[PATCH 2/2] Improve performance of peaks resampler

[mkbosmans@xxxxxxxxx (Maarten Bosmans)]

This is mainly achieved by special-casing the common 1ch float case,
which is used by applications such as pavucontrol.

Performance is improved by 35% for a mono 48000->25Hz peaks resampling.
---
 src/pulsecore/resampler.c |   37 ++++++++++++++++++++++++++-----------
 1 files changed, 26 insertions(+), 11 deletions(-)

diff --git a/src/pulsecore/resampler.c b/src/pulsecore/resampler.c
index cbf5782..dd9c1aa 100644
--- a/src/pulsecore/resampler.c
+++ b/src/pulsecore/resampler.c
@@ -1443,7 +1443,9 @@ static void peaks_resample(pa_resampler *r, const pa_memchunk *input, unsigned i
     pa_assert(input);
     pa_assert(output);
     pa_assert(out_n_frames);
+    pa_assert(r->i_ss.rate >= r->o_ss.rate);
     pa_assert(r->i_ss.channels == r->o_ss.channels);
+    pa_assert(r->work_format == PA_SAMPLE_S16NE || r->work_format == PA_SAMPLE_FLOAT32NE);
 
     src = (uint8_t*) pa_memblock_acquire(input->memblock) + input->index;
     dst = (uint8_t*) pa_memblock_acquire(output->memblock) + output->index;
@@ -1455,19 +1457,34 @@ static void peaks_resample(pa_resampler *r, const pa_memchunk *input, unsigned i
         i_end = ((r->peaks.o_counter+1) * r->i_ss.rate) / r->o_ss.rate;
         i_end = i_end > r->peaks.i_counter ? i_end - r->peaks.i_counter : 0;
 
-        pa_assert(o_index * r->w_sz * r->o_ss.channels < pa_memblock_get_length(output->memblock));
+        pa_assert_fp(o_index * r->w_sz * r->o_ss.channels < pa_memblock_get_length(output->memblock));
 
-        if (r->work_format == PA_SAMPLE_S16NE) {
+        /* 1ch float is treated separately, because that is the common case */
+        if (r->o_ss.channels == 1 && r->work_format == PA_SAMPLE_FLOAT32NE) {
+            float *s = (float*) src + i;
+            float *d = (float*) dst + o_index;
+
+            for (; i < i_end && i < in_n_frames; i++) {
+                float n = fabsf(*s++);
+
+                if (n > r->peaks.max_f[0])
+                    r->peaks.max_f[0] = n;
+            }
+
+            if (i == i_end) {
+                *d = r->peaks.max_f[0];
+                r->peaks.max_f[0] = 0;
+                o_index++, r->peaks.o_counter++;
+            }
+        } else if (r->work_format == PA_SAMPLE_S16NE) {
             int16_t *s = (int16_t*) src + r->i_ss.channels * i;
             int16_t *d = (int16_t*) dst + r->o_ss.channels * o_index;
 
             for (; i < i_end && i < in_n_frames; i++)
-                for (c = 0; c < r->o_ss.channels; c++, s++) {
-                    int16_t n;
+                for (c = 0; c < r->o_ss.channels; c++) {
+                    int16_t n = abs(*s++);
 
-                    n = (int16_t) (*s < 0 ? -*s : *s);
-
-                    if (PA_UNLIKELY(n > r->peaks.max_i[c]))
+                    if (n > r->peaks.max_i[c])
                         r->peaks.max_i[c] = n;
                 }
 
@@ -1482,11 +1499,9 @@ static void peaks_resample(pa_resampler *r, const pa_memchunk *input, unsigned i
             float *s = (float*) src + r->i_ss.channels * i;
             float *d = (float*) dst + r->o_ss.channels * o_index;
 
-            pa_assert(r->work_format == PA_SAMPLE_FLOAT32NE);
-
             for (; i < i_end && i < in_n_frames; i++)
-                for (c = 0; c < r->o_ss.channels; c++, s++) {
-                    float n = fabsf(*s);
+                for (c = 0; c < r->o_ss.channels; c++) {
+                    float n = fabsf(*s++);
 
                     if (n > r->peaks.max_f[c])
                         r->peaks.max_f[c] = n;
-- 
1.7.4.1