1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/common/x86/recon_mmx.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,274 @@
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 +
1.18 +;void copy_mem8x8_mmx(
1.19 +; unsigned char *src,
1.20 +; int src_stride,
1.21 +; unsigned char *dst,
1.22 +; int dst_stride
1.23 +; )
1.24 +global sym(vp8_copy_mem8x8_mmx) PRIVATE
1.25 +sym(vp8_copy_mem8x8_mmx):
1.26 + push rbp
1.27 + mov rbp, rsp
1.28 + SHADOW_ARGS_TO_STACK 4
1.29 + push rsi
1.30 + push rdi
1.31 + ; end prolog
1.32 +
1.33 + mov rsi, arg(0) ;src;
1.34 + movq mm0, [rsi]
1.35 +
1.36 + movsxd rax, dword ptr arg(1) ;src_stride;
1.37 + mov rdi, arg(2) ;dst;
1.38 +
1.39 + movq mm1, [rsi+rax]
1.40 + movq mm2, [rsi+rax*2]
1.41 +
1.42 + movsxd rcx, dword ptr arg(3) ;dst_stride
1.43 + lea rsi, [rsi+rax*2]
1.44 +
1.45 + movq [rdi], mm0
1.46 + add rsi, rax
1.47 +
1.48 + movq [rdi+rcx], mm1
1.49 + movq [rdi+rcx*2], mm2
1.50 +
1.51 +
1.52 + lea rdi, [rdi+rcx*2]
1.53 + movq mm3, [rsi]
1.54 +
1.55 + add rdi, rcx
1.56 + movq mm4, [rsi+rax]
1.57 +
1.58 + movq mm5, [rsi+rax*2]
1.59 + movq [rdi], mm3
1.60 +
1.61 + lea rsi, [rsi+rax*2]
1.62 + movq [rdi+rcx], mm4
1.63 +
1.64 + movq [rdi+rcx*2], mm5
1.65 + lea rdi, [rdi+rcx*2]
1.66 +
1.67 + movq mm0, [rsi+rax]
1.68 + movq mm1, [rsi+rax*2]
1.69 +
1.70 + movq [rdi+rcx], mm0
1.71 + movq [rdi+rcx*2],mm1
1.72 +
1.73 + ; begin epilog
1.74 + pop rdi
1.75 + pop rsi
1.76 + UNSHADOW_ARGS
1.77 + pop rbp
1.78 + ret
1.79 +
1.80 +
1.81 +;void copy_mem8x4_mmx(
1.82 +; unsigned char *src,
1.83 +; int src_stride,
1.84 +; unsigned char *dst,
1.85 +; int dst_stride
1.86 +; )
1.87 +global sym(vp8_copy_mem8x4_mmx) PRIVATE
1.88 +sym(vp8_copy_mem8x4_mmx):
1.89 + push rbp
1.90 + mov rbp, rsp
1.91 + SHADOW_ARGS_TO_STACK 4
1.92 + push rsi
1.93 + push rdi
1.94 + ; end prolog
1.95 +
1.96 + mov rsi, arg(0) ;src;
1.97 + movq mm0, [rsi]
1.98 +
1.99 + movsxd rax, dword ptr arg(1) ;src_stride;
1.100 + mov rdi, arg(2) ;dst;
1.101 +
1.102 + movq mm1, [rsi+rax]
1.103 + movq mm2, [rsi+rax*2]
1.104 +
1.105 + movsxd rcx, dword ptr arg(3) ;dst_stride
1.106 + lea rsi, [rsi+rax*2]
1.107 +
1.108 + movq [rdi], mm0
1.109 + movq [rdi+rcx], mm1
1.110 +
1.111 + movq [rdi+rcx*2], mm2
1.112 + lea rdi, [rdi+rcx*2]
1.113 +
1.114 + movq mm3, [rsi+rax]
1.115 + movq [rdi+rcx], mm3
1.116 +
1.117 + ; begin epilog
1.118 + pop rdi
1.119 + pop rsi
1.120 + UNSHADOW_ARGS
1.121 + pop rbp
1.122 + ret
1.123 +
1.124 +
1.125 +;void copy_mem16x16_mmx(
1.126 +; unsigned char *src,
1.127 +; int src_stride,
1.128 +; unsigned char *dst,
1.129 +; int dst_stride
1.130 +; )
1.131 +global sym(vp8_copy_mem16x16_mmx) PRIVATE
1.132 +sym(vp8_copy_mem16x16_mmx):
1.133 + push rbp
1.134 + mov rbp, rsp
1.135 + SHADOW_ARGS_TO_STACK 4
1.136 + push rsi
1.137 + push rdi
1.138 + ; end prolog
1.139 +
1.140 + mov rsi, arg(0) ;src;
1.141 + movsxd rax, dword ptr arg(1) ;src_stride;
