Text archives Help

---

Author: wald
Date: Tue Oct 16 15:44:37 2007
New Revision: 1778

Modified:
   trunk/Core/Math/SSEDefs.h
   trunk/Model/Primitives/Sphere.cc
Log:
renamed "cast_f2i" and "cast_i2f" to (DynRT-style) cast4_f2i/i2f -- all other 
sse-macros use an explicit "4" in the name, so this is more consistent.

also added oneOver4 (though I kept "oneOver" intact, in case somebody needs 
it). 

Suggest to either add a "4" to _all_ sse-related fcts (even those that are 
not macros, ie, where it is not absolutely necessary), or to replace all 
"4-macros" with "non-4" inline fct counterparts....




Modified: trunk/Core/Math/SSEDefs.h
==============================================================================
--- trunk/Core/Math/SSEDefs.h   (original)
+++ trunk/Core/Math/SSEDefs.h   Tue Oct 16 15:44:37 2007
@@ -54,8 +54,8 @@
 #define store44i _mm_store_si128
 // The cast_x2y are more like reinterpret casts.  For convertions of
 // types use the convert4 functions below.
-#define cast_i2f _mm_castsi128_ps
-#define cast_f2i _mm_castps_si128
+#define cast4_i2f _mm_castsi128_ps
+#define cast4_f2i _mm_castps_si128
 #define convert4_f2i _mm_cvttps_epi32
 #define convert4_i2f _mm_cvtepi32_ps
 #define sqrt4 _mm_sqrt_ps
@@ -197,6 +197,11 @@
   }
 
   inline sse_t oneOver(const sse_t v)
+  {
+    const sse_t rcp = _mm_rcp_ps(v);
+    return _mm_sub_ps(_mm_add_ps(rcp,rcp),_mm_mul_ps(_mm_mul_ps(rcp,rcp),v));
+  }
+  inline sse_t oneOver4(const sse_t v)
   {
     const sse_t rcp = _mm_rcp_ps(v);
     return _mm_sub_ps(_mm_add_ps(rcp,rcp),_mm_mul_ps(_mm_mul_ps(rcp,rcp),v));

Modified: trunk/Model/Primitives/Sphere.cc
==============================================================================
--- trunk/Model/Primitives/Sphere.cc    (original)
+++ trunk/Model/Primitives/Sphere.cc    Tue Oct 16 15:44:37 2007
@@ -132,7 +132,7 @@
 
           sse_t r = sqrt4(disc);
           // -(r+B)   The xor negates the value
-          sse_t t0 = xor4(add4(r, B), cast_i2f(set4i(0x80000000)));
+          sse_t t0 = xor4(add4(r, B), cast4_i2f(set4i(0x80000000)));
           sse_t hit0 = and4(hit, cmp4_gt(t0, set4(T_EPSILON)));
           rays.hitWithoutTminCheck(i, hit0, t0, getMaterial(), this, 
getTexCoordMapper());
 
@@ -262,7 +262,7 @@
 
           sse_t r = _mm_sqrt_ps(disc);
           // -(r+B)   The xor negates the value
-          sse_t t0 = xor4(add4(r, B), cast_i2f(set4i(0x80000000)));
+          sse_t t0 = xor4(add4(r, B), cast4_i2f(set4i(0x80000000)));
           sse_t hit0 = and4(hit, cmp4_gt(t0, set4(T_EPSILON)));
           rays.hitWithoutTminCheck(i, hit0, t0, getMaterial(), this, 
getTexCoordMapper());
 

---

• [Manta] r1778 - in trunk: Core/Math Model/Primitives, wald, 10/16/2007

---

Archive powered by MHonArc 2.6.16.