Cycles: merging features from tomato branch.

=== BVH build time optimizations ===

* BVH building was multithreaded. Not all building is multithreaded, packing
  and the initial bounding/splitting is still single threaded, but recursive
  splitting is, which was the main bottleneck.

* Object splitting now uses binning rather than sorting of all elements, using
  code from the Embree raytracer from Intel.
  http://software.intel.com/en-us/articles/embree-photo-realistic-ray-tracing-kernels/

* Other small changes to avoid allocations, pack memory more tightly, avoid
  some unnecessary operations, ...

These optimizations do not work yet when Spatial Splits are enabled, for that
more work is needed. There's also other optimizations still needed, in
particular for the case of many low poly objects, the packing step and node
memory allocation.

BVH raytracing time should remain about the same, but BVH build time should be
significantly reduced, test here show speedup of about 5x to 10x on a dual core
and 5x to 25x on an 8-core machine, depending on the scene.

=== Threads ===

Centralized task scheduler for multithreading, which is basically the
CPU device threading code wrapped into something reusable.

Basic idea is that there is a single TaskScheduler that keeps a pool of threads,
one for each core. Other places in the code can then create a TaskPool that they
can drop Tasks in to be executed by the scheduler, and wait for them to complete
or cancel them early.

=== Normal ====

Added a Normal output to the texture coordinate node. This currently
gives the object space normal, which is the same under object animation.

In the future this might become a "generated" normal so it's also stable for
deforming objects, but for now it's already useful for non-deforming objects.

=== Render Layers ===

Per render layer Samples control, leaving it to 0 will use the common scene
setting.

Environment pass will now render environment even if film is set to transparent.

Exclude Layers" added. Scene layers (all object that influence the render,
directly or indirectly) are shared between all render layers. However sometimes
it's useful to leave out some object influence for a particular render layer.
That's what this option allows you to do.

=== Filter Glossy ===

When using a value higher than 0.0, this will blur glossy reflections after
blurry bounces, to reduce noise at the cost of accuracy. 1.0 is a good
starting value to tweak.

Some light paths have a low probability of being found while contributing much
light to the pixel. As a result these light paths will be found in some pixels
and not in others, causing fireflies. An example of such a difficult path might
be a small light that is causing a small specular highlight on a sharp glossy
material, which we are seeing through a rough glossy material. With path tracing
it is difficult to find the specular highlight, but if we increase the roughness
on the material the highlight gets bigger and softer, and so easier to find.

Often this blurring will be hardly noticeable, because we are seeing it through
a blurry material anyway, but there are also cases where this will lead to a
loss of detail in lighting.

8 files changed, 992 insertions, 344 deletions

diff --git a/intern/cycles/util/CMakeLists.txt b/intern/cycles/util/CMakeLists.txt
index 9182ee4cbe1..87bd84b4e0f 100644
--- a/intern/cycles/util/CMakeLists.txt
+++ b/intern/cycles/util/CMakeLists.txt
@@ -15,6 +15,7 @@ set(SRC
 	util_path.cpp
 	util_string.cpp
 	util_system.cpp
+	util_task.cpp
 	util_time.cpp
 	util_transform.cpp
 )
@@ -50,6 +51,7 @@ set(SRC_HEADERS
 	util_set.h
 	util_string.h
 	util_system.h
+	util_task.h
 	util_thread.h
 	util_time.h
 	util_transform.h
diff --git a/intern/cycles/util/util_boundbox.h b/intern/cycles/util/util_boundbox.h
index bb1df0b220f..9511b48e103 100644
--- a/intern/cycles/util/util_boundbox.h
+++ b/intern/cycles/util/util_boundbox.h
@@ -23,6 +23,7 @@
 #include <float.h>
 
 #include "util_math.h"
+#include "util_string.h"
 #include "util_transform.h"
 #include "util_types.h"
 
@@ -35,45 +36,81 @@ class BoundBox
 public:
 	float3 min, max;
 
-	BoundBox(void)
+	__forceinline BoundBox()
 	{
-		min = make_float3(FLT_MAX, FLT_MAX, FLT_MAX);
-		max = make_float3(-FLT_MAX, -FLT_MAX, -FLT_MAX);
 	}
 
-	BoundBox(const float3& min_, const float3& max_)
+	__forceinline BoundBox(const float3& pt)
+	: min(pt), max(pt)
+	{
+	}
+
+	__forceinline BoundBox(const float3& min_, const float3& max_)
 	: min(min_), max(max_)
 	{
 	}
 
-	void grow(const float3& pt)  
+	static struct empty_t {} empty;
+
+	__forceinline BoundBox(empty_t)
+	: min(make_float3(FLT_MAX, FLT_MAX, FLT_MAX)), max(make_float3(-FLT_MAX, -FLT_MAX, -FLT_MAX))
+	{
+	}
+
+	__forceinline void grow(const float3& pt)  
 	{
 		min = ccl::min(min, pt);
 		max = ccl::max(max, pt);
 	}
 
-	void grow(const BoundBox& bbox)
+	__forceinline void grow(const BoundBox& bbox)
 	{
 		grow(bbox.min);
 		grow(bbox.max);
 	}
 
-	void intersect(const BoundBox& bbox) 
+	__forceinline void intersect(const BoundBox& bbox) 
 	{
 		min = ccl::max(min, bbox.min);
 		max = ccl::min(max, bbox.max);
 	}
 
-	float area(void) const
+	/* todo: avoid using this */
+	__forceinline float safe_area() const
 	{
-		if(!valid())
+		if(!((min.x <= max.x) && (min.y <= max.y) && (min.z <= max.z)))
 			return 0.0f;
 
+		return area();
+	}
+
+	__forceinline float area() const
+	{
+		return half_area()*2.0f;
+	}
+
+	__forceinline float half_area() const
+	{
 		float3 d = max - min;
-		return dot(d, d)*2.0f;
+		return (d.x*d.z + d.y*d.z + d.x*d.y);
+	}
+
+	__forceinline float3 center() const
+	{
+		return 0.5f*(min + max);
 	}
 
-	bool valid(void) const
+	__forceinline float3 center2() const
+	{
+		return min + max;
+	}
+
+	__forceinline float3 size() const
+	{
+		return max - min;
+	}
+	
+	__forceinline bool valid() const
 	{
 		return (min.x <= max.x) && (min.y <= max.y) && (min.z <= max.z) &&
 		       (isfinite(min.x) && isfinite(min.y) && isfinite(min.z)) &&
@@ -82,7 +119,7 @@ public:
 
 	BoundBox transformed(const Transform *tfm)
 	{
-		BoundBox result;
+		BoundBox result = BoundBox::empty;
 
 		for(int i = 0; i < 8; i++) {
 			float3 p;
@@ -98,6 +135,31 @@ public:
 	}
 };
 
+__forceinline BoundBox merge(const BoundBox& bbox, const float3& pt)
+{
+	return BoundBox(min(bbox.min, pt), max(bbox.max, pt));
+}
+
+__forceinline BoundBox merge(const BoundBox& a, const BoundBox& b)
+{
+	return BoundBox(min(a.min, b.min), max(a.max, b.max));
+}
+
+__forceinline BoundBox merge(const BoundBox& a, const BoundBox& b, const BoundBox& c, const BoundBox& d)
+{
+	return merge(merge(a, b), merge(c, d));
+}
+
+__forceinline BoundBox intersect(const BoundBox& a, const BoundBox& b)
+{
+	return BoundBox(max(a.min, b.min), min(a.max, b.max));
+}
+
+__forceinline BoundBox intersect(const BoundBox& a, const BoundBox& b, const BoundBox& c)
+{
+	return intersect(a, intersect(b, c));
+}
+
 CCL_NAMESPACE_END
 
