1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/common/x86/vp9_postproc_sse2.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,695 @@
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 +;void vp9_post_proc_down_and_across_xmm
1.18 +;(
1.19 +; unsigned char *src_ptr,
1.20 +; unsigned char *dst_ptr,
1.21 +; int src_pixels_per_line,
1.22 +; int dst_pixels_per_line,
1.23 +; int rows,
1.24 +; int cols,
1.25 +; int flimit
1.26 +;)
1.27 +global sym(vp9_post_proc_down_and_across_xmm) PRIVATE
1.28 +sym(vp9_post_proc_down_and_across_xmm):
1.29 + push rbp
1.30 + mov rbp, rsp
1.31 + SHADOW_ARGS_TO_STACK 7
1.32 + SAVE_XMM 7
1.33 + GET_GOT rbx
1.34 + push rsi
1.35 + push rdi
1.36 + ; end prolog
1.37 +
1.38 +%if ABI_IS_32BIT=1 && CONFIG_PIC=1
1.39 + ALIGN_STACK 16, rax
1.40 + ; move the global rd onto the stack, since we don't have enough registers
1.41 + ; to do PIC addressing
1.42 + movdqa xmm0, [GLOBAL(rd42)]
1.43 + sub rsp, 16
1.44 + movdqa [rsp], xmm0
1.45 +%define RD42 [rsp]
1.46 +%else
1.47 +%define RD42 [GLOBAL(rd42)]
1.48 +%endif
1.49 +
1.50 +
1.51 + movd xmm2, dword ptr arg(6) ;flimit
1.52 + punpcklwd xmm2, xmm2
1.53 + punpckldq xmm2, xmm2
1.54 + punpcklqdq xmm2, xmm2
1.55 +
1.56 + mov rsi, arg(0) ;src_ptr
1.57 + mov rdi, arg(1) ;dst_ptr
1.58 +
1.59 + movsxd rcx, DWORD PTR arg(4) ;rows
1.60 + movsxd rax, DWORD PTR arg(2) ;src_pixels_per_line ; destination pitch?
1.61 + pxor xmm0, xmm0 ; mm0 = 00000000
1.62 +
1.63 +.nextrow:
1.64 +
1.65 + xor rdx, rdx ; clear out rdx for use as loop counter
1.66 +.nextcol:
1.67 + movq xmm3, QWORD PTR [rsi] ; mm4 = r0 p0..p7
1.68 + punpcklbw xmm3, xmm0 ; mm3 = p0..p3
1.69 + movdqa xmm1, xmm3 ; mm1 = p0..p3
1.70 + psllw xmm3, 2 ;
1.71 +
1.72 + movq xmm5, QWORD PTR [rsi + rax] ; mm4 = r1 p0..p7
1.73 + punpcklbw xmm5, xmm0 ; mm5 = r1 p0..p3
1.74 + paddusw xmm3, xmm5 ; mm3 += mm6
1.75 +
1.76 + ; thresholding
1.77 + movdqa xmm7, xmm1 ; mm7 = r0 p0..p3
1.78 + psubusw xmm7, xmm5 ; mm7 = r0 p0..p3 - r1 p0..p3
1.79 + psubusw xmm5, xmm1 ; mm5 = r1 p0..p3 - r0 p0..p3
1.80 + paddusw xmm7, xmm5 ; mm7 = abs(r0 p0..p3 - r1 p0..p3)
1.81 + pcmpgtw xmm7, xmm2
1.82 +
1.83 + movq xmm5, QWORD PTR [rsi + 2*rax] ; mm4 = r2 p0..p7
1.84 + punpcklbw xmm5, xmm0 ; mm5 = r2 p0..p3
1.85 + paddusw xmm3, xmm5 ; mm3 += mm5
1.86 +
1.87 + ; thresholding
1.88 + movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
