1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/encoder/x86/vp9_sad_mmx.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,427 @@
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 +global sym(vp9_sad16x16_mmx) PRIVATE
1.18 +global sym(vp9_sad8x16_mmx) PRIVATE
1.19 +global sym(vp9_sad8x8_mmx) PRIVATE
1.20 +global sym(vp9_sad4x4_mmx) PRIVATE
1.21 +global sym(vp9_sad16x8_mmx) PRIVATE
1.22 +
1.23 +;unsigned int vp9_sad16x16_mmx(
1.24 +; unsigned char *src_ptr,
1.25 +; int src_stride,
1.26 +; unsigned char *ref_ptr,
1.27 +; int ref_stride)
1.28 +sym(vp9_sad16x16_mmx):
1.29 + push rbp
1.30 + mov rbp, rsp
1.31 + SHADOW_ARGS_TO_STACK 4
1.32 + push rsi
1.33 + push rdi
1.34 + ; end prolog
1.35 +
1.36 + mov rsi, arg(0) ;src_ptr
1.37 + mov rdi, arg(2) ;ref_ptr
1.38 +
1.39 + movsxd rax, dword ptr arg(1) ;src_stride
1.40 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.41 +
1.42 + lea rcx, [rsi+rax*8]
1.43 +
1.44 + lea rcx, [rcx+rax*8]
1.45 + pxor mm7, mm7
1.46 +
1.47 + pxor mm6, mm6
1.48 +
1.49 +.x16x16sad_mmx_loop:
1.50 +
1.51 + movq mm0, QWORD PTR [rsi]
1.52 + movq mm2, QWORD PTR [rsi+8]
1.53 +
1.54 + movq mm1, QWORD PTR [rdi]
1.55 + movq mm3, QWORD PTR [rdi+8]
1.56 +
1.57 + movq mm4, mm0
1.58 + movq mm5, mm2
1.59 +
1.60 + psubusb mm0, mm1
1.61 + psubusb mm1, mm4
1.62 +
1.63 + psubusb mm2, mm3
1.64 + psubusb mm3, mm5
1.65 +
1.66 + por mm0, mm1
1.67 + por mm2, mm3
1.68 +
1.69 + movq mm1, mm0
1.70 + movq mm3, mm2
1.71 +
1.72 + punpcklbw mm0, mm6
1.73 + punpcklbw mm2, mm6
1.74 +
1.75 + punpckhbw mm1, mm6
1.76 + punpckhbw mm3, mm6
1.77 +
1.78 + paddw mm0, mm2
1.79 + paddw mm1, mm3
1.80 +
1.81 +
1.82 + lea rsi, [rsi+rax]
1.83 + add rdi, rdx
1.84 +
1.85 + paddw mm7, mm0
1.86 + paddw mm7, mm1
1.87 +
1.88 + cmp rsi, rcx
1.89 + jne .x16x16sad_mmx_loop
1.90 +
1.91 +
1.92 + movq mm0, mm7
1.93 +
1.94 + punpcklwd mm0, mm6
1.95 + punpckhwd mm7, mm6
1.96 +
1.97 + paddw mm0, mm7
1.98 + movq mm7, mm0
1.99 +
1.100 +
1.101 + psrlq mm0, 32
1.102 + paddw mm7, mm0
1.103 +
1.104 + movq rax, mm7
1.105 +
1.106 + pop rdi
1.107 + pop rsi
1.108 + mov rsp, rbp
1.109 + ; begin epilog
1.110 + UNSHADOW_ARGS
1.111 + pop rbp
1.112 + ret
1.113 +
1.114 +
1.115 +;unsigned int vp9_sad8x16_mmx(
1.116 +; unsigned char *src_ptr,
1.117 +; int src_stride,
1.118 +; unsigned char *ref_ptr,
1.119 +; int ref_stride)
1.120 +sym(vp9_sad8x16_mmx):
1.121 + push rbp
1.122 + mov rbp, rsp
1.123 + SHADOW_ARGS_TO_STACK 4
1.124 + push rsi
1.125 + push rdi
1.126 + ; end prolog
1.127 +
1.128 + mov rsi, arg(0) ;src_ptr
1.129 + mov rdi, arg(2) ;ref_ptr
1.130 +
1.131 + movsxd rax, dword ptr arg(1) ;src_stride
1.132 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.133 +
1.134 + lea rcx, [rsi+rax*8]
1.135 +
1.136 + lea rcx, [rcx+rax*8]
1.137 + pxor mm7, mm7
1.138 +
1.139 + pxor mm6, mm6
1.140 +
1.141 +.x8x16sad_mmx_loop:
1.142 +
1.143 + movq mm0, QWORD PTR [rsi]
1.144 + movq mm1, QWORD PTR [rdi]
