bevel.h

Go to the documentation of this file.

/* SPDX-FileCopyrightText: 2011-2022 Blender Foundation
 *
 * SPDX-License-Identifier: Apache-2.0 */
 
#pragma once
 
#include "kernel/bvh/bvh.h"
 
#include "kernel/geom/motion_triangle.h"
#include "kernel/geom/triangle.h"
#include "kernel/geom/triangle_intersect.h"
 
#include "kernel/integrator/path_state.h"
 
#include "kernel/svm/util.h"
 
CCL_NAMESPACE_BEGIN
 
#ifdef __SHADER_RAYTRACE__
 
/* Planar Cubic BSSRDF falloff, reused for bevel.
 *
 * This is basically (Rm - x)^3, with some factors to normalize it. For sampling
 * we integrate 2*pi*x * (Rm - x)^3, which gives us a quintic equation that as
 * far as I can tell has no closed form solution. So we get an iterative solution
 * instead with newton-raphson. */
 
ccl_device float svm_bevel_cubic_eval(const float radius, const float r)
{
  const float Rm = radius;
 
  if (r >= Rm) {
    return 0.0f;
  }
 
  /* integrate (2*pi*r * 10*(R - r)^3)/(pi * R^5) from 0 to R = 1 */
  const float Rm5 = (Rm * Rm) * (Rm * Rm) * Rm;
  const float f = Rm - r;
  const float num = f * f * f;
 
  return (10.0f * num) / (Rm5 * M_PI_F);
}
 
ccl_device float svm_bevel_cubic_pdf(const float radius, const float r)
{
  return svm_bevel_cubic_eval(radius, r);
}
 
/* solve 10x^2 - 20x^3 + 15x^4 - 4x^5 - xi == 0 */
ccl_device_forceinline float svm_bevel_cubic_quintic_root_find(const float xi)
{
  /* newton-raphson iteration, usually succeeds in 2-4 iterations, except
   * outside 0.02 ... 0.98 where it can go up to 10, so overall performance
   * should not be too bad */
  const float tolerance = 1e-6f;
  const int max_iteration_count = 10;
  float x = 0.25f;
  int i;
 
  for (i = 0; i < max_iteration_count; i++) {
    const float x2 = x * x;
    const float x3 = x2 * x;
    const float nx = (1.0f - x);
 
    const float f = 10.0f * x2 - 20.0f * x3 + 15.0f * x2 * x2 - 4.0f * x2 * x3 - xi;
    const float f_ = 20.0f * (x * nx) * (nx * nx);
 
    if (fabsf(f) < tolerance || f_ == 0.0f) {
      break;
    }
 
    x = saturatef(x - f / f_);
  }
 
  return x;
}
 
ccl_device void svm_bevel_cubic_sample(const float radius,
                                       const float xi,
                                       ccl_private float *r,
                                       ccl_private float *h)
{
  const float Rm = radius;
  float r_ = svm_bevel_cubic_quintic_root_find(xi);
 
  r_ *= Rm;
  *r = r_;
 
  /* h^2 + r^2 = Rm^2 */
  *h = safe_sqrtf(Rm * Rm - r_ * r_);
}
 
/* Bevel shader averaging normals from nearby surfaces.
 *
 * Sampling strategy from: BSSRDF Importance Sampling, SIGGRAPH 2013
 * http://library.imageworks.com/pdfs/imageworks-library-BSSRDF-sampling.pdf
 */
 
#  ifdef __KERNEL_OPTIX__
extern "C" __device__ float3 __direct_callable__svm_node_bevel(
#  else
ccl_device float3 svm_bevel(
#  endif
    KernelGlobals kg,
    ConstIntegratorState state,
    ccl_private ShaderData *sd,
    const float radius,
    const int num_samples)
{
  /* Early out if no sampling needed. */
  if (radius <= 0.0f || num_samples < 1 || sd->object == OBJECT_NONE) {
    return sd->N;
  }
 
  /* Can't ray-trace from shaders like displacement, before BVH exists. */
  if (kernel_data.bvh.bvh_layout == BVH_LAYOUT_NONE) {
    return sd->N;
  }
 
  /* Don't bevel for blurry indirect rays. */
  if (INTEGRATOR_STATE(state, path, min_ray_pdf) < 8.0f) {
    return sd->N;
  }
 
