1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/encoder/x86/subtract_mmx.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,223 @@
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 +;void vp8_subtract_b_mmx_impl(unsigned char *z, int src_stride,
1.18 +; short *diff, unsigned char *Predictor,
1.19 +; int pitch);
1.20 +global sym(vp8_subtract_b_mmx_impl) PRIVATE
1.21 +sym(vp8_subtract_b_mmx_impl):
1.22 + push rbp
1.23 + mov rbp, rsp
1.24 + SHADOW_ARGS_TO_STACK 5
1.25 + push rsi
1.26 + push rdi
1.27 + ; end prolog
1.28 +
1.29 +
1.30 + mov rdi, arg(2) ;diff
1.31 + mov rax, arg(3) ;Predictor
1.32 + mov rsi, arg(0) ;z
1.33 + movsxd rdx, dword ptr arg(1);src_stride;
1.34 + movsxd rcx, dword ptr arg(4);pitch
1.35 + pxor mm7, mm7
1.36 +
1.37 + movd mm0, [rsi]
1.38 + movd mm1, [rax]
1.39 + punpcklbw mm0, mm7
1.40 + punpcklbw mm1, mm7
1.41 + psubw mm0, mm1
1.42 + movq [rdi], mm0
1.43 +
1.44 +
1.45 + movd mm0, [rsi+rdx]
1.46 + movd mm1, [rax+rcx]
1.47 + punpcklbw mm0, mm7
1.48 + punpcklbw mm1, mm7
1.49 + psubw mm0, mm1
1.50 + movq [rdi+rcx*2],mm0
1.51 +
1.52 +
1.53 + movd mm0, [rsi+rdx*2]
1.54 + movd mm1, [rax+rcx*2]
1.55 + punpcklbw mm0, mm7
1.56 + punpcklbw mm1, mm7
1.57 + psubw mm0, mm1
1.58 + movq [rdi+rcx*4], mm0
1.59 +
1.60 + lea rsi, [rsi+rdx*2]
1.61 + lea rcx, [rcx+rcx*2]
1.62 +
1.63 +
1.64 +
1.65 + movd mm0, [rsi+rdx]
1.66 + movd mm1, [rax+rcx]
1.67 + punpcklbw mm0, mm7
1.68 + punpcklbw mm1, mm7
1.69 + psubw mm0, mm1
1.70 + movq [rdi+rcx*2], mm0
1.71 +
1.72 + ; begin epilog
1.73 + pop rdi
1.74 + pop rsi
1.75 + UNSHADOW_ARGS
1.76 + pop rbp
1.77 + ret
1.78 +
1.79 +;void vp8_subtract_mby_mmx(short *diff, unsigned char *src, int src_stride,
1.80 +;unsigned char *pred, int pred_stride)
1.81 +global sym(vp8_subtract_mby_mmx) PRIVATE
1.82 +sym(vp8_subtract_mby_mmx):
1.83 + push rbp
1.84 + mov rbp, rsp
1.85 + SHADOW_ARGS_TO_STACK 5
1.86 + push rsi
1.87 + push rdi
1.88 + ; end prolog
1.89 +
1.90 + mov rdi, arg(0) ;diff
1.91 + mov rsi, arg(1) ;src
1.92 + movsxd rdx, dword ptr arg(2);src_stride
1.93 + mov rax, arg(3) ;pred
1.94 + push rbx
1.95 + movsxd rbx, dword ptr arg(4);pred_stride
1.96 +
1.97 + pxor mm0, mm0
1.98 + mov rcx, 16
1.99 +
1.100 +
1.101 +.submby_loop:
1.102 + movq mm1, [rsi]
1.103 + movq mm3, [rax]
1.104 +
1.105 + movq mm2, mm1
1.106 + movq mm4, mm3
1.107 +
1.108 + punpcklbw mm1, mm0
1.109 + punpcklbw mm3, mm0
1.110 +
1.111 + punpckhbw mm2, mm0
1.112 + punpckhbw mm4, mm0
1.113 +
1.114 + psubw mm1, mm3
1.115 + psubw mm2, mm4
