Resync

Author: inactive123

audio/audio_mix.c

@@ -114,7 +114,8 @@ void audio_mix_free_chunk(audio_chunk_t *chunk)
    free(chunk);
 }
 
-audio_chunk_t* audio_mix_load_wav_file(const char *path, int sample_rate)
+audio_chunk_t* audio_mix_load_wav_file(const char *path, int sample_rate,
+      const char *resampler_ident, enum resampler_quality quality)
 {
 #ifdef HAVE_RWAV
    int sample_size;
@@ -233,8 +234,8 @@ audio_chunk_t* audio_mix_load_wav_file(const char *path, int sample_rate)
 
       retro_resampler_realloc(&chunk->resampler_data,
             &chunk->resampler,
-            NULL,
-            RESAMPLER_QUALITY_DONTCARE,
+            resampler_ident,
+            quality,
             chunk->ratio);
 
       if (chunk->resampler && chunk->resampler_data)

audio/audio_mixer.c

@@ -267,15 +267,16 @@ static bool wav_to_float(const rwav_t* wav, float** pcm, size_t samples_out)
 }
 
 static bool one_shot_resample(const float* in, size_t samples_in,
-      unsigned rate, float** out, size_t* samples_out)
+      unsigned rate, const char *resampler_ident, enum resampler_quality quality,
+      float** out, size_t* samples_out)
 {
    struct resampler_data info;
    void* data                         = NULL;
    const retro_resampler_t* resampler = NULL;
    float ratio                        = (double)s_rate / (double)rate;
 
-   if (!retro_resampler_realloc(&data, &resampler, NULL,
-            RESAMPLER_QUALITY_DONTCARE, ratio))
+   if (!retro_resampler_realloc(&data, &resampler,
+         resampler_ident, quality, ratio))
       return false;
 
    /*
@@ -323,7 +324,8 @@ void audio_mixer_done(void)
       s_voices[i].type = AUDIO_MIXER_TYPE_NONE;
 }
 
-audio_mixer_sound_t* audio_mixer_load_wav(void *buffer, int32_t size)
+audio_mixer_sound_t* audio_mixer_load_wav(void *buffer, int32_t size,
+      const char *resampler_ident, enum resampler_quality quality)
 {
 #ifdef HAVE_RWAV
    /* WAV data */
@@ -353,8 +355,9 @@ audio_mixer_sound_t* audio_mixer_load_wav(void *buffer, int32_t size)
    {
       float* resampled           = NULL;
 
-      if (!one_shot_resample(pcm, samples,
-               wav.samplerate, &resampled, &samples))
+      if (!one_shot_resample(pcm, samples, wav.samplerate,
+            resampler_ident, quality,
+            &resampled, &samples))
          return NULL;
 
       memalign_free((void*)pcm);
@@ -514,6 +517,8 @@ static bool audio_mixer_play_ogg(
       audio_mixer_sound_t* sound,
       audio_mixer_voice_t* voice,
       bool repeat, float volume,
+      const char *resampler_ident,
+      enum resampler_quality quality,
       audio_mixer_stop_cb_t stop_cb)
 {
    stb_vorbis_info info;
@@ -537,7 +542,7 @@ static bool audio_mixer_play_ogg(
       ratio = (double)s_rate / (double)info.sample_rate;
 
       if (!retro_resampler_realloc(&resampler_data,
-               &resamp, NULL, RESAMPLER_QUALITY_DONTCARE,
+               &resamp, resampler_ident, quality,
                ratio))
          goto error;
    }
@@ -662,6 +667,8 @@ static bool audio_mixer_play_flac(
       audio_mixer_sound_t* sound,
       audio_mixer_voice_t* voice,
       bool repeat, float volume,
+      const char *resampler_ident,
+      enum resampler_quality quality,
       audio_mixer_stop_cb_t stop_cb)
 {
    float ratio                     = 1.0f;
@@ -678,7 +685,7 @@ static bool audio_mixer_play_flac(
       ratio = (double)s_rate / (double)(dr_flac->sampleRate);
 
       if (!retro_resampler_realloc(&resampler_data,
-               &resamp, NULL, RESAMPLER_QUALITY_DONTCARE,
+               &resamp, resampler_ident, quality,
                ratio))
          goto error;
    }
@@ -723,6 +730,8 @@ static bool audio_mixer_play_mp3(
       audio_mixer_sound_t* sound,
       audio_mixer_voice_t* voice,
       bool repeat, float volume,
+      const char *resampler_ident,
+      enum resampler_quality quality,
       audio_mixer_stop_cb_t stop_cb)
 {
    float ratio                     = 1.0f;
@@ -748,7 +757,7 @@ static bool audio_mixer_play_mp3(
       ratio = (double)s_rate / (double)(voice->types.mp3.stream.sampleRate);
 
       if (!retro_resampler_realloc(&resampler_data,
-               &resamp, NULL, RESAMPLER_QUALITY_DONTCARE,
+               &resamp, resampler_ident, quality,
                ratio))
          goto error;
    }
@@ -786,8 +795,11 @@ static bool audio_mixer_play_mp3(
 }
 #endif
 
