lib/3rdparty/libmpg123/synth_sse_float.S

   1 /*
   2         synth_sse_float: SSE optimized synth (float output version)
   3
   4         copyright 1995-2009 by the mpg123 project - free software under the terms of the LGPL 2.1
   5         see COPYING and AUTHORS files in distribution or http://mpg123.org
   6         initially written by Taihei Monma
   7 */
   8
   9 #include "mangle.h"
  10
  11 /* real *window; */
  12 #define WINDOW %ebx
  13 /* real *b0; */
  14 #define B0 %edx
  15 /* real *samples; */
  16 #define SAMPLES %esi
  17
  18 /*
  19         int synth_1to1_real_sse_asm(real *window, real *b0, real *samples, int bo1);
  20         return value: number of clipped samples (0)
  21 */
  22
  23 #ifndef __APPLE__
  24         .section        .rodata
  25 #else
  26         .data
  27 #endif
  28         ALIGN32
  29 ASM_NAME(scale_sse):
  30         .long   939524096
  31         .long   939524096
  32         .long   939524096
  33         .long   939524096
  34         .text
  35         ALIGN16
  36 .globl ASM_NAME(synth_1to1_real_sse_asm)
  37 ASM_NAME(synth_1to1_real_sse_asm):
  38         pushl           %ebp
  39         movl            %esp, %ebp
  40         pushl           %ebx
  41         pushl           %esi
  42
  43         movl            8(%ebp), WINDOW
  44         movl            12(%ebp), B0
  45         movl            16(%ebp), SAMPLES
  46         movl            20(%ebp), %eax
  47         shll            $2, %eax
  48
  49         leal            64(WINDOW), WINDOW
  50         subl            %eax, WINDOW
  51
  52         movl            $4, %ecx
  53
  54         ALIGN16
  55 Loop_start_1:
  56         movups          (WINDOW), %xmm0
  57         movups          16(WINDOW), %xmm1
  58         movups          32(WINDOW), %xmm2
  59         movups          48(WINDOW), %xmm3
  60         movups          128(WINDOW), %xmm4
  61         movups          144(WINDOW), %xmm5
  62         movups          160(WINDOW), %xmm6
  63         movups          176(WINDOW), %xmm7
  64         mulps           0(B0), %xmm0
  65         mulps           16(B0), %xmm1
  66         mulps           32(B0), %xmm2
  67         mulps           48(B0), %xmm3
  68         mulps           64(B0), %xmm4
  69         mulps           80(B0), %xmm5
  70         mulps           96(B0), %xmm6
  71         mulps           112(B0), %xmm7
  72         addps           %xmm1, %xmm0
  73         addps           %xmm3, %xmm2
  74         addps           %xmm5, %xmm4
  75         addps           %xmm7, %xmm6
  76         addps           %xmm2, %xmm0
  77         addps           %xmm6, %xmm4
  78         movaps          %xmm4, %xmm5
  79         movaps          %xmm0, %xmm4
  80
  81         leal            256(WINDOW), WINDOW
  82         leal            128(B0), B0
  83
  84         movups          (WINDOW), %xmm0
  85         movups          16(WINDOW), %xmm1
  86         movups          32(WINDOW), %xmm2
  87         movups          48(WINDOW), %xmm3
  88         movups          128(WINDOW), %xmm6
  89         movups          144(WINDOW), %xmm7
  90         mulps           (B0), %xmm0
  91         mulps           16(B0), %xmm1
  92         mulps           32(B0), %xmm2
  93         mulps           48(B0), %xmm3
  94         mulps           64(B0), %xmm6
  95         mulps           80(B0), %xmm7
  96         addps           %xmm1, %xmm0
  97         addps           %xmm3, %xmm2
  98         addps           %xmm7, %xmm6
  99         movups          160(WINDOW), %xmm1
 100         movups          176(WINDOW), %xmm3
 101         mulps           96(B0), %xmm1
 102         mulps           112(B0), %xmm3
 103         addps           %xmm2, %xmm0
 104         addps           %xmm3, %xmm1
 105         addps           %xmm1, %xmm6
 106         movaps          %xmm6, %xmm7
 107         movaps          %xmm0, %xmm6
 108
 109         leal            256(WINDOW), WINDOW
 110         leal            128(B0), B0
 111
 112         movaps          %xmm4, %xmm0
 113         movaps          %xmm6, %xmm1
 114         unpcklps        %xmm5, %xmm4
 115         unpcklps        %xmm7, %xmm6
 116         unpckhps        %xmm5, %xmm0
 117         unpckhps        %xmm7, %xmm1
 118         movaps          %xmm4, %xmm2
 119         movaps          %xmm0, %xmm3
 120         movlhps         %xmm6, %xmm4
 121         movhlps         %xmm2, %xmm6
 122         movlhps         %xmm1, %xmm0
 123         movhlps         %xmm3, %xmm1
 124         subps           %xmm6, %xmm4
