1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/encoder/x86/vp9_sad_sse3.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,378 @@
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 +%include "vpx_ports/x86_abi_support.asm"
1.15 +
1.16 +%macro STACK_FRAME_CREATE_X3 0
1.17 +%if ABI_IS_32BIT
1.18 + %define src_ptr rsi
1.19 + %define src_stride rax
1.20 + %define ref_ptr rdi
1.21 + %define ref_stride rdx
1.22 + %define end_ptr rcx
1.23 + %define ret_var rbx
1.24 + %define result_ptr arg(4)
1.25 + %define max_err arg(4)
1.26 + %define height dword ptr arg(4)
1.27 + push rbp
1.28 + mov rbp, rsp
1.29 + push rsi
1.30 + push rdi
1.31 + push rbx
1.32 +
1.33 + mov rsi, arg(0) ; src_ptr
1.34 + mov rdi, arg(2) ; ref_ptr
1.35 +
1.36 + movsxd rax, dword ptr arg(1) ; src_stride
1.37 + movsxd rdx, dword ptr arg(3) ; ref_stride
1.38 +%else
1.39 + %if LIBVPX_YASM_WIN64
1.40 + SAVE_XMM 7, u
1.41 + %define src_ptr rcx
1.42 + %define src_stride rdx
1.43 + %define ref_ptr r8
1.44 + %define ref_stride r9
1.45 + %define end_ptr r10
1.46 + %define ret_var r11
1.47 + %define result_ptr [rsp+xmm_stack_space+8+4*8]
1.48 + %define max_err [rsp+xmm_stack_space+8+4*8]
1.49 + %define height dword ptr [rsp+xmm_stack_space+8+4*8]
1.50 + %else
1.51 + %define src_ptr rdi
1.52 + %define src_stride rsi
1.53 + %define ref_ptr rdx
1.54 + %define ref_stride rcx
1.55 + %define end_ptr r9
1.56 + %define ret_var r10
1.57 + %define result_ptr r8
1.58 + %define max_err r8
1.59 + %define height r8
1.60 + %endif
1.61 +%endif
1.62 +
1.63 +%endmacro
1.64 +
1.65 +%macro STACK_FRAME_DESTROY_X3 0
1.66 + %define src_ptr
1.67 + %define src_stride
1.68 + %define ref_ptr
1.69 + %define ref_stride
1.70 + %define end_ptr
1.71 + %define ret_var
1.72 + %define result_ptr
1.73 + %define max_err
1.74 + %define height
1.75 +
1.76 +%if ABI_IS_32BIT
1.77 + pop rbx
1.78 + pop rdi
1.79 + pop rsi
1.80 + pop rbp
1.81 +%else
1.82 + %if LIBVPX_YASM_WIN64
1.83 + RESTORE_XMM
1.84 + %endif
1.85 +%endif
1.86 + ret
1.87 +%endmacro
1.88 +
1.89 +%macro PROCESS_16X2X3 5
1.90 +%if %1==0
1.91 + movdqa xmm0, XMMWORD PTR [%2]
1.92 + lddqu xmm5, XMMWORD PTR [%3]
1.93 + lddqu xmm6, XMMWORD PTR [%3+1]
1.94 + lddqu xmm7, XMMWORD PTR [%3+2]
1.95 +
1.96 + psadbw xmm5, xmm0
1.97 + psadbw xmm6, xmm0
1.98 + psadbw xmm7, xmm0
1.99 +%else
1.100 + movdqa xmm0, XMMWORD PTR [%2]
1.101 + lddqu xmm1, XMMWORD PTR [%3]
1.102 + lddqu xmm2, XMMWORD PTR [%3+1]
1.103 + lddqu xmm3, XMMWORD PTR [%3+2]
1.104 +
1.105 + psadbw xmm1, xmm0
1.106 + psadbw xmm2, xmm0
1.107 + psadbw xmm3, xmm0
1.108 +
1.109 + paddw xmm5, xmm1
1.110 + paddw xmm6, xmm2
1.111 + paddw xmm7, xmm3
1.112 +%endif
1.113 + movdqa xmm0, XMMWORD PTR [%2+%4]
1.114 + lddqu xmm1, XMMWORD PTR [%3+%5]
1.115 + lddqu xmm2, XMMWORD PTR [%3+%5+1]
1.116 + lddqu xmm3, XMMWORD PTR [%3+%5+2]
1.117 +
1.118 +%if %1==0 || %1==1
1.119 + lea %2, [%2+%4*2]
1.120 + lea %3, [%3+%5*2]
1.121 +%endif
1.122 +
1.123 + psadbw xmm1, xmm0
1.124 + psadbw xmm2, xmm0
1.125 + psadbw xmm3, xmm0
1.126 +
1.127 + paddw xmm5, xmm1
1.128 + paddw xmm6, xmm2
1.129 + paddw xmm7, xmm3
1.130 +%endmacro
1.131 +
1.132 +%macro PROCESS_8X2X3 5
1.133 +%if %1==0
1.134 + movq mm0, QWORD PTR [%2]
1.135 + movq mm5, QWORD PTR [%3]
1.136 + movq mm6, QWORD PTR [%3+1]
1.137 + movq mm7, QWORD PTR [%3+2]
