1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/encoder/x86/vp9_sad_sse4.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,359 @@
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 PROCESS_16X2X8 1
1.18 +%if %1
1.19 + movdqa xmm0, XMMWORD PTR [rsi]
1.20 + movq xmm1, MMWORD PTR [rdi]
1.21 + movq xmm3, MMWORD PTR [rdi+8]
1.22 + movq xmm2, MMWORD PTR [rdi+16]
1.23 + punpcklqdq xmm1, xmm3
1.24 + punpcklqdq xmm3, xmm2
1.25 +
1.26 + movdqa xmm2, xmm1
1.27 + mpsadbw xmm1, xmm0, 0x0
1.28 + mpsadbw xmm2, xmm0, 0x5
1.29 +
1.30 + psrldq xmm0, 8
1.31 +
1.32 + movdqa xmm4, xmm3
1.33 + mpsadbw xmm3, xmm0, 0x0
1.34 + mpsadbw xmm4, xmm0, 0x5
1.35 +
1.36 + paddw xmm1, xmm2
1.37 + paddw xmm1, xmm3
1.38 + paddw xmm1, xmm4
1.39 +%else
1.40 + movdqa xmm0, XMMWORD PTR [rsi]
1.41 + movq xmm5, MMWORD PTR [rdi]
1.42 + movq xmm3, MMWORD PTR [rdi+8]
1.43 + movq xmm2, MMWORD PTR [rdi+16]
1.44 + punpcklqdq xmm5, xmm3
1.45 + punpcklqdq xmm3, xmm2
1.46 +
1.47 + movdqa xmm2, xmm5
1.48 + mpsadbw xmm5, xmm0, 0x0
1.49 + mpsadbw xmm2, xmm0, 0x5
1.50 +
1.51 + psrldq xmm0, 8
1.52 +
1.53 + movdqa xmm4, xmm3
1.54 + mpsadbw xmm3, xmm0, 0x0
1.55 + mpsadbw xmm4, xmm0, 0x5
1.56 +
1.57 + paddw xmm5, xmm2
1.58 + paddw xmm5, xmm3
1.59 + paddw xmm5, xmm4
1.60 +
1.61 + paddw xmm1, xmm5
1.62 +%endif
1.63 + movdqa xmm0, XMMWORD PTR [rsi + rax]
1.64 + movq xmm5, MMWORD PTR [rdi+ rdx]
1.65 + movq xmm3, MMWORD PTR [rdi+ rdx+8]
1.66 + movq xmm2, MMWORD PTR [rdi+ rdx+16]
1.67 + punpcklqdq xmm5, xmm3
1.68 + punpcklqdq xmm3, xmm2
1.69 +
1.70 + lea rsi, [rsi+rax*2]
1.71 + lea rdi, [rdi+rdx*2]
1.72 +
1.73 + movdqa xmm2, xmm5
1.74 + mpsadbw xmm5, xmm0, 0x0
1.75 + mpsadbw xmm2, xmm0, 0x5
1.76 +
1.77 + psrldq xmm0, 8
1.78 + movdqa xmm4, xmm3
1.79 + mpsadbw xmm3, xmm0, 0x0
1.80 + mpsadbw xmm4, xmm0, 0x5
1.81 +
1.82 + paddw xmm5, xmm2
1.83 + paddw xmm5, xmm3
1.84 + paddw xmm5, xmm4
1.85 +
1.86 + paddw xmm1, xmm5
1.87 +%endmacro
1.88 +
1.89 +%macro PROCESS_8X2X8 1
1.90 +%if %1
1.91 + movq xmm0, MMWORD PTR [rsi]
1.92 + movq xmm1, MMWORD PTR [rdi]
1.93 + movq xmm3, MMWORD PTR [rdi+8]
1.94 + punpcklqdq xmm1, xmm3
1.95 +
1.96 + movdqa xmm2, xmm1
1.97 + mpsadbw xmm1, xmm0, 0x0
1.98 + mpsadbw xmm2, xmm0, 0x5
1.99 + paddw xmm1, xmm2
1.100 +%else
1.101 + movq xmm0, MMWORD PTR [rsi]
1.102 + movq xmm5, MMWORD PTR [rdi]
1.103 + movq xmm3, MMWORD PTR [rdi+8]
1.104 + punpcklqdq xmm5, xmm3
1.105 +
1.106 + movdqa xmm2, xmm5
1.107 + mpsadbw xmm5, xmm0, 0x0
1.108 + mpsadbw xmm2, xmm0, 0x5
1.109 + paddw xmm5, xmm2
1.110 +
1.111 + paddw xmm1, xmm5
1.112 +%endif
1.113 + movq xmm0, MMWORD PTR [rsi + rax]
1.114 + movq xmm5, MMWORD PTR [rdi+ rdx]
1.115 + movq xmm3, MMWORD PTR [rdi+ rdx+8]
1.116 + punpcklqdq xmm5, xmm3
1.117 +
1.118 + lea rsi, [rsi+rax*2]
1.119 + lea rdi, [rdi+rdx*2]
1.120 +
1.121 + movdqa xmm2, xmm5
