summaryrefslogtreecommitdiff
path: root/thirdparty/meshoptimizer
diff options
context:
space:
mode:
Diffstat (limited to 'thirdparty/meshoptimizer')
-rw-r--r--thirdparty/meshoptimizer/LICENSE.md21
-rw-r--r--thirdparty/meshoptimizer/allocator.cpp8
-rw-r--r--thirdparty/meshoptimizer/clusterizer.cpp351
-rw-r--r--thirdparty/meshoptimizer/indexcodec.cpp752
-rw-r--r--thirdparty/meshoptimizer/indexgenerator.cpp347
-rw-r--r--thirdparty/meshoptimizer/meshoptimizer.h951
-rw-r--r--thirdparty/meshoptimizer/overdrawanalyzer.cpp230
-rw-r--r--thirdparty/meshoptimizer/overdrawoptimizer.cpp333
-rw-r--r--thirdparty/meshoptimizer/patches/simplifier_get_resulting_error.patch96
-rw-r--r--thirdparty/meshoptimizer/simplifier.cpp1562
-rw-r--r--thirdparty/meshoptimizer/spatialorder.cpp194
-rw-r--r--thirdparty/meshoptimizer/stripifier.cpp295
-rw-r--r--thirdparty/meshoptimizer/vcacheanalyzer.cpp73
-rw-r--r--thirdparty/meshoptimizer/vcacheoptimizer.cpp473
-rw-r--r--thirdparty/meshoptimizer/vertexcodec.cpp1265
-rw-r--r--thirdparty/meshoptimizer/vertexfilter.cpp825
-rw-r--r--thirdparty/meshoptimizer/vfetchanalyzer.cpp58
-rw-r--r--thirdparty/meshoptimizer/vfetchoptimizer.cpp74
18 files changed, 7908 insertions, 0 deletions
diff --git a/thirdparty/meshoptimizer/LICENSE.md b/thirdparty/meshoptimizer/LICENSE.md
new file mode 100644
index 0000000000..4fcd766d22
--- /dev/null
+++ b/thirdparty/meshoptimizer/LICENSE.md
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2016-2020 Arseny Kapoulkine
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/thirdparty/meshoptimizer/allocator.cpp b/thirdparty/meshoptimizer/allocator.cpp
new file mode 100644
index 0000000000..da7cc540b2
--- /dev/null
+++ b/thirdparty/meshoptimizer/allocator.cpp
@@ -0,0 +1,8 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+void meshopt_setAllocator(void* (*allocate)(size_t), void (*deallocate)(void*))
+{
+ meshopt_Allocator::Storage::allocate = allocate;
+ meshopt_Allocator::Storage::deallocate = deallocate;
+}
diff --git a/thirdparty/meshoptimizer/clusterizer.cpp b/thirdparty/meshoptimizer/clusterizer.cpp
new file mode 100644
index 0000000000..f7d88c5136
--- /dev/null
+++ b/thirdparty/meshoptimizer/clusterizer.cpp
@@ -0,0 +1,351 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <math.h>
+#include <string.h>
+
+// This work is based on:
+// Graham Wihlidal. Optimizing the Graphics Pipeline with Compute. 2016
+// Matthaeus Chajdas. GeometryFX 1.2 - Cluster Culling. 2016
+// Jack Ritter. An Efficient Bounding Sphere. 1990
+namespace meshopt
+{
+
+static void computeBoundingSphere(float result[4], const float points[][3], size_t count)
+{
+ assert(count > 0);
+
+ // find extremum points along all 3 axes; for each axis we get a pair of points with min/max coordinates
+ size_t pmin[3] = {0, 0, 0};
+ size_t pmax[3] = {0, 0, 0};
+
+ for (size_t i = 0; i < count; ++i)
+ {
+ const float* p = points[i];
+
+ for (int axis = 0; axis < 3; ++axis)
+ {
+ pmin[axis] = (p[axis] < points[pmin[axis]][axis]) ? i : pmin[axis];
+ pmax[axis] = (p[axis] > points[pmax[axis]][axis]) ? i : pmax[axis];
+ }
+ }
+
+ // find the pair of points with largest distance
+ float paxisd2 = 0;
+ int paxis = 0;
+
+ for (int axis = 0; axis < 3; ++axis)
+ {
+ const float* p1 = points[pmin[axis]];
+ const float* p2 = points[pmax[axis]];
+
+ float d2 = (p2[0] - p1[0]) * (p2[0] - p1[0]) + (p2[1] - p1[1]) * (p2[1] - p1[1]) + (p2[2] - p1[2]) * (p2[2] - p1[2]);
+
+ if (d2 > paxisd2)
+ {
+ paxisd2 = d2;
+ paxis = axis;
+ }
+ }
+
+ // use the longest segment as the initial sphere diameter
+ const float* p1 = points[pmin[paxis]];
+ const float* p2 = points[pmax[paxis]];
+
+ float center[3] = {(p1[0] + p2[0]) / 2, (p1[1] + p2[1]) / 2, (p1[2] + p2[2]) / 2};
+ float radius = sqrtf(paxisd2) / 2;
+
+ // iteratively adjust the sphere up until all points fit
+ for (size_t i = 0; i < count; ++i)
+ {
+ const float* p = points[i];
+ float d2 = (p[0] - center[0]) * (p[0] - center[0]) + (p[1] - center[1]) * (p[1] - center[1]) + (p[2] - center[2]) * (p[2] - center[2]);
+
+ if (d2 > radius * radius)
+ {
+ float d = sqrtf(d2);
+ assert(d > 0);
+
+ float k = 0.5f + (radius / d) / 2;
+
+ center[0] = center[0] * k + p[0] * (1 - k);
+ center[1] = center[1] * k + p[1] * (1 - k);
+ center[2] = center[2] * k + p[2] * (1 - k);
+ radius = (radius + d) / 2;
+ }
+ }
+
+ result[0] = center[0];
+ result[1] = center[1];
+ result[2] = center[2];
+ result[3] = radius;
+}
+
+} // namespace meshopt
+
+size_t meshopt_buildMeshletsBound(size_t index_count, size_t max_vertices, size_t max_triangles)
+{
+ assert(index_count % 3 == 0);
+ assert(max_vertices >= 3);
+ assert(max_triangles >= 1);
+
+ // meshlet construction is limited by max vertices and max triangles per meshlet
+ // the worst case is that the input is an unindexed stream since this equally stresses both limits
+ // note that we assume that in the worst case, we leave 2 vertices unpacked in each meshlet - if we have space for 3 we can pack any triangle
+ size_t max_vertices_conservative = max_vertices - 2;
+ size_t meshlet_limit_vertices = (index_count + max_vertices_conservative - 1) / max_vertices_conservative;
+ size_t meshlet_limit_triangles = (index_count / 3 + max_triangles - 1) / max_triangles;
+
+ return meshlet_limit_vertices > meshlet_limit_triangles ? meshlet_limit_vertices : meshlet_limit_triangles;
+}
+
+size_t meshopt_buildMeshlets(meshopt_Meshlet* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles)
+{
+ assert(index_count % 3 == 0);
+ assert(max_vertices >= 3);
+ assert(max_triangles >= 1);
+
+ meshopt_Allocator allocator;
+
+ meshopt_Meshlet meshlet;
+ memset(&meshlet, 0, sizeof(meshlet));
+
+ assert(max_vertices <= sizeof(meshlet.vertices) / sizeof(meshlet.vertices[0]));
+ assert(max_triangles <= sizeof(meshlet.indices) / 3);
+
+ // index of the vertex in the meshlet, 0xff if the vertex isn't used
+ unsigned char* used = allocator.allocate<unsigned char>(vertex_count);
+ memset(used, -1, vertex_count);
+
+ size_t offset = 0;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int a = indices[i + 0], b = indices[i + 1], c = indices[i + 2];
+ assert(a < vertex_count && b < vertex_count && c < vertex_count);
+
+ unsigned char& av = used[a];
+ unsigned char& bv = used[b];
+ unsigned char& cv = used[c];
+
+ unsigned int used_extra = (av == 0xff) + (bv == 0xff) + (cv == 0xff);
+
+ if (meshlet.vertex_count + used_extra > max_vertices || meshlet.triangle_count >= max_triangles)
+ {
+ destination[offset++] = meshlet;
+
+ for (size_t j = 0; j < meshlet.vertex_count; ++j)
+ used[meshlet.vertices[j]] = 0xff;
+
+ memset(&meshlet, 0, sizeof(meshlet));
+ }
+
+ if (av == 0xff)
+ {
+ av = meshlet.vertex_count;
+ meshlet.vertices[meshlet.vertex_count++] = a;
+ }
+
+ if (bv == 0xff)
+ {
+ bv = meshlet.vertex_count;
+ meshlet.vertices[meshlet.vertex_count++] = b;
+ }
+
+ if (cv == 0xff)
+ {
+ cv = meshlet.vertex_count;
+ meshlet.vertices[meshlet.vertex_count++] = c;
+ }
+
+ meshlet.indices[meshlet.triangle_count][0] = av;
+ meshlet.indices[meshlet.triangle_count][1] = bv;
+ meshlet.indices[meshlet.triangle_count][2] = cv;
+ meshlet.triangle_count++;
+ }
+
+ if (meshlet.triangle_count)
+ destination[offset++] = meshlet;
+
+ assert(offset <= meshopt_buildMeshletsBound(index_count, max_vertices, max_triangles));
+
+ return offset;
+}
+
+meshopt_Bounds meshopt_computeClusterBounds(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+
+ assert(index_count / 3 <= 256);
+
+ (void)vertex_count;
+
+ size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
+
+ // compute triangle normals and gather triangle corners
+ float normals[256][3];
+ float corners[256][3][3];
+ size_t triangles = 0;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int a = indices[i + 0], b = indices[i + 1], c = indices[i + 2];
+ assert(a < vertex_count && b < vertex_count && c < vertex_count);
+
+ const float* p0 = vertex_positions + vertex_stride_float * a;
+ const float* p1 = vertex_positions + vertex_stride_float * b;
+ const float* p2 = vertex_positions + vertex_stride_float * c;
+
+ float p10[3] = {p1[0] - p0[0], p1[1] - p0[1], p1[2] - p0[2]};
+ float p20[3] = {p2[0] - p0[0], p2[1] - p0[1], p2[2] - p0[2]};
+
+ float normalx = p10[1] * p20[2] - p10[2] * p20[1];
+ float normaly = p10[2] * p20[0] - p10[0] * p20[2];
+ float normalz = p10[0] * p20[1] - p10[1] * p20[0];
+
+ float area = sqrtf(normalx * normalx + normaly * normaly + normalz * normalz);
+
+ // no need to include degenerate triangles - they will be invisible anyway
+ if (area == 0.f)
+ continue;
+
+ // record triangle normals & corners for future use; normal and corner 0 define a plane equation
+ normals[triangles][0] = normalx / area;
+ normals[triangles][1] = normaly / area;
+ normals[triangles][2] = normalz / area;
+ memcpy(corners[triangles][0], p0, 3 * sizeof(float));
+ memcpy(corners[triangles][1], p1, 3 * sizeof(float));
+ memcpy(corners[triangles][2], p2, 3 * sizeof(float));
+ triangles++;
+ }
+
+ meshopt_Bounds bounds = {};
+
+ // degenerate cluster, no valid triangles => trivial reject (cone data is 0)
+ if (triangles == 0)
+ return bounds;
+
+ // compute cluster bounding sphere; we'll use the center to determine normal cone apex as well
+ float psphere[4] = {};
+ computeBoundingSphere(psphere, corners[0], triangles * 3);
+
+ float center[3] = {psphere[0], psphere[1], psphere[2]};
+
+ // treating triangle normals as points, find the bounding sphere - the sphere center determines the optimal cone axis
+ float nsphere[4] = {};
+ computeBoundingSphere(nsphere, normals, triangles);
+
+ float axis[3] = {nsphere[0], nsphere[1], nsphere[2]};
+ float axislength = sqrtf(axis[0] * axis[0] + axis[1] * axis[1] + axis[2] * axis[2]);
+ float invaxislength = axislength == 0.f ? 0.f : 1.f / axislength;
+
+ axis[0] *= invaxislength;
+ axis[1] *= invaxislength;
+ axis[2] *= invaxislength;
+
+ // compute a tight cone around all normals, mindp = cos(angle/2)
+ float mindp = 1.f;
+
+ for (size_t i = 0; i < triangles; ++i)
+ {
+ float dp = normals[i][0] * axis[0] + normals[i][1] * axis[1] + normals[i][2] * axis[2];
+
+ mindp = (dp < mindp) ? dp : mindp;
+ }
+
+ // fill bounding sphere info; note that below we can return bounds without cone information for degenerate cones
+ bounds.center[0] = center[0];
+ bounds.center[1] = center[1];
+ bounds.center[2] = center[2];
+ bounds.radius = psphere[3];
+
+ // degenerate cluster, normal cone is larger than a hemisphere => trivial accept
+ // note that if mindp is positive but close to 0, the triangle intersection code below gets less stable
+ // we arbitrarily decide that if a normal cone is ~168 degrees wide or more, the cone isn't useful
+ if (mindp <= 0.1f)
+ {
+ bounds.cone_cutoff = 1;
+ bounds.cone_cutoff_s8 = 127;
+ return bounds;
+ }
+
+ float maxt = 0;
+
+ // we need to find the point on center-t*axis ray that lies in negative half-space of all triangles
+ for (size_t i = 0; i < triangles; ++i)
+ {
+ // dot(center-t*axis-corner, trinormal) = 0
+ // dot(center-corner, trinormal) - t * dot(axis, trinormal) = 0
+ float cx = center[0] - corners[i][0][0];
+ float cy = center[1] - corners[i][0][1];
+ float cz = center[2] - corners[i][0][2];
+
+ float dc = cx * normals[i][0] + cy * normals[i][1] + cz * normals[i][2];
+ float dn = axis[0] * normals[i][0] + axis[1] * normals[i][1] + axis[2] * normals[i][2];
+
+ // dn should be larger than mindp cutoff above
+ assert(dn > 0.f);
+ float t = dc / dn;
+
+ maxt = (t > maxt) ? t : maxt;
+ }
+
+ // cone apex should be in the negative half-space of all cluster triangles by construction
+ bounds.cone_apex[0] = center[0] - axis[0] * maxt;
+ bounds.cone_apex[1] = center[1] - axis[1] * maxt;
+ bounds.cone_apex[2] = center[2] - axis[2] * maxt;
+
+ // note: this axis is the axis of the normal cone, but our test for perspective camera effectively negates the axis
+ bounds.cone_axis[0] = axis[0];
+ bounds.cone_axis[1] = axis[1];
+ bounds.cone_axis[2] = axis[2];
+
+ // cos(a) for normal cone is mindp; we need to add 90 degrees on both sides and invert the cone
+ // which gives us -cos(a+90) = -(-sin(a)) = sin(a) = sqrt(1 - cos^2(a))
+ bounds.cone_cutoff = sqrtf(1 - mindp * mindp);
+
+ // quantize axis & cutoff to 8-bit SNORM format
+ bounds.cone_axis_s8[0] = (signed char)(meshopt_quantizeSnorm(bounds.cone_axis[0], 8));
+ bounds.cone_axis_s8[1] = (signed char)(meshopt_quantizeSnorm(bounds.cone_axis[1], 8));
+ bounds.cone_axis_s8[2] = (signed char)(meshopt_quantizeSnorm(bounds.cone_axis[2], 8));
+
+ // for the 8-bit test to be conservative, we need to adjust the cutoff by measuring the max. error
+ float cone_axis_s8_e0 = fabsf(bounds.cone_axis_s8[0] / 127.f - bounds.cone_axis[0]);
+ float cone_axis_s8_e1 = fabsf(bounds.cone_axis_s8[1] / 127.f - bounds.cone_axis[1]);
+ float cone_axis_s8_e2 = fabsf(bounds.cone_axis_s8[2] / 127.f - bounds.cone_axis[2]);
+
+ // note that we need to round this up instead of rounding to nearest, hence +1
+ int cone_cutoff_s8 = int(127 * (bounds.cone_cutoff + cone_axis_s8_e0 + cone_axis_s8_e1 + cone_axis_s8_e2) + 1);
+
+ bounds.cone_cutoff_s8 = (cone_cutoff_s8 > 127) ? 127 : (signed char)(cone_cutoff_s8);
+
+ return bounds;
+}
+
+meshopt_Bounds meshopt_computeMeshletBounds(const meshopt_Meshlet* meshlet, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+
+ unsigned int indices[sizeof(meshlet->indices) / sizeof(meshlet->indices[0][0])];
+
+ for (size_t i = 0; i < meshlet->triangle_count; ++i)
+ {
+ unsigned int a = meshlet->vertices[meshlet->indices[i][0]];
+ unsigned int b = meshlet->vertices[meshlet->indices[i][1]];
+ unsigned int c = meshlet->vertices[meshlet->indices[i][2]];
+
+ assert(a < vertex_count && b < vertex_count && c < vertex_count);
+
+ indices[i * 3 + 0] = a;
+ indices[i * 3 + 1] = b;
+ indices[i * 3 + 2] = c;
+ }
+
+ return meshopt_computeClusterBounds(indices, meshlet->triangle_count * 3, vertex_positions, vertex_count, vertex_positions_stride);
+}
diff --git a/thirdparty/meshoptimizer/indexcodec.cpp b/thirdparty/meshoptimizer/indexcodec.cpp
new file mode 100644
index 0000000000..eeb541e5be
--- /dev/null
+++ b/thirdparty/meshoptimizer/indexcodec.cpp
@@ -0,0 +1,752 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <string.h>
+
+#ifndef TRACE
+#define TRACE 0
+#endif
+
+#if TRACE
+#include <stdio.h>
+#endif
+
+// This work is based on:
+// Fabian Giesen. Simple lossless index buffer compression & follow-up. 2013
+// Conor Stokes. Vertex Cache Optimised Index Buffer Compression. 2014
+namespace meshopt
+{
+
+const unsigned char kIndexHeader = 0xe0;
+const unsigned char kSequenceHeader = 0xd0;
+
+static int gEncodeIndexVersion = 0;
+
+typedef unsigned int VertexFifo[16];
+typedef unsigned int EdgeFifo[16][2];
+
+static const unsigned int kTriangleIndexOrder[3][3] = {
+ {0, 1, 2},
+ {1, 2, 0},
+ {2, 0, 1},
+};
+
+static const unsigned char kCodeAuxEncodingTable[16] = {
+ 0x00, 0x76, 0x87, 0x56, 0x67, 0x78, 0xa9, 0x86, 0x65, 0x89, 0x68, 0x98, 0x01, 0x69,
+ 0, 0, // last two entries aren't used for encoding
+};
+
+static int rotateTriangle(unsigned int a, unsigned int b, unsigned int c, unsigned int next)
+{
+ (void)a;
+
+ return (b == next) ? 1 : (c == next) ? 2 : 0;
+}
+
+static int getEdgeFifo(EdgeFifo fifo, unsigned int a, unsigned int b, unsigned int c, size_t offset)
+{
+ for (int i = 0; i < 16; ++i)
+ {
+ size_t index = (offset - 1 - i) & 15;
+
+ unsigned int e0 = fifo[index][0];
+ unsigned int e1 = fifo[index][1];
+
+ if (e0 == a && e1 == b)
+ return (i << 2) | 0;
+ if (e0 == b && e1 == c)
+ return (i << 2) | 1;
+ if (e0 == c && e1 == a)
+ return (i << 2) | 2;
+ }
+
+ return -1;
+}
+
+static void pushEdgeFifo(EdgeFifo fifo, unsigned int a, unsigned int b, size_t& offset)
+{
+ fifo[offset][0] = a;
+ fifo[offset][1] = b;
+ offset = (offset + 1) & 15;
+}
+
+static int getVertexFifo(VertexFifo fifo, unsigned int v, size_t offset)
+{
+ for (int i = 0; i < 16; ++i)
+ {
+ size_t index = (offset - 1 - i) & 15;
+
+ if (fifo[index] == v)
+ return i;
+ }
+
+ return -1;
+}
+
+static void pushVertexFifo(VertexFifo fifo, unsigned int v, size_t& offset, int cond = 1)
+{
+ fifo[offset] = v;
+ offset = (offset + cond) & 15;
+}
+
+static void encodeVByte(unsigned char*& data, unsigned int v)
+{
+ // encode 32-bit value in up to 5 7-bit groups
+ do
+ {
+ *data++ = (v & 127) | (v > 127 ? 128 : 0);