1.89 + psubusw xmm6, xmm5 ; mm6 = r0 p0..p3 - r2 p0..p3
1.90 + psubusw xmm5, xmm1 ; mm5 = r2 p0..p3 - r2 p0..p3
1.91 + paddusw xmm6, xmm5 ; mm6 = abs(r0 p0..p3 - r2 p0..p3)
1.92 + pcmpgtw xmm6, xmm2
1.93 + por xmm7, xmm6 ; accumulate thresholds
1.94 +
1.95 +
1.96 + neg rax
1.97 + movq xmm5, QWORD PTR [rsi+2*rax] ; mm4 = r-2 p0..p7
1.98 + punpcklbw xmm5, xmm0 ; mm5 = r-2 p0..p3
1.99 + paddusw xmm3, xmm5 ; mm3 += mm5
1.100 +
1.101 + ; thresholding
1.102 + movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
1.103 + psubusw xmm6, xmm5 ; mm6 = p0..p3 - r-2 p0..p3
1.104 + psubusw xmm5, xmm1 ; mm5 = r-2 p0..p3 - p0..p3
1.105 + paddusw xmm6, xmm5 ; mm6 = abs(r0 p0..p3 - r-2 p0..p3)
1.106 + pcmpgtw xmm6, xmm2
1.107 + por xmm7, xmm6 ; accumulate thresholds
1.108 +
1.109 + movq xmm4, QWORD PTR [rsi+rax] ; mm4 = r-1 p0..p7
1.110 + punpcklbw xmm4, xmm0 ; mm4 = r-1 p0..p3
1.111 + paddusw xmm3, xmm4 ; mm3 += mm5
1.112 +
1.113 + ; thresholding
1.114 + movdqa xmm6, xmm1 ; mm6 = r0 p0..p3
1.115 + psubusw xmm6, xmm4 ; mm6 = p0..p3 - r-2 p0..p3
1.116 + psubusw xmm4, xmm1 ; mm5 = r-1 p0..p3 - p0..p3
1.117 + paddusw xmm6, xmm4 ; mm6 = abs(r0 p0..p3 - r-1 p0..p3)
1.118 + pcmpgtw xmm6, xmm2
1.119 + por xmm7, xmm6 ; accumulate thresholds
1.120 +
1.121 +
1.122 + paddusw xmm3, RD42 ; mm3 += round value
1.123 + psraw xmm3, 3 ; mm3 /= 8
1.124 +
1.125 + pand xmm1, xmm7 ; mm1 select vals > thresh from source
1.126 + pandn xmm7, xmm3 ; mm7 select vals < thresh from blurred result
1.127 + paddusw xmm1, xmm7 ; combination
1.128 +
1.129 + packuswb xmm1, xmm0 ; pack to bytes
1.130 + movq QWORD PTR [rdi], xmm1 ;
1.131 +
1.132 + neg rax ; pitch is positive
1.133 + add rsi, 8
1.134 + add rdi, 8
1.135 +
1.136 + add rdx, 8
1.137 + cmp edx, dword arg(5) ;cols
1.138 +
1.139 + jl .nextcol
1.140 +
1.141 + ; done with the all cols, start the across filtering in place
1.142 + sub rsi, rdx
1.143 + sub rdi, rdx
1.144 +
1.145 + xor rdx, rdx
1.146 + movq mm0, QWORD PTR [rdi-8];
1.147 +
1.148 +.acrossnextcol:
1.149 + movq xmm7, QWORD PTR [rdi +rdx -2]
1.150 + movd xmm4, DWORD PTR [rdi +rdx +6]
1.151 +
1.152 + pslldq xmm4, 8
1.153 + por xmm4, xmm7
1.154 +
1.155 + movdqa xmm3, xmm4
1.156 + psrldq xmm3, 2
1.157 + punpcklbw xmm3, xmm0 ; mm3 = p0..p3
1.158 + movdqa xmm1, xmm3 ; mm1 = p0..p3
1.159 + psllw xmm3, 2
1.160 +
1.161 +
1.162 + movdqa xmm5, xmm4
1.163 + psrldq xmm5, 3
1.164 + punpcklbw xmm5, xmm0 ; mm5 = p1..p4