1.122 + mpsadbw xmm5, xmm0, 0x0
1.123 + mpsadbw xmm2, xmm0, 0x5
1.124 + paddw xmm5, xmm2
1.125 +
1.126 + paddw xmm1, xmm5
1.127 +%endmacro
1.128 +
1.129 +%macro PROCESS_4X2X8 1
1.130 +%if %1
1.131 + movd xmm0, [rsi]
1.132 + movq xmm1, MMWORD PTR [rdi]
1.133 + movq xmm3, MMWORD PTR [rdi+8]
1.134 + punpcklqdq xmm1, xmm3
1.135 +
1.136 + mpsadbw xmm1, xmm0, 0x0
1.137 +%else
1.138 + movd xmm0, [rsi]
1.139 + movq xmm5, MMWORD PTR [rdi]
1.140 + movq xmm3, MMWORD PTR [rdi+8]
1.141 + punpcklqdq xmm5, xmm3
1.142 +
1.143 + mpsadbw xmm5, xmm0, 0x0
1.144 +
1.145 + paddw xmm1, xmm5
1.146 +%endif
1.147 + movd xmm0, [rsi + rax]
1.148 + movq xmm5, MMWORD PTR [rdi+ rdx]
1.149 + movq xmm3, MMWORD PTR [rdi+ rdx+8]
1.150 + punpcklqdq xmm5, xmm3
1.151 +
1.152 + lea rsi, [rsi+rax*2]
1.153 + lea rdi, [rdi+rdx*2]
1.154 +
1.155 + mpsadbw xmm5, xmm0, 0x0
1.156 +
1.157 + paddw xmm1, xmm5
1.158 +%endmacro
1.159 +
1.160 +%macro WRITE_AS_INTS 0
1.161 + mov rdi, arg(4) ;Results
1.162 + pxor xmm0, xmm0
1.163 + movdqa xmm2, xmm1
1.164 + punpcklwd xmm1, xmm0
1.165 + punpckhwd xmm2, xmm0
1.166 +
1.167 + movdqa [rdi], xmm1
1.168 + movdqa [rdi + 16], xmm2
1.169 +%endmacro
1.170 +
1.171 +;void vp9_sad16x16x8_sse4(
1.172 +; const unsigned char *src_ptr,
1.173 +; int src_stride,
1.174 +; const unsigned char *ref_ptr,
1.175 +; int ref_stride,
1.176 +; unsigned short *sad_array);
1.177 +global sym(vp9_sad16x16x8_sse4) PRIVATE
1.178 +sym(vp9_sad16x16x8_sse4):
1.179 + push rbp
1.180 + mov rbp, rsp
1.181 + SHADOW_ARGS_TO_STACK 5
1.182 + push rsi
1.183 + push rdi
1.184 + ; end prolog
1.185 +
1.186 + mov rsi, arg(0) ;src_ptr
1.187 + mov rdi, arg(2) ;ref_ptr
1.188 +
1.189 + movsxd rax, dword ptr arg(1) ;src_stride
1.190 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.191 +
1.192 + PROCESS_16X2X8 1
1.193 + PROCESS_16X2X8 0
1.194 + PROCESS_16X2X8 0
1.195 + PROCESS_16X2X8 0
1.196 + PROCESS_16X2X8 0
1.197 + PROCESS_16X2X8 0
1.198 + PROCESS_16X2X8 0
1.199 + PROCESS_16X2X8 0
1.200 +
1.201 + WRITE_AS_INTS
1.202 +
1.203 + ; begin epilog
1.204 + pop rdi
1.205 + pop rsi
1.206 + UNSHADOW_ARGS
1.207 + pop rbp
1.208 + ret
1.209 +
1.210 +
1.211 +;void vp9_sad16x8x8_sse4(
1.212 +; const unsigned char *src_ptr,
1.213 +; int src_stride,
1.214 +; const unsigned char *ref_ptr,
1.215 +; int ref_stride,
1.216 +; unsigned short *sad_array
1.217 +;);
1.218 +global sym(vp9_sad16x8x8_sse4) PRIVATE
1.219 +sym(vp9_sad16x8x8_sse4):
1.220 + push rbp
1.221 + mov rbp, rsp
1.222 + SHADOW_ARGS_TO_STACK 5
1.223 + push rsi
1.224 + push rdi
1.225 + ; end prolog
1.226 +
1.227 + mov rsi, arg(0) ;src_ptr
1.228 + mov rdi, arg(2) ;ref_ptr
1.229 +
1.230 + movsxd rax, dword ptr arg(1) ;src_stride
1.231 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.232 +
1.233 + PROCESS_16X2X8 1
1.234 + PROCESS_16X2X8 0
1.235 + PROCESS_16X2X8 0
