Optimized the SSE2 path for hard-thresholding filter

Thanks for the idea from MonoS: MonoS@421fa78
HomeOfVapourSynthEvolution · Jan 23, 2016 · 1263139 · 1263139
1 parent 1fc6137
commit 1263139
Show file tree

Hide file tree

Showing 3 changed files with 15 additions and 12 deletions.
diff --git a/include/Helper.h b/include/Helper.h
@@ -104,6 +104,15 @@ class ClockCounter
 #endif
 
 
+#if defined(__SSE2__)
+inline __m128 _mm_abs_ps(const __m128 &x)
+{
+    static const __m128 mask = _mm_castsi128_ps(_mm_set1_epi32(~0x80000000));
+    return _mm_and_ps(x, mask);
+}
+#endif
+
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
 // Exception handle
 

diff --git a/source/BM3D_Basic.cpp b/source/BM3D_Basic.cpp
@@ -88,18 +88,15 @@ void BM3D_Basic_Process::CollaborativeFilter(int plane,
     const ptrdiff_t simd_residue = srcGroup.size() % simd_step;
     const ptrdiff_t simd_width = srcGroup.size() - simd_residue;
 
-    static const __m128 zero_ps = _mm_setzero_ps();
     __m128i cmp_sum = _mm_setzero_si128();
 
     for (const auto upper1 = srcp + simd_width; srcp < upper1; srcp += simd_step, thrp += simd_step)
     {
         const __m128 s1 = _mm_load_ps(srcp);
-        const __m128 t1p = _mm_load_ps(thrp);
-        const __m128 t1n = _mm_sub_ps(zero_ps, t1p);
+        const __m128 t1 = _mm_load_ps(thrp);
 
-        const __m128 cmp1 = _mm_cmpgt_ps(s1, t1p);
-        const __m128 cmp2 = _mm_cmplt_ps(s1, t1n);
-        const __m128 cmp = _mm_or_ps(cmp1, cmp2);
+        const __m128 s1abs = _mm_abs_ps(s1);
+        const __m128 cmp = _mm_cmpgt_ps(s1abs, t1);
 
         const __m128 d1 = _mm_and_ps(cmp, s1);
         _mm_store_ps(srcp, d1);

diff --git a/source/VBM3D_Basic.cpp b/source/VBM3D_Basic.cpp
@@ -88,18 +88,15 @@ void VBM3D_Basic_Process::CollaborativeFilter(int plane,
     const ptrdiff_t simd_residue = srcGroup.size() % simd_step;
     const ptrdiff_t simd_width = srcGroup.size() - simd_residue;
 
-    static const __m128 zero_ps = _mm_setzero_ps();
     __m128i cmp_sum = _mm_setzero_si128();
 
     for (const auto upper1 = srcp + simd_width; srcp < upper1; srcp += simd_step, thrp += simd_step)
     {
         const __m128 s1 = _mm_load_ps(srcp);
-        const __m128 t1p = _mm_load_ps(thrp);
-        const __m128 t1n = _mm_sub_ps(zero_ps, t1p);
+        const __m128 t1 = _mm_load_ps(thrp);
 
-        const __m128 cmp1 = _mm_cmpgt_ps(s1, t1p);
-        const __m128 cmp2 = _mm_cmplt_ps(s1, t1n);
-        const __m128 cmp = _mm_or_ps(cmp1, cmp2);
+        const __m128 s1abs = _mm_abs_ps(s1);
+        const __m128 cmp = _mm_cmpgt_ps(s1abs, t1);
 
         const __m128 d1 = _mm_and_ps(cmp, s1);
         _mm_store_ps(srcp, d1);