me_cmp.asm 12.3 KB
Newer Older
1
;*****************************************************************************
2
;* SIMD-optimized motion compensation estimation
3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
;*****************************************************************************
;* Copyright (c) 2000, 2001 Fabrice Bellard
;* Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
;*
;* This file is part of FFmpeg.
;*
;* FFmpeg is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* FFmpeg is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with FFmpeg; if not, write to the Free Software
21
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 23
;*****************************************************************************

24
%include "libavutil/x86/x86util.asm"
25 26 27

SECTION .text

28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
%macro DIFF_PIXELS_1 4
    movh            %1, %3
    movh            %2, %4
    punpcklbw       %2, %1
    punpcklbw       %1, %1
    psubw           %1, %2
%endmacro

; %1=uint8_t *pix1, %2=uint8_t *pix2, %3=static offset, %4=stride, %5=stride*3
; %6=temporary storage location
; this macro requires $mmsize stack space (aligned) on %6 (except on SSE+x86-64)
%macro DIFF_PIXELS_8 6
    DIFF_PIXELS_1   m0, m7, [%1     +%3], [%2     +%3]
    DIFF_PIXELS_1   m1, m7, [%1+%4  +%3], [%2+%4  +%3]
    DIFF_PIXELS_1   m2, m7, [%1+%4*2+%3], [%2+%4*2+%3]
    add             %1, %5
    add             %2, %5
    DIFF_PIXELS_1   m3, m7, [%1     +%3], [%2     +%3]
    DIFF_PIXELS_1   m4, m7, [%1+%4  +%3], [%2+%4  +%3]
    DIFF_PIXELS_1   m5, m7, [%1+%4*2+%3], [%2+%4*2+%3]
    DIFF_PIXELS_1   m6, m7, [%1+%5  +%3], [%2+%5  +%3]
%ifdef m8
    DIFF_PIXELS_1   m7, m8, [%1+%4*4+%3], [%2+%4*4+%3]
%else
    mova          [%6], m0
    DIFF_PIXELS_1   m7, m0, [%1+%4*4+%3], [%2+%4*4+%3]
    mova            m0, [%6]
%endif
    sub             %1, %5
    sub             %2, %5
%endmacro

%macro HADAMARD8 0
61 62 63 64 65 66
    SUMSUB_BADC       w, 0, 1, 2, 3
    SUMSUB_BADC       w, 4, 5, 6, 7
    SUMSUB_BADC       w, 0, 2, 1, 3
    SUMSUB_BADC       w, 4, 6, 5, 7
    SUMSUB_BADC       w, 0, 4, 1, 5
    SUMSUB_BADC       w, 2, 6, 3, 7
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101
%endmacro

%macro ABS1_SUM 3
    ABS1            %1, %2
    paddusw         %3, %1
%endmacro

%macro ABS2_SUM 6
    ABS2            %1, %2, %3, %4
    paddusw         %5, %1
    paddusw         %6, %2
%endmacro

%macro ABS_SUM_8x8_64 1
    ABS2            m0, m1, m8, m9
    ABS2_SUM        m2, m3, m8, m9, m0, m1
    ABS2_SUM        m4, m5, m8, m9, m0, m1
    ABS2_SUM        m6, m7, m8, m9, m0, m1
    paddusw         m0, m1
%endmacro

%macro ABS_SUM_8x8_32 1
    mova          [%1], m7
    ABS1            m0, m7
    ABS1            m1, m7
    ABS1_SUM        m2, m7, m0
    ABS1_SUM        m3, m7, m1
    ABS1_SUM        m4, m7, m0
    ABS1_SUM        m5, m7, m1
    ABS1_SUM        m6, m7, m0
    mova            m2, [%1]
    ABS1_SUM        m2, m7, m1
    paddusw         m0, m1
%endmacro

102
; FIXME: HSUM saturates at 64k, while an 8x8 hadamard or dct block can get up to
103 104
; about 100k on extreme inputs. But that's very unlikely to occur in natural video,
; and it's even more unlikely to not have any alternative mvs/modes with lower cost.
105 106 107
%macro HSUM 3
%if cpuflag(sse2)
    movhlps         %2, %1
108
    paddusw         %1, %2
109 110 111
    pshuflw         %2, %1, 0xE
    paddusw         %1, %2
    pshuflw         %2, %1, 0x1
112 113
    paddusw         %1, %2
    movd            %3, %1
114
%elif cpuflag(mmxext)
115 116 117 118 119
    pshufw          %2, %1, 0xE
    paddusw         %1, %2
    pshufw          %2, %1, 0x1
    paddusw         %1, %2
    movd            %3, %1
120 121 122
%elif cpuflag(mmx)
    mova            %2, %1
    psrlq           %1, 32
123
    paddusw         %1, %2
124 125
    mova            %2, %1
    psrlq           %1, 16
126 127
    paddusw         %1, %2
    movd            %3, %1
128
%endif
129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144
%endmacro