1.165 + paddusw xmm3, xmm5 ; mm3 += mm6
1.166 +
1.167 + ; thresholding
1.168 + movdqa xmm7, xmm1 ; mm7 = p0..p3
1.169 + psubusw xmm7, xmm5 ; mm7 = p0..p3 - p1..p4
1.170 + psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
1.171 + paddusw xmm7, xmm5 ; mm7 = abs(p0..p3 - p1..p4)
1.172 + pcmpgtw xmm7, xmm2
1.173 +
1.174 + movdqa xmm5, xmm4
1.175 + psrldq xmm5, 4
1.176 + punpcklbw xmm5, xmm0 ; mm5 = p2..p5
1.177 + paddusw xmm3, xmm5 ; mm3 += mm5
1.178 +
1.179 + ; thresholding
1.180 + movdqa xmm6, xmm1 ; mm6 = p0..p3
1.181 + psubusw xmm6, xmm5 ; mm6 = p0..p3 - p1..p4
1.182 + psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
1.183 + paddusw xmm6, xmm5 ; mm6 = abs(p0..p3 - p1..p4)
1.184 + pcmpgtw xmm6, xmm2
1.185 + por xmm7, xmm6 ; accumulate thresholds
1.186 +
1.187 +
1.188 + movdqa xmm5, xmm4 ; mm5 = p-2..p5
1.189 + punpcklbw xmm5, xmm0 ; mm5 = p-2..p1
1.190 + paddusw xmm3, xmm5 ; mm3 += mm5
1.191 +
1.192 + ; thresholding
1.193 + movdqa xmm6, xmm1 ; mm6 = p0..p3
1.194 + psubusw xmm6, xmm5 ; mm6 = p0..p3 - p1..p4
1.195 + psubusw xmm5, xmm1 ; mm5 = p1..p4 - p0..p3
1.196 + paddusw xmm6, xmm5 ; mm6 = abs(p0..p3 - p1..p4)
1.197 + pcmpgtw xmm6, xmm2
1.198 + por xmm7, xmm6 ; accumulate thresholds
1.199 +
1.200 + psrldq xmm4, 1 ; mm4 = p-1..p5
1.201 + punpcklbw xmm4, xmm0 ; mm4 = p-1..p2
1.202 + paddusw xmm3, xmm4 ; mm3 += mm5
1.203 +
1.204 + ; thresholding
1.205 + movdqa xmm6, xmm1 ; mm6 = p0..p3
1.206 + psubusw xmm6, xmm4 ; mm6 = p0..p3 - p1..p4
1.207 + psubusw xmm4, xmm1 ; mm5 = p1..p4 - p0..p3
1.208 + paddusw xmm6, xmm4 ; mm6 = abs(p0..p3 - p1..p4)
1.209 + pcmpgtw xmm6, xmm2
1.210 + por xmm7, xmm6 ; accumulate thresholds
1.211 +
1.212 + paddusw xmm3, RD42 ; mm3 += round value
1.213 + psraw xmm3, 3 ; mm3 /= 8
1.214 +
1.215 + pand xmm1, xmm7 ; mm1 select vals > thresh from source
1.216 + pandn xmm7, xmm3 ; mm7 select vals < thresh from blurred result
1.217 + paddusw xmm1, xmm7 ; combination
1.218 +
1.219 + packuswb xmm1, xmm0 ; pack to bytes
1.220 + movq QWORD PTR [rdi+rdx-8], mm0 ; store previous four bytes
1.221 + movdq2q mm0, xmm1
1.222 +
1.223 + add rdx, 8
1.224 + cmp edx, dword arg(5) ;cols
1.225 + jl .acrossnextcol;
1.226 +
1.227 + ; last 8 pixels
1.228 + movq QWORD PTR [rdi+rdx-8], mm0
1.229 +
1.230 + ; done with this rwo
1.231 + add rsi,rax ; next line
1.232 + mov eax, dword arg(3) ;dst_pixels_per_line ; destination pitch?