+ v >>= 7;
+ } while (v);
+}
+
+static unsigned int decodeVByte(const unsigned char*& data)
+{
+ unsigned char lead = *data++;
+
+ // fast path: single byte
+ if (lead < 128)
+ return lead;
+
+ // slow path: up to 4 extra bytes
+ // note that this loop always terminates, which is important for malformed data
+ unsigned int result = lead & 127;
+ unsigned int shift = 7;
+
+ for (int i = 0; i < 4; ++i)
+ {
+ unsigned char group = *data++;
+ result |= (group & 127) << shift;
+ shift += 7;
+
+ if (group < 128)
+ break;
+ }
+
+ return result;
+}
+
+static void encodeIndex(unsigned char*& data, unsigned int index, unsigned int last)
+{
+ unsigned int d = index - last;
+ unsigned int v = (d << 1) ^ (int(d) >> 31);
+
+ encodeVByte(data, v);
+}
+
+static unsigned int decodeIndex(const unsigned char*& data, unsigned int last)
+{
+ unsigned int v = decodeVByte(data);
+ unsigned int d = (v >> 1) ^ -int(v & 1);
+
+ return last + d;
+}
+
+static int getCodeAuxIndex(unsigned char v, const unsigned char* table)
+{
+ for (int i = 0; i < 16; ++i)
+ if (table[i] == v)
+ return i;
+
+ return -1;
+}
+
+static void writeTriangle(void* destination, size_t offset, size_t index_size, unsigned int a, unsigned int b, unsigned int c)
+{
+ if (index_size == 2)
+ {
+ static_cast<unsigned short*>(destination)[offset + 0] = (unsigned short)(a);
+ static_cast<unsigned short*>(destination)[offset + 1] = (unsigned short)(b);
+ static_cast<unsigned short*>(destination)[offset + 2] = (unsigned short)(c);
+ }
+ else
+ {
+ static_cast<unsigned int*>(destination)[offset + 0] = a;
+ static_cast<unsigned int*>(destination)[offset + 1] = b;
+ static_cast<unsigned int*>(destination)[offset + 2] = c;
+ }
+}
+
+#if TRACE
+static size_t sortTop16(unsigned char dest[16], size_t stats[256])
+{
+ size_t destsize = 0;
+
+ for (size_t i = 0; i < 256; ++i)
+ {
+ size_t j = 0;
+ for (; j < destsize; ++j)
+ {
+ if (stats[i] >= stats[dest[j]])
+ {
+ if (destsize < 16)
+ destsize++;
+
+ memmove(&dest[j + 1], &dest[j], destsize - 1 - j);
+ dest[j] = (unsigned char)i;
+ break;
+ }
+ }
+
+ if (j == destsize && destsize < 16)
+ {
+ dest[destsize] = (unsigned char)i;
+ destsize++;
+ }
+ }
+
+ return destsize;
+}
+#endif
+
+} // namespace meshopt
+
+size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+
+#if TRACE
+ size_t codestats[256] = {};
+ size_t codeauxstats[256] = {};
+#endif
+
+ // the minimum valid encoding is header, 1 byte per triangle and a 16-byte codeaux table
+ if (buffer_size < 1 + index_count / 3 + 16)
+ return 0;
+
+ int version = gEncodeIndexVersion;
+
+ buffer[0] = (unsigned char)(kIndexHeader | version);
+
+ EdgeFifo edgefifo;
+ memset(edgefifo, -1, sizeof(edgefifo));
+
+ VertexFifo vertexfifo;
+ memset(vertexfifo, -1, sizeof(vertexfifo));
+
+ size_t edgefifooffset = 0;
+ size_t vertexfifooffset = 0;
+
+ unsigned int next = 0;
+ unsigned int last = 0;
+
+ unsigned char* code = buffer + 1;
+ unsigned char* data = code + index_count / 3;
+ unsigned char* data_safe_end = buffer + buffer_size - 16;
+
+ int fecmax = version >= 1 ? 13 : 15;
+
+ // use static encoding table; it's possible to pack the result and then build an optimal table and repack
+ // for now we keep it simple and use the table that has been generated based on symbol frequency on a training mesh set
+ const unsigned char* codeaux_table = kCodeAuxEncodingTable;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ // make sure we have enough space to write a triangle
+ // each triangle writes at most 16 bytes: 1b for codeaux and 5b for each free index
+ // after this we can be sure we can write without extra bounds checks
+ if (data > data_safe_end)
+ return 0;
+
+ int fer = getEdgeFifo(edgefifo, indices[i + 0], indices[i + 1], indices[i + 2], edgefifooffset);
+
+ if (fer >= 0 && (fer >> 2) < 15)
+ {
+ const unsigned int* order = kTriangleIndexOrder[fer & 3];
+
+ unsigned int a = indices[i + order[0]], b = indices[i + order[1]], c = indices[i + order[2]];
+
+ // encode edge index and vertex fifo index, next or free index
+ int fe = fer >> 2;
+ int fc = getVertexFifo(vertexfifo, c, vertexfifooffset);
+
+ int fec = (fc >= 1 && fc < fecmax) ? fc : (c == next) ? (next++, 0) : 15;
+
+ if (fec == 15 && version >= 1)
+ {
+ // encode last-1 and last+1 to optimize strip-like sequences
+ if (c + 1 == last)
+ fec = 13, last = c;
+ if (c == last + 1)
+ fec = 14, last = c;
+ }
+
+ *code++ = (unsigned char)((fe << 4) | fec);
+
+#if TRACE
+ codestats[code[-1]]++;
+#endif
+
+ // note that we need to update the last index since free indices are delta-encoded
+ if (fec == 15)
+ encodeIndex(data, c, last), last = c;
+
+ // we only need to push third vertex since first two are likely already in the vertex fifo
+ if (fec == 0 || fec >= fecmax)
+ pushVertexFifo(vertexfifo, c, vertexfifooffset);
+
+ // we only need to push two new edges to edge fifo since the third one is already there
+ pushEdgeFifo(edgefifo, c, b, edgefifooffset);
+ pushEdgeFifo(edgefifo, a, c, edgefifooffset);
+ }
+ else
+ {
+ int rotation = rotateTriangle(indices[i + 0], indices[i + 1], indices[i + 2], next);
+ const unsigned int* order = kTriangleIndexOrder[rotation];
+
+ unsigned int a = indices[i + order[0]], b = indices[i + order[1]], c = indices[i + order[2]];
+
+ // if a/b/c are 0/1/2, we emit a reset code
+ bool reset = false;
+
+ if (a == 0 && b == 1 && c == 2 && next > 0 && version >= 1)
+ {
+ reset = true;
+ next = 0;
+
+ // reset vertex fifo to make sure we don't accidentally reference vertices from that in the future
+ // this makes sure next continues to get incremented instead of being stuck
+ memset(vertexfifo, -1, sizeof(vertexfifo));
+ }
+
+ int fb = getVertexFifo(vertexfifo, b, vertexfifooffset);
+ int fc = getVertexFifo(vertexfifo, c, vertexfifooffset);
+
+ // after rotation, a is almost always equal to next, so we don't waste bits on FIFO encoding for a
+ int fea = (a == next) ? (next++, 0) : 15;
+ int feb = (fb >= 0 && fb < 14) ? (fb + 1) : (b == next) ? (next++, 0) : 15;
+ int fec = (fc >= 0 && fc < 14) ? (fc + 1) : (c == next) ? (next++, 0) : 15;
+
+ // we encode feb & fec in 4 bits using a table if possible, and as a full byte otherwise
+ unsigned char codeaux = (unsigned char)((feb << 4) | fec);
+ int codeauxindex = getCodeAuxIndex(codeaux, codeaux_table);
+
+ // <14 encodes an index into codeaux table, 14 encodes fea=0, 15 encodes fea=15
+ if (fea == 0 && codeauxindex >= 0 && codeauxindex < 14 && !reset)
+ {
+ *code++ = (unsigned char)((15 << 4) | codeauxindex);
+ }
+ else
+ {
+ *code++ = (unsigned char)((15 << 4) | 14 | fea);
+ *data++ = codeaux;
+ }
+
+#if TRACE
+ codestats[code[-1]]++;
+ codeauxstats[codeaux]++;
+#endif
+
+ // note that we need to update the last index since free indices are delta-encoded
+ if (fea == 15)
+ encodeIndex(data, a, last), last = a;
+
+ if (feb == 15)
+ encodeIndex(data, b, last), last = b;
+
+ if (fec == 15)
+ encodeIndex(data, c, last), last = c;
+
+ // only push vertices that weren't already in fifo
+ if (fea == 0 || fea == 15)
+ pushVertexFifo(vertexfifo, a, vertexfifooffset);
+
+ if (feb == 0 || feb == 15)
+ pushVertexFifo(vertexfifo, b, vertexfifooffset);
+
+ if (fec == 0 || fec == 15)
+ pushVertexFifo(vertexfifo, c, vertexfifooffset);
+
+ // all three edges aren't in the fifo; pushing all of them is important so that we can match them for later triangles
+ pushEdgeFifo(edgefifo, b, a, edgefifooffset);
+ pushEdgeFifo(edgefifo, c, b, edgefifooffset);
+ pushEdgeFifo(edgefifo, a, c, edgefifooffset);
+ }
+ }
+
+ // make sure we have enough space to write codeaux table
+ if (data > data_safe_end)
+ return 0;
+
+ // add codeaux encoding table to the end of the stream; this is used for decoding codeaux *and* as padding
+ // we need padding for decoding to be able to assume that each triangle is encoded as <= 16 bytes of extra data
+ // this is enough space for aux byte + 5 bytes per varint index which is the absolute worst case for any input
+ for (size_t i = 0; i < 16; ++i)
+ {
+ // decoder assumes that table entries never refer to separately encoded indices
+ assert((codeaux_table[i] & 0xf) != 0xf && (codeaux_table[i] >> 4) != 0xf);
+
+ *data++ = codeaux_table[i];
+ }
+
+ // since we encode restarts as codeaux without a table reference, we need to make sure 00 is encoded as a table reference
+ assert(codeaux_table[0] == 0);
+
+ assert(data >= buffer + index_count / 3 + 16);
+ assert(data <= buffer + buffer_size);
+
+#if TRACE
+ unsigned char codetop[16], codeauxtop[16];
+ size_t codetopsize = sortTop16(codetop, codestats);
+ size_t codeauxtopsize = sortTop16(codeauxtop, codeauxstats);
+
+ size_t sumcode = 0, sumcodeaux = 0;
+ for (size_t i = 0; i < 256; ++i)
+ sumcode += codestats[i], sumcodeaux += codeauxstats[i];
+
+ size_t acccode = 0, acccodeaux = 0;
+
+ printf("code\t\t\t\t\tcodeaux\n");
+
+ for (size_t i = 0; i < codetopsize && i < codeauxtopsize; ++i)
+ {
+ acccode += codestats[codetop[i]];
+ acccodeaux += codeauxstats[codeauxtop[i]];
+
+ printf("%2d: %02x = %d (%.1f%% ..%.1f%%)\t\t%2d: %02x = %d (%.1f%% ..%.1f%%)\n",
+ int(i), codetop[i], int(codestats[codetop[i]]), double(codestats[codetop[i]]) / double(sumcode) * 100, double(acccode) / double(sumcode) * 100,
+ int(i), codeauxtop[i], int(codeauxstats[codeauxtop[i]]), double(codeauxstats[codeauxtop[i]]) / double(sumcodeaux) * 100, double(acccodeaux) / double(sumcodeaux) * 100);
+ }
+#endif
+
+ return data - buffer;
+}
+
+size_t meshopt_encodeIndexBufferBound(size_t index_count, size_t vertex_count)
+{
+ assert(index_count % 3 == 0);
+
+ // compute number of bits required for each index
+ unsigned int vertex_bits = 1;
+
+ while (vertex_bits < 32 && vertex_count > size_t(1) << vertex_bits)
+ vertex_bits++;
+
+ // worst-case encoding is 2 header bytes + 3 varint-7 encoded index deltas
+ unsigned int vertex_groups = (vertex_bits + 1 + 6) / 7;
+
+ return 1 + (index_count / 3) * (2 + 3 * vertex_groups) + 16;
+}
+
+void meshopt_encodeIndexVersion(int version)
+{
+ assert(unsigned(version) <= 1);
+
+ meshopt::gEncodeIndexVersion = version;
+}
+
+int meshopt_decodeIndexBuffer(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(index_size == 2 || index_size == 4);
+
+ // the minimum valid encoding is header, 1 byte per triangle and a 16-byte codeaux table
+ if (buffer_size < 1 + index_count / 3 + 16)
+ return -2;
+
+ if ((buffer[0] & 0xf0) != kIndexHeader)
+ return -1;
+
+ int version = buffer[0] & 0x0f;
+ if (version > 1)
+ return -1;
+
+ EdgeFifo edgefifo;
+ memset(edgefifo, -1, sizeof(edgefifo));
+
+ VertexFifo vertexfifo;
+ memset(vertexfifo, -1, sizeof(vertexfifo));
+
+ size_t edgefifooffset = 0;
+ size_t vertexfifooffset = 0;
+
+ unsigned int next = 0;
+ unsigned int last = 0;
+
+ int fecmax = version >= 1 ? 13 : 15;
+
+ // since we store 16-byte codeaux table at the end, triangle data has to begin before data_safe_end
+ const unsigned char* code = buffer + 1;
+ const unsigned char* data = code + index_count / 3;
+ const unsigned char* data_safe_end = buffer + buffer_size - 16;
+
+ const unsigned char* codeaux_table = data_safe_end;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ // make sure we have enough data to read for a triangle
+ // each triangle reads at most 16 bytes of data: 1b for codeaux and 5b for each free index
+ // after this we can be sure we can read without extra bounds checks
+ if (data > data_safe_end)
+ return -2;
+
+ unsigned char codetri = *code++;
+
+ if (codetri < 0xf0)
+ {
+ int fe = codetri >> 4;
+
+ // fifo reads are wrapped around 16 entry buffer
+ unsigned int a = edgefifo[(edgefifooffset - 1 - fe) & 15][0];
+ unsigned int b = edgefifo[(edgefifooffset - 1 - fe) & 15][1];
+
+ int fec = codetri & 15;
+
+ // note: this is the most common path in the entire decoder
+ // inside this if we try to stay branchless (by using cmov/etc.) since these aren't predictable
+ if (fec < fecmax)
+ {
+ // fifo reads are wrapped around 16 entry buffer
+ unsigned int cf = vertexfifo[(vertexfifooffset - 1 - fec) & 15];
+ unsigned int c = (fec == 0) ? next : cf;
+
+ int fec0 = fec == 0;
+ next += fec0;
+
+ // output triangle
+ writeTriangle(destination, i, index_size, a, b, c);
+
+ // push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
+ pushVertexFifo(vertexfifo, c, vertexfifooffset, fec0);
+
+ pushEdgeFifo(edgefifo, c, b, edgefifooffset);
+ pushEdgeFifo(edgefifo, a, c, edgefifooffset);
+ }
+ else
+ {
+ unsigned int c = 0;
+
+ // fec - (fec ^ 3) decodes 13, 14 into -1, 1
+ // note that we need to update the last index since free indices are delta-encoded
+ last = c = (fec != 15) ? last + (fec - (fec ^ 3)) : decodeIndex(data, last);
+
+ // output triangle
+ writeTriangle(destination, i, index_size, a, b, c);
+
+ // push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
+ pushVertexFifo(vertexfifo, c, vertexfifooffset);
+
+ pushEdgeFifo(edgefifo, c, b, edgefifooffset);
+ pushEdgeFifo(edgefifo, a, c, edgefifooffset);
+ }
+ }
+ else
+ {
+ // fast path: read codeaux from the table
+ if (codetri < 0xfe)
+ {
+ unsigned char codeaux = codeaux_table[codetri & 15];
+
+ // note: table can't contain feb/fec=15
+ int feb = codeaux >> 4;
+ int fec = codeaux & 15;
+
+ // fifo reads are wrapped around 16 entry buffer
+ // also note that we increment next for all three vertices before decoding indices - this matches encoder behavior
+ unsigned int a = next++;
+
+ unsigned int bf = vertexfifo[(vertexfifooffset - feb) & 15];
+ unsigned int b = (feb == 0) ? next : bf;
+
+ int feb0 = feb == 0;
+ next += feb0;
+
+ unsigned int cf = vertexfifo[(vertexfifooffset - fec) & 15];
+ unsigned int c = (fec == 0) ? next : cf;
+
+ int fec0 = fec == 0;
+ next += fec0;
+
+ // output triangle
+ writeTriangle(destination, i, index_size, a, b, c);
+
+ // push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
+ pushVertexFifo(vertexfifo, a, vertexfifooffset);
+ pushVertexFifo(vertexfifo, b, vertexfifooffset, feb0);
+ pushVertexFifo(vertexfifo, c, vertexfifooffset, fec0);
+
+ pushEdgeFifo(edgefifo, b, a, edgefifooffset);
+ pushEdgeFifo(edgefifo, c, b, edgefifooffset);
+ pushEdgeFifo(edgefifo, a, c, edgefifooffset);
+ }
+ else
+ {
+ // slow path: read a full byte for codeaux instead of using a table lookup
+ unsigned char codeaux = *data++;
+
+ int fea = codetri == 0xfe ? 0 : 15;
+ int feb = codeaux >> 4;
+ int fec = codeaux & 15;
+
+ // reset: codeaux is 0 but encoded as not-a-table
+ if (codeaux == 0)
+ next = 0;
+
+ // fifo reads are wrapped around 16 entry buffer
+ // also note that we increment next for all three vertices before decoding indices - this matches encoder behavior
+ unsigned int a = (fea == 0) ? next++ : 0;
+ unsigned int b = (feb == 0) ? next++ : vertexfifo[(vertexfifooffset - feb) & 15];
+ unsigned int c = (fec == 0) ? next++ : vertexfifo[(vertexfifooffset - fec) & 15];
+
+ // note that we need to update the last index since free indices are delta-encoded
+ if (fea == 15)
+ last = a = decodeIndex(data, last);
+
+ if (feb == 15)
+ last = b = decodeIndex(data, last);
+
+ if (fec == 15)
+ last = c = decodeIndex(data, last);
+
+ // output triangle
+ writeTriangle(destination, i, index_size, a, b, c);
+
+ // push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
+ pushVertexFifo(vertexfifo, a, vertexfifooffset);
+ pushVertexFifo(vertexfifo, b, vertexfifooffset, (feb == 0) | (feb == 15));
+ pushVertexFifo(vertexfifo, c, vertexfifooffset, (fec == 0) | (fec == 15));
+
+ pushEdgeFifo(edgefifo, b, a, edgefifooffset);
+ pushEdgeFifo(edgefifo, c, b, edgefifooffset);
+ pushEdgeFifo(edgefifo, a, c, edgefifooffset);
+ }
+ }
+ }
+
+ // we should've read all data bytes and stopped at the boundary between data and codeaux table
+ if (data != data_safe_end)
+ return -3;
+
+ return 0;
+}
+
+size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count)
+{
+ using namespace meshopt;
+
+ // the minimum valid encoding is header, 1 byte per index and a 4-byte tail
+ if (buffer_size < 1 + index_count + 4)
+ return 0;
+
+ int version = gEncodeIndexVersion;
+
+ buffer[0] = (unsigned char)(kSequenceHeader | version);
+
+ unsigned int last[2] = {};
+ unsigned int current = 0;
+
+ unsigned char* data = buffer + 1;
+ unsigned char* data_safe_end = buffer + buffer_size - 4;
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ // make sure we have enough data to write