 #endif /* __UTIL_BOUNDBOX_H__ */
diff --git a/intern/cycles/util/util_math.h b/intern/cycles/util/util_math.h
index 019dede07fa..25d81481d12 100644
--- a/intern/cycles/util/util_math.h
+++ b/intern/cycles/util/util_math.h
@@ -182,93 +182,74 @@ __device_inline float average(const float2 a)
 
 __device_inline float2 operator-(const float2 a)
 {
-	float2 r = {-a.x, -a.y};
-	return r;
+	return make_float2(-a.x, -a.y);
 }
 
 __device_inline float2 operator*(const float2 a, const float2 b)
 {
-	float2 r = {a.x*b.x, a.y*b.y};
-	return r;
+	return make_float2(a.x*b.x, a.y*b.y);
 }
 
 __device_inline float2 operator*(const float2 a, float f)
 {
-	float2 r = {a.x*f, a.y*f};
-	return r;
+	return make_float2(a.x*f, a.y*f);
 }
 
 __device_inline float2 operator*(float f, const float2 a)
 {
-	float2 r = {a.x*f, a.y*f};
-	return r;
+	return make_float2(a.x*f, a.y*f);
 }
 
 __device_inline float2 operator/(float f, const float2 a)
 {
-	float2 r = {f/a.x, f/a.y};
-	return r;
+	return make_float2(f/a.x, f/a.y);
 }
 
 __device_inline float2 operator/(const float2 a, float f)
 {
 	float invf = 1.0f/f;
-	float2 r = {a.x*invf, a.y*invf};
-	return r;
+	return make_float2(a.x*invf, a.y*invf);
 }
 
 __device_inline float2 operator/(const float2 a, const float2 b)
 {
-	float2 r = {a.x/b.x, a.y/b.y};
-	return r;
+	return make_float2(a.x/b.x, a.y/b.y);
 }
 
 __device_inline float2 operator+(const float2 a, const float2 b)
 {
-	float2 r = {a.x+b.x, a.y+b.y};
-	return r;
+	return make_float2(a.x+b.x, a.y+b.y);
 }
 
 __device_inline float2 operator-(const float2 a, const float2 b)
 {
-	float2 r = {a.x-b.x, a.y-b.y};
-	return r;
+	return make_float2(a.x-b.x, a.y-b.y);
 }
 
 __device_inline float2 operator+=(float2& a, const float2 b)
 {
-	a.x += b.x;
-	a.y += b.y;
-	return a;
+	return a = a + b;
 }
 
 __device_inline float2 operator*=(float2& a, const float2 b)
 {
-	a.x *= b.x;
-	a.y *= b.y;
-	return a;
+	return a = a * b;
 }
 
 __device_inline float2 operator*=(float2& a, float f)
 {
-	a.x *= f;
-	a.y *= f;
-	return a;
+	return a = a * f;
 }
 
 __device_inline float2 operator/=(float2& a, const float2 b)
 {
-	a.x /= b.x;
-	a.y /= b.y;
-	return a;
+	return a = a / b;
 }
 
 __device_inline float2 operator/=(float2& a, float f)
 {
 	float invf = 1.0f/f;
-	a.x *= invf;
-	a.y *= invf;
-	return a;
+	return a = a * invf;
 }
 
 
@@ -314,14 +295,12 @@ __device_inline bool operator!=(const float2 a, const float2 b)
 
 __device_inline float2 min(float2 a, float2 b)
 {
-	float2 r = {min(a.x, b.x), min(a.y, b.y)};
-	return r;
+	return make_float2(min(a.x, b.x), min(a.y, b.y));
 }
 
 __device_inline float2 max(float2 a, float2 b)
 {
-	float2 r = {max(a.x, b.x), max(a.y, b.y)};
-	return r;
+	return make_float2(max(a.x, b.x), max(a.y, b.y));
 }
 
 __device_inline float2 clamp(float2 a, float2 mn, float2 mx)
@@ -361,112 +340,78 @@ __device_inline float2 interp(float2 a, float2 b, float t)
 
 /* Float3 Vector */
 
-__device_inline bool is_zero(const float3 a)
-{
-	return (a.x == 0.0f && a.y == 0.0f && a.z == 0.0f);
-}
-
-__device_inline float average(const float3 a)
-{
-	return (a.x + a.y + a.z)*(1.0f/3.0f);
-}
-
 #ifndef __KERNEL_OPENCL__
 
 __device_inline float3 operator-(const float3 a)
 {
-	float3 r = make_float3(-a.x, -a.y, -a.z);
-	return r;
+	return make_float3(-a.x, -a.y, -a.z);
 }
 
 __device_inline float3 operator*(const float3 a, const float3 b)
 {
-	float3 r = make_float3(a.x*b.x, a.y*b.y, a.z*b.z);
-	return r;
+	return make_float3(a.x*b.x, a.y*b.y, a.z*b.z);
 }
 
 __device_inline float3 operator*(const float3 a, float f)
 {
-	float3 r = make_float3(a.x*f, a.y*f, a.z*f);
-	return r;
+	return make_float3(a.x*f, a.y*f, a.z*f);
 }
 
 __device_inline float3 operator*(float f, const float3 a)
 {
-	float3 r = make_float3(a.x*f, a.y*f, a.z*f);
-	return r;
+	return make_float3(a.x*f, a.y*f, a.z*f);
 }
 
 __device_inline float3 operator/(float f, const float3 a)
 {
-	float3 r = make_float3(f/a.x, f/a.y, f/a.z);
-	return r;
+	return make_float3(f/a.x, f/a.y, f/a.z);
 }
 
 __device_inline float3 operator/(const float3 a, float f)
 {
 	float invf = 1.0f/f;
-	float3 r = make_float3(a.x*invf, a.y*invf, a.z*invf);
-	return r;
+	return make_float3(a.x*invf, a.y*invf, a.z*invf);
 }
 
 __device_inline float3 operator/(const float3 a, const float3 b)
 {
-	float3 r = make_float3(a.x/b.x, a.y/b.y, a.z/b.z);
-	return r;
+	return make_float3(a.x/b.x, a.y/b.y, a.z/b.z);
 }
 
 __device_inline float3 operator+(const float3 a, const float3 b)
 {
-	float3 r = make_float3(a.x+b.x, a.y+b.y, a.z+b.z);
-	return r;
+	return make_float3(a.x+b.x, a.y+b.y, a.z+b.z);
 }
 
 __device_inline float3 operator-(const float3 a, const float3 b)
 {
-	float3 r = make_float3(a.x-b.x, a.y-b.y, a.z-b.z);
-	return r;
+	return make_float3(a.x-b.x, a.y-b.y, a.z-b.z);
 }
 
 __device_inline float3 operator+=(float3& a, const float3 b)
 {
-	a.x += b.x;
-	a.y += b.y;
-	a.z += b.z;
-	return a;
+	return a = a + b;
 }
 
 __device_inline float3 operator*=(float3& a, const float3 b)
 {
-	a.x *= b.x;
-	a.y *= b.y;
-	a.z *= b.z;
-	return a;
+	return a = a * b;
 }
 
 __device_inline float3 operator*=(float3& a, float f)
 {
-	a.x *= f;
-	a.y *= f;
-	a.z *= f;
-	return a;
+	return a = a * f;
 }
 
 __device_inline float3 operator/=(float3& a, const float3 b)
 {
-	a.x /= b.x;
-	a.y /= b.y;
-	a.z /= b.z;
-	return a;
+	return a = a / b;
 }
 
