forked from bartvdbraak/blender
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.
This commit is contained in:
parent
5e5ec4c138
commit
e7c2578576
@ -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__)
|
||||
|
@ -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__)
|
||||
|
@ -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
|
||||
|
@ -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);
|
||||
@ -324,10 +324,9 @@ ccl_device_inline bool motion_triangle_intersect(KernelGlobals *kg, Intersection
|
||||
motion_triangle_vertices(kg, fobject, prim, time, verts);
|
||||
|
||||
/* ray-triangle intersection, unoptimized */
|
||||
float3 D = 1.0f/idir;
|
||||
float t, u, v;
|
||||
|
||||
if(ray_triangle_intersect_uv(P, D, isect->t, verts[2], verts[0], verts[1], &u, &v, &t)) {
|
||||
if(ray_triangle_intersect_uv(P, dir, isect->t, verts[2], verts[0], verts[1], &u, &v, &t)) {
|
||||
isect->prim = triAddr;
|
||||
isect->object = object;
|
||||
isect->type = PRIMITIVE_MOTION_TRIANGLE;
|
||||
@ -347,7 +346,7 @@ ccl_device_inline bool motion_triangle_intersect(KernelGlobals *kg, Intersection
|
||||
|
||||
#ifdef __SUBSURFACE__
|
||||
ccl_device_inline void motion_triangle_intersect_subsurface(KernelGlobals *kg, Intersection *isect_array,
|
||||
float3 P, float3 idir, float time, int object, int triAddr, float tmax, uint *num_hits, uint *lcg_state, int max_hits)
|
||||
float3 P, float3 dir, float time, int object, int triAddr, float tmax, uint *num_hits, uint *lcg_state, int max_hits)
|
||||
{
|
||||
/* primitive index for vertex location lookup */
|
||||
int prim = kernel_tex_fetch(__prim_index, triAddr);
|
||||
@ -358,10 +357,9 @@ ccl_device_inline void motion_triangle_intersect_subsurface(KernelGlobals *kg, I
|
||||
motion_triangle_vertices(kg, fobject, prim, time, verts);
|
||||
|
||||
/* ray-triangle intersection, unoptimized */
|
||||
float3 D = 1.0f/idir;
|
||||
float t, u, v;
|
||||
|
||||
if(ray_triangle_intersect_uv(P, D, tmax, verts[2], verts[0], verts[1], &u, &v, &t)) {
|
||||
if(ray_triangle_intersect_uv(P, dir, tmax, verts[2], verts[0], verts[1], &u, &v, &t)) {
|
||||
(*num_hits)++;
|
||||
|
||||
int hit;
|
||||
|
@ -361,33 +361,31 @@ ccl_device float3 particle_angular_velocity(KernelGlobals *kg, int particle)
|
||||
|
||||
/* Object intersection in BVH */
|
||||
|
||||
ccl_device_inline float3 bvh_clamp_direction(float3 dir)
|
||||
{
|
||||
/* clamp absolute values by exp2f(-80.0f) to avoid division by zero when calculating inverse direction */
|
||||
float ooeps = 8.271806E-25;
|
||||
return make_float3((fabsf(dir.x) > ooeps)? dir.x: copysignf(ooeps, dir.x),
|
||||
(fabsf(dir.y) > ooeps)? dir.y: copysignf(ooeps, dir.y),
|
||||
(fabsf(dir.z) > ooeps)? dir.z: copysignf(ooeps, dir.z));
|
||||
}
|
||||
|
||||
ccl_device_inline float3 bvh_inverse_direction(float3 dir)
|
||||
{
|
||||
/* avoid divide by zero (ooeps = exp2f(-80.0f)) */
|
||||
float ooeps = 0.00000000000000000000000082718061255302767487140869206996285356581211090087890625f;
|
||||
float3 idir;
|
||||
|
||||
idir.x = 1.0f/((fabsf(dir.x) > ooeps)? dir.x: copysignf(ooeps, dir.x));
|
||||
idir.y = 1.0f/((fabsf(dir.y) > ooeps)? dir.y: copysignf(ooeps, dir.y));
|
||||
idir.z = 1.0f/((fabsf(dir.z) > ooeps)? dir.z: copysignf(ooeps, dir.z));
|
||||
|
||||
return idir;
|
||||
return 1.0f / dir;
|
||||
}
|
||||
|
||||
/* Transform ray into object space to enter static object in BVH */
|
||||
|
||||
ccl_device_inline void bvh_instance_push(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *idir, float *t, const float tmax)
|
||||
ccl_device_inline void bvh_instance_push(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, float *t, const float tmax)
|
||||
{
|
||||
Transform tfm = object_fetch_transform(kg, object, OBJECT_INVERSE_TRANSFORM);
|
||||
|
||||
*P = transform_point(&tfm, ray->P);
|
||||
|
||||
float3 dir = transform_direction(&tfm, ray->D);
|
||||
|
||||
float len;
|
||||
dir = normalize_len(dir, &len);
|
||||
|
||||
*idir = bvh_inverse_direction(dir);
|
||||
*dir = bvh_clamp_direction(normalize_len(transform_direction(&tfm, ray->D), &len));
|
||||
*idir = bvh_inverse_direction(*dir);
|
||||
|
||||
if(*t != FLT_MAX)
|
||||
*t *= len;
|
||||
@ -395,7 +393,7 @@ ccl_device_inline void bvh_instance_push(KernelGlobals *kg, int object, const Ra
