1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp8/common/x86/dequantize_mmx.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,258 @@
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 vp8_dequantize_b_impl_mmx(short *sq, short *dq, short *q)
1.19 +global sym(vp8_dequantize_b_impl_mmx) PRIVATE
1.20 +sym(vp8_dequantize_b_impl_mmx):
1.21 + push rbp
1.22 + mov rbp, rsp
1.23 + SHADOW_ARGS_TO_STACK 3
1.24 + push rsi
1.25 + push rdi
1.26 + ; end prolog
1.27 +
1.28 + mov rsi, arg(0) ;sq
1.29 + mov rdi, arg(1) ;dq
1.30 + mov rax, arg(2) ;q
1.31 +
1.32 + movq mm1, [rsi]
1.33 + pmullw mm1, [rax+0] ; mm4 *= kernel 0 modifiers.
1.34 + movq [rdi], mm1
1.35 +
1.36 + movq mm1, [rsi+8]
1.37 + pmullw mm1, [rax+8] ; mm4 *= kernel 0 modifiers.
1.38 + movq [rdi+8], mm1
1.39 +
1.40 + movq mm1, [rsi+16]
1.41 + pmullw mm1, [rax+16] ; mm4 *= kernel 0 modifiers.
1.42 + movq [rdi+16], mm1
1.43 +
1.44 + movq mm1, [rsi+24]
1.45 + pmullw mm1, [rax+24] ; mm4 *= kernel 0 modifiers.
1.46 + movq [rdi+24], mm1
1.47 +
1.48 + ; begin epilog
1.49 + pop rdi
1.50 + pop rsi
1.51 + UNSHADOW_ARGS
1.52 + pop rbp
1.53 + ret
1.54 +
1.55 +
1.56 +;void dequant_idct_add_mmx(
1.57 +;short *input, 0
1.58 +;short *dq, 1
1.59 +;unsigned char *dest, 2
1.60 +;int stride) 3
1.61 +global sym(vp8_dequant_idct_add_mmx) PRIVATE
1.62 +sym(vp8_dequant_idct_add_mmx):
1.63 + push rbp
1.64 + mov rbp, rsp
1.65 + SHADOW_ARGS_TO_STACK 4
1.66 + GET_GOT rbx
1.67 + push rdi
1.68 + ; end prolog
1.69 +
1.70 + mov rax, arg(0) ;input
1.71 + mov rdx, arg(1) ;dq
1.72 +
1.73 +
1.74 + movq mm0, [rax ]
1.75 + pmullw mm0, [rdx]
1.76 +
1.77 + movq mm1, [rax +8]
1.78 + pmullw mm1, [rdx +8]
1.79 +
1.80 + movq mm2, [rax+16]
1.81 + pmullw mm2, [rdx+16]
1.82 +
1.83 + movq mm3, [rax+24]
1.84 + pmullw mm3, [rdx+24]
1.85 +
1.86 + mov rdx, arg(2) ;dest
1.87 +
1.88 + pxor mm7, mm7
1.89 +
1.90 +
1.91 + movq [rax], mm7
1.92 + movq [rax+8], mm7
1.93 +
1.94 + movq [rax+16],mm7
1.95 + movq [rax+24],mm7
1.96 +
1.97 +
1.98 + movsxd rdi, dword ptr arg(3) ;stride
1.99 +
1.100 + psubw mm0, mm2 ; b1= 0-2
1.101 + paddw mm2, mm2 ;
1.102 +
1.103 + movq mm5, mm1
1.104 + paddw mm2, mm0 ; a1 =0+2
1.105 +
1.106 + pmulhw mm5, [GLOBAL(x_s1sqr2)];
1.107 + paddw mm5, mm1 ; ip1 * sin(pi/8) * sqrt(2)
1.108 +
1.109 + movq mm7, mm3 ;
1.110 + pmulhw mm7, [GLOBAL(x_c1sqr2less1)];
1.111 +
1.112 + paddw mm7, mm3 ; ip3 * cos(pi/8) * sqrt(2)
1.113 + psubw mm7, mm5 ; c1
1.114 +
1.115 + movq mm5, mm1
1.116 + movq mm4, mm3
1.117 +
1.118 + pmulhw mm5, [GLOBAL(x_c1sqr2less1)]
1.119 + paddw mm5, mm1
1.120 +
1.121 + pmulhw mm3, [GLOBAL(x_s1sqr2)]
1.122 + paddw mm3, mm4
1.123 +
1.124 + paddw mm3, mm5 ; d1
1.125 + movq mm6, mm2 ; a1
1.126 +
1.127 + movq mm4, mm0 ; b1
1.128 + paddw mm2, mm3 ;0
1.129 +
1.130 + paddw mm4, mm7 ;1
1.131 + psubw mm0, mm7 ;2
1.132 +
1.133 + psubw mm6, mm3 ;3
1.134 +
1.135 + movq mm1, mm2 ; 03 02 01 00
1.136 + movq mm3, mm4 ; 23 22 21 20
1.137 +