 __device_inline float3 operator/=(float3& a, float f)
 {
 	float invf = 1.0f/f;
-	a.x *= invf;
-	a.y *= invf;
-	a.z *= invf;
-	return a;
+	return a = a * invf;
 }
 
 __device_inline float dot(const float3 a, const float3 b)
@@ -506,7 +451,11 @@ __device_inline float3 normalize_len(const float3 a, float *t)
 
 __device_inline bool operator==(const float3 a, const float3 b)
 {
+#ifdef __KERNEL_SSE__
+	return (_mm_movemask_ps(_mm_cmpeq_ps(a.m128, b.m128)) & 7) == 7;
+#else
 	return (a.x == b.x && a.y == b.y && a.z == b.z);
+#endif
 }
 
 __device_inline bool operator!=(const float3 a, const float3 b)
@@ -516,14 +465,20 @@ __device_inline bool operator!=(const float3 a, const float3 b)
 
 __device_inline float3 min(float3 a, float3 b)
 {
-	float3 r = make_float3(min(a.x, b.x), min(a.y, b.y), min(a.z, b.z));
-	return r;
+#ifdef __KERNEL_SSE__
+	return _mm_min_ps(a.m128, b.m128);
+#else
+	return make_float3(min(a.x, b.x), min(a.y, b.y), min(a.z, b.z));
+#endif
 }
 
 __device_inline float3 max(float3 a, float3 b)
 {
-	float3 r = make_float3(max(a.x, b.x), max(a.y, b.y), max(a.z, b.z));
-	return r;
+#ifdef __KERNEL_SSE__
+	return _mm_max_ps(a.m128, b.m128);
+#else
+	return make_float3(max(a.x, b.x), max(a.y, b.y), max(a.z, b.z));
+#endif
 }
 
 __device_inline float3 clamp(float3 a, float3 mn, float3 mx)
@@ -533,7 +488,12 @@ __device_inline float3 clamp(float3 a, float3 mn, float3 mx)
 
 __device_inline float3 fabs(float3 a)
 {
+#ifdef __KERNEL_SSE__
+	__m128 mask = _mm_castsi128_ps(_mm_set1_epi32(0x7fffffff));
+	return _mm_and_ps(a.m128, mask);
+#else
 	return make_float3(fabsf(a.x), fabsf(a.y), fabsf(a.z));
+#endif
 }
 
 #endif
@@ -555,6 +515,25 @@ __device_inline void print_float3(const char *label, const float3& a)
 	printf("%s: %.8f %.8f %.8f\n", label, a.x, a.y, a.z);
 }
 
+__device_inline float reduce_add(const float3& a)
+{
+#ifdef __KERNEL_SSE__
+	return (a.x + a.y + a.z);
+#else
+	return (a.x + a.y + a.z);
+#endif
+}
+
+__device_inline float3 rcp(const float3& a)
+{
+#ifdef __KERNEL_SSE__
+	float4 r = _mm_rcp_ps(a.m128);
+	return _mm_sub_ps(_mm_add_ps(r, r), _mm_mul_ps(_mm_mul_ps(r, r), a));
+#else
+	return make_float3(1.0f/a.x, 1.0f/a.y, 1.0f/a.z);
+#endif
+}
+
 #endif
 
 __device_inline float3 interp(float3 a, float3 b, float t)
@@ -562,122 +541,258 @@ __device_inline float3 interp(float3 a, float3 b, float t)
 	return a + t*(b - a);
 }
 
+__device_inline bool is_zero(const float3 a)
+{
+#ifdef __KERNEL_SSE__
+	return a == make_float3(0.0f);
+#else
+	return (a.x == 0.0f && a.y == 0.0f && a.z == 0.0f);
+#endif
+}
+
+__device_inline float average(const float3 a)
+{
+	return reduce_add(a)*(1.0f/3.0f);
+}
+
 /* Float4 Vector */
 
-#ifndef __KERNEL_OPENCL__
+#ifdef __KERNEL_SSE__
 
-__device_inline bool is_zero(const float4& a)
+template<size_t index_0, size_t index_1, size_t index_2, size_t index_3> __forceinline const float4 shuffle(const float4& b)
 {
-	return (a.x == 0.0f && a.y == 0.0f && a.z == 0.0f && a.w == 0.0f);
+	return _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(b), _MM_SHUFFLE(index_3, index_2, index_1, index_0)));
 }
 
-__device_inline float average(const float4& a)
+template<> __forceinline const float4 shuffle<0, 0, 2, 2>(const float4& b)
 {
-	return (a.x + a.y + a.z + a.w)*(1.0f/4.0f);
+	return _mm_moveldup_ps(b);
 }
 
+template<> __forceinline const float4 shuffle<1, 1, 3, 3>(const float4& b)
+{
+	return _mm_movehdup_ps(b);
+}
+
+template<> __forceinline const float4 shuffle<0, 1, 0, 1>(const float4& b)
+{
+	return _mm_castpd_ps(_mm_movedup_pd(_mm_castps_pd(b)));
+}
+
+#endif
+
+#ifndef __KERNEL_OPENCL__
+
 __device_inline float4 operator-(const float4& a)
 {
-	float4 r = {-a.x, -a.y, -a.z, -a.w};
-	return r;
+#ifdef __KERNEL_SSE__
+	__m128 mask = _mm_castsi128_ps(_mm_set1_epi32(0x80000000));
+	return _mm_xor_ps(a.m128, mask);
+#else
+	return make_float4(-a.x, -a.y, -a.z, -a.w);
+#endif
 }
 
 __device_inline float4 operator*(const float4& a, const float4& b)
 {
-	float4 r = {a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w};
-	return r;
+#ifdef __KERNEL_SSE__
+	return _mm_mul_ps(a.m128, b.m128);
+#else
+	return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w);
+#endif
 }
 
 __device_inline float4 operator*(const float4& a, float f)
 {
-	float4 r = {a.x*f, a.y*f, a.z*f, a.w*f};
-	return r;
+#ifdef __KERNEL_SSE__
+	return a * make_float4(f);
+#else
+	return make_float4(a.x*f, a.y*f, a.z*f, a.w*f);
+#endif
 }
 
 __device_inline float4 operator*(float f, const float4& a)
 {
-	float4 r = {a.x*f, a.y*f, a.z*f, a.w*f};
-	return r;
+	return a * f;
+}
+
+__device_inline float4 rcp(const float4& a)
+{
+#ifdef __KERNEL_SSE__
+	float4 r = _mm_rcp_ps(a.m128);
+	return _mm_sub_ps(_mm_add_ps(r, r), _mm_mul_ps(_mm_mul_ps(r, r), a));
+#else
+	return make_float4(1.0f/a.x, 1.0f/a.y, 1.0f/a.z, 1.0f/a.w);
+#endif
 }
 
 __device_inline float4 operator/(const float4& a, float f)
 {
-	float invf = 1.0f/f;
-	float4 r = {a.x*invf, a.y*invf, a.z*invf, a.w*invf};
-	return r;
+	return a * (1.0f/f);
 }
 
 __device_inline float4 operator/(const float4& a, const float4& b)
 {
-	float4 r = {a.x/b.x, a.y/b.y, a.z/b.z, a.w/b.w};
-	return r;
+#ifdef __KERNEL_SSE__
+	return a * rcp(b);
+#else
+	return make_float4(a.x/b.x, a.y/b.y, a.z/b.z, a.w/b.w);
+#endif
+
 }
 