1.233 + add rdi,rax ; next destination
1.234 + mov eax, dword arg(2) ;src_pixels_per_line ; destination pitch?
1.235 +
1.236 + dec rcx ; decrement count
1.237 + jnz .nextrow ; next row
1.238 +
1.239 +%if ABI_IS_32BIT=1 && CONFIG_PIC=1
1.240 + add rsp,16
1.241 + pop rsp
1.242 +%endif
1.243 + ; begin epilog
1.244 + pop rdi
1.245 + pop rsi
1.246 + RESTORE_GOT
1.247 + RESTORE_XMM
1.248 + UNSHADOW_ARGS
1.249 + pop rbp
1.250 + ret
1.251 +%undef RD42
1.252 +
1.253 +
1.254 +;void vp9_mbpost_proc_down_xmm(unsigned char *dst,
1.255 +; int pitch, int rows, int cols,int flimit)
1.256 +extern sym(vp9_rv)
1.257 +global sym(vp9_mbpost_proc_down_xmm) PRIVATE
1.258 +sym(vp9_mbpost_proc_down_xmm):
1.259 + push rbp
1.260 + mov rbp, rsp
1.261 + SHADOW_ARGS_TO_STACK 5
1.262 + SAVE_XMM 7
1.263 + GET_GOT rbx
1.264 + push rsi
1.265 + push rdi
1.266 + ; end prolog
1.267 +
1.268 + ALIGN_STACK 16, rax
1.269 + sub rsp, 128+16
1.270 +
1.271 + ; unsigned char d[16][8] at [rsp]
1.272 + ; create flimit2 at [rsp+128]
1.273 + mov eax, dword ptr arg(4) ;flimit
1.274 + mov [rsp+128], eax
1.275 + mov [rsp+128+4], eax
1.276 + mov [rsp+128+8], eax
1.277 + mov [rsp+128+12], eax
1.278 +%define flimit4 [rsp+128]
1.279 +
1.280 +%if ABI_IS_32BIT=0
1.281 + lea r8, [GLOBAL(sym(vp9_rv))]
1.282 +%endif
1.283 +
1.284 + ;rows +=8;
1.285 + add dword arg(2), 8
1.286 +
1.287 + ;for(c=0; c<cols; c+=8)
1.288 +.loop_col:
1.289 + mov rsi, arg(0) ; s
1.290 + pxor xmm0, xmm0 ;
1.291 +
1.292 + movsxd rax, dword ptr arg(1) ;pitch ;
1.293 + neg rax ; rax = -pitch
1.294 +
1.295 + lea rsi, [rsi + rax*8]; ; rdi = s[-pitch*8]
1.296 + neg rax
1.297 +
1.298 +
1.299 + pxor xmm5, xmm5
1.300 + pxor xmm6, xmm6 ;
1.301 +
1.302 + pxor xmm7, xmm7 ;
1.303 + mov rdi, rsi
1.304 +
1.305 + mov rcx, 15 ;
1.306 +
1.307 +.loop_initvar:
1.308 + movq xmm1, QWORD PTR [rdi];
1.309 + punpcklbw xmm1, xmm0 ;
1.310 +
1.311 + paddw xmm5, xmm1 ;
1.312 + pmullw xmm1, xmm1 ;
1.313 +
1.314 + movdqa xmm2, xmm1 ;
1.315 + punpcklwd xmm1, xmm0 ;
1.316 +
1.317 + punpckhwd xmm2, xmm0 ;
1.318 + paddd xmm6, xmm1 ;
1.319 +
1.320 + paddd xmm7, xmm2 ;
1.321 + lea rdi, [rdi+rax] ;
1.322 +
1.323 + dec rcx
1.324 + jne .loop_initvar
1.325 + ;save the var and sum
1.326 + xor rdx, rdx
