intern/cycles/blender/mesh.cpp

@@ -26,7 +26,7 @@
 #include "util/log.h"
 #include "util/math.h"
 
-#include "mikktspace.hh"
+#include "mikktspace_ref.hh"
 
 #include "BKE_attribute.hh"
 #include "BKE_attribute_math.hh"
@@ -229,12 +229,12 @@ static void mikk_compute_tangents(
   if (is_subd) {
     MikkMeshWrapper<true> userdata(b_mesh, layer_name, mesh, tangent, tangent_sign);
     /* Compute tangents. */
-    mikk::Mikktspace(userdata).genTangSpace();
+    mikk::RefMikktspace(userdata).genTangSpace();
   }
   else {
     MikkMeshWrapper<false> userdata(b_mesh, layer_name, mesh, tangent, tangent_sign);
     /* Compute tangents. */
-    mikk::Mikktspace(userdata).genTangSpace();
+    mikk::RefMikktspace(userdata).genTangSpace();
   }
 }
 

intern/mikktspace/mikk_atomic_hash_set.hh

@@ -77,6 +77,14 @@ class AtomicHashSet {
   typedef typename std::conditional<isAtomic, std::atomic<KeyT>, KeyT>::type cell_type;
   std::vector<cell_type> cells_;
 
+  static inline size_t power_of_two(size_t x)
+  {
+    size_t y = 1;
+    while (y < x)
+      y *= 2;
+    return y;
+  }
+
  public:
   struct Config {
     KeyT emptyKey;
@@ -97,7 +105,7 @@ class AtomicHashSet {
                 KeyHash hasher = KeyHash(),
                 KeyEqual equalityChecker = KeyEqual(),
                 const Config &c = Config())
-      : capacity_(size_t(double(maxSize) / c.maxLoadFactor) + 1),
+      : capacity_(power_of_two(size_t(double(maxSize) / c.maxLoadFactor) + 1)),
         kEmptyKey_(c.emptyKey),
         hasher_(hasher),
         equalityChecker_(equalityChecker),