-audio_mixer_voice_t* audio_mixer_play(audio_mixer_sound_t* sound, bool repeat,
-      float volume, audio_mixer_stop_cb_t stop_cb)
+audio_mixer_voice_t* audio_mixer_play(audio_mixer_sound_t* sound,
+      bool repeat, float volume,
+      const char *resampler_ident,
+      enum resampler_quality quality,
+      audio_mixer_stop_cb_t stop_cb)
 {
    unsigned i;
    bool res                   = false;
@@ -808,7 +820,8 @@ audio_mixer_voice_t* audio_mixer_play(audio_mixer_sound_t* sound, bool repeat,
             break;
          case AUDIO_MIXER_TYPE_OGG:
 #ifdef HAVE_STB_VORBIS
-            res = audio_mixer_play_ogg(sound, voice, repeat, volume, stop_cb);
+            res = audio_mixer_play_ogg(sound, voice, repeat, volume,
+                  resampler_ident, quality, stop_cb);
 #endif
             break;
          case AUDIO_MIXER_TYPE_MOD:
@@ -818,12 +831,14 @@ audio_mixer_voice_t* audio_mixer_play(audio_mixer_sound_t* sound, bool repeat,
             break;
          case AUDIO_MIXER_TYPE_FLAC:
 #ifdef HAVE_DR_FLAC
-            res = audio_mixer_play_flac(sound, voice, repeat, volume, stop_cb);
+            res = audio_mixer_play_flac(sound, voice, repeat, volume,
+                  resampler_ident, quality, stop_cb);
 #endif
             break;
          case AUDIO_MIXER_TYPE_MP3:
 #ifdef HAVE_DR_MP3
-            res = audio_mixer_play_mp3(sound, voice, repeat, volume, stop_cb);
+            res = audio_mixer_play_mp3(sound, voice, repeat, volume,
+                  resampler_ident, quality, stop_cb);
 #endif
             break;
          case AUDIO_MIXER_TYPE_NONE:

audio/conversion/float_to_s16.c

@@ -1,4 +1,4 @@
-/* Copyright  (C) 2010-2020 The RetroArch team
+/* Copyright  (C) 2010-2021 The RetroArch team
  *
  * ---------------------------------------------------------------------------------------
  * The following license statement only applies to this file (float_to_s16.c).
@@ -28,37 +28,75 @@
 #include <altivec.h>
 #endif
 
-#if (defined(__ARM_NEON__) && !defined(DONT_WANT_ARM_ASM_OPTIMIZATIONS)) || defined(HAVE_NEON)
-#ifndef HAVE_ARM_NEON_ASM_OPTIMIZATIONS
-#define HAVE_ARM_NEON_ASM_OPTIMIZATIONS
-#endif
-#endif
-
 #include <features/features_cpu.h>
 #include <audio/conversion/float_to_s16.h>
 
-#if defined(HAVE_ARM_NEON_ASM_OPTIMIZATIONS)
+#if (defined(__ARM_NEON__) || defined(HAVE_NEON))
 static bool float_to_s16_neon_enabled = false;
-void convert_float_s16_asm(int16_t *out, const float *in, size_t samples);
+#ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS
+void convert_float_s16_asm(int16_t *out,
+      const float *in, size_t samples);
+#else
+#include <arm_neon.h>
 #endif
 
-/**
- * convert_float_to_s16:
- * @out               : output buffer
- * @in                : input buffer
- * @samples           : size of samples to be converted
- *
- * Converts floating point
- * to signed integer 16-bit.
- *
- * C implementation callback function.
- **/
 void convert_float_to_s16(int16_t *out,
       const float *in, size_t samples)
 {
-   size_t i      = 0;
+   size_t i           = 0;
+   if (float_to_s16_neon_enabled)
+   {
+      float        gf = (1<<15);
+      float32x4_t vgf = {gf, gf, gf, gf};
+      while (samples >= 8)
+      {
+#ifdef HAVE_ARM_NEON_ASM_OPTIMIZATIONS
+         size_t aligned_samples = samples & ~7;
+         if (aligned_samples)
+            convert_float_s16_asm(out, in, aligned_samples);
+
+         out                += aligned_samples;
+         in                 += aligned_samples;
+         samples            -= aligned_samples;
+         i                   = 0;
+#else
+         int16x4x2_t oreg;
+         int32x4x2_t creg;
+         float32x4x2_t inreg = vld2q_f32(in);
+         creg.val[0]         = vcvtq_s32_f32(vmulq_f32(inreg.val[0], vgf));
+         creg.val[1]         = vcvtq_s32_f32(vmulq_f32(inreg.val[1], vgf));
+         oreg.val[0]         = vqmovn_s32(creg.val[0]);
+         oreg.val[1]         = vqmovn_s32(creg.val[1]);
+         vst2_s16(out, oreg);
+         in                 += 8;
+         out                += 8;
+         samples            -= 8;
+#endif
+      }
+   }
+
+   for (; i < samples; i++)
+   {
+      int32_t val = (int32_t)(in[i] * 0x8000);
+      out[i]      = (val > 0x7FFF) ? 0x7FFF :
+         (val < -0x8000 ? -0x8000 : (int16_t)val);
+   }
+}
+
+void convert_float_to_s16_init_simd(void)
+{
+   unsigned cpu = cpu_features_get();
+
+   if (cpu & RETRO_SIMD_NEON)
+      float_to_s16_neon_enabled = true;
+}
+#else
+void convert_float_to_s16(int16_t *out,
+      const float *in, size_t samples)
+{
+   size_t i          = 0;
 #if defined(__SSE2__)
-   __m128 factor = _mm_set1_ps((float)0x8000);
+   __m128 factor     = _mm_set1_ps((float)0x8000);
 