+ // each index writes at most 5 bytes of data; there's a 4 byte tail after data_safe_end
+ // after this we can be sure we can write without extra bounds checks
+ if (data >= data_safe_end)
+ return 0;
+
+ unsigned int index = indices[i];
+
+ // this is a heuristic that switches between baselines when the delta grows too large
+ // we want the encoded delta to fit into one byte (7 bits), but 2 bits are used for sign and baseline index
+ // for now we immediately switch the baseline when delta grows too large - this can be adjusted arbitrarily
+ int cd = int(index - last[current]);
+ current ^= ((cd < 0 ? -cd : cd) >= 30);
+
+ // encode delta from the last index
+ unsigned int d = index - last[current];
+ unsigned int v = (d << 1) ^ (int(d) >> 31);
+
+ // note: low bit encodes the index of the last baseline which will be used for reconstruction
+ encodeVByte(data, (v << 1) | current);
+
+ // update last for the next iteration that uses it
+ last[current] = index;
+ }
+
+ // make sure we have enough space to write tail
+ if (data > data_safe_end)
+ return 0;
+
+ for (int k = 0; k < 4; ++k)
+ *data++ = 0;
+
+ return data - buffer;
+}
+
+size_t meshopt_encodeIndexSequenceBound(size_t index_count, size_t vertex_count)
+{
+ // compute number of bits required for each index
+ unsigned int vertex_bits = 1;
+
+ while (vertex_bits < 32 && vertex_count > size_t(1) << vertex_bits)
+ vertex_bits++;
+
+ // worst-case encoding is 1 varint-7 encoded index delta for a K bit value and an extra bit
+ unsigned int vertex_groups = (vertex_bits + 1 + 1 + 6) / 7;
+
+ return 1 + index_count * vertex_groups + 4;
+}
+
+int meshopt_decodeIndexSequence(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size)
+{
+ using namespace meshopt;
+
+ // the minimum valid encoding is header, 1 byte per index and a 4-byte tail
+ if (buffer_size < 1 + index_count + 4)
+ return -2;
+
+ if ((buffer[0] & 0xf0) != kSequenceHeader)
+ return -1;
+
+ int version = buffer[0] & 0x0f;
+ if (version > 1)
+ return -1;
+
+ const unsigned char* data = buffer + 1;
+ const unsigned char* data_safe_end = buffer + buffer_size - 4;
+
+ unsigned int last[2] = {};
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ // make sure we have enough data to read
+ // each index reads at most 5 bytes of data; there's a 4 byte tail after data_safe_end
+ // after this we can be sure we can read without extra bounds checks
+ if (data >= data_safe_end)
+ return -2;
+
+ unsigned int v = decodeVByte(data);
+
+ // decode the index of the last baseline
+ unsigned int current = v & 1;
+ v >>= 1;
+
+ // reconstruct index as a delta
+ unsigned int d = (v >> 1) ^ -int(v & 1);
+ unsigned int index = last[current] + d;
+
+ // update last for the next iteration that uses it
+ last[current] = index;
+
+ if (index_size == 2)
+ {
+ static_cast<unsigned short*>(destination)[i] = (unsigned short)(index);
+ }
+ else
+ {
+ static_cast<unsigned int*>(destination)[i] = index;
+ }
+ }
+
+ // we should've read all data bytes and stopped at the boundary between data and tail
+ if (data != data_safe_end)
+ return -3;
+
+ return 0;
+}
diff --git a/thirdparty/meshoptimizer/indexgenerator.cpp b/thirdparty/meshoptimizer/indexgenerator.cpp
new file mode 100644
index 0000000000..aa4a30efa4
--- /dev/null
+++ b/thirdparty/meshoptimizer/indexgenerator.cpp
@@ -0,0 +1,347 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <string.h>
+
+namespace meshopt
+{
+
+static unsigned int hashUpdate4(unsigned int h, const unsigned char* key, size_t len)
+{
+ // MurmurHash2
+ const unsigned int m = 0x5bd1e995;
+ const int r = 24;
+
+ while (len >= 4)
+ {
+ unsigned int k = *reinterpret_cast<const unsigned int*>(key);
+
+ k *= m;
+ k ^= k >> r;
+ k *= m;
+
+ h *= m;
+ h ^= k;
+
+ key += 4;
+ len -= 4;
+ }
+
+ return h;
+}
+
+struct VertexHasher
+{
+ const unsigned char* vertices;
+ size_t vertex_size;
+ size_t vertex_stride;
+
+ size_t hash(unsigned int index) const
+ {
+ return hashUpdate4(0, vertices + index * vertex_stride, vertex_size);
+ }
+
+ bool equal(unsigned int lhs, unsigned int rhs) const
+ {
+ return memcmp(vertices + lhs * vertex_stride, vertices + rhs * vertex_stride, vertex_size) == 0;
+ }
+};
+
+struct VertexStreamHasher
+{
+ const meshopt_Stream* streams;
+ size_t stream_count;
+
+ size_t hash(unsigned int index) const
+ {
+ unsigned int h = 0;
+
+ for (size_t i = 0; i < stream_count; ++i)
+ {
+ const meshopt_Stream& s = streams[i];
+ const unsigned char* data = static_cast<const unsigned char*>(s.data);
+
+ h = hashUpdate4(h, data + index * s.stride, s.size);
+ }
+
+ return h;
+ }
+
+ bool equal(unsigned int lhs, unsigned int rhs) const
+ {
+ for (size_t i = 0; i < stream_count; ++i)
+ {
+ const meshopt_Stream& s = streams[i];
+ const unsigned char* data = static_cast<const unsigned char*>(s.data);
+
+ if (memcmp(data + lhs * s.stride, data + rhs * s.stride, s.size) != 0)
+ return false;
+ }
+
+ return true;
+ }
+};
+
+static size_t hashBuckets(size_t count)
+{
+ size_t buckets = 1;
+ while (buckets < count)
+ buckets *= 2;
+
+ return buckets;
+}
+
+template <typename T, typename Hash>
+static T* hashLookup(T* table, size_t buckets, const Hash& hash, const T& key, const T& empty)
+{
+ assert(buckets > 0);
+ assert((buckets & (buckets - 1)) == 0);
+
+ size_t hashmod = buckets - 1;
+ size_t bucket = hash.hash(key) & hashmod;
+
+ for (size_t probe = 0; probe <= hashmod; ++probe)
+ {
+ T& item = table[bucket];
+
+ if (item == empty)
+ return &item;
+
+ if (hash.equal(item, key))
+ return &item;
+
+ // hash collision, quadratic probing
+ bucket = (bucket + probe + 1) & hashmod;
+ }
+
+ assert(false && "Hash table is full"); // unreachable
+ return 0;
+}
+
+} // namespace meshopt
+
+size_t meshopt_generateVertexRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
+{
+ using namespace meshopt;
+
+ assert(indices || index_count == vertex_count);
+ assert(index_count % 3 == 0);
+ assert(vertex_size > 0 && vertex_size <= 256);
+
+ meshopt_Allocator allocator;
+
+ memset(destination, -1, vertex_count * sizeof(unsigned int));
+
+ VertexHasher hasher = {static_cast<const unsigned char*>(vertices), vertex_size, vertex_size};
+
+ size_t table_size = hashBuckets(vertex_count);
+ unsigned int* table = allocator.allocate<unsigned int>(table_size);
+ memset(table, -1, table_size * sizeof(unsigned int));
+
+ unsigned int next_vertex = 0;
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices ? indices[i] : unsigned(i);
+ assert(index < vertex_count);
+
+ if (destination[index] == ~0u)
+ {
+ unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
+
+ if (*entry == ~0u)
+ {
+ *entry = index;
+
+ destination[index] = next_vertex++;
+ }
+ else
+ {
+ assert(destination[*entry] != ~0u);
+
+ destination[index] = destination[*entry];
+ }
+ }
+ }
+
+ assert(next_vertex <= vertex_count);
+
+ return next_vertex;
+}
+
+size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count)
+{
+ using namespace meshopt;
+
+ assert(indices || index_count == vertex_count);
+ assert(index_count % 3 == 0);
+ assert(stream_count > 0 && stream_count <= 16);
+
+ for (size_t i = 0; i < stream_count; ++i)
+ {
+ assert(streams[i].size > 0 && streams[i].size <= 256);
+ assert(streams[i].size <= streams[i].stride);
+ }
+
+ meshopt_Allocator allocator;
+
+ memset(destination, -1, vertex_count * sizeof(unsigned int));
+
+ VertexStreamHasher hasher = {streams, stream_count};
+
+ size_t table_size = hashBuckets(vertex_count);
+ unsigned int* table = allocator.allocate<unsigned int>(table_size);
+ memset(table, -1, table_size * sizeof(unsigned int));
+
+ unsigned int next_vertex = 0;
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices ? indices[i] : unsigned(i);
+ assert(index < vertex_count);
+
+ if (destination[index] == ~0u)
+ {
+ unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
+
+ if (*entry == ~0u)
+ {
+ *entry = index;
+
+ destination[index] = next_vertex++;
+ }
+ else
+ {
+ assert(destination[*entry] != ~0u);
+
+ destination[index] = destination[*entry];
+ }
+ }
+ }
+
+ assert(next_vertex <= vertex_count);
+
+ return next_vertex;
+}
+
+void meshopt_remapVertexBuffer(void* destination, const void* vertices, size_t vertex_count, size_t vertex_size, const unsigned int* remap)
+{
+ assert(vertex_size > 0 && vertex_size <= 256);
+
+ meshopt_Allocator allocator;
+
+ // support in-place remap
+ if (destination == vertices)
+ {
+ unsigned char* vertices_copy = allocator.allocate<unsigned char>(vertex_count * vertex_size);
+ memcpy(vertices_copy, vertices, vertex_count * vertex_size);
+ vertices = vertices_copy;
+ }
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ if (remap[i] != ~0u)
+ {
+ assert(remap[i] < vertex_count);
+
+ memcpy(static_cast<unsigned char*>(destination) + remap[i] * vertex_size, static_cast<const unsigned char*>(vertices) + i * vertex_size, vertex_size);
+ }
+ }
+}
+
+void meshopt_remapIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const unsigned int* remap)
+{
+ assert(index_count % 3 == 0);
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices ? indices[i] : unsigned(i);
+ assert(remap[index] != ~0u);
+
+ destination[i] = remap[index];
+ }
+}
+
+void meshopt_generateShadowIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride)
+{
+ using namespace meshopt;
+
+ assert(indices);
+ assert(index_count % 3 == 0);
+ assert(vertex_size > 0 && vertex_size <= 256);
+ assert(vertex_size <= vertex_stride);
+
+ meshopt_Allocator allocator;
+
+ unsigned int* remap = allocator.allocate<unsigned int>(vertex_count);
+ memset(remap, -1, vertex_count * sizeof(unsigned int));
+
+ VertexHasher hasher = {static_cast<const unsigned char*>(vertices), vertex_size, vertex_stride};
+
+ size_t table_size = hashBuckets(vertex_count);
+ unsigned int* table = allocator.allocate<unsigned int>(table_size);
+ memset(table, -1, table_size * sizeof(unsigned int));
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices[i];
+ assert(index < vertex_count);
+
+ if (remap[index] == ~0u)
+ {
+ unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
+
+ if (*entry == ~0u)
+ *entry = index;
+
+ remap[index] = *entry;
+ }
+
+ destination[i] = remap[index];
+ }
+}
+
+void meshopt_generateShadowIndexBufferMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count)
+{
+ using namespace meshopt;
+
+ assert(indices);
+ assert(index_count % 3 == 0);
+ assert(stream_count > 0 && stream_count <= 16);
+
+ for (size_t i = 0; i < stream_count; ++i)
+ {
+ assert(streams[i].size > 0 && streams[i].size <= 256);
+ assert(streams[i].size <= streams[i].stride);
+ }
+
+ meshopt_Allocator allocator;
+
+ unsigned int* remap = allocator.allocate<unsigned int>(vertex_count);
+ memset(remap, -1, vertex_count * sizeof(unsigned int));
+
+ VertexStreamHasher hasher = {streams, stream_count};
+
+ size_t table_size = hashBuckets(vertex_count);
+ unsigned int* table = allocator.allocate<unsigned int>(table_size);
+ memset(table, -1, table_size * sizeof(unsigned int));
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices[i];
+ assert(index < vertex_count);
+
+ if (remap[index] == ~0u)
+ {
+ unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
+
+ if (*entry == ~0u)
+ *entry = index;
+
+ remap[index] = *entry;
+ }
+
+ destination[i] = remap[index];
+ }
+}
diff --git a/thirdparty/meshoptimizer/meshoptimizer.h b/thirdparty/meshoptimizer/meshoptimizer.h
new file mode 100644
index 0000000000..fde00f9c82
--- /dev/null
+++ b/thirdparty/meshoptimizer/meshoptimizer.h
@@ -0,0 +1,951 @@
+/**
+ * meshoptimizer - version 0.15
+ *
+ * Copyright (C) 2016-2020, by Arseny Kapoulkine (arseny.kapoulkine@gmail.com)
+ * Report bugs and download new versions at https://github.com/zeux/meshoptimizer
+ *
+ * This library is distributed under the MIT License. See notice at the end of this file.
+ */
+#pragma once
+
+#include <assert.h>
+#include <stddef.h>
+
+/* Version macro; major * 1000 + minor * 10 + patch */
+#define MESHOPTIMIZER_VERSION 150 /* 0.15 */
+
+/* If no API is defined, assume default */
+#ifndef MESHOPTIMIZER_API
+#define MESHOPTIMIZER_API
+#endif
+
+/* Experimental APIs have unstable interface and might have implementation that's not fully tested or optimized */
+#define MESHOPTIMIZER_EXPERIMENTAL MESHOPTIMIZER_API
+
+/* C interface */
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * Vertex attribute stream, similar to glVertexPointer
+ * Each element takes size bytes, with stride controlling the spacing between successive elements.
+ */
+struct meshopt_Stream
+{
+ const void* data;
+ size_t size;
+ size_t stride;
+};
+
+/**
+ * Generates a vertex remap table from the vertex buffer and an optional index buffer and returns number of unique vertices
+ * As a result, all vertices that are binary equivalent map to the same (new) location, with no gaps in the resulting sequence.
+ * Resulting remap table maps old vertices to new vertices and can be used in meshopt_remapVertexBuffer/meshopt_remapIndexBuffer.
+ * Note that binary equivalence considers all vertex_size bytes, including padding which should be zero-initialized.
+ *
+ * destination must contain enough space for the resulting remap table (vertex_count elements)
+ * indices can be NULL if the input is unindexed
+ */
+MESHOPTIMIZER_API size_t meshopt_generateVertexRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
+
+/**
+ * Generates a vertex remap table from multiple vertex streams and an optional index buffer and returns number of unique vertices
+ * As a result, all vertices that are binary equivalent map to the same (new) location, with no gaps in the resulting sequence.
+ * Resulting remap table maps old vertices to new vertices and can be used in meshopt_remapVertexBuffer/meshopt_remapIndexBuffer.
+ * To remap vertex buffers, you will need to call meshopt_remapVertexBuffer for each vertex stream.
+ * Note that binary equivalence considers all size bytes in each stream, including padding which should be zero-initialized.
+ *
+ * destination must contain enough space for the resulting remap table (vertex_count elements)
+ * indices can be NULL if the input is unindexed
+ */
+MESHOPTIMIZER_API size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count);
+
+/**
+ * Generates vertex buffer from the source vertex buffer and remap table generated by meshopt_generateVertexRemap
+ *
+ * destination must contain enough space for the resulting vertex buffer (unique_vertex_count elements, returned by meshopt_generateVertexRemap)
+ * vertex_count should be the initial vertex count and not the value returned by meshopt_generateVertexRemap
+ */
+MESHOPTIMIZER_API void meshopt_remapVertexBuffer(void* destination, const void* vertices, size_t vertex_count, size_t vertex_size, const unsigned int* remap);
+
+/**
+ * Generate index buffer from the source index buffer and remap table generated by meshopt_generateVertexRemap
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ * indices can be NULL if the input is unindexed
+ */
+MESHOPTIMIZER_API void meshopt_remapIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const unsigned int* remap);
+
+/**
+ * Generate index buffer that can be used for more efficient rendering when only a subset of the vertex attributes is necessary
+ * All vertices that are binary equivalent (wrt first vertex_size bytes) map to the first vertex in the original vertex buffer.
+ * This makes it possible to use the index buffer for Z pre-pass or shadowmap rendering, while using the original index buffer for regular rendering.
+ * Note that binary equivalence considers all vertex_size bytes, including padding which should be zero-initialized.
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ */
+MESHOPTIMIZER_API void meshopt_generateShadowIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride);
+
+/**
+ * Generate index buffer that can be used for more efficient rendering when only a subset of the vertex attributes is necessary
+ * All vertices that are binary equivalent (wrt specified streams) map to the first vertex in the original vertex buffer.
+ * This makes it possible to use the index buffer for Z pre-pass or shadowmap rendering, while using the original index buffer for regular rendering.
+ * Note that binary equivalence considers all size bytes in each stream, including padding which should be zero-initialized.