intern/mikktspace/mikk_simd.hh

@@ -0,0 +1,402 @@
+#pragma once
+
+namespace mikk {
+
+/* Only functions using v8sf require AVX2 (and some of those only require AVX).
+ * The rest requires at most SSE4.1, which has been included in all x86 CPUs since
+ * 2013 or so, and may be assumed to exist.
+ *
+ * We must separately test for AVX2 CPU support at runtime, becase Intel
+ * was still making non-AVX2 CPUs as late as 2020. */
+
+#ifdef __GNUC__
+#  pragma GCC push_options
+#  pragma GCC target("avx2", "sse4.1")
+#endif
+
+#ifdef __GNUC__
+typedef float v4sf __attribute__((vector_size(16)));
+typedef float v8sf __attribute__((vector_size(32)));
+
+inline v4sf to_float4(const float3 &x)
+{
+  return v4sf{x.x, x.y, x.z, 0.0f};
+}
+
+inline v4sf to_float4_masked(const float3 &x)
+{
+  typedef int v4si __attribute__((vector_size(16)));
+  return __builtin_ia32_maskloadps((const v4sf *)&x, v4si{-1, -1, -1, 0});
+}
+
+inline float3 to_float3(v4sf x)
+{
+  return float3(x[0], x[1], x[2]);
+}
+inline v8sf from_2xv4sf(v4sf a, v4sf b)
+{
+  v8sf a2 = __builtin_ia32_ps256_ps(a);
+  return __builtin_ia32_vinsertf128_ps256(a2, b, 1);
+}
+
+inline v8sf from_v4sf(v4sf a)
+{
+  v8sf a2 = __builtin_ia32_ps256_ps(a);
+  return __builtin_ia32_vinsertf128_ps256(a2, a, 1);
+}
+
+inline v4sf v4sf_set(float x)
+{
+  return v4sf{x, x, x, x};
+}
+inline v4sf v4sf_setzero()
+{
+  return v4sf{0, 0, 0, 0};
+}
+inline v8sf v8sf_set(float x)
+{
+  return v8sf{x, x, x, x, x, x, x, x};
+}
+inline v8sf v8sf_setzero()
+{
+  return v8sf{0, 0, 0, 0, 0, 0, 0, 0};
+}
+inline v8sf v8sf_set2(float x, float y)
+{
+  return v8sf{x, x, x, x, y, y, y, y};
+}
+
+inline float extract(v4sf v, uint x)
+{
+  return v[x];
+}
+inline float extract(v8sf v, uint x)
+{
+  return v[x];
+}
+
+inline v4sf insert(v4sf v, int x, float y)
+{
+  v[x] = y;
+  return v;
+}
+inline v8sf insert(v8sf v, int x, float y)
+{
+  v[x] = y;
+  return v;
+}
+
+#elif _MSC_VER
+
+/* Without this line, the compiler may use FMA ops to rewrite float arithmetics.
+ * Sometimes this results in slight changes in numerical results.
+ *
+ * For example,
+ * 'a*b-c*d' might compile into  "temp = mul(c*d); out = fmsub(a,b,temp);"
+ *
+ * Here, temp is truncated to float32, but fmsub does everything internally in
+ * float64 precision, so executing this for a=c and b=d no longer produces a zero
+ *   (but something on the order of |ab| * 10^-13).
+ *
+ * This would break our reference compatibility, because, e.g., the output of initTriangle()
+ * changes fundamentally if some internal values go from exactly zero to almost zero. */
+#  pragma fp_contract(off)
+
+typedef __m128 v4sf;
+typedef __m128i v4si;
+typedef __m256 v8sf;
+
+inline v4sf to_float4(const float3 &x)
+{
+  return _mm_set_ps(0, x.z, x.y, x.x);
+}
+inline v4sf to_float4_masked(const float3 &x)
+{
+  return _mm_maskload_ps(&x.x, _mm_set_epi32(0, -1, -1, -1));
+}
+
+inline float3 to_float3(v4sf x)
+{
+  return float3(x.m128_f32[0], x.m128_f32[1], x.m128_f32[2]);
+}
+inline v8sf from_2xv4sf(v4sf a, v4sf b)
+{
+  return _mm256_set_m128(b, a);
+}
+
+inline v8sf from_v4sf(v4sf a)
+{
+  return _mm256_set_m128(a, a);
+}
+
+inline v4sf v4sf_set(float x)
+{
+  return _mm_set1_ps(x);
+}
+inline v4si v4si_set(int x)
+{
+  return _mm_set1_epi32(x);
+}
+inline v4sf v4sf_setzero()
+{
+  return _mm_setzero_ps();
+}
+inline v8sf v8sf_setzero()
+{
+  return _mm256_setzero_ps();
+}
+inline v8sf v8sf_set(float x)
+{
+  return _mm256_set1_ps(x);
+}
+inline v8sf v8sf_set2(float x, float y)
+{
+  return _mm256_set_ps(y, y, y, y, x, x, x, x);
+}
+
+#  define __builtin_ia32_dpps _mm_dp_ps
+#  define __builtin_ia32_hsubps _mm_hsub_ps
+#  define __builtin_ia32_maxss _mm_max_ss
+#  define __builtin_ia32_maxps _mm_max_ps
+#  define __builtin_ia32_minps _mm_min_ps
+#  define __builtin_ia32_mulps _mm_mul_ps
+#  define __builtin_ia32_shufps _mm_shuffle_ps
+#  define __builtin_ia32_cmpltps _mm_cmplt_ps
+#  define __builtin_ia32_andps _mm_and_ps
+#  define __builtin_ia32_andnps _mm_andnot_ps
+#  define __builtin_ia32_rsqrtss _mm_rsqrt_ss
+#  define __builtin_ia32_sqrtps _mm_sqrt_ps
+#  define __builtin_ia32_maxps256 _mm256_max_ps
+#  define __builtin_ia32_sqrtps256 _mm256_sqrt_ps
+#  define __builtin_ia32_dpps256 _mm256_dp_ps
+#  define __builtin_ia32_vpermilps256 _mm256_permute_ps
+#  define __builtin_ia32_vextractf128_ps256 _mm256_extractf128_ps
+#  define __builtin_ia32_vperm2f128_ps256 _mm256_permute2f128_ps
+#  define __builtin_ia32_blendps256 _mm256_blend_ps
+#  define __builtin_ia32_permvarsf256 _mm256_permutevar8x32_ps
+#  define __builtin_ia32_cmpgtps _mm_cmpgt_ps
+#  define __builtin_ia32_cmpps256 _mm256_cmp_ps
+#  define __builtin_ia32_andps256 _mm256_and_ps
+#  define __builtin_ia32_andnps256 _mm256_andnot_ps
+#  define __builtin_ia32_haddps256 _mm256_hadd_ps
+
+inline float extract(__m128 v, uint x)
+{
+  return v.m128_f32[x];
+}
+inline int extract(__m128i v, uint x)
+{
+  return v.m128i_i32[x];
+}
+inline float extract(__m256 v, uint x)
+{
+  return v.m256_f32[x];
+}
+
+inline __m128 insert(__m128 v, int x, float y)
+{
+  v.m128_f32[x] = y;
+  return v;
+}
+inline __m256 insert(__m256 v, int x, float y)
+{
+  v.m256_f32[x] = y;
+  return v;
+}
+
+// GCC provides all these automatically
+inline v4sf operator*(v4sf a, v4sf b)
+{
+  return _mm_mul_ps(a, b);
+}
+inline v4sf operator*(v4sf a, float b)
+{
+  return _mm_mul_ps(a, v4sf_set(b));
+}
+inline v4sf operator+(float a, v4sf b)
+{
+  return _mm_add_ps(v4sf_set(a), b);
+}
+inline v4sf operator-(float a, v4sf b)
+{
+  return _mm_sub_ps(v4sf_set(a), b);
+}
+inline v8sf operator*(v8sf a, v8sf b)
+{
+  return _mm256_mul_ps(a, b);
+}
+inline v4sf operator-(v4sf a, v4sf b)
+{
+  return _mm_sub_ps(a, b);
+}
+inline v4sf operator+(v4sf a, v4sf b)
+{
+  return _mm_add_ps(a, b);
+}
+inline v4si operator+(v4si a, v4si b)
+{
+  return _mm_add_epi32(a, b);
+}
+inline v8sf operator+(v8sf a, v8sf b)
+{
+  return _mm256_add_ps(a, b);
+}
+inline v8sf operator-(v8sf a, v8sf b)
+{
+  return _mm256_sub_ps(a, b);
+}
+inline v8sf operator/(v8sf a, v8sf b)
+{
+  return _mm256_div_ps(a, b);
+}
+inline v4sf operator/(float a, v4sf b)
+{
+  return _mm_div_ps(v4sf_set(a), b);
+}
+inline v8sf operator/(float a, v8sf b)
+{
+  return _mm256_div_ps(v8sf_set(a), b);
+}
+inline void operator+=(v8sf &a, v8sf b)
+{
+  a = _mm256_add_ps(a, b);
+}
+inline v4si operator*(v4si a, int b)
+{
+  return _mm_mul_epi32(a, v4si_set(b));
+}
+inline v4si operator^(v4si a, v4si b)
+{
+  return _mm_xor_si128(a, b);
+}
+
+#endif
+
+inline v4sf to_float4(v4sf x)
+{
+  return x;
+}
+
+inline void unpack(v4sf &p0, v4sf &p1, v8sf v)
+{
+  p0 = __builtin_ia32_vextractf128_ps256(v, 0);
+  p1 = __builtin_ia32_vextractf128_ps256(v, 1);
+}
+
+template<int i> inline v4sf bcast(v4sf x)
+{
+  return __builtin_ia32_shufps(x, x, i * 0x55);
+}
+
+inline float dot(const v4sf &a, const v4sf &b)
+{
+  return extract(__builtin_ia32_dpps(a, b, 255), 0);
+}
+
+inline v8sf dot_single(v8sf a, v8sf b)
+{
+  return __builtin_ia32_dpps256(a, b, 0xFF);
+}
+
+inline float dot_inner(const v8sf &a)
+{
+  return extract(dot_single(a, __builtin_ia32_vperm2f128_ps256(a, a, 1)), 0);
+}
+
+/* Potentially somewhat expensive (creates a global constant and adds a memory access, though the
+ * constant may be expected to stay in cache), use sparingly. Sometimes unavoidable since few AVX2
+ * ops send data across the 128-bit boundary */
+inline v8sf cross_lane_permute(
+    v8sf v, int i0, int i1, int i2, int i3, int i4, int i5, int i6, int i7)
+{
+#if __GNUC__
+  typedef int v8si __attribute__((vector_size(32)));
+  v8si perm_mask{i0, i1, i2, i3, i4, i5, i6, i7};
+  return __builtin_ia32_permvarsf256(v, perm_mask);
+#else
+  return __builtin_ia32_permvarsf256(v, _mm256_set_epi32(i7, i6, i5, i4, i3, i2, i1, i0));
+#endif
+}
+
+/* input: [... i j ..... ], with <i> in position <offset> and <j> in position <offset+1>
+ * output: [ i i i i j j j j ] */
+template<int offset> inline v8sf unpack4x(v8sf a)
+{
+  return cross_lane_permute(
+      a, offset, offset, offset, offset, offset + 1, offset + 1, offset + 1, offset + 1);
+}
+
+inline float length_squared(const v4sf &x)
+{
+  return dot(x, x);
+}
+inline float length(const v4sf &x)
+{
+  v4sf v = __builtin_ia32_dpps(x, x, 255);
+  v = __builtin_ia32_sqrtps(v);
+  return extract(v, 0);
+}
+
+inline float inv_length(const v4sf &x)
+{
+  v4sf v = __builtin_ia32_dpps(x, x, 255);
+  v4sf one = v4sf_set(1.0f), zero = v4sf_setzero();
+  v4sf rv = 1.0f / __builtin_ia32_sqrtps(v);
+  v4sf nonzero = __builtin_ia32_cmpgtps(v, zero);
+  rv = __builtin_ia32_andps(nonzero, rv) + __builtin_ia32_andnps(nonzero, one);
+  return extract(rv, 0);
+}
+
+/* input: 8 float4's packed into 4 v8sf
+ * output: 8 floats containing inverse lengths of each */
+inline v8sf inv_length8(v8sf px0, v8sf px1, v8sf px2, v8sf px3)
+{
+  v8sf v;
+  v = __builtin_ia32_dpps256(px0, px0, 0xF1);
+  v += __builtin_ia32_dpps256(px1, px1, 0xF2);
+  v += __builtin_ia32_dpps256(px2, px2, 0xF4);
+  v += __builtin_ia32_dpps256(px3, px3, 0xF8);
+
+  /* v is [0 2 4 6 1 3 5 7], and we want [0 1 2 3 4 5 6 7] */
+  v = cross_lane_permute(v, 0, 4, 1, 5, 2, 6, 3, 7);
+
+  /* There is an intrinsic for rsqrt, but it has relative accuracy of ~1/1000 */
+
+  v8sf one = v8sf_set(1.0f), zero = v8sf_setzero();
+  v8sf nonzero = __builtin_ia32_cmpps256(v, zero, 4);
+  v8sf rv = 1.0f / __builtin_ia32_sqrtps256(v);
+  /* we must match reference behavior (return 1 if input is zero) or else
+   * we'll have significant mismatches in some situations */
+  return __builtin_ia32_andps256(nonzero, rv) + __builtin_ia32_andnps256(nonzero, one);
+}
+
+/* Projects v onto the surface with normal n. */
+inline v4sf project1(v4sf n, v4sf v)
+{
+  return v - (n * dot(n, v));
+}
+
+/* Projects two packed float3's in v8 onto surfaces in n8. */
+inline v8sf project2(v8sf n8, v8sf v8)
+{
+  v8sf d8 = dot_single(n8, v8);
+  return v8 - n8 * d8;
+}
+
+inline v4sf fast_acosf_4x(v4sf x)
+{
+  v4sf f = __builtin_ia32_maxps(x, x * -1.0f);
+  v4sf one = v4sf_set(1.0f), zero = v4sf_setzero();
+  v4sf m = __builtin_ia32_minps(f, one);
+  v4sf a = __builtin_ia32_sqrtps(one - m) *
+           (1.5707963267f + m * (-0.213300989f + m * (0.077980478f + m * -0.02164095f)));
+
+  v4sf sign = __builtin_ia32_cmpltps(x, zero);
+  return __builtin_ia32_andps(sign, 3.1415926535897932f - a) + __builtin_ia32_andnps(sign, a);
+}
+
+#ifdef __GNUC__
+#  pragma GCC pop_options
+#endif
+
+}  // namespace mikk

