| 1 |
|
; Originally provided by Intel at AP-922 |
| 2 |
|
; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm |
| 3 |
|
; (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) |
| 4 |
|
; but in a limited edition. |
| 5 |
|
; New macro implements a column part for precise iDCT |
| 6 |
|
; The routine precision now satisfies IEEE standard 1180-1990. |
| 7 |
|
; |
| 8 |
|
; Copyright (c) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
| 9 |
|
; Rounding trick Copyright (c) 2000 Michel Lespinasse <walken@zoy.org> |
| 10 |
|
; |
| 11 |
|
; http://www.elecard.com/peter/idct.html |
| 12 |
|
; http://www.linuxvideo.org/mpeg2dec/ |
| 13 |
|
; |
| 14 |
|
;============================================================================= |
| 15 |
|
; |
| 16 |
|
; These examples contain code fragments for first stage iDCT 8x8 |
| 17 |
|
; (for rows) and first stage DCT 8x8 (for columns) |
| 18 |
|
; |
| 19 |
|
;============================================================================= |
| 20 |
|
; |
| 21 |
|
; 04.11.2001 nasm conversion; peter ross <pross@cs.rmit.edu.au> |
| 22 |
|
; |
| 23 |
|
|
| 24 |
|
; this 3dne function is compatible with iSSE, but is optimized specifically for |
| 25 |
|
; K7 pipelines (ca 5% gain) |
| 26 |
|
; |
| 27 |
|
;------------------------------------------------------------------------------ |
| 28 |
|
; 09.12.2002 Athlon optimizations contributed by Jaan Kalda |
| 29 |
|
;------------------------------------------------------------------------------ |
| 30 |
|
|
| 31 |
|
bits 32 |
| 32 |
|
|
| 33 |
|
%macro cglobal 1 |
| 34 |
|
%ifdef PREFIX |
| 35 |
|
global _%1 |
| 36 |
|
%define %1 _%1 |
| 37 |
|
%else |
| 38 |
|
global %1 |
| 39 |
|
%endif |
| 40 |
|
%endmacro |
| 41 |
|
|
| 42 |
|
%define BITS_INV_ACC 5 ; 4 or 5 for IEEE |
| 43 |
|
%define SHIFT_INV_ROW 16 - BITS_INV_ACC |
| 44 |
|
%define SHIFT_INV_COL 1 + BITS_INV_ACC |
| 45 |
|
%define RND_INV_ROW 1024 * (6 - BITS_INV_ACC) ; 1 << (SHIFT_INV_ROW-1) |
| 46 |
|
%define RND_INV_COL 16 * (BITS_INV_ACC - 3) ; 1 << (SHIFT_INV_COL-1) |
| 47 |
|
%define RND_INV_CORR RND_INV_COL - 1 ; correction -1.0 and round |
| 48 |
|
|
| 49 |
|
%define BITS_FRW_ACC 3 ; 2 or 3 for accuracy |
| 50 |
|
%define SHIFT_FRW_COL BITS_FRW_ACC |
| 51 |
|
%define SHIFT_FRW_ROW BITS_FRW_ACC + 17 |
| 52 |
|
%define RND_FRW_ROW 262144 * (BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) |
| 53 |
|
|
| 54 |
|
%ifdef FORMAT_COFF |
| 55 |
|
section .data data |
| 56 |
|
%else |
| 57 |
|
section .data data align=16 |
| 58 |
|
%endif |
| 59 |
|
|
| 60 |
|
align 16 |
| 61 |
|
|
| 62 |
|
one_corr dw 1, 1, 1, 1 |
| 63 |
|
round_inv_row dd RND_INV_ROW, RND_INV_ROW |
| 64 |
|
round_inv_col dw RND_INV_COL, RND_INV_COL, RND_INV_COL, RND_INV_COL |
| 65 |
|
round_inv_corr dw RND_INV_CORR, RND_INV_CORR, RND_INV_CORR, RND_INV_CORR |
| 66 |
|
round_frw_row dd RND_FRW_ROW, RND_FRW_ROW |
| 67 |
|
tg_1_16 dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 |
| 68 |
|
tg_2_16 dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 |
| 69 |
|
tg_3_16 dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 |
| 70 |
|
cos_4_16 dw -19195, -19195, -19195, -19195 ; cos * (2<<16) + 0.5 |
| 71 |
|
ocos_4_16 dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 |
| 72 |
|
|
| 73 |
|
otg_3_16 dw 21895, 21895, 21895, 21895 ; tg * (2<<16) + 0.5 |
| 74 |
|
|
| 75 |
|
%if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 |
| 76 |
|
rounder_0 dd 65536, 65536 |
| 77 |
|
rounder_4 dd 0, 0 |
| 78 |
|
rounder_1 dd 7195, 7195 |
| 79 |
|
rounder_7 dd 1024, 1024 |
| 80 |
|
rounder_2 dd 4520, 4520 |
| 81 |
|
rounder_6 dd 1024, 1024 |
| 82 |
|
rounder_3 dd 2407, 2407 |
| 83 |
|
rounder_5 dd 240, 240 |
| 84 |
|
|
| 85 |
|
%elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 |
| 86 |
|
rounder_0 dd 65536, 65536 |
| 87 |
|
rounder_4 dd 0, 0 |
| 88 |
|
rounder_1 dd 3597, 3597 |
| 89 |
|
rounder_7 dd 512, 512 |
| 90 |
|
rounder_2 dd 2260, 2260 |
| 91 |
|
rounder_6 dd 512, 512 |
| 92 |
|
rounder_3 dd 1203, 1203 |
| 93 |
|
rounder_5 dd 120, 120 |
| 94 |
|
%else |
| 95 |
|
|
| 96 |
|
%error invalid _SHIFT_INV_ROW_ |
| 97 |
|
|
| 98 |
|
%endif |
| 99 |
|
|
| 100 |
|
;============================================================================= |
| 101 |
|
; |
| 102 |
|
; The first stage iDCT 8x8 - inverse DCTs of rows |
| 103 |
|
; |
| 104 |
|
;----------------------------------------------------------------------------- |
| 105 |
|
; The 8-point inverse DCT direct algorithm |
| 106 |
|
;----------------------------------------------------------------------------- |
| 107 |
|
; |
| 108 |
|
; static const short w[32] = { |
| 109 |
|
; FIX(cos_4_16), FIX(cos_2_16), FIX(cos_4_16), FIX(cos_6_16), |
| 110 |
|
; FIX(cos_4_16), FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16), |
| 111 |
|
; FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16), FIX(cos_2_16), |
| 112 |
|
; FIX(cos_4_16), -FIX(cos_2_16), FIX(cos_4_16), -FIX(cos_6_16), |
| 113 |
|
; FIX(cos_1_16), FIX(cos_3_16), FIX(cos_5_16), FIX(cos_7_16), |
| 114 |
|
; FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16), |
| 115 |
|
; FIX(cos_5_16), -FIX(cos_1_16), FIX(cos_7_16), FIX(cos_3_16), |
| 116 |
|
; FIX(cos_7_16), -FIX(cos_5_16), FIX(cos_3_16), -FIX(cos_1_16) }; |
| 117 |
|
; |
| 118 |
|
; #define DCT_8_INV_ROW(x, y) |
| 119 |
|
; { |
| 120 |
|
; int a0, a1, a2, a3, b0, b1, b2, b3; |
| 121 |
|
; |
| 122 |
|
; a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3]; |
| 123 |
|
; a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7]; |
| 124 |
|
; a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11]; |
| 125 |
|
; a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15]; |
| 126 |
|
; b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19]; |
| 127 |
|
; b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23]; |
| 128 |
|
; b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27]; |
| 129 |
|
; b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31]; |
| 130 |
|
; |
| 131 |
|
; y[0] = SHIFT_ROUND ( a0 + b0 ); |
| 132 |
|
; y[1] = SHIFT_ROUND ( a1 + b1 ); |
| 133 |
|
; y[2] = SHIFT_ROUND ( a2 + b2 ); |
| 134 |
|
; y[3] = SHIFT_ROUND ( a3 + b3 ); |
| 135 |
|
; y[4] = SHIFT_ROUND ( a3 - b3 ); |
| 136 |
|
; y[5] = SHIFT_ROUND ( a2 - b2 ); |
| 137 |
|
; y[6] = SHIFT_ROUND ( a1 - b1 ); |
| 138 |
|
; y[7] = SHIFT_ROUND ( a0 - b0 ); |
| 139 |
|
; } |
| 140 |
|
; |
| 141 |
|
;----------------------------------------------------------------------------- |
| 142 |
|
; |
| 143 |
|
; In this implementation the outputs of the iDCT-1D are multiplied |
| 144 |
|
; for rows 0,4 - by cos_4_16, |
