)
1.359 +
1.360 + paddd xmm4,xmm1
1.361 + paddd xmm7,xmm1
1.362 + psrad xmm4,DESCALE_P2_4
1.363 + psrad xmm7,DESCALE_P2_4
1.364 + paddd xmm5,xmm1
1.365 + paddd xmm3,xmm1
1.366 + psrad xmm5,DESCALE_P2_4
1.367 + psrad xmm3,DESCALE_P2_4
1.368 +
1.369 + packssdw xmm4,xmm3 ; xmm4=(00 10 20 30 02 12 22 32)
1.370 + packssdw xmm7,xmm5 ; xmm7=(01 11 21 31 03 13 23 33)
1.371 +
1.372 + movdqa xmm0,xmm4 ; transpose coefficients(phase 1)
1.373 + punpcklwd xmm4,xmm7 ; xmm4=(00 01 10 11 20 21 30 31)
1.374 + punpckhwd xmm0,xmm7 ; xmm0=(02 03 12 13 22 23 32 33)
1.375 +
1.376 + movdqa xmm6,xmm4 ; transpose coefficients(phase 2)
1.377 + punpckldq xmm4,xmm0 ; xmm4=(00 01 02 03 10 11 12 13)
1.378 + punpckhdq xmm6,xmm0 ; xmm6=(20 21 22 23 30 31 32 33)
1.379 +
1.380 + packsswb xmm4,xmm6 ; xmm4=(00 01 02 03 10 11 12 13 20 ..)
1.381 + paddb xmm4,[rel PB_CENTERJSAMP]
1.382 +
1.383 + pshufd xmm2,xmm4,0x39 ; xmm2=(10 11 12 13 20 21 22 23 30 ..)
1.384 + pshufd xmm1,xmm4,0x4E ; xmm1=(20 21 22 23 30 31 32 33 00 ..)
1.385 + pshufd xmm3,xmm4,0x93 ; xmm3=(30 31 32 33 00 01 02 03 10 ..)
1.386 +
1.387 + mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
1.388 + mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
1.389 + movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
1.390 + movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
1.391 + mov rdx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
1.392 + mov rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
1.393 + movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
1.394 + movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
1.395 +
1.396 + uncollect_args
1.397 + mov rsp,rbp ; rsp <- aligned rbp
1.398 + pop rsp ; rsp <- original rbp
1.399 + pop rbp
1.400 + ret
1.401 +
1.402 +
1.403 +; --------------------------------------------------------------------------
1.404 +;
1.405 +; Perform dequantization and inverse DCT on one block of coefficients,
1.406 +; producing a reduced-size 2x2 output block.
1.407 +;
1.408 +; GLOBAL(void)
1.409 +; jsimd_idct_2x2_sse2 (void * dct_table, JCOEFPTR coef_block,
1.410 +; JSAMPARRAY output_buf, JDIMENSION output_col)
1.411 +;
1.412 +
1.413 +; r10 = void * dct_table
1.414 +; r11 = JCOEFPTR coef_block
1.415 +; r12 = JSAMPARRAY output_buf
1.416 +; r13 = JDIMENSION output_col
1.417 +
1.418 + align 16
1.419 + global EXTN(jsimd_idct_2x2_sse2)
1.420 +
1.421 +EXTN(jsimd_idct_2x2_sse2):
1.422 + push rbp
1.423 + mov rax,rsp
1.424 + mov rbp,rsp
1.425 + collect_args
1.426 + push rbx
1.427 +
1.428 + ; ---- Pass 1: process columns from input.
1.429 +
1.430 + mov rdx, r10 ; quantptr
1.431 + mov rsi, r11 ; inptr
1.432 +
1.433 + ; | input: | result: |
1.434 + ; | 00 01 ** 03 ** 05 ** 07 | |
1.435 + ; | 10 11 ** 13 ** 15 ** 17 | |
1.436 + ; | ** ** ** ** ** ** ** ** | |
1.437 + ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
1.438 + ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
1.439 + ; | 50 51 ** 53 ** 55 ** 57 | |
1.440 + ; | ** ** ** ** ** ** ** ** | |
1.441 + ; | 70 71 ** 73 ** 75 ** 77 | |
1.442 +
1.443 + ; -- Odd part
1.444 +
1.445 + movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
1.446 + movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
1.447 + pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.448 + pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.449 + movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
1.450 + movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
1.451 + pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.452 + pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.453 +
1.454 + ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
1.455 + ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
1.456 +
1.457 + pcmpeqd xmm7,xmm7
1.458 + pslld xmm7,WORD_BIT ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
1.459 +
1.460 + movdqa xmm4,xmm0 ; xmm4=(10 11 ** 13 ** 15 ** 17)
1.461 + movdqa xmm5,xmm2 ; xmm5=(50 51 ** 53 ** 55 ** 57)
1.462 + punpcklwd xmm4,xmm1 ; xmm4=(10 30 11 31 ** ** 13 33)
1.463 + punpcklwd xmm5,xmm3 ; xmm5=(50 70 51 71 ** ** 53 73)
1.464 + pmaddwd xmm4,[rel PW_F362_MF127]
1.465 + pmaddwd xmm5,[rel PW_F085_MF072]
1.466 +