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ */
+MESHOPTIMIZER_API void meshopt_generateShadowIndexBufferMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count);
+
+/**
+ * Vertex transform cache optimizer
+ * Reorders indices to reduce the number of GPU vertex shader invocations
+ * If index buffer contains multiple ranges for multiple draw calls, this functions needs to be called on each range individually.
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ */
+MESHOPTIMIZER_API void meshopt_optimizeVertexCache(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count);
+
+/**
+ * Vertex transform cache optimizer for strip-like caches
+ * Produces inferior results to meshopt_optimizeVertexCache from the GPU vertex cache perspective
+ * However, the resulting index order is more optimal if the goal is to reduce the triangle strip length or improve compression efficiency
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ */
+MESHOPTIMIZER_API void meshopt_optimizeVertexCacheStrip(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count);
+
+/**
+ * Vertex transform cache optimizer for FIFO caches
+ * Reorders indices to reduce the number of GPU vertex shader invocations
+ * Generally takes ~3x less time to optimize meshes but produces inferior results compared to meshopt_optimizeVertexCache
+ * If index buffer contains multiple ranges for multiple draw calls, this functions needs to be called on each range individually.
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ * cache_size should be less than the actual GPU cache size to avoid cache thrashing
+ */
+MESHOPTIMIZER_API void meshopt_optimizeVertexCacheFifo(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size);
+
+/**
+ * Overdraw optimizer
+ * Reorders indices to reduce the number of GPU vertex shader invocations and the pixel overdraw
+ * If index buffer contains multiple ranges for multiple draw calls, this functions needs to be called on each range individually.
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ * indices must contain index data that is the result of meshopt_optimizeVertexCache (*not* the original mesh indices!)
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ * threshold indicates how much the overdraw optimizer can degrade vertex cache efficiency (1.05 = up to 5%) to reduce overdraw more efficiently
+ */
+MESHOPTIMIZER_API void meshopt_optimizeOverdraw(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold);
+
+/**
+ * Vertex fetch cache optimizer
+ * Reorders vertices and changes indices to reduce the amount of GPU memory fetches during vertex processing
+ * Returns the number of unique vertices, which is the same as input vertex count unless some vertices are unused
+ * This functions works for a single vertex stream; for multiple vertex streams, use meshopt_optimizeVertexFetchRemap + meshopt_remapVertexBuffer for each stream.
+ *
+ * destination must contain enough space for the resulting vertex buffer (vertex_count elements)
+ * indices is used both as an input and as an output index buffer
+ */
+MESHOPTIMIZER_API size_t meshopt_optimizeVertexFetch(void* destination, unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
+
+/**
+ * Vertex fetch cache optimizer
+ * Generates vertex remap to reduce the amount of GPU memory fetches during vertex processing
+ * Returns the number of unique vertices, which is the same as input vertex count unless some vertices are unused
+ * The resulting remap table should be used to reorder vertex/index buffers using meshopt_remapVertexBuffer/meshopt_remapIndexBuffer
+ *
+ * destination must contain enough space for the resulting remap table (vertex_count elements)
+ */
+MESHOPTIMIZER_API size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count);
+
+/**
+ * Index buffer encoder
+ * Encodes index data into an array of bytes that is generally much smaller (<1.5 bytes/triangle) and compresses better (<1 bytes/triangle) compared to original.
+ * Input index buffer must represent a triangle list.
+ * Returns encoded data size on success, 0 on error; the only error condition is if buffer doesn't have enough space
+ * For maximum efficiency the index buffer being encoded has to be optimized for vertex cache and vertex fetch first.
+ *
+ * buffer must contain enough space for the encoded index buffer (use meshopt_encodeIndexBufferBound to compute worst case size)
+ */
+MESHOPTIMIZER_API size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count);
+MESHOPTIMIZER_API size_t meshopt_encodeIndexBufferBound(size_t index_count, size_t vertex_count);
+
+/**
+ * Experimental: Set index encoder format version
+ * version must specify the data format version to encode; valid values are 0 (decodable by all library versions) and 1 (decodable by 0.14+)
+ */
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_encodeIndexVersion(int version);
+
+/**
+ * Index buffer decoder
+ * Decodes index data from an array of bytes generated by meshopt_encodeIndexBuffer
+ * Returns 0 if decoding was successful, and an error code otherwise
+ * The decoder is safe to use for untrusted input, but it may produce garbage data (e.g. out of range indices).
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ */
+MESHOPTIMIZER_API int meshopt_decodeIndexBuffer(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size);
+
+/**
+ * Experimental: Index sequence encoder
+ * Encodes index sequence into an array of bytes that is generally smaller and compresses better compared to original.
+ * Input index sequence can represent arbitrary topology; for triangle lists meshopt_encodeIndexBuffer is likely to be better.
+ * Returns encoded data size on success, 0 on error; the only error condition is if buffer doesn't have enough space
+ *
+ * buffer must contain enough space for the encoded index sequence (use meshopt_encodeIndexSequenceBound to compute worst case size)
+ */
+MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count);
+MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_encodeIndexSequenceBound(size_t index_count, size_t vertex_count);
+
+/**
+ * Index sequence decoder
+ * Decodes index data from an array of bytes generated by meshopt_encodeIndexSequence
+ * Returns 0 if decoding was successful, and an error code otherwise
+ * The decoder is safe to use for untrusted input, but it may produce garbage data (e.g. out of range indices).
+ *
+ * destination must contain enough space for the resulting index sequence (index_count elements)
+ */
+MESHOPTIMIZER_EXPERIMENTAL int meshopt_decodeIndexSequence(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size);
+
+/**
+ * Vertex buffer encoder
+ * Encodes vertex data into an array of bytes that is generally smaller and compresses better compared to original.
+ * Returns encoded data size on success, 0 on error; the only error condition is if buffer doesn't have enough space
+ * This function works for a single vertex stream; for multiple vertex streams, call meshopt_encodeVertexBuffer for each stream.
+ * Note that all vertex_size bytes of each vertex are encoded verbatim, including padding which should be zero-initialized.
+ *
+ * buffer must contain enough space for the encoded vertex buffer (use meshopt_encodeVertexBufferBound to compute worst case size)
+ */
+MESHOPTIMIZER_API size_t meshopt_encodeVertexBuffer(unsigned char* buffer, size_t buffer_size, const void* vertices, size_t vertex_count, size_t vertex_size);
+MESHOPTIMIZER_API size_t meshopt_encodeVertexBufferBound(size_t vertex_count, size_t vertex_size);
+
+/**
+ * Experimental: Set vertex encoder format version
+ * version must specify the data format version to encode; valid values are 0 (decodable by all library versions)
+ */
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_encodeVertexVersion(int version);
+
+/**
+ * Vertex buffer decoder
+ * Decodes vertex data from an array of bytes generated by meshopt_encodeVertexBuffer
+ * Returns 0 if decoding was successful, and an error code otherwise
+ * The decoder is safe to use for untrusted input, but it may produce garbage data.
+ *
+ * destination must contain enough space for the resulting vertex buffer (vertex_count * vertex_size bytes)
+ */
+MESHOPTIMIZER_API int meshopt_decodeVertexBuffer(void* destination, size_t vertex_count, size_t vertex_size, const unsigned char* buffer, size_t buffer_size);
+
+/**
+ * Vertex buffer filters
+ * These functions can be used to filter output of meshopt_decodeVertexBuffer in-place.
+ * count must be aligned by 4 and stride is fixed for each function to facilitate SIMD implementation.
+ *
+ * meshopt_decodeFilterOct decodes octahedral encoding of a unit vector with K-bit (K <= 16) signed X/Y as an input; Z must store 1.0f.
+ * Each component is stored as an 8-bit or 16-bit normalized integer; stride must be equal to 4 or 8. W is preserved as is.
+ *
+ * meshopt_decodeFilterQuat decodes 3-component quaternion encoding with K-bit (4 <= K <= 16) component encoding and a 2-bit component index indicating which component to reconstruct.
+ * Each component is stored as an 16-bit integer; stride must be equal to 8.
+ *
+ * meshopt_decodeFilterExp decodes exponential encoding of floating-point data with 8-bit exponent and 24-bit integer mantissa as 2^E*M.
+ * Each 32-bit component is decoded in isolation; stride must be divisible by 4.
+ */
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_decodeFilterOct(void* buffer, size_t vertex_count, size_t vertex_size);
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_decodeFilterQuat(void* buffer, size_t vertex_count, size_t vertex_size);
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_decodeFilterExp(void* buffer, size_t vertex_count, size_t vertex_size);
+
+/**
+ * Experimental: Mesh simplifier
+ * Reduces the number of triangles in the mesh, attempting to preserve mesh appearance as much as possible
+ * The algorithm tries to preserve mesh topology and can stop short of the target goal based on topology constraints or target error.
+ * If not all attributes from the input mesh are required, it's recommended to reindex the mesh using meshopt_generateShadowIndexBuffer prior to simplification.
+ * Returns the number of indices after simplification, with destination containing new index data
+ * The resulting index buffer references vertices from the original vertex buffer.
+ * If the original vertex data isn't required, creating a compact vertex buffer using meshopt_optimizeVertexFetch is recommended.
+ *
+ * destination must contain enough space for the *source* index buffer (since optimization is iterative, this means index_count elements - *not* target_index_count!)
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ */
+// -- GODOT start --
+//MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error);
+MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplify(unsigned int *destination, const unsigned int *indices, size_t index_count, const float *vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error, float *r_resulting_error);
+// -- GODOT end --
+
+/**
+ * Experimental: Mesh simplifier (sloppy)
+ * Reduces the number of triangles in the mesh, sacrificing mesh apperance for simplification performance
+ * The algorithm doesn't preserve mesh topology but is always able to reach target triangle count.
+ * Returns the number of indices after simplification, with destination containing new index data
+ * The resulting index buffer references vertices from the original vertex buffer.
+ * If the original vertex data isn't required, creating a compact vertex buffer using meshopt_optimizeVertexFetch is recommended.
+ *
+ * destination must contain enough space for the target index buffer
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ */
+MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplifySloppy(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count);
+
+/**
+ * Experimental: Point cloud simplifier
+ * Reduces the number of points in the cloud to reach the given target
+ * Returns the number of points after simplification, with destination containing new index data
+ * The resulting index buffer references vertices from the original vertex buffer.
+ * If the original vertex data isn't required, creating a compact vertex buffer using meshopt_optimizeVertexFetch is recommended.
+ *
+ * destination must contain enough space for the target index buffer
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ */
+MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplifyPoints(unsigned int* destination, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_vertex_count);
+
+/**
+ * Mesh stripifier
+ * Converts a previously vertex cache optimized triangle list to triangle strip, stitching strips using restart index or degenerate triangles
+ * Returns the number of indices in the resulting strip, with destination containing new index data
+ * For maximum efficiency the index buffer being converted has to be optimized for vertex cache first.
+ * Using restart indices can result in ~10% smaller index buffers, but on some GPUs restart indices may result in decreased performance.
+ *
+ * destination must contain enough space for the target index buffer, worst case can be computed with meshopt_stripifyBound
+ * restart_index should be 0xffff or 0xffffffff depending on index size, or 0 to use degenerate triangles
+ */
+MESHOPTIMIZER_API size_t meshopt_stripify(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int restart_index);
+MESHOPTIMIZER_API size_t meshopt_stripifyBound(size_t index_count);
+
+/**
+ * Mesh unstripifier
+ * Converts a triangle strip to a triangle list
+ * Returns the number of indices in the resulting list, with destination containing new index data
+ *
+ * destination must contain enough space for the target index buffer, worst case can be computed with meshopt_unstripifyBound
+ */
+MESHOPTIMIZER_API size_t meshopt_unstripify(unsigned int* destination, const unsigned int* indices, size_t index_count, unsigned int restart_index);
+MESHOPTIMIZER_API size_t meshopt_unstripifyBound(size_t index_count);
+
+struct meshopt_VertexCacheStatistics
+{
+ unsigned int vertices_transformed;
+ unsigned int warps_executed;
+ float acmr; /* transformed vertices / triangle count; best case 0.5, worst case 3.0, optimum depends on topology */
+ float atvr; /* transformed vertices / vertex count; best case 1.0, worst case 6.0, optimum is 1.0 (each vertex is transformed once) */
+};
+
+/**
+ * Vertex transform cache analyzer
+ * Returns cache hit statistics using a simplified FIFO model
+ * Results may not match actual GPU performance
+ */
+MESHOPTIMIZER_API struct meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int primgroup_size);
+
+struct meshopt_OverdrawStatistics
+{
+ unsigned int pixels_covered;
+ unsigned int pixels_shaded;
+ float overdraw; /* shaded pixels / covered pixels; best case 1.0 */
+};
+
+/**
+ * Overdraw analyzer
+ * Returns overdraw statistics using a software rasterizer
+ * Results may not match actual GPU performance
+ *
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ */
+MESHOPTIMIZER_API struct meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+
+struct meshopt_VertexFetchStatistics
+{
+ unsigned int bytes_fetched;
+ float overfetch; /* fetched bytes / vertex buffer size; best case 1.0 (each byte is fetched once) */
+};
+
+/**
+ * Vertex fetch cache analyzer
+ * Returns cache hit statistics using a simplified direct mapped model
+ * Results may not match actual GPU performance
+ */
+MESHOPTIMIZER_API struct meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const unsigned int* indices, size_t index_count, size_t vertex_count, size_t vertex_size);
+
+struct meshopt_Meshlet
+{
+ unsigned int vertices[64];
+ unsigned char indices[126][3];
+ unsigned char triangle_count;
+ unsigned char vertex_count;
+};
+
+/**
+ * Experimental: Meshlet builder
+ * Splits the mesh into a set of meshlets where each meshlet has a micro index buffer indexing into meshlet vertices that refer to the original vertex buffer
+ * The resulting data can be used to render meshes using NVidia programmable mesh shading pipeline, or in other cluster-based renderers.
+ * For maximum efficiency the index buffer being converted has to be optimized for vertex cache first.
+ *
+ * destination must contain enough space for all meshlets, worst case size can be computed with meshopt_buildMeshletsBound
+ * max_vertices and max_triangles can't exceed limits statically declared in meshopt_Meshlet (max_vertices <= 64, max_triangles <= 126)
+ */
+MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_buildMeshlets(struct meshopt_Meshlet* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles);
+MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_buildMeshletsBound(size_t index_count, size_t max_vertices, size_t max_triangles);
+
+struct meshopt_Bounds
+{
+ /* bounding sphere, useful for frustum and occlusion culling */
+ float center[3];
+ float radius;
+
+ /* normal cone, useful for backface culling */
+ float cone_apex[3];
+ float cone_axis[3];
+ float cone_cutoff; /* = cos(angle/2) */
+
+ /* normal cone axis and cutoff, stored in 8-bit SNORM format; decode using x/127.0 */
+ signed char cone_axis_s8[3];
+ signed char cone_cutoff_s8;
+};
+
+/**
+ * Experimental: Cluster bounds generator
+ * Creates bounding volumes that can be used for frustum, backface and occlusion culling.
+ *
+ * For backface culling with orthographic projection, use the following formula to reject backfacing clusters:
+ * dot(view, cone_axis) >= cone_cutoff
+ *
+ * For perspective projection, you can the formula that needs cone apex in addition to axis & cutoff:
+ * dot(normalize(cone_apex - camera_position), cone_axis) >= cone_cutoff
+ *
+ * Alternatively, you can use the formula that doesn't need cone apex and uses bounding sphere instead:
+ * dot(normalize(center - camera_position), cone_axis) >= cone_cutoff + radius / length(center - camera_position)
+ * or an equivalent formula that doesn't have a singularity at center = camera_position:
+ * dot(center - camera_position, cone_axis) >= cone_cutoff * length(center - camera_position) + radius
+ *
+ * The formula that uses the apex is slightly more accurate but needs the apex; if you are already using bounding sphere
+ * to do frustum/occlusion culling, the formula that doesn't use the apex may be preferable.
+ *
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ * index_count should be less than or equal to 256*3 (the function assumes clusters of limited size)
+ */
+MESHOPTIMIZER_EXPERIMENTAL struct meshopt_Bounds meshopt_computeClusterBounds(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+MESHOPTIMIZER_EXPERIMENTAL struct meshopt_Bounds meshopt_computeMeshletBounds(const struct meshopt_Meshlet* meshlet, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+
+/**
+ * Experimental: Spatial sorter
+ * Generates a remap table that can be used to reorder points for spatial locality.
+ * Resulting remap table maps old vertices to new vertices and can be used in meshopt_remapVertexBuffer.
+ *
+ * destination must contain enough space for the resulting remap table (vertex_count elements)
+ */
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_spatialSortRemap(unsigned int* destination, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+
+/**
+ * Experimental: Spatial sorter
+ * Reorders triangles for spatial locality, and generates a new index buffer. The resulting index buffer can be used with other functions like optimizeVertexCache.
+ *
+ * destination must contain enough space for the resulting index buffer (index_count elements)
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ */
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_spatialSortTriangles(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+
+/**
+ * Set allocation callbacks
+ * These callbacks will be used instead of the default operator new/operator delete for all temporary allocations in the library.
+ * Note that all algorithms only allocate memory for temporary use.
+ * allocate/deallocate are always called in a stack-like order - last pointer to be allocated is deallocated first.
+ */
+MESHOPTIMIZER_API void meshopt_setAllocator(void* (*allocate)(size_t), void (*deallocate)(void*));
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
+/* Quantization into commonly supported data formats */
+#ifdef __cplusplus
+/**
+ * Quantize a float in [0..1] range into an N-bit fixed point unorm value
+ * Assumes reconstruction function (q / (2^N-1)), which is the case for fixed-function normalized fixed point conversion
+ * Maximum reconstruction error: 1/2^(N+1)
+ */
+inline int meshopt_quantizeUnorm(float v, int N);
+
+/**
+ * Quantize a float in [-1..1] range into an N-bit fixed point snorm value
+ * Assumes reconstruction function (q / (2^(N-1)-1)), which is the case for fixed-function normalized fixed point conversion (except early OpenGL versions)
+ * Maximum reconstruction error: 1/2^N
+ */
+inline int meshopt_quantizeSnorm(float v, int N);
+
+/**
+ * Quantize a float into half-precision floating point value
+ * Generates +-inf for overflow, preserves NaN, flushes denormals to zero, rounds to nearest
+ * Representable magnitude range: [6e-5; 65504]
+ * Maximum relative reconstruction error: 5e-4
+ */
+inline unsigned short meshopt_quantizeHalf(float v);
+
+/**
+ * Quantize a float into a floating point value with a limited number of significant mantissa bits
+ * Generates +-inf for overflow, preserves NaN, flushes denormals to zero, rounds to nearest
+ * Assumes N is in a valid mantissa precision range, which is 1..23
+ */
+inline float meshopt_quantizeFloat(float v, int N);
+#endif
+
+/**
+ * C++ template interface
+ *
+ * These functions mirror the C interface the library provides, providing template-based overloads so that
+ * the caller can use an arbitrary type for the index data, both for input and output.
+ * When the supplied type is the same size as that of unsigned int, the wrappers are zero-cost; when it's not,
+ * the wrappers end up allocating memory and copying index data to convert from one type to another.