| 145 |
|
; for rows 1,7 - by cos_1_16, |
| 146 |
|
; for rows 2,6 - by cos_2_16, |
| 147 |
|
; for rows 3,5 - by cos_3_16 |
| 148 |
|
; and are shifted to the left for better accuracy |
| 149 |
|
; |
| 150 |
|
; For the constants used, |
| 151 |
|
; FIX(float_const) = (short) (float_const * (1<<15) + 0.5) |
| 152 |
|
; |
| 153 |
|
;============================================================================= |
| 154 |
|
|
| 155 |
|
;============================================================================= |
| 156 |
|
; MMX code |
| 157 |
|
;============================================================================= |
| 158 |
|
|
| 159 |
|
; Table for rows 0,4 - constants are multiplied by cos_4_16 |
| 160 |
|
|
| 161 |
|
tab_i_04 dw 16384, 16384, 16384, -16384 ; movq-> w06 w04 w02 w00 |
| 162 |
|
dw 21407, 8867, 8867, -21407 ; w07 w05 w03 w01 |
| 163 |
|
dw 16384, -16384, 16384, 16384 ; w14 w12 w10 w08 |
| 164 |
|
dw -8867, 21407, -21407, -8867 ; w15 w13 w11 w09 |
| 165 |
|
dw 22725, 12873, 19266, -22725 ; w22 w20 w18 w16 |
| 166 |
|
dw 19266, 4520, -4520, -12873 ; w23 w21 w19 w17 |
| 167 |
|
dw 12873, 4520, 4520, 19266 ; w30 w28 w26 w24 |
| 168 |
|
dw -22725, 19266, -12873, -22725 ; w31 w29 w27 w25 |
| 169 |
|
|
| 170 |
|
; Table for rows 1,7 - constants are multiplied by cos_1_16 |
| 171 |
|
|
| 172 |
|
tab_i_17 dw 22725, 22725, 22725, -22725 ; movq-> w06 w04 w02 w00 |
| 173 |
|
dw 29692, 12299, 12299, -29692 ; w07 w05 w03 w01 |
| 174 |
|
dw 22725, -22725, 22725, 22725 ; w14 w12 w10 w08 |
| 175 |
|
dw -12299, 29692, -29692, -12299 ; w15 w13 w11 w09 |
| 176 |
|
dw 31521, 17855, 26722, -31521 ; w22 w20 w18 w16 |
| 177 |
|
dw 26722, 6270, -6270, -17855 ; w23 w21 w19 w17 |
| 178 |
|
dw 17855, 6270, 6270, 26722 ; w30 w28 w26 w24 |
| 179 |
|
dw -31521, 26722, -17855, -31521 ; w31 w29 w27 w25 |
| 180 |
|
|
| 181 |
|
; Table for rows 2,6 - constants are multiplied by cos_2_16 |
| 182 |
|
|
| 183 |
|
tab_i_26 dw 21407, 21407, 21407, -21407 ; movq-> w06 w04 w02 w00 |
| 184 |
|
dw 27969, 11585, 11585, -27969 ; w07 w05 w03 w01 |
| 185 |
|
dw 21407, -21407, 21407, 21407 ; w14 w12 w10 w08 |
| 186 |
|
dw -11585, 27969, -27969, -11585 ; w15 w13 w11 w09 |
| 187 |
|
dw 29692, 16819, 25172, -29692 ; w22 w20 w18 w16 |
| 188 |
|
dw 25172, 5906, -5906, -16819 ; w23 w21 w19 w17 |
| 189 |
|
dw 16819, 5906, 5906, 25172 ; w30 w28 w26 w24 |
| 190 |
|
dw -29692, 25172, -16819, -29692 ; w31 w29 w27 w25 |
| 191 |
|
|
| 192 |
|
; Table for rows 3,5 - constants are multiplied by cos_3_16 |
| 193 |
|
|
| 194 |
|
tab_i_35 dw 19266, 19266, 19266, -19266 ; movq-> w06 w04 w02 w00 |
| 195 |
|
dw 25172, 10426, 10426, -25172 ; w07 w05 w03 w01 |
| 196 |
|
dw 19266, -19266, 19266, 19266 ; w14 w12 w10 w08 |
| 197 |
|
dw -10426, 25172, -25172, -10426 ; w15 w13 w11 w09 |
| 198 |
|
dw 26722, 15137, 22654, -26722 ; w22 w20 w18 w16 |
| 199 |
|
dw 22654, 5315, -5315, -15137 ; w23 w21 w19 w17 |
| 200 |
|
dw 15137, 5315, 5315, 22654 ; w30 w28 w26 w24 |
| 201 |
|
dw -26722, 22654, -15137, -26722 ; w31 w29 w27 w25 |
| 202 |
|
|
| 203 |
|
;----------------------------------------------------------------------------- |
| 204 |
|
|
| 205 |
|
; |
| 206 |
|
; DCT_8_INV_ROW_1 INP, OUT, TABLE, ROUNDER |
| 207 |
|
; |
| 208 |
|
|
| 209 |
|
%macro DCT_8_INV_ROW_1 4 |
| 210 |
|
|
| 211 |
|
movq mm0, [%1] ; 0 ; x3 x2 x1 x0 |
| 212 |
|
|
| 213 |
|
movq mm1, [%1+8] ; 1 ; x7 x6 x5 x4 |
| 214 |
|
movq mm2, mm0 ; 2 ; x3 x2 x1 x0 |
| 215 |
|
|
| 216 |
|
movq mm3, [%3] ; 3 ; w06 w04 w02 w00 |
| 217 |
|
punpcklwd mm0, mm1 ; x5 x1 x4 x0 |
| 218 |
|
|
| 219 |
|
movq mm5, mm0 ; 5 ; x5 x1 x4 x0 |
| 220 |
|
punpckldq mm0, mm0 ; x4 x0 x4 x0 |
| 221 |
|
|
| 222 |
|
movq mm4, [%3+8] ; 4 ; w07 w05 w03 w01 |
| 223 |
|
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 |
| 224 |
|
|
| 225 |
|
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 |
| 226 |
|
movq mm6, mm2 ; 6 ; x7 x3 x6 x2 |
| 227 |
|
|
| 228 |
|
movq mm1, [%3+32] ; 1 ; w22 w20 w18 w16 |
| 229 |
|
punpckldq mm2, mm2 ; x6 x2 x6 x2 |
| 230 |
|
|
| 231 |
|
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 |
| 232 |
|
punpckhdq mm5, mm5 ; x5 x1 x5 x1 |
| 233 |
|
|
| 234 |
|
pmaddwd mm0, [%3+16] ; x4*w14+x0*w12 x4*w10+x0*w08 |
| 235 |
|
punpckhdq mm6, mm6 ; x7 x3 x7 x3 |
| 236 |
|
|
| 237 |
|
movq mm7, [%3+40] ; 7 ; w23 w21 w19 w17 |
| 238 |
|
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 |
| 239 |
|
|
| 240 |
|
paddd mm3, [%4] ; +%4 |
| 241 |
|
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 |
| 242 |
|
|
| 243 |
|
pmaddwd mm2, [%3+24] ; x6*w15+x2*w13 x6*w11+x2*w09 |
| 244 |
|
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) |
| 245 |
|
|
| 246 |
|
pmaddwd mm5, [%3+48] ; x5*w30+x1*w28 x5*w26+x1*w24 |
| 247 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 248 |
|
|
| 249 |
|
pmaddwd mm6, [%3+56] ; x7*w31+x3*w29 x7*w27+x3*w25 |
| 250 |
|
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) |
| 251 |
|
|
| 252 |
|
paddd mm0, [%4] ; +%4 |
| 253 |
|
psubd mm3, mm1 ; a1-b1 a0-b0 |
| 254 |
|
|
| 255 |
|
psrad mm3, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 256 |
|
paddd mm1, mm4 ; 4 ; a1+b1 a0+b0 |
| 257 |
|
|
| 258 |
|
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) |
| 259 |
|
psrad mm1, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 260 |
|
|
| 261 |
|
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) |
| 262 |
|
movq mm4, mm0 ; 4 ; a3 a2 |
| 263 |
|
|
| 264 |
|
paddd mm0, mm5 ; a3+b3 a2+b2 |
| 265 |
|
psubd mm4, mm5 ; 5 ; a3-b3 a2-b2 |
| 266 |
|
|
| 267 |
|
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 268 |
|
psrad mm4, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 269 |
|
|
| 270 |
|
packssdw mm1, mm0 ; 0 ; y3 y2 y1 y0 |
| 271 |
|
packssdw mm4, mm3 ; 3 ; y6 y7 y4 y5 |
| 272 |
|
|
| 273 |
|
movq mm7, mm4 ; 7 ; y6 y7 y4 y5 |
| 274 |
|
psrld mm4, 16 ; 0 y6 0 y4 |
| 275 |
|
|
| 276 |
|
pslld mm7, 16 ; y7 0 y5 0 |
| 277 |
|
movq [%2], mm1 ; 1 ; save y3 y2 y1 y0 |
| 278 |
|
|
| 279 |
|
por mm7, mm4 ; 4 ; y7 y6 y5 y4 |
| 280 |
|
movq [%2+8], mm7 ; 7 ; save y7 y6 y5 y4 |
| 281 |
|
%endmacro |
| 282 |
|
|
| 283 |
|
|
| 284 |
|
|
| 285 |
|
|
| 286 |
|
;============================================================================= |
| 287 |
|
; code for Pentium III |
| 288 |
|
;============================================================================= |
| 289 |
|
|
| 290 |
|
; %3 for rows 0,4 - constants are multiplied by cos_4_16 |
| 291 |
|
|
| 292 |
|
tab_i_04_sse dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 |
| 293 |
|
dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 |
| 294 |
|
dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 |
| 295 |
|
dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 |
| 296 |
|
dw 22725, 19266, 19266, -4520 ; w21 w20 w17 w16 |
| 297 |
|
dw 12873, 4520, -22725, -12873 ; w23 w22 w19 w18 |
| 298 |
|
dw 12873, -22725, 4520, -12873 ; w29 w28 w25 w24 |
| 299 |
|
dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 |
| 300 |
|
|
| 301 |
|
; %3 for rows 1,7 - constants are multiplied by cos_1_16 |
| 302 |
|
|
| 303 |
|
tab_i_17_sse dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 |
| 304 |
|
dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 |
| 305 |
|
dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 |
| 306 |
|
dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 |
| 307 |
|
dw 31521, 26722, 26722, -6270 ; w21 w20 w17 w16 |
| 308 |
|
dw 17855, 6270, -31521, -17855 ; w23 w22 w19 w18 |
| 309 |
|
dw 17855, -31521, 6270, -17855 ; w29 w28 w25 w24 |
| 310 |
|
dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 |
| 311 |
|
|
| 312 |
|
; %3 for rows 2,6 - constants are multiplied by cos_2_16 |
| 313 |
|
|
| 314 |
|
tab_i_26_sse dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 |
| 315 |
|
dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 |
| 316 |
|
dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 |
| 317 |
|
dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 |
| 318 |
|
dw 29692, 25172, 25172, -5906 ; w21 w20 w17 w16 |
| 319 |
|
dw 16819, 5906, -29692, -16819 ; w23 w22 w19 w18 |
| 320 |
|
dw 16819, -29692, 5906, -16819 ; w29 w28 w25 w24 |
| 321 |
|
dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 |
| 322 |
|
|
| 323 |
|
; %3 for rows 3,5 - constants are multiplied by cos_3_16 |
| 324 |
|
|
| 325 |
|
tab_i_35_sse dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 |
| 326 |
|
dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 |
| 327 |
|
dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 |
| 328 |
|
dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 |
| 329 |
|
dw 26722, 22654, 22654, -5315 ; w21 w20 w17 w16 |
| 330 |
|
dw 15137, 5315, -26722, -15137 ; w23 w22 w19 w18 |
| 331 |
|
dw 15137, -26722, 5315, -15137 ; w29 w28 w25 w24 |
| 332 |
|
dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 |
| 333 |
|
|
| 334 |
|
;============================================================================= |
| 335 |
|
; |
| 336 |
|
;============================================================================= |
| 337 |
|
|
| 338 |
|
;============================================================================= |
| 339 |
|
; |
| 340 |
|
; The first stage DCT 8x8 - forward DCTs of columns |
| 341 |
|
; |
| 342 |
|
; The %2puts are multiplied |
| 343 |
|
; for rows 0,4 - on cos_4_16, |
| 344 |
|
; for rows 1,7 - on cos_1_16, |
| 345 |
|
; for rows 2,6 - on cos_2_16, |
| 346 |
|
; for rows 3,5 - on cos_3_16 |
| 347 |
|
; and are shifted to the left for rise of accuracy |
| 348 |
|
; |
| 349 |
|
;----------------------------------------------------------------------------- |
| 350 |
|
; |
| 351 |
|
; The 8-point scaled forward DCT algorithm (26a8m) |
| 352 |
|
; |
| 353 |
|
;----------------------------------------------------------------------------- |
| 354 |
|
; |
| 355 |
|
; #define DCT_8_FRW_COL(x, y) |
| 356 |
|
;{ |
| 357 |
|
; short t0, t1, t2, t3, t4, t5, t6, t7; |
| 358 |
|
; short tp03, tm03, tp12, tm12, tp65, tm65; |
| 359 |
|
; short tp465, tm465, tp765, tm765; |
| 360 |
|
; |
| 361 |
|
; t0 = LEFT_SHIFT ( x[0] + x[7] ); |
| 362 |
|
; t1 = LEFT_SHIFT ( x[1] + x[6] ); |
| 363 |
|
; t2 = LEFT_SHIFT ( x[2] + x[5] ); |
| 364 |
|
; t3 = LEFT_SHIFT ( x[3] + x[4] ); |
| 365 |
|
; t4 = LEFT_SHIFT ( x[3] - x[4] ); |
| 366 |
|
; t5 = LEFT_SHIFT ( x[2] - x[5] ); |
| 367 |
|
; t6 = LEFT_SHIFT ( x[1] - x[6] ); |
| 368 |
|
; t7 = LEFT_SHIFT ( x[0] - x[7] ); |
| 369 |
|
; |
| 370 |
|
; tp03 = t0 + t3; |
| 371 |
|
; tm03 = t0 - t3; |
| 372 |
|
; tp12 = t1 + t2; |
| 373 |
|
; tm12 = t1 - t2; |
| 374 |
|
; |
| 375 |
|
; y[0] = tp03 + tp12; |
| 376 |
|
; y[4] = tp03 - tp12; |
| 377 |
|
; |
| 378 |
|
; y[2] = tm03 + tm12 * tg_2_16; |
| 379 |
|
; y[6] = tm03 * tg_2_16 - tm12; |
| 380 |
|
; |
| 381 |
|
; tp65 =(t6 +t5 )*cos_4_16; |
| 382 |
|
; tm65 =(t6 -t5 )*cos_4_16; |
| 383 |
|
; |
| 384 |
|
; tp765 = t7 + tp65; |
| 385 |
|
; tm765 = t7 - tp65; |
| 386 |
|
; tp465 = t4 + tm65; |
| 387 |
|
; tm465 = t4 - tm65; |
| 388 |
|
; |
| 389 |
|
; y[1] = tp765 + tp465 * tg_1_16; |
| 390 |
|
; y[7] = tp765 * tg_1_16 - tp465; |
| 391 |
|
; y[5] = tm765 * tg_3_16 + tm465; |
| 392 |
|
; y[3] = tm765 - tm465 * tg_3_16; |
| 393 |
|
;} |
| 394 |
|
; |
| 395 |
|
;============================================================================= |
| 396 |
|
|
| 397 |
|
|
| 398 |
|
; |
| 399 |
|
; DCT_8_FRW_COL_4 INP, OUT |
| 400 |
|
; |
| 401 |
|
|
| 402 |
|
%macro DCT_8_FRW_COL_4 2 |
| 403 |
|
|
| 404 |
|
LOCAL x0, x1, x2, x3, x4, x5, x6, x7 |
| 405 |
|
LOCAL y0, y1, y2, y3, y4, y5, y6, y7 |
| 406 |
|
x0 equ [%1 + 0*16] |
| 407 |
|
x1 equ [%1 + 1*16] |
| 408 |
|
x2 equ [%1 + 2*16] |
| 409 |
|
x3 equ [%1 + 3*16] |
| 410 |
|
x4 equ [%1 + 4*16] |
| 411 |
|
x5 equ [%1 + 5*16] |
| 412 |
|
x6 equ [%1 + 6*16] |
| 413 |
|
x7 equ [%1 + 7*16] |
| 414 |
|
y0 equ [%2 + 0*16] |
| 415 |
|
y1 equ [%2 + 1*16] |
| 416 |
|
y2 equ [%2 + 2*16] |
| 417 |
|
y3 equ [%2 + 3*16] |
| 418 |
|
y4 equ [%2 + 4*16] |
| 419 |
|
y5 equ [%2 + 5*16] |
| 420 |
|
y6 equ [%2 + 6*16] |
| 421 |
|
y7 equ [%2 + 7*16] |
| 422 |
|
movq mm0, x1 ; 0 ; x1 |
| 423 |
|
movq mm1, x6 ; 1 ; x6 |
| 424 |
|
movq mm2, mm0 ; 2 ; x1 |
| 425 |
|
movq mm3, x2 ; 3 ; x2 |
| 426 |
|
paddsw mm0, mm1 ; t1 = x[1] + x[6] |
| 427 |
|
movq mm4, x5 ; 4 ; x5 |
| 428 |
|
psllw mm0, SHIFT_FRW_COL ; t1 |
| 429 |
|
movq mm5, x0 ; 5 ; x0 |
| 430 |
|
paddsw mm4, mm3 ; t2 = x[2] + x[5] |
| 431 |
|
paddsw mm5, x7 ; t0 = x[0] + x[7] |
| 432 |
|
psllw mm4, SHIFT_FRW_COL ; t2 |
| 433 |
|
movq mm6, mm0 ; 6 ; t1 |
| 434 |
|
psubsw mm2, mm1 ; 1 ; t6 = x[1] - x[6] |
| 435 |
|
movq mm1, [tg_2_16] ; 1 ; tg_2_16 |
| 436 |
|
psubsw mm0, mm4 ; tm12 = t1 - t2 |
| 437 |
|
movq mm7, x3 ; 7 ; x3 |
| 438 |
|
pmulhw mm1, mm0 ; tm12*tg_2_16 |
| 439 |
|
paddsw mm7, x4 ; t3 = x[3] + x[4] |
| 440 |
|
psllw mm5, SHIFT_FRW_COL ; t0 |
| 441 |
|
paddsw mm6, mm4 ; 4 ; tp12 = t1 + t2 |
| 442 |
|
psllw mm7, SHIFT_FRW_COL ; t3 |
| 443 |
|
movq mm4, mm5 ; 4 ; t0 |
| 444 |
|
psubsw mm5, mm7 ; tm03 = t0 - t3 |
| 445 |
|
paddsw mm1, mm5 ; y2 = tm03 + tm12*tg_2_16 |
| 446 |
|
paddsw mm4, mm7 ; 7 ; tp03 = t0 + t3 |
| 447 |
|
por mm1, [one_corr] ; correction y2 +0.5 |
| 448 |
|
psllw mm2, SHIFT_FRW_COL+1 ; t6 |
| 449 |
|
pmulhw mm5, [tg_2_16] ; tm03*tg_2_16 |
