yading@10: /*
yading@10:  * MMX optimized LPC DSP utils
yading@10:  * Copyright (c) 2007 Loren Merritt
yading@10:  *
yading@10:  * This file is part of FFmpeg.
yading@10:  *
yading@10:  * FFmpeg is free software; you can redistribute it and/or
yading@10:  * modify it under the terms of the GNU Lesser General Public
yading@10:  * License as published by the Free Software Foundation; either
yading@10:  * version 2.1 of the License, or (at your option) any later version.
yading@10:  *
yading@10:  * FFmpeg is distributed in the hope that it will be useful,
yading@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@10:  * Lesser General Public License for more details.
yading@10:  *
yading@10:  * You should have received a copy of the GNU Lesser General Public
yading@10:  * License along with FFmpeg; if not, write to the Free Software
yading@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10:  */
yading@10: 
yading@10: #include "libavutil/x86/asm.h"
yading@10: #include "libavutil/attributes.h"
yading@10: #include "libavutil/cpu.h"
yading@10: #include "libavcodec/lpc.h"
yading@10: 
yading@10: #if HAVE_SSE2_INLINE
yading@10: 
yading@10: static void lpc_apply_welch_window_sse2(const int32_t *data, int len,
yading@10:                                         double *w_data)
yading@10: {
yading@10:     double c = 2.0 / (len-1.0);
yading@10:     int n2 = len>>1;
yading@10:     x86_reg i = -n2*sizeof(int32_t);
yading@10:     x86_reg j =  n2*sizeof(int32_t);
yading@10:     __asm__ volatile(
yading@10:         "movsd   %4,     %%xmm7                \n\t"
yading@10:         "movapd  "MANGLE(ff_pd_1)", %%xmm6     \n\t"
yading@10:         "movapd  "MANGLE(ff_pd_2)", %%xmm5     \n\t"
yading@10:         "movlhps %%xmm7, %%xmm7                \n\t"
yading@10:         "subpd   %%xmm5, %%xmm7                \n\t"
yading@10:         "addsd   %%xmm6, %%xmm7                \n\t"
yading@10:         "test    $1,     %5                    \n\t"
yading@10:         "jz      2f                            \n\t"
yading@10: #define WELCH(MOVPD, offset)\
yading@10:         "1:                                    \n\t"\
yading@10:         "movapd   %%xmm7,  %%xmm1              \n\t"\
yading@10:         "mulpd    %%xmm1,  %%xmm1              \n\t"\
yading@10:         "movapd   %%xmm6,  %%xmm0              \n\t"\
yading@10:         "subpd    %%xmm1,  %%xmm0              \n\t"\
yading@10:         "pshufd   $0x4e,   %%xmm0, %%xmm1      \n\t"\
yading@10:         "cvtpi2pd (%3,%0), %%xmm2              \n\t"\
yading@10:         "cvtpi2pd "#offset"*4(%3,%1), %%xmm3   \n\t"\
yading@10:         "mulpd    %%xmm0,  %%xmm2              \n\t"\
yading@10:         "mulpd    %%xmm1,  %%xmm3              \n\t"\
yading@10:         "movapd   %%xmm2, (%2,%0,2)            \n\t"\
yading@10:         MOVPD"    %%xmm3, "#offset"*8(%2,%1,2) \n\t"\
yading@10:         "subpd    %%xmm5,  %%xmm7              \n\t"\
yading@10:         "sub      $8,      %1                  \n\t"\
yading@10:         "add      $8,      %0                  \n\t"\
yading@10:         "jl 1b                                 \n\t"\
yading@10: 
yading@10:         WELCH("movupd", -1)
yading@10:         "jmp 3f                                \n\t"
yading@10:         "2:                                    \n\t"
yading@10:         WELCH("movapd", -2)
yading@10:         "3:                                    \n\t"
yading@10:         :"+&r"(i), "+&r"(j)
yading@10:         :"r"(w_data+n2), "r"(data+n2), "m"(c), "r"(len)
yading@10:          XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
yading@10:                                     "%xmm5", "%xmm6", "%xmm7")
yading@10:     );
yading@10: #undef WELCH
yading@10: }
yading@10: 
yading@10: static void lpc_compute_autocorr_sse2(const double *data, int len, int lag,
yading@10:                                       double *autoc)
yading@10: {
yading@10:     int j;
yading@10: 
yading@10:     if((x86_reg)data & 15)
yading@10:         data++;
