Text archives Help

---

Author: bigler
Date: Tue Oct 31 21:27:35 2006
New Revision: 1231

Modified:
   trunk/Model/Materials/Phong.cc
Log:
Tried to make use of the or(and, andnot) masked move functionality.  
Functional, but not complete.

Modified: trunk/Model/Materials/Phong.cc
==============================================================================
--- trunk/Model/Materials/Phong.cc      (original)
+++ trunk/Model/Materials/Phong.cc      Tue Oct 31 21:27:35 2006
@@ -49,6 +49,18 @@
 
 using namespace Manta;
 
+#define USE_MASKEDSTORE 1
+static inline void maskedStore(__m128i mask, __m128i* oldD, __m128i newD) {
+  _mm_store_si128(oldD,
+                  _mm_or_si128(_mm_and_si128(mask, newD),
+                               _mm_andnot_si128(mask, 
_mm_load_si128(oldD))));
+}
+static inline void maskedStore(__m128 mask, float* oldD, __m128 newD) {
+  _mm_store_ps(oldD,
+               _mm_or_ps(_mm_and_ps(mask, newD),
+                         _mm_andnot_ps(mask, _mm_load_ps(oldD))));
+}
+
 Phong::Phong(const Color& diffuse, const Color& specular,
              int specpow, ColorComponent refl)
   : specpow(specpow)
@@ -211,9 +223,33 @@
           _mm_store_ps(&ambientAndDiffuseLight[1][i], 
_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[1][i]), _mm_mul_ps(lightg, 
cos_theta)));
           _mm_store_ps(&ambientAndDiffuseLight[2][i], 
_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[2][i]), _mm_mul_ps(lightb, 
cos_theta)));
         } else {
+#if USE_MASKEDSTORE
+#if 0
+          maskedStore(_mm_castps_si128(mask),
+                      (__m128i*)&ambientAndDiffuseLight[0][i],
+                      
_mm_castps_si128(_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[0][i]), 
_mm_mul_ps(lightr, cos_theta))));
+          maskedStore(_mm_castps_si128(mask),
+                      (__m128i*)&ambientAndDiffuseLight[1][i],
+                      
_mm_castps_si128(_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[1][i]), 
_mm_mul_ps(lightg, cos_theta))));
+          maskedStore(_mm_castps_si128(mask),
+                      (__m128i*)&ambientAndDiffuseLight[2][i],
+                      
_mm_castps_si128(_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[2][i]), 
_mm_mul_ps(lightb, cos_theta))));
+#else
+          maskedStore(mask,(float*)&ambientAndDiffuseLight[0][i],
+                      _mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[0][i]),
+                                 _mm_mul_ps(lightr, cos_theta)));
+          maskedStore(mask,(float*)&ambientAndDiffuseLight[1][i],
+                      _mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[1][i]),
+                                 _mm_mul_ps(lightg, cos_theta)));
+          maskedStore(mask,(float*)&ambientAndDiffuseLight[2][i],
+                      _mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[2][i]),
+                                 _mm_mul_ps(lightb, cos_theta)));
+#endif
+#else // USE_MASKEDSTORE
           _mm_maskmoveu_si128((__m128i) 
_mm_castps_si128(_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[0][i]), 
_mm_mul_ps(lightr, cos_theta))), (__m128i) _mm_castps_si128(mask), 
(char*)&ambientAndDiffuseLight[0][i]);
           _mm_maskmoveu_si128((__m128i) 
_mm_castps_si128(_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[1][i]), 
_mm_mul_ps(lightg, cos_theta))), (__m128i) _mm_castps_si128(mask), 
(char*)&ambientAndDiffuseLight[1][i]);
           _mm_maskmoveu_si128((__m128i) 
_mm_castps_si128(_mm_add_ps(_mm_load_ps(&ambientAndDiffuseLight[2][i]), 
_mm_mul_ps(lightb, cos_theta))), (__m128i) _mm_castps_si128(mask), 
(char*)&ambientAndDiffuseLight[2][i]);
+#endif
         }
         
         __m128 Hx = _mm_sub_ps(sdx, _mm_load_ps(&data->direction[0][i]));

---

• [MANTA] r1231 - trunk/Model/Materials, bigler, 10/31/2006

---

Archive powered by MHonArc 2.6.16.