intern/mikktspace/mikk_util.hh

@@ -19,7 +19,7 @@ namespace mikk {
 
 inline bool not_zero(const float fX)
 {
-  return fabsf(fX) > FLT_MIN;
+  return (fX > FLT_MIN) || (-fX > FLT_MIN);
 }
 
 /* Helpers for (un)packing a 2-bit vertex index and a 30-bit face index to one integer. */

intern/mikktspace/mikktspace_ref.hh

@@ -0,0 +1,880 @@
+/* SPDX-FileCopyrightText: 2011 Morten S. Mikkelsen
+ * SPDX-FileCopyrightText: 2022 Blender Authors
+ *
+ * SPDX-License-Identifier: Apache-2.0 */
+
+/** \file
+ * \ingroup mikktspace
+ */
+
+#include <algorithm>
+#include <cassert>
+#include <unordered_map>
+
+#ifdef WITH_TBB
+#  include <tbb/parallel_for.h>
+#endif
+
+#include "mikk_atomic_hash_set.hh"
+#include "mikk_float3.hh"
+#include "mikk_util.hh"
+
+namespace mikk {
+
+static constexpr uint UNSET_ENTRY = 0xffffffffu;
+
+template<typename Mesh> class RefMikktspace {
+  struct Triangle {
+    /* Stores neighboring triangle for group assignment. */
+    std::array<uint, 3> neighbor;
+    /* Stores assigned group of each vertex. */
+    std::array<uint, 3> group;
+    /* Stores vertex indices that make up the triangle. */
+    std::array<uint, 3> vertices;
+
+    /* Computed face tangent, will be accumulated into group. */
+    float3 tangent;
+
+    /* Index of the face that this triangle belongs to. */
+    uint faceIdx;
+    /* Index of the first of this triangle's vertices' TSpaces. */
+    uint tSpaceIdx;
+
+    /* Stores mapping from this triangle's vertices to the original
+     * face's vertices (relevant for quads). */
+    std::array<uint8_t, 3> faceVertex;
+
+    // flags
+    bool markDegenerate : 1;
+    bool quadOneDegenTri : 1;
+    bool groupWithAny : 1;
+    bool orientPreserving : 1;
+
+    Triangle(uint faceIdx_, uint tSpaceIdx_)
+        : tangent{0.0f, 0.0f, 0.0f},
+          faceIdx{faceIdx_},
+          tSpaceIdx{tSpaceIdx_},
+          markDegenerate{false},
+          quadOneDegenTri{false},
+          groupWithAny{true},
+          orientPreserving{false}
+    {
+      neighbor.fill(UNSET_ENTRY);
+      group.fill(UNSET_ENTRY);
+    }
+
+    void setVertices(uint8_t i0, uint8_t i1, uint8_t i2)
+    {
+      faceVertex[0] = i0;
+      faceVertex[1] = i1;
+      faceVertex[2] = i2;
+      vertices[0] = pack_index(faceIdx, i0);
+      vertices[1] = pack_index(faceIdx, i1);
+      vertices[2] = pack_index(faceIdx, i2);
+    }
+  };
+
+  struct Group {
+    float3 tangent;
+    uint vertexRepresentative;
+    bool orientPreserving;
+
+    Group(uint vertexRepresentative_, bool orientPreserving_)
+        : tangent{0.0f},
+          vertexRepresentative{vertexRepresentative_},
+          orientPreserving{orientPreserving_}
+    {
+    }
+
+    void normalizeTSpace()
+    {
+      tangent = tangent.normalize();
+    }
+
+    void accumulateTSpaceAtomic(float3 v_tangent)
+    {
+      float_add_atomic(&tangent.x, v_tangent.x);
+      float_add_atomic(&tangent.y, v_tangent.y);
+      float_add_atomic(&tangent.z, v_tangent.z);
+    }
+
+    void accumulateTSpace(float3 v_tangent)
+    {
+      tangent += v_tangent;
+    }
+  };
+
+  struct TSpace {
+    float3 tangent = float3(1.0f, 0.0f, 0.0f);
+    uint counter = 0;
+    bool orientPreserving = false;
+
+    void accumulateGroup(const Group &group)
+    {
+      assert(counter < 2);
+
+      if (counter == 0) {
+        tangent = group.tangent;
+      }
+      else if (tangent == group.tangent) {
+        // this if is important. Due to floating point precision
+        // averaging when ts0==ts1 will cause a slight difference
+        // which results in tangent space splits later on, so do nothing
+      }
+      else {
+        tangent = (tangent + group.tangent).normalize();
+      }
+
+      counter++;
+      orientPreserving = group.orientPreserving;
+    }
+  };
+
+  Mesh &mesh;
+
+  std::vector<Triangle> triangles;
+  std::vector<TSpace> tSpaces;
+  std::vector<Group> groups;
+
+  uint nrTSpaces, nrFaces, nrTriangles, totalTriangles;
+
+  int nrThreads;
+  bool isParallel;
+
+ public:
+  bool profile = false;
+  bool trace_on = false;
+  RefMikktspace(Mesh &mesh_) : mesh(mesh_) {}
+
+  void genTangSpace()
+  {
+    nrFaces = uint(mesh.GetNumFaces());
+
+#ifdef WITH_TBB
+    nrThreads = tbb::this_task_arena::max_concurrency();
+    isParallel = (nrThreads > 1) && (nrFaces > 10000);
+    // isParallel = false;
+#else
+    nrThreads = 1;
+    isParallel = false;
+#endif
+    // make an initial triangle --> face index list
+    generateInitialVerticesIndexList();
+
+    if (nrTriangles == 0) {
+      return;
+    }
+
+    // make a welded index list of identical positions and attributes (pos, norm, texc)
+    generateSharedVerticesIndexList();
+
+    // mark all triangle pairs that belong to a quad with only one
+    // good triangle. These need special treatment in degenEpilogue().
+    // Additionally, move all good triangles to the start of
+    // triangles[] without changing order and
+    // put the degenerate triangles last.
+    degenPrologue();
+
+    if (nrTriangles == 0) {
+      // No point in building tangents if there are no non-degenerate triangles, so just zero them
+      tSpaces.resize(nrTSpaces);
+    }
+    else {
+      // evaluate triangle level attributes and neighbor list
+      initTriangle();
+
+      // match up edge pairs
+      buildNeighbors();
+
+      // based on the 4 rules, identify groups based on connectivity
+      build4RuleGroups();
+
+      // make tspaces, each group is split up into subgroups.
+      // Finally a tangent space is made for every resulting subgroup
+      generateTSpaces();
+
+      // degenerate quads with one good triangle will be fixed by copying a space from
+      // the good triangle to the coinciding vertex.
+      // all other degenerate triangles will just copy a space from any good triangle
+      // with the same welded index in vertices[].
+      degenEpilogue();
+    }
+
+    uint index = 0;
+
+    for (uint f = 0; f < nrFaces; f++) {
+      const uint verts = mesh.GetNumVerticesOfFace(f);
+      if (verts != 3 && verts != 4) {
+        continue;
+      }
+
+      // set data
+      for (uint i = 0; i < verts; i++) {
+        const TSpace &tSpace = tSpaces[index++];
+        mesh.SetTangentSpace(f, i, tSpace.tangent, tSpace.orientPreserving);
+      }
+    }
+  }
+
+ protected:
+  template<typename F> void runParallel(uint start, uint end, F func)
+  {
+#ifdef WITH_TBB
+    if (isParallel) {
+      tbb::parallel_for(start, end, func);
+    }
+    else
+#endif
+    {
+      for (uint i = start; i < end; i++) {
+        func(i);
+      }
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+
+  float3 getPosition(uint vertexID)
+  {
+    uint f, v;
+    unpack_index(f, v, vertexID);
+    return mesh.GetPosition(f, v);
+  }
+
+  float3 getNormal(uint vertexID)
+  {
+    uint f, v;
+    unpack_index(f, v, vertexID);
+
+    return mesh.GetNormal(f, v);
+  }
+
+  float3 getTexCoord(uint vertexID)
+  {
+    uint f, v;
+    unpack_index(f, v, vertexID);
+    return mesh.GetTexCoord(f, v);
+  }
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+
+  void generateInitialVerticesIndexList()
+  {
+    nrTriangles = 0;
+    for (uint f = 0; f < nrFaces; f++) {