1.142 +
1.143 + mov rdi, arg(2) ;dst;
1.144 + movsxd rcx, dword ptr arg(3) ;dst_stride
1.145 +
1.146 + movq mm0, [rsi]
1.147 + movq mm3, [rsi+8];
1.148 +
1.149 + movq mm1, [rsi+rax]
1.150 + movq mm4, [rsi+rax+8]
1.151 +
1.152 + movq mm2, [rsi+rax*2]
1.153 + movq mm5, [rsi+rax*2+8]
1.154 +
1.155 + lea rsi, [rsi+rax*2]
1.156 + add rsi, rax
1.157 +
1.158 + movq [rdi], mm0
1.159 + movq [rdi+8], mm3
1.160 +
1.161 + movq [rdi+rcx], mm1
1.162 + movq [rdi+rcx+8], mm4
1.163 +
1.164 + movq [rdi+rcx*2], mm2
1.165 + movq [rdi+rcx*2+8], mm5
1.166 +
1.167 + lea rdi, [rdi+rcx*2]
1.168 + add rdi, rcx
1.169 +
1.170 + movq mm0, [rsi]
1.171 + movq mm3, [rsi+8];
1.172 +
1.173 + movq mm1, [rsi+rax]
1.174 + movq mm4, [rsi+rax+8]
1.175 +
1.176 + movq mm2, [rsi+rax*2]
1.177 + movq mm5, [rsi+rax*2+8]
1.178 +
1.179 + lea rsi, [rsi+rax*2]
1.180 + add rsi, rax
1.181 +
1.182 + movq [rdi], mm0
1.183 + movq [rdi+8], mm3
1.184 +
1.185 + movq [rdi+rcx], mm1
1.186 + movq [rdi+rcx+8], mm4
1.187 +
1.188 + movq [rdi+rcx*2], mm2
1.189 + movq [rdi+rcx*2+8], mm5
1.190 +
1.191 + lea rdi, [rdi+rcx*2]
1.192 + add rdi, rcx
1.193 +
1.194 + movq mm0, [rsi]
1.195 + movq mm3, [rsi+8];
1.196 +
1.197 + movq mm1, [rsi+rax]
1.198 + movq mm4, [rsi+rax+8]
1.199 +
1.200 + movq mm2, [rsi+rax*2]
1.201 + movq mm5, [rsi+rax*2+8]
1.202 +
1.203 + lea rsi, [rsi+rax*2]
1.204 + add rsi, rax
1.205 +
1.206 + movq [rdi], mm0
1.207 + movq [rdi+8], mm3
1.208 +
1.209 + movq [rdi+rcx], mm1
1.210 + movq [rdi+rcx+8], mm4
1.211 +
1.212 + movq [rdi+rcx*2], mm2
1.213 + movq [rdi+rcx*2+8], mm5
1.214 +
1.215 + lea rdi, [rdi+rcx*2]
1.216 + add rdi, rcx
1.217 +
1.218 + movq mm0, [rsi]
1.219 + movq mm3, [rsi+8];
1.220 +
1.221 + movq mm1, [rsi+rax]
1.222 + movq mm4, [rsi+rax+8]
1.223 +
1.224 + movq mm2, [rsi+rax*2]
1.225 + movq mm5, [rsi+rax*2+8]
1.226 +
1.227 + lea rsi, [rsi+rax*2]
1.228 + add rsi, rax
1.229 +
1.230 + movq [rdi], mm0
1.231 + movq [rdi+8], mm3
1.232 +
1.233 + movq [rdi+rcx], mm1
1.234 + movq [rdi+rcx+8], mm4
1.235 +
1.236 + movq [rdi+rcx*2], mm2
1.237 + movq [rdi+rcx*2+8], mm5
1.238 +
1.239 + lea rdi, [rdi+rcx*2]
1.240 + add rdi, rcx
1.241 +
1.242 + movq mm0, [rsi]
1.243 + movq mm3, [rsi+8];
1.244 +
1.245 + movq mm1, [rsi+rax]
1.246 + movq mm4, [rsi+rax+8]
1.247 +
1.248 + movq mm2, [rsi+rax*2]
1.249 + movq mm5, [rsi+rax*2+8]
1.250 +
1.251 + lea rsi, [rsi+rax*2]
1.252 + add rsi, rax
1.253 +
1.254 + movq [rdi], mm0
1.255 + movq [rdi+8], mm3
1.256 +
1.257 + movq [rdi+rcx], mm1
1.258 + movq [rdi+rcx+8], mm4
1.259 +
1.260 + movq [rdi+rcx*2], mm2
1.261 + movq [rdi+rcx*2+8], mm5
1.262 +
1.263 + lea rdi, [rdi+rcx*2]
1.264 + add rdi, rcx
1.265 +
1.266 + movq mm0, [rsi]
1.267 + movq mm3, [rsi+8];
1.268 +
1.269 + movq [rdi], mm0
1.270 + movq [rdi+8], mm3
1.271 +
1.272 + ; begin epilog
1.273 + pop rdi
1.274 + pop rsi
1.275 + UNSHADOW_ARGS
1.276 + pop rbp
1.277 + ret