1.467 + psrld xmm0,WORD_BIT ; xmm0=(11 -- 13 -- 15 -- 17 --)
1.468 + pand xmm1,xmm7 ; xmm1=(-- 31 -- 33 -- 35 -- 37)
1.469 + psrld xmm2,WORD_BIT ; xmm2=(51 -- 53 -- 55 -- 57 --)
1.470 + pand xmm3,xmm7 ; xmm3=(-- 71 -- 73 -- 75 -- 77)
1.471 + por xmm0,xmm1 ; xmm0=(11 31 13 33 15 35 17 37)
1.472 + por xmm2,xmm3 ; xmm2=(51 71 53 73 55 75 57 77)
1.473 + pmaddwd xmm0,[rel PW_F362_MF127]
1.474 + pmaddwd xmm2,[rel PW_F085_MF072]
1.475 +
1.476 + paddd xmm4,xmm5 ; xmm4=tmp0[col0 col1 **** col3]
1.477 + paddd xmm0,xmm2 ; xmm0=tmp0[col1 col3 col5 col7]
1.478 +
1.479 + ; -- Even part
1.480 +
1.481 + movdqa xmm6, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
1.482 + pmullw xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
1.483 +
1.484 + ; xmm6=(00 01 ** 03 ** 05 ** 07)
1.485 +
1.486 + movdqa xmm1,xmm6 ; xmm1=(00 01 ** 03 ** 05 ** 07)
1.487 + pslld xmm6,WORD_BIT ; xmm6=(-- 00 -- ** -- ** -- **)
1.488 + pand xmm1,xmm7 ; xmm1=(-- 01 -- 03 -- 05 -- 07)
1.489 + psrad xmm6,(WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****]
1.490 + psrad xmm1,(WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7]
1.491 +
1.492 + ; -- Final output stage
1.493 +
1.494 + movdqa xmm3,xmm6
1.495 + movdqa xmm5,xmm1
1.496 + paddd xmm6,xmm4 ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
1.497 + paddd xmm1,xmm0 ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
1.498 + psubd xmm3,xmm4 ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
1.499 + psubd xmm5,xmm0 ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
1.500 +
1.501 + movdqa xmm2,[rel PD_DESCALE_P1_2] ; xmm2=[rel PD_DESCALE_P1_2]
1.502 +
1.503 + punpckldq xmm6,xmm3 ; xmm6=(A0 B0 ** **)
1.504 +
1.505 + movdqa xmm7,xmm1
1.506 + punpcklqdq xmm1,xmm5 ; xmm1=(A1 A3 B1 B3)
1.507 + punpckhqdq xmm7,xmm5 ; xmm7=(A5 A7 B5 B7)
1.508 +
1.509 + paddd xmm6,xmm2
1.510 + psrad xmm6,DESCALE_P1_2
1.511 +
1.512 + paddd xmm1,xmm2
1.513 + paddd xmm7,xmm2
1.514 + psrad xmm1,DESCALE_P1_2
1.515 + psrad xmm7,DESCALE_P1_2
1.516 +
1.517 + ; -- Prefetch the next coefficient block
1.518 +
1.519 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
1.520 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
1.521 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
1.522 + prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
1.523 +
1.524 + ; ---- Pass 2: process rows, store into output array.
1.525 +
1.526 + mov rdi, r12 ; (JSAMPROW *)
1.527 + mov rax, r13
1.528 +
1.529 + ; | input:| result:|
1.530 + ; | A0 B0 | |
1.531 + ; | A1 B1 | C0 C1 |
1.532 + ; | A3 B3 | D0 D1 |
1.533 + ; | A5 B5 | |
1.534 + ; | A7 B7 | |
1.535 +
1.536 + ; -- Odd part
1.537 +
1.538 + packssdw xmm1,xmm1 ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
1.539 + packssdw xmm7,xmm7 ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
1.540 + pmaddwd xmm1,[rel PW_F362_MF127]
1.541 + pmaddwd xmm7,[rel PW_F085_MF072]
1.542 +
1.543 + paddd xmm1,xmm7 ; xmm1=tmp0[row0 row1 row0 row1]
1.544 +
1.545 + ; -- Even part
1.546 +
1.547 + pslld xmm6,(CONST_BITS+2) ; xmm6=tmp10[row0 row1 **** ****]
1.548 +
1.549 + ; -- Final output stage
1.550 +
1.551 + movdqa xmm4,xmm6
1.552 + paddd xmm6,xmm1 ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
1.553 + psubd xmm4,xmm1 ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
1.554 +
1.555 + punpckldq xmm6,xmm4 ; xmm6=(C0 D0 C1 D1)
1.556 +
1.557 + paddd xmm6,[rel PD_DESCALE_P2_2]
1.558 + psrad xmm6,DESCALE_P2_2
1.559 +
1.560 + packssdw xmm6,xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
1.561 + packsswb xmm6,xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
1.562 + paddb xmm6,[rel PB_CENTERJSAMP]
1.563 +
1.564 + pextrw ebx,xmm6,0x00 ; ebx=(C0 D0 -- --)
1.565 + pextrw ecx,xmm6,0x01 ; ecx=(C1 D1 -- --)
1.566 +
1.567 + mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
1.568 + mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
1.569 + mov WORD [rdx+rax*SIZEOF_JSAMPLE], bx
1.570 + mov WORD [rsi+rax*SIZEOF_JSAMPLE], cx
1.571 +
1.572 + pop rbx
1.573 + uncollect_args
1.574 + pop rbp
1.575 + ret
1.576 +
1.577 +; For some reason, the OS X linker does not honor the request to align the
1.578 +; segment unless we do this.
1.579 + align 16