1.116 +
1.117 + movq [rdi], mm1
1.118 + movq [rdi+8], mm2
1.119 +
1.120 + movq mm1, [rsi+8]
1.121 + movq mm3, [rax+8]
1.122 +
1.123 + movq mm2, mm1
1.124 + movq mm4, mm3
1.125 +
1.126 + punpcklbw mm1, mm0
1.127 + punpcklbw mm3, mm0
1.128 +
1.129 + punpckhbw mm2, mm0
1.130 + punpckhbw mm4, mm0
1.131 +
1.132 + psubw mm1, mm3
1.133 + psubw mm2, mm4
1.134 +
1.135 + movq [rdi+16], mm1
1.136 + movq [rdi+24], mm2
1.137 + add rdi, 32
1.138 + lea rax, [rax+rbx]
1.139 + lea rsi, [rsi+rdx]
1.140 + dec rcx
1.141 + jnz .submby_loop
1.142 +
1.143 + pop rbx
1.144 + pop rdi
1.145 + pop rsi
1.146 + ; begin epilog
1.147 + UNSHADOW_ARGS
1.148 + pop rbp
1.149 + ret
1.150 +
1.151 +
1.152 +;vp8_subtract_mbuv_mmx(short *diff, unsigned char *usrc, unsigned char *vsrc,
1.153 +; int src_stride, unsigned char *upred,
1.154 +; unsigned char *vpred, int pred_stride)
1.155 +
1.156 +global sym(vp8_subtract_mbuv_mmx) PRIVATE
1.157 +sym(vp8_subtract_mbuv_mmx):
1.158 + push rbp
1.159 + mov rbp, rsp
1.160 + SHADOW_ARGS_TO_STACK 7
1.161 + push rsi
1.162 + push rdi
1.163 + ; end prolog
1.164 +
1.165 + mov rdi, arg(0) ;diff
1.166 + mov rsi, arg(1) ;usrc
1.167 + movsxd rdx, dword ptr arg(3);src_stride;
1.168 + mov rax, arg(4) ;upred
1.169 + add rdi, 256*2 ;diff = diff + 256 (shorts)
1.170 + mov rcx, 8
1.171 + push rbx
1.172 + movsxd rbx, dword ptr arg(6);pred_stride
1.173 +
1.174 + pxor mm7, mm7
1.175 +
1.176 +.submbu_loop:
1.177 + movq mm0, [rsi]
1.178 + movq mm1, [rax]
1.179 + movq mm3, mm0
1.180 + movq mm4, mm1
1.181 + punpcklbw mm0, mm7
1.182 + punpcklbw mm1, mm7
1.183 + punpckhbw mm3, mm7
1.184 + punpckhbw mm4, mm7
1.185 + psubw mm0, mm1
1.186 + psubw mm3, mm4
1.187 + movq [rdi], mm0
1.188 + movq [rdi+8], mm3
1.189 + add rdi, 16
1.190 + add rsi, rdx
1.191 + add rax, rbx
1.192 +
1.193 + dec rcx
1.194 + jnz .submbu_loop
1.195 +
1.196 + mov rsi, arg(2) ;vsrc
1.197 + mov rax, arg(5) ;vpred
1.198 + mov rcx, 8
1.199 +
1.200 +.submbv_loop:
1.201 + movq mm0, [rsi]
1.202 + movq mm1, [rax]
1.203 + movq mm3, mm0
1.204 + movq mm4, mm1
1.205 + punpcklbw mm0, mm7
1.206 + punpcklbw mm1, mm7
1.207 + punpckhbw mm3, mm7
1.208 + punpckhbw mm4, mm7
1.209 + psubw mm0, mm1
1.210 + psubw mm3, mm4
1.211 + movq [rdi], mm0
1.212 + movq [rdi+8], mm3
1.213 + add rdi, 16
1.214 + add rsi, rdx
1.215 + add rax, rbx
1.216 +
1.217 + dec rcx
1.218 + jnz .submbv_loop
1.219 +
1.220 + pop rbx
1.221 + ; begin epilog
1.222 + pop rdi
1.223 + pop rsi
1.224 + UNSHADOW_ARGS
1.225 + pop rbp
1.226 + ret