%macro STORE4 5
    mova [%1+mmsize*0], %2
    mova [%1+mmsize*1], %3
    mova [%1+mmsize*2], %4
    mova [%1+mmsize*3], %5
%endmacro

%macro LOAD4 5
    mova            %2, [%1+mmsize*0]
    mova            %3, [%1+mmsize*1]
    mova            %4, [%1+mmsize*2]
    mova            %5, [%1+mmsize*3]
%endmacro

145 146
%macro hadamard8_16_wrapper 2
cglobal hadamard8_diff, 4, 4, %1
147
%ifndef m8
148
    %assign pad %2*mmsize-(4+stack_offset&(mmsize-1))
149 150
    SUB            rsp, pad
%endif
151
    call hadamard8x8_diff %+ SUFFIX
152 153 154 155 156
%ifndef m8
    ADD            rsp, pad
%endif
    RET

157
cglobal hadamard8_diff16, 5, 6, %1
158
%ifndef m8
159
    %assign pad %2*mmsize-(4+stack_offset&(mmsize-1))
160 161 162
    SUB            rsp, pad
%endif

163
    call hadamard8x8_diff %+ SUFFIX
164 165 166 167
    mov            r5d, eax

    add             r1, 8
    add             r2, 8
168
    call hadamard8x8_diff %+ SUFFIX
169 170 171 172 173 174 175
    add            r5d, eax

    cmp            r4d, 16
    jne .done

    lea             r1, [r1+r3*8-8]
    lea             r2, [r2+r3*8-8]
176
    call hadamard8x8_diff %+ SUFFIX
177 178 179 180
    add            r5d, eax

    add             r1, 8
    add             r2, 8
181
    call hadamard8x8_diff %+ SUFFIX
182 183
    add            r5d, eax

184
.done:
185 186 187 188 189 190 191
    mov            eax, r5d
%ifndef m8
    ADD            rsp, pad
%endif
    RET
%endmacro

192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210
%macro HADAMARD8_DIFF 0-1
%if cpuflag(sse2)
hadamard8x8_diff %+ SUFFIX:
    lea                          r0, [r3*3]
    DIFF_PIXELS_8                r1, r2,  0, r3, r0, rsp+gprsize
    HADAMARD8
%if ARCH_X86_64
    TRANSPOSE8x8W                 0,  1,  2,  3,  4,  5,  6,  7,  8
%else
    TRANSPOSE8x8W                 0,  1,  2,  3,  4,  5,  6,  7, [rsp+gprsize], [rsp+mmsize+gprsize]
%endif
    HADAMARD8
    ABS_SUM_8x8         rsp+gprsize
    HSUM                        m0, m1, eax
    and                         eax, 0xFFFF
    ret

hadamard8_16_wrapper %1, 3
%elif cpuflag(mmx)
211
ALIGN 16
212 213
; int ff_hadamard8_diff_ ## cpu(MpegEncContext *s, uint8_t *src1,
;                               uint8_t *src2, int stride, int h)
214 215 216 217
; r0 = void *s = unused, int h = unused (always 8)
; note how r1, r2 and r3 are not clobbered in this function, so 16x16
; can simply call this 2x2x (and that's why we access rsp+gprsize
; everywhere, which is rsp of calling func
218
hadamard8x8_diff %+ SUFFIX:
219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254
    lea                          r0, [r3*3]

    ; first 4x8 pixels
    DIFF_PIXELS_8                r1, r2,  0, r3, r0, rsp+gprsize+0x60
    HADAMARD8
    mova         [rsp+gprsize+0x60], m7
    TRANSPOSE4x4W                 0,  1,  2,  3,  7
    STORE4              rsp+gprsize, m0, m1, m2, m3
    mova                         m7, [rsp+gprsize+0x60]
    TRANSPOSE4x4W                 4,  5,  6,  7,  0
    STORE4         rsp+gprsize+0x40, m4, m5, m6, m7

