Commit: e7c2578576380288befcd77e88edd8ae508ed01a
Author: Sv. Lockal
Date: Thu Apr 3 22:08:53 2014 +0400
https://developer.blender.org/rBe7c2578576380288befcd77e88edd8ae508ed01a
Cycles: avoid 1.0f/(1.0f/x) divisions, which msvc (only) can't optimize.
This makes bmw scene in msvc 12 builds 6% faster.
It also gives a minor speedup for SSE hair in all compilers.
===================================================================
M intern/cycles/kernel/geom/geom_bvh_subsurface.h
M intern/cycles/kernel/geom/geom_bvh_traversal.h
M intern/cycles/kernel/geom/geom_curve.h
M intern/cycles/kernel/geom/geom_motion_triangle.h
M intern/cycles/kernel/geom/geom_object.h
M intern/cycles/kernel/geom/geom_triangle.h
===================================================================
diff --git a/intern/cycles/kernel/geom/geom_bvh_subsurface.h
b/intern/cycles/kernel/geom/geom_bvh_subsurface.h
index ae4641d..6b71ffc 100644
--- a/intern/cycles/kernel/geom/geom_bvh_subsurface.h
+++ b/intern/cycles/kernel/geom/geom_bvh_subsurface.h
@@ -50,7 +50,8 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const
Ray *ray, Intersectio
/* ray parameters in registers */
const float tmax = ray->t;
float3 P = ray->P;
- float3 idir = bvh_inverse_direction(ray->D);
+ float3 dir = bvh_clamp_direction(ray->D);
+ float3 idir = bvh_inverse_direction(dir);
int object = OBJECT_NONE;
float isect_t = tmax;
@@ -215,11 +216,11 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg,
const Ray *ray, Intersectio
switch(type & PRIMITIVE_ALL) {
case
PRIMITIVE_TRIANGLE: {
-
triangle_intersect_subsurface(kg, isect_array, P, idir, object, primAddr,
isect_t, &num_hits, lcg_state, max_hits);
+
triangle_intersect_subsurface(kg, isect_array, P, dir, object, primAddr,
isect_t, &num_hits, lcg_state, max_hits);
break;
}
case
PRIMITIVE_MOTION_TRIANGLE: {
-
motion_triangle_intersect_subsurface(kg, isect_array, P, idir, ray->time,
object, primAddr, isect_t, &num_hits, lcg_state, max_hits);
+
motion_triangle_intersect_subsurface(kg, isect_array, P, dir, ray->time,
object, primAddr, isect_t, &num_hits, lcg_state, max_hits);
break;
}
default: {
@@ -235,9 +236,9 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const
Ray *ray, Intersectio
object = subsurface_object;
#if FEATURE(BVH_MOTION)
- bvh_instance_motion_push(kg,
object, ray, &P, &idir, &isect_t, &ob_tfm, tmax);
+ bvh_instance_motion_push(kg,
object, ray, &P, &dir, &idir, &isect_t, &ob_tfm, tmax);
#else
- bvh_instance_push(kg, object,
ray, &P, &idir, &isect_t, tmax);
+ bvh_instance_push(kg, object,
ray, &P, &dir, &idir, &isect_t, tmax);
#endif
#if defined(__KERNEL_SSE2__)
@@ -271,9 +272,9 @@ ccl_device uint BVH_FUNCTION_NAME(KernelGlobals *kg, const
Ray *ray, Intersectio
/* instance pop */
#if FEATURE(BVH_MOTION)
- bvh_instance_motion_pop(kg, object, ray, &P, &idir,
&isect_t, &ob_tfm, tmax);
+ bvh_instance_motion_pop(kg, object, ray, &P, &dir,
&idir, &isect_t, &ob_tfm, tmax);
#else
- bvh_instance_pop(kg, object, ray, &P, &idir, &isect_t,
tmax);
+ bvh_instance_pop(kg, object, ray, &P, &dir, &idir,
&isect_t, tmax);
#endif
#if defined(__KERNEL_SSE2__)
diff --git a/intern/cycles/kernel/geom/geom_bvh_traversal.h
b/intern/cycles/kernel/geom/geom_bvh_traversal.h
index 153efe3..566aa42 100644
--- a/intern/cycles/kernel/geom/geom_bvh_traversal.h
+++ b/intern/cycles/kernel/geom/geom_bvh_traversal.h
@@ -55,7 +55,8 @@ ccl_device bool BVH_FUNCTION_NAME
/* ray parameters in registers */
const float tmax = ray->t;
float3 P = ray->P;
- float3 idir = bvh_inverse_direction(ray->D);