| 450 |
|
movq mm7, mm4 ; 7 ; tp03 |
| 451 |
|
psubsw mm3, x5 ; t5 = x[2] - x[5] |
| 452 |
|
psubsw mm4, mm6 ; y4 = tp03 - tp12 |
| 453 |
|
movq y2, mm1 ; 1 ; save y2 |
| 454 |
|
paddsw mm7, mm6 ; 6 ; y0 = tp03 + tp12 |
| 455 |
|
movq mm1, x3 ; 1 ; x3 |
| 456 |
|
psllw mm3, SHIFT_FRW_COL+1 ; t5 |
| 457 |
|
psubsw mm1, x4 ; t4 = x[3] - x[4] |
| 458 |
|
movq mm6, mm2 ; 6 ; t6 |
| 459 |
|
movq y4, mm4 ; 4 ; save y4 |
| 460 |
|
paddsw mm2, mm3 ; t6 + t5 |
| 461 |
|
pmulhw mm2, [ocos_4_16] ; tp65 = (t6 + t5)*cos_4_16 |
| 462 |
|
psubsw mm6, mm3 ; 3 ; t6 - t5 |
| 463 |
|
pmulhw mm6, [ocos_4_16] ; tm65 = (t6 - t5)*cos_4_16 |
| 464 |
|
psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12 |
| 465 |
|
por mm5, [one_corr] ; correction y6 +0.5 |
| 466 |
|
psllw mm1, SHIFT_FRW_COL ; t4 |
| 467 |
|
por mm2, [one_corr] ; correction tp65 +0.5 |
| 468 |
|
movq mm4, mm1 ; 4 ; t4 |
| 469 |
|
movq mm3, x0 ; 3 ; x0 |
| 470 |
|
paddsw mm1, mm6 ; tp465 = t4 + tm65 |
| 471 |
|
psubsw mm3, x7 ; t7 = x[0] - x[7] |
| 472 |
|
psubsw mm4, mm6 ; 6 ; tm465 = t4 - tm65 |
| 473 |
|
movq mm0, [tg_1_16] ; 0 ; tg_1_16 |
| 474 |
|
psllw mm3, SHIFT_FRW_COL ; t7 |
| 475 |
|
movq mm6, [tg_3_16] ; 6 ; tg_3_16 |
| 476 |
|
pmulhw mm0, mm1 ; tp465*tg_1_16 |
| 477 |
|
movq y0, mm7 ; 7 ; save y0 |
| 478 |
|
pmulhw mm6, mm4 ; tm465*tg_3_16 |
| 479 |
|
movq y6, mm5 ; 5 ; save y6 |
| 480 |
|
movq mm7, mm3 ; 7 ; t7 |
| 481 |
|
movq mm5, [tg_3_16] ; 5 ; tg_3_16 |
| 482 |
|
psubsw mm7, mm2 ; tm765 = t7 - tp65 |
| 483 |
|
paddsw mm3, mm2 ; 2 ; tp765 = t7 + tp65 |
| 484 |
|
pmulhw mm5, mm7 ; tm765*tg_3_16 |
| 485 |
|
paddsw mm0, mm3 ; y1 = tp765 + tp465*tg_1_16 |
| 486 |
|
paddsw mm6, mm4 ; tm465*tg_3_16 |
| 487 |
|
pmulhw mm3, [tg_1_16] ; tp765*tg_1_16 |
| 488 |
|
por mm0, [one_corr] ; correction y1 +0.5 |
| 489 |
|
paddsw mm5, mm7 ; tm765*tg_3_16 |
| 490 |
|
psubsw mm7, mm6 ; 6 ; y3 = tm765 - tm465*tg_3_16 |
| 491 |
|
movq y1, mm0 ; 0 ; save y1 |
| 492 |
|
paddsw mm5, mm4 ; 4 ; y5 = tm765*tg_3_16 + tm465 |
| 493 |
|
movq y3, mm7 ; 7 ; save y3 |
| 494 |
|
psubsw mm3, mm1 ; 1 ; y7 = tp765*tg_1_16 - tp465 |
| 495 |
|
movq y5, mm5 ; 5 ; save y5 |
| 496 |
|
movq y7, mm3 ; 3 ; save y7 |
| 497 |
|
%endmacro |
| 498 |
|
|
| 499 |
|
|
| 500 |
|
; |
| 501 |
|
; DCT_8_INV_COL_4 INP,OUT |
| 502 |
|
; |
| 503 |
|
|
| 504 |
|
%macro DCT_8_INV_COL_4 2 |
| 505 |
|
movq mm0, [tg_3_16] |
| 506 |
|
|
| 507 |
|
movq mm3, [%1+16*3] |
| 508 |
|
movq mm1, mm0 ; tg_3_16 |
| 509 |
|
|
| 510 |
|
movq mm5, [%1+16*5] |
| 511 |
|
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
| 512 |
|
|
| 513 |
|
movq mm4, [tg_1_16] |
| 514 |
|
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
| 515 |
|
|
| 516 |
|
movq mm7, [%1+16*7] |
| 517 |
|
movq mm2, mm4 ; tg_1_16 |
| 518 |
|
|
| 519 |
|
movq mm6, [%1+16*1] |
| 520 |
|
pmulhw mm4, mm7 ; x7*tg_1_16 |
| 521 |
|
|
| 522 |
|
paddsw mm0, mm3 ; x3*tg_3_16 |
| 523 |
|
pmulhw mm2, mm6 ; x1*tg_1_16 |
| 524 |
|
|
| 525 |
|
paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) |
| 526 |
|
psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 |
| 527 |
|
|
| 528 |
|
movq mm3, [ocos_4_16] |
| 529 |
|
paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
| 530 |
|
|
| 531 |
|
paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 |
| 532 |
|
psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 |
| 533 |
|
|
| 534 |
|
movq mm5, mm4 ; tp17 |
| 535 |
|
movq mm6, mm2 ; tm17 |
| 536 |
|
|
| 537 |
|
paddsw mm5, mm1 ; tp17+tp35 = b0 |
| 538 |
|
psubsw mm6, mm0 ; tm17-tm35 = b3 |
| 539 |
|
|
| 540 |
|
psubsw mm4, mm1 ; tp17-tp35 = t1 |
| 541 |
|
paddsw mm2, mm0 ; tm17+tm35 = t2 |
| 542 |
|
|
| 543 |
|
movq mm7, [tg_2_16] |
| 544 |
|
movq mm1, mm4 ; t1 |
| 545 |
|
|
| 546 |
|
; movq [SCRATCH+0], mm5 ; save b0 |
| 547 |
|
movq [%2+3*16], mm5 ; save b0 |
| 548 |
|
paddsw mm1, mm2 ; t1+t2 |
| 549 |
|
|
| 550 |
|
; movq [SCRATCH+8], mm6 ; save b3 |
| 551 |
|
movq [%2+5*16], mm6 ; save b3 |
| 552 |
|
psubsw mm4, mm2 ; t1-t2 |
| 553 |
|
|
| 554 |
|
movq mm5, [%1+2*16] |
| 555 |
|
movq mm0, mm7 ; tg_2_16 |
| 556 |
|
|
| 557 |
|
movq mm6, [%1+6*16] |
| 558 |
|
pmulhw mm0, mm5 ; x2*tg_2_16 |
| 559 |
|
|
| 560 |
|
pmulhw mm7, mm6 ; x6*tg_2_16 |
| 561 |
|
; slot |
| 562 |
|
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
| 563 |
|
; slot |
| 564 |
|
movq mm2, [%1+0*16] |
| 565 |
|
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
| 566 |
|
|
| 567 |
|
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
| 568 |
|
movq mm3, mm2 ; x0 |
| 569 |
|
|
| 570 |
|
movq mm6, [%1+4*16] |
| 571 |
|
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
| 572 |
|
|
| 573 |
|
paddsw mm2, mm6 ; x0+x4 = tp04 |
| 574 |
|
psubsw mm3, mm6 ; x0-x4 = tm04 |
| 575 |
|
|
| 576 |
|
movq mm5, mm2 ; tp04 |
| 577 |
|
movq mm6, mm3 ; tm04 |
| 578 |
|
|
| 579 |
|
psubsw mm2, mm7 ; tp04-tp26 = a3 |
| 580 |
|
paddsw mm3, mm0 ; tm04+tm26 = a1 |
| 581 |
|
|
| 582 |
|
paddsw mm1, mm1 ; b1 |
| 583 |
|
paddsw mm4, mm4 ; b2 |
| 584 |
|
|
| 585 |
|
paddsw mm5, mm7 ; tp04+tp26 = a0 |
| 586 |
|
psubsw mm6, mm0 ; tm04-tm26 = a2 |
| 587 |
|
|
| 588 |
|
movq mm7, mm3 ; a1 |
| 589 |
|
movq mm0, mm6 ; a2 |
| 590 |
|
|
| 591 |
|
paddsw mm3, mm1 ; a1+b1 |
| 592 |
|
paddsw mm6, mm4 ; a2+b2 |
| 593 |
|
|
| 594 |
|
psraw mm3, SHIFT_INV_COL ; dst1 |
| 595 |
|
psubsw mm7, mm1 ; a1-b1 |
| 596 |
|
|
| 597 |
|
psraw mm6, SHIFT_INV_COL ; dst2 |
| 598 |
|
psubsw mm0, mm4 ; a2-b2 |
| 599 |
|
|
| 600 |
|
; movq mm1, [SCRATCH+0] ; load b0 |
| 601 |
|
movq mm1, [%2+3*16] ; load b0 |
| 602 |
|
psraw mm7, SHIFT_INV_COL ; dst6 |
| 603 |
|
|
| 604 |
|
movq mm4, mm5 ; a0 |
| 605 |
|
psraw mm0, SHIFT_INV_COL ; dst5 |
| 606 |
|
|
| 607 |
|
movq [%2+1*16], mm3 |
| 608 |
|
paddsw mm5, mm1 ; a0+b0 |
| 609 |
|
|
| 610 |
|
movq [%2+2*16], mm6 |
| 611 |
|
psubsw mm4, mm1 ; a0-b0 |
| 612 |
|
|
| 613 |
|
; movq mm3, [SCRATCH+8] ; load b3 |
| 614 |
|
movq mm3, [%2+5*16] ; load b3 |
| 615 |
|
psraw mm5, SHIFT_INV_COL ; dst0 |
| 616 |
|
|
| 617 |
|
movq mm6, mm2 ; a3 |
| 618 |
|
psraw mm4, SHIFT_INV_COL ; dst7 |
| 619 |
|
|
| 620 |
|
movq [%2+5*16], mm0 |
| 621 |
|
paddsw mm2, mm3 ; a3+b3 |
| 622 |
|
|
| 623 |
|
movq [%2+6*16], mm7 |
| 624 |
|
psubsw mm6, mm3 ; a3-b3 |
| 625 |
|
|
| 626 |
|
movq [%2+0*16], mm5 |
| 627 |
|
psraw mm2, SHIFT_INV_COL ; dst3 |
| 628 |
|
|
| 629 |
|
movq [%2+7*16], mm4 |
| 630 |
|
psraw mm6, SHIFT_INV_COL ; dst4 |
| 631 |
|
|
| 632 |
|
movq [%2+3*16], mm2 |
| 633 |
|
|
| 634 |
|
movq [%2+4*16], mm6 |
| 635 |
|
%endmacro |
| 636 |
|
|
| 637 |
|
|
| 638 |
|
;============================================================================= |
| 639 |
|
; |