yading@10: 
yading@10:     for(j=0; j<lag; j+=2){
yading@10:         x86_reg i = -len*sizeof(double);
yading@10:         if(j == lag-2) {
yading@10:             __asm__ volatile(
yading@10:                 "movsd    "MANGLE(ff_pd_1)", %%xmm0 \n\t"
yading@10:                 "movsd    "MANGLE(ff_pd_1)", %%xmm1 \n\t"
yading@10:                 "movsd    "MANGLE(ff_pd_1)", %%xmm2 \n\t"
yading@10:                 "1:                                 \n\t"
yading@10:                 "movapd   (%2,%0), %%xmm3           \n\t"
yading@10:                 "movupd -8(%3,%0), %%xmm4           \n\t"
yading@10:                 "movapd   (%3,%0), %%xmm5           \n\t"
yading@10:                 "mulpd     %%xmm3, %%xmm4           \n\t"
yading@10:                 "mulpd     %%xmm3, %%xmm5           \n\t"
yading@10:                 "mulpd -16(%3,%0), %%xmm3           \n\t"
yading@10:                 "addpd     %%xmm4, %%xmm1           \n\t"
yading@10:                 "addpd     %%xmm5, %%xmm0           \n\t"
yading@10:                 "addpd     %%xmm3, %%xmm2           \n\t"
yading@10:                 "add       $16,    %0               \n\t"
yading@10:                 "jl 1b                              \n\t"
yading@10:                 "movhlps   %%xmm0, %%xmm3           \n\t"
yading@10:                 "movhlps   %%xmm1, %%xmm4           \n\t"
yading@10:                 "movhlps   %%xmm2, %%xmm5           \n\t"
yading@10:                 "addsd     %%xmm3, %%xmm0           \n\t"
yading@10:                 "addsd     %%xmm4, %%xmm1           \n\t"
yading@10:                 "addsd     %%xmm5, %%xmm2           \n\t"
yading@10:                 "movsd     %%xmm0,   (%1)           \n\t"
yading@10:                 "movsd     %%xmm1,  8(%1)           \n\t"
yading@10:                 "movsd     %%xmm2, 16(%1)           \n\t"
yading@10:                 :"+&r"(i)
yading@10:                 :"r"(autoc+j), "r"(data+len), "r"(data+len-j)
yading@10:                 :"memory"
yading@10:             );
yading@10:         } else {
yading@10:             __asm__ volatile(
yading@10:                 "movsd    "MANGLE(ff_pd_1)", %%xmm0 \n\t"
yading@10:                 "movsd    "MANGLE(ff_pd_1)", %%xmm1 \n\t"
yading@10:                 "1:                                 \n\t"
yading@10:                 "movapd   (%3,%0), %%xmm3           \n\t"
yading@10:                 "movupd -8(%4,%0), %%xmm4           \n\t"
yading@10:                 "mulpd     %%xmm3, %%xmm4           \n\t"
yading@10:                 "mulpd    (%4,%0), %%xmm3           \n\t"
yading@10:                 "addpd     %%xmm4, %%xmm1           \n\t"
yading@10:                 "addpd     %%xmm3, %%xmm0           \n\t"
yading@10:                 "add       $16,    %0               \n\t"
yading@10:                 "jl 1b                              \n\t"
yading@10:                 "movhlps   %%xmm0, %%xmm3           \n\t"
yading@10:                 "movhlps   %%xmm1, %%xmm4           \n\t"
yading@10:                 "addsd     %%xmm3, %%xmm0           \n\t"
yading@10:                 "addsd     %%xmm4, %%xmm1           \n\t"
yading@10:                 "movsd     %%xmm0, %1               \n\t"
yading@10:                 "movsd     %%xmm1, %2               \n\t"
yading@10:                 :"+&r"(i), "=m"(autoc[j]), "=m"(autoc[j+1])
yading@10:                 :"r"(data+len), "r"(data+len-j)
yading@10:             );
yading@10:         }
yading@10:     }
yading@10: }
yading@10: 
yading@10: #endif /* HAVE_SSE2_INLINE */
yading@10: 
yading@10: av_cold void ff_lpc_init_x86(LPCContext *c)
yading@10: {
yading@10: #if HAVE_SSE2_INLINE
yading@10:     int mm_flags = av_get_cpu_flags();
yading@10: 
yading@10:     if (mm_flags & (AV_CPU_FLAG_SSE2|AV_CPU_FLAG_SSE2SLOW)) {
yading@10:         c->lpc_apply_welch_window = lpc_apply_welch_window_sse2;
yading@10:         c->lpc_compute_autocorr   = lpc_compute_autocorr_sse2;
yading@10:     }
yading@10: #endif /* HAVE_SSE2_INLINE */
yading@10: }