    ; second 4x8 pixels
    DIFF_PIXELS_8                r1, r2,  4, r3, r0, rsp+gprsize+0x60
    HADAMARD8
    mova         [rsp+gprsize+0x60], m7
    TRANSPOSE4x4W                 0,  1,  2,  3,  7
    STORE4         rsp+gprsize+0x20, m0, m1, m2, m3
    mova                         m7, [rsp+gprsize+0x60]
    TRANSPOSE4x4W                 4,  5,  6,  7,  0

    LOAD4          rsp+gprsize+0x40, m0, m1, m2, m3
    HADAMARD8
    ABS_SUM_8x8_32 rsp+gprsize+0x60
    mova         [rsp+gprsize+0x60], m0

    LOAD4          rsp+gprsize     , m0, m1, m2, m3
    LOAD4          rsp+gprsize+0x20, m4, m5, m6, m7
    HADAMARD8
    ABS_SUM_8x8_32 rsp+gprsize
    paddusw                      m0, [rsp+gprsize+0x60]

    HSUM                         m0, m1, eax
    and                         rax, 0xFFFF
    ret

255
hadamard8_16_wrapper 0, 14
256 257 258
%endif
%endmacro

259 260
INIT_MMX mmx
HADAMARD8_DIFF
261

262 263
INIT_MMX mmxext
HADAMARD8_DIFF
264

265
INIT_XMM sse2
266
%if ARCH_X86_64
267 268 269 270
%define ABS_SUM_8x8 ABS_SUM_8x8_64
%else
%define ABS_SUM_8x8 ABS_SUM_8x8_32
%endif
271
HADAMARD8_DIFF 10
272

273
INIT_XMM ssse3
274
%define ABS_SUM_8x8 ABS_SUM_8x8_64
275
HADAMARD8_DIFF 9
276

277 278 279 280 281 282 283 284
; int ff_sse*_*(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
;               int line_size, int h)

%macro SUM_SQUARED_ERRORS 1
cglobal sse%1, 5,5,8, v, pix1, pix2, lsize, h
%if %1 == mmsize
    shr       hd, 1
%endif
285 286 287
    pxor      m0, m0         ; mm0 = 0
    pxor      m7, m7         ; mm7 holds the sum

288
.next2lines: ; FIXME why are these unaligned movs? pix1[] is aligned
289 290 291 292 293 294 295 296 297
    movu      m1, [pix1q]    ; m1 = pix1[0][0-15], [0-7] for mmx
    movu      m2, [pix2q]    ; m2 = pix2[0][0-15], [0-7] for mmx
%if %1 == mmsize
    movu      m3, [pix1q+lsizeq] ; m3 = pix1[1][0-15], [0-7] for mmx
    movu      m4, [pix2q+lsizeq] ; m4 = pix2[1][0-15], [0-7] for mmx
%else  ; %1 / 2 == mmsize; mmx only
    mova      m3, [pix1q+8]  ; m3 = pix1[0][8-15]
    mova      m4, [pix2q+8]  ; m4 = pix2[0][8-15]
%endif
298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330

    ; todo: mm1-mm2, mm3-mm4
    ; algo: subtract mm1 from mm2 with saturation and vice versa
    ;       OR the result to get the absolute difference
    mova      m5, m1
    mova      m6, m3
    psubusb   m1, m2
    psubusb   m3, m4
    psubusb   m2, m5
    psubusb   m4, m6

    por       m2, m1
    por       m4, m3

    ; now convert to 16-bit vectors so we can square them
    mova      m1, m2
    mova      m3, m4

    punpckhbw m2, m0
    punpckhbw m4, m0
    punpcklbw m1, m0         ; mm1 not spread over (mm1,mm2)
    punpcklbw m3, m0         ; mm4 not spread over (mm3,mm4)

    pmaddwd   m2, m2
    pmaddwd   m4, m4
    pmaddwd   m1, m1
    pmaddwd   m3, m3

    paddd     m1, m2
    paddd     m3, m4
    paddd     m7, m1
    paddd     m7, m3

331 332 333 334 335 336 337 338
%if %1 == mmsize
    lea    pix1q, [pix1q + 2*lsizeq]
    lea    pix2q, [pix2q + 2*lsizeq]
%else
    add    pix1q, lsizeq
    add    pix2q, lsizeq
%endif
    dec       hd
339 340
    jnz .next2lines

