code.delx.au - pulseaudio/blob - src/pulsecore/sconv_neon.c

   1 /***
   2   This file is part of PulseAudio.
   3
   4   Copyright 2012 Peter Meerwald <p.meerwald@bct-electronic.com>
   5
   6   PulseAudio is free software; you can redistribute it and/or modify
   7   it under the terms of the GNU Lesser General Public License as published
   8   by the Free Software Foundation; either version 2.1 of the License,
   9   or (at your option) any later version.
  10
  11   PulseAudio is distributed in the hope that it will be useful, but
  12   WITHOUT ANY WARRANTY; without even the implied warranty of
  13   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14   General Public License for more details.
  15 ***/
  16
  17 #ifdef HAVE_CONFIG_H
  18 #include <config.h>
  19 #endif
  20
  21 #include <pulse/rtclock.h>
  22
  23 #include <pulsecore/macro.h>
  24 #include <pulsecore/endianmacros.h>
  25
  26 #include "cpu-arm.h"
  27 #include "sconv.h"
  28
  29 #include <math.h>
  30 #include <arm_neon.h>
  31
  32 static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t *dst) {
  33     unsigned i = n & 3;
  34
  35     __asm__ __volatile__ (
  36         "movs       %[n], %[n], lsr #2      \n\t"
  37         "beq        2f                      \n\t"
  38
  39         "1:                                 \n\t"
  40         "vld1.32    {q0}, [%[src]]!         \n\t"
  41         "vcvt.s32.f32 q0, q0, #31           \n\t" /* s32<-f32 as 16:16 fixed-point, with implicit multiplication by 32768 */
  42         "vqrshrn.s32 d0, q0, #16            \n\t" /* shift, round, narrow */
  43         "subs       %[n], %[n], #1          \n\t"
  44         "vst1.16    {d0}, [%[dst]]!         \n\t"
  45         "bgt        1b                      \n\t"
  46
  47         "2:                                 \n\t"
  48
  49         : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */
  50         : /* input operands */
  51         : "memory", "cc", "q0" /* clobber list */
  52     );
  53
  54     /* leftovers */
  55     while (i--) {
  56         *dst++ = (int16_t) PA_CLAMP_UNLIKELY(lrintf(*src * (1 << 15)), -0x8000, 0x7FFF);
  57         src++;
  58     }
  59 }
  60
  61 static void pa_sconv_s16le_to_f32ne_neon(unsigned n, const int16_t *src, float *dst) {
  62     unsigned i = n & 3;
  63
  64     const float invscale = 1.0f / (1 << 15);
  65
  66     __asm__ __volatile__ (
  67         "movs        %[n], %[n], lsr #2     \n\t"
  68         "beq        2f                      \n\t"
  69
  70         "vdup.f32   q1, %[invscale]         \n\t"
  71
  72         "1:                                 \n\t"
  73         "vld1.16    {d0}, [%[src]]!         \n\t"
  74         "vmovl.s16  q0, d0                  \n\t" /* widen */
  75         "vcvt.f32.s32 q0, q0                \n\t" /* f32<-s32 */
  76         "vmul.f32   q0, q0, q1              \n\t"
  77         "subs       %[n], %[n], #1          \n\t"
  78         "vst1.32    {q0}, [%[dst]]!         \n\t"
  79         "bgt        1b                      \n\t"
  80
  81         "2:                                 \n\t"
  82
  83         : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */
  84         : [invscale] "r" (invscale) /* input operands */
  85         : "memory", "cc", "q0", "q1" /* clobber list */
  86     );
  87
  88     /* leftovers */
  89     while (i--) {
  90         *dst++ = *src++ * invscale;
  91     }
  92 }
  93
  94 void pa_convert_func_init_neon(pa_cpu_arm_flag_t flags) {
  95     pa_log_info("Initialising ARM NEON optimized conversions.");
  96     pa_set_convert_from_float32ne_function(PA_SAMPLE_S16LE, (pa_convert_func_t) pa_sconv_s16le_from_f32ne_neon);
  97     pa_set_convert_to_float32ne_function(PA_SAMPLE_S16LE, (pa_convert_func_t) pa_sconv_s16le_to_f32ne_neon);
  98 }