+      const uint verts = mesh.GetNumVerticesOfFace(f);
+      if (verts == 3) {
+        nrTriangles += 1;
+      }
+      else if (verts == 4) {
+        nrTriangles += 2;
+      }
+    }
+
+    triangles.reserve(nrTriangles);
+
+    nrTSpaces = 0;
+    for (uint f = 0; f < nrFaces; f++) {
+      const uint verts = mesh.GetNumVerticesOfFace(f);
+      if (verts != 3 && verts != 4)
+        continue;
+
+      uint tA = uint(triangles.size());
+      triangles.emplace_back(f, nrTSpaces);
+
+      if (verts == 3) {
+        Triangle &triA = triangles[tA];
+        triA.setVertices(0, 1, 2);
+      }
+      else {
+        uint tB = uint(triangles.size());
+        triangles.emplace_back(f, nrTSpaces);
+        Triangle &triA = triangles[tA];
+        Triangle &triB = triangles[tB];
+
+        // need an order independent way to evaluate
+        // tspace on quads. This is done by splitting
+        // along the shortest diagonal.
+        float distSQ_02 = (mesh.GetTexCoord(f, 2) - mesh.GetTexCoord(f, 0)).length_squared();
+        float distSQ_13 = (mesh.GetTexCoord(f, 3) - mesh.GetTexCoord(f, 1)).length_squared();
+        bool quadDiagIs_02;
+        if (distSQ_02 != distSQ_13)
+          quadDiagIs_02 = (distSQ_02 < distSQ_13);
+        else {
+          distSQ_02 = (mesh.GetPosition(f, 2) - mesh.GetPosition(f, 0)).length_squared();
+          distSQ_13 = (mesh.GetPosition(f, 3) - mesh.GetPosition(f, 1)).length_squared();
+          quadDiagIs_02 = !(distSQ_13 < distSQ_02);
+        }
+
+        if (quadDiagIs_02) {
+          triA.setVertices(0, 1, 2);
+          triB.setVertices(0, 2, 3);
+        }
+        else {
+          triA.setVertices(0, 1, 3);
+          triB.setVertices(1, 2, 3);
+        }
+      }
+
+      nrTSpaces += verts;
+    }
+  }
+
+  struct VertexHash {
+    RefMikktspace<Mesh> *mikk;
+    inline uint operator()(const uint &k) const
+    {
+      return hash_float3x3(mikk->getPosition(k), mikk->getNormal(k), mikk->getTexCoord(k));
+    }
+  };
+
+  struct VertexEqual {
+    RefMikktspace<Mesh> *mikk;
+    inline bool operator()(const uint &kA, const uint &kB) const
+    {
+      return mikk->getTexCoord(kA) == mikk->getTexCoord(kB) &&
+             mikk->getNormal(kA) == mikk->getNormal(kB) &&
+             mikk->getPosition(kA) == mikk->getPosition(kB);
+    }
+  };
+
+  /* Merge identical vertices.
+   * To find vertices with identical position, normal and texcoord, we calculate a hash of the 9
+   * values. Then, by sorting based on that hash, identical elements (having identical hashes) will
+   * be moved next to each other. Since there might be hash collisions, the elements of each block
+   * are then compared with each other and duplicates are merged.
+   */
+  template<bool isAtomic> void generateSharedVerticesIndexList_impl()
+  {
+    uint numVertices = nrTriangles * 3;
+    AtomicHashSet<uint, isAtomic, VertexHash, VertexEqual> set(numVertices, {this}, {this});
+    runParallel(0u, nrTriangles, [&](uint t) {
+      for (uint i = 0; i < 3; i++) {
+        auto res = set.emplace(triangles[t].vertices[i]);
+        if (!res.second) {
+          triangles[t].vertices[i] = res.first;
+        }
+      }
+    });
+  }
+  void generateSharedVerticesIndexList()
+  {
+    if (isParallel) {
+      generateSharedVerticesIndexList_impl<true>();
+    }
+    else {
+      generateSharedVerticesIndexList_impl<false>();
+    }
+  }
+
+  /////////////////////////////////////////////////////////////////////////////////////////////
+  /////////////////////////////////// Degenerate triangles ////////////////////////////////////
+
+  void degenPrologue()
+  {
+    // Mark all degenerate triangles
+    totalTriangles = nrTriangles;
+    std::atomic<uint> degenTriangles(0);
+    runParallel(0u, totalTriangles, [&](uint t) {
+      const float3 p0 = getPosition(triangles[t].vertices[0]);
+      const float3 p1 = getPosition(triangles[t].vertices[1]);
+      const float3 p2 = getPosition(triangles[t].vertices[2]);
+      if (p0 == p1 || p0 == p2 || p1 == p2)  // degenerate
+      {
+        triangles[t].markDegenerate = true;
+        degenTriangles.fetch_add(1);
+      }
+    });
+    nrTriangles -= degenTriangles.load();
+
+    if (totalTriangles == nrTriangles) {
+      return;
+    }
+
+    // locate quads with only one good triangle
+    runParallel(0u, totalTriangles - 1, [&](uint t) {
+      Triangle &triangleA = triangles[t], &triangleB = triangles[t + 1];
+      if (triangleA.faceIdx != triangleB.faceIdx) {
+        /* Individual triangle, skip. */
+        return;
+      }
+      if (triangleA.markDegenerate != triangleB.markDegenerate) {
+        triangleA.quadOneDegenTri = true;
+        triangleB.quadOneDegenTri = true;
+      }
+    });
+
+    std::stable_partition(triangles.begin(), triangles.end(), [](const Triangle &tri) {
+      return !tri.markDegenerate;
+    });
+  }
+
+  void degenEpilogue()
+  {
+    if (nrTriangles == totalTriangles) {
+      return;
+    }
+
+    std::unordered_map<uint, uint> goodTriangleMap;
+    for (uint t = 0; t < nrTriangles; t++) {
+      for (uint i = 0; i < 3; i++) {
+        goodTriangleMap.emplace(triangles[t].vertices[i], pack_index(t, i));
+      }
+    }
+
+    // deal with degenerate triangles
+    // punishment for degenerate triangles is O(nrTriangles) extra memory.
+    for (uint t = nrTriangles; t < totalTriangles; t++) {
+      // degenerate triangles on a quad with one good triangle are skipped
+      // here but processed in the next loop
+      if (triangles[t].quadOneDegenTri) {
+        continue;
+      }
+
+      for (uint i = 0; i < 3; i++) {
+        const auto entry = goodTriangleMap.find(triangles[t].vertices[i]);
+        if (entry == goodTriangleMap.end()) {
+          // Matching vertex from good triangle is not found.
+          continue;
+        }
+
+        uint tSrc, iSrc;
+        unpack_index(tSrc, iSrc, entry->second);
+        const uint iSrcVert = triangles[tSrc].faceVertex[iSrc];
+        const uint iSrcOffs = triangles[tSrc].tSpaceIdx;
+        const uint iDstVert = triangles[t].faceVertex[i];
+        const uint iDstOffs = triangles[t].tSpaceIdx;
+        // copy tspace
+        tSpaces[iDstOffs + iDstVert] = tSpaces[iSrcOffs + iSrcVert];
+      }
+    }
+
+    // deal with degenerate quads with one good triangle
+    for (uint t = 0; t < nrTriangles; t++) {
+      // this triangle belongs to a quad where the
+      // other triangle is degenerate
+      if (!triangles[t].quadOneDegenTri) {
+        continue;
+      }
+      uint vertFlag = (1u << triangles[t].faceVertex[0]) | (1u << triangles[t].faceVertex[1]) |
+                      (1u << triangles[t].faceVertex[2]);
+      uint missingFaceVertex = 0;
+      if ((vertFlag & 2) == 0)
+        missingFaceVertex = 1;
+      else if ((vertFlag & 4) == 0)
+        missingFaceVertex = 2;
+      else if ((vertFlag & 8) == 0)
+        missingFaceVertex = 3;
+
+      uint faceIdx = triangles[t].faceIdx;
+      float3 dstP = mesh.GetPosition(faceIdx, missingFaceVertex);
+      bool found = false;
+      for (uint i = 0; i < 3; i++) {
+        const uint faceVertex = triangles[t].faceVertex[i];