| 640 |
|
; void idct_sse (short * const src_result); |
| 641 |
|
; |
| 642 |
|
;============================================================================= |
| 643 |
|
|
| 644 |
|
align 16 |
| 645 |
|
cglobal idct_3dne |
| 646 |
|
idct_3dne |
| 647 |
|
mov eax,[esp+4] |
| 648 |
|
; DCT_8_INV_ROW_1_s [eax+64], [eax+64], tab_i_04_sse, rounder_4 ;rounder_4=0 |
| 649 |
|
pshufw mm0, [eax+64],10001000b ; x2 x0 x2 x0 |
| 650 |
|
movq mm3, [tab_i_04_sse] ; 3 ; w05 w04 w01 w00 |
| 651 |
|
pshufw mm1, [eax+64+8],10001000b ; x6 x4 x6 x4 |
| 652 |
|
movq mm4, [tab_i_04_sse+8] ; 4 ; w07 w06 w03 w02 |
| 653 |
|
pshufw mm2, [eax+64],11011101b ; x3 x1 x3 x1 |
| 654 |
|
pshufw mm5, [eax+64+8],11011101b ; x7 x5 x7 x5 |
| 655 |
|
movq mm6, [tab_i_04_sse+32] ; 6 ; w21 w20 w17 w16 |
| 656 |
|
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 657 |
|
movq mm7, [tab_i_04_sse+40] ; 7 ; w23 w22 w19 w18 ; |
| 658 |
|
pmaddwd mm0, [tab_i_04_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 659 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 660 |
|
pmaddwd mm1, [tab_i_04_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 661 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 662 |
|
pmaddwd mm2, [tab_i_04_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 663 |
|
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 664 |
|
pmaddwd mm5, [tab_i_04_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 665 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 666 |
|
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
| 667 |
|
pshufw mm1, [eax+80+8],10001000b ; x6 x4 x6 x4 |
| 668 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 669 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 670 |
|
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 671 |
|
pshufw mm5, [eax+80],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
| 672 |
|
movq mm7, mm0 ; 7 ; a3 a2 |
| 673 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
| 674 |
|
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 675 |
|
movq mm3, [tab_i_35_sse] ; 3 ; w05 w04 w01 w00 |
| 676 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 677 |
|
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
| 678 |
|
pshufw mm2, [eax+80],11011101b ; x3 x1 x3 x1 |
| 679 |
|
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 680 |
|
pmaddwd mm5, [tab_i_35_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 681 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 682 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 683 |
|
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 684 |
|
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 685 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 686 |
|
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
| 687 |
|
pshufw mm0, [eax+80+8],11011101b ; x7 x5 x7 x5 |
| 688 |
|
movq [eax+64], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
| 689 |
|
; DCT_8_INV_ROW_1_s [eax+80], [eax+80], tab_i_35_sse, rounder_5 |
| 690 |
|
movq mm4, [tab_i_35_sse+8] ; 4 ; w07 w06 w03 w02 |
| 691 |
|
movq mm6, [tab_i_35_sse+32] ; 6 ; w21 w20 w17 w16 |
| 692 |
|
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
| 693 |
|
paddd mm3, [rounder_5] ; +rounder stall 6 |
| 694 |
|
paddd mm5, [rounder_5] ; +rounder |
| 695 |
|
movq [eax+64+8], mm7 ; 7 ; save y7 y6 y5 y4 |
| 696 |
|
movq mm7, [tab_i_35_sse+40] ; 7 ; w23 w22 w19 w18 |
| 697 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 698 |
|
pmaddwd mm1, [tab_i_35_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 699 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 700 |
|
pmaddwd mm2, [tab_i_35_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 701 |
|
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 702 |
|
pmaddwd mm0, [tab_i_35_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 703 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 704 |
|
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
| 705 |
|
pshufw mm1, [eax+96+8],10001000b ; x6 x4 x6 x4 |
| 706 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 707 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 708 |
|
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 709 |
|
pshufw mm0, [eax+96],10001000b ; x2 x0 x2 x0 |
| 710 |
|
movq mm7, mm5 ; 7 ; a3 a2 |
| 711 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
| 712 |
|
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 713 |
|
movq mm3, [tab_i_26_sse] ; 3 ; w05 w04 w01 w00 |
| 714 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 715 |
|
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
| 716 |
|
pshufw mm2, [eax+96],11011101b ; x3 x1 x3 x1 |
| 717 |
|
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 718 |
|
pmaddwd mm0, [tab_i_26_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 719 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 720 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 721 |
|
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 722 |
|
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 723 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 724 |
|
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
| 725 |
|
pshufw mm5, [eax+96+8],11011101b ; x7 x5 x7 x5 |
| 726 |
|
movq [eax+80], mm6 ; 3 ; save y3 y2 y1 y0 |
| 727 |
|
; DCT_8_INV_ROW_1_s [eax+96], [eax+96], tab_i_26_sse, rounder_6 |
| 728 |
|
movq mm4, [tab_i_26_sse+8] ; 4 ; w07 w06 w03 w02 |
| 729 |
|
movq mm6, [tab_i_26_sse+32] ; 6 ; w21 w20 w17 w16 |
| 730 |
|
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
| 731 |
|
paddd mm3, [rounder_6] ; +rounder |
| 732 |
|
paddd mm0, [rounder_6] ; +rounder |
| 733 |
|
movq [eax+80+8], mm7 ; 7 ; save y7 y6 |
| 734 |
|
movq mm7, [tab_i_26_sse+40] ; 7 ; w23 w22 w19 w18 |
| 735 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 736 |
|
pmaddwd mm1, [tab_i_26_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 737 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 738 |
|
pmaddwd mm2, [tab_i_26_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 739 |
|
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 740 |
|
pmaddwd mm5, [tab_i_26_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 741 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 742 |
|
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
| 743 |
|
pshufw mm1, [eax+112+8],10001000b ; x6 x4 x6 x4 |
| 744 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 745 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 746 |
|
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 747 |
|
pshufw mm5, [eax+112],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