1.327 +.loop_row:
1.328 + movq xmm1, QWORD PTR [rsi] ; [s-pitch*8]
1.329 + movq xmm2, QWORD PTR [rdi] ; [s+pitch*7]
1.330 +
1.331 + punpcklbw xmm1, xmm0
1.332 + punpcklbw xmm2, xmm0
1.333 +
1.334 + paddw xmm5, xmm2
1.335 + psubw xmm5, xmm1
1.336 +
1.337 + pmullw xmm2, xmm2
1.338 + movdqa xmm4, xmm2
1.339 +
1.340 + punpcklwd xmm2, xmm0
1.341 + punpckhwd xmm4, xmm0
1.342 +
1.343 + paddd xmm6, xmm2
1.344 + paddd xmm7, xmm4
1.345 +
1.346 + pmullw xmm1, xmm1
1.347 + movdqa xmm2, xmm1
1.348 +
1.349 + punpcklwd xmm1, xmm0
1.350 + psubd xmm6, xmm1
1.351 +
1.352 + punpckhwd xmm2, xmm0
1.353 + psubd xmm7, xmm2
1.354 +
1.355 +
1.356 + movdqa xmm3, xmm6
1.357 + pslld xmm3, 4
1.358 +
1.359 + psubd xmm3, xmm6
1.360 + movdqa xmm1, xmm5
1.361 +
1.362 + movdqa xmm4, xmm5
1.363 + pmullw xmm1, xmm1
1.364 +
1.365 + pmulhw xmm4, xmm4
1.366 + movdqa xmm2, xmm1
1.367 +
1.368 + punpcklwd xmm1, xmm4
1.369 + punpckhwd xmm2, xmm4
1.370 +
1.371 + movdqa xmm4, xmm7
1.372 + pslld xmm4, 4
1.373 +
1.374 + psubd xmm4, xmm7
1.375 +
1.376 + psubd xmm3, xmm1
1.377 + psubd xmm4, xmm2
1.378 +
1.379 + psubd xmm3, flimit4
1.380 + psubd xmm4, flimit4
1.381 +
1.382 + psrad xmm3, 31
1.383 + psrad xmm4, 31
1.384 +
1.385 + packssdw xmm3, xmm4
1.386 + packsswb xmm3, xmm0
1.387 +
1.388 + movq xmm1, QWORD PTR [rsi+rax*8]
1.389 +
1.390 + movq xmm2, xmm1
1.391 + punpcklbw xmm1, xmm0
1.392 +
1.393 + paddw xmm1, xmm5
1.394 + mov rcx, rdx
1.395 +
1.396 + and rcx, 127
1.397 +%if ABI_IS_32BIT=1 && CONFIG_PIC=1
1.398 + push rax
1.399 + lea rax, [GLOBAL(sym(vp9_rv))]
1.400 + movdqu xmm4, [rax + rcx*2] ;vp9_rv[rcx*2]
1.401 + pop rax
1.402 +%elif ABI_IS_32BIT=0
1.403 + movdqu xmm4, [r8 + rcx*2] ;vp9_rv[rcx*2]
1.404 +%else
1.405 + movdqu xmm4, [sym(vp9_rv) + rcx*2]
1.406 +%endif
1.407 +
1.408 + paddw xmm1, xmm4
1.409 + ;paddw xmm1, eight8s
1.410 + psraw xmm1, 4
1.411 +
1.412 + packuswb xmm1, xmm0
1.413 + pand xmm1, xmm3
1.414 +
1.415 + pandn xmm3, xmm2
1.416 + por xmm1, xmm3
1.417 +
1.418 + and rcx, 15
1.419 + movq QWORD PTR [rsp + rcx*8], xmm1 ;d[rcx*8]
1.420 +
1.421 + mov rcx, rdx
1.422 + sub rcx, 8
1.423 +
1.424 + and rcx, 15
1.425 + movq mm0, [rsp + rcx*8] ;d[rcx*8]
1.426 +
1.427 + movq [rsi], mm0
1.428 + lea rsi, [rsi+rax]
1.429 +