+        const float3 srcP = mesh.GetPosition(faceIdx, faceVertex);
+        if (srcP == dstP) {
+          const uint offset = triangles[t].tSpaceIdx;
+          tSpaces[offset + missingFaceVertex] = tSpaces[offset + faceVertex];
+          found = true;
+          break;
+        }
+      }
+      assert(found);
+      (void)found;
+    }
+  }
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+
+  // returns the texture area times 2
+  float calcTexArea(uint tri)
+  {
+    const float3 t1 = getTexCoord(triangles[tri].vertices[0]);
+    const float3 t2 = getTexCoord(triangles[tri].vertices[1]);
+    const float3 t3 = getTexCoord(triangles[tri].vertices[2]);
+
+    const float t21x = t2.x - t1.x;
+    const float t21y = t2.y - t1.y;
+    const float t31x = t3.x - t1.x;
+    const float t31y = t3.y - t1.y;
+
+    const float signedAreaSTx2 = t21x * t31y - t21y * t31x;
+    return fabsf(signedAreaSTx2);
+  }
+
+  void initTriangle()
+  {
+    // triangles[f].iFlag is cleared in generateInitialVerticesIndexList()
+    // which is called before this function.
+    //    printf("initTriangle %d\n", nrTriangles);
+    // evaluate first order derivatives
+    runParallel(0u, nrTriangles, [&](uint t) {
+      Triangle &triangle = triangles[t];
+      // if(t==1844 && nrFaces==923)
+      //   printf("Tri: %d Vertices: %d %d %d\n", t, triangle.vertices[0], triangle.vertices[1],
+      //   triangle.vertices[2]);
+
+      // initial values
+      const float3 v1 = getPosition(triangle.vertices[0]);
+      const float3 v2 = getPosition(triangle.vertices[1]);
+      const float3 v3 = getPosition(triangle.vertices[2]);
+      const float3 t1 = getTexCoord(triangle.vertices[0]);
+      const float3 t2 = getTexCoord(triangle.vertices[1]);
+      const float3 t3 = getTexCoord(triangle.vertices[2]);
+
+      const float t21x = t2.x - t1.x;
+      const float t21y = t2.y - t1.y;
+      const float t31x = t3.x - t1.x;
+      const float t31y = t3.y - t1.y;
+      const float3 d1 = v2 - v1, d2 = v3 - v1;
+
+      const float signedAreaSTx2 = t21x * t31y - t21y * t31x;
+      const float3 vOs = (t31y * d1) - (t21y * d2);   // eq 18
+      const float3 vOt = (-t31x * d1) + (t21x * d2);  // eq 19
+
+      triangle.orientPreserving = (signedAreaSTx2 > 0);
+#if 0      
+        if(t==1844 && nrFaces==923)
+        {
+          /*
+          const uint32_t* p1 = (const uint32_t*)&v1.x;
+          const uint32_t* p2 = (const uint32_t*)&v2.x;
+          const uint32_t* p3 = (const uint32_t*)&v3.x;
+          const uint32_t* q1 = (const uint32_t*)&t1.x;
+          const uint32_t* q2 = (const uint32_t*)&t2.x;
+          const uint32_t* q3 = (const uint32_t*)&t3.x;
+          printf("Coord: \n%08x %08x %08x\n%08x %08x %08x\n%08x %08x %08x\n",
+            p1[0], p1[1], p1[2],
+            p2[0], p2[1], p2[2],
+            p3[0], p3[1], p3[2]);
+          printf("Tex: \n%08x %08x %08x\n%08x %08x %08x\n%08x %08x %08x\n",
+            q1[0], q1[1], q1[2],
+            q2[0], q2[1], q2[2],
+            q3[0], q3[1], q3[2]);
+          printf("signedAreaSTx2 %e\n", signedAreaSTx2);
+          */
+          printf("Coords:\n%f %f %f\n%f %f %f\n%f %f %f\n",
+            v1.x,v1.y,v1.z,
+            v2.x,v2.y,v2.z,
+            v3.x,v3.y,v3.z);
+          printf("Tex:\n%f %f %f\n%f %f %f\n%f %f %f\n",
+            t1.x,t1.y,t1.z,
+            t2.x,t2.y,t2.z,
+            t3.x,t3.y,t3.z);
+        }
+#endif
+      if (not_zero(signedAreaSTx2)) {
+        const float lenOs2 = vOs.length_squared();
+        const float lenOt2 = vOt.length_squared();
+        // if(t==5858)
+        // printf("len0s2 %e, len0t2 %e\n", lenOs2, lenOt2);
+        const float fS = triangle.orientPreserving ? 1.0f : (-1.0f);
+        if (not_zero(lenOs2))
+          triangle.tangent = vOs * (fS / sqrtf(lenOs2));
+
+        // if this is a good triangle
+        if (not_zero(lenOs2) && not_zero(lenOt2))
+          triangle.groupWithAny = false;
+      }
+    });
+
+    // force otherwise healthy quads to a fixed orientation
+    runParallel(0u, nrTriangles - 1, [&](uint t) {
+      Triangle &triangleA = triangles[t], &triangleB = triangles[t + 1];
+      if (triangleA.faceIdx != triangleB.faceIdx) {
+        // this is not a quad
+        return;
+      }
+
+      // bad triangles should already have been removed by
+      // degenPrologue(), but just in case check that neither are degenerate
+      if (!(triangleA.markDegenerate || triangleB.markDegenerate)) {
+        // if this happens the quad has extremely bad mapping!!
+        if (triangleA.orientPreserving != triangleB.orientPreserving) {
+          bool chooseOrientFirstTri = false;
+          if (triangleB.groupWithAny)
+            chooseOrientFirstTri = true;
+          else if (calcTexArea(t) >= calcTexArea(t + 1))
+            chooseOrientFirstTri = true;
+
+          // force match
+          const uint t0 = chooseOrientFirstTri ? t : (t + 1);
+          const uint t1 = chooseOrientFirstTri ? (t + 1) : t;
+          triangles[t1].orientPreserving = triangles[t0].orientPreserving;
+        }
+      }
+    });
+  }
+
+  /////////////////////////////////////////////////////////////////////////////////////////////
+  /////////////////////////////////////////// Edges ///////////////////////////////////////////
+
+  struct NeighborShard {
+    struct Entry {
+      Entry(uint32_t key_, uint data_) : key(key_), data(data_) {}
+      uint key, data;
+    };
+    std::vector<Entry> entries;
+    uint id;
+
+    NeighborShard(size_t capacity)
+    {
+      entries.reserve(capacity);
+    }
+
+    void buildNeighbors(RefMikktspace<Mesh> *mikk)
+    {
+      /* Entries are added by iterating over t, so by using a stable sort,
+       * we don't have to compare based on t as well. */
+      {
+        std::vector<Entry> tempEntries(entries.size(), {0, 0});
+        radixsort(entries, tempEntries, [](const Entry &e) { return e.key; });
+      }
+
+      for (uint i = 0; i < entries.size(); i++) {
+        const Entry &a = entries[i];
+        uint tA, iA;
+        unpack_index(tA, iA, a.data);
+        RefMikktspace<Mesh>::Triangle &triA = mikk->triangles[tA];
+
+        if (triA.neighbor[iA] != UNSET_ENTRY) {
+          continue;
+        }
+
+        uint i0A = triA.vertices[iA], i1A = triA.vertices[(iA != 2) ? (iA + 1) : 0];
+        for (uint j = i + 1; j < entries.size(); j++) {
+          const Entry &b = entries[j];
+          uint tB, iB;
+          unpack_index(tB, iB, b.data);
+          RefMikktspace<Mesh>::Triangle &triB = mikk->triangles[tB];
+
+          if (b.key != a.key)
+            break;
+
+          if (triB.neighbor[iB] != UNSET_ENTRY) {
+            continue;
+          }
+
+          uint i1B = triB.vertices[iB], i0B = triB.vertices[(iB != 2) ? (iB + 1) : 0];
+          if (i0A == i0B && i1A == i1B) {
+            triA.neighbor[iA] = tB;
+            triB.neighbor[iB] = tA;
+            break;
+          }
+        }
+      }
+    }
+  };
+
+  void buildNeighbors()
+  {
+    /* In order to parallelize the processing, we divide the vertices into shards.
+     * Since only vertex pairs with the same key will be checked, we can process