  /* Setup for multi intersection. */
  LocalIntersection isect;
  uint lcg_state = lcg_state_init(INTEGRATOR_STATE(state, path, rng_pixel),
                                  INTEGRATOR_STATE(state, path, rng_offset),
                                  INTEGRATOR_STATE(state, path, sample),
                                  0x64c6a40e);
 
  /* Sample normals from surrounding points on surface. */
  float3 sum_N = make_float3(0.0f, 0.0f, 0.0f);
 
  /* TODO: support ray-tracing in shadow shader evaluation? */
  RNGState rng_state;
  path_state_rng_load(state, &rng_state);
 
  for (int sample = 0; sample < num_samples; sample++) {
    float2 rand_disk = path_branched_rng_2D(
        kg, &rng_state, sample, num_samples, PRNG_SURFACE_BEVEL);
 
    /* Pick random axis in local frame and point on disk. */
    float3 disk_N;
    float3 disk_T;
    float3 disk_B;
    float pick_pdf_N;
    float pick_pdf_T;
    float pick_pdf_B;
 
    disk_N = sd->Ng;
    make_orthonormals(disk_N, &disk_T, &disk_B);
 
    const float axisu = rand_disk.x;
 
    if (axisu < 0.5f) {
      pick_pdf_N = 0.5f;
      pick_pdf_T = 0.25f;
      pick_pdf_B = 0.25f;
      rand_disk.x *= 2.0f;
    }
    else if (axisu < 0.75f) {
      const float3 tmp = disk_N;
      disk_N = disk_T;
      disk_T = tmp;
      pick_pdf_N = 0.25f;
      pick_pdf_T = 0.5f;
      pick_pdf_B = 0.25f;
      rand_disk.x = (rand_disk.x - 0.5f) * 4.0f;
    }
    else {
      const float3 tmp = disk_N;
      disk_N = disk_B;
      disk_B = tmp;
      pick_pdf_N = 0.25f;
      pick_pdf_T = 0.25f;
      pick_pdf_B = 0.5f;
      rand_disk.x = (rand_disk.x - 0.75f) * 4.0f;
    }
 
    /* Sample point on disk. */
    const float phi = M_2PI_F * rand_disk.x;
    float disk_r = rand_disk.y;
    float disk_height;
 
    /* Perhaps find something better than Cubic BSSRDF, but happens to work well. */
    svm_bevel_cubic_sample(radius, disk_r, &disk_r, &disk_height);
 
    const float3 disk_P = to_global(polar_to_cartesian(disk_r, phi), disk_T, disk_B);
 
    /* Create ray. */
    Ray ray ccl_optional_struct_init;
    ray.P = sd->P + disk_N * disk_height + disk_P;
    ray.D = -disk_N;
    ray.tmin = 0.0f;
    ray.tmax = 2.0f * disk_height;
    ray.dP = differential_zero_compact();
    ray.dD = differential_zero_compact();
    ray.time = sd->time;
    ray.self.object = OBJECT_NONE;
    ray.self.prim = PRIM_NONE;
    ray.self.light_object = OBJECT_NONE;
    ray.self.light_prim = PRIM_NONE;
 
    /* Intersect with the same object. if multiple intersections are found it
     * will use at most LOCAL_MAX_HITS hits, a random subset of all hits. */
    scene_intersect_local(kg, &ray, &isect, sd->object, &lcg_state, LOCAL_MAX_HITS);
 
    const int num_eval_hits = min(isect.num_hits, LOCAL_MAX_HITS);
 
    for (int hit = 0; hit < num_eval_hits; hit++) {
      /* Quickly retrieve P and Ng without setting up ShaderData. */
      float3 hit_P;
      if (sd->type == PRIMITIVE_TRIANGLE) {
        hit_P = triangle_point_from_uv(
            kg, sd, isect.hits[hit].prim, isect.hits[hit].u, isect.hits[hit].v);
      }
#  ifdef __OBJECT_MOTION__
      else if (sd->type == PRIMITIVE_MOTION_TRIANGLE) {
        float3 verts[3];
        motion_triangle_vertices(kg, sd->object, isect.hits[hit].prim, sd->time, verts);
        hit_P = motion_triangle_point_from_uv(kg, sd, isect.hits[hit].u, isect.hits[hit].v, verts);
      }
#  endif /* __OBJECT_MOTION__ */
 