    for (i = 0; i + 8 <= samples; i += 8, in += 8, out += 8)
    {
@@ -73,10 +111,10 @@ void convert_float_to_s16(int16_t *out,
       _mm_storeu_si128((__m128i *)out, packed);
    }
 
-   samples = samples - i;
-   i       = 0;
+   samples           = samples - i;
+   i                 = 0;
 #elif defined(__ALTIVEC__)
-   int samples_in = samples;
+   int samples_in    = samples;
 
    /* Unaligned loads/store is a bit expensive,
     * so we optimize for the good path (very likely). */
@@ -92,25 +130,12 @@ void convert_float_to_s16(int16_t *out,
          vec_st(vec_packs(result0, result1), 0, out);
       }
 
-      samples_in -= i;
+      samples_in    -= i;
    }
 
-   samples = samples_in;
-   i       = 0;
-#elif defined(HAVE_ARM_NEON_ASM_OPTIMIZATIONS)
-   if (float_to_s16_neon_enabled)
-   {
-      size_t aligned_samples = samples & ~7;
-      if (aligned_samples)
-         convert_float_s16_asm(out, in, aligned_samples);
-
-      out     = out     + aligned_samples;
-      in      = in      + aligned_samples;
-      samples = samples - aligned_samples;
-      i       = 0;
-   }
+   samples           = samples_in;
+   i                 = 0;
 #elif defined(_MIPS_ARCH_ALLEGREX)
-
 #ifdef DEBUG
    /* Make sure the buffers are 16 byte aligned, this should be
     * the default behaviour of malloc in the PSPSDK.
@@ -138,29 +163,16 @@ void convert_float_to_s16(int16_t *out,
             ".set    pop                  \n"
             :: "r"(in + i), "r"(out + i));
    }
-
 #endif
 
    for (; i < samples; i++)
    {
-      int32_t val = (int32_t)(in[i] * 0x8000);
-      out[i]      = (val > 0x7FFF) ? 0x7FFF :
-         (val < -0x8000 ? -0x8000 : (int16_t)val);
+      int32_t val    = (int32_t)(in[i] * 0x8000);
+      out[i]         = (val > 0x7FFF) 
+         ? 0x7FFF 
+         : (val < -0x8000 ? -0x8000 : (int16_t)val);
    }
 }
 
-/**
- * convert_float_to_s16_init_simd:
- *
- * Sets up function pointers for conversion
- * functions based on CPU features.
- **/
-void convert_float_to_s16_init_simd(void)
-{
-#if defined(HAVE_ARM_NEON_ASM_OPTIMIZATIONS)
-   unsigned cpu = cpu_features_get();
-
-   if (cpu & RETRO_SIMD_NEON)
-      float_to_s16_neon_enabled = true;
+void convert_float_to_s16_init_simd(void) { }
 #endif
-}

audio/conversion/float_to_s16_neon.S

@@ -19,7 +19,7 @@
  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  */
-#if defined(__ARM_NEON__) && !defined(DONT_WANT_ARM_ASM_OPTIMIZATIONS)
+#if defined(__ARM_NEON__) && defined(HAVE_ARM_NEON_ASM_OPTIMIZATIONS)
 
 #ifndef __MACH__
 .arm

audio/conversion/float_to_s16_neon.c

@@ -19,7 +19,7 @@
  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  */
-#if defined(__ARM_NEON__) && !defined(DONT_WANT_ARM_ASM_OPTIMIZATIONS)
+#if defined(__ARM_NEON__) && defined(HAVE_ARM_NEON_ASM_OPTIMIZATIONS)
 
 #if defined(__thumb__)
 #define DECL_ARMMODE(x) "  .align 2\n" "  .global " x "\n" "  .thumb\n" "  .thumb_func\n" "  .type " x ", %function\n" x ":\n"