1.138 +
1.139 + psadbw mm5, mm0
1.140 + psadbw mm6, mm0
1.141 + psadbw mm7, mm0
1.142 +%else
1.143 + movq mm0, QWORD PTR [%2]
1.144 + movq mm1, QWORD PTR [%3]
1.145 + movq mm2, QWORD PTR [%3+1]
1.146 + movq mm3, QWORD PTR [%3+2]
1.147 +
1.148 + psadbw mm1, mm0
1.149 + psadbw mm2, mm0
1.150 + psadbw mm3, mm0
1.151 +
1.152 + paddw mm5, mm1
1.153 + paddw mm6, mm2
1.154 + paddw mm7, mm3
1.155 +%endif
1.156 + movq mm0, QWORD PTR [%2+%4]
1.157 + movq mm1, QWORD PTR [%3+%5]
1.158 + movq mm2, QWORD PTR [%3+%5+1]
1.159 + movq mm3, QWORD PTR [%3+%5+2]
1.160 +
1.161 +%if %1==0 || %1==1
1.162 + lea %2, [%2+%4*2]
1.163 + lea %3, [%3+%5*2]
1.164 +%endif
1.165 +
1.166 + psadbw mm1, mm0
1.167 + psadbw mm2, mm0
1.168 + psadbw mm3, mm0
1.169 +
1.170 + paddw mm5, mm1
1.171 + paddw mm6, mm2
1.172 + paddw mm7, mm3
1.173 +%endmacro
1.174 +
1.175 +;void int vp9_sad16x16x3_sse3(
1.176 +; unsigned char *src_ptr,
1.177 +; int src_stride,
1.178 +; unsigned char *ref_ptr,
1.179 +; int ref_stride,
1.180 +; int *results)
1.181 +global sym(vp9_sad16x16x3_sse3) PRIVATE
1.182 +sym(vp9_sad16x16x3_sse3):
1.183 +
1.184 + STACK_FRAME_CREATE_X3
1.185 +
1.186 + PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
1.187 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.188 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.189 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.190 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.191 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.192 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.193 + PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
1.194 +
1.195 + mov rcx, result_ptr
1.196 +
1.197 + movq xmm0, xmm5
1.198 + psrldq xmm5, 8
1.199 +
1.200 + paddw xmm0, xmm5
1.201 + movd [rcx], xmm0
1.202 +;-
1.203 + movq xmm0, xmm6
1.204 + psrldq xmm6, 8
1.205 +
1.206 + paddw xmm0, xmm6
1.207 + movd [rcx+4], xmm0
1.208 +;-
1.209 + movq xmm0, xmm7
1.210 + psrldq xmm7, 8
1.211 +
1.212 + paddw xmm0, xmm7
1.213 + movd [rcx+8], xmm0
1.214 +
1.215 + STACK_FRAME_DESTROY_X3
1.216 +
1.217 +;void int vp9_sad16x8x3_sse3(
1.218 +; unsigned char *src_ptr,
1.219 +; int src_stride,
1.220 +; unsigned char *ref_ptr,
1.221 +; int ref_stride,
1.222 +; int *results)
1.223 +global sym(vp9_sad16x8x3_sse3) PRIVATE
1.224 +sym(vp9_sad16x8x3_sse3):
1.225 +
1.226 + STACK_FRAME_CREATE_X3
1.227 +
1.228 + PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
1.229 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.230 + PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.231 + PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
1.232 +
1.233 + mov rcx, result_ptr
1.234 +
1.235 + movq xmm0, xmm5
1.236 + psrldq xmm5, 8
1.237 +
1.238 + paddw xmm0, xmm5
1.239 + movd [rcx], xmm0
1.240 +;-
1.241 + movq xmm0, xmm6
1.242 + psrldq xmm6, 8
1.243 +
1.244 + paddw xmm0, xmm6
1.245 + movd [rcx+4], xmm0
1.246 +;-
1.247 + movq xmm0, xmm7
1.248 + psrldq xmm7, 8
1.249 +
1.250 + paddw xmm0, xmm7
1.251 + movd [rcx+8], xmm0
1.252 +
1.253 + STACK_FRAME_DESTROY_X3
1.254 +
1.255 +;void int vp9_sad8x16x3_sse3(
1.256 +; unsigned char *src_ptr,
1.257 +; int src_stride,
1.258 +; unsigned char *ref_ptr,
1.259 +; int ref_stride,
1.260 +; int *results)
1.261 +global sym(vp9_sad8x16x3_sse3) PRIVATE
1.262 +sym(vp9_sad8x16x3_sse3):