+     * shards independently as long as we ensure that all vertices with the same
+     * key go into the same shard.
+     * This is done by hashing the key to get the shard index of each vertex.
+     */
+    // TODO: Two-step filling that first counts and then fills? Could be parallel then.
+    uint targetNrShards = isParallel ? uint(4 * nrThreads) : 1;
+    uint nrShards = 1, hashShift = 32;
+    while (nrShards < targetNrShards) {
+      nrShards *= 2;
+      hashShift -= 1;
+    }
+
+    /* Reserve 25% extra to account for variation due to hashing. */
+    size_t reserveSize = size_t(double(3 * nrTriangles) * 1.25 / nrShards);
+    std::vector<NeighborShard> shards(nrShards, {reserveSize});
+    for (uint t = 0; t < nrShards; t++)
+      shards[t].id = t;
+
+    for (uint t = 0; t < nrTriangles; t++) {
+      Triangle &triangle = triangles[t];
+      for (uint i = 0; i < 3; i++) {
+        const uint i0 = triangle.vertices[i];
+        const uint i1 = triangle.vertices[(i != 2) ? (i + 1) : 0];
+        const uint high = std::max(i0, i1), low = std::min(i0, i1);
+        const uint hash = hash_uint3(high, low, 0);
+        /* TODO: Reusing the hash here means less hash space inside each shard.
+         * Computing a second hash with a different seed it probably not worth it? */
+        const uint shard = isParallel ? (hash >> hashShift) : 0;
+        shards[shard].entries.emplace_back(hash, pack_index(t, i));
+      }
+    }
+
+    runParallel(0u, nrShards, [&](uint s) { shards[s].buildNeighbors(this); });
+  }
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+
+  void assignRecur(const uint t, uint groupId)
+  {
+    if (t == UNSET_ENTRY) {
+      return;
+    }
+
+    Triangle &triangle = triangles[t];
+    Group &group = groups[groupId];
+
+    // track down vertex
+    const uint vertRep = group.vertexRepresentative;
+    uint i = 3;
+    if (triangle.vertices[0] == vertRep)
+      i = 0;
+    else if (triangle.vertices[1] == vertRep)
+      i = 1;
+    else if (triangle.vertices[2] == vertRep)
+      i = 2;
+    assert(i < 3);
+
+    // early out
+    if (triangle.group[i] != UNSET_ENTRY) {
+      return;
+    }
+
+    if (triangle.groupWithAny) {
+      // first to group with a group-with-anything triangle
+      // determines its orientation.
+      // This is the only existing order dependency in the code!!
+      if (triangle.group[0] == UNSET_ENTRY && triangle.group[1] == UNSET_ENTRY &&
+          triangle.group[2] == UNSET_ENTRY)
+      {
+        triangle.orientPreserving = group.orientPreserving;
+      }
+      else {
+      }
+    }
+
+    if (triangle.orientPreserving != group.orientPreserving) {
+      return;
+    }
+    triangle.group[i] = groupId;
+    //    if(groupId==0)
+    //      printf("triangle %d -> group %d\n", t, groupId);
+
+    const uint t_L = triangle.neighbor[i];
+    const uint t_R = triangle.neighbor[i > 0 ? (i - 1) : 2];
+    if (groupId == 0) {
+      //      printf("<2>Propagating group %d to %d, %d\n", groupId, t_L, t_R);
+    }
+    assignRecur(t_L, groupId);
+    assignRecur(t_R, groupId);
+  }
+
+  void build4RuleGroups()
+  {
+    /* NOTE: This could be parallelized by grouping all [t, i] pairs into
+     * shards by hash(triangles[t].vertices[i]). This way, each shard can be processed
+     * independently and in parallel.
+     * However, the `groupWithAny` logic needs special handling (e.g. lock a mutex when
+     * encountering a `groupWithAny` triangle, then sort it out, then unlock and proceed). */
+    for (uint t = 0; t < nrTriangles; t++) {
+      Triangle &triangle = triangles[t];
+      for (uint i = 0; i < 3; i++) {
+        // if not assigned to a group
+        if (triangle.groupWithAny || triangle.group[i] != UNSET_ENTRY) {
+          // if(t==230 || t==231)
+          //   printf("<ref> t %d i %d: skipping (groupWithAny %d, group %d, neighbors %d %d
+          //   %d)\n",
+          //     t, i, triangle.groupWithAny, triangle.group[i], triangle.neighbor[0],
+          //     triangle.neighbor[1], triangle.neighbor[2]);
+          continue;
+        }
+
+        const uint newGroupId = uint(groups.size());
+        triangle.group[i] = newGroupId;
+        groups.emplace_back(triangle.vertices[i], bool(triangle.orientPreserving));
+
+        if (newGroupId == 1151 && nrFaces == 923) {
+          // printf("Initializing group %d: tri %d vert %d\n", newGroupId, t, i);
+        }
+
+        const uint t_L = triangle.neighbor[i];
+        const uint t_R = triangle.neighbor[i > 0 ? (i - 1) : 2];
+        assignRecur(t_L, newGroupId);
+        assignRecur(t_R, newGroupId);
+      }
+    }
+    //    printf("%d triangles -> %d groups\n", nrTriangles, groups.size());
+  }
+
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+  ///////////////////////////////////////////////////////////////////////////////////////////////////
+
+  template<bool atomic> void accumulateTSpaces(uint t)
+  {
+    const Triangle &triangle = triangles[t];
+    // only valid triangles get to add their contribution
+    if (triangle.groupWithAny) {
+      return;
+    }
+
+    /* Todo: Vectorize?
+     * Also: Could add special case for flat shading, when all normals are equal half of the fCos
+     * projections and two of the three tangent projections are unnecessary. */
+    std::array<float3, 3> n, p;
+    for (uint i = 0; i < 3; i++) {
+      n[i] = getNormal(triangle.vertices[i]);
+      p[i] = getPosition(triangle.vertices[i]);
+    }
+
+    std::array<float, 3> fCos = {dot(project(n[0], p[1] - p[0]), project(n[0], p[2] - p[0])),
+                                 dot(project(n[1], p[2] - p[1]), project(n[1], p[0] - p[1])),
+                                 dot(project(n[2], p[0] - p[2]), project(n[2], p[1] - p[2]))};
+
+    for (uint i = 0; i < 3; i++) {
+      uint groupId = triangle.group[i];
+      if (groupId != UNSET_ENTRY) {
+        float3 tangent = project(n[i], triangle.tangent) *
+                         fast_acosf(std::clamp(fCos[i], -1.0f, 1.0f));
+        if constexpr (atomic) {
+          groups[groupId].accumulateTSpaceAtomic(tangent);
+        }
+        else {
+          groups[groupId].accumulateTSpace(tangent);
+        }
+      }
+    }
+  }
+
+  void generateTSpaces()
+  {
+    if (isParallel) {
+      runParallel(0u, nrTriangles, [&](uint t) { accumulateTSpaces<true>(t); });
+    }
+    else {
+      for (uint t = 0; t < nrTriangles; t++) {
+        accumulateTSpaces<false>(t);
+      }
+    }
+
+    /* TODO: Worth parallelizing? Probably not. */
+    for (Group &group : groups) {
+      group.normalizeTSpace();
+    }
+
+    tSpaces.resize(nrTSpaces);
+
+    for (uint t = 0; t < nrTriangles; t++) {
+      Triangle &triangle = triangles[t];
+      for (uint i = 0; i < 3; i++) {
+        const uint offset = triangle.tSpaceIdx;
+        const uint faceVertex = triangle.faceVertex[i];
+
+        uint groupId = triangle.group[i];
+        if (groupId == UNSET_ENTRY) {
+          continue;
+        }
+        const Group group = groups[groupId];
+        assert(triangle.orientPreserving == group.orientPreserving);
+        tSpaces[offset + faceVertex].accumulateGroup(group);
+      }
+    }
+  }
+};
+
+}  // namespace mikk