 125         subps           %xmm1, %xmm0
 126         addps           %xmm4, %xmm0
 127
 128         movups          (SAMPLES), %xmm1
 129         movups          16(SAMPLES), %xmm2
 130         mulps           ASM_NAME(scale_sse), %xmm0
 131         shufps          $0xdd, %xmm2, %xmm1
 132         movaps          %xmm0, %xmm2
 133         unpcklps        %xmm1, %xmm0
 134         unpckhps        %xmm1, %xmm2
 135         movups          %xmm0, (SAMPLES)
 136         movups          %xmm2, 16(SAMPLES)
 137
 138         leal            32(SAMPLES), SAMPLES
 139         decl            %ecx
 140         jnz                     Loop_start_1
 141
 142         movl            $4, %ecx
 143
 144         ALIGN16
 145 Loop_start_2:
 146         movups          (WINDOW), %xmm0
 147         movups          16(WINDOW), %xmm1
 148         movups          32(WINDOW), %xmm2
 149         movups          48(WINDOW), %xmm3
 150         movups          128(WINDOW), %xmm4
 151         movups          144(WINDOW), %xmm5
 152         movups          160(WINDOW), %xmm6
 153         movups          176(WINDOW), %xmm7
 154         mulps           0(B0), %xmm0
 155         mulps           16(B0), %xmm1
 156         mulps           32(B0), %xmm2
 157         mulps           48(B0), %xmm3
 158         mulps           -64(B0), %xmm4
 159         mulps           -48(B0), %xmm5
 160         mulps           -32(B0), %xmm6
 161         mulps           -16(B0), %xmm7
 162         addps           %xmm1, %xmm0
 163         addps           %xmm3, %xmm2
 164         addps           %xmm5, %xmm4
 165         addps           %xmm7, %xmm6
 166         addps           %xmm2, %xmm0
 167         addps           %xmm6, %xmm4
 168         movaps          %xmm4, %xmm5
 169         movaps          %xmm0, %xmm4
 170
 171         leal            256(WINDOW), WINDOW
 172         leal            -128(B0), B0
 173
 174         movups          (WINDOW), %xmm0
 175         movups          16(WINDOW), %xmm1
 176         movups          32(WINDOW), %xmm2
 177         movups          48(WINDOW), %xmm3
 178         movups          128(WINDOW), %xmm6
 179         movups          144(WINDOW), %xmm7
 180         mulps           (B0), %xmm0
 181         mulps           16(B0), %xmm1
 182         mulps           32(B0), %xmm2
 183         mulps           48(B0), %xmm3
 184         mulps           -64(B0), %xmm6
 185         mulps           -48(B0), %xmm7
 186         addps           %xmm1, %xmm0
 187         addps           %xmm3, %xmm2
 188         addps           %xmm7, %xmm6
 189         movups          160(WINDOW), %xmm1
 190         movups          176(WINDOW), %xmm3
 191         mulps           -32(B0), %xmm1
 192         mulps           -16(B0), %xmm3
 193         addps           %xmm2, %xmm0
 194         addps           %xmm3, %xmm1
 195         addps           %xmm1, %xmm6
 196         movaps          %xmm6, %xmm7
 197         movaps          %xmm0, %xmm6
 198
 199         leal            256(WINDOW), WINDOW
 200         leal            -128(B0), B0
 201
 202         movaps          %xmm4, %xmm0
 203         movaps          %xmm6, %xmm1
 204         unpcklps        %xmm5, %xmm4
 205         unpcklps        %xmm7, %xmm6
 206         unpckhps        %xmm5, %xmm0
 207         unpckhps        %xmm7, %xmm1
 208         movaps          %xmm4, %xmm2
 209         movaps          %xmm0, %xmm3
 210         movlhps         %xmm6, %xmm4
 211         movhlps         %xmm2, %xmm6
 212         movlhps         %xmm1, %xmm0
 213         movhlps         %xmm3, %xmm1
 214         addps           %xmm6, %xmm4
 215         addps           %xmm1, %xmm0
 216         addps           %xmm4, %xmm0
 217
 218         movups          (SAMPLES), %xmm1
 219         movups          16(SAMPLES), %xmm2
 220         mulps           ASM_NAME(scale_sse), %xmm0
 221         shufps          $0xdd, %xmm2, %xmm1
 222         movaps          %xmm0, %xmm2
 223         unpcklps        %xmm1, %xmm0
 224         unpckhps        %xmm1, %xmm2
 225         movups          %xmm0, (SAMPLES)
 226         movups          %xmm2, 16(SAMPLES)
 227
 228         leal            32(SAMPLES), SAMPLES
 229         decl            %ecx
 230         jnz                     Loop_start_2
 231
 232         xorl            %eax, %eax
 233
 234         popl            %esi
 235         popl            %ebx
 236         movl            %ebp, %esp
 237         popl            %ebp
 238
 239         ret
 240
 241 NONEXEC_STACK