1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/common/x86/postproc_mmx.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,313 @@
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 +%define VP8_FILTER_WEIGHT 128
1.18 +%define VP8_FILTER_SHIFT 7
1.19 +
1.20 +;void vp8_mbpost_proc_down_mmx(unsigned char *dst,
1.21 +; int pitch, int rows, int cols,int flimit)
1.22 +extern sym(vp8_rv)
1.23 +global sym(vp8_mbpost_proc_down_mmx) PRIVATE
1.24 +sym(vp8_mbpost_proc_down_mmx):
1.25 + push rbp
1.26 + mov rbp, rsp
1.27 + SHADOW_ARGS_TO_STACK 5
1.28 + GET_GOT rbx
1.29 + push rsi
1.30 + push rdi
1.31 + ; end prolog
1.32 +
1.33 + ALIGN_STACK 16, rax
1.34 + sub rsp, 136
1.35 +
1.36 + ; unsigned char d[16][8] at [rsp]
1.37 + ; create flimit2 at [rsp+128]
1.38 + mov eax, dword ptr arg(4) ;flimit
1.39 + mov [rsp+128], eax
1.40 + mov [rsp+128+4], eax
1.41 +%define flimit2 [rsp+128]
1.42 +
1.43 +%if ABI_IS_32BIT=0
1.44 + lea r8, [GLOBAL(sym(vp8_rv))]
1.45 +%endif
1.46 +
1.47 + ;rows +=8;
1.48 + add dword ptr arg(2), 8
1.49 +
1.50 + ;for(c=0; c<cols; c+=4)
1.51 +.loop_col:
1.52 + mov rsi, arg(0) ;s
1.53 + pxor mm0, mm0 ;
1.54 +
1.55 + movsxd rax, dword ptr arg(1) ;pitch ;
1.56 +
1.57 + ; this copies the last row down into the border 8 rows
1.58 + mov rdi, rsi
1.59 + mov rdx, arg(2)
1.60 + sub rdx, 9
1.61 + imul rdx, rax
1.62 + lea rdi, [rdi+rdx]
1.63 + movq mm1, QWORD ptr[rdi] ; first row
1.64 + mov rcx, 8
1.65 +.init_borderd ; initialize borders
1.66 + lea rdi, [rdi + rax]
1.67 + movq [rdi], mm1
1.68 +
1.69 + dec rcx
1.70 + jne .init_borderd
1.71 +
1.72 + neg rax ; rax = -pitch
1.73 +
1.74 + ; this copies the first row up into the border 8 rows
1.75 + mov rdi, rsi
1.76 + movq mm1, QWORD ptr[rdi] ; first row
1.77 + mov rcx, 8
1.78 +.init_border ; initialize borders
1.79 + lea rdi, [rdi + rax]
1.80 + movq [rdi], mm1
1.81 +
1.82 + dec rcx
1.83 + jne .init_border
1.84 +
1.85 +
1.86 + lea rsi, [rsi + rax*8]; ; rdi = s[-pitch*8]
1.87 + neg rax
1.88 +
1.89 +
1.90 + pxor mm5, mm5
1.91 + pxor mm6, mm6 ;
1.92 +
1.93 + pxor mm7, mm7 ;
1.94 + mov rdi, rsi
1.95 +
1.96 + mov rcx, 15 ;
1.97 +
1.98 +.loop_initvar:
1.99 + movd mm1, DWORD PTR [rdi];
1.100 + punpcklbw mm1, mm0 ;
1.101 +
1.102 + paddw mm5, mm1 ;
1.103 + pmullw mm1, mm1 ;
1.104 +
1.105 + movq mm2, mm1 ;
1.106 + punpcklwd mm1, mm0 ;
1.107 +
1.108 + punpckhwd mm2, mm0 ;
1.109 + paddd mm6, mm1 ;