1.236 + PROCESS_16X2X8 0
1.237 +
1.238 + WRITE_AS_INTS
1.239 +
1.240 + ; begin epilog
1.241 + pop rdi
1.242 + pop rsi
1.243 + UNSHADOW_ARGS
1.244 + pop rbp
1.245 + ret
1.246 +
1.247 +
1.248 +;void vp9_sad8x8x8_sse4(
1.249 +; const unsigned char *src_ptr,
1.250 +; int src_stride,
1.251 +; const unsigned char *ref_ptr,
1.252 +; int ref_stride,
1.253 +; unsigned short *sad_array
1.254 +;);
1.255 +global sym(vp9_sad8x8x8_sse4) PRIVATE
1.256 +sym(vp9_sad8x8x8_sse4):
1.257 + push rbp
1.258 + mov rbp, rsp
1.259 + SHADOW_ARGS_TO_STACK 5
1.260 + push rsi
1.261 + push rdi
1.262 + ; end prolog
1.263 +
1.264 + mov rsi, arg(0) ;src_ptr
1.265 + mov rdi, arg(2) ;ref_ptr
1.266 +
1.267 + movsxd rax, dword ptr arg(1) ;src_stride
1.268 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.269 +
1.270 + PROCESS_8X2X8 1
1.271 + PROCESS_8X2X8 0
1.272 + PROCESS_8X2X8 0
1.273 + PROCESS_8X2X8 0
1.274 +
1.275 + WRITE_AS_INTS
1.276 +
1.277 + ; begin epilog
1.278 + pop rdi
1.279 + pop rsi
1.280 + UNSHADOW_ARGS
1.281 + pop rbp
1.282 + ret
1.283 +
1.284 +
1.285 +;void vp9_sad8x16x8_sse4(
1.286 +; const unsigned char *src_ptr,
1.287 +; int src_stride,
1.288 +; const unsigned char *ref_ptr,
1.289 +; int ref_stride,
1.290 +; unsigned short *sad_array
1.291 +;);
1.292 +global sym(vp9_sad8x16x8_sse4) PRIVATE
1.293 +sym(vp9_sad8x16x8_sse4):
1.294 + push rbp
1.295 + mov rbp, rsp
1.296 + SHADOW_ARGS_TO_STACK 5
1.297 + push rsi
1.298 + push rdi
1.299 + ; end prolog
1.300 +
1.301 + mov rsi, arg(0) ;src_ptr
1.302 + mov rdi, arg(2) ;ref_ptr
1.303 +
1.304 + movsxd rax, dword ptr arg(1) ;src_stride
1.305 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.306 +
1.307 + PROCESS_8X2X8 1
1.308 + PROCESS_8X2X8 0
1.309 + PROCESS_8X2X8 0
1.310 + PROCESS_8X2X8 0
1.311 + PROCESS_8X2X8 0
1.312 + PROCESS_8X2X8 0
1.313 + PROCESS_8X2X8 0
1.314 + PROCESS_8X2X8 0
1.315 +
1.316 + WRITE_AS_INTS
1.317 +
1.318 + ; begin epilog
1.319 + pop rdi
1.320 + pop rsi
1.321 + UNSHADOW_ARGS
1.322 + pop rbp
1.323 + ret
1.324 +
1.325 +
1.326 +;void vp9_sad4x4x8_c(
1.327 +; const unsigned char *src_ptr,
1.328 +; int src_stride,
1.329 +; const unsigned char *ref_ptr,
1.330 +; int ref_stride,
1.331 +; unsigned short *sad_array
1.332 +;);
1.333 +global sym(vp9_sad4x4x8_sse4) PRIVATE
1.334 +sym(vp9_sad4x4x8_sse4):
1.335 + push rbp
1.336 + mov rbp, rsp
1.337 + SHADOW_ARGS_TO_STACK 5
1.338 + push rsi
1.339 + push rdi
1.340 + ; end prolog
1.341 +
1.342 + mov rsi, arg(0) ;src_ptr
1.343 + mov rdi, arg(2) ;ref_ptr
1.344 +
1.345 + movsxd rax, dword ptr arg(1) ;src_stride
1.346 + movsxd rdx, dword ptr arg(3) ;ref_stride
1.347 +
1.348 + PROCESS_4X2X8 1
1.349 + PROCESS_4X2X8 0
1.350 +
1.351 + WRITE_AS_INTS
1.352 +
1.353 + ; begin epilog
1.354 + pop rdi
1.355 + pop rsi
1.356 + UNSHADOW_ARGS
1.357 + pop rbp
1.358 + ret
1.359 +
1.360 +
1.361 +
1.362 +