| 748 |
|
movq mm7, mm0 ; 7 ; a3 a2 |
| 749 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
| 750 |
|
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 751 |
|
movq mm3, [tab_i_17_sse] ; 3 ; w05 w04 w01 w00 |
| 752 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 753 |
|
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
| 754 |
|
pshufw mm2, [eax+112],11011101b ; x3 x1 x3 x1 |
| 755 |
|
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 756 |
|
pmaddwd mm5, [tab_i_17_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 757 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 758 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 759 |
|
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 760 |
|
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 761 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 762 |
|
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
| 763 |
|
pshufw mm0, [eax+112+8],11011101b ; x7 x5 x7 x5 |
| 764 |
|
movq [eax+96], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
| 765 |
|
; DCT_8_INV_ROW_1_s [eax+112], [eax+112], tab_i_17_sse, rounder_7 |
| 766 |
|
movq mm4, [tab_i_17_sse+8] ; 4 ; w07 w06 w03 w02 |
| 767 |
|
movq mm6, [tab_i_17_sse+32] ; 6 ; w21 w20 w17 w16 |
| 768 |
|
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
| 769 |
|
paddd mm3, [rounder_7] ; +rounder stall 6 |
| 770 |
|
paddd mm5, [rounder_7] ; +rounder |
| 771 |
|
movq [eax+96+8], mm7 ; 7 ; save y7 y6 y5 y4 |
| 772 |
|
movq mm7, [tab_i_17_sse+40] ; 7 ; w23 w22 w19 w18 |
| 773 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 774 |
|
pmaddwd mm1, [tab_i_17_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 775 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 776 |
|
pmaddwd mm2, [tab_i_17_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 777 |
|
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 778 |
|
pmaddwd mm0, [tab_i_17_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 779 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 780 |
|
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
| 781 |
|
pshufw mm1, [eax+0+8],10001000b ; x6 x4 x6 x4 |
| 782 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 783 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 784 |
|
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 785 |
|
pshufw mm0, [eax+0],10001000b ; x2 x0 x2 x0 |
| 786 |
|
movq mm7, mm5 ; 7 ; a3 a2 |
| 787 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
| 788 |
|
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 789 |
|
movq mm3, [tab_i_04_sse] ; 3 ; w05 w04 w01 w00 |
| 790 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 791 |
|
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
| 792 |
|
pshufw mm2, [eax+0],11011101b ; x3 x1 x3 x1 |
| 793 |
|
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 794 |
|
pmaddwd mm0, [tab_i_04_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 795 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 796 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 797 |
|
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 798 |
|
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 799 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 800 |
|
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
| 801 |
|
pshufw mm5, [eax+0+8],11011101b ; x7 x5 x7 x5 |
| 802 |
|
movq [eax+112], mm6 ; 3 ; save y3 y2 y1 y0 |
| 803 |
|
; DCT_8_INV_ROW_1_s [eax+0], 0, tab_i_04_sse, rounder_0 |
| 804 |
|
movq mm4, [tab_i_04_sse+8] ; 4 ; w07 w06 w03 w02 |
| 805 |
|
movq mm6, [tab_i_04_sse+32] ; 6 ; w21 w20 w17 w16 |
| 806 |
|
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
| 807 |
|
paddd mm3, [rounder_0] ; +rounder |
| 808 |
|
paddd mm0, [rounder_0] ; +rounder |
| 809 |
|
movq [eax+112+8], mm7 ; 7 ; save y7 y6 |
| 810 |
|
movq mm7, [tab_i_04_sse+40] ; 7 ; w23 w22 w19 w18 |
| 811 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 812 |
|
pmaddwd mm1, [tab_i_04_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 813 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 814 |
|
pmaddwd mm2, [tab_i_04_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 815 |
|
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 816 |
|
pmaddwd mm5, [tab_i_04_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 817 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 818 |
|
paddd mm0, mm1 ; 1 |
| 819 |
|
pshufw mm1, [eax+16+8],10001000b ; x6 x4 x6 x4 |
| 820 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 821 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 822 |
|
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 823 |
|
pshufw mm5, [eax+16],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
| 824 |
|
movq mm7, mm0 ; 7 ; a3 a2 |
| 825 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
| 826 |
|
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 827 |
|
movq mm3, [tab_i_17_sse] ; 3 ; w05 w04 w01 w00 |
| 828 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 829 |
|
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
| 830 |
|
pshufw mm2, [eax+16],11011101b ; x3 x1 x3 x1 |
| 831 |
|
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 832 |
|
pmaddwd mm5, [tab_i_17_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 833 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 834 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 835 |
|
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 836 |
|
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 837 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 838 |
|
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
| 839 |
|
pshufw mm0, [eax+16+8],11011101b ; x7 x5 x7 x5 |
| 840 |
|
movq [eax+0], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
| 841 |
|
; DCT_8_INV_ROW_1_s [eax+16], 16, tab_i_17_sse, rounder_1 |
| 842 |
|
movq mm4, [tab_i_17_sse+8] ; 4 ; w07 w06 w03 w02 |
| 843 |
|
movq mm6, [tab_i_17_sse+32] ; 6 ; w21 w20 w17 w16 |
| 844 |
|
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
| 845 |
|
paddd mm3, [rounder_1] ; +rounder stall 6 |
| 846 |
|
paddd mm5, [rounder_1] ; +rounder |
| 847 |
|
movq [eax+0+8], mm7 ; 7 ; save y7 y6 y5 y4 |
| 848 |
|
movq mm7, [tab_i_17_sse+40] ; 7 ; w23 w22 w19 w18 |
| 849 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 850 |
|
pmaddwd mm1, [tab_i_17_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 851 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 852 |
|
pmaddwd mm2, [tab_i_17_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 853 |
|
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 854 |
|