1.145 +
1.146 + movq mm2, mm0
1.147 + psubusb mm0, mm1
1.148 +
1.149 + psubusb mm1, mm2
1.150 + por mm0, mm1
1.151 +
1.152 + movq mm2, mm0
1.153 + punpcklbw mm0, mm6
1.154 +
1.155 + punpckhbw mm2, mm6
1.156 + lea rsi, [rsi+rax]
1.157 +
1.158 + add rdi, rdx
1.159 + paddw mm7, mm0
1.160 +
1.161 + paddw mm7, mm2
1.162 + cmp rsi, rcx
1.163 +
1.164 + jne .x8x16sad_mmx_loop
1.165 +
1.166 + movq mm0, mm7
1.167 + punpcklwd mm0, mm6
1.168 +
1.169 + punpckhwd mm7, mm6
1.170 + paddw mm0, mm7
1.171 +
1.172 + movq mm7, mm0
1.173 + psrlq mm0, 32
1.174 +
1.175 + paddw mm7, mm0
1.176 + movq rax, mm7
1.177 +
1.178 + pop rdi
1.179 + pop rsi
1.180 + mov rsp, rbp
1.181 + ; begin epilog
1.182 + UNSHADOW_ARGS
1.183 + pop rbp
1.184 + ret
1.185 +
1.186 +
1.187 +;unsigned int vp9_sad8x8_mmx(
1.188 +; unsigned char *src_ptr,
1.189 +; int src_stride,
1.190 +; unsigned char *ref_ptr,
1.191 +; int ref_stride)
1.192 +sym(vp9_sad8x8_mmx):
1.193 + push rbp
1.194 + mov rbp, rsp
1.195 + SHADOW_ARGS_TO_STACK 4
1.196 + push rsi
1.197 + push rdi
1.198 + ; end prolog
1.199 +
1.200 + mov rsi, arg(0) ;src_ptr
1.201 + mov rdi, arg(2) ;ref_ptr
1.202 +
1.203 + movsxd rax, dword ptr arg(1) ;src_stride
1.204 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.205 +
1.206 + lea rcx, [rsi+rax*8]
1.207 + pxor mm7, mm7
1.208 +
1.209 + pxor mm6, mm6
1.210 +
1.211 +.x8x8sad_mmx_loop:
1.212 +
1.213 + movq mm0, QWORD PTR [rsi]
1.214 + movq mm1, QWORD PTR [rdi]
1.215 +
1.216 + movq mm2, mm0
1.217 + psubusb mm0, mm1
1.218 +
1.219 + psubusb mm1, mm2
1.220 + por mm0, mm1
1.221 +
1.222 + movq mm2, mm0
1.223 + punpcklbw mm0, mm6
1.224 +
1.225 + punpckhbw mm2, mm6
1.226 + paddw mm0, mm2
1.227 +
1.228 + lea rsi, [rsi+rax]
1.229 + add rdi, rdx
1.230 +
1.231 + paddw mm7, mm0
1.232 + cmp rsi, rcx
1.233 +
1.234 + jne .x8x8sad_mmx_loop
1.235 +
1.236 + movq mm0, mm7
1.237 + punpcklwd mm0, mm6
1.238 +
1.239 + punpckhwd mm7, mm6
1.240 + paddw mm0, mm7
1.241 +
1.242 + movq mm7, mm0
1.243 + psrlq mm0, 32
1.244 +
1.245 + paddw mm7, mm0
1.246 + movq rax, mm7
1.247 +
1.248 + pop rdi
1.249 + pop rsi
1.250 + mov rsp, rbp
1.251 + ; begin epilog
1.252 + UNSHADOW_ARGS
1.253 + pop rbp
1.254 + ret
1.255 +
1.256 +
1.257 +;unsigned int vp9_sad4x4_mmx(
1.258 +; unsigned char *src_ptr,
1.259 +; int src_stride,
1.260 +; unsigned char *ref_ptr,
1.261 +; int ref_stride)
1.262 +sym(vp9_sad4x4_mmx):
1.263 + push rbp
1.264 + mov rbp, rsp
1.265 + SHADOW_ARGS_TO_STACK 4
1.266 + push rsi
1.267 + push rdi
1.268 + ; end prolog
1.269 +
1.270 + mov rsi, arg(0) ;src_ptr
1.271 + mov rdi, arg(2) ;ref_ptr
1.272 +
1.273 + movsxd rax, dword ptr arg(1) ;src_stride
1.274 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.275 +
1.276 + movd mm0, DWORD PTR [rsi]
1.277 + movd mm1, DWORD PTR [rdi]
1.278 +
1.279 + movd mm2, DWORD PTR [rsi+rax]
1.280 + movd mm3, DWORD PTR [rdi+rdx]
1.281 +
1.282 + punpcklbw mm0, mm2
1.283 + punpcklbw mm1, mm3