341
    HADDD     m7, m1
342 343
    movd     eax, m7         ; return value
    RET
344 345 346 347 348 349 350 351 352 353
%endmacro

INIT_MMX mmx
SUM_SQUARED_ERRORS 8

INIT_MMX mmx
SUM_SQUARED_ERRORS 16

INIT_XMM sse2
SUM_SQUARED_ERRORS 16
354

355 356 357 358
;-----------------------------------------------
;int ff_sum_abs_dctelem(int16_t *block)
;-----------------------------------------------
; %1 = number of xmm registers used
359
; %2 = number of inline loops
360

361
%macro SUM_ABS_DCTELEM 2
362 363 364
cglobal sum_abs_dctelem, 1, 1, %1, block
    pxor    m0, m0
    pxor    m1, m1
365 366 367 368 369 370 371 372 373 374 375 376
%assign %%i 0
%rep %2
    mova      m2, [blockq+mmsize*(0+%%i)]
    mova      m3, [blockq+mmsize*(1+%%i)]
    mova      m4, [blockq+mmsize*(2+%%i)]
    mova      m5, [blockq+mmsize*(3+%%i)]
    ABS1_SUM  m2, m6, m0
    ABS1_SUM  m3, m6, m1
    ABS1_SUM  m4, m6, m0
    ABS1_SUM  m5, m6, m1
%assign %%i %%i+4
%endrep
377 378 379 380 381 382 383
    paddusw m0, m1
    HSUM    m0, m1, eax
    and     eax, 0xFFFF
    RET
%endmacro

INIT_MMX mmx
384
SUM_ABS_DCTELEM 0, 4
385
INIT_MMX mmxext
386
SUM_ABS_DCTELEM 0, 4
387
INIT_XMM sse2
388
SUM_ABS_DCTELEM 7, 2
389
INIT_XMM ssse3
390
SUM_ABS_DCTELEM 6, 2
391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467

;------------------------------------------------------------------------------
; int ff_hf_noise*_mmx(uint8_t *pix1, int lsize, int h)
;------------------------------------------------------------------------------
; %1 = 8/16. %2-5=m#
%macro HF_NOISE_PART1 5
    mova      m%2, [pix1q]
%if %1 == 8
    mova      m%3, m%2
    psllq     m%2, 8
    psrlq     m%3, 8
    psrlq     m%2, 8
%else
    mova      m%3, [pix1q+1]
%endif
    mova      m%4, m%2
    mova      m%5, m%3
    punpcklbw m%2, m7
    punpcklbw m%3, m7
    punpckhbw m%4, m7
    punpckhbw m%5, m7
    psubw     m%2, m%3
    psubw     m%4, m%5
%endmacro

; %1-2 = m#
%macro HF_NOISE_PART2 4
    psubw     m%1, m%3
    psubw     m%2, m%4
    pxor       m3, m3
    pxor       m1, m1
    pcmpgtw    m3, m%1
    pcmpgtw    m1, m%2
    pxor      m%1, m3
    pxor      m%2, m1
    psubw     m%1, m3
    psubw     m%2, m1
    paddw     m%2, m%1
    paddw      m6, m%2
%endmacro

; %1 = 8/16
%macro HF_NOISE 1
cglobal hf_noise%1, 3,3,0, pix1, lsize, h
    movsxdifnidn lsizeq, lsized
    sub        hd, 2
    pxor       m7, m7
    pxor       m6, m6
    HF_NOISE_PART1 %1, 0, 1, 2, 3
    add     pix1q, lsizeq
    HF_NOISE_PART1 %1, 4, 1, 5, 3
    HF_NOISE_PART2     0, 2, 4, 5
    add     pix1q, lsizeq
.loop:
    HF_NOISE_PART1 %1, 0, 1, 2, 3
    HF_NOISE_PART2     4, 5, 0, 2
    add     pix1q, lsizeq
    HF_NOISE_PART1 %1, 4, 1, 5, 3
    HF_NOISE_PART2     0, 2, 4, 5
    add     pix1q, lsizeq
    sub        hd, 2
        jne .loop

    mova       m0, m6
    punpcklwd  m0, m7
    punpckhwd  m6, m7
    paddd      m6, m0
    mova       m0, m6
    psrlq      m6, 32
    paddd      m0, m6
    movd      eax, m0   ; eax = result of hf_noise8;
    REP_RET                 ; return eax;
%endmacro

INIT_MMX mmx
HF_NOISE 8
HF_NOISE 16