1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/encoder/x86/dct_sse2.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,432 @@
1.4 +;
1.5 +; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
1.6 +;
1.7 +; Use of this source code is governed by a BSD-style license
1.8 +; that can be found in the LICENSE file in the root of the source
1.9 +; tree. An additional intellectual property rights grant can be found
1.10 +; in the file PATENTS. All contributing project authors may
1.11 +; be found in the AUTHORS file in the root of the source tree.
1.12 +;
1.13 +
1.14 +
1.15 +%include "vpx_ports/x86_abi_support.asm"
1.16 +
1.17 +%macro STACK_FRAME_CREATE 0
1.18 +%if ABI_IS_32BIT
1.19 + %define input rsi
1.20 + %define output rdi
1.21 + %define pitch rax
1.22 + push rbp
1.23 + mov rbp, rsp
1.24 + GET_GOT rbx
1.25 + push rsi
1.26 + push rdi
1.27 + ; end prolog
1.28 +
1.29 + mov rsi, arg(0)
1.30 + mov rdi, arg(1)
1.31 +
1.32 + movsxd rax, dword ptr arg(2)
1.33 + lea rcx, [rsi + rax*2]
1.34 +%else
1.35 + %if LIBVPX_YASM_WIN64
1.36 + %define input rcx
1.37 + %define output rdx
1.38 + %define pitch r8
1.39 + SAVE_XMM 7, u
1.40 + %else
1.41 + %define input rdi
1.42 + %define output rsi
1.43 + %define pitch rdx
1.44 + %endif
1.45 +%endif
1.46 +%endmacro
1.47 +
1.48 +%macro STACK_FRAME_DESTROY 0
1.49 + %define input
1.50 + %define output
1.51 + %define pitch
1.52 +
1.53 +%if ABI_IS_32BIT
1.54 + pop rdi
1.55 + pop rsi
1.56 + RESTORE_GOT
1.57 + pop rbp
1.58 +%else
1.59 + %if LIBVPX_YASM_WIN64
1.60 + RESTORE_XMM
1.61 + %endif
1.62 +%endif
1.63 + ret
1.64 +%endmacro
1.65 +
1.66 +;void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch)
1.67 +global sym(vp8_short_fdct4x4_sse2) PRIVATE
1.68 +sym(vp8_short_fdct4x4_sse2):
1.69 +
1.70 + STACK_FRAME_CREATE
1.71 +
1.72 + movq xmm0, MMWORD PTR[input ] ;03 02 01 00
1.73 + movq xmm2, MMWORD PTR[input+ pitch] ;13 12 11 10
1.74 + lea input, [input+2*pitch]
1.75 + movq xmm1, MMWORD PTR[input ] ;23 22 21 20
1.76 + movq xmm3, MMWORD PTR[input+ pitch] ;33 32 31 30
1.77 +
1.78 + punpcklqdq xmm0, xmm2 ;13 12 11 10 03 02 01 00
1.79 + punpcklqdq xmm1, xmm3 ;33 32 31 30 23 22 21 20
1.80 +
1.81 + movdqa xmm2, xmm0
1.82 + punpckldq xmm0, xmm1 ;23 22 03 02 21 20 01 00
