recipes/openal/openal-soft/neon_write_converter.patch

   1 diff -ur openal-soft-1.15.1.orig/Alc/ALu.c openal-soft-1.15.1/Alc/ALu.c
   2 --- openal-soft-1.15.1.orig/Alc/ALu.c   2014-07-02 03:36:17.874323362 +0300
   3 +++ openal-soft-1.15.1/Alc/ALu.c        2014-07-03 02:41:18.000116397 +0300
   4 @@ -956,11 +956,78 @@
   5      return SamplesToDo*numchans*sizeof(T);                                    \
   6  }
   7
   8 +static int Write_ALshort(ALCdevice *device, ALshort *RESTRICT buffer, ALuint SamplesToDo)
   9 +{
  10 +    ALfloat (*RESTRICT DryBuffer)[BUFFERSIZE] = device->DryBuffer;
  11 +    ALuint numchans = ChannelsFromDevFmt(device->FmtChans);
  12 +    const ALuint *offsets = device->ChannelOffsets;
  13 +    ALuint i, j;
  14 +
  15 +#ifdef __ARM_NEON__
  16 +    if (numchans == 2 && offsets[0] == 0 && offsets[1] == 1)
  17 +    {
  18 +        ALfloat *cl = DryBuffer[0];
  19 +        ALfloat *cr = DryBuffer[1];
  20 +        ALuint samples = SamplesToDo;
  21 +        asm volatile (
  22 +             "movw       r3, #0x0000fe00\n"
  23 +             "movt       r3, #0x46ff\n"
  24 +             "vdup.32    d4, r3 @ 32767.0\n"
  25 +            "0:\n"
  26 +             "vld1.32    {q0}, [%1, :128]!\n"
  27 +             "vld1.32    {q1}, [%2, :128]!\n"
  28 +             "subs       %3, #4\n"
  29 +             "vmul.f32   q0, d4[0]\n"
  30 +             "vmul.f32   q1, d4[0]\n"
  31 +             "vcvt.s32.f32 q0, q0\n"
  32 +             "vcvt.s32.f32 q1, q1\n"
  33 +             "pld        [%1, #64*2]\n"
  34 +             "pld        [%2, #64*2]\n"
  35 +             "vqmovn.s32 d0, q0\n"
  36 +             "vqmovn.s32 d1, q1\n"
  37 +             "blt        1f\n"
  38 +             "vst2.16    {d0,d1}, [%0]!\n"
  39 +             "bgt        0b\n"
  40 +             "nop\n"
  41 +             "b          2f\n" /* eq 4 - all done */
  42 +            "1:\n"
  43 +             "vzip.16    q0, q0\n"
  44 +             "add        %3, #4\n"
  45 +             "vst1.32    {d0[0]}, [%0]!\n"
  46 +             "cmp        %3, #1\n"
  47 +             "ble        2f\n"
  48 +             "vst1.32    {d0[1]}, [%0]!\n"
  49 +             "cmp        %3, #2\n"
  50 +             "ble        2f\n"
  51 +             "vst1.32    {d1[0]}, [%0]!\n"
  52 +            "2:\n"
  53 +            : "=&r"(buffer), "=&r"(cl), "=&r"(cr), "=&r"(samples)
  54 +            : "0"(buffer), "1"(cl), "2"(cr), "3"(samples)
  55 +            : "r3", "q0", "q1", "d4", "cc", "memory"
  56 +        );
  57 +        return SamplesToDo * numchans * sizeof(ALshort);
  58 +    }
  59 +#endif
  60 +
  61 +    for(j = 0;j < MaxChannels;j++)
  62 +    {
  63 +        ALshort *RESTRICT out;
  64 +
  65 +        if(offsets[j] == INVALID_OFFSET)
  66 +            continue;
  67 +
  68 +        out = buffer + offsets[j];
  69 +        for(i = 0;i < SamplesToDo;i++)
  70 +            out[i * numchans] = aluF2S(DryBuffer[j][i]);
  71 +    }
  72 +    return SamplesToDo * numchans * sizeof(ALshort);
  73 +}
  74 +
  75  DECL_TEMPLATE(ALfloat, aluF2F)
  76  DECL_TEMPLATE(ALuint, aluF2UI)
  77  DECL_TEMPLATE(ALint, aluF2I)
  78  DECL_TEMPLATE(ALushort, aluF2US)
  79 -DECL_TEMPLATE(ALshort, aluF2S)
  80 +//DECL_TEMPLATE(ALshort, aluF2S)
  81  DECL_TEMPLATE(ALubyte, aluF2UB)
  82  DECL_TEMPLATE(ALbyte, aluF2B)
  83