1.430 + lea rdi, [rdi+rax]
1.431 + add rdx, 1
1.432 +
1.433 + cmp edx, dword arg(2) ;rows
1.434 + jl .loop_row
1.435 +
1.436 + add dword arg(0), 8 ; s += 8
1.437 + sub dword arg(3), 8 ; cols -= 8
1.438 + cmp dword arg(3), 0
1.439 + jg .loop_col
1.440 +
1.441 + add rsp, 128+16
1.442 + pop rsp
1.443 +
1.444 + ; begin epilog
1.445 + pop rdi
1.446 + pop rsi
1.447 + RESTORE_GOT
1.448 + RESTORE_XMM
1.449 + UNSHADOW_ARGS
1.450 + pop rbp
1.451 + ret
1.452 +%undef flimit4
1.453 +
1.454 +
1.455 +;void vp9_mbpost_proc_across_ip_xmm(unsigned char *src,
1.456 +; int pitch, int rows, int cols,int flimit)
1.457 +global sym(vp9_mbpost_proc_across_ip_xmm) PRIVATE
1.458 +sym(vp9_mbpost_proc_across_ip_xmm):
1.459 + push rbp
1.460 + mov rbp, rsp
1.461 + SHADOW_ARGS_TO_STACK 5
1.462 + SAVE_XMM 7
1.463 + GET_GOT rbx
1.464 + push rsi
1.465 + push rdi
1.466 + ; end prolog
1.467 +
1.468 + ALIGN_STACK 16, rax
1.469 + sub rsp, 16
1.470 +
1.471 + ; create flimit4 at [rsp]
1.472 + mov eax, dword ptr arg(4) ;flimit
1.473 + mov [rsp], eax
1.474 + mov [rsp+4], eax
1.475 + mov [rsp+8], eax
1.476 + mov [rsp+12], eax
1.477 +%define flimit4 [rsp]
1.478 +
1.479 +
1.480 + ;for(r=0;r<rows;r++)
1.481 +.ip_row_loop:
1.482 +
1.483 + xor rdx, rdx ;sumsq=0;
1.484 + xor rcx, rcx ;sum=0;
1.485 + mov rsi, arg(0); s
1.486 + mov rdi, -8
1.487 +.ip_var_loop:
1.488 + ;for(i=-8;i<=6;i++)
1.489 + ;{
1.490 + ; sumsq += s[i]*s[i];
1.491 + ; sum += s[i];
1.492 + ;}
1.493 + movzx eax, byte [rsi+rdi]
1.494 + add ecx, eax
1.495 + mul al
1.496 + add edx, eax
1.497 + add rdi, 1
1.498 + cmp rdi, 6
1.499 + jle .ip_var_loop
1.500 +
1.501 +
1.502 + ;mov rax, sumsq
1.503 + ;movd xmm7, rax
1.504 + movd xmm7, edx
1.505 +
1.506 + ;mov rax, sum
1.507 + ;movd xmm6, rax
1.508 + movd xmm6, ecx
1.509 +
1.510 + mov rsi, arg(0) ;s
1.511 + xor rcx, rcx
1.512 +
1.513 + movsxd rdx, dword arg(3) ;cols
1.514 + add rdx, 8
1.515 + pxor mm0, mm0
1.516 + pxor mm1, mm1
1.517 +
1.518 + pxor xmm0, xmm0
1.519 +.nextcol4:
1.520 +
1.521 + movd xmm1, DWORD PTR [rsi+rcx-8] ; -8 -7 -6 -5
1.522 + movd xmm2, DWORD PTR [rsi+rcx+7] ; +7 +8 +9 +10
1.523 +
1.524 + punpcklbw xmm1, xmm0 ; expanding
1.525 + punpcklbw xmm2, xmm0 ; expanding
1.526 +
1.527 + punpcklwd xmm1, xmm0 ; expanding to dwords