 __device_inline float4 operator+(const float4& a, const float4& b)
 {
-	float4 r = {a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w};
-	return r;
+#ifdef __KERNEL_SSE__
+	return _mm_add_ps(a.m128, b.m128);
+#else
+	return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w);
+#endif
 }
 
 __device_inline float4 operator-(const float4& a, const float4& b)
 {
-	float4 r = {a.x-b.x, a.y-b.y, a.z-b.z, a.w-b.w};
-	return r;
+#ifdef __KERNEL_SSE__
+	return _mm_sub_ps(a.m128, b.m128);
+#else
+	return make_float4(a.x-b.x, a.y-b.y, a.z-b.z, a.w-b.w);
+#endif
 }
 
 __device_inline float4 operator+=(float4& a, const float4& b)
 {
-	a.x += b.x;
-	a.y += b.y;
-	a.z += b.z;
-	a.w += b.w;
-	return a;
+	return a = a + b;
 }
 
 __device_inline float4 operator*=(float4& a, const float4& b)
 {
-	a.x *= b.x;
-	a.y *= b.y;
-	a.z *= b.z;
-	a.w *= b.w;
-	return a;
+	return a = a * b;
 }
 
 __device_inline float4 operator/=(float4& a, float f)
 {
-	float invf = 1.0f/f;
-	a.x *= invf;
-	a.y *= invf;
-	a.z *= invf;
-	a.w *= invf;
-	return a;
+	return a = a / f;
 }
 
-__device_inline float dot(const float4& a, const float4& b)
+__device_inline int4 operator<(const float4& a, const float4& b)
 {
-	return a.x*b.x + a.y*b.y + a.z*b.z + a.w*b.w;
+#ifdef __KERNEL_SSE__
+	return _mm_cvtps_epi32(_mm_cmplt_ps(a.m128, b.m128)); /* todo: avoid cvt */
+#else
+	return make_int4(a.x < b.x, a.y < b.y, a.z < b.z, a.w < b.w);
+#endif
+}
+
+__device_inline int4 operator>=(float4 a, float4 b)
+{
+#ifdef __KERNEL_SSE__
+	return _mm_cvtps_epi32(_mm_cmpge_ps(a.m128, b.m128)); /* todo: avoid cvt */
+#else
+	return make_int4(a.x >= b.x, a.y >= b.y, a.z >= b.z, a.w >= b.w);
+#endif
+}
+
+__device_inline int4 operator<=(const float4& a, const float4& b)
+{
+#ifdef __KERNEL_SSE__
+	return _mm_cvtps_epi32(_mm_cmple_ps(a.m128, b.m128)); /* todo: avoid cvt */
+#else
+	return make_int4(a.x <= b.x, a.y <= b.y, a.z <= b.z, a.w <= b.w);
+#endif
+}
+
+__device_inline bool operator==(const float4 a, const float4 b)
+{
+#ifdef __KERNEL_SSE__
+	return (_mm_movemask_ps(_mm_cmpeq_ps(a.m128, b.m128)) & 15) == 15;
+#else
+	return (a.x == b.x && a.y == b.y && a.z == b.z && a.w == b.w);
+#endif
 }
 
 __device_inline float4 cross(const float4& a, const float4& b)
 {
-	float4 r = {a.y*b.z - a.z*b.y, a.z*b.x - a.x*b.z, a.x*b.y - a.y*b.x, 0.0f};
-	return r;
+#ifdef __KERNEL_SSE__
+	return (shuffle<1,2,0,0>(a)*shuffle<2,0,1,0>(b)) - (shuffle<2,0,1,0>(a)*shuffle<1,2,0,0>(b));
+#else
+	return make_float4(a.y*b.z - a.z*b.y, a.z*b.x - a.x*b.z, a.x*b.y - a.y*b.x, 0.0f);
+#endif
 }
 
 __device_inline float4 min(float4 a, float4 b)
 {
+#ifdef __KERNEL_SSE__
+	return _mm_min_ps(a.m128, b.m128);
+#else
 	return make_float4(min(a.x, b.x), min(a.y, b.y), min(a.z, b.z), min(a.w, b.w));
+#endif
 }
 
 __device_inline float4 max(float4 a, float4 b)
 {
+#ifdef __KERNEL_SSE__
+	return _mm_max_ps(a.m128, b.m128);
+#else
 	return make_float4(max(a.x, b.x), max(a.y, b.y), max(a.z, b.z), max(a.w, b.w));
+#endif
 }
 
 #endif
 
 #ifndef __KERNEL_GPU__
 
+__device_inline float4 select(const int4& mask, const float4& a, const float4& b)
+{
+#ifdef __KERNEL_SSE__
+	/* blendv is sse4, and apparently broken on vs2008 */
+	return _mm_or_ps(_mm_and_ps(_mm_cvtepi32_ps(mask), a), _mm_andnot_ps(_mm_cvtepi32_ps(mask), b)); /* todo: avoid cvt */
+#else
+	return make_float4((mask.x)? a.x: b.x, (mask.y)? a.y: b.y, (mask.z)? a.z: b.z, (mask.w)? a.w: b.w);
+#endif
+}
+
+__device_inline float4 reduce_min(const float4& a)
+{
+#ifdef __KERNEL_SSE__
+	float4 h = min(shuffle<1,0,3,2>(a), a);
+	return min(shuffle<2,3,0,1>(h), h);
+#else
+	return make_float4(min(min(a.x, a.y), min(a.z, a.w)));
+#endif
+}
+
+__device_inline float4 reduce_max(const float4& a)
+{
+#ifdef __KERNEL_SSE__
+	float4 h = max(shuffle<1,0,3,2>(a), a);
+	return max(shuffle<2,3,0,1>(h), h);
+#else
+	return make_float4(max(max(a.x, a.y), max(a.z, a.w)));
+#endif
+}
+
+#if 0
+__device_inline float4 reduce_add(const float4& a)
+{
+#ifdef __KERNEL_SSE__
+	float4 h = shuffle<1,0,3,2>(a) + a;
+	return shuffle<2,3,0,1>(h) + h;
+#else
+	return make_float4((a.x + a.y) + (a.z + a.w));
+#endif
+}
+#endif
+
+__device_inline float reduce_add(const float4& a)
+{
+#ifdef __KERNEL_SSE__
+	float4 h = shuffle<1,0,3,2>(a) + a;
+	return _mm_cvtss_f32(shuffle<2,3,0,1>(h) + h); /* todo: efficiency? */
+#else
+	return ((a.x + a.y) + (a.z + a.w));
+#endif
+}
+
 __device_inline void print_float4(const char *label, const float4& a)
 {
 	printf("%s: %.8f %.8f %.8f %.8f\n", label, a.x, a.y, a.z, a.w);
@@ -685,26 +800,67 @@ __device_inline void print_float4(const char *label, const float4& a)
 
 #endif
 
+#ifndef __KERNEL_OPENCL__
+
+__device_inline bool is_zero(const float4& a)
+{
+#ifdef __KERNEL_SSE__
+	return a == make_float4(0.0f);
+#else
+	return (a.x == 0.0f && a.y == 0.0f && a.z == 0.0f && a.w == 0.0f);
+#endif
+}
+
+__device_inline float average(const float4& a)
+{
+	return reduce_add(a) * 0.25f;
+}
+
+__device_inline float dot(const float4& a, const float4& b)
+{
+	return reduce_add(a * b);
+}
+
+#endif
+
 /* Int3 */
 