1.83 + punpckhdq xmm2, xmm1 ;33 32 13 12 31 30 11 10
1.84 + movdqa xmm1, xmm0
1.85 + punpckldq xmm0, xmm2 ;31 21 30 20 11 10 01 00
1.86 + pshufhw xmm1, xmm1, 0b1h ;22 23 02 03 xx xx xx xx
1.87 + pshufhw xmm2, xmm2, 0b1h ;32 33 12 13 xx xx xx xx
1.88 +
1.89 + punpckhdq xmm1, xmm2 ;32 33 22 23 12 13 02 03
1.90 + movdqa xmm3, xmm0
1.91 + paddw xmm0, xmm1 ;b1 a1 b1 a1 b1 a1 b1 a1
1.92 + psubw xmm3, xmm1 ;c1 d1 c1 d1 c1 d1 c1 d1
1.93 + psllw xmm0, 3 ;b1 <<= 3 a1 <<= 3
1.94 + psllw xmm3, 3 ;c1 <<= 3 d1 <<= 3
1.95 +
1.96 + movdqa xmm1, xmm0
1.97 + pmaddwd xmm0, XMMWORD PTR[GLOBAL(_mult_add)] ;a1 + b1
1.98 + pmaddwd xmm1, XMMWORD PTR[GLOBAL(_mult_sub)] ;a1 - b1
1.99 + movdqa xmm4, xmm3
1.100 + pmaddwd xmm3, XMMWORD PTR[GLOBAL(_5352_2217)] ;c1*2217 + d1*5352
1.101 + pmaddwd xmm4, XMMWORD PTR[GLOBAL(_2217_neg5352)];d1*2217 - c1*5352
1.102 +
1.103 + paddd xmm3, XMMWORD PTR[GLOBAL(_14500)]
1.104 + paddd xmm4, XMMWORD PTR[GLOBAL(_7500)]
1.105 + psrad xmm3, 12 ;(c1 * 2217 + d1 * 5352 + 14500)>>12
1.106 + psrad xmm4, 12 ;(d1 * 2217 - c1 * 5352 + 7500)>>12
1.107 +
1.108 + packssdw xmm0, xmm1 ;op[2] op[0]
1.109 + packssdw xmm3, xmm4 ;op[3] op[1]
1.110 + ; 23 22 21 20 03 02 01 00
1.111 + ;
1.112 + ; 33 32 31 30 13 12 11 10
1.113 + ;
1.114 + movdqa xmm2, xmm0
1.115 + punpcklqdq xmm0, xmm3 ;13 12 11 10 03 02 01 00
1.116 + punpckhqdq xmm2, xmm3 ;23 22 21 20 33 32 31 30
1.117 +
1.118 + movdqa xmm3, xmm0
1.119 + punpcklwd xmm0, xmm2 ;32 30 22 20 12 10 02 00
1.120 + punpckhwd xmm3, xmm2 ;33 31 23 21 13 11 03 01
1.121 + movdqa xmm2, xmm0
1.122 + punpcklwd xmm0, xmm3 ;13 12 11 10 03 02 01 00
1.123 + punpckhwd xmm2, xmm3 ;33 32 31 30 23 22 21 20
1.124 +
1.125 + movdqa xmm5, XMMWORD PTR[GLOBAL(_7)]
1.126 + pshufd xmm2, xmm2, 04eh
1.127 + movdqa xmm3, xmm0
1.128 + paddw xmm0, xmm2 ;b1 b1 b1 b1 a1 a1 a1 a1
1.129 + psubw xmm3, xmm2 ;c1 c1 c1 c1 d1 d1 d1 d1
1.130 +
1.131 + pshufd xmm0, xmm0, 0d8h ;b1 b1 a1 a1 b1 b1 a1 a1
1.132 + movdqa xmm2, xmm3 ;save d1 for compare
1.133 + pshufd xmm3, xmm3, 0d8h ;c1 c1 d1 d1 c1 c1 d1 d1
1.134 + pshuflw xmm0, xmm0, 0d8h ;b1 b1 a1 a1 b1 a1 b1 a1
1.135 + pshuflw xmm3, xmm3, 0d8h ;c1 c1 d1 d1 c1 d1 c1 d1
1.136 + pshufhw xmm0, xmm0, 0d8h ;b1 a1 b1 a1 b1 a1 b1 a1
1.137 + pshufhw xmm3, xmm3, 0d8h ;c1 d1 c1 d1 c1 d1 c1 d1
