From 90349b384fe538b0d160b5ff0485d6235b1c6f9d Mon Sep 17 00:00:00 2001
From: Chris Robinson <chris.kcat@gmail.com>
Date: Sun, 25 Oct 2015 14:07:58 -0700
Subject: Use the correct position in the SSE resamplers for left-over
 processing

---
 Alc/mixer_sse.c   |  4 ++++
 Alc/mixer_sse2.c  |  5 ++++-
 Alc/mixer_sse3.c  | 11 +++++++----
 Alc/mixer_sse41.c | 13 ++++++++-----
 4 files changed, 23 insertions(+), 10 deletions(-)

(limited to 'Alc')

diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index c618f966..e67c9cca 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -182,6 +182,10 @@ void Mix_SSE(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)
                     _mm_store_ps(&OutBuffer[c][OutPos+pos], dry4);
                     pos += 4;
                 } while(minsize-pos > 3);
+                /* NOTE: gain4 now represents the next four gains after the
+                 * last four mixed samples, so the lowest element represents
+                 * the next gain to apply.
+                 */
                 gain = _mm_cvtss_f32(gain4);
             }
             /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
diff --git a/Alc/mixer_sse2.c b/Alc/mixer_sse2.c
index 1787c196..a134b9a6 100644
--- a/Alc/mixer_sse2.c
+++ b/Alc/mixer_sse2.c
@@ -63,7 +63,10 @@ const ALfloat *Resample_lerp32_SSE2(const ALfloat *src, ALuint frac, ALuint incr
         _mm_store_ps(pos_.f, _mm_castsi128_ps(pos4));
     }
 
-    pos = pos_.i[3];
+    /* NOTE: These four elements represent the position *after* the last four
+     * samples, so the lowest element is the next position to resample.
+     */
+    pos = pos_.i[0];
     frac = _mm_cvtsi128_si32(frac4);
 
     for(;i < numsamples;i++)
diff --git a/Alc/mixer_sse3.c b/Alc/mixer_sse3.c
index acb0338a..b673315b 100644
--- a/Alc/mixer_sse3.c
+++ b/Alc/mixer_sse3.c
@@ -78,8 +78,11 @@ const ALfloat *Resample_fir4_32_SSE3(const ALfloat *src, ALuint frac, ALuint inc
         _mm_store_ps(frac_.f, _mm_castsi128_ps(frac4));
     }
 
-    pos = pos_.i[3];
-    frac = frac_.i[3];
+    /* NOTE: These four elements represent the position *after* the last four
+     * samples, so the lowest element is the next position to resample.
+     */
+    pos = pos_.i[0];
+    frac = frac_.i[0];
 
     for(;i < numsamples;i++)
     {
@@ -143,8 +146,8 @@ const ALfloat *Resample_fir8_32_SSE3(const ALfloat *src, ALuint frac, ALuint inc
         _mm_store_ps(frac_.f, _mm_castsi128_ps(frac4));
     }
 
-    pos = pos_.i[3];
-    frac = frac_.i[3];
+    pos = pos_.i[0];
+    frac = frac_.i[0];
 
     for(;i < numsamples;i++)
     {
diff --git a/Alc/mixer_sse41.c b/Alc/mixer_sse41.c
index 4b347d32..59b99b2b 100644
--- a/Alc/mixer_sse41.c
+++ b/Alc/mixer_sse41.c
@@ -67,7 +67,10 @@ const ALfloat *Resample_lerp32_SSE41(const ALfloat *src, ALuint frac, ALuint inc
         pos_.i[3] = _mm_extract_epi32(pos4, 3);
     }
 
-    pos = pos_.i[3];
+    /* NOTE: These four elements represent the position *after* the last four
+     * samples, so the lowest element is the next position to resample.
+     */
+    pos = pos_.i[0];
     frac = _mm_cvtsi128_si32(frac4);
 
     for(;i < numsamples;i++)
@@ -134,8 +137,8 @@ const ALfloat *Resample_fir4_32_SSE41(const ALfloat *src, ALuint frac, ALuint in
         frac_.i[3] = _mm_extract_epi32(frac4, 3);
     }
 
-    pos = pos_.i[3];
-    frac = frac_.i[3];
+    pos = pos_.i[0];
+    frac = frac_.i[0];
 
     for(;i < numsamples;i++)
     {
@@ -205,8 +208,8 @@ const ALfloat *Resample_fir8_32_SSE41(const ALfloat *src, ALuint frac, ALuint in
         frac_.i[3] = _mm_extract_epi32(frac4, 3);
     }
 
-    pos = pos_.i[3];
-    frac = frac_.i[3];
+    pos = pos_.i[0];
+    frac = frac_.i[0];
 
     for(;i < numsamples;i++)
     {
-- 
cgit v1.2.3