+ float3 dir = bvh_clamp_direction(ray->D);
+ float3 idir = bvh_inverse_direction(dir);
int object = OBJECT_NONE;
#if FEATURE(BVH_MOTION)
@@ -253,11 +254,11 @@ ccl_device bool BVH_FUNCTION_NAME
switch(type & PRIMITIVE_ALL) {
case
PRIMITIVE_TRIANGLE: {
- hit =
triangle_intersect(kg, isect, P, idir, visibility, object, primAddr);
+ hit =
triangle_intersect(kg, isect, P, dir, visibility, object, primAddr);
break;
}
case
PRIMITIVE_MOTION_TRIANGLE: {
- hit =
motion_triangle_intersect(kg, isect, P, idir, ray->time, visibility, object,
primAddr);
+ hit =
motion_triangle_intersect(kg, isect, P, dir, ray->time, visibility, object,
primAddr);
break;
}
#if FEATURE(BVH_HAIR)
@@ -265,14 +266,14 @@ ccl_device bool BVH_FUNCTION_NAME
case
PRIMITIVE_MOTION_CURVE: {
#if FEATURE(BVH_HAIR_MINIMUM_WIDTH)
if(kernel_data.curve.curveflags & CURVE_KN_INTERPOLATE)
- hit =
bvh_cardinal_curve_intersect(kg, isect, P, idir, visibility, object, primAddr,
ray->time, type, lcg_state, difl, extmax);
+ hit =
bvh_cardinal_curve_intersect(kg, isect, P, dir, visibility, object, primAddr,
ray->time, type, lcg_state, difl, extmax);
else
- hit =
bvh_curve_intersect(kg, isect, P, idir, visibility, object, primAddr,
ray->time, type, lcg_state, difl, extmax);
+ hit =
bvh_curve_intersect(kg, isect, P, dir, visibility, object, primAddr, ray->time,
type, lcg_state, difl, extmax);
#else
if(kernel_data.curve.curveflags & CURVE_KN_INTERPOLATE)
- hit =
bvh_cardinal_curve_intersect(kg, isect, P, idir, visibility, object, primAddr,
ray->time, type);
+ hit =
bvh_cardinal_curve_intersect(kg, isect, P, dir, visibility, object, primAddr,
ray->time, type);
else
- hit =
bvh_curve_intersect(kg, isect, P, idir, visibility, object, primAddr,
ray->time, type);
+ hit =
bvh_curve_intersect(kg, isect, P, dir, visibility, object, primAddr, ray->time,
type);
#endif
break;
@@ -306,9 +307,9 @@ ccl_device bool BVH_FUNCTION_NAME
object =
kernel_tex_fetch(__prim_object, -primAddr-1);
#if FEATURE(BVH_MOTION)
- bvh_instance_motion_push(kg, object,
ray, &P, &idir, &isect->t, &ob_tfm, tmax);
+ bvh_instance_motion_push(kg, object,
ray, &P, &dir, &idir, &isect->t, &ob_tfm, tmax);
#else
- bvh_instance_push(kg, object, ray, &P,
&idir, &isect->t, tmax);
+ bvh_instance_push(kg, object, ray, &P,
&dir, &idir, &isect->t, tmax);
#endif
#if defined(__KERNEL_SSE2__)
@@ -336,9 +337,9 @@ ccl_device bool BVH_FUNCTION_NAME
/* instance pop */
#if FEATURE(BVH_MOTION)
- bvh_instance_motion_pop(kg, object, ray, &P, &idir,
&isect->t, &ob_tfm, tmax);
+ bvh_instance_motion_pop(kg, object, ray, &P, &dir,
&idir, &isect->t, &ob_tfm, tmax);
#else
- bvh_instance_pop(kg, object, ray, &P, &idir, &isect->t,
tmax);
+ bvh_instance_pop(kg, object, ray, &P, &dir, &idir,
&isect->t, tmax);
#endif
#if defined(__KERNEL_SSE2__)
diff --git a/intern/cycles/kernel/geom/geom_curve.h
b/intern/cycles/kernel/geom/geom_curve.h
index b508f50..e57bcd8 100644
--- a/intern/cycles/kernel/geom/geom_curve.h
+++ b/intern/cycles/kernel/geom/geom_curve.h
@@ -205,12 +205,12 @@ ccl_device_inline __m128 transform_point_T3(const __m128
t[3], const __m128 &a)
#endif
#ifdef __KERNEL_SSE2__
-/* Pass P and idir by reference to aligned vector */
+/* Pass P and dir by reference to aligned vector */
ccl_device_inline bool bvh_cardinal_curve_intersect(KernelGlobals *kg,
Intersection *isect,
- const float3 &P, const float3 &idir, uint visibility, int object, int
curveAddr, float time, int type, uint *lcg_state, float difl, float extmax)
+ const float3 &P, const float3 &dir, uint visibility, int object, int