1.110 +
1.111 + paddd mm7, mm2 ;
1.112 + lea rdi, [rdi+rax] ;
1.113 +
1.114 + dec rcx
1.115 + jne .loop_initvar
1.116 + ;save the var and sum
1.117 + xor rdx, rdx
1.118 +.loop_row:
1.119 + movd mm1, DWORD PTR [rsi] ; [s-pitch*8]
1.120 + movd mm2, DWORD PTR [rdi] ; [s+pitch*7]
1.121 +
1.122 + punpcklbw mm1, mm0
1.123 + punpcklbw mm2, mm0
1.124 +
1.125 + paddw mm5, mm2
1.126 + psubw mm5, mm1
1.127 +
1.128 + pmullw mm2, mm2
1.129 + movq mm4, mm2
1.130 +
1.131 + punpcklwd mm2, mm0
1.132 + punpckhwd mm4, mm0
1.133 +
1.134 + paddd mm6, mm2
1.135 + paddd mm7, mm4
1.136 +
1.137 + pmullw mm1, mm1
1.138 + movq mm2, mm1
1.139 +
1.140 + punpcklwd mm1, mm0
1.141 + psubd mm6, mm1
1.142 +
1.143 + punpckhwd mm2, mm0
1.144 + psubd mm7, mm2
1.145 +
1.146 +
1.147 + movq mm3, mm6
1.148 + pslld mm3, 4
1.149 +
1.150 + psubd mm3, mm6
1.151 + movq mm1, mm5
1.152 +
1.153 + movq mm4, mm5
1.154 + pmullw mm1, mm1
1.155 +
1.156 + pmulhw mm4, mm4
1.157 + movq mm2, mm1
1.158 +
1.159 + punpcklwd mm1, mm4
1.160 + punpckhwd mm2, mm4
1.161 +
1.162 + movq mm4, mm7
1.163 + pslld mm4, 4
1.164 +
1.165 + psubd mm4, mm7
1.166 +
1.167 + psubd mm3, mm1
1.168 + psubd mm4, mm2
1.169 +
1.170 + psubd mm3, flimit2
1.171 + psubd mm4, flimit2
1.172 +
1.173 + psrad mm3, 31
1.174 + psrad mm4, 31
1.175 +
1.176 + packssdw mm3, mm4
1.177 + packsswb mm3, mm0
1.178 +
1.179 + movd mm1, DWORD PTR [rsi+rax*8]
1.180 +
1.181 + movq mm2, mm1
1.182 + punpcklbw mm1, mm0
1.183 +
1.184 + paddw mm1, mm5
1.185 + mov rcx, rdx
1.186 +
1.187 + and rcx, 127
1.188 +%if ABI_IS_32BIT=1 && CONFIG_PIC=1
1.189 + push rax
1.190 + lea rax, [GLOBAL(sym(vp8_rv))]
1.191 + movq mm4, [rax + rcx*2] ;vp8_rv[rcx*2]
1.192 + pop rax
1.193 +%elif ABI_IS_32BIT=0
1.194 + movq mm4, [r8 + rcx*2] ;vp8_rv[rcx*2]
1.195 +%else
1.196 + movq mm4, [sym(vp8_rv) + rcx*2]
1.197 +%endif
1.198 + paddw mm1, mm4
1.199 + psraw mm1, 4
1.200 +
1.201 + packuswb mm1, mm0
1.202 + pand mm1, mm3
1.203 +
1.204 + pandn mm3, mm2
1.205 + por mm1, mm3
1.206 +
1.207 + and rcx, 15
1.208 + movd DWORD PTR [rsp+rcx*4], mm1 ;d[rcx*4]
1.209 +
1.210 + mov rcx, rdx
1.211 + sub rcx, 8
1.212 +
1.213 + and rcx, 15
1.214 + movd mm1, DWORD PTR [rsp+rcx*4] ;d[rcx*4]
1.215 +
1.216 + movd [rsi], mm1
1.217 + lea rsi, [rsi+rax]
1.218 +
1.219 + lea rdi, [rdi+rax]
1.220 + add rdx, 1
1.221 +
1.222 + cmp edx, dword arg(2) ;rows
1.223 + jl .loop_row
1.224 +
1.225 +
