Libav 0.7.1
|
00001 /* 00002 * MMX optimized LPC DSP utils 00003 * Copyright (c) 2007 Loren Merritt 00004 * 00005 * This file is part of Libav. 00006 * 00007 * Libav is free software; you can redistribute it and/or 00008 * modify it under the terms of the GNU Lesser General Public 00009 * License as published by the Free Software Foundation; either 00010 * version 2.1 of the License, or (at your option) any later version. 00011 * 00012 * Libav is distributed in the hope that it will be useful, 00013 * but WITHOUT ANY WARRANTY; without even the implied warranty of 00014 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 00015 * Lesser General Public License for more details. 00016 * 00017 * You should have received a copy of the GNU Lesser General Public 00018 * License along with Libav; if not, write to the Free Software 00019 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 00020 */ 00021 00022 #include "libavutil/x86_cpu.h" 00023 #include "libavutil/cpu.h" 00024 #include "libavcodec/lpc.h" 00025 00026 static void lpc_apply_welch_window_sse2(const int32_t *data, int len, 00027 double *w_data) 00028 { 00029 double c = 2.0 / (len-1.0); 00030 int n2 = len>>1; 00031 x86_reg i = -n2*sizeof(int32_t); 00032 x86_reg j = n2*sizeof(int32_t); 00033 __asm__ volatile( 00034 "movsd %4, %%xmm7 \n\t" 00035 "movapd "MANGLE(ff_pd_1)", %%xmm6 \n\t" 00036 "movapd "MANGLE(ff_pd_2)", %%xmm5 \n\t" 00037 "movlhps %%xmm7, %%xmm7 \n\t" 00038 "subpd %%xmm5, %%xmm7 \n\t" 00039 "addsd %%xmm6, %%xmm7 \n\t" 00040 "test $1, %5 \n\t" 00041 "jz 2f \n\t" 00042 #define WELCH(MOVPD, offset)\ 00043 "1: \n\t"\ 00044 "movapd %%xmm7, %%xmm1 \n\t"\ 00045 "mulpd %%xmm1, %%xmm1 \n\t"\ 00046 "movapd %%xmm6, %%xmm0 \n\t"\ 00047 "subpd %%xmm1, %%xmm0 \n\t"\ 00048 "pshufd $0x4e, %%xmm0, %%xmm1 \n\t"\ 00049 "cvtpi2pd (%3,%0), %%xmm2 \n\t"\ 00050 "cvtpi2pd "#offset"*4(%3,%1), %%xmm3 \n\t"\ 00051 "mulpd %%xmm0, %%xmm2 \n\t"\ 00052 "mulpd %%xmm1, %%xmm3 \n\t"\ 00053 "movapd %%xmm2, (%2,%0,2) \n\t"\ 00054 MOVPD" %%xmm3, "#offset"*8(%2,%1,2) \n\t"\ 00055 "subpd %%xmm5, %%xmm7 \n\t"\ 00056 "sub $8, %1 \n\t"\ 00057 "add $8, %0 \n\t"\ 00058 "jl 1b \n\t"\ 00059 00060 WELCH("movupd", -1) 00061 "jmp 3f \n\t" 00062 "2: \n\t" 00063 WELCH("movapd", -2) 00064 "3: \n\t" 00065 :"+&r"(i), "+&r"(j) 00066 :"r"(w_data+n2), "r"(data+n2), "m"(c), "r"(len) 00067 XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3", 00068 "%xmm5", "%xmm6", "%xmm7") 00069 ); 00070 #undef WELCH 00071 } 00072 00073 static void lpc_compute_autocorr_sse2(const double *data, int len, int lag, 00074 double *autoc) 00075 { 00076 int j; 00077 00078 if((x86_reg)data & 15) 00079 data++; 00080 00081 for(j=0; j<lag; j+=2){ 00082 x86_reg i = -len*sizeof(double); 00083 if(j == lag-2) { 00084 __asm__ volatile( 00085 "movsd "MANGLE(ff_pd_1)", %%xmm0 \n\t" 00086 "movsd "MANGLE(ff_pd_1)", %%xmm1 \n\t" 00087 "movsd "MANGLE(ff_pd_1)", %%xmm2 \n\t" 00088 "1: \n\t" 00089 "movapd (%2,%0), %%xmm3 \n\t" 00090 "movupd -8(%3,%0), %%xmm4 \n\t" 00091 "movapd (%3,%0), %%xmm5 \n\t" 00092 "mulpd %%xmm3, %%xmm4 \n\t" 00093 "mulpd %%xmm3, %%xmm5 \n\t" 00094 "mulpd -16(%3,%0), %%xmm3 \n\t" 00095 "addpd %%xmm4, %%xmm1 \n\t" 00096 "addpd %%xmm5, %%xmm0 \n\t" 00097 "addpd %%xmm3, %%xmm2 \n\t" 00098 "add $16, %0 \n\t" 00099 "jl 1b \n\t" 00100 "movhlps %%xmm0, %%xmm3 \n\t" 00101 "movhlps %%xmm1, %%xmm4 \n\t" 00102 "movhlps %%xmm2, %%xmm5 \n\t" 00103 "addsd %%xmm3, %%xmm0 \n\t" 00104 "addsd %%xmm4, %%xmm1 \n\t" 00105 "addsd %%xmm5, %%xmm2 \n\t" 00106 "movsd %%xmm0, (%1) \n\t" 00107 "movsd %%xmm1, 8(%1) \n\t" 00108 "movsd %%xmm2, 16(%1) \n\t" 00109 :"+&r"(i) 00110 :"r"(autoc+j), "r"(data+len), "r"(data+len-j) 00111 :"memory" 00112 ); 00113 } else { 00114 __asm__ volatile( 00115 "movsd "MANGLE(ff_pd_1)", %%xmm0 \n\t" 00116 "movsd "MANGLE(ff_pd_1)", %%xmm1 \n\t" 00117 "1: \n\t" 00118 "movapd (%3,%0), %%xmm3 \n\t" 00119 "movupd -8(%4,%0), %%xmm4 \n\t" 00120 "mulpd %%xmm3, %%xmm4 \n\t" 00121 "mulpd (%4,%0), %%xmm3 \n\t" 00122 "addpd %%xmm4, %%xmm1 \n\t" 00123 "addpd %%xmm3, %%xmm0 \n\t" 00124 "add $16, %0 \n\t" 00125 "jl 1b \n\t" 00126 "movhlps %%xmm0, %%xmm3 \n\t" 00127 "movhlps %%xmm1, %%xmm4 \n\t" 00128 "addsd %%xmm3, %%xmm0 \n\t" 00129 "addsd %%xmm4, %%xmm1 \n\t" 00130 "movsd %%xmm0, %1 \n\t" 00131 "movsd %%xmm1, %2 \n\t" 00132 :"+&r"(i), "=m"(autoc[j]), "=m"(autoc[j+1]) 00133 :"r"(data+len), "r"(data+len-j) 00134 ); 00135 } 00136 } 00137 } 00138 00139 av_cold void ff_lpc_init_x86(LPCContext *c) 00140 { 00141 int mm_flags = av_get_cpu_flags(); 00142 00143 if (mm_flags & (AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW)) { 00144 c->lpc_apply_welch_window = lpc_apply_welch_window_sse2; 00145 c->lpc_compute_autocorr = lpc_compute_autocorr_sse2; 00146 } 00147 }