|
||||
|
||||
/* Transorm ray to exit static object in BVH */
|
||||
|
||||
ccl_device_inline void bvh_instance_pop(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *idir, float *t, const float tmax)
|
||||
ccl_device_inline void bvh_instance_pop(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, float *t, const float tmax)
|
||||
{
|
||||
if(*t != FLT_MAX) {
|
||||
Transform tfm = object_fetch_transform(kg, object, OBJECT_TRANSFORM);
|
||||
@ -403,25 +401,23 @@ ccl_device_inline void bvh_instance_pop(KernelGlobals *kg, int object, const Ray
|
||||
}
|
||||
|
||||
*P = ray->P;
|
||||
*idir = bvh_inverse_direction(ray->D);
|
||||
*dir = bvh_clamp_direction(ray->D);
|
||||
*idir = bvh_inverse_direction(*dir);
|
||||
}
|
||||
|
||||
#ifdef __OBJECT_MOTION__
|
||||
/* Transform ray into object space to enter motion blurred object in BVH */
|
||||
|
||||
ccl_device_inline void bvh_instance_motion_push(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *idir, float *t, Transform *tfm, const float tmax)
|
||||
ccl_device_inline void bvh_instance_motion_push(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, float *t, Transform *tfm, const float tmax)
|
||||
{
|
||||
Transform itfm;
|
||||
*tfm = object_fetch_transform_motion_test(kg, object, ray->time, &itfm);
|
||||
|
||||
*P = transform_point(&itfm, ray->P);
|
||||
|
||||
float3 dir = transform_direction(&itfm, ray->D);
|
||||
|
||||
float len;
|
||||
dir = normalize_len(dir, &len);
|
||||
|
||||
*idir = bvh_inverse_direction(dir);
|
||||
*dir = bvh_clamp_direction(normalize_len(transform_direction(&itfm, ray->D), &len));
|
||||
*idir = bvh_inverse_direction(*dir);
|
||||
|
||||
if(*t != FLT_MAX)
|
||||
*t *= len;
|
||||
@ -429,13 +425,14 @@ ccl_device_inline void bvh_instance_motion_push(KernelGlobals *kg, int object, c
|
||||
|
||||
/* Transorm ray to exit motion blurred object in BVH */
|
||||
|
||||
ccl_device_inline void bvh_instance_motion_pop(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *idir, float *t, Transform *tfm, const float tmax)
|
||||
ccl_device_inline void bvh_instance_motion_pop(KernelGlobals *kg, int object, const Ray *ray, float3 *P, float3 *dir, float3 *idir, float *t, Transform *tfm, const float tmax)
|
||||
{
|
||||
if(*t != FLT_MAX)
|
||||
*t *= len(transform_direction(tfm, 1.0f/(*idir)));
|
||||
|
||||
*P = ray->P;
|
||||
*idir = bvh_inverse_direction(ray->D);
|
||||
*dir = bvh_clamp_direction(ray->D);
|
||||
*idir = bvh_inverse_direction(*dir);
|
||||
}
|
||||
#endif
|
||||
|
||||
|
@ -269,12 +269,11 @@ ccl_device float3 triangle_attribute_float3(KernelGlobals *kg, const ShaderData
|
||||
* Based on Sven Woop's algorithm with precomputed triangle storage */
|
||||
|
||||
ccl_device_inline bool triangle_intersect(KernelGlobals *kg, Intersection *isect,
|
||||
float3 P, float3 idir, uint visibility, int object, int triAddr)
|
||||
float3 P, float3 dir, uint visibility, int object, int triAddr)
|
||||
{
|
||||
/* compute and check intersection t-value */
|
||||
float4 v00 = kernel_tex_fetch(__tri_woop, triAddr*TRI_NODE_SIZE+0);
|
||||
float4 v11 = kernel_tex_fetch(__tri_woop, triAddr*TRI_NODE_SIZE+1);
|
||||
float3 dir = 1.0f/idir;
|
||||
|
||||
float Oz = v00.w - P.x*v00.x - P.y*v00.y - P.z*v00.z;
|
||||
float invDz = 1.0f/(dir.x*v00.x + dir.y*v00.y + dir.z*v00.z);
|
||||
@ -322,12 +321,11 @@ ccl_device_inline bool triangle_intersect(KernelGlobals *kg, Intersection *isect
|
||||
|
||||
#ifdef __SUBSURFACE__
|
||||
ccl_device_inline void triangle_intersect_subsurface(KernelGlobals *kg, Intersection *isect_array,
|
||||
float3 P, float3 idir, int object, int triAddr, float tmax, uint *num_hits, uint *lcg_state, int max_hits)
|
||||
float3 P, float3 dir, int object, int triAddr, float tmax, uint *num_hits, uint *lcg_state, int max_hits)
|
||||
{
|
||||
/* compute and check intersection t-value */
|
||||
float4 v00 = kernel_tex_fetch(__tri_woop, triAddr*TRI_NODE_SIZE+0);
|
||||
float4 v11 = kernel_tex_fetch(__tri_woop, triAddr*TRI_NODE_SIZE+1);
|
||||
float3 dir = 1.0f/idir;
|
||||
|
||||
float Oz = v00.w - P.x*v00.x - P.y*v00.y - P.z*v00.z;
|
||||
float invDz = 1.0f/(dir.x*v00.x + dir.y*v00.y + dir.z*v00.z);
|
||||
|
Loading…
Reference in New Issue
Block a user