1.528 + punpcklwd xmm2, xmm0 ; expanding to dwords
1.529 +
1.530 + psubd xmm2, xmm1 ; 7--8 8--7 9--6 10--5
1.531 + paddd xmm1, xmm1 ; -8*2 -7*2 -6*2 -5*2
1.532 +
1.533 + paddd xmm1, xmm2 ; 7+-8 8+-7 9+-6 10+-5
1.534 + pmaddwd xmm1, xmm2 ; squared of 7+-8 8+-7 9+-6 10+-5
1.535 +
1.536 + paddd xmm6, xmm2
1.537 + paddd xmm7, xmm1
1.538 +
1.539 + pshufd xmm6, xmm6, 0 ; duplicate the last ones
1.540 + pshufd xmm7, xmm7, 0 ; duplicate the last ones
1.541 +
1.542 + psrldq xmm1, 4 ; 8--7 9--6 10--5 0000
1.543 + psrldq xmm2, 4 ; 8--7 9--6 10--5 0000
1.544 +
1.545 + pshufd xmm3, xmm1, 3 ; 0000 8--7 8--7 8--7 squared
1.546 + pshufd xmm4, xmm2, 3 ; 0000 8--7 8--7 8--7 squared
1.547 +
1.548 + paddd xmm6, xmm4
1.549 + paddd xmm7, xmm3
1.550 +
1.551 + pshufd xmm3, xmm1, 01011111b ; 0000 0000 9--6 9--6 squared
1.552 + pshufd xmm4, xmm2, 01011111b ; 0000 0000 9--6 9--6 squared
1.553 +
1.554 + paddd xmm7, xmm3
1.555 + paddd xmm6, xmm4
1.556 +
1.557 + pshufd xmm3, xmm1, 10111111b ; 0000 0000 8--7 8--7 squared
1.558 + pshufd xmm4, xmm2, 10111111b ; 0000 0000 8--7 8--7 squared
1.559 +
1.560 + paddd xmm7, xmm3
1.561 + paddd xmm6, xmm4
1.562 +
1.563 + movdqa xmm3, xmm6
1.564 + pmaddwd xmm3, xmm3
1.565 +
1.566 + movdqa xmm5, xmm7
1.567 + pslld xmm5, 4
1.568 +
1.569 + psubd xmm5, xmm7
1.570 + psubd xmm5, xmm3
1.571 +
1.572 + psubd xmm5, flimit4
1.573 + psrad xmm5, 31
1.574 +
1.575 + packssdw xmm5, xmm0
1.576 + packsswb xmm5, xmm0
1.577 +
1.578 + movd xmm1, DWORD PTR [rsi+rcx]
1.579 + movq xmm2, xmm1
1.580 +
1.581 + punpcklbw xmm1, xmm0
1.582 + punpcklwd xmm1, xmm0
1.583 +
1.584 + paddd xmm1, xmm6
1.585 + paddd xmm1, [GLOBAL(four8s)]
1.586 +
1.587 + psrad xmm1, 4
1.588 + packssdw xmm1, xmm0
1.589 +
1.590 + packuswb xmm1, xmm0
1.591 + pand xmm1, xmm5
1.592 +
1.593 + pandn xmm5, xmm2
1.594 + por xmm5, xmm1
1.595 +
1.596 + movd [rsi+rcx-8], mm0
1.597 + movq mm0, mm1
1.598 +
1.599 + movdq2q mm1, xmm5
1.600 + psrldq xmm7, 12
1.601 +
1.602 + psrldq xmm6, 12
1.603 + add rcx, 4
1.604 +
1.605 + cmp rcx, rdx
1.606 + jl .nextcol4
1.607 +
1.608 + ;s+=pitch;
1.609 + movsxd rax, dword arg(1)
1.610 + add arg(0), rax
1.611 +
1.612 + sub dword arg(2), 1 ;rows-=1
1.613 + cmp dword arg(2), 0
1.614 + jg .ip_row_loop
