openal-soft: use NEON for float conversion on output

diff --git a/recipes/openal/openal-soft/neon_write_converter.patch b/recipes/openal/openal-soft/neon_write_converter.patch
new file mode 100644 (file)
index 0000000..76865a4
--- /dev/null
+++ b/recipes/openal/openal-soft/neon_write_converter.patch
@@ -0,0 +1,83 @@
+diff -ur openal-soft-1.15.1.orig/Alc/ALu.c openal-soft-1.15.1/Alc/ALu.c
+--- openal-soft-1.15.1.orig/Alc/ALu.c  2014-07-02 03:36:17.874323362 +0300
++++ openal-soft-1.15.1/Alc/ALu.c       2014-07-03 02:41:18.000116397 +0300
+@@ -956,11 +956,78 @@
+     return SamplesToDo*numchans*sizeof(T);                                    \
+ }
+ 
++static int Write_ALshort(ALCdevice *device, ALshort *RESTRICT buffer, ALuint SamplesToDo)
++{
++    ALfloat (*RESTRICT DryBuffer)[BUFFERSIZE] = device->DryBuffer;
++    ALuint numchans = ChannelsFromDevFmt(device->FmtChans);
++    const ALuint *offsets = device->ChannelOffsets;
++    ALuint i, j;
++
++#ifdef __ARM_NEON__
++    if (numchans == 2 && offsets[0] == 0 && offsets[1] == 1)
++    {
++        ALfloat *cl = DryBuffer[0];
++        ALfloat *cr = DryBuffer[1];
++        ALuint samples = SamplesToDo;
++        asm volatile (
++             "movw       r3, #0x0000fe00\n"
++             "movt       r3, #0x46ff\n"
++             "vdup.32    d4, r3 @ 32767.0\n"
++            "0:\n"
++             "vld1.32    {q0}, [%1, :128]!\n"
++             "vld1.32    {q1}, [%2, :128]!\n"
++             "subs       %3, #4\n"
++             "vmul.f32   q0, d4[0]\n"
++             "vmul.f32   q1, d4[0]\n"
++             "vcvt.s32.f32 q0, q0\n"
++             "vcvt.s32.f32 q1, q1\n"
++             "pld        [%1, #64*2]\n"
++             "pld        [%2, #64*2]\n"
++             "vqmovn.s32 d0, q0\n"
++             "vqmovn.s32 d1, q1\n"
++             "blt        1f\n"
++             "vst2.16    {d0,d1}, [%0]!\n"
++             "bgt        0b\n"
++             "nop\n"
++             "b          2f\n" /* eq 4 - all done */
++            "1:\n"
++             "vzip.16    q0, q0\n"
++             "add        %3, #4\n"
++             "vst1.32    {d0[0]}, [%0]!\n"
++             "cmp        %3, #1\n"
++             "ble        2f\n"
++             "vst1.32    {d0[1]}, [%0]!\n"
++             "cmp        %3, #2\n"
++             "ble        2f\n"
++             "vst1.32    {d1[0]}, [%0]!\n"
++            "2:\n"
++            : "=&r"(buffer), "=&r"(cl), "=&r"(cr), "=&r"(samples)
++            : "0"(buffer), "1"(cl), "2"(cr), "3"(samples)
++            : "r3", "q0", "q1", "d4", "cc", "memory"
++        );
++        return SamplesToDo * numchans * sizeof(ALshort);
++    }
++#endif
++
++    for(j = 0;j < MaxChannels;j++)
++    {
++        ALshort *RESTRICT out;
++
++        if(offsets[j] == INVALID_OFFSET)
++            continue;
++
++        out = buffer + offsets[j];
++        for(i = 0;i < SamplesToDo;i++)
++            out[i * numchans] = aluF2S(DryBuffer[j][i]);
++    }
++    return SamplesToDo * numchans * sizeof(ALshort);
++}
++
+ DECL_TEMPLATE(ALfloat, aluF2F)
+ DECL_TEMPLATE(ALuint, aluF2UI)
+ DECL_TEMPLATE(ALint, aluF2I)
+ DECL_TEMPLATE(ALushort, aluF2US)
+-DECL_TEMPLATE(ALshort, aluF2S)
++//DECL_TEMPLATE(ALshort, aluF2S)
+ DECL_TEMPLATE(ALubyte, aluF2UB)
+ DECL_TEMPLATE(ALbyte, aluF2B)
+ 

diff --git a/recipes/openal/openal-soft_1.15.1.bb b/recipes/openal/openal-soft_1.15.1.bb
index 077a258..5f4f69a 100644 (file)
--- a/recipes/openal/openal-soft_1.15.1.bb
+++ b/recipes/openal/openal-soft_1.15.1.bb
@@ -3,7 +3,7 @@ HOMEPAGE = "http://kcat.strangesoft.net/openal.html"
 LICENSE = "GPLv2"
 LIC_FILES_CHKSUM = "file://COPYING;md5=facc3a8f452930083bbb95d82b989c35"
 
-PR = "r1"
+PR = "r2"
 
 inherit cmake
 
@@ -14,6 +14,7 @@ DEPENDS = "${@base_contains('DISTRO_FEATURES', 'alsa', 'alsa-lib', '', d)} \
 SRC_URI = " \
        http://kcat.strangesoft.net/openal-releases/${PN}-${PV}.tar.bz2 \
        file://default_to_short.patch;patch=1 \
+       file://neon_write_converter.patch;patch=1 \
 "
 
 SRC_URI[md5sum] = "ea83dec3b9655a27d28e7bc7cae9cd71"