8 files changed, 48 insertions, 21 deletions

diff --git a/intern/cycles/kernel/kernel_bvh.h b/intern/cycles/kernel/kernel_bvh.h
index 93e546eaece..17791f4f35a 100644
--- a/intern/cycles/kernel/kernel_bvh.h
+++ b/intern/cycles/kernel/kernel_bvh.h
@@ -235,7 +235,7 @@ ccl_device_inline bool bvh_cardinal_curve_intersect(KernelGlobals *kg, Intersect
 	int prim = kernel_tex_fetch(__prim_index, curveAddr);
 
 #ifdef __KERNEL_SSE2__
-	__m128 vdir = _mm_div_ps(_mm_set1_ps(1.0f), (__m128 &)idir);
+	__m128 vdir = _mm_div_ps(_mm_set1_ps(1.0f), load_m128(idir));
 	__m128 vcurve_coef[4];
 	const float3 *curve_coef = (float3 *)vcurve_coef;
 	
@@ -268,10 +268,11 @@ ccl_device_inline bool bvh_cardinal_curve_intersect(KernelGlobals *kg, Intersect
 		__m128 htfm2 = shuffle<1, 3, 2, 3>(mul_shuf, vdir0);
 
 		__m128 htfm[] = { htfm0, htfm1, htfm2 };
-		__m128 p0 = transform_point_T3(htfm, _mm_sub_ps(P0, (__m128 &)P));
-		__m128 p1 = transform_point_T3(htfm, _mm_sub_ps(P1, (__m128 &)P));
-		__m128 p2 = transform_point_T3(htfm, _mm_sub_ps(P2, (__m128 &)P));
-		__m128 p3 = transform_point_T3(htfm, _mm_sub_ps(P3, (__m128 &)P));
+		__m128 vP = load_m128(P);
+		__m128 p0 = transform_point_T3(htfm, _mm_sub_ps(P0, vP));
+		__m128 p1 = transform_point_T3(htfm, _mm_sub_ps(P1, vP));
+		__m128 p2 = transform_point_T3(htfm, _mm_sub_ps(P2, vP));
+		__m128 p3 = transform_point_T3(htfm, _mm_sub_ps(P3, vP));
 
 		float fc = 0.71f;
 		__m128 vfc = _mm_set1_ps(fc);
diff --git a/intern/cycles/kernel/kernel_bvh_subsurface.h b/intern/cycles/kernel/kernel_bvh_subsurface.h
index 6b691f3364b..40683a2da57 100644
--- a/intern/cycles/kernel/kernel_bvh_subsurface.h
+++ b/intern/cycles/kernel/kernel_bvh_subsurface.h
@@ -49,8 +49,8 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const Ray *ray, Intersectio
 
 	/* ray parameters in registers */
 	const float tmax = ray->t;
-	ccl_align(16) float3 P = ray->P;
-	ccl_align(16) float3 idir = bvh_inverse_direction(ray->D);
+	float3 P = ray->P;
+	float3 idir = bvh_inverse_direction(ray->D);
 	int object = ~0;
 	float isect_t = tmax;
 
diff --git a/intern/cycles/kernel/kernel_bvh_traversal.h b/intern/cycles/kernel/kernel_bvh_traversal.h
index bfd72b0aa16..0515a9e0fa7 100644
--- a/intern/cycles/kernel/kernel_bvh_traversal.h
+++ b/intern/cycles/kernel/kernel_bvh_traversal.h
@@ -55,8 +55,8 @@ ccl_device bool BVH_FUNCTION_NAME
 
 	/* ray parameters in registers */
 	const float tmax = ray->t;
-	ccl_align(16) float3 P = ray->P;
-	ccl_align(16) float3 idir = bvh_inverse_direction(ray->D);
+	float3 P = ray->P;
+	float3 idir = bvh_inverse_direction(ray->D);
 	int object = ~0;
 
 #if FEATURE(BVH_MOTION)
diff --git a/intern/cycles/kernel/kernel_compat_cpu.h b/intern/cycles/kernel/kernel_compat_cpu.h
index b213e91274d..a9c66ec2d68 100644
--- a/intern/cycles/kernel/kernel_compat_cpu.h
+++ b/intern/cycles/kernel/kernel_compat_cpu.h
@@ -20,9 +20,9 @@
 #define __KERNEL_CPU__
 
 #include "util_debug.h"
-#include "util_half.h"
 #include "util_math.h"
 #include "util_simd.h"
+#include "util_half.h"
 #include "util_types.h"
 
 CCL_NAMESPACE_BEGIN
diff --git a/intern/cycles/kernel/kernel_film.h b/intern/cycles/kernel/kernel_film.h
index cbd875e994c..dc5f6e7ce38 100644
--- a/intern/cycles/kernel/kernel_film.h
+++ b/intern/cycles/kernel/kernel_film.h
@@ -75,7 +75,7 @@ ccl_device void kernel_film_convert_to_half_float(KernelGlobals *kg,
 
 	float exposure = kernel_data.film.exposure;
 
-	ccl_align(16) float4 rgba_in = *in;
+	float4 rgba_in = *in;
 
 	if(exposure != 1.0f) {
 		rgba_in.x *= exposure;
@@ -83,7 +83,7 @@ ccl_device void kernel_film_convert_to_half_float(KernelGlobals *kg,
 		rgba_in.z *= exposure;
 	}
 