1.226 + add dword arg(0), 4 ; s += 4
1.227 + sub dword arg(3), 4 ; cols -= 4
1.228 + cmp dword arg(3), 0
1.229 + jg .loop_col
1.230 +
1.231 + add rsp, 136
1.232 + pop rsp
1.233 +
1.234 + ; begin epilog
1.235 + pop rdi
1.236 + pop rsi
1.237 + RESTORE_GOT
1.238 + UNSHADOW_ARGS
1.239 + pop rbp
1.240 + ret
1.241 +%undef flimit2
1.242 +
1.243 +
1.244 +;void vp8_plane_add_noise_mmx (unsigned char *Start, unsigned char *noise,
1.245 +; unsigned char blackclamp[16],
1.246 +; unsigned char whiteclamp[16],
1.247 +; unsigned char bothclamp[16],
1.248 +; unsigned int Width, unsigned int Height, int Pitch)
1.249 +extern sym(rand)
1.250 +global sym(vp8_plane_add_noise_mmx) PRIVATE
1.251 +sym(vp8_plane_add_noise_mmx):
1.252 + push rbp
1.253 + mov rbp, rsp
1.254 + SHADOW_ARGS_TO_STACK 8
1.255 + GET_GOT rbx
1.256 + push rsi
1.257 + push rdi
1.258 + ; end prolog
1.259 +
1.260 +.addnoise_loop:
1.261 + call sym(rand) WRT_PLT
1.262 + mov rcx, arg(1) ;noise
1.263 + and rax, 0xff
1.264 + add rcx, rax
1.265 +
1.266 + ; we rely on the fact that the clamping vectors are stored contiguously
1.267 + ; in black/white/both order. Note that we have to reload this here because
1.268 + ; rdx could be trashed by rand()
1.269 + mov rdx, arg(2) ; blackclamp
1.270 +
1.271 +
1.272 + mov rdi, rcx
1.273 + movsxd rcx, dword arg(5) ;[Width]
1.274 + mov rsi, arg(0) ;Pos
1.275 + xor rax,rax
1.276 +
1.277 +.addnoise_nextset:
1.278 + movq mm1,[rsi+rax] ; get the source
1.279 +
1.280 + psubusb mm1, [rdx] ;blackclamp ; clamp both sides so we don't outrange adding noise
1.281 + paddusb mm1, [rdx+32] ;bothclamp
1.282 + psubusb mm1, [rdx+16] ;whiteclamp
1.283 +
1.284 + movq mm2,[rdi+rax] ; get the noise for this line
1.285 + paddb mm1,mm2 ; add it in
1.286 + movq [rsi+rax],mm1 ; store the result
1.287 +
1.288 + add rax,8 ; move to the next line
1.289 +
1.290 + cmp rax, rcx
1.291 + jl .addnoise_nextset
1.292 +
1.293 + movsxd rax, dword arg(7) ; Pitch
1.294 + add arg(0), rax ; Start += Pitch
1.295 + sub dword arg(6), 1 ; Height -= 1
1.296 + jg .addnoise_loop
1.297 +
1.298 + ; begin epilog
1.299 + pop rdi
1.300 + pop rsi
1.301 + RESTORE_GOT
1.302 + UNSHADOW_ARGS
1.303 + pop rbp
1.304 + ret
1.305 +
1.306 +
1.307 +SECTION_RODATA
1.308 +align 16
1.309 +Blur:
1.310 + times 16 dw 16
1.311 + times 8 dw 64
1.312 + times 16 dw 16
1.313 + times 8 dw 0
1.314 +
1.315 +rd:
1.316 + times 4 dw 0x40