1.263 +
1.264 + STACK_FRAME_CREATE_X3
1.265 +
1.266 + PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
1.267 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.268 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.269 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.270 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.271 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.272 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.273 + PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
1.274 +
1.275 + mov rcx, result_ptr
1.276 +
1.277 + punpckldq mm5, mm6
1.278 +
1.279 + movq [rcx], mm5
1.280 + movd [rcx+8], mm7
1.281 +
1.282 + STACK_FRAME_DESTROY_X3
1.283 +
1.284 +;void int vp9_sad8x8x3_sse3(
1.285 +; unsigned char *src_ptr,
1.286 +; int src_stride,
1.287 +; unsigned char *ref_ptr,
1.288 +; int ref_stride,
1.289 +; int *results)
1.290 +global sym(vp9_sad8x8x3_sse3) PRIVATE
1.291 +sym(vp9_sad8x8x3_sse3):
1.292 +
1.293 + STACK_FRAME_CREATE_X3
1.294 +
1.295 + PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
1.296 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.297 + PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
1.298 + PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
1.299 +
1.300 + mov rcx, result_ptr
1.301 +
1.302 + punpckldq mm5, mm6
1.303 +
1.304 + movq [rcx], mm5
1.305 + movd [rcx+8], mm7
1.306 +
1.307 + STACK_FRAME_DESTROY_X3
1.308 +
1.309 +;void int vp9_sad4x4x3_sse3(
1.310 +; unsigned char *src_ptr,
1.311 +; int src_stride,
1.312 +; unsigned char *ref_ptr,
1.313 +; int ref_stride,
1.314 +; int *results)
1.315 +global sym(vp9_sad4x4x3_sse3) PRIVATE
1.316 +sym(vp9_sad4x4x3_sse3):
1.317 +
1.318 + STACK_FRAME_CREATE_X3
1.319 +
1.320 + movd mm0, DWORD PTR [src_ptr]
1.321 + movd mm1, DWORD PTR [ref_ptr]
1.322 +
1.323 + movd mm2, DWORD PTR [src_ptr+src_stride]
1.324 + movd mm3, DWORD PTR [ref_ptr+ref_stride]
1.325 +
1.326 + punpcklbw mm0, mm2
1.327 + punpcklbw mm1, mm3
1.328 +
1.329 + movd mm4, DWORD PTR [ref_ptr+1]
1.330 + movd mm5, DWORD PTR [ref_ptr+2]
1.331 +
1.332 + movd mm2, DWORD PTR [ref_ptr+ref_stride+1]
1.333 + movd mm3, DWORD PTR [ref_ptr+ref_stride+2]
1.334 +
1.335 + psadbw mm1, mm0
1.336 +
1.337 + punpcklbw mm4, mm2
1.338 + punpcklbw mm5, mm3
1.339 +
1.340 + psadbw mm4, mm0
1.341 + psadbw mm5, mm0
1.342 +
1.343 + lea src_ptr, [src_ptr+src_stride*2]
1.344 + lea ref_ptr, [ref_ptr+ref_stride*2]
1.345 +
1.346 + movd mm0, DWORD PTR [src_ptr]
1.347 + movd mm2, DWORD PTR [ref_ptr]
1.348 +
1.349 + movd mm3, DWORD PTR [src_ptr+src_stride]
1.350 + movd mm6, DWORD PTR [ref_ptr+ref_stride]
1.351 +
1.352 + punpcklbw mm0, mm3
1.353 + punpcklbw mm2, mm6
1.354 +
1.355 + movd mm3, DWORD PTR [ref_ptr+1]
1.356 + movd mm7, DWORD PTR [ref_ptr+2]
1.357 +
1.358 + psadbw mm2, mm0
1.359 +
1.360 + paddw mm1, mm2
1.361 +
1.362 + movd mm2, DWORD PTR [ref_ptr+ref_stride+1]
1.363 + movd mm6, DWORD PTR [ref_ptr+ref_stride+2]
1.364 +
1.365 + punpcklbw mm3, mm2
1.366 + punpcklbw mm7, mm6
1.367 +
1.368 + psadbw mm3, mm0
1.369 + psadbw mm7, mm0
1.370 +
1.371 + paddw mm3, mm4
1.372 + paddw mm7, mm5
1.373 +
1.374 + mov rcx, result_ptr
1.375 +
1.376 + punpckldq mm1, mm3
1.377 +
1.378 + movq [rcx], mm1
1.379 + movd [rcx+8], mm7
1.380 +
1.381 + STACK_FRAME_DESTROY_X3