1.138 + punpcklwd mm1, mm0 ; 11 01 10 00
1.139 + punpckhwd mm2, mm0 ; 13 03 12 02
1.140 +
1.141 + punpcklwd mm3, mm6 ; 31 21 30 20
1.142 + punpckhwd mm4, mm6 ; 33 23 32 22
1.143 +
1.144 + movq mm0, mm1 ; 11 01 10 00
1.145 + movq mm5, mm2 ; 13 03 12 02
1.146 +
1.147 + punpckldq mm0, mm3 ; 30 20 10 00
1.148 + punpckhdq mm1, mm3 ; 31 21 11 01
1.149 +
1.150 + punpckldq mm2, mm4 ; 32 22 12 02
1.151 + punpckhdq mm5, mm4 ; 33 23 13 03
1.152 +
1.153 + movq mm3, mm5 ; 33 23 13 03
1.154 +
1.155 + psubw mm0, mm2 ; b1= 0-2
1.156 + paddw mm2, mm2 ;
1.157 +
1.158 + movq mm5, mm1
1.159 + paddw mm2, mm0 ; a1 =0+2
1.160 +
1.161 + pmulhw mm5, [GLOBAL(x_s1sqr2)];
1.162 + paddw mm5, mm1 ; ip1 * sin(pi/8) * sqrt(2)
1.163 +
1.164 + movq mm7, mm3 ;
1.165 + pmulhw mm7, [GLOBAL(x_c1sqr2less1)];
1.166 +
1.167 + paddw mm7, mm3 ; ip3 * cos(pi/8) * sqrt(2)
1.168 + psubw mm7, mm5 ; c1
1.169 +
1.170 + movq mm5, mm1
1.171 + movq mm4, mm3
1.172 +
1.173 + pmulhw mm5, [GLOBAL(x_c1sqr2less1)]
1.174 + paddw mm5, mm1
1.175 +
1.176 + pmulhw mm3, [GLOBAL(x_s1sqr2)]
1.177 + paddw mm3, mm4
1.178 +
1.179 + paddw mm3, mm5 ; d1
1.180 + paddw mm0, [GLOBAL(fours)]
1.181 +
1.182 + paddw mm2, [GLOBAL(fours)]
1.183 + movq mm6, mm2 ; a1
1.184 +
1.185 + movq mm4, mm0 ; b1
1.186 + paddw mm2, mm3 ;0
1.187 +
1.188 + paddw mm4, mm7 ;1
1.189 + psubw mm0, mm7 ;2
1.190 +
1.191 + psubw mm6, mm3 ;3
1.192 + psraw mm2, 3
1.193 +
1.194 + psraw mm0, 3
1.195 + psraw mm4, 3
1.196 +
1.197 + psraw mm6, 3
1.198 +
1.199 + movq mm1, mm2 ; 03 02 01 00
1.200 + movq mm3, mm4 ; 23 22 21 20
1.201 +
1.202 + punpcklwd mm1, mm0 ; 11 01 10 00
1.203 + punpckhwd mm2, mm0 ; 13 03 12 02
1.204 +
1.205 + punpcklwd mm3, mm6 ; 31 21 30 20
1.206 + punpckhwd mm4, mm6 ; 33 23 32 22
1.207 +
1.208 + movq mm0, mm1 ; 11 01 10 00
1.209 + movq mm5, mm2 ; 13 03 12 02
1.210 +
1.211 + punpckldq mm0, mm3 ; 30 20 10 00
1.212 + punpckhdq mm1, mm3 ; 31 21 11 01
1.213 +
1.214 + punpckldq mm2, mm4 ; 32 22 12 02
1.215 + punpckhdq mm5, mm4 ; 33 23 13 03
1.216 +
1.217 + pxor mm7, mm7
1.218 +
1.219 + movd mm4, [rdx]
1.220 + punpcklbw mm4, mm7
1.221 + paddsw mm0, mm4
1.222 + packuswb mm0, mm7
1.223 + movd [rdx], mm0
1.224 +
1.225 + movd mm4, [rdx+rdi]
1.226 + punpcklbw mm4, mm7
1.227 + paddsw mm1, mm4
1.228 + packuswb mm1, mm7
1.229 + movd [rdx+rdi], mm1
1.230 +
1.231 + movd mm4, [rdx+2*rdi]
1.232 + punpcklbw mm4, mm7
1.233 + paddsw mm2, mm4
1.234 + packuswb mm2, mm7
1.235 + movd [rdx+rdi*2], mm2
1.236 +
1.237 + add rdx, rdi
1.238 +
1.239 + movd mm4, [rdx+2*rdi]
1.240 + punpcklbw mm4, mm7
1.241 + paddsw mm5, mm4
1.242 + packuswb mm5, mm7
1.243 + movd [rdx+rdi*2], mm5
1.244 +
1.245 + ; begin epilog
1.246 + pop rdi
1.247 + RESTORE_GOT
1.248 + UNSHADOW_ARGS
1.249 + pop rbp
1.250 + ret
1.251 +
1.252 +SECTION_RODATA
1.253 +align 16
1.254 +x_s1sqr2:
1.255 + times 4 dw 0x8A8C
1.256 +align 16
1.257 +x_c1sqr2less1:
1.258 + times 4 dw 0x4E7B
1.259 +align 16
1.260 +fours:
1.261 + times 4 dw 0x0004
