From 2555317a71dff05431db42dc5b2fce7fc0419265 Mon Sep 17 00:00:00 2001
From: Wim Taymans <wtaymans@redhat.com>
Date: Mon, 18 Jan 2016 12:52:41 +0100
Subject: [PATCH] audio-resampler: add float stereo SSE function

---
 gst-libs/gst/audio/audio-resampler-x86.h | 30 ++++++++++++++++++++++++++++++
 gst-libs/gst/audio/audio-resampler.c     | 31 ++++++++++++++++++++++++++++---
 2 files changed, 58 insertions(+), 3 deletions(-)

diff --git a/gst-libs/gst/audio/audio-resampler-x86.h b/gst-libs/gst/audio/audio-resampler-x86.h
index d77726b..c660aa0 100644
--- a/gst-libs/gst/audio/audio-resampler-x86.h
+++ b/gst-libs/gst/audio/audio-resampler-x86.h
@@ -39,7 +39,35 @@ inner_product_gfloat_1_sse (gfloat * o, const gfloat * a, const gfloat * b, gint
   _mm_store_ss (o, sum);
 }
 
+static inline void
+inner_product_gfloat_2_sse (gfloat * o, const gfloat * a, const gfloat * b, gint len)
+{
+  gint i = 0;
+  __m128 sum = _mm_setzero_ps (), t;
+
+  for (; i < len; i += 8) {
+    t = _mm_load_ps (b + i);
+    sum =
+        _mm_add_ps (sum, _mm_mul_ps (_mm_loadu_ps (a + 2 * i + 0),
+            _mm_unpacklo_ps (t, t)));
+    sum =
+        _mm_add_ps (sum, _mm_mul_ps (_mm_loadu_ps (a + 2 * i + 4),
+            _mm_unpackhi_ps (t, t)));
+
+    t = _mm_load_ps (b + i + 4);
+    sum =
+        _mm_add_ps (sum, _mm_mul_ps (_mm_loadu_ps (a + 2 * i + 8),
+            _mm_unpacklo_ps (t, t)));
+    sum =
+        _mm_add_ps (sum, _mm_mul_ps (_mm_loadu_ps (a + 2 * i + 12),
+            _mm_unpackhi_ps (t, t)));
+  }
+  sum = _mm_add_ps (sum, _mm_movehl_ps (sum, sum));
+  *(gint64*)o = _mm_cvtsi128_si64 ((__m128i)sum);
+}
+
 MAKE_RESAMPLE_FUNC (gfloat, 1, sse);
+MAKE_RESAMPLE_FUNC (gfloat, 2, sse);
 #endif
 
 #if defined (HAVE_EMMINTRIN_H) && defined(__SSE2__)
@@ -212,12 +240,14 @@ audio_resampler_check_x86 (const gchar *option)
 #if defined (HAVE_XMMINTRIN_H) && defined(__SSE__)
     GST_DEBUG ("enable SSE optimisations");
     resample_gfloat_1 = resample_gfloat_1_sse;
+    resample_gfloat_2 = resample_gfloat_2_sse;
 #endif
   } else if (!strcmp (option, "sse2")) {
 #if defined (HAVE_EMMINTRIN_H) && defined(__SSE2__)
     GST_DEBUG ("enable SSE2 optimisations");
     resample_gint16_1 = resample_gint16_1_sse2;
     resample_gfloat_1 = resample_gfloat_1_sse;
+    resample_gfloat_2 = resample_gfloat_2_sse;
     resample_gdouble_1 = resample_gdouble_1_sse2;
     resample_gint16_2 = resample_gint16_2_sse2;
     resample_gdouble_2 = resample_gdouble_2_sse2;
diff --git a/gst-libs/gst/audio/audio-resampler.c b/gst-libs/gst/audio/audio-resampler.c
index 5a2f643..68e192a 100644
--- a/gst-libs/gst/audio/audio-resampler.c
+++ b/gst-libs/gst/audio/audio-resampler.c
@@ -424,6 +424,21 @@ inner_product_gfloat_1_c (gfloat * o, const gfloat * a, const gfloat * b,
 }
 
 static inline void
+inner_product_gfloat_2_c (gfloat * o, const gfloat * a, const gfloat * b,
+    gint len)
+{
+  gint i;
+  gfloat r[2] = { 0.0, 0.0 };
+
+  for (i = 0; i < len; i++) {
+    r[0] += a[2 * i] * b[i];
+    r[1] += a[2 * i + 1] * b[i];
+  }
+  o[0] = r[0];
+  o[1] = r[1];
+}
+
+static inline void
 inner_product_gdouble_1_c (gdouble * o, const gdouble * a, const gdouble * b,
     gint len)
 {
@@ -498,6 +513,7 @@ MAKE_RESAMPLE_FUNC (gint32, 1, c);
 MAKE_RESAMPLE_FUNC (gfloat, 1, c);
 MAKE_RESAMPLE_FUNC (gdouble, 1, c);
 MAKE_RESAMPLE_FUNC (gint16, 2, c);
+MAKE_RESAMPLE_FUNC (gfloat, 2, c);
 MAKE_RESAMPLE_FUNC (gdouble, 2, c);
 
 static ResampleFunc resample_funcs[] = {
@@ -506,6 +522,7 @@ static ResampleFunc resample_funcs[] = {
   resample_gfloat_1_c,
   resample_gdouble_1_c,
   resample_gint16_2_c,
+  resample_gfloat_2_c,
   resample_gdouble_2_c,
 };
 
@@ -514,7 +531,8 @@ static ResampleFunc resample_funcs[] = {
 #define resample_gfloat_1 resample_funcs[2]
 #define resample_gdouble_1 resample_funcs[3]
 #define resample_gint16_2 resample_funcs[4]
-#define resample_gdouble_2 resample_funcs[5]
+#define resample_gfloat_2 resample_funcs[5]
+#define resample_gdouble_2 resample_funcs[6]
 
 #if defined HAVE_ORC && !defined DISABLE_ORC
 # if defined (__i386__) || defined (__x86_64__)
@@ -739,8 +757,15 @@ resampler_calculate_taps (GstAudioResampler * resampler)
       }
       break;
     case GST_AUDIO_FORMAT_F32:
-      resampler->resample = resample_gfloat_1;
-      resampler->deinterleave = deinterleave_gfloat;
+      if (!non_interleaved && resampler->channels == 2 && n_taps >= 4) {
+        resampler->resample = resample_gfloat_2;
+        resampler->deinterleave = deinterleave_copy;
+        resampler->blocks = 1;
+        resampler->inc = resampler->channels;;
+      } else {
+        resampler->resample = resample_gfloat_1;
+        resampler->deinterleave = deinterleave_gfloat;
+      }
       break;
     case GST_AUDIO_FORMAT_S32:
       resampler->resample = resample_gint32_1;
-- 
2.7.4