+ */
+#if defined(__cplusplus) && !defined(MESHOPTIMIZER_NO_WRAPPERS)
+template <typename T>
+inline size_t meshopt_generateVertexRemap(unsigned int* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
+template <typename T>
+inline size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count);
+template <typename T>
+inline void meshopt_remapIndexBuffer(T* destination, const T* indices, size_t index_count, const unsigned int* remap);
+template <typename T>
+inline void meshopt_generateShadowIndexBuffer(T* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride);
+template <typename T>
+inline void meshopt_generateShadowIndexBufferMulti(T* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count);
+template <typename T>
+inline void meshopt_optimizeVertexCache(T* destination, const T* indices, size_t index_count, size_t vertex_count);
+template <typename T>
+inline void meshopt_optimizeVertexCacheStrip(T* destination, const T* indices, size_t index_count, size_t vertex_count);
+template <typename T>
+inline void meshopt_optimizeVertexCacheFifo(T* destination, const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size);
+template <typename T>
+inline void meshopt_optimizeOverdraw(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold);
+template <typename T>
+inline size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count);
+template <typename T>
+inline size_t meshopt_optimizeVertexFetch(void* destination, T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
+template <typename T>
+inline size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count);
+template <typename T>
+inline int meshopt_decodeIndexBuffer(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size);
+template <typename T>
+inline size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count);
+template <typename T>
+inline int meshopt_decodeIndexSequence(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size);
+template <typename T>
+inline size_t meshopt_simplify(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error);
+template <typename T>
+inline size_t meshopt_simplifySloppy(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count);
+template <typename T>
+inline size_t meshopt_stripify(T* destination, const T* indices, size_t index_count, size_t vertex_count, T restart_index);
+template <typename T>
+inline size_t meshopt_unstripify(T* destination, const T* indices, size_t index_count, T restart_index);
+template <typename T>
+inline meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int buffer_size);
+template <typename T>
+inline meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+template <typename T>
+inline meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const T* indices, size_t index_count, size_t vertex_count, size_t vertex_size);
+template <typename T>
+inline size_t meshopt_buildMeshlets(meshopt_Meshlet* destination, const T* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles);
+template <typename T>
+inline meshopt_Bounds meshopt_computeClusterBounds(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+template <typename T>
+inline void meshopt_spatialSortTriangles(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
+#endif
+
+/* Inline implementation */
+#ifdef __cplusplus
+inline int meshopt_quantizeUnorm(float v, int N)
+{
+ const float scale = float((1 << N) - 1);
+
+ v = (v >= 0) ? v : 0;
+ v = (v <= 1) ? v : 1;
+
+ return int(v * scale + 0.5f);
+}
+
+inline int meshopt_quantizeSnorm(float v, int N)
+{
+ const float scale = float((1 << (N - 1)) - 1);
+
+ float round = (v >= 0 ? 0.5f : -0.5f);
+
+ v = (v >= -1) ? v : -1;
+ v = (v <= +1) ? v : +1;
+
+ return int(v * scale + round);
+}
+
+inline unsigned short meshopt_quantizeHalf(float v)
+{
+ union { float f; unsigned int ui; } u = {v};
+ unsigned int ui = u.ui;
+
+ int s = (ui >> 16) & 0x8000;
+ int em = ui & 0x7fffffff;
+
+ /* bias exponent and round to nearest; 112 is relative exponent bias (127-15) */
+ int h = (em - (112 << 23) + (1 << 12)) >> 13;
+
+ /* underflow: flush to zero; 113 encodes exponent -14 */
+ h = (em < (113 << 23)) ? 0 : h;
+
+ /* overflow: infinity; 143 encodes exponent 16 */
+ h = (em >= (143 << 23)) ? 0x7c00 : h;
+
+ /* NaN; note that we convert all types of NaN to qNaN */
+ h = (em > (255 << 23)) ? 0x7e00 : h;
+
+ return (unsigned short)(s | h);
+}
+
+inline float meshopt_quantizeFloat(float v, int N)
+{
+ union { float f; unsigned int ui; } u = {v};
+ unsigned int ui = u.ui;
+
+ const int mask = (1 << (23 - N)) - 1;
+ const int round = (1 << (23 - N)) >> 1;
+
+ int e = ui & 0x7f800000;
+ unsigned int rui = (ui + round) & ~mask;
+
+ /* round all numbers except inf/nan; this is important to make sure nan doesn't overflow into -0 */
+ ui = e == 0x7f800000 ? ui : rui;
+
+ /* flush denormals to zero */
+ ui = e == 0 ? 0 : ui;
+
+ u.ui = ui;
+ return u.f;
+}
+#endif
+
+/* Internal implementation helpers */
+#ifdef __cplusplus
+class meshopt_Allocator
+{
+public:
+ template <typename T>
+ struct StorageT
+ {
+ static void* (*allocate)(size_t);
+ static void (*deallocate)(void*);
+ };
+
+ typedef StorageT<void> Storage;
+
+ meshopt_Allocator()
+ : blocks()
+ , count(0)
+ {
+ }
+
+ ~meshopt_Allocator()
+ {
+ for (size_t i = count; i > 0; --i)
+ Storage::deallocate(blocks[i - 1]);
+ }
+
+ template <typename T> T* allocate(size_t size)
+ {
+ assert(count < sizeof(blocks) / sizeof(blocks[0]));
+ T* result = static_cast<T*>(Storage::allocate(size > size_t(-1) / sizeof(T) ? size_t(-1) : size * sizeof(T)));
+ blocks[count++] = result;
+ return result;
+ }
+
+private:
+ void* blocks[24];
+ size_t count;
+};
+
+// This makes sure that allocate/deallocate are lazily generated in translation units that need them and are deduplicated by the linker
+template <typename T> void* (*meshopt_Allocator::StorageT<T>::allocate)(size_t) = operator new;
+template <typename T> void (*meshopt_Allocator::StorageT<T>::deallocate)(void*) = operator delete;
+#endif
+
+/* Inline implementation for C++ templated wrappers */
+#if defined(__cplusplus) && !defined(MESHOPTIMIZER_NO_WRAPPERS)
+template <typename T, bool ZeroCopy = sizeof(T) == sizeof(unsigned int)>
+struct meshopt_IndexAdapter;
+
+template <typename T>
+struct meshopt_IndexAdapter<T, false>
+{
+ T* result;
+ unsigned int* data;
+ size_t count;
+
+ meshopt_IndexAdapter(T* result_, const T* input, size_t count_)
+ : result(result_)
+ , data(0)
+ , count(count_)
+ {
+ size_t size = count > size_t(-1) / sizeof(unsigned int) ? size_t(-1) : count * sizeof(unsigned int);
+
+ data = static_cast<unsigned int*>(meshopt_Allocator::Storage::allocate(size));
+
+ if (input)
+ {
+ for (size_t i = 0; i < count; ++i)
+ data[i] = input[i];
+ }
+ }
+
+ ~meshopt_IndexAdapter()
+ {
+ if (result)
+ {
+ for (size_t i = 0; i < count; ++i)
+ result[i] = T(data[i]);
+ }
+
+ meshopt_Allocator::Storage::deallocate(data);
+ }
+};
+
+template <typename T>
+struct meshopt_IndexAdapter<T, true>
+{
+ unsigned int* data;
+
+ meshopt_IndexAdapter(T* result, const T* input, size_t)
+ : data(reinterpret_cast<unsigned int*>(result ? result : const_cast<T*>(input)))
+ {
+ }
+};
+
+template <typename T>
+inline size_t meshopt_generateVertexRemap(unsigned int* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
+{
+ meshopt_IndexAdapter<T> in(0, indices, indices ? index_count : 0);
+
+ return meshopt_generateVertexRemap(destination, indices ? in.data : 0, index_count, vertices, vertex_count, vertex_size);
+}
+
+template <typename T>
+inline size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, indices ? index_count : 0);
+
+ return meshopt_generateVertexRemapMulti(destination, indices ? in.data : 0, index_count, vertex_count, streams, stream_count);
+}
+
+template <typename T>
+inline void meshopt_remapIndexBuffer(T* destination, const T* indices, size_t index_count, const unsigned int* remap)
+{
+ meshopt_IndexAdapter<T> in(0, indices, indices ? index_count : 0);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_remapIndexBuffer(out.data, indices ? in.data : 0, index_count, remap);
+}
+
+template <typename T>
+inline void meshopt_generateShadowIndexBuffer(T* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_generateShadowIndexBuffer(out.data, in.data, index_count, vertices, vertex_count, vertex_size, vertex_stride);
+}
+
+template <typename T>
+inline void meshopt_generateShadowIndexBufferMulti(T* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_generateShadowIndexBufferMulti(out.data, in.data, index_count, vertex_count, streams, stream_count);
+}
+
+template <typename T>
+inline void meshopt_optimizeVertexCache(T* destination, const T* indices, size_t index_count, size_t vertex_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_optimizeVertexCache(out.data, in.data, index_count, vertex_count);
+}
+
+template <typename T>
+inline void meshopt_optimizeVertexCacheStrip(T* destination, const T* indices, size_t index_count, size_t vertex_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_optimizeVertexCacheStrip(out.data, in.data, index_count, vertex_count);
+}
+
+template <typename T>
+inline void meshopt_optimizeVertexCacheFifo(T* destination, const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_optimizeVertexCacheFifo(out.data, in.data, index_count, vertex_count, cache_size);
+}
+
+template <typename T>
+inline void meshopt_optimizeOverdraw(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_optimizeOverdraw(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride, threshold);
+}
+
+template <typename T>
+inline size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_optimizeVertexFetchRemap(destination, in.data, index_count, vertex_count);
+}
+
+template <typename T>
+inline size_t meshopt_optimizeVertexFetch(void* destination, T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
+{
+ meshopt_IndexAdapter<T> inout(indices, indices, index_count);
+
+ return meshopt_optimizeVertexFetch(destination, inout.data, index_count, vertices, vertex_count, vertex_size);
+}
+
+template <typename T>
+inline size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_encodeIndexBuffer(buffer, buffer_size, in.data, index_count);
+}
+
+template <typename T>
+inline int meshopt_decodeIndexBuffer(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size)
+{
+ char index_size_valid[sizeof(T) == 2 || sizeof(T) == 4 ? 1 : -1];
+ (void)index_size_valid;
+
+ return meshopt_decodeIndexBuffer(destination, index_count, sizeof(T), buffer, buffer_size);
+}
+
+template <typename T>
+inline size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_encodeIndexSequence(buffer, buffer_size, in.data, index_count);
+}
+
+template <typename T>
+inline int meshopt_decodeIndexSequence(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size)
+{
+ char index_size_valid[sizeof(T) == 2 || sizeof(T) == 4 ? 1 : -1];
+ (void)index_size_valid;
+
+ return meshopt_decodeIndexSequence(destination, index_count, sizeof(T), buffer, buffer_size);
+}
+
+template <typename T>
+inline size_t meshopt_simplify(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ return meshopt_simplify(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride, target_index_count, target_error);
+}
+
+template <typename T>
+inline size_t meshopt_simplifySloppy(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, target_index_count);
+
+ return meshopt_simplifySloppy(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride, target_index_count);
+}
+
+template <typename T>
+inline size_t meshopt_stripify(T* destination, const T* indices, size_t index_count, size_t vertex_count, T restart_index)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, (index_count / 3) * 5);
+
+ return meshopt_stripify(out.data, in.data, index_count, vertex_count, unsigned(restart_index));
+}
+
+template <typename T>
+inline size_t meshopt_unstripify(T* destination, const T* indices, size_t index_count, T restart_index)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, (index_count - 2) * 3);
+
+ return meshopt_unstripify(out.data, in.data, index_count, unsigned(restart_index));
+}
+
+template <typename T>
+inline meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int buffer_size)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_analyzeVertexCache(in.data, index_count, vertex_count, cache_size, warp_size, buffer_size);
+}
+
+template <typename T>
+inline meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_analyzeOverdraw(in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride);
+}
+
+template <typename T>
+inline meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const T* indices, size_t index_count, size_t vertex_count, size_t vertex_size)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_analyzeVertexFetch(in.data, index_count, vertex_count, vertex_size);
+}
+
+template <typename T>
+inline size_t meshopt_buildMeshlets(meshopt_Meshlet* destination, const T* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_buildMeshlets(destination, in.data, index_count, vertex_count, max_vertices, max_triangles);
+}
+
+template <typename T>
+inline meshopt_Bounds meshopt_computeClusterBounds(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+
+ return meshopt_computeClusterBounds(in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride);
+}
+
+template <typename T>
+inline void meshopt_spatialSortTriangles(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ meshopt_IndexAdapter<T> in(0, indices, index_count);
+ meshopt_IndexAdapter<T> out(destination, 0, index_count);
+
+ meshopt_spatialSortTriangles(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride);
+}
+#endif
+
+/**
+ * Copyright (c) 2016-2020 Arseny Kapoulkine
+ *
+ * Permission is hereby granted, free of charge, to any person
+ * obtaining a copy of this software and associated documentation
+ * files (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use,
+ * copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following
+ * conditions:
+ *
+ * The above copyright notice and this permission notice shall be
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+ * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+ * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
diff --git a/thirdparty/meshoptimizer/overdrawanalyzer.cpp b/thirdparty/meshoptimizer/overdrawanalyzer.cpp
new file mode 100644
index 0000000000..8d5859ba39
--- /dev/null
+++ b/thirdparty/meshoptimizer/overdrawanalyzer.cpp
@@ -0,0 +1,230 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <float.h>
+#include <string.h>
+
+// This work is based on:
+// Nicolas Capens. Advanced Rasterization. 2004
+namespace meshopt
+{
+
+const int kViewport = 256;
+
+struct OverdrawBuffer
+{
+ float z[kViewport][kViewport][2];
+ unsigned int overdraw[kViewport][kViewport][2];
+};
+
+#ifndef min
+#define min(a, b) ((a) < (b) ? (a) : (b))
+#endif
+
+#ifndef max
+#define max(a, b) ((a) > (b) ? (a) : (b))
+#endif
+
+static float computeDepthGradients(float& dzdx, float& dzdy, float x1, float y1, float z1, float x2, float y2, float z2, float x3, float y3, float z3)
+{
+ // z2 = z1 + dzdx * (x2 - x1) + dzdy * (y2 - y1)
+ // z3 = z1 + dzdx * (x3 - x1) + dzdy * (y3 - y1)
+ // (x2-x1 y2-y1)(dzdx) = (z2-z1)
+ // (x3-x1 y3-y1)(dzdy) (z3-z1)
+ // we'll solve it with Cramer's rule
+ float det = (x2 - x1) * (y3 - y1) - (y2 - y1) * (x3 - x1);
+ float invdet = (det == 0) ? 0 : 1 / det;
+
+ dzdx = (z2 - z1) * (y3 - y1) - (y2 - y1) * (z3 - z1) * invdet;
+ dzdy = (x2 - x1) * (z3 - z1) - (z2 - z1) * (x3 - x1) * invdet;
+
+ return det;
+}
+
+// half-space fixed point triangle rasterizer
+static void rasterize(OverdrawBuffer* buffer, float v1x, float v1y, float v1z, float v2x, float v2y, float v2z, float v3x, float v3y, float v3z)
+{
+ // compute depth gradients
+ float DZx, DZy;
+ float det = computeDepthGradients(DZx, DZy, v1x, v1y, v1z, v2x, v2y, v2z, v3x, v3y, v3z);
+ int sign = det > 0;
+
+ // flip backfacing triangles to simplify rasterization logic
+ if (sign)
+ {
+ // flipping v2 & v3 preserves depth gradients since they're based on v1
+ float t;
+ t = v2x, v2x = v3x, v3x = t;
+ t = v2y, v2y = v3y, v3y = t;
+ t = v2z, v2z = v3z, v3z = t;
+
+ // flip depth since we rasterize backfacing triangles to second buffer with reverse Z; only v1z is used below
+ v1z = kViewport - v1z;
+ DZx = -DZx;
+ DZy = -DZy;
+ }
+
+ // coordinates, 28.4 fixed point
+ int X1 = int(16.0f * v1x + 0.5f);
+ int X2 = int(16.0f * v2x + 0.5f);
+ int X3 = int(16.0f * v3x + 0.5f);
+
+ int Y1 = int(16.0f * v1y + 0.5f);