1.284 +
1.285 + movq mm2, mm0
1.286 + psubusb mm0, mm1
1.287 +
1.288 + psubusb mm1, mm2
1.289 + por mm0, mm1
1.290 +
1.291 + movq mm2, mm0
1.292 + pxor mm3, mm3
1.293 +
1.294 + punpcklbw mm0, mm3
1.295 + punpckhbw mm2, mm3
1.296 +
1.297 + paddw mm0, mm2
1.298 +
1.299 + lea rsi, [rsi+rax*2]
1.300 + lea rdi, [rdi+rdx*2]
1.301 +
1.302 + movd mm4, DWORD PTR [rsi]
1.303 + movd mm5, DWORD PTR [rdi]
1.304 +
1.305 + movd mm6, DWORD PTR [rsi+rax]
1.306 + movd mm7, DWORD PTR [rdi+rdx]
1.307 +
1.308 + punpcklbw mm4, mm6
1.309 + punpcklbw mm5, mm7
1.310 +
1.311 + movq mm6, mm4
1.312 + psubusb mm4, mm5
1.313 +
1.314 + psubusb mm5, mm6
1.315 + por mm4, mm5
1.316 +
1.317 + movq mm5, mm4
1.318 + punpcklbw mm4, mm3
1.319 +
1.320 + punpckhbw mm5, mm3
1.321 + paddw mm4, mm5
1.322 +
1.323 + paddw mm0, mm4
1.324 + movq mm1, mm0
1.325 +
1.326 + punpcklwd mm0, mm3
1.327 + punpckhwd mm1, mm3
1.328 +
1.329 + paddw mm0, mm1
1.330 + movq mm1, mm0
1.331 +
1.332 + psrlq mm0, 32
1.333 + paddw mm0, mm1
1.334 +
1.335 + movq rax, mm0
1.336 +
1.337 + pop rdi
1.338 + pop rsi
1.339 + mov rsp, rbp
1.340 + ; begin epilog
1.341 + UNSHADOW_ARGS
1.342 + pop rbp
1.343 + ret
1.344 +
1.345 +
1.346 +;unsigned int vp9_sad16x8_mmx(
1.347 +; unsigned char *src_ptr,
1.348 +; int src_stride,
1.349 +; unsigned char *ref_ptr,
1.350 +; int ref_stride)
1.351 +sym(vp9_sad16x8_mmx):
1.352 + push rbp
1.353 + mov rbp, rsp
1.354 + SHADOW_ARGS_TO_STACK 4
1.355 + push rsi
1.356 + push rdi
1.357 + ; end prolog
1.358 +
1.359 + mov rsi, arg(0) ;src_ptr
1.360 + mov rdi, arg(2) ;ref_ptr
1.361 +
1.362 + movsxd rax, dword ptr arg(1) ;src_stride
1.363 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.364 +
1.365 + lea rcx, [rsi+rax*8]
1.366 + pxor mm7, mm7
1.367 +
1.368 + pxor mm6, mm6
1.369 +
1.370 +.x16x8sad_mmx_loop:
1.371 +
1.372 + movq mm0, [rsi]
1.373 + movq mm1, [rdi]
1.374 +
1.375 + movq mm2, [rsi+8]
1.376 + movq mm3, [rdi+8]
1.377 +
1.378 + movq mm4, mm0
1.379 + movq mm5, mm2
1.380 +
1.381 + psubusb mm0, mm1
1.382 + psubusb mm1, mm4
1.383 +
1.384 + psubusb mm2, mm3
1.385 + psubusb mm3, mm5
1.386 +
1.387 + por mm0, mm1
1.388 + por mm2, mm3
1.389 +
1.390 + movq mm1, mm0
1.391 + movq mm3, mm2
1.392 +
1.393 + punpcklbw mm0, mm6
1.394 + punpckhbw mm1, mm6
1.395 +
1.396 + punpcklbw mm2, mm6
1.397 + punpckhbw mm3, mm6
1.398 +
1.399 +
1.400 + paddw mm0, mm2
1.401 + paddw mm1, mm3
1.402 +
1.403 + paddw mm0, mm1
1.404 + lea rsi, [rsi+rax]
1.405 +
1.406 + add rdi, rdx
1.407 + paddw mm7, mm0
1.408 +
1.409 + cmp rsi, rcx
1.410 + jne .x16x8sad_mmx_loop
1.411 +
1.412 + movq mm0, mm7
1.413 + punpcklwd mm0, mm6
1.414 +
1.415 + punpckhwd mm7, mm6
1.416 + paddw mm0, mm7
1.417 +
1.418 + movq mm7, mm0
1.419 + psrlq mm0, 32
1.420 +
1.421 + paddw mm7, mm0
1.422 + movq rax, mm7
1.423 +
1.424 + pop rdi
1.425 + pop rsi
1.426 + mov rsp, rbp
1.427 + ; begin epilog
1.428 + UNSHADOW_ARGS
1.429 + pop rbp
1.430 + ret
