svolume: Fix ARM alignment issues

As Peter Meerwald <p.meerwald@bct-electronic.com> discovered, our ARM
svolume code performance is quite terrible when the incoming samples are
not word-aligned. This can very easily be the case, since the
architecture only requires that the samples be 16-bit aligned, and we
might end up running the innermost loop after processing modulo-4
samples. The performance degradation was ~50x on a Cortex A9
(Pandaboard).

This reworks the svolume logic to first consume enough samples to make
sure the rest is word aligned, and reordering the processing to work
with 4 samples at a time first, and then finally deal with the
remainder.

With this, performance is comparable for arbitrary alignments (~3x
faster than the C code).

1 files changed, 60 insertions, 41 deletions

diff --git a/src/pulsecore/svolume_arm.c b/src/pulsecore/svolume_arm.c
index 08b33888e..f325d774c 100644
--- a/src/pulsecore/svolume_arm.c
+++ b/src/pulsecore/svolume_arm.c
@@ -40,101 +40,117 @@
     " addcs r0, %1                  \n\t" \
     " movcs r6, r0                  \n\t"
 
+static pa_do_volume_func_t _volume_ref;
+
 static void pa_volume_s16ne_arm(int16_t *samples, const int32_t *volumes, unsigned channels, unsigned length) {
     /* Channels must be at least 4, and always a multiple of the original number.
      * This is also the max amount we overread the volume array, which should
      * have enough padding. */
     const int32_t *ve = volumes + (channels == 3 ? 6 : PA_MAX (4U, channels));
+    unsigned rem = PA_ALIGN((size_t) samples) - (size_t) samples;
 
-    __asm__ __volatile__ (
-        " mov r6, %1                      \n\t" /* r6 = volumes */
-        " mov %3, %3, LSR #1              \n\t" /* length /= sizeof (int16_t) */
-        " tst %3, #1                      \n\t" /* check for odd samples */
-        " beq  2f                         \n\t"
+    /* Make sure we're word-aligned, else performance _really_ sucks */
+    if (rem) {
+        _volume_ref(samples, volumes, channels, rem < length ? rem : length);
 
-        "1:                               \n\t" /* odd samples volumes */
-        " ldr  r0, [r6], #4               \n\t" /* r0 = volume */
-        " ldrh r2, [%0]                   \n\t" /* r2 = sample */
-
-        " smulwb r0, r0, r2               \n\t" /* r0 = (r0 * r2) >> 16 */
-        " ssat r0, #16, r0                \n\t" /* r0 = PA_CLAMP(r0, 0x7FFF) */
+        if (rem < length) {
+            length -= rem;
+            samples += rem / sizeof(*samples);
+        } else
+            return; /* we're done */
+    }
 
-        " strh r0, [%0], #2               \n\t" /* sample = r0 */
+    __asm__ __volatile__ (
+        " mov r6, %4                      \n\t" /* r6 = volumes + rem */
+        " mov %3, %3, LSR #1              \n\t" /* length /= sizeof (int16_t) */
 
-        MOD_INC()
+        " cmp %3, #4                      \n\t" /* check for 4+ samples */
+        " blt 2f                          \n\t"
 
-        "2:                               \n\t"
-        " mov %3, %3, LSR #1              \n\t"
-        " tst %3, #1                      \n\t" /* check for odd samples */
-        " beq  4f                         \n\t"
+        /* See final case for how the multiplication works */
 
-        "3:                               \n\t"
-        " ldrd r2, [r6], #8               \n\t" /* 2 samples at a time */
-        " ldr  r0, [%0]                   \n\t"
+        "1:                               \n\t"
+        " ldrd r2, [r6], #8               \n\t" /* 4 samples at a time */
+        " ldrd r4, [r6], #8               \n\t"
+        " ldrd r0, [%0]                   \n\t"
 
 #ifdef WORDS_BIGENDIAN
         " smulwt r2, r2, r0               \n\t"
         " smulwb r3, r3, r0               \n\t"
+        " smulwt r4, r4, r1               \n\t"
+        " smulwb r5, r5, r1               \n\t"
 #else
         " smulwb r2, r2, r0               \n\t"
         " smulwt r3, r3, r0               \n\t"
+        " smulwb r4, r4, r1               \n\t"
+        " smulwt r5, r5, r1               \n\t"
 #endif
 