1.138 + movdqa xmm1, xmm0
1.139 + pmaddwd xmm0, XMMWORD PTR[GLOBAL(_mult_add)] ;a1 + b1
1.140 + pmaddwd xmm1, XMMWORD PTR[GLOBAL(_mult_sub)] ;a1 - b1
1.141 +
1.142 + pxor xmm4, xmm4 ;zero out for compare
1.143 + paddd xmm0, xmm5
1.144 + paddd xmm1, xmm5
1.145 + pcmpeqw xmm2, xmm4
1.146 + psrad xmm0, 4 ;(a1 + b1 + 7)>>4
1.147 + psrad xmm1, 4 ;(a1 - b1 + 7)>>4
1.148 + pandn xmm2, XMMWORD PTR[GLOBAL(_cmp_mask)] ;clear upper,
1.149 + ;and keep bit 0 of lower
1.150 +
1.151 + movdqa xmm4, xmm3
1.152 + pmaddwd xmm3, XMMWORD PTR[GLOBAL(_5352_2217)] ;c1*2217 + d1*5352
1.153 + pmaddwd xmm4, XMMWORD PTR[GLOBAL(_2217_neg5352)] ;d1*2217 - c1*5352
1.154 + paddd xmm3, XMMWORD PTR[GLOBAL(_12000)]
1.155 + paddd xmm4, XMMWORD PTR[GLOBAL(_51000)]
1.156 + packssdw xmm0, xmm1 ;op[8] op[0]
1.157 + psrad xmm3, 16 ;(c1 * 2217 + d1 * 5352 + 12000)>>16
1.158 + psrad xmm4, 16 ;(d1 * 2217 - c1 * 5352 + 51000)>>16
1.159 +
1.160 + packssdw xmm3, xmm4 ;op[12] op[4]
1.161 + movdqa xmm1, xmm0
1.162 + paddw xmm3, xmm2 ;op[4] += (d1!=0)
1.163 + punpcklqdq xmm0, xmm3 ;op[4] op[0]
1.164 + punpckhqdq xmm1, xmm3 ;op[12] op[8]
1.165 +
1.166 + movdqa XMMWORD PTR[output + 0], xmm0
1.167 + movdqa XMMWORD PTR[output + 16], xmm1
1.168 +
1.169 + STACK_FRAME_DESTROY
1.170 +
1.171 +;void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch)
1.172 +global sym(vp8_short_fdct8x4_sse2) PRIVATE
1.173 +sym(vp8_short_fdct8x4_sse2):
1.174 +
1.175 + STACK_FRAME_CREATE
1.176 +
1.177 + ; read the input data
1.178 + movdqa xmm0, [input ]
1.179 + movdqa xmm2, [input+ pitch]
1.180 + lea input, [input+2*pitch]
1.181 + movdqa xmm4, [input ]
1.182 + movdqa xmm3, [input+ pitch]
1.183 +
1.184 + ; transpose for the first stage
1.185 + movdqa xmm1, xmm0 ; 00 01 02 03 04 05 06 07
1.186 + movdqa xmm5, xmm4 ; 20 21 22 23 24 25 26 27
1.187 +
1.188 + punpcklwd xmm0, xmm2 ; 00 10 01 11 02 12 03 13
1.189 + punpckhwd xmm1, xmm2 ; 04 14 05 15 06 16 07 17
1.190 +
1.191 + punpcklwd xmm4, xmm3 ; 20 30 21 31 22 32 23 33
1.192 + punpckhwd xmm5, xmm3 ; 24 34 25 35 26 36 27 37
1.193 +
1.194 + movdqa xmm2, xmm0 ; 00 10 01 11 02 12 03 13
1.195 + punpckldq xmm0, xmm4 ; 00 10 20 30 01 11 21 31
1.196 +
1.197 + punpckhdq xmm2, xmm4 ; 02 12 22 32 03 13 23 33
1.198 +