-	float4_store_half(out, &rgba_in, sample_scale);
+	float4_store_half(out, rgba_in, sample_scale);
 }
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/util/util_half.h b/intern/cycles/util/util_half.h
index 21192024f7f..88709955b32 100644
--- a/intern/cycles/util/util_half.h
+++ b/intern/cycles/util/util_half.h
@@ -25,7 +25,7 @@ CCL_NAMESPACE_BEGIN
 
 #ifdef __KERNEL_OPENCL__
 
-#define float4_store_half(h, f, scale) vstore_half4(*(f) * (scale), 0, h);
+#define float4_store_half(h, f, scale) vstore_half4(f * (scale), 0, h);
 
 #else
 
@@ -34,24 +34,24 @@ struct half4 { half x, y, z, w; };
 
 #ifdef __KERNEL_CUDA__
 
-ccl_device_inline void float4_store_half(half *h, const float4 *f, float scale)
+ccl_device_inline void float4_store_half(half *h, float4 f, float scale)
 {
-	h[0] = __float2half_rn(f->x * scale);
-	h[1] = __float2half_rn(f->y * scale);
-	h[2] = __float2half_rn(f->z * scale);
-	h[3] = __float2half_rn(f->w * scale);
+	h[0] = __float2half_rn(f.x * scale);
+	h[1] = __float2half_rn(f.y * scale);
+	h[2] = __float2half_rn(f.z * scale);
+	h[3] = __float2half_rn(f.w * scale);
 }
 
 #else
 
-ccl_device_inline void float4_store_half(half *h, const float4 *f, float scale)
+ccl_device_inline void float4_store_half(half *h, float4 f, float scale)
 {
 #ifndef __KERNEL_SSE2__
 	for(int i = 0; i < 4; i++) {
 		/* optimized float to half for pixels:
 		 * assumes no negative, no nan, no inf, and sets denormal to 0 */
 		union { uint i; float f; } in;
-		float fscale = (*f)[i] * scale;
+		float fscale = f[i] * scale;
 		in.f = (fscale > 0.0f)? ((fscale < 65500.0f)? fscale: 65500.0f): 0.0f;
 		int x = in.i;
 
@@ -70,7 +70,7 @@ ccl_device_inline void float4_store_half(half *h, const float4 *f, float scale)
 	const __m128i mm_7FFFFFFF = _mm_set1_epi32(0x7FFFFFFF);
 	const __m128i mm_C8000000 = _mm_set1_epi32(0xC8000000);
 
-	__m128 mm_fscale = _mm_mul_ps(*(__m128*)f, mm_scale);
+	__m128 mm_fscale = _mm_mul_ps(load_m128(f), mm_scale);
 	__m128i x = _mm_castps_si128(_mm_min_ps(_mm_max_ps(mm_fscale, _mm_set_ps1(0.0f)), _mm_set_ps1(65500.0f)));
 	__m128i absolute = _mm_and_si128(x, mm_7FFFFFFF);
 	__m128i Z = _mm_add_epi32(absolute, mm_C8000000);
diff --git a/intern/cycles/util/util_simd.h b/intern/cycles/util/util_simd.h
index fd5ba1de37b..fff682bb436 100644
--- a/intern/cycles/util/util_simd.h
+++ b/intern/cycles/util/util_simd.h
@@ -180,6 +180,30 @@ ccl_device_inline const __m128 set_sign_bit(const __m128 &a)
 	return _mm_xor_ps(a, _mm_castsi128_ps(_mm_setr_epi32(S1 << 31, S2 << 31, S3 << 31, S4 << 31)));
 }
 
+#ifdef __KERNEL_WITH_SSE_ALIGN__
+ccl_device_inline const __m128 load_m128(const float4 &vec)
+{
+	return _mm_load_ps(&vec.x);
+}
+
+ccl_device_inline const __m128 load_m128(const float3 &vec)
+{
+	return _mm_load_ps(&vec.x);
+}
+
+#else
+
+ccl_device_inline const __m128 load_m128(const float4 &vec)
+{
+	return _mm_loadu_ps(&vec.x);
+}
+
+ccl_device_inline const __m128 load_m128(const float3 &vec)
+{
+	return _mm_loadu_ps(&vec.x);
+}
+#endif /* __KERNEL_WITH_SSE_ALIGN__ */
+
 #endif /* __KERNEL_SSE2__ */
 
 CCL_NAMESPACE_END
diff --git a/intern/cycles/util/util_types.h b/intern/cycles/util/util_types.h
index c770931c69b..241da40975c 100644
--- a/intern/cycles/util/util_types.h
+++ b/intern/cycles/util/util_types.h
@@ -37,6 +37,7 @@
 #define ccl_device_noinline static
 #define ccl_global
 #define ccl_constant
+#define __KERNEL_WITH_SSE_ALIGN__
 
 #if defined(_WIN32) && !defined(FREE_WINDOWS)
 
@@ -45,6 +46,7 @@
 #ifdef __KERNEL_64_BIT__
 #define ccl_try_align(...) __declspec(align(__VA_ARGS__))
 #else
+#undef __KERNEL_WITH_SSE_ALIGN__
 #define ccl_try_align(...) /* not support for function arguments (error C2719) */
 #endif
 #define ccl_may_alias