         " ssat r2, #16, r2                \n\t"
         " ssat r3, #16, r3                \n\t"
+        " ssat r4, #16, r4                \n\t"
+        " ssat r5, #16, r5                \n\t"
 
 #ifdef WORDS_BIGENDIAN
         " pkhbt r0, r3, r2, LSL #16       \n\t"
+        " pkhbt r1, r5, r4, LSL #16       \n\t"
 #else
         " pkhbt r0, r2, r3, LSL #16       \n\t"
+        " pkhbt r1, r4, r5, LSL #16       \n\t"
 #endif
-        " str  r0, [%0], #4               \n\t"
+        " strd  r0, [%0], #8              \n\t"
 
         MOD_INC()
 
-        "4:                               \n\t"
-        " movs %3, %3, LSR #1             \n\t"
-        " beq  6f                         \n\t"
+        " subs %3, %3, #4                 \n\t"
+        " cmp %3, #4                      \n\t"
+        " bge 1b                          \n\t"
 
-        "5:                               \n\t"
-        " ldrd r2, [r6], #8               \n\t" /* 4 samples at a time */
-        " ldrd r4, [r6], #8               \n\t"
-        " ldrd r0, [%0]                   \n\t"
+        "2:                               \n\t"
+        " cmp %3, #2                      \n\t"
+        " blt 3f                          \n\t"
+
+        " ldrd r2, [r6], #8               \n\t" /* 2  samples at a time */
+        " ldr  r0, [%0]                   \n\t"
 
 #ifdef WORDS_BIGENDIAN
         " smulwt r2, r2, r0               \n\t"
         " smulwb r3, r3, r0               \n\t"
-        " smulwt r4, r4, r1               \n\t"
-        " smulwb r5, r5, r1               \n\t"
 #else
         " smulwb r2, r2, r0               \n\t"
         " smulwt r3, r3, r0               \n\t"
-        " smulwb r4, r4, r1               \n\t"
-        " smulwt r5, r5, r1               \n\t"
 #endif
 
         " ssat r2, #16, r2                \n\t"
         " ssat r3, #16, r3                \n\t"
-        " ssat r4, #16, r4                \n\t"
-        " ssat r5, #16, r5                \n\t"
 
 #ifdef WORDS_BIGENDIAN
         " pkhbt r0, r3, r2, LSL #16       \n\t"
-        " pkhbt r1, r5, r4, LSL #16       \n\t"
 #else
         " pkhbt r0, r2, r3, LSL #16       \n\t"
-        " pkhbt r1, r4, r5, LSL #16       \n\t"
 #endif
-        " strd  r0, [%0], #8              \n\t"
+        " str  r0, [%0], #4               \n\t"
 
         MOD_INC()
 
-        " subs %3, %3, #1                 \n\t"
-        " bne 5b                          \n\t"
-        "6:                               \n\t"
+        " subs %3, %3, #2                 \n\t"
+
+        "3:                               \n\t" /* check for odd # of samples */
+        " cmp %3, #1                      \n\t"
+        " bne 4f                          \n\t"
+
+        " ldr  r0, [r6], #4               \n\t" /* r0 = volume */
+        " ldrh r2, [%0]                   \n\t" /* r2 = sample */
+
+        " smulwb r0, r0, r2               \n\t" /* r0 = (r0 * r2) >> 16 */
+        " ssat r0, #16, r0                \n\t" /* r0 = PA_CLAMP(r0, 0x7FFF) */
+
+        " strh r0, [%0], #2               \n\t" /* sample = r0 */
+
+        "4:                               \n\t"
 
         : "+r" (samples), "+r" (volumes), "+r" (ve), "+r" (length)
-        :
+        : "r" (volumes + ((rem / sizeof(*samples)) % channels))
         : "r6", "r5", "r4", "r3", "r2", "r1", "r0", "cc"
     );
 }
@@ -145,6 +161,9 @@ void pa_volume_func_init_arm(pa_cpu_arm_flag_t flags) {
 #if defined (__arm__) && defined (HAVE_ARMV6)
     pa_log_info("Initialising ARM optimized volume functions.");
 
+    if (!_volume_ref)
+        _volume_ref = pa_get_volume_func(PA_SAMPLE_S16NE);
+
     pa_set_volume_func(PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_arm);
 #endif /* defined (__arm__) && defined (HAVE_ARMV6) */
 }