Subject: [PATCH] Fix SSE optimized inner product function.

The original implementation assumes the filter length is multiples of 8,
but it is only multiples of 4.
---
 libspeex/resample_sse.h | 46 ++++++++++++++++++++++++++++++++++------------
 1 file changed, 34 insertions(+), 12 deletions(-)

diff --git a/libspeex/resample_sse.h b/libspeex/resample_sse.h
index 4bd35a2..5d8a3f3 100644
--- a/libspeex/resample_sse.h
+++ b/libspeex/resample_sse.h
@@ -39,18 +39,40 @@
 #define OVERRIDE_INNER_PRODUCT_SINGLE
 static inline float inner_product_single(const float *a, const float *b, unsigned int len)
 {
-   int i;
-   float ret;
-   __m128 sum = _mm_setzero_ps();
-   for (i=0;i<len;i+=8)
-   {
-      sum = _mm_add_ps(sum, _mm_mul_ps(_mm_loadu_ps(a+i), _mm_loadu_ps(b+i)));
-      sum = _mm_add_ps(sum, _mm_mul_ps(_mm_loadu_ps(a+i+4), _mm_loadu_ps(b+i+4)));
-   }
-   sum = _mm_add_ps(sum, _mm_movehl_ps(sum, sum));
-   sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55));
-   _mm_store_ss(&ret, sum);
-   return ret;
+  float sum = 0;
+  unsigned int chunk = len / 4;
+  if (chunk) {
+    __asm__ __volatile__ (
+        "xorps %%xmm0, %%xmm0                       \n"
+        "1:                                         \n"
+        "lddqu (%[a]), %%xmm1                       \n"
+        "lddqu (%[b]), %%xmm2                       \n"
+        "mulps %%xmm1, %%xmm2                       \n"
+        "addps %%xmm2, %%xmm0                       \n"
+        "add $16, %[a]                              \n"
+        "add $16, %[b]                              \n"
+        "sub $1, %[chunk]                           \n"
+        "jnz 1b                                     \n"
+        "movhlps %%xmm0, %%xmm1                     \n"
+        "addps %%xmm1, %%xmm0                       \n"
+        "movaps %%xmm0, %%xmm1                      \n"
+        "shufps $1, %%xmm0, %%xmm0                  \n"
+        "addss %%xmm1, %%xmm0                       \n"
+        "movss %%xmm0, %[sum]                       \n"
+       : /* output */
+          "=r"(chunk),
+          "=r"(a),
+          "=r"(b),
+          [sum]"=m"(sum)
+        : /* input */
+          [chunk]"0"(chunk),
+          [a]"1"(a),
+          [b]"2"(b)
+        : /* clobber */
+          "xmm0", "xmm1", "xmm2", "memory", "cc"
+        );
+  }
+  return sum;
 }
 
 #define OVERRIDE_INTERPOLATE_PRODUCT_SINGLE
-- 
1.7.12.4