+ int Y2 = int(16.0f * v2y + 0.5f);
+ int Y3 = int(16.0f * v3y + 0.5f);
+
+ // bounding rectangle, clipped against viewport
+ // since we rasterize pixels with covered centers, min >0.5 should round up
+ // as for max, due to top-left filling convention we will never rasterize right/bottom edges
+ // so max >= 0.5 should round down
+ int minx = max((min(X1, min(X2, X3)) + 7) >> 4, 0);
+ int maxx = min((max(X1, max(X2, X3)) + 7) >> 4, kViewport);
+ int miny = max((min(Y1, min(Y2, Y3)) + 7) >> 4, 0);
+ int maxy = min((max(Y1, max(Y2, Y3)) + 7) >> 4, kViewport);
+
+ // deltas, 28.4 fixed point
+ int DX12 = X1 - X2;
+ int DX23 = X2 - X3;
+ int DX31 = X3 - X1;
+
+ int DY12 = Y1 - Y2;
+ int DY23 = Y2 - Y3;
+ int DY31 = Y3 - Y1;
+
+ // fill convention correction
+ int TL1 = DY12 < 0 || (DY12 == 0 && DX12 > 0);
+ int TL2 = DY23 < 0 || (DY23 == 0 && DX23 > 0);
+ int TL3 = DY31 < 0 || (DY31 == 0 && DX31 > 0);
+
+ // half edge equations, 24.8 fixed point
+ // note that we offset minx/miny by half pixel since we want to rasterize pixels with covered centers
+ int FX = (minx << 4) + 8;
+ int FY = (miny << 4) + 8;
+ int CY1 = DX12 * (FY - Y1) - DY12 * (FX - X1) + TL1 - 1;
+ int CY2 = DX23 * (FY - Y2) - DY23 * (FX - X2) + TL2 - 1;
+ int CY3 = DX31 * (FY - Y3) - DY31 * (FX - X3) + TL3 - 1;
+ float ZY = v1z + (DZx * float(FX - X1) + DZy * float(FY - Y1)) * (1 / 16.f);
+
+ for (int y = miny; y < maxy; y++)
+ {
+ int CX1 = CY1;
+ int CX2 = CY2;
+ int CX3 = CY3;
+ float ZX = ZY;
+
+ for (int x = minx; x < maxx; x++)
+ {
+ // check if all CXn are non-negative
+ if ((CX1 | CX2 | CX3) >= 0)
+ {
+ if (ZX >= buffer->z[y][x][sign])
+ {
+ buffer->z[y][x][sign] = ZX;
+ buffer->overdraw[y][x][sign]++;
+ }
+ }
+
+ // signed left shift is UB for negative numbers so use unsigned-signed casts
+ CX1 -= int(unsigned(DY12) << 4);
+ CX2 -= int(unsigned(DY23) << 4);
+ CX3 -= int(unsigned(DY31) << 4);
+ ZX += DZx;
+ }
+
+ // signed left shift is UB for negative numbers so use unsigned-signed casts
+ CY1 += int(unsigned(DX12) << 4);
+ CY2 += int(unsigned(DX23) << 4);
+ CY3 += int(unsigned(DX31) << 4);
+ ZY += DZy;
+ }
+}
+
+} // namespace meshopt
+
+meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+
+ meshopt_Allocator allocator;
+
+ size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
+
+ meshopt_OverdrawStatistics result = {};
+
+ float minv[3] = {FLT_MAX, FLT_MAX, FLT_MAX};
+ float maxv[3] = {-FLT_MAX, -FLT_MAX, -FLT_MAX};
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ const float* v = vertex_positions + i * vertex_stride_float;
+
+ for (int j = 0; j < 3; ++j)
+ {
+ minv[j] = min(minv[j], v[j]);
+ maxv[j] = max(maxv[j], v[j]);
+ }
+ }
+
+ float extent = max(maxv[0] - minv[0], max(maxv[1] - minv[1], maxv[2] - minv[2]));
+ float scale = kViewport / extent;
+
+ float* triangles = allocator.allocate<float>(index_count * 3);
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices[i];
+ assert(index < vertex_count);
+
+ const float* v = vertex_positions + index * vertex_stride_float;
+
+ triangles[i * 3 + 0] = (v[0] - minv[0]) * scale;
+ triangles[i * 3 + 1] = (v[1] - minv[1]) * scale;
+ triangles[i * 3 + 2] = (v[2] - minv[2]) * scale;
+ }
+
+ OverdrawBuffer* buffer = allocator.allocate<OverdrawBuffer>(1);
+
+ for (int axis = 0; axis < 3; ++axis)
+ {
+ memset(buffer, 0, sizeof(OverdrawBuffer));
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ const float* vn0 = &triangles[3 * (i + 0)];
+ const float* vn1 = &triangles[3 * (i + 1)];
+ const float* vn2 = &triangles[3 * (i + 2)];
+
+ switch (axis)
+ {
+ case 0:
+ rasterize(buffer, vn0[2], vn0[1], vn0[0], vn1[2], vn1[1], vn1[0], vn2[2], vn2[1], vn2[0]);
+ break;
+ case 1:
+ rasterize(buffer, vn0[0], vn0[2], vn0[1], vn1[0], vn1[2], vn1[1], vn2[0], vn2[2], vn2[1]);
+ break;
+ case 2:
+ rasterize(buffer, vn0[1], vn0[0], vn0[2], vn1[1], vn1[0], vn1[2], vn2[1], vn2[0], vn2[2]);
+ break;
+ }
+ }
+
+ for (int y = 0; y < kViewport; ++y)
+ for (int x = 0; x < kViewport; ++x)
+ for (int s = 0; s < 2; ++s)
+ {
+ unsigned int overdraw = buffer->overdraw[y][x][s];
+
+ result.pixels_covered += overdraw > 0;
+ result.pixels_shaded += overdraw;
+ }
+ }
+
+ result.overdraw = result.pixels_covered ? float(result.pixels_shaded) / float(result.pixels_covered) : 0.f;
+
+ return result;
+}
diff --git a/thirdparty/meshoptimizer/overdrawoptimizer.cpp b/thirdparty/meshoptimizer/overdrawoptimizer.cpp
new file mode 100644
index 0000000000..143656ed76
--- /dev/null
+++ b/thirdparty/meshoptimizer/overdrawoptimizer.cpp
@@ -0,0 +1,333 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <math.h>
+#include <string.h>
+
+// This work is based on:
+// Pedro Sander, Diego Nehab and Joshua Barczak. Fast Triangle Reordering for Vertex Locality and Reduced Overdraw. 2007
+namespace meshopt
+{
+
+static void calculateSortData(float* sort_data, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_positions_stride, const unsigned int* clusters, size_t cluster_count)
+{
+ size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
+
+ float mesh_centroid[3] = {};
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ const float* p = vertex_positions + vertex_stride_float * indices[i];
+
+ mesh_centroid[0] += p[0];
+ mesh_centroid[1] += p[1];
+ mesh_centroid[2] += p[2];
+ }
+
+ mesh_centroid[0] /= index_count;
+ mesh_centroid[1] /= index_count;
+ mesh_centroid[2] /= index_count;
+
+ for (size_t cluster = 0; cluster < cluster_count; ++cluster)
+ {
+ size_t cluster_begin = clusters[cluster] * 3;
+ size_t cluster_end = (cluster + 1 < cluster_count) ? clusters[cluster + 1] * 3 : index_count;
+ assert(cluster_begin < cluster_end);
+
+ float cluster_area = 0;
+ float cluster_centroid[3] = {};
+ float cluster_normal[3] = {};
+
+ for (size_t i = cluster_begin; i < cluster_end; i += 3)
+ {
+ const float* p0 = vertex_positions + vertex_stride_float * indices[i + 0];
+ const float* p1 = vertex_positions + vertex_stride_float * indices[i + 1];
+ const float* p2 = vertex_positions + vertex_stride_float * indices[i + 2];
+
+ float p10[3] = {p1[0] - p0[0], p1[1] - p0[1], p1[2] - p0[2]};
+ float p20[3] = {p2[0] - p0[0], p2[1] - p0[1], p2[2] - p0[2]};
+
+ float normalx = p10[1] * p20[2] - p10[2] * p20[1];
+ float normaly = p10[2] * p20[0] - p10[0] * p20[2];
+ float normalz = p10[0] * p20[1] - p10[1] * p20[0];
+
+ float area = sqrtf(normalx * normalx + normaly * normaly + normalz * normalz);
+
+ cluster_centroid[0] += (p0[0] + p1[0] + p2[0]) * (area / 3);
+ cluster_centroid[1] += (p0[1] + p1[1] + p2[1]) * (area / 3);
+ cluster_centroid[2] += (p0[2] + p1[2] + p2[2]) * (area / 3);
+ cluster_normal[0] += normalx;
+ cluster_normal[1] += normaly;
+ cluster_normal[2] += normalz;
+ cluster_area += area;
+ }
+
+ float inv_cluster_area = cluster_area == 0 ? 0 : 1 / cluster_area;
+
+ cluster_centroid[0] *= inv_cluster_area;
+ cluster_centroid[1] *= inv_cluster_area;
+ cluster_centroid[2] *= inv_cluster_area;
+
+ float cluster_normal_length = sqrtf(cluster_normal[0] * cluster_normal[0] + cluster_normal[1] * cluster_normal[1] + cluster_normal[2] * cluster_normal[2]);
+ float inv_cluster_normal_length = cluster_normal_length == 0 ? 0 : 1 / cluster_normal_length;
+
+ cluster_normal[0] *= inv_cluster_normal_length;
+ cluster_normal[1] *= inv_cluster_normal_length;
+ cluster_normal[2] *= inv_cluster_normal_length;
+
+ float centroid_vector[3] = {cluster_centroid[0] - mesh_centroid[0], cluster_centroid[1] - mesh_centroid[1], cluster_centroid[2] - mesh_centroid[2]};
+
+ sort_data[cluster] = centroid_vector[0] * cluster_normal[0] + centroid_vector[1] * cluster_normal[1] + centroid_vector[2] * cluster_normal[2];
+ }
+}
+
+static void calculateSortOrderRadix(unsigned int* sort_order, const float* sort_data, unsigned short* sort_keys, size_t cluster_count)
+{
+ // compute sort data bounds and renormalize, using fixed point snorm
+ float sort_data_max = 1e-3f;
+
+ for (size_t i = 0; i < cluster_count; ++i)
+ {
+ float dpa = fabsf(sort_data[i]);
+
+ sort_data_max = (sort_data_max < dpa) ? dpa : sort_data_max;
+ }
+
+ const int sort_bits = 11;
+
+ for (size_t i = 0; i < cluster_count; ++i)
+ {
+ // note that we flip distribution since high dot product should come first
+ float sort_key = 0.5f - 0.5f * (sort_data[i] / sort_data_max);
+
+ sort_keys[i] = meshopt_quantizeUnorm(sort_key, sort_bits) & ((1 << sort_bits) - 1);
+ }
+
+ // fill histogram for counting sort
+ unsigned int histogram[1 << sort_bits];
+ memset(histogram, 0, sizeof(histogram));
+
+ for (size_t i = 0; i < cluster_count; ++i)
+ {
+ histogram[sort_keys[i]]++;
+ }
+
+ // compute offsets based on histogram data
+ size_t histogram_sum = 0;
+
+ for (size_t i = 0; i < 1 << sort_bits; ++i)
+ {
+ size_t count = histogram[i];
+ histogram[i] = unsigned(histogram_sum);
+ histogram_sum += count;
+ }
+
+ assert(histogram_sum == cluster_count);
+
+ // compute sort order based on offsets
+ for (size_t i = 0; i < cluster_count; ++i)
+ {
+ sort_order[histogram[sort_keys[i]]++] = unsigned(i);
+ }
+}
+
+static unsigned int updateCache(unsigned int a, unsigned int b, unsigned int c, unsigned int cache_size, unsigned int* cache_timestamps, unsigned int& timestamp)
+{
+ unsigned int cache_misses = 0;
+
+ // if vertex is not in cache, put it in cache
+ if (timestamp - cache_timestamps[a] > cache_size)
+ {
+ cache_timestamps[a] = timestamp++;
+ cache_misses++;
+ }
+
+ if (timestamp - cache_timestamps[b] > cache_size)
+ {
+ cache_timestamps[b] = timestamp++;
+ cache_misses++;
+ }
+
+ if (timestamp - cache_timestamps[c] > cache_size)
+ {
+ cache_timestamps[c] = timestamp++;
+ cache_misses++;
+ }
+
+ return cache_misses;
+}
+
+static size_t generateHardBoundaries(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int* cache_timestamps)
+{
+ memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
+
+ unsigned int timestamp = cache_size + 1;
+
+ size_t face_count = index_count / 3;
+
+ size_t result = 0;
+
+ for (size_t i = 0; i < face_count; ++i)
+ {
+ unsigned int m = updateCache(indices[i * 3 + 0], indices[i * 3 + 1], indices[i * 3 + 2], cache_size, &cache_timestamps[0], timestamp);
+
+ // when all three vertices are not in the cache it's usually relatively safe to assume that this is a new patch in the mesh
+ // that is disjoint from previous vertices; sometimes it might come back to reference existing vertices but that frequently
+ // suggests an inefficiency in the vertex cache optimization algorithm
+ // usually the first triangle has 3 misses unless it's degenerate - thus we make sure the first cluster always starts with 0
+ if (i == 0 || m == 3)
+ {
+ destination[result++] = unsigned(i);
+ }
+ }
+
+ assert(result <= index_count / 3);
+
+ return result;
+}
+
+static size_t generateSoftBoundaries(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const unsigned int* clusters, size_t cluster_count, unsigned int cache_size, float threshold, unsigned int* cache_timestamps)
+{
+ memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
+
+ unsigned int timestamp = 0;
+
+ size_t result = 0;
+
+ for (size_t it = 0; it < cluster_count; ++it)
+ {
+ size_t start = clusters[it];
+ size_t end = (it + 1 < cluster_count) ? clusters[it + 1] : index_count / 3;
+ assert(start < end);
+
+ // reset cache
+ timestamp += cache_size + 1;
+
+ // measure cluster ACMR
+ unsigned int cluster_misses = 0;
+
+ for (size_t i = start; i < end; ++i)
+ {
+ unsigned int m = updateCache(indices[i * 3 + 0], indices[i * 3 + 1], indices[i * 3 + 2], cache_size, &cache_timestamps[0], timestamp);
+
+ cluster_misses += m;
+ }
+
+ float cluster_threshold = threshold * (float(cluster_misses) / float(end - start));
+
+ // first cluster always starts from the hard cluster boundary
+ destination[result++] = unsigned(start);
+
+ // reset cache
+ timestamp += cache_size + 1;
+
+ unsigned int running_misses = 0;
+ unsigned int running_faces = 0;
+
+ for (size_t i = start; i < end; ++i)
+ {
+ unsigned int m = updateCache(indices[i * 3 + 0], indices[i * 3 + 1], indices[i * 3 + 2], cache_size, &cache_timestamps[0], timestamp);
+
+ running_misses += m;
+ running_faces += 1;
+
+ if (float(running_misses) / float(running_faces) <= cluster_threshold)
+ {
+ // we have reached the target ACMR with the current triangle so we need to start a new cluster on the next one
+ // note that this may mean that we add 'end` to destination for the last triangle, which will imply that the last
+ // cluster is empty; however, the 'pop_back' after the loop will clean it up
+ destination[result++] = unsigned(i + 1);
+
+ // reset cache
+ timestamp += cache_size + 1;
+
+ running_misses = 0;
+ running_faces = 0;
+ }
+ }
+
+ // each time we reach the target ACMR we flush the cluster
+ // this means that the last cluster is by definition not very good - there are frequent cases where we are left with a few triangles
+ // in the last cluster, producing a very bad ACMR and significantly penalizing the overall results
+ // thus we remove the last cluster boundary, merging the last complete cluster with the last incomplete one
+ // there are sometimes cases when the last cluster is actually good enough - in which case the code above would have added 'end'
+ // to the cluster boundary array which we need to remove anyway - this code will do that automatically
+ if (destination[result - 1] != start)
+ {
+ result--;
+ }
+ }
+
+ assert(result >= cluster_count);
+ assert(result <= index_count / 3);
+
+ return result;
+}
+
+} // namespace meshopt
+
+void meshopt_optimizeOverdraw(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+
+ meshopt_Allocator allocator;
+
+ // guard for empty meshes
+ if (index_count == 0 || vertex_count == 0)
+ return;
+
+ // support in-place optimization
+ if (destination == indices)
+ {
+ unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
+ memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
+ indices = indices_copy;
+ }
+
+ unsigned int cache_size = 16;
+
+ unsigned int* cache_timestamps = allocator.allocate<unsigned int>(vertex_count);
+
+ // generate hard boundaries from full-triangle cache misses
+ unsigned int* hard_clusters = allocator.allocate<unsigned int>(index_count / 3);
+ size_t hard_cluster_count = generateHardBoundaries(hard_clusters, indices, index_count, vertex_count, cache_size, cache_timestamps);
+
+ // generate soft boundaries
+ unsigned int* soft_clusters = allocator.allocate<unsigned int>(index_count / 3 + 1);
+ size_t soft_cluster_count = generateSoftBoundaries(soft_clusters, indices, index_count, vertex_count, hard_clusters, hard_cluster_count, cache_size, threshold, cache_timestamps);
+
+ const unsigned int* clusters = soft_clusters;
+ size_t cluster_count = soft_cluster_count;
+
+ // fill sort data
+ float* sort_data = allocator.allocate<float>(cluster_count);
+ calculateSortData(sort_data, indices, index_count, vertex_positions, vertex_positions_stride, clusters, cluster_count);
+
+ // sort clusters using sort data
+ unsigned short* sort_keys = allocator.allocate<unsigned short>(cluster_count);
+ unsigned int* sort_order = allocator.allocate<unsigned int>(cluster_count);
+ calculateSortOrderRadix(sort_order, sort_data, sort_keys, cluster_count);
+
+ // fill output buffer
+ size_t offset = 0;
+
+ for (size_t it = 0; it < cluster_count; ++it)
+ {
+ unsigned int cluster = sort_order[it];
+ assert(cluster < cluster_count);
+
+ size_t cluster_begin = clusters[cluster] * 3;
+ size_t cluster_end = (cluster + 1 < cluster_count) ? clusters[cluster + 1] * 3 : index_count;
+ assert(cluster_begin < cluster_end);
+
+ memcpy(destination + offset, indices + cluster_begin, (cluster_end - cluster_begin) * sizeof(unsigned int));
+ offset += cluster_end - cluster_begin;
+ }
+
+ assert(offset == index_count);
+}
diff --git a/thirdparty/meshoptimizer/patches/simplifier_get_resulting_error.patch b/thirdparty/meshoptimizer/patches/simplifier_get_resulting_error.patch
new file mode 100644
index 0000000000..1be38e45d2
--- /dev/null
+++ b/thirdparty/meshoptimizer/patches/simplifier_get_resulting_error.patch
@@ -0,0 +1,96 @@
+diff --git a/thirdparty/meshoptimizer/meshoptimizer.h b/thirdparty/meshoptimizer/meshoptimizer.h
+index a442d103c8..fde00f9c82 100644
+--- a/thirdparty/meshoptimizer/meshoptimizer.h
++++ b/thirdparty/meshoptimizer/meshoptimizer.h
+@@ -266,7 +266,10 @@ MESHOPTIMIZER_EXPERIMENTAL void meshopt_decodeFilterExp(void* buffer, size_t ver
+ * destination must contain enough space for the *source* index buffer (since optimization is iterative, this means index_count elements - *not* target_index_count!)
+ * vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
+ */
+-MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error);
++// -- GODOT start --
++//MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error);
++MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplify(unsigned int *destination, const unsigned int *indices, size_t index_count, const float *vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error, float *r_resulting_error);
++// -- GODOT end --
+
+ /**
+ * Experimental: Mesh simplifier (sloppy)
+diff --git a/thirdparty/meshoptimizer/simplifier.cpp b/thirdparty/meshoptimizer/simplifier.cpp
+index bd523275ce..51cf634186 100644
+--- a/thirdparty/meshoptimizer/simplifier.cpp
++++ b/thirdparty/meshoptimizer/simplifier.cpp
+@@ -1143,7 +1143,10 @@ unsigned int* meshopt_simplifyDebugLoop = 0;
+ unsigned int* meshopt_simplifyDebugLoopBack = 0;
+ #endif
+
+-size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error)
++// -- GODOT start --
++//size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error)
++size_t meshopt_simplify(unsigned int *destination, const unsigned int *indices, size_t index_count, const float *vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error, float *r_resulting_error)
++// -- GODOT end --
+ {
+ using namespace meshopt;
+
+@@ -1198,10 +1201,13 @@ size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices,
+ if (result != indices)
+ memcpy(result, indices, index_count * sizeof(unsigned int));
+
++// -- GODOT start --
+ #if TRACE
+ size_t pass_count = 0;
+- float worst_error = 0;
++ //float worst_error = 0;
+ #endif
++ float worst_error = 0;
++// -- GODOT end --
+
+ Collapse* edge_collapses = allocator.allocate<Collapse>(index_count);
+ unsigned int* collapse_order = allocator.allocate<unsigned int>(index_count);
+@@ -1213,6 +1219,12 @@ size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices,
+ // target_error input is linear; we need to adjust it to match quadricError units
+ float error_limit = target_error * target_error;
+
++// -- GODOT start --
++ if (r_resulting_error) {
++ *r_resulting_error = 1.0;
++ }
++// -- GODOT end --
++
+ while (result_count > target_index_count)
+ {
+ size_t edge_collapse_count = pickEdgeCollapses(edge_collapses, result, result_count, remap, vertex_kind, loop);
+@@ -1257,7 +1269,8 @@ size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices,
+ size_t new_count = remapIndexBuffer(result, result_count, collapse_remap);
+ assert(new_count < result_count);
+
+-#if TRACE
++// -- GODOT start --
++//#if TRACE
+ float pass_error = 0.f;
+ for (size_t i = 0; i < edge_collapse_count; ++i)
+ {
+@@ -1267,15 +1280,24 @@ size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices,
+ pass_error = c.error;
+ }
+
+- pass_count++;
++ //pass_count++;
+ worst_error = (worst_error < pass_error) ? pass_error : worst_error;
+
++#if TRACE
++ pass_count++;
+ printf("pass %d: triangles: %d -> %d, collapses: %d/%d (goal: %d), error: %e (limit %e goal %e)\n", int(pass_count), int(result_count / 3), int(new_count / 3), int(collapses), int(edge_collapse_count), int(edge_collapse_goal), pass_error, error_limit, error_goal);
+ #endif
++// -- GODOT end --
+
+ result_count = new_count;
+ }
+
++// -- GODOT start --
++ if (r_resulting_error) {
++ *r_resulting_error = sqrt(worst_error);
++ }
++// -- GODOT end --
++
+ #if TRACE
+ printf("passes: %d, worst error: %e\n", int(pass_count), worst_error);
+ #endif
diff --git a/thirdparty/meshoptimizer/simplifier.cpp b/thirdparty/meshoptimizer/simplifier.cpp
new file mode 100644
index 0000000000..b195a8cb5d
--- /dev/null
+++ b/thirdparty/meshoptimizer/simplifier.cpp
@@ -0,0 +1,1562 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <float.h>
+#include <math.h>
+#include <string.h>
+
+
+#ifndef TRACE
+#define TRACE 0
+#endif
+
+#if TRACE
+#include <stdio.h>
+#endif
+
+// This work is based on:
+// Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. 1997
+// Michael Garland. Quadric-based polygonal surface simplification. 1999
+// Peter Lindstrom. Out-of-Core Simplification of Large Polygonal Models. 2000
+// Matthias Teschner, Bruno Heidelberger, Matthias Mueller, Danat Pomeranets, Markus Gross. Optimized Spatial Hashing for Collision Detection of Deformable Objects. 2003
+// Peter Van Sandt, Yannis Chronis, Jignesh M. Patel. Efficiently Searching In-Memory Sorted Arrays: Revenge of the Interpolation Search? 2019
+namespace meshopt
+{
+
+struct EdgeAdjacency
+{
+ unsigned int* counts;
+ unsigned int* offsets;
+ unsigned int* data;
+};
+
+static void buildEdgeAdjacency(EdgeAdjacency& adjacency, const unsigned int* indices, size_t index_count, size_t vertex_count, meshopt_Allocator& allocator)
+{
+ size_t face_count = index_count / 3;
+
+ // allocate arrays
+ adjacency.counts = allocator.allocate<unsigned int>(vertex_count);
+ adjacency.offsets = allocator.allocate<unsigned int>(vertex_count);
+ adjacency.data = allocator.allocate<unsigned int>(index_count);
+
+ // fill edge counts
+ memset(adjacency.counts, 0, vertex_count * sizeof(unsigned int));
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ assert(indices[i] < vertex_count);
+
+ adjacency.counts[indices[i]]++;
+ }
+
+ // fill offset table
+ unsigned int offset = 0;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ adjacency.offsets[i] = offset;
+ offset += adjacency.counts[i];
+ }
+
+ assert(offset == index_count);
+
+ // fill edge data
+ for (size_t i = 0; i < face_count; ++i)
+ {
+ unsigned int a = indices[i * 3 + 0], b = indices[i * 3 + 1], c = indices[i * 3 + 2];
+
+ adjacency.data[adjacency.offsets[a]++] = b;
+ adjacency.data[adjacency.offsets[b]++] = c;
+ adjacency.data[adjacency.offsets[c]++] = a;
+ }
+
+ // fix offsets that have been disturbed by the previous pass
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ assert(adjacency.offsets[i] >= adjacency.counts[i]);
+
+ adjacency.offsets[i] -= adjacency.counts[i];
+ }
+}
+
+struct PositionHasher
+{
+ const float* vertex_positions;
+ size_t vertex_stride_float;
+
+ size_t hash(unsigned int index) const
+ {
+ const unsigned int* key = reinterpret_cast<const unsigned int*>(vertex_positions + index * vertex_stride_float);
+
+ // Optimized Spatial Hashing for Collision Detection of Deformable Objects
+ return (key[0] * 73856093) ^ (key[1] * 19349663) ^ (key[2] * 83492791);
+ }
+
+ bool equal(unsigned int lhs, unsigned int rhs) const
+ {
+ return memcmp(vertex_positions + lhs * vertex_stride_float, vertex_positions + rhs * vertex_stride_float, sizeof(float) * 3) == 0;
+ }
+};
+
+static size_t hashBuckets2(size_t count)
+{
+ size_t buckets = 1;
+ while (buckets < count)
+ buckets *= 2;
+
+ return buckets;
+}
+
+template <typename T, typename Hash>
+static T* hashLookup2(T* table, size_t buckets, const Hash& hash, const T& key, const T& empty)
+{
+ assert(buckets > 0);
+ assert((buckets & (buckets - 1)) == 0);
+
+ size_t hashmod = buckets - 1;
+ size_t bucket = hash.hash(key) & hashmod;
+
+ for (size_t probe = 0; probe <= hashmod; ++probe)
+ {
+ T& item = table[bucket];
+
+ if (item == empty)
+ return &item;
+
+ if (hash.equal(item, key))
+ return &item;
+
+ // hash collision, quadratic probing
+ bucket = (bucket + probe + 1) & hashmod;
+ }
+
+ assert(false && "Hash table is full"); // unreachable
+ return 0;
+}
+
+static void buildPositionRemap(unsigned int* remap, unsigned int* wedge, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, meshopt_Allocator& allocator)
+{
+ PositionHasher hasher = {vertex_positions_data, vertex_positions_stride / sizeof(float)};
+
+ size_t table_size = hashBuckets2(vertex_count);
+ unsigned int* table = allocator.allocate<unsigned int>(table_size);
+ memset(table, -1, table_size * sizeof(unsigned int));
+
+ // build forward remap: for each vertex, which other (canonical) vertex does it map to?
+ // we use position equivalence for this, and remap vertices to other existing vertices
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ unsigned int index = unsigned(i);
+ unsigned int* entry = hashLookup2(table, table_size, hasher, index, ~0u);
+
+ if (*entry == ~0u)
+ *entry = index;
+
+ remap[index] = *entry;
+ }
+
+ // build wedge table: for each vertex, which other vertex is the next wedge that also maps to the same vertex?
+ // entries in table form a (cyclic) wedge loop per vertex; for manifold vertices, wedge[i] == remap[i] == i
+ for (size_t i = 0; i < vertex_count; ++i)
+ wedge[i] = unsigned(i);
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ if (remap[i] != i)
+ {
+ unsigned int r = remap[i];
+
+ wedge[i] = wedge[r];
+ wedge[r] = unsigned(i);
+ }
+}
+
+enum VertexKind
+{
+ Kind_Manifold, // not on an attribute seam, not on any boundary
+ Kind_Border, // not on an attribute seam, has exactly two open edges
+ Kind_Seam, // on an attribute seam with exactly two attribute seam edges
+ Kind_Complex, // none of the above; these vertices can move as long as all wedges move to the target vertex
+ Kind_Locked, // none of the above; these vertices can't move
+
+ Kind_Count
+};
+
+// manifold vertices can collapse onto anything
+// border/seam vertices can only be collapsed onto border/seam respectively
+// complex vertices can collapse onto complex/locked
+// a rule of thumb is that collapsing kind A into kind B preserves the kind B in the target vertex
+// for example, while we could collapse Complex into Manifold, this would mean the target vertex isn't Manifold anymore
+const unsigned char kCanCollapse[Kind_Count][Kind_Count] = {
+ {1, 1, 1, 1, 1},
+ {0, 1, 0, 0, 0},
+ {0, 0, 1, 0, 0},
+ {0, 0, 0, 1, 1},
+ {0, 0, 0, 0, 0},
+};
+
+// if a vertex is manifold or seam, adjoining edges are guaranteed to have an opposite edge
+// note that for seam edges, the opposite edge isn't present in the attribute-based topology
+// but is present if you consider a position-only mesh variant
+const unsigned char kHasOpposite[Kind_Count][Kind_Count] = {
+ {1, 1, 1, 0, 1},
+ {1, 0, 1, 0, 0},
+ {1, 1, 1, 0, 1},
+ {0, 0, 0, 0, 0},
+ {1, 0, 1, 0, 0},
+};
+
+static bool hasEdge(const EdgeAdjacency& adjacency, unsigned int a, unsigned int b)
+{
+ unsigned int count = adjacency.counts[a];
+ const unsigned int* data = adjacency.data + adjacency.offsets[a];
+
+ for (size_t i = 0; i < count; ++i)
+ if (data[i] == b)
+ return true;
+
+ return false;
+}
+
+static void classifyVertices(unsigned char* result, unsigned int* loop, unsigned int* loopback, size_t vertex_count, const EdgeAdjacency& adjacency, const unsigned int* remap, const unsigned int* wedge)
+{
+ memset(loop, -1, vertex_count * sizeof(unsigned int));
+ memset(loopback, -1, vertex_count * sizeof(unsigned int));
+
+ // incoming & outgoing open edges: ~0u if no open edges, i if there are more than 1
+ // note that this is the same data as required in loop[] arrays; loop[] data is only valid for border/seam
+ // but here it's okay to fill the data out for other types of vertices as well
+ unsigned int* openinc = loopback;
+ unsigned int* openout = loop;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ unsigned int vertex = unsigned(i);
+
+ unsigned int count = adjacency.counts[vertex];
+ const unsigned int* data = adjacency.data + adjacency.offsets[vertex];
+
+ for (size_t j = 0; j < count; ++j)
+ {
+ unsigned int target = data[j];
+
+ if (!hasEdge(adjacency, target, vertex))
+ {
+ openinc[target] = (openinc[target] == ~0u) ? vertex : target;
+ openout[vertex] = (openout[vertex] == ~0u) ? target : vertex;
+ }
+ }
+ }
+
+#if TRACE
+ size_t lockedstats[4] = {};
+#define TRACELOCKED(i) lockedstats[i]++;
+#else
+#define TRACELOCKED(i) (void)0
+#endif
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ if (remap[i] == i)
+ {
+ if (wedge[i] == i)
+ {
+ // no attribute seam, need to check if it's manifold
+ unsigned int openi = openinc[i], openo = openout[i];
+
+ // note: we classify any vertices with no open edges as manifold
+ // this is technically incorrect - if 4 triangles share an edge, we'll classify vertices as manifold
+ // it's unclear if this is a problem in practice
+ if (openi == ~0u && openo == ~0u)
+ {
+ result[i] = Kind_Manifold;
+ }
+ else if (openi != i && openo != i)
+ {
+ result[i] = Kind_Border;
+ }
+ else
+ {
+ result[i] = Kind_Locked;
+ TRACELOCKED(0);
+ }
+ }
+ else if (wedge[wedge[i]] == i)
+ {
+ // attribute seam; need to distinguish between Seam and Locked
+ unsigned int w = wedge[i];
+ unsigned int openiv = openinc[i], openov = openout[i];
+ unsigned int openiw = openinc[w], openow = openout[w];
+
+ // seam should have one open half-edge for each vertex, and the edges need to "connect" - point to the same vertex post-remap
+ if (openiv != ~0u && openiv != i && openov != ~0u && openov != i &&
+ openiw != ~0u && openiw != w && openow != ~0u && openow != w)
+ {
+ if (remap[openiv] == remap[openow] && remap[openov] == remap[openiw])
+ {
+ result[i] = Kind_Seam;
+ }
+ else
+ {
+ result[i] = Kind_Locked;
+ TRACELOCKED(1);
+ }
+ }
+ else
+ {
+ result[i] = Kind_Locked;
+ TRACELOCKED(2);
+ }
+ }
+ else
+ {
+ // more than one vertex maps to this one; we don't have classification available
+ result[i] = Kind_Locked;
+ TRACELOCKED(3);
+ }
+ }
+ else
+ {
+ assert(remap[i] < i);
+
+ result[i] = result[remap[i]];
+ }
+ }
+
+#if TRACE
+ printf("locked: many open edges %d, disconnected seam %d, many seam edges %d, many wedges %d\n",
+ int(lockedstats[0]), int(lockedstats[1]), int(lockedstats[2]), int(lockedstats[3]));
+#endif
+}
+
+struct Vector3
+{
+ float x, y, z;
+};
+// -- GODOT start --
+//static void rescalePositions(Vector3* result, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride)
+static float rescalePositions(Vector3* result, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride)
+// -- GODOT end --
+
+{
+ size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
+
+ float minv[3] = {FLT_MAX, FLT_MAX, FLT_MAX};
+ float maxv[3] = {-FLT_MAX, -FLT_MAX, -FLT_MAX};
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ const float* v = vertex_positions_data + i * vertex_stride_float;
+
+ result[i].x = v[0];
+ result[i].y = v[1];
+ result[i].z = v[2];
+
+ for (int j = 0; j < 3; ++j)
+ {
+ float vj = v[j];
+
+ minv[j] = minv[j] > vj ? vj : minv[j];
+ maxv[j] = maxv[j] < vj ? vj : maxv[j];
+ }
+ }
+
+ float extent = 0.f;
+
+ extent = (maxv[0] - minv[0]) < extent ? extent : (maxv[0] - minv[0]);
+ extent = (maxv[1] - minv[1]) < extent ? extent : (maxv[1] - minv[1]);
+ extent = (maxv[2] - minv[2]) < extent ? extent : (maxv[2] - minv[2]);
+
+ float scale = extent == 0 ? 0.f : 1.f / extent;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ result[i].x = (result[i].x - minv[0]) * scale;
+ result[i].y = (result[i].y - minv[1]) * scale;
+ result[i].z = (result[i].z - minv[2]) * scale;
+ }
+// -- GODOT start --
+ return extent;
+// -- GODOT end --
+
+}
+
+struct Quadric
+{
+ float a00, a11, a22;
+ float a10, a20, a21;
+ float b0, b1, b2, c;
+ float w;
+};
+
+struct Collapse
+{
+ unsigned int v0;
+ unsigned int v1;
+
+ union
+ {
+ unsigned int bidi;
+ float error;
+ unsigned int errorui;
+ };
+};
+
+static float normalize(Vector3& v)
+{
+ float length = sqrtf(v.x * v.x + v.y * v.y + v.z * v.z);
+
+ if (length > 0)
+ {
+ v.x /= length;
+ v.y /= length;
+ v.z /= length;
+ }
+
+ return length;
+}
+
+static void quadricAdd(Quadric& Q, const Quadric& R)
+{
+ Q.a00 += R.a00;
+ Q.a11 += R.a11;
+ Q.a22 += R.a22;
+ Q.a10 += R.a10;
+ Q.a20 += R.a20;
+ Q.a21 += R.a21;
+ Q.b0 += R.b0;
+ Q.b1 += R.b1;
+ Q.b2 += R.b2;
+ Q.c += R.c;
+ Q.w += R.w;
+}
+
+static float quadricError(const Quadric& Q, const Vector3& v)
+{
+ float rx = Q.b0;
+ float ry = Q.b1;
+ float rz = Q.b2;
+
+ rx += Q.a10 * v.y;
+ ry += Q.a21 * v.z;
+ rz += Q.a20 * v.x;
+
+ rx *= 2;
+ ry *= 2;
+ rz *= 2;
+
+ rx += Q.a00 * v.x;
+ ry += Q.a11 * v.y;
+ rz += Q.a22 * v.z;
+
+ float r = Q.c;
+ r += rx * v.x;
+ r += ry * v.y;
+ r += rz * v.z;
+
+ float s = Q.w == 0.f ? 0.f : 1.f / Q.w;
+
+ return fabsf(r) * s;
+}
+
+static void quadricFromPlane(Quadric& Q, float a, float b, float c, float d, float w)
+{
+ float aw = a * w;
+ float bw = b * w;
+ float cw = c * w;
+ float dw = d * w;
+
+ Q.a00 = a * aw;
+ Q.a11 = b * bw;
+ Q.a22 = c * cw;
+ Q.a10 = a * bw;
+ Q.a20 = a * cw;
+ Q.a21 = b * cw;
+ Q.b0 = a * dw;
+ Q.b1 = b * dw;
+ Q.b2 = c * dw;
+ Q.c = d * dw;
+ Q.w = w;
+}
+
+static void quadricFromPoint(Quadric& Q, float x, float y, float z, float w)
+{
+ // we need to encode (x - X) ^ 2 + (y - Y)^2 + (z - Z)^2 into the quadric
+ Q.a00 = w;
+ Q.a11 = w;
+ Q.a22 = w;
+ Q.a10 = 0.f;
+ Q.a20 = 0.f;
+ Q.a21 = 0.f;
+ Q.b0 = -2.f * x * w;
+ Q.b1 = -2.f * y * w;
+ Q.b2 = -2.f * z * w;
+ Q.c = (x * x + y * y + z * z) * w;
+ Q.w = w;
+}
+
+static void quadricFromTriangle(Quadric& Q, const Vector3& p0, const Vector3& p1, const Vector3& p2, float weight)
+{
+ Vector3 p10 = {p1.x - p0.x, p1.y - p0.y, p1.z - p0.z};
+ Vector3 p20 = {p2.x - p0.x, p2.y - p0.y, p2.z - p0.z};
+
+ // normal = cross(p1 - p0, p2 - p0)