 #ifndef __KERNEL_OPENCL__
 
+__device_inline int3 min(int3 a, int3 b)
+{
+#ifdef __KERNEL_SSE__
+	return _mm_min_epi32(a.m128, b.m128);
+#else
+	return make_int3(min(a.x, b.x), min(a.y, b.y), min(a.z, b.z));
+#endif
+}
+
 __device_inline int3 max(int3 a, int3 b)
 {
-	int3 r = {max(a.x, b.x), max(a.y, b.y), max(a.z, b.z)};
-	return r;
+#ifdef __KERNEL_SSE__
+	return _mm_max_epi32(a.m128, b.m128);
+#else
+	return make_int3(max(a.x, b.x), max(a.y, b.y), max(a.z, b.z));
+#endif
 }
 
 __device_inline int3 clamp(const int3& a, int mn, int mx)
 {
-	int3 r = {clamp(a.x, mn, mx), clamp(a.y, mn, mx), clamp(a.z, mn, mx)};
-	return r;
+#ifdef __KERNEL_SSE__
+	return min(max(a, make_int3(mn)), make_int3(mx));
+#else
+	return make_int3(clamp(a.x, mn, mx), clamp(a.y, mn, mx), clamp(a.z, mn, mx));
+#endif
 }
 
 __device_inline int3 clamp(const int3& a, int3& mn, int mx)
 {
-	int3 r = {clamp(a.x, mn.x, mx), clamp(a.y, mn.y, mx), clamp(a.z, mn.z, mx)};
-	return r;
+#ifdef __KERNEL_SSE__
+	return min(max(a, mn), make_int3(mx));
+#else
+	return make_int3(clamp(a.x, mn.x, mx), clamp(a.y, mn.y, mx), clamp(a.z, mn.z, mx));
+#endif
 }
 
 #endif
@@ -720,16 +876,63 @@ __device_inline void print_int3(const char *label, const int3& a)
 
 /* Int4 */
 
-#ifndef __KERNEL_OPENCL__
+#ifndef __KERNEL_GPU__
 
-__device_inline int4 operator>=(float4 a, float4 b)
+__device_inline int4 operator+(const int4& a, const int4& b)
 {
-	return make_int4(a.x >= b.x, a.y >= b.y, a.z >= b.z, a.w >= b.w);
+#ifdef __KERNEL_SSE__
+	return _mm_add_epi32(a.m128, b.m128);
+#else
+	return make_int4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w);
+#endif
+}
+
+__device_inline int4 operator+=(int4& a, const int4& b)
+{
+	return a = a + b;
 }
 
+__device_inline int4 operator>>(const int4& a, int i)
+{
+#ifdef __KERNEL_SSE__
+	return _mm_srai_epi32(a.m128, i);
+#else
+	return make_int4(a.x >> i, a.y >> i, a.z >> i, a.w >> i);
 #endif
+}
 
-#ifndef __KERNEL_GPU__
+__device_inline int4 min(int4 a, int4 b)
+{
+#ifdef __KERNEL_SSE__
+	return _mm_min_epi32(a.m128, b.m128);
+#else
+	return make_int4(min(a.x, b.x), min(a.y, b.y), min(a.z, b.z), min(a.w, b.w));
+#endif
+}
+
+__device_inline int4 max(int4 a, int4 b)
+{
+#ifdef __KERNEL_SSE__
+	return _mm_max_epi32(a.m128, b.m128);
+#else
+	return make_int4(max(a.x, b.x), max(a.y, b.y), max(a.z, b.z), max(a.w, b.w));
+#endif
+}
+
+__device_inline int4 clamp(const int4& a, const int4& mn, const int4& mx)
+{
+	return min(max(a, mn), mx);
+}
+
+__device_inline int4 select(const int4& mask, const int4& a, const int4& b)
+{
+#ifdef __KERNEL_SSE__
+	__m128 m = _mm_cvtepi32_ps(mask);
+	return _mm_castps_si128(_mm_or_ps(_mm_and_ps(m, _mm_castsi128_ps(a)), _mm_andnot_ps(m, _mm_castsi128_ps(b)))); /* todo: avoid cvt */
+#else
+	return make_int4((mask.x)? a.x: b.x, (mask.y)? a.y: b.y, (mask.z)? a.z: b.z, (mask.w)? a.w: b.w);
+#endif
+}
 