pmaddwd mm0, [tab_i_17_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 855 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 856 |
|
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
| 857 |
|
pshufw mm1, [eax+32+8],10001000b ; x6 x4 x6 x4 |
| 858 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 859 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 860 |
|
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 861 |
|
pshufw mm0, [eax+32],10001000b ; x2 x0 x2 x0 |
| 862 |
|
movq mm7, mm5 ; 7 ; a3 a2 |
| 863 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
| 864 |
|
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 865 |
|
movq mm3, [tab_i_26_sse] ; 3 ; w05 w04 w01 w00 |
| 866 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 867 |
|
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
| 868 |
|
pshufw mm2, [eax+32],11011101b ; x3 x1 x3 x1 |
| 869 |
|
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 870 |
|
pmaddwd mm0, [tab_i_26_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 871 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 872 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 873 |
|
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 874 |
|
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 875 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 876 |
|
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
| 877 |
|
pshufw mm5, [eax+32+8],11011101b ; x7 x5 x7 x5 |
| 878 |
|
movq [eax+16], mm6 ; 3 ; save y3 y2 y1 y0 |
| 879 |
|
; DCT_8_INV_ROW_1_s [eax+32], 32, tab_i_26_sse, rounder_2 |
| 880 |
|
movq mm4, [tab_i_26_sse+8] ; 4 ; w07 w06 w03 w02 |
| 881 |
|
movq mm6, [tab_i_26_sse+32] ; 6 ; w21 w20 w17 w16 |
| 882 |
|
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
| 883 |
|
paddd mm3, [rounder_2] ; +rounder |
| 884 |
|
paddd mm0, [rounder_2] ; +rounder |
| 885 |
|
movq [eax+16+8], mm7 ; 7 ; save y7 y6 |
| 886 |
|
movq mm7, [tab_i_26_sse+40] ; 7 ; w23 w22 w19 w18 |
| 887 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 888 |
|
pmaddwd mm1, [tab_i_26_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 889 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 890 |
|
pmaddwd mm2, [tab_i_26_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 891 |
|
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 892 |
|
pmaddwd mm5, [tab_i_26_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 893 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 894 |
|
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
| 895 |
|
pshufw mm1, [eax+48+8],10001000b ; x6 x4 x6 x4 |
| 896 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 897 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 898 |
|
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 899 |
|
pshufw mm5, [eax+48],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
| 900 |
|
movq mm7, mm0 ; 7 ; a3 a2 |
| 901 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
| 902 |
|
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 903 |
|
movq mm3, [tab_i_35_sse] ; 3 ; w05 w04 w01 w00 |
| 904 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 905 |
|
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
| 906 |
|
pshufw mm2, [eax+48],11011101b ; x3 x1 x3 x1 |
| 907 |
|
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
| 908 |
|
pmaddwd mm5, [tab_i_35_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
| 909 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 910 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 911 |
|
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 912 |
|
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 913 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 914 |
|
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
| 915 |
|
pshufw mm0, [eax+48+8],11011101b ; x7 x5 x7 x5 |
| 916 |
|
movq [eax+32], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
| 917 |
|
; DCT_8_INV_ROW_1_s [eax+48], [eax+48], tab_i_35_sse, rounder_3 |
| 918 |
|
movq mm4, [tab_i_35_sse+8] ; 4 ; w07 w06 w03 w02 |
| 919 |
|
movq mm6, [tab_i_35_sse+32] ; 6 ; w21 w20 w17 w16 |
| 920 |
|
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
| 921 |
|
paddd mm3, [rounder_3] ; +rounder stall 6 |
| 922 |
|
paddd mm5, [rounder_3] ; +rounder |
| 923 |
|
movq [eax+32+8], mm7 ; 7 ; save y7 y6 y5 y4 |
| 924 |
|
movq mm7, [tab_i_35_sse+40] ; 7 ; w23 w22 w19 w18 |
| 925 |
|
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
| 926 |
|
pmaddwd mm1, [tab_i_35_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
| 927 |
|
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
| 928 |
|
pmaddwd mm2, [tab_i_35_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
| 929 |
|
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
| 930 |
|
pmaddwd mm0, [tab_i_35_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
| 931 |
|
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
| 932 |
|
paddd mm5, mm1 ; mm1 free ; a3=sum(even3) a2=sum(even2) |
| 933 |
|
movq mm1, [tg_3_16] |
| 934 |
|
movq mm4, mm3 ; 4 ; a1 a0 |
| 935 |
|
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
| 936 |
|
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
| 937 |
|
movq mm0, [tg_3_16] |
| 938 |
|
movq mm7, mm5 ; 7 ; a3 a2 |
| 939 |
|
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
| 940 |
|
paddd mm3, mm6 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
| 941 |
|
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
| 942 |
|
paddd mm2, mm5 ; 0 free a3+b3 a2+b2 |
| 943 |
|
movq mm5, [eax+16*5] |
| 944 |
|
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
| 945 |
|
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
| 946 |
|
psrad mm3, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
| 947 |
|
psrad mm2, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
| 948 |
|
movq mm6, [eax+16*1] |
| 949 |
|
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
| 950 |
|
movq mm4, [tg_1_16] |
| 951 |
|
packssdw mm3, mm2 ; 0 free ; y3 y2 y1 y0 |
| 952 |
|
pshufw mm2, mm7, 10110001b ; y7 y6 y5 y4 |
| 953 |
|
; DCT_8_INV_COL_4 [eax+0],[eax+0] |
| 954 |
|
; movq mm3,mmword ptr [eax+16*3] |
| 955 |
|
movq mm7, [eax+16*7] |
| 956 |
|
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
| 957 |
|
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
| 958 |
|
movq [eax+48+8], mm2 ; 7 ; save y7 y6 y5 y4 |
| 959 |
|
movq mm2, mm4 ; tg_1_16 |
| 960 |
|
pmulhw mm4, mm7 ; x7*tg_1_16 |
| 961 |
|
paddsw mm0, mm3 ; x3*tg_3_16 |
| 962 |
|
pmulhw mm2, mm6 ; x1*tg_1_16 |
| 963 |
|
paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) |
| 964 |
|
psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 |