+ Vector3 normal = {p10.y * p20.z - p10.z * p20.y, p10.z * p20.x - p10.x * p20.z, p10.x * p20.y - p10.y * p20.x};
+ float area = normalize(normal);
+
+ float distance = normal.x * p0.x + normal.y * p0.y + normal.z * p0.z;
+
+ // we use sqrtf(area) so that the error is scaled linearly; this tends to improve silhouettes
+ quadricFromPlane(Q, normal.x, normal.y, normal.z, -distance, sqrtf(area) * weight);
+}
+
+static void quadricFromTriangleEdge(Quadric& Q, const Vector3& p0, const Vector3& p1, const Vector3& p2, float weight)
+{
+ Vector3 p10 = {p1.x - p0.x, p1.y - p0.y, p1.z - p0.z};
+ float length = normalize(p10);
+
+ // p20p = length of projection of p2-p0 onto normalize(p1 - p0)
+ Vector3 p20 = {p2.x - p0.x, p2.y - p0.y, p2.z - p0.z};
+ float p20p = p20.x * p10.x + p20.y * p10.y + p20.z * p10.z;
+
+ // normal = altitude of triangle from point p2 onto edge p1-p0
+ Vector3 normal = {p20.x - p10.x * p20p, p20.y - p10.y * p20p, p20.z - p10.z * p20p};
+ normalize(normal);
+
+ float distance = normal.x * p0.x + normal.y * p0.y + normal.z * p0.z;
+
+ // note: the weight is scaled linearly with edge length; this has to match the triangle weight
+ quadricFromPlane(Q, normal.x, normal.y, normal.z, -distance, length * weight);
+}
+
+static void fillFaceQuadrics(Quadric* vertex_quadrics, const unsigned int* indices, size_t index_count, const Vector3* vertex_positions, const unsigned int* remap)
+{
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int i0 = indices[i + 0];
+ unsigned int i1 = indices[i + 1];
+ unsigned int i2 = indices[i + 2];
+
+ Quadric Q;
+ quadricFromTriangle(Q, vertex_positions[i0], vertex_positions[i1], vertex_positions[i2], 1.f);
+
+ quadricAdd(vertex_quadrics[remap[i0]], Q);
+ quadricAdd(vertex_quadrics[remap[i1]], Q);
+ quadricAdd(vertex_quadrics[remap[i2]], Q);
+ }
+}
+
+static void fillEdgeQuadrics(Quadric* vertex_quadrics, const unsigned int* indices, size_t index_count, const Vector3* vertex_positions, const unsigned int* remap, const unsigned char* vertex_kind, const unsigned int* loop, const unsigned int* loopback)
+{
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ static const int next[3] = {1, 2, 0};
+
+ for (int e = 0; e < 3; ++e)
+ {
+ unsigned int i0 = indices[i + e];
+ unsigned int i1 = indices[i + next[e]];
+
+ unsigned char k0 = vertex_kind[i0];
+ unsigned char k1 = vertex_kind[i1];
+
+ // check that either i0 or i1 are border/seam and are on the same edge loop
+ // note that we need to add the error even for edged that connect e.g. border & locked
+ // if we don't do that, the adjacent border->border edge won't have correct errors for corners
+ if (k0 != Kind_Border && k0 != Kind_Seam && k1 != Kind_Border && k1 != Kind_Seam)
+ continue;
+
+ if ((k0 == Kind_Border || k0 == Kind_Seam) && loop[i0] != i1)
+ continue;
+
+ if ((k1 == Kind_Border || k1 == Kind_Seam) && loopback[i1] != i0)
+ continue;
+
+ // seam edges should occur twice (i0->i1 and i1->i0) - skip redundant edges
+ if (kHasOpposite[k0][k1] && remap[i1] > remap[i0])
+ continue;
+
+ unsigned int i2 = indices[i + next[next[e]]];
+
+ // we try hard to maintain border edge geometry; seam edges can move more freely
+ // due to topological restrictions on collapses, seam quadrics slightly improves collapse structure but aren't critical
+ const float kEdgeWeightSeam = 1.f;
+ const float kEdgeWeightBorder = 10.f;
+
+ float edgeWeight = (k0 == Kind_Border || k1 == Kind_Border) ? kEdgeWeightBorder : kEdgeWeightSeam;
+
+ Quadric Q;
+ quadricFromTriangleEdge(Q, vertex_positions[i0], vertex_positions[i1], vertex_positions[i2], edgeWeight);
+
+ quadricAdd(vertex_quadrics[remap[i0]], Q);
+ quadricAdd(vertex_quadrics[remap[i1]], Q);
+ }
+ }
+}
+
+static size_t pickEdgeCollapses(Collapse* collapses, const unsigned int* indices, size_t index_count, const unsigned int* remap, const unsigned char* vertex_kind, const unsigned int* loop)
+{
+ size_t collapse_count = 0;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ static const int next[3] = {1, 2, 0};
+
+ for (int e = 0; e < 3; ++e)
+ {
+ unsigned int i0 = indices[i + e];
+ unsigned int i1 = indices[i + next[e]];
+
+ // this can happen either when input has a zero-length edge, or when we perform collapses for complex
+ // topology w/seams and collapse a manifold vertex that connects to both wedges onto one of them
+ // we leave edges like this alone since they may be important for preserving mesh integrity
+ if (remap[i0] == remap[i1])
+ continue;
+
+ unsigned char k0 = vertex_kind[i0];
+ unsigned char k1 = vertex_kind[i1];
+
+ // the edge has to be collapsible in at least one direction
+ if (!(kCanCollapse[k0][k1] | kCanCollapse[k1][k0]))
+ continue;
+
+ // manifold and seam edges should occur twice (i0->i1 and i1->i0) - skip redundant edges
+ if (kHasOpposite[k0][k1] && remap[i1] > remap[i0])
+ continue;
+
+ // two vertices are on a border or a seam, but there's no direct edge between them
+ // this indicates that they belong to two different edge loops and we should not collapse this edge
+ // loop[] tracks half edges so we only need to check i0->i1
+ if (k0 == k1 && (k0 == Kind_Border || k0 == Kind_Seam) && loop[i0] != i1)
+ continue;
+
+ // edge can be collapsed in either direction - we will pick the one with minimum error
+ // note: we evaluate error later during collapse ranking, here we just tag the edge as bidirectional
+ if (kCanCollapse[k0][k1] & kCanCollapse[k1][k0])
+ {
+ Collapse c = {i0, i1, {/* bidi= */ 1}};
+ collapses[collapse_count++] = c;
+ }
+ else
+ {
+ // edge can only be collapsed in one direction
+ unsigned int e0 = kCanCollapse[k0][k1] ? i0 : i1;
+ unsigned int e1 = kCanCollapse[k0][k1] ? i1 : i0;
+
+ Collapse c = {e0, e1, {/* bidi= */ 0}};
+ collapses[collapse_count++] = c;
+ }
+ }
+ }
+
+ return collapse_count;
+}
+
+static void rankEdgeCollapses(Collapse* collapses, size_t collapse_count, const Vector3* vertex_positions, const Quadric* vertex_quadrics, const unsigned int* remap)
+{
+ for (size_t i = 0; i < collapse_count; ++i)
+ {
+ Collapse& c = collapses[i];
+
+ unsigned int i0 = c.v0;
+ unsigned int i1 = c.v1;
+
+ // most edges are bidirectional which means we need to evaluate errors for two collapses
+ // to keep this code branchless we just use the same edge for unidirectional edges
+ unsigned int j0 = c.bidi ? i1 : i0;
+ unsigned int j1 = c.bidi ? i0 : i1;
+
+ const Quadric& qi = vertex_quadrics[remap[i0]];
+ const Quadric& qj = vertex_quadrics[remap[j0]];
+
+ float ei = quadricError(qi, vertex_positions[i1]);
+ float ej = quadricError(qj, vertex_positions[j1]);
+
+ // pick edge direction with minimal error
+ c.v0 = ei <= ej ? i0 : j0;
+ c.v1 = ei <= ej ? i1 : j1;
+ c.error = ei <= ej ? ei : ej;
+ }
+}
+
+#if TRACE > 1
+static void dumpEdgeCollapses(const Collapse* collapses, size_t collapse_count, const unsigned char* vertex_kind)
+{
+ size_t ckinds[Kind_Count][Kind_Count] = {};
+ float cerrors[Kind_Count][Kind_Count] = {};
+
+ for (int k0 = 0; k0 < Kind_Count; ++k0)
+ for (int k1 = 0; k1 < Kind_Count; ++k1)
+ cerrors[k0][k1] = FLT_MAX;
+
+ for (size_t i = 0; i < collapse_count; ++i)
+ {
+ unsigned int i0 = collapses[i].v0;
+ unsigned int i1 = collapses[i].v1;
+
+ unsigned char k0 = vertex_kind[i0];
+ unsigned char k1 = vertex_kind[i1];
+
+ ckinds[k0][k1]++;
+ cerrors[k0][k1] = (collapses[i].error < cerrors[k0][k1]) ? collapses[i].error : cerrors[k0][k1];
+ }
+
+ for (int k0 = 0; k0 < Kind_Count; ++k0)
+ for (int k1 = 0; k1 < Kind_Count; ++k1)
+ if (ckinds[k0][k1])
+ printf("collapses %d -> %d: %d, min error %e\n", k0, k1, int(ckinds[k0][k1]), cerrors[k0][k1]);
+}
+
+static void dumpLockedCollapses(const unsigned int* indices, size_t index_count, const unsigned char* vertex_kind)
+{
+ size_t locked_collapses[Kind_Count][Kind_Count] = {};
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ static const int next[3] = {1, 2, 0};
+
+ for (int e = 0; e < 3; ++e)
+ {
+ unsigned int i0 = indices[i + e];
+ unsigned int i1 = indices[i + next[e]];
+
+ unsigned char k0 = vertex_kind[i0];
+ unsigned char k1 = vertex_kind[i1];
+
+ locked_collapses[k0][k1] += !kCanCollapse[k0][k1] && !kCanCollapse[k1][k0];
+ }
+ }
+
+ for (int k0 = 0; k0 < Kind_Count; ++k0)
+ for (int k1 = 0; k1 < Kind_Count; ++k1)
+ if (locked_collapses[k0][k1])
+ printf("locked collapses %d -> %d: %d\n", k0, k1, int(locked_collapses[k0][k1]));
+}
+#endif
+
+static void sortEdgeCollapses(unsigned int* sort_order, const Collapse* collapses, size_t collapse_count)
+{
+ const int sort_bits = 11;
+
+ // fill histogram for counting sort
+ unsigned int histogram[1 << sort_bits];
+ memset(histogram, 0, sizeof(histogram));
+
+ for (size_t i = 0; i < collapse_count; ++i)
+ {
+ // skip sign bit since error is non-negative
+ unsigned int key = (collapses[i].errorui << 1) >> (32 - sort_bits);
+
+ histogram[key]++;
+ }
+
+ // compute offsets based on histogram data
+ size_t histogram_sum = 0;
+
+ for (size_t i = 0; i < 1 << sort_bits; ++i)
+ {
+ size_t count = histogram[i];
+ histogram[i] = unsigned(histogram_sum);
+ histogram_sum += count;
+ }
+
+ assert(histogram_sum == collapse_count);
+
+ // compute sort order based on offsets
+ for (size_t i = 0; i < collapse_count; ++i)
+ {
+ // skip sign bit since error is non-negative
+ unsigned int key = (collapses[i].errorui << 1) >> (32 - sort_bits);
+
+ sort_order[histogram[key]++] = unsigned(i);
+ }
+}
+
+static size_t performEdgeCollapses(unsigned int* collapse_remap, unsigned char* collapse_locked, Quadric* vertex_quadrics, const Collapse* collapses, size_t collapse_count, const unsigned int* collapse_order, const unsigned int* remap, const unsigned int* wedge, const unsigned char* vertex_kind, size_t triangle_collapse_goal, float error_goal, float error_limit)
+{
+ size_t edge_collapses = 0;
+ size_t triangle_collapses = 0;
+
+ for (size_t i = 0; i < collapse_count; ++i)
+ {
+ const Collapse& c = collapses[collapse_order[i]];
+
+ if (c.error > error_limit)
+ break;
+
+ if (c.error > error_goal && triangle_collapses > triangle_collapse_goal / 10)
+ break;
+
+ if (triangle_collapses >= triangle_collapse_goal)
+ break;
+
+ unsigned int i0 = c.v0;
+ unsigned int i1 = c.v1;
+
+ unsigned int r0 = remap[i0];
+ unsigned int r1 = remap[i1];
+
+ // we don't collapse vertices that had source or target vertex involved in a collapse
+ // it's important to not move the vertices twice since it complicates the tracking/remapping logic
+ // it's important to not move other vertices towards a moved vertex to preserve error since we don't re-rank collapses mid-pass
+ if (collapse_locked[r0] | collapse_locked[r1])
+ continue;
+
+ assert(collapse_remap[r0] == r0);
+ assert(collapse_remap[r1] == r1);
+
+ quadricAdd(vertex_quadrics[r1], vertex_quadrics[r0]);
+
+ if (vertex_kind[i0] == Kind_Complex)
+ {
+ unsigned int v = i0;
+
+ do
+ {
+ collapse_remap[v] = r1;
+ v = wedge[v];
+ } while (v != i0);
+ }
+ else if (vertex_kind[i0] == Kind_Seam)
+ {
+ // remap v0 to v1 and seam pair of v0 to seam pair of v1
+ unsigned int s0 = wedge[i0];
+ unsigned int s1 = wedge[i1];
+
+ assert(s0 != i0 && s1 != i1);
+ assert(wedge[s0] == i0 && wedge[s1] == i1);
+
+ collapse_remap[i0] = i1;
+ collapse_remap[s0] = s1;
+ }
+ else
+ {
+ assert(wedge[i0] == i0);
+
+ collapse_remap[i0] = i1;
+ }
+
+ collapse_locked[r0] = 1;
+ collapse_locked[r1] = 1;
+
+ // border edges collapse 1 triangle, other edges collapse 2 or more
+ triangle_collapses += (vertex_kind[i0] == Kind_Border) ? 1 : 2;
+ edge_collapses++;
+ }
+
+ return edge_collapses;
+}
+
+static size_t remapIndexBuffer(unsigned int* indices, size_t index_count, const unsigned int* collapse_remap)
+{
+ size_t write = 0;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int v0 = collapse_remap[indices[i + 0]];
+ unsigned int v1 = collapse_remap[indices[i + 1]];
+ unsigned int v2 = collapse_remap[indices[i + 2]];
+
+ // we never move the vertex twice during a single pass
+ assert(collapse_remap[v0] == v0);
+ assert(collapse_remap[v1] == v1);
+ assert(collapse_remap[v2] == v2);
+
+ if (v0 != v1 && v0 != v2 && v1 != v2)
+ {
+ indices[write + 0] = v0;
+ indices[write + 1] = v1;
+ indices[write + 2] = v2;
+ write += 3;
+ }
+ }
+
+ return write;
+}
+
+static void remapEdgeLoops(unsigned int* loop, size_t vertex_count, const unsigned int* collapse_remap)
+{
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ if (loop[i] != ~0u)
+ {
+ unsigned int l = loop[i];
+ unsigned int r = collapse_remap[l];
+
+ // i == r is a special case when the seam edge is collapsed in a direction opposite to where loop goes
+ loop[i] = (i == r) ? loop[l] : r;
+ }
+ }
+}
+
+struct CellHasher
+{
+ const unsigned int* vertex_ids;
+
+ size_t hash(unsigned int i) const
+ {
+ unsigned int h = vertex_ids[i];
+
+ // MurmurHash2 finalizer
+ h ^= h >> 13;
+ h *= 0x5bd1e995;
+ h ^= h >> 15;
+ return h;
+ }
+
+ bool equal(unsigned int lhs, unsigned int rhs) const
+ {
+ return vertex_ids[lhs] == vertex_ids[rhs];
+ }
+};
+
+struct IdHasher
+{
+ size_t hash(unsigned int id) const
+ {
+ unsigned int h = id;
+
+ // MurmurHash2 finalizer
+ h ^= h >> 13;
+ h *= 0x5bd1e995;
+ h ^= h >> 15;
+ return h;
+ }
+
+ bool equal(unsigned int lhs, unsigned int rhs) const
+ {
+ return lhs == rhs;
+ }
+};
+
+struct TriangleHasher
+{
+ unsigned int* indices;
+
+ size_t hash(unsigned int i) const
+ {
+ const unsigned int* tri = indices + i * 3;
+
+ // Optimized Spatial Hashing for Collision Detection of Deformable Objects
+ return (tri[0] * 73856093) ^ (tri[1] * 19349663) ^ (tri[2] * 83492791);
+ }
+
+ bool equal(unsigned int lhs, unsigned int rhs) const
+ {
+ const unsigned int* lt = indices + lhs * 3;
+ const unsigned int* rt = indices + rhs * 3;
+
+ return lt[0] == rt[0] && lt[1] == rt[1] && lt[2] == rt[2];
+ }
+};
+
+static void computeVertexIds(unsigned int* vertex_ids, const Vector3* vertex_positions, size_t vertex_count, int grid_size)
+{
+ assert(grid_size >= 1 && grid_size <= 1024);
+ float cell_scale = float(grid_size - 1);
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ const Vector3& v = vertex_positions[i];
+
+ int xi = int(v.x * cell_scale + 0.5f);
+ int yi = int(v.y * cell_scale + 0.5f);
+ int zi = int(v.z * cell_scale + 0.5f);
+
+ vertex_ids[i] = (xi << 20) | (yi << 10) | zi;
+ }
+}
+
+static size_t countTriangles(const unsigned int* vertex_ids, const unsigned int* indices, size_t index_count)
+{
+ size_t result = 0;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int id0 = vertex_ids[indices[i + 0]];
+ unsigned int id1 = vertex_ids[indices[i + 1]];
+ unsigned int id2 = vertex_ids[indices[i + 2]];
+
+ result += (id0 != id1) & (id0 != id2) & (id1 != id2);
+ }
+
+ return result;
+}
+
+static size_t fillVertexCells(unsigned int* table, size_t table_size, unsigned int* vertex_cells, const unsigned int* vertex_ids, size_t vertex_count)
+{
+ CellHasher hasher = {vertex_ids};
+
+ memset(table, -1, table_size * sizeof(unsigned int));
+
+ size_t result = 0;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ unsigned int* entry = hashLookup2(table, table_size, hasher, unsigned(i), ~0u);
+
+ if (*entry == ~0u)
+ {
+ *entry = unsigned(i);
+ vertex_cells[i] = unsigned(result++);
+ }
+ else
+ {
+ vertex_cells[i] = vertex_cells[*entry];
+ }
+ }
+
+ return result;
+}
+
+static size_t countVertexCells(unsigned int* table, size_t table_size, const unsigned int* vertex_ids, size_t vertex_count)
+{
+ IdHasher hasher;
+
+ memset(table, -1, table_size * sizeof(unsigned int));
+
+ size_t result = 0;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ unsigned int id = vertex_ids[i];
+ unsigned int* entry = hashLookup2(table, table_size, hasher, id, ~0u);
+
+ result += (*entry == ~0u);
+ *entry = id;
+ }
+
+ return result;
+}
+
+static void fillCellQuadrics(Quadric* cell_quadrics, const unsigned int* indices, size_t index_count, const Vector3* vertex_positions, const unsigned int* vertex_cells)
+{
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int i0 = indices[i + 0];
+ unsigned int i1 = indices[i + 1];
+ unsigned int i2 = indices[i + 2];
+
+ unsigned int c0 = vertex_cells[i0];
+ unsigned int c1 = vertex_cells[i1];
+ unsigned int c2 = vertex_cells[i2];
+
+ bool single_cell = (c0 == c1) & (c0 == c2);
+
+ Quadric Q;
+ quadricFromTriangle(Q, vertex_positions[i0], vertex_positions[i1], vertex_positions[i2], single_cell ? 3.f : 1.f);
+
+ if (single_cell)
+ {
+ quadricAdd(cell_quadrics[c0], Q);
+ }
+ else
+ {
+ quadricAdd(cell_quadrics[c0], Q);
+ quadricAdd(cell_quadrics[c1], Q);
+ quadricAdd(cell_quadrics[c2], Q);
+ }
+ }
+}
+
+static void fillCellQuadrics(Quadric* cell_quadrics, const Vector3* vertex_positions, size_t vertex_count, const unsigned int* vertex_cells)
+{
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ unsigned int c = vertex_cells[i];
+ const Vector3& v = vertex_positions[i];
+
+ Quadric Q;
+ quadricFromPoint(Q, v.x, v.y, v.z, 1.f);
+
+ quadricAdd(cell_quadrics[c], Q);
+ }
+}
+
+static void fillCellRemap(unsigned int* cell_remap, float