1.199 + movdqa xmm4, xmm1 ; 04 14 05 15 06 16 07 17
1.200 + punpckldq xmm4, xmm5 ; 04 14 24 34 05 15 25 35
1.201 +
1.202 + punpckhdq xmm1, xmm5 ; 06 16 26 36 07 17 27 37
1.203 + movdqa xmm3, xmm2 ; 02 12 22 32 03 13 23 33
1.204 +
1.205 + punpckhqdq xmm3, xmm1 ; 03 13 23 33 07 17 27 37
1.206 + punpcklqdq xmm2, xmm1 ; 02 12 22 32 06 16 26 36
1.207 +
1.208 + movdqa xmm1, xmm0 ; 00 10 20 30 01 11 21 31
1.209 + punpcklqdq xmm0, xmm4 ; 00 10 20 30 04 14 24 34
1.210 +
1.211 + punpckhqdq xmm1, xmm4 ; 01 11 21 32 05 15 25 35
1.212 +
1.213 + ; xmm0 0
1.214 + ; xmm1 1
1.215 + ; xmm2 2
1.216 + ; xmm3 3
1.217 +
1.218 + ; first stage
1.219 + movdqa xmm5, xmm0
1.220 + movdqa xmm4, xmm1
1.221 +
1.222 + paddw xmm0, xmm3 ; a1 = 0 + 3
1.223 + paddw xmm1, xmm2 ; b1 = 1 + 2
1.224 +
1.225 + psubw xmm4, xmm2 ; c1 = 1 - 2
1.226 + psubw xmm5, xmm3 ; d1 = 0 - 3
1.227 +
1.228 + psllw xmm5, 3
1.229 + psllw xmm4, 3
1.230 +
1.231 + psllw xmm0, 3
1.232 + psllw xmm1, 3
1.233 +
1.234 + ; output 0 and 2
1.235 + movdqa xmm2, xmm0 ; a1
1.236 +
1.237 + paddw xmm0, xmm1 ; op[0] = a1 + b1
1.238 + psubw xmm2, xmm1 ; op[2] = a1 - b1
1.239 +
1.240 + ; output 1 and 3
1.241 + ; interleave c1, d1
1.242 + movdqa xmm1, xmm5 ; d1
1.243 + punpcklwd xmm1, xmm4 ; c1 d1
1.244 + punpckhwd xmm5, xmm4 ; c1 d1
1.245 +
1.246 + movdqa xmm3, xmm1
1.247 + movdqa xmm4, xmm5
1.248 +
1.249 + pmaddwd xmm1, XMMWORD PTR[GLOBAL (_5352_2217)] ; c1*2217 + d1*5352
1.250 + pmaddwd xmm4, XMMWORD PTR[GLOBAL (_5352_2217)] ; c1*2217 + d1*5352
1.251 +
1.252 + pmaddwd xmm3, XMMWORD PTR[GLOBAL(_2217_neg5352)] ; d1*2217 - c1*5352
1.253 + pmaddwd xmm5, XMMWORD PTR[GLOBAL(_2217_neg5352)] ; d1*2217 - c1*5352
1.254 +
1.255 + paddd xmm1, XMMWORD PTR[GLOBAL(_14500)]
1.256 + paddd xmm4, XMMWORD PTR[GLOBAL(_14500)]
1.257 + paddd xmm3, XMMWORD PTR[GLOBAL(_7500)]
1.258 + paddd xmm5, XMMWORD PTR[GLOBAL(_7500)]
1.259 +
1.260 + psrad xmm1, 12 ; (c1 * 2217 + d1 * 5352 + 14500)>>12
1.261 + psrad xmm4, 12 ; (c1 * 2217 + d1 * 5352 + 14500)>>12
1.262 + psrad xmm3, 12 ; (d1 * 2217 - c1 * 5352 + 7500)>>12
1.263 + psrad xmm5, 12 ; (d1 * 2217 - c1 * 5352 + 7500)>>12
1.264 +
1.265 + packssdw xmm1, xmm4 ; op[1]
1.266 + packssdw xmm3, xmm5 ; op[3]
1.267 +
1.268 + ; done with vertical
1.269 + ; transpose for the second stage