| 965 |
|
movq [eax+48], mm3 ; 3 ; save y3 y2 y1 y0 |
| 966 |
|
movq mm3, [ocos_4_16] |
| 967 |
|
paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
| 968 |
|
paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 |
| 969 |
|
psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 |
| 970 |
|
movq mm5, mm4 ; tp17 |
| 971 |
|
movq mm6, mm2 ; tm17 |
| 972 |
|
paddsw mm5, mm1 ; tp17+tp35 = b0 |
| 973 |
|
psubsw mm6, mm0 ; tm17-tm35 = b3 |
| 974 |
|
psubsw mm4, mm1 ; tp17-tp35 = t1 |
| 975 |
|
paddsw mm2, mm0 ; tm17+tm35 = t2 |
| 976 |
|
movq mm7, [tg_2_16] |
| 977 |
|
movq mm1, mm4 ; t1 |
| 978 |
|
movq [eax+3*16], mm5 ; save b0 |
| 979 |
|
paddsw mm1, mm2 ; t1+t2 |
| 980 |
|
movq [eax+5*16], mm6 ; save b3 |
| 981 |
|
psubsw mm4, mm2 ; t1-t2 |
| 982 |
|
movq mm5, [eax+2*16] |
| 983 |
|
movq mm0, mm7 ; tg_2_16 |
| 984 |
|
movq mm6, [eax+6*16] |
| 985 |
|
pmulhw mm0, mm5 ; x2*tg_2_16 |
| 986 |
|
pmulhw mm7, mm6 ; x6*tg_2_16 |
| 987 |
|
; slot |
| 988 |
|
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
| 989 |
|
; slot |
| 990 |
|
movq mm2, [eax+0*16] |
| 991 |
|
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
| 992 |
|
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
| 993 |
|
movq mm3, [eax+0*16] ; x0 |
| 994 |
|
movq mm6, [eax+4*16] |
| 995 |
|
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
| 996 |
|
paddsw mm2, mm6 ; x0+x4 = tp04 |
| 997 |
|
psubsw mm3, mm6 ; x0-x4 = tm04 |
| 998 |
|
movq mm5, mm2 ; tp04 |
| 999 |
|
movq mm6, mm3 ; tm04 |
| 1000 |
|
psubsw mm2, mm7 ; tp04-tp26 = a3 |
| 1001 |
|
paddsw mm3, mm0 ; tm04+tm26 = a1 |
| 1002 |
|
paddsw mm1, mm1 ; b1 |
| 1003 |
|
paddsw mm4, mm4 ; b2 |
| 1004 |
|
paddsw mm5, mm7 ; tp04+tp26 = a0 |
| 1005 |
|
psubsw mm6, mm0 ; tm04-tm26 = a2 |
| 1006 |
|
movq mm7, mm3 ; a1 |
| 1007 |
|
movq mm0, mm6 ; a2 |
| 1008 |
|
paddsw mm3, mm1 ; a1+b1 |
| 1009 |
|
paddsw mm6, mm4 ; a2+b2 |
| 1010 |
|
psraw mm3, SHIFT_INV_COL ; dst1 |
| 1011 |
|
psubsw mm7, mm1 ; a1-b1 |
| 1012 |
|
psraw mm6, SHIFT_INV_COL ; dst2 |
| 1013 |
|
psubsw mm0, mm4 ; a2-b2 |
| 1014 |
|
movq mm1, [eax+3*16] ; load b0 |
| 1015 |
|
psraw mm7, SHIFT_INV_COL ; dst6 |
| 1016 |
|
movq mm4, mm5 ; a0 |
| 1017 |
|
psraw mm0, SHIFT_INV_COL ; dst5 |
| 1018 |
|
movq [eax+1*16], mm3 |
| 1019 |
|
paddsw mm5, mm1 ; a0+b0 |
| 1020 |
|
movq [eax+2*16], mm6 |
| 1021 |
|
psubsw mm4, mm1 ; a0-b0 |
| 1022 |
|
movq mm3, [eax+5*16] ; load b3 |
| 1023 |
|
psraw mm5, SHIFT_INV_COL ; dst0 |
| 1024 |
|
movq mm6, mm2 ; a3 |
| 1025 |
|
psraw mm4, SHIFT_INV_COL ; dst7 |
| 1026 |
|
movq [eax+5*16], mm0 |
| 1027 |
|
movq mm0, [tg_3_16] |
| 1028 |
|
paddsw mm2, mm3 ; a3+b3 |
| 1029 |
|
movq [eax+6*16], mm7 |
| 1030 |
|
psubsw mm6, mm3 ; a3-b3 |
| 1031 |
|
movq mm3, [eax+8+16*3] |
| 1032 |
|
movq [eax+0*16], mm5 |
| 1033 |
|
psraw mm2, SHIFT_INV_COL ; dst3 |
| 1034 |
|
movq [eax+7*16], mm4 |
| 1035 |
|
; DCT_8_INV_COL_4 [eax+8],[eax+8] |
| 1036 |
|
|
| 1037 |
|
movq mm1, mm0 ; tg_3_16 |
| 1038 |
|
movq mm5, [eax+8+16*5] |
| 1039 |
|
psraw mm6, SHIFT_INV_COL ; dst4 |
| 1040 |
|
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
| 1041 |
|
movq mm4, [tg_1_16] |
| 1042 |
|
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
| 1043 |
|
movq mm7, [eax+8+16*7] |
| 1044 |
|
movq [eax+3*16], mm2 |
| 1045 |
|
movq mm2, mm4 ; tg_1_16 |
| 1046 |
|
movq [eax+4*16], mm6 |
| 1047 |
|
movq mm6, [eax+8+16*1] |
| 1048 |
|
pmulhw mm4, mm7 ; x7*tg_1_16 |
| 1049 |
|
paddsw mm0, mm3 ; x3*tg_3_16 |
| 1050 |
|
pmulhw mm2, mm6 ; x1*tg_1_16 |
| 1051 |
|
paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) |
| 1052 |
|
psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 |
| 1053 |
|
movq mm3, [ocos_4_16] |
| 1054 |
|
paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
| 1055 |
|
paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 |
| 1056 |
|
psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 |
| 1057 |
|
movq mm5, mm4 ; tp17 |
| 1058 |
|
movq mm6, mm2 ; tm17 |
| 1059 |
|
paddsw mm5, mm1 ; tp17+tp35 = b0 |
| 1060 |
|
psubsw mm4, mm1 ; tp17-tp35 = t1 |
| 1061 |
|
paddsw mm2, mm0 ; tm17+tm35 = t2 |
| 1062 |
|
movq mm7, [tg_2_16] |
| 1063 |
|
movq mm1, mm4 ; t1 |
| 1064 |
|
psubsw mm6, mm0 ; tm17-tm35 = b3 |
| 1065 |
|
movq [eax+8+3*16], mm5 ; save b0 |
| 1066 |
|
movq [eax+8+5*16], mm6 ; save b3 |
| 1067 |
|
psubsw mm4, mm2 ; t1-t2 |
| 1068 |
|
movq mm5, [eax+8+2*16] |
| 1069 |
|
movq mm0, mm7 ; tg_2_16 |
| 1070 |
|
movq mm6, [eax+8+6*16] |
| 1071 |
|
paddsw mm1, mm2 ; t1+t2 |
| 1072 |
|
pmulhw mm0, mm5 ; x2*tg_2_16 |
| 1073 |
|
pmulhw mm7, mm6 ; x6*tg_2_16 |
| 1074 |
|
movq mm2, [eax+8+0*16] |
| 1075 |
|
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
| 1076 |
|
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
| 1077 |
|
; slot |
| 1078 |
|
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
| 1079 |
|
; slot |
| 1080 |
|
movq mm3, [eax+8+0*16] ; x0 |
| 1081 |
|
movq mm6, [eax+8+4*16] |
| 1082 |
|
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
| 1083 |
|
paddsw mm2, mm6 ; x0+x4 = tp04 |
| 1084 |
|
psubsw mm3, mm6 ; x0-x4 = tm04 |
| 1085 |
|
movq mm5, mm2 ; tp04 |
| 1086 |
|
movq mm6, mm3 ; tm04 |
| 1087 |
|
psubsw mm2, mm7 ; tp04-tp26 = a3 |
| 1088 |
|
paddsw mm3, mm0 ; tm04+tm26 = a1 |
| 1089 |
|
paddsw mm1, mm1 ; b1 |
| 1090 |
|
paddsw mm4, mm4 ; b2 |
| 1091 |
|
paddsw mm5, mm7 ; tp04+tp26 = a0 |
| 1092 |
|
psubsw mm6, mm0 ; tm04-tm26 = a2 |
| 1093 |
|
movq mm7, mm3 ; a1 |
| 1094 |
|
movq mm0, mm6 ; a2 |
| 1095 |
|
paddsw mm3, mm1 ; a1+b1 |
| 1096 |
|
paddsw mm6, mm4 ; a2+b2 |
| 1097 |
|
psraw mm3, SHIFT_INV_COL ; dst1 |
| 1098 |
|
psubsw mm7, mm1 ; a1-b1 |
| 1099 |
|
psraw mm6, SHIFT_INV_COL ; dst2 |
| 1100 |
|
psubsw mm0, mm4 ; a2-b2 |
| 1101 |
|
movq mm1, [eax+8+3*16] ; load b0 |
| 1102 |
|
psraw mm7, SHIFT_INV_COL ; dst6 |
| 1103 |
|
movq mm4, mm5 ; a0 |
| 1104 |
|
psraw mm0, SHIFT_INV_COL ; dst5 |
| 1105 |
|
movq [eax+8+1*16], mm3 |
| 1106 |
|
paddsw mm5, mm1 ; a0+b0 |
| 1107 |
|
movq [eax+8+2*16], mm6 |
| 1108 |
|
psubsw mm4, mm1 ; a0-b0 |
| 1109 |
|
movq mm3, [eax+8+5*16] ; load b3 |
| 1110 |
|
psraw mm5, SHIFT_INV_COL ; dst0 |
| 1111 |
|
movq mm6, mm2 ; a3 |
| 1112 |
|
psraw mm4, SHIFT_INV_COL ; dst7 |
| 1113 |
|
movq [eax+8+5*16], mm0 |
| 1114 |
|
paddsw mm2, mm3 ; a3+b3 |
| 1115 |
|
movq [eax+8+6*16], mm7 |
| 1116 |
|
psubsw mm6, mm3 ; a3-b3 |
| 1117 |
|
movq [eax+8+0*16], mm5 |
| 1118 |
|
psraw mm2, SHIFT_INV_COL ; dst3 |
| 1119 |
|
movq [eax+8+7*16], mm4 |
| 1120 |
|
psraw mm6, SHIFT_INV_COL ; dst4 |
| 1121 |
|
movq [eax+8+3*16], mm2 |
| 1122 |
|
movq [eax+8+4*16], mm6 |
| 1123 |
|
ret |
| 1124 |
|
|
Parent Directory
| 
Revision Log
| 
Patch