1.615 +
1.616 + add rsp, 16
1.617 + pop rsp
1.618 +
1.619 + ; begin epilog
1.620 + pop rdi
1.621 + pop rsi
1.622 + RESTORE_GOT
1.623 + RESTORE_XMM
1.624 + UNSHADOW_ARGS
1.625 + pop rbp
1.626 + ret
1.627 +%undef flimit4
1.628 +
1.629 +
1.630 +;void vp9_plane_add_noise_wmt (unsigned char *start, unsigned char *noise,
1.631 +; unsigned char blackclamp[16],
1.632 +; unsigned char whiteclamp[16],
1.633 +; unsigned char bothclamp[16],
1.634 +; unsigned int width, unsigned int height, int pitch)
1.635 +extern sym(rand)
1.636 +global sym(vp9_plane_add_noise_wmt) PRIVATE
1.637 +sym(vp9_plane_add_noise_wmt):
1.638 + push rbp
1.639 + mov rbp, rsp
1.640 + SHADOW_ARGS_TO_STACK 8
1.641 + GET_GOT rbx
1.642 + push rsi
1.643 + push rdi
1.644 + ; end prolog
1.645 +
1.646 +.addnoise_loop:
1.647 + call sym(rand) WRT_PLT
1.648 + mov rcx, arg(1) ;noise
1.649 + and rax, 0xff
1.650 + add rcx, rax
1.651 +
1.652 + ; we rely on the fact that the clamping vectors are stored contiguously
1.653 + ; in black/white/both order. Note that we have to reload this here because
1.654 + ; rdx could be trashed by rand()
1.655 + mov rdx, arg(2) ; blackclamp
1.656 +
1.657 +
1.658 + mov rdi, rcx
1.659 + movsxd rcx, dword arg(5) ;[Width]
1.660 + mov rsi, arg(0) ;Pos
1.661 + xor rax,rax
1.662 +
1.663 +.addnoise_nextset:
1.664 + movdqu xmm1,[rsi+rax] ; get the source
1.665 +
1.666 + psubusb xmm1, [rdx] ;blackclamp ; clamp both sides so we don't outrange adding noise
1.667 + paddusb xmm1, [rdx+32] ;bothclamp
1.668 + psubusb xmm1, [rdx+16] ;whiteclamp
1.669 +
1.670 + movdqu xmm2,[rdi+rax] ; get the noise for this line
1.671 + paddb xmm1,xmm2 ; add it in
1.672 + movdqu [rsi+rax],xmm1 ; store the result
1.673 +
1.674 + add rax,16 ; move to the next line
1.675 +
1.676 + cmp rax, rcx
1.677 + jl .addnoise_nextset
1.678 +
1.679 + movsxd rax, dword arg(7) ; Pitch
1.680 + add arg(0), rax ; Start += Pitch
1.681 + sub dword arg(6), 1 ; Height -= 1
1.682 + jg .addnoise_loop
1.683 +
1.684 + ; begin epilog
1.685 + pop rdi
1.686 + pop rsi
1.687 + RESTORE_GOT
1.688 + UNSHADOW_ARGS
1.689 + pop rbp
1.690 + ret
1.691 +
1.692 +
1.693 +SECTION_RODATA
1.694 +align 16
1.695 +rd42:
1.696 + times 8 dw 0x04
1.697 +four8s:
1.698 + times 4 dd 8