1.270 + movdqa xmm4, xmm0 ; 00 10 20 30 04 14 24 34
1.271 + movdqa xmm5, xmm2 ; 02 12 22 32 06 16 26 36
1.272 +
1.273 + punpcklwd xmm0, xmm1 ; 00 01 10 11 20 21 30 31
1.274 + punpckhwd xmm4, xmm1 ; 04 05 14 15 24 25 34 35
1.275 +
1.276 + punpcklwd xmm2, xmm3 ; 02 03 12 13 22 23 32 33
1.277 + punpckhwd xmm5, xmm3 ; 06 07 16 17 26 27 36 37
1.278 +
1.279 + movdqa xmm1, xmm0 ; 00 01 10 11 20 21 30 31
1.280 + punpckldq xmm0, xmm2 ; 00 01 02 03 10 11 12 13
1.281 +
1.282 + punpckhdq xmm1, xmm2 ; 20 21 22 23 30 31 32 33
1.283 +
1.284 + movdqa xmm2, xmm4 ; 04 05 14 15 24 25 34 35
1.285 + punpckldq xmm2, xmm5 ; 04 05 06 07 14 15 16 17
1.286 +
1.287 + punpckhdq xmm4, xmm5 ; 24 25 26 27 34 35 36 37
1.288 + movdqa xmm3, xmm1 ; 20 21 22 23 30 31 32 33
1.289 +
1.290 + punpckhqdq xmm3, xmm4 ; 30 31 32 33 34 35 36 37
1.291 + punpcklqdq xmm1, xmm4 ; 20 21 22 23 24 25 26 27
1.292 +
1.293 + movdqa xmm4, xmm0 ; 00 01 02 03 10 11 12 13
1.294 + punpcklqdq xmm0, xmm2 ; 00 01 02 03 04 05 06 07
1.295 +
1.296 + punpckhqdq xmm4, xmm2 ; 10 11 12 13 14 15 16 17
1.297 +
1.298 + ; xmm0 0
1.299 + ; xmm1 4
1.300 + ; xmm2 1
1.301 + ; xmm3 3
1.302 +
1.303 + movdqa xmm5, xmm0
1.304 + movdqa xmm2, xmm1
1.305 +
1.306 + paddw xmm0, xmm3 ; a1 = 0 + 3
1.307 + paddw xmm1, xmm4 ; b1 = 1 + 2
1.308 +
1.309 + psubw xmm4, xmm2 ; c1 = 1 - 2
1.310 + psubw xmm5, xmm3 ; d1 = 0 - 3
1.311 +
1.312 + pxor xmm6, xmm6 ; zero out for compare
1.313 +
1.314 + pcmpeqw xmm6, xmm5 ; d1 != 0
1.315 +
1.316 + pandn xmm6, XMMWORD PTR[GLOBAL(_cmp_mask8x4)] ; clear upper,
1.317 + ; and keep bit 0 of lower
1.318 +
1.319 + ; output 0 and 2
1.320 + movdqa xmm2, xmm0 ; a1
1.321 +
1.322 + paddw xmm0, xmm1 ; a1 + b1
1.323 + psubw xmm2, xmm1 ; a1 - b1
1.324 +
1.325 + paddw xmm0, XMMWORD PTR[GLOBAL(_7w)]
1.326 + paddw xmm2, XMMWORD PTR[GLOBAL(_7w)]
1.327 +
1.328 + psraw xmm0, 4 ; op[0] = (a1 + b1 + 7)>>4
1.329 + psraw xmm2, 4 ; op[8] = (a1 - b1 + 7)>>4
1.330 +
1.331 + ; output 1 and 3
1.332 + ; interleave c1, d1
1.333 + movdqa xmm1, xmm5 ; d1
1.334 + punpcklwd xmm1, xmm4 ; c1 d1
1.335 + punpckhwd xmm5, xmm4 ; c1 d1
1.336 +
1.337 + movdqa xmm3, xmm1
1.338 + movdqa xmm4, xmm5
1.339 +
1.340 + pmaddwd xmm1, XMMWORD PTR[GLOBAL (_5352_2217)] ; c1*2217 + d1*5352