curveAddr, float time, int type, uint *lcg_state, float difl, float extmax)
#else
ccl_device_inline bool bvh_cardinal_curve_intersect(KernelGlobals *kg,
Intersection *isect,
- float3 P, float3 idir, uint visibility, int object, int curveAddr,
float time,int type, uint *lcg_state, float difl, float extmax)
+ float3 P, float3 dir, uint visibility, int object, int curveAddr, float
time,int type, uint *lcg_state, float difl, float extmax)
#endif
{
int segment = PRIMITIVE_UNPACK_SEGMENT(type);
@@ -222,7 +222,7 @@ 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), load_m128(idir));
+ __m128 vdir = load_m128(dir);
__m128 vcurve_coef[4];
const float3 *curve_coef = (float3 *)vcurve_coef;
@@ -285,8 +285,6 @@ ccl_device_inline bool
bvh_cardinal_curve_intersect(KernelGlobals *kg, Intersect
float3 curve_coef[4];
/* curve Intersection check */
- float3 dir = 1.0f/idir;
-
/* obtain curve parameters */
{
/* ray transform created - this should be created at beginning
of intersection loop */
@@ -597,7 +595,7 @@ ccl_device_inline bool
bvh_cardinal_curve_intersect(KernelGlobals *kg, Intersect
}
ccl_device_inline bool bvh_curve_intersect(KernelGlobals *kg, Intersection
*isect,
- float3 P, float3 idir, uint visibility, int object, int curveAddr,
float time, int type, uint *lcg_state, float difl, float extmax)
+ float3 P, float3 direction, uint visibility, int object, int curveAddr,
float time, int type, uint *lcg_state, float difl, float extmax)
{
/* define few macros to minimize code duplication for SSE */
#ifndef __KERNEL_SSE2__
@@ -647,9 +645,9 @@ ccl_device_inline bool bvh_curve_intersect(KernelGlobals
*kg, Intersection *isec
}
/* --- */
- float3 dir = 1.0f / idir;
float3 p21_diff = p2 - p1;
float3 sphere_dif1 = (dif + dif_second) * 0.5f;
+ float3 dir = direction;
float sphere_b_tmp = dot3(dir, sphere_dif1);
float3 sphere_dif2 = sphere_dif1 - sphere_b_tmp * dir;
#else
@@ -680,9 +678,9 @@ ccl_device_inline bool bvh_curve_intersect(KernelGlobals
*kg, Intersection *isec
float or1 = _mm_cvtss_f32(or12), or2 =
_mm_cvtss_f32(broadcast<2>(or12));
float r1 = _mm_cvtss_f32(r12), r2 = _mm_cvtss_f32(broadcast<2>(r12));
- const __m128 dir = _mm_div_ps(_mm_set1_ps(1.0f), load_m128(idir));
const __m128 p21_diff = _mm_sub_ps(P_curve[1], P_curve[0]);
const __m128 sphere_dif1 = _mm_mul_ps(_mm_add_ps(dif, dif_second),
_mm_set1_ps(0.5f));
+ const __m128 dir = load_m128(direction);
const __m128 sphere_b_tmp = dot3_splat(dir, sphere_dif1);
const __m128 sphere_dif2 = fnma(sphere_b_tmp, dir, sphere_dif1);
#endif
diff --git a/intern/cycles/kernel/geom/geom_motion_triangle.h
b/intern/cycles/kernel/geom/geom_motion_triangle.h
index c5eb097..73338bb 100644
--- a/intern/cycles/kernel/geom/geom_motion_triangle.h
+++ b/intern/cycles/kernel/geom/geom_motion_triangle.h
@@ -313,7 +313,7 @@ ccl_device_noinline void
motion_triangle_shader_setup(KernelGlobals *kg, ShaderD
* time and do a ray intersection with the resulting triangle */
ccl_device_inline bool motion_triangle_intersect(KernelGlobals *kg,
Intersection *isect,
- float3 P, float3 idir, float time, uint visibility, int object, int
triAddr)
+ float3 P, float3 dir, float time, uint visibility, int object, int
triAddr)
{
/* primitive index for vertex location lookup */
int prim = kernel_tex_fetch(__prim_index, triAddr);
@@ -324,
@@ Diff output truncated at 10240 characters. @@
_______________________________________________
Bf-blender-cvs mailing list
[email protected]
http://lists.blender.org/mailman/listinfo/bf-blender-cvs