source/blender/blenkernel/BKE_mesh_tangent.hh

@@ -64,7 +64,11 @@ void BKE_mesh_calc_loop_tangent_ex(const float (*vert_positions)[3],
                                    /* result */
                                    CustomData *loopdata_out,
                                    uint loopdata_out_len,
-                                   short *tangent_mask_curr_p);
+                                   short *tangent_mask_curr_p,
+                                   int len_vert_positions,
+                                   int len_corner_verts,
+                                   int len_vert_normals,
+                                   int len_face_normals);
 
 void BKE_mesh_calc_loop_tangents(Mesh *mesh_eval,
                                  bool calc_active_tangent,

source/blender/blenkernel/intern/editmesh_tangent.cc

@@ -23,6 +23,7 @@
 
 /* interface */
 #include "mikktspace.hh"
+#include "mikktspace_ref.hh"
 
 /* -------------------------------------------------------------------- */
 /** \name Tangent Space Calculation
@@ -147,7 +148,8 @@ static void emDM_calc_loop_tangents_thread(TaskPool *__restrict /*pool*/, void *
 {
   SGLSLEditMeshToTangent *mesh_data = static_cast<SGLSLEditMeshToTangent *>(taskdata);
 
-  mikk::Mikktspace<SGLSLEditMeshToTangent> mikk(*mesh_data);
+  mikk::RefMikktspace<SGLSLEditMeshToTangent> mikk(*mesh_data);
+  printf("EditMesh::tangent\n");
   mikk.genTangSpace();
 }
 

source/blender/blenkernel/intern/mesh_tangent.cc

@@ -25,6 +25,7 @@
 #include "BKE_report.hh"
 
 #include "mikktspace.hh"
+#include "mikktspace_ref.hh"
 
 #include "BLI_strict_flags.h" /* Keep last. */
 
@@ -75,6 +76,8 @@ struct BKEMeshToTangent {
   const float (*corner_normals)[3];  /* loops' normals */
   float (*tangents)[4];              /* output tangents */
   int num_faces;                     /* number of polygons */
+  size_t tangent_len;
+  int num_face_as_quad_map;
 };
 
 void BKE_mesh_calc_loop_tangent_single_ex(const float (*vert_positions)[3],
@@ -83,7 +86,7 @@ void BKE_mesh_calc_loop_tangent_single_ex(const float (*vert_positions)[3],
                                           float (*r_looptangent)[4],
                                           const float (*corner_normals)[3],
                                           const float (*loop_uvs)[2],
-                                          const int /*numLoops*/,
+                                          const int numLoops,
                                           const blender::OffsetIndices<int> faces,
                                           ReportList *reports)
 {
@@ -96,8 +99,9 @@ void BKE_mesh_calc_loop_tangent_single_ex(const float (*vert_positions)[3],
   mesh_to_tangent.corner_normals = corner_normals;
   mesh_to_tangent.tangents = r_looptangent;
   mesh_to_tangent.num_faces = int(faces.size());
+  mesh_to_tangent.tangent_len = (size_t)numLoops * 4;
 