1.341 + pmaddwd xmm4, XMMWORD PTR[GLOBAL (_5352_2217)] ; c1*2217 + d1*5352
1.342 +
1.343 + pmaddwd xmm3, XMMWORD PTR[GLOBAL(_2217_neg5352)] ; d1*2217 - c1*5352
1.344 + pmaddwd xmm5, XMMWORD PTR[GLOBAL(_2217_neg5352)] ; d1*2217 - c1*5352
1.345 +
1.346 + paddd xmm1, XMMWORD PTR[GLOBAL(_12000)]
1.347 + paddd xmm4, XMMWORD PTR[GLOBAL(_12000)]
1.348 + paddd xmm3, XMMWORD PTR[GLOBAL(_51000)]
1.349 + paddd xmm5, XMMWORD PTR[GLOBAL(_51000)]
1.350 +
1.351 + psrad xmm1, 16 ; (c1 * 2217 + d1 * 5352 + 14500)>>16
1.352 + psrad xmm4, 16 ; (c1 * 2217 + d1 * 5352 + 14500)>>16
1.353 + psrad xmm3, 16 ; (d1 * 2217 - c1 * 5352 + 7500)>>16
1.354 + psrad xmm5, 16 ; (d1 * 2217 - c1 * 5352 + 7500)>>16
1.355 +
1.356 + packssdw xmm1, xmm4 ; op[4]
1.357 + packssdw xmm3, xmm5 ; op[12]
1.358 +
1.359 + paddw xmm1, xmm6 ; op[4] += (d1!=0)
1.360 +
1.361 + movdqa xmm4, xmm0
1.362 + movdqa xmm5, xmm2
1.363 +
1.364 + punpcklqdq xmm0, xmm1
1.365 + punpckhqdq xmm4, xmm1
1.366 +
1.367 + punpcklqdq xmm2, xmm3
1.368 + punpckhqdq xmm5, xmm3
1.369 +
1.370 + movdqa XMMWORD PTR[output + 0 ], xmm0
1.371 + movdqa XMMWORD PTR[output + 16], xmm2
1.372 + movdqa XMMWORD PTR[output + 32], xmm4
1.373 + movdqa XMMWORD PTR[output + 48], xmm5
1.374 +
1.375 + STACK_FRAME_DESTROY
1.376 +
1.377 +SECTION_RODATA
1.378 +align 16
1.379 +_5352_2217:
1.380 + dw 5352
1.381 + dw 2217
1.382 + dw 5352
1.383 + dw 2217
1.384 + dw 5352
1.385 + dw 2217
1.386 + dw 5352
1.387 + dw 2217
1.388 +align 16
1.389 +_2217_neg5352:
1.390 + dw 2217
1.391 + dw -5352
1.392 + dw 2217
1.393 + dw -5352
1.394 + dw 2217
1.395 + dw -5352
1.396 + dw 2217
1.397 + dw -5352
1.398 +align 16
1.399 +_mult_add:
1.400 + times 8 dw 1
1.401 +align 16
1.402 +_cmp_mask:
1.403 + times 4 dw 1
1.404 + times 4 dw 0
1.405 +align 16
1.406 +_cmp_mask8x4:
1.407 + times 8 dw 1
1.408 +align 16
1.409 +_mult_sub:
1.410 + dw 1
1.411 + dw -1
1.412 + dw 1
1.413 + dw -1
1.414 + dw 1
1.415 + dw -1
1.416 + dw 1
1.417 + dw -1
1.418 +align 16
1.419 +_7:
1.420 + times 4 dd 7
1.421 +align 16
1.422 +_7w:
1.423 + times 8 dw 7
1.424 +align 16
1.425 +_14500:
1.426 + times 4 dd 14500
1.427 +align 16
1.428 +_7500:
1.429 + times 4 dd 7500
1.430 +align 16
1.431 +_12000:
1.432 + times 4 dd 12000
1.433 +align 16
1.434 +_51000:
1.435 + times 4 dd 51000