      /* Get geometric normal. */
      float3 hit_Ng = isect.Ng[hit];
      const int object = isect.hits[hit].object;
      const int object_flag = kernel_data_fetch(object_flag, object);
      if (object_negative_scale_applied(object_flag)) {
        hit_Ng = -hit_Ng;
      }
 
      /* Compute smooth normal. */
      float3 N = hit_Ng;
      const int prim = isect.hits[hit].prim;
      const int shader = kernel_data_fetch(tri_shader, prim);
 
      if (shader & SHADER_SMOOTH_NORMAL) {
        const float u = isect.hits[hit].u;
        const float v = isect.hits[hit].v;
 
        if (sd->type == PRIMITIVE_TRIANGLE) {
          N = triangle_smooth_normal(kg, N, prim, u, v);
        }
#  ifdef __OBJECT_MOTION__
        else if (sd->type == PRIMITIVE_MOTION_TRIANGLE) {
          N = motion_triangle_smooth_normal(kg, N, sd->object, prim, u, v, sd->time);
        }
#  endif /* __OBJECT_MOTION__ */
      }
 
      /* Transform normals to world space. */
      if (!(object_flag & SD_OBJECT_TRANSFORM_APPLIED)) {
        object_normal_transform(kg, sd, &N);
        object_normal_transform(kg, sd, &hit_Ng);
      }
 
      /* Probability densities for local frame axes. */
      const float pdf_N = pick_pdf_N * fabsf(dot(disk_N, hit_Ng));
      const float pdf_T = pick_pdf_T * fabsf(dot(disk_T, hit_Ng));
      const float pdf_B = pick_pdf_B * fabsf(dot(disk_B, hit_Ng));
 
      /* Multiple importance sample between 3 axes, power heuristic
       * found to be slightly better than balance heuristic. pdf_N
       * in the MIS weight and denominator canceled out. */
      float w = pdf_N / (sqr(pdf_N) + sqr(pdf_T) + sqr(pdf_B));
      if (isect.num_hits > LOCAL_MAX_HITS) {
        w *= isect.num_hits / (float)LOCAL_MAX_HITS;
      }
 
      /* Real distance to sampled point. */
      const float r = len(hit_P - sd->P);
 
      /* Compute weight. */
      const float pdf = svm_bevel_cubic_pdf(radius, r);
      const float disk_pdf = svm_bevel_cubic_pdf(radius, disk_r);
 
      w *= pdf / disk_pdf;
 
      /* Sum normal and weight. */
      sum_N += w * N;
    }
  }
 
  /* Normalize. */
  const float3 N = safe_normalize(sum_N);
  return is_zero(N) ? sd->N : (sd->flag & SD_BACKFACING) ? -N : N;
}
 
template<uint node_feature_mask, typename ConstIntegratorGenericState>
#  if defined(__KERNEL_OPTIX__)
ccl_device_inline
#  else
ccl_device_noinline
#  endif
    void
    svm_node_bevel(KernelGlobals kg,
                   ConstIntegratorGenericState state,
                   ccl_private ShaderData *sd,
                   ccl_private float *stack,
                   const uint4 node)
{
  uint num_samples;
  uint radius_offset;
  uint normal_offset;
  uint out_offset;
  svm_unpack_node_uchar4(node.y, &num_samples, &radius_offset, &normal_offset, &out_offset);
 
  float3 bevel_N = sd->N;
 
  IF_KERNEL_NODES_FEATURE(RAYTRACE)
  {
    float radius = stack_load_float(stack, radius_offset);
 
#  ifdef __KERNEL_OPTIX__
    bevel_N = optixDirectCall<float3>(1, kg, state, sd, radius, num_samples);
#  else
    bevel_N = svm_bevel(kg, state, sd, radius, num_samples);
#  endif
 
    if (stack_valid(normal_offset)) {
      /* Preserve input normal. */
      const float3 ref_N = stack_load_float3(stack, normal_offset);
      bevel_N = normalize(ref_N + (bevel_N - sd->N));
    }
  }
 
  stack_store_float3(stack, out_offset, bevel_N);
}
 
#endif /* __SHADER_RAYTRACE__ */
 
CCL_NAMESPACE_END