-  mikk::Mikktspace<BKEMeshToTangent> mikk(mesh_to_tangent);
+  mikk::RefMikktspace<BKEMeshToTangent> mikk(mesh_to_tangent);
 
   /* First check we do have a tris/quads only mesh. */
   for (const int64_t i : faces.index_range()) {
@@ -200,6 +204,68 @@ struct SGLSLMeshToTangent {
     return uint(tri[int(vert_num)]);
   }
 
+  mikk::float3 GetPositionDirect(const uint loop_index)
+  {
+    return mikk::float3(positions[corner_verts[loop_index]]);
+  }
+
+  inline mikk::float3 GetTexCoordDirect(const uint loop_index)
+  {
+    if (mloopuv != nullptr) {
+      const float2 &uv = mloopuv[loop_index];
+      return mikk::float3(uv[0], uv[1], 1.0f);
+    }
+    const float *l_orco = orco[corner_verts[loop_index]];
+    float u, v;
+    map_to_sphere(&u, &v, l_orco[0], l_orco[1], l_orco[2]);
+    return mikk::float3(u, v, 1.0f);
+  }
+
+  inline mikk::float3 GetNormalDirect(const int face_index, const uint loop_index)
+  {
+    blender::int3 tri;
+    if (precomputedLoopNormals) {
+      return mikk::float3(precomputedLoopNormals[loop_index]);
+    }
+    if (!sharp_faces.is_empty() && sharp_faces[face_index]) { /* flat */
+      if (precomputedFaceNormals) {
+        return mikk::float3(precomputedFaceNormals[face_index]);
+      }
+#ifdef USE_TRI_DETECT_QUADS
+      const blender::IndexRange face = faces[face_index];
+      float normal[3];
+      if (face.size() == 4) {
+        normal_quad_v3(normal,
+                       positions[corner_verts[face[0]]],
+                       positions[corner_verts[face[1]]],
+                       positions[corner_verts[face[2]]],
+                       positions[corner_verts[face[3]]]);
+      }
+      else
+#endif
+      {
+        normal_tri_v3(normal,
+                      positions[corner_verts[tri[0]]],
+                      positions[corner_verts[tri[1]]],
+                      positions[corner_verts[tri[2]]]);
+      }
+      return mikk::float3(normal);
+    }
+    return mikk::float3(vert_normals[corner_verts[loop_index]]);
+  }
+
+  uint GetStoreIndex(const uint face_num, const uint vert_num)
+  {
+    blender::int3 tri;
+    int face_index;
+    return GetLoop(face_num, vert_num, tri, face_index);
+  }
+
+  void SetTangentSpaceDirect(const uint loop_index, mikk::float3 T, bool orientation)
+  {
+    copy_v4_fl4(tangent[loop_index], T.x, T.y, T.z, orientation ? 1.0f : -1.0f);
+  }
+
   mikk::float3 GetPosition(const uint face_num, const uint vert_num)
   {
     blender::int3 tri;
@@ -280,6 +346,12 @@ struct SGLSLMeshToTangent {
   float (*tangent)[4]; /* destination */
   blender::Span<bool> sharp_faces;
   int numTessFaces;
+  size_t tangent_len;
+
+  int len_corner_verts;
+  int len_positions;
+  int len_face_normals;
+  int len_vert_normals;
 
 #ifdef USE_TRI_DETECT_QUADS
   /* map from 'fake' face index to corner_tris,
@@ -292,7 +364,7 @@ struct SGLSLMeshToTangent {
 static void DM_calc_loop_tangents_thread(TaskPool *__restrict /*pool*/, void *taskdata)
 {
   SGLSLMeshToTangent *mesh_data = static_cast<SGLSLMeshToTangent *>(taskdata);
-
+  // mikk::RefMikktspace<SGLSLMeshToTangent> mikk(*mesh_data);
   mikk::Mikktspace<SGLSLMeshToTangent> mikk(*mesh_data);
   mikk.genTangSpace();
 }
@@ -392,7 +464,6 @@ void BKE_mesh_calc_loop_tangent_ex(const float (*vert_positions)[3],
                                    const int *corner_tri_faces,
                                    const uint corner_tris_len,
                                    const blender::Span<bool> sharp_faces,
-
                                    const CustomData *loopdata,
                                    bool calc_active_tangent,
                                    const char (*tangent_names)[MAX_CUSTOMDATA_LAYER_NAME],
@@ -404,7 +475,12 @@ void BKE_mesh_calc_loop_tangent_ex(const float (*vert_positions)[3],
                                    /* result */
                                    CustomData *loopdata_out,
                                    const uint loopdata_out_len,
-                                   short *tangent_mask_curr_p)
+                                   short *tangent_mask_curr_p,
+
+                                   int len_vert_positions,
+                                   int len_corner_verts,
+                                   [[maybe_unused]] int len_vert_normals,
+                                   int len_face_normals)
 {
   int act_uv_n = -1;
   int ren_uv_n = -1;
@@ -508,6 +584,11 @@ void BKE_mesh_calc_loop_tangent_ex(const float (*vert_positions)[3],
         mesh2tangent->mloopuv = static_cast<const float2 *>(CustomData_get_layer_named(
             loopdata, CD_PROP_FLOAT2, loopdata_out->layers[index].name));
 
+        mesh2tangent->len_corner_verts = len_corner_verts;
+        mesh2tangent->len_positions = len_vert_positions;
+        mesh2tangent->len_face_normals = len_face_normals;
+        mesh2tangent->tangent_len = loopdata_out_len;
+
         /* Fill the resulting tangent_mask */
         if (!mesh2tangent->mloopuv) {
           mesh2tangent->orco = vert_orco;
@@ -591,9 +672,10 @@ void BKE_mesh_calc_loop_tangents(Mesh *mesh_eval,
       /* may be nullptr */
       static_cast<const float(*)[3]>(CustomData_get_layer(&mesh_eval->vert_data, CD_ORCO)),
       /* result */
-      &mesh_eval->corner_data,
-      uint(mesh_eval->corners_num),
-      &tangent_mask);
+      (int)me_eval->vert_positions().size(),
+      (int)me_eval->corner_verts().size(),
+      (int)me_eval->vert_normals().size(),
+      (int)me_eval->face_normals().size());
 }
 
 /** \} */

source/blender/draw/intern/mesh_extractors/extract_mesh_vbo_tan.cc

@@ -137,7 +137,11 @@ static void extract_tan_init_common(const MeshRenderData &mr,
                                     orco,
                                     r_loop_data,
                                     mr.corner_verts.size(),
-                                    &tangent_mask);
+                                    &tangent_mask,
+                                    (uint)mr.vert_positions.size(),
+                                    (uint)mr.corner_verts.size(),
+                                    (uint)mr.vert_normals.size(),
+                                    (uint)mr.face_normals.size());
     }
   }
 

source/blender/render/intern/multires_bake.cc

@@ -539,7 +539,11 @@ static void do_multires_bake(MultiresBakeRender *bkr,
           /* result */
           &dm->loopData,
           dm->getNumLoops(dm),
-          &dm->tangent_mask);
+          &dm->tangent_mask,
+          (uint)positions.size(),
+          (uint)corner_verts.size(),
+          (uint)vert_normals.size(),
+          (uint)face_normals.size());
     }
 
     pvtangent = static_cast<float *>(DM_get_loop_data_layer(dm, CD_TANGENT));