 __device_inline void print_int4(const char *label, const int4& a)
 {
diff --git a/intern/cycles/util/util_task.cpp b/intern/cycles/util/util_task.cpp
new file mode 100644
index 00000000000..6da9a70ec0c
--- /dev/null
+++ b/intern/cycles/util/util_task.cpp
@@ -0,0 +1,223 @@
+/*
+ * Copyright 2011, Blender Foundation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+ */
+
+#include "util_debug.h"
+#include "util_foreach.h"
+#include "util_system.h"
+#include "util_task.h"
+
+CCL_NAMESPACE_BEGIN
+
+/* Task Pool */
+
+TaskPool::TaskPool(const TaskRunFunction& run_)
+{
+	num = 0;
+	num_done = 0;
+
+	do_cancel = false;
+
+	run = run_;
+}
+
+TaskPool::~TaskPool()
+{
+	stop();
+}
+
+void TaskPool::push(Task *task, bool front)
+{
+	TaskScheduler::Entry entry;
+
+	entry.task = task;
+	entry.pool = this;
+
+	TaskScheduler::push(entry, front);
+}
+
+void TaskPool::wait()
+{
+	thread_scoped_lock lock(done_mutex);
+
+	while(num_done != num)
+		done_cond.wait(lock);
+}
+
+void TaskPool::cancel()
+{
+	TaskScheduler::clear(this);
+
+	do_cancel = true;
+	wait();
+	do_cancel = false;
+}
+
+void TaskPool::stop()
+{
+	TaskScheduler::clear(this);
+
+	assert(num_done == num);
+}
+
+bool TaskPool::cancelled()
+{
+	return do_cancel;
+}
+
+void TaskPool::done_increase(int done)
+{
+	done_mutex.lock();
+	num_done += done;
+	done_mutex.unlock();
+
+	assert(num_done <= num);
+	done_cond.notify_all();
+}
+
+/* Task Scheduler */
+
+thread_mutex TaskScheduler::mutex;
+int TaskScheduler::users = 0;
+vector<thread*> TaskScheduler::threads;
+volatile bool TaskScheduler::do_exit = false;
+
+list<TaskScheduler::Entry> TaskScheduler::queue;
+thread_mutex TaskScheduler::queue_mutex;
+thread_condition_variable TaskScheduler::queue_cond;
+
+void TaskScheduler::init(int num_threads)
+{
+	thread_scoped_lock lock(mutex);
+
+	/* multiple cycles instances can use this task scheduler, sharing the same
+	   threads, so we keep track of the number of users. */
+	if(users == 0) {
+		do_exit = false;
+
+		/* launch threads that will be waiting for work */
+		if(num_threads == 0)
+			num_threads = system_cpu_thread_count();
+
+		threads.resize(num_threads);
+
+		for(size_t i = 0; i < threads.size(); i++)
+			threads[i] = new thread(function_bind(&TaskScheduler::thread_run, i));
+	}
+	
+	users++;
+}
+
+void TaskScheduler::exit()
+{
+	thread_scoped_lock lock(mutex);
+
+	users--;
+
+	if(users == 0) {
+		/* stop all waiting threads */
+		do_exit = true;
+		TaskScheduler::queue_cond.notify_all();
+
+		/* delete threads */
+		foreach(thread *t, threads) {
+			t->join();
+			delete t;
+		}
+
+		threads.clear();
+	}
+}
+
+bool TaskScheduler::thread_wait_pop(Entry& entry)
+{
+	thread_scoped_lock lock(queue_mutex);
+
+	while(queue.empty() && !do_exit)
+		queue_cond.wait(lock);
+
+	if(queue.empty()) {
+		assert(do_exit);
+		return false;
+	}
+	
+	entry = queue.front();
+	queue.pop_front();
+
+	return true;
+}
+
+void TaskScheduler::thread_run(int thread_id)
+{
+	Entry entry;
+
+	/* todo: test affinity/denormal mask */
+
+	/* keep popping off tasks */
+	while(thread_wait_pop(entry)) {
+		/* run task */
+		entry.pool->run(entry.task, thread_id);
+
+		/* delete task */
+		delete entry.task;
+
+		/* notify pool task was done */
+		entry.pool->done_increase(1);
+	}
+}
+
+void TaskScheduler::push(Entry& entry, bool front)
+{
+	/* add entry to queue */
+	TaskScheduler::queue_mutex.lock();
+	if(front)
+		TaskScheduler::queue.push_front(entry);
+	else
+		TaskScheduler::queue.push_back(entry);
+	entry.pool->num++;
+	TaskScheduler::queue_mutex.unlock();
+
+	TaskScheduler::queue_cond.notify_one();
+}
+
+void TaskScheduler::clear(TaskPool *pool)
+{
+	thread_scoped_lock lock(TaskScheduler::queue_mutex);
+
+	/* erase all tasks from this pool from the queue */
+	list<TaskScheduler::Entry>::iterator it = TaskScheduler::queue.begin();
+	int done = 0;
+
+	while(it != TaskScheduler::queue.end()) {
+		TaskScheduler::Entry& entry = *it;
+
+		if(entry.pool == pool) {
+			done++;
+			delete entry.task;
+
+			it = TaskScheduler::queue.erase(it);
+		}
+		else
+			it++;
+	}
+
+	/* notify done */
+	pool->done_increase(done);
+}
+
+CCL_NAMESPACE_END
+
diff --git a/intern/cycles/util/util_task.h b/intern/cycles/util/util_task.h
new file mode 100644
index 00000000000..acdb2cb50a2
--- /dev/null
+++ b/intern/cycles/util/util_task.h
@@ -0,0 +1,122 @@
+/*
+ * Copyright 2011, Blender Foundation.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
+ */
+
+#ifndef __UTIL_TASK_H__
+#define __UTIL_TASK_H__
+
+#include "util_list.h"
+#include "util_thread.h"
+#include "util_vector.h"
+
+CCL_NAMESPACE_BEGIN
+
+class Task;
+class TaskPool;
+class TaskScheduler;
+
+typedef boost::function<void(Task*,int)> TaskRunFunction;
+
+/* Task
+ *
+ * Base class for tasks to be executed in threads. */
+
+class Task
+{
+public:
+	Task() {};
+	virtual ~Task() {}
+};
+
+/* Task Pool
+ *
+ * Pool of tasks that will be executed by the central TaskScheduler.For each
+ * pool, we can wait for all tasks to be done, or cancel them before they are
+ * done.
+ *
+ * The run callback that actually executes the task may be create like this:
+ * function_bind(&MyClass::task_execute, this, _1, _2) */
+
+class TaskPool
+{
+public:
+	TaskPool(const TaskRunFunction& run);
+	~TaskPool();
+
+	void push(Task *task, bool front = false);
+
+	void wait();		/* wait until all tasks are done */
+	void cancel();		/* cancel all tasks, keep worker threads running */
+	void stop();		/* stop all worker threads */
+
+	bool cancelled();	/* for worker threads, test if cancelled */
+
+protected:
+	friend class TaskScheduler;
+
+	void done_increase(int done);
+
+	TaskRunFunction run;
+
+	thread_mutex done_mutex;
+	thread_condition_variable done_cond;
+
+	volatile int num, num_done;
+	volatile bool do_cancel;
+};
+
+/* Task Scheduler
+ * 
+ * Central scheduler that holds running threads ready to execute tasks. A singe
+ * queue holds the task from all pools. */
+
+class TaskScheduler
+{
+public:
+	static void init(int num_threads = 0);
+	static void exit();
+
+	static int num_threads() { return threads.size(); }
+
+protected:
+	friend class TaskPool;
+
+	struct Entry {
+		Task *task;
+		TaskPool *pool;
+	};
+
+	static thread_mutex mutex;
+	static int users;
+	static vector<thread*> threads;
+	static volatile bool do_exit;
+
+	static list<Entry> queue;
+	static thread_mutex queue_mutex;
+	static thread_condition_variable queue_cond;
+
+	static void thread_run(int thread_id);
+	static bool thread_wait_pop(Entry& entry);
+
+	static void push(Entry& entry, bool front);
+	static void clear(TaskPool *pool);
+};
+
+CCL_NAMESPACE_END
+
+#endif
+
diff --git a/intern/cycles/util/util_thread.h b/intern/cycles/util/util_thread.h
index 6836be203f5..3d15b342fe5 100644
--- a/intern/cycles/util/util_thread.h
+++ b/intern/cycles/util/util_thread.h
@@ -69,133 +69,6 @@ protected:
 	bool joined;
 };
 
-/* Thread Safe Queue to pass tasks from one thread to another. Tasks should be
- * pushed into the queue, while the worker thread waits to pop the next task
- * off the queue. Once all tasks are into the queue, calling stop() will stop
- * the worker threads from waiting for more tasks once all tasks are done. */
-
-template<typename T> class ThreadQueue
-{
-public:
-	ThreadQueue()
-	{
-		tot = 0;
-		tot_done = 0;
-		do_stop = false;
-		do_cancel = false;
-	}
-
-	/* Main thread functions */
-
-	/* push a task to be executed */
-	void push(const T& value)
-	{
-		thread_scoped_lock lock(queue_mutex);
-		queue.push(value);
-		tot++;
-		lock.unlock();
-
-		queue_cond.notify_one();
-	}
-
-	/* wait until all tasks are done */
-	void wait_done()
-	{
-		thread_scoped_lock lock(done_mutex);
-
-		while(tot_done != tot)
-			done_cond.wait(lock);
-	}
-
-	/* stop all worker threads */
-	void stop()
-	{
-		clear();
-		do_stop = true;
-		queue_cond.notify_all();
-	}
-
-	/* cancel all tasks, but keep worker threads running */
-	void cancel()
-	{
-		clear();
-		do_cancel = true;
-		wait_done();
-		do_cancel = false;
-	}
-
-	/* Worker thread functions
-     *
-	 * while(queue.worker_wait_pop(task)) {
-	 *		for(..) {
-	 *			... do work ...
-	 *
-	 *			if(queue.worker_cancel())
-	 *				break;
-	 *      }
-	 *		
-	 *		queue.worker_done();
-	 * }
-	 */
-
-	bool worker_wait_pop(T& value)
-	{
-		thread_scoped_lock lock(queue_mutex);
-
-		while(queue.empty() && !do_stop)
-			queue_cond.wait(lock);
-
-		if(queue.empty())
-			return false;
-		
-		value = queue.front();
-		queue.pop();
-
-		return true;
-	}
-
-	void worker_done()
-	{
-		thread_scoped_lock lock(done_mutex);
-		tot_done++;
-		lock.unlock();
-
-		assert(tot_done <= tot);
-
-		done_cond.notify_all();
-	}
-
-	bool worker_cancel()
-	{
-		return do_cancel;
-	}
-
-protected:
-	void clear()
-	{
-		thread_scoped_lock lock(queue_mutex);
-
-		while(!queue.empty()) {
-			thread_scoped_lock done_lock(done_mutex);
-			tot_done++;
-			done_lock.unlock();
-
-			queue.pop();
-		}
-
-		done_cond.notify_all();
-	}
-
-	std::queue<T> queue;
-	thread_mutex queue_mutex;
-	thread_mutex done_mutex;
-	thread_condition_variable queue_cond;
-	thread_condition_variable done_cond;
-	volatile bool do_stop;
-	volatile bool do_cancel;
-	volatile int tot, tot_done;
-};
-
 /* Thread Local Storage
  *
  * Boost implementation is a bit slow, and Mac OS X __thread is not supported
diff --git a/intern/cycles/util/util_transform.cpp b/intern/cycles/util/util_transform.cpp
index 61bc36ae888..0fd26825911 100644
--- a/intern/cycles/util/util_transform.cpp
+++ b/intern/cycles/util/util_transform.cpp
@@ -129,23 +129,26 @@ static bool transform_matrix4_gj_inverse(float R[][4], float M[][4])
 
 Transform transform_inverse(const Transform& tfm)
 {
-	union { Transform T; float M[4][4]; } R, M;
-	
-	R.T = transform_identity();
-	M.T = tfm;
+	Transform tfmR = transform_identity();
+	float M[4][4], R[4][4];
 
-	if(!transform_matrix4_gj_inverse(R.M, M.M)) {
+	memcpy(R, &tfmR, sizeof(R));
+	memcpy(M, &tfm, sizeof(M));
+
+	if(!transform_matrix4_gj_inverse(R, M)) {
 		/* matrix is degenerate (e.g. 0 scale on some axis), ideally we should
 		   never be in this situation, but try to invert it anyway with tweak */
-		M.M[0][0] += 1e-8f;
-		M.M[1][1] += 1e-8f;
-		M.M[2][2] += 1e-8f;
+		M[0][0] += 1e-8f;
+		M[1][1] += 1e-8f;
+		M[2][2] += 1e-8f;
 
-		if(!transform_matrix4_gj_inverse(R.M, M.M))
+		if(!transform_matrix4_gj_inverse(R, M))
 			return transform_identity();
 	}
 
-	return R.T;
+	memcpy(&tfmR, R, sizeof(R));
+
+	return tfmR;
 }
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/util/util_types.h b/intern/cycles/util/util_types.h
index efdda98571a..cf167707e47 100644
--- a/intern/cycles/util/util_types.h
+++ b/intern/cycles/util/util_types.h
@@ -36,23 +36,37 @@
 #define __shared
 #define __constant
 
-#ifdef __GNUC__
-#define __device_inline static inline __attribute__((always_inline))
-#else
+#ifdef _WIN32
 #define __device_inline static __forceinline
+#define __align(...) __declspec(align(__VA_ARGS__))
+#else
+#define __device_inline static inline __attribute__((always_inline))
+#define __forceinline inline __attribute__((always_inline))
+#define __align(...) __attribute__((aligned(__VA_ARGS__)))
 #endif
 
 #endif
 
+/* Bitness */
+
+#if defined(__ppc64__) || defined(__PPC64__) || defined(__x86_64__) || defined(__ia64__) || defined(_M_X64)
+#define __KERNEL_64_BIT__
+#endif
+
 /* SIMD Types */
 
-/* not needed yet, will be for qbvh
-#ifndef __KERNEL_GPU__
+/* not enabled, globally applying it just gives slowdown,
+ * but useful for testing. */
+//#define __KERNEL_SSE__
+#ifdef __KERNEL_SSE__
 
-#include <emmintrin.h>
-#include <xmmintrin.h>
+#include <xmmintrin.h> /* SSE 1 */
+#include <emmintrin.h> /* SSE 2 */
+#include <pmmintrin.h> /* SSE 3 */
+#include <tmmintrin.h> /* SSE 3 */
+#include <smmintrin.h> /* SSE 4 */
 
-#endif*/
+#endif
 
 #ifndef _WIN32
 #ifndef __KERNEL_GPU__
@@ -97,6 +111,12 @@ typedef unsigned int uint32_t;
 typedef long long int64_t;
 typedef unsigned long long uint64_t;
 
+#ifdef __KERNEL_64_BIT__
+typedef int64_t ssize_t;
+#else
+typedef int32_t ssize_t;
+#endif
+
 #endif
 
 /* Generic Memory Pointer */
@@ -108,89 +128,137 @@ typedef uint64_t device_ptr;
 struct uchar2 {
 	uchar x, y;
 
-	uchar operator[](int i) const { return *(&x + i); }
-	uchar& operator[](int i) { return *(&x + i); }
+	__forceinline uchar operator[](int i) const { return *(&x + i); }
+	__forceinline uchar& operator[](int i) { return *(&x + i); }
 };
 
 struct uchar3 {
 	uchar x, y, z;
 
-	uchar operator[](int i) const { return *(&x + i); }
-	uchar& operator[](int i) { return *(&x + i); }
+	__forceinline uchar operator[](int i) const { return *(&x + i); }
+	__forceinline uchar& operator[](int i) { return *(&x + i); }
 };
 
 struct uchar4 {
 	uchar x, y, z, w;
 
-	uchar operator[](int i) const { return *(&x + i); }
-	uchar& operator[](int i) { return *(&x + i); }
+	__forceinline uchar operator[](int i) const { return *(&x + i); }
+	__forceinline uchar& operator[](int i) { return *(&x + i); }
 };
 
 struct int2 {
 	int x, y;
 
-	int operator[](int i) const { return *(&x + i); }
-	int& operator[](int i) { return *(&x + i); }
+	__forceinline int operator[](int i) const { return *(&x + i); }
+	__forceinline int& operator[](int i) { return *(&x + i); }
 };
 
+#ifdef __KERNEL_SSE__
+struct __align(16) int3 {
+	union {
+		__m128i m128;
+		struct { int x, y, z, w; };
+	};
+
+	__forceinline int3() {}
+	__forceinline int3(const __m128i a) : m128(a) {}
+	__forceinline operator const __m128i&(void) const { return m128; }
+	__forceinline operator __m128i&(void) { return m128; }
+#else
 struct int3 {
-	int x, y, z;
+	int x, y, z, w;
+#endif
 
-	int operator[](int i) const { return *(&x + i); }
-	int& operator[](int i) { return *(&x + i); }
+	__forceinline int operator[](int i) const { return *(&x + i); }
+	__forceinline int& operator[](int i) { return *(&x + i); }
 };
 
+#ifdef __KERNEL_SSE__
+struct __align(16) int4 {
+	union {
+		__m128i m128;
+		struct { int x, y, z, w; };
+	};
+
+	__forceinline int4() {}
+	__forceinline int4(const __m128i a) : m128(a) {}
+	__forceinline operator const __m128i&(void) const { return m128; }
+	__forceinline operator __m128i&(void) { return m128; }
+#else
 struct int4 {
 	int x, y, z, w;
+#endif
 
-	int operator[](int i) const { return *(&x + i); }
-	int& operator[](int i) { return *(&x + i); }
+	__forceinline int operator[](int i) const { return *(&x + i); }
+	__forceinline int& operator[](int i) { return *(&x + i); }
 };
 
 struct uint2 {
 	uint x, y;
 
-	uint operator[](int i) const { return *(&x + i); }
-	uint& operator[](int i) { return *(&x + i); }
+	__forceinline uint operator[](uint i) const { return *(&x + i); }
+	__forceinline uint& operator[](uint i) { return *(&x + i); }
 };
 
 struct uint3 {
 	uint x, y, z;
 
-	uint operator[](int i) const { return *(&x + i); }
-	uint& operator[](int i) { return *(&x + i); }
+	__forceinline uint operator[](uint i) const { return *(&x + i); }
+	__forceinline uint& operator[](uint i) { return *(&x + i); }
 };
 
 struct uint4 {
 	uint x, y, z, w;
 
-	uint operator[](int i) const { return *(&x + i); }
-	uint& operator[](int i) { return *(&x + i); }
+	__forceinline uint operator[](uint i) const { return *(&x + i); }
+	__forceinline uint& operator[](uint i) { return *(&x + i); }
 };
 
 struct float2 {
 	float x, y;
 
-	float operator[](int i) const { return *(&x + i); }
-	float& operator[](int i) { return *(&x + i); }
+	__forceinline float operator[](int i) const { return *(&x + i); }
+	__forceinline float& operator[](int i) { return *(&x + i); }
 };
 
+#ifdef __KERNEL_SSE__
+struct __align(16) float3 {
+	union {
+		__m128 m128;
+		struct { float x, y, z, w; };
+	};
+
+	__forceinline float3() {}
+	__forceinline float3(const __m128 a) : m128(a) {}
+	__forceinline operator const __m128&(void) const { return m128; }
+	__forceinline operator __m128&(void) { return m128; }
+#else
 struct float3 {
-	float x, y, z;
-
-#ifdef WITH_OPENCL
-	float w;
+	float x, y, z, w;
 #endif
 
-	float operator[](int i) const { return *(&x + i); }
-	float& operator[](int i) { return *(&x + i); }
+	__forceinline float operator[](int i) const { return *(&x + i); }
+	__forceinline float& operator[](int i) { return *(&x + i); }
 };
 
+#ifdef __KERNEL_SSE__
+struct __align(16) float4 {
+	union {
+		__m128 m128;
+		struct { float x, y, z, w; };
+	};
+
+	__forceinline float4() {}
+	__forceinline float4(const __m128 a) : m128(a) {}
+	__forceinline operator const __m128&(void) const { return m128; }
+	__forceinline operator __m128&(void) { return m128; }
+#else
 struct float4 {
 	float x, y, z, w;
+#endif
 
-	float operator[](int i) const { return *(&x + i); }
-	float& operator[](int i) { return *(&x + i); }
+	__forceinline float operator[](int i) const { return *(&x + i); }
+	__forceinline float& operator[](int i) { return *(&x + i); }
 };
 
 #endif
@@ -201,87 +269,179 @@ struct float4 {
  * 
  * OpenCL does not support C++ class, so we use these instead. */
 
-__device uchar2 make_uchar2(uchar x, uchar y)
+__device_inline uchar2 make_uchar2(uchar x, uchar y)
 {
 	uchar2 a = {x, y};
 	return a;
 }
 
-__device uchar3 make_uchar3(uchar x, uchar y, uchar z)
+__device_inline uchar3 make_uchar3(uchar x, uchar y, uchar z)
 {
 	uchar3 a = {x, y, z};
 	return a;
 }
 
-__device uchar4 make_uchar4(uchar x, uchar y, uchar z, uchar w)
+__device_inline uchar4 make_uchar4(uchar x, uchar y, uchar z, uchar w)
 {
 	uchar4 a = {x, y, z, w};
 	return a;
 }
 
-__device int2 make_int2(int x, int y)
+__device_inline int2 make_int2(int x, int y)
 {
 	int2 a = {x, y};
 	return a;
 }
 
-__device int3 make_int3(int x, int y, int z)
+__device_inline int3 make_int3(int x, int y, int z)
 {
-	int3 a = {x, y, z};
+#ifdef __KERNEL_SSE__
+	int3 a;
+	a.m128 = _mm_set_epi32(0, z, y, x);
+#else
+	int3 a = {x, y, z, 0};
+#endif
+
 	return a;
 }
 
-__device int4 make_int4(int x, int y, int z, int w)
+__device_inline int4 make_int4(int x, int y, int z, int w)
 {
+#ifdef __KERNEL_SSE__
+	int4 a;
+	a.m128 = _mm_set_epi32(w, z, y, x);
+#else
 	int4 a = {x, y, z, w};
+#endif
+
 	return a;
 }
 
-__device uint2 make_uint2(uint x, uint y)
+__device_inline uint2 make_uint2(uint x, uint y)
 {
 	uint2 a = {x, y};
 	return a;
 }
 
-__device uint3 make_uint3(uint x, uint y, uint z)
+__device_inline uint3 make_uint3(uint x, uint y, uint z)
 {
 	uint3 a = {x, y, z};
 	return a;
 }
 
-__device uint4 make_uint4(uint x, uint y, uint z, uint w)
+__device_inline uint4 make_uint4(uint x, uint y, uint z, uint w)
 {
 	uint4 a = {x, y, z, w};
 	return a;
 }
 
-__device float2 make_float2(float x, float y)
+__device_inline float2 make_float2(float x, float y)
 {
 	float2 a = {x, y};
 	return a;
 }
 
-__device float3 make_float3(float x, float y, float z)
+__device_inline float3 make_float3(float x, float y, float z)
 {
-#ifdef WITH_OPENCL
-	float3 a = {x, y, z, 0.0f};
+#ifdef __KERNEL_SSE__
+	float3 a;
+	a.m128 = _mm_set_ps(0.0f, z, y, x);
 #else
-	float3 a = {x, y, z};
+	float3 a = {x, y, z, 0.0f};
 #endif
+
 	return a;
 }
 
-__device float4 make_float4(float x, float y, float z, float w)
+__device_inline float4 make_float4(float x, float y, float z, float w)
 {
+#ifdef __KERNEL_SSE__
+	float4 a;
+	a.m128 = _mm_set_ps(w, z, y, x);
+#else
 	float4 a = {x, y, z, w};
+#endif
+
 	return a;
 }
 
-__device int align_up(int offset, int alignment)
+__device_inline int align_up(int offset, int alignment)
 {
 	return (offset + alignment - 1) & ~(alignment - 1);
 }
 
+__device_inline int3 make_int3(int i)
+{
+#ifdef __KERNEL_SSE__
+	int3 a;
+	a.m128 = _mm_set1_epi32(i);
+#else
+	int3 a = {i, i, i, i};
+#endif
+
+	return a;
+}
+
+__device_inline int4 make_int4(int i)
+{
+#ifdef __KERNEL_SSE__
+	int4 a;
+	a.m128 = _mm_set1_epi32(i);
+#else
+	int4 a = {i, i, i, i};
+#endif
+
+	return a;
+}
+
+__device_inline float3 make_float3(float f)
+{
+#ifdef __KERNEL_SSE__
+	float3 a;
+	a.m128 = _mm_set1_ps(f);
+#else
+	float3 a = {f, f, f, f};
+#endif
+
+	return a;
+}
+
+__device_inline float4 make_float4(float f)
+{
+#ifdef __KERNEL_SSE__
+	float4 a;
+	a.m128 = _mm_set1_ps(f);
+#else
+	float4 a = {f, f, f, f};
+#endif
+
+	return a;
+}
+
+__device_inline float4 make_float4(const int4& i)
+{
+#ifdef __KERNEL_SSE__
+	float4 a;
+	a.m128 = _mm_cvtepi32_ps(i.m128);
+#else
+	float4 a = {(float)i.x, (float)i.y, (float)i.z, (float)i.w};
+#endif
+
+	return a;
+}
+
+__device_inline int4 make_int4(const float3& f)
+{
+#ifdef __KERNEL_SSE__
+	int4 a;
+	a.m128 = _mm_cvtps_epi32(f.m128);
+#else
+	int4 a = {(int)f.x, (int)f.y, (int)f.z, (int)f.w};
+#endif
+
+	return a;
+}
+
 #endif
 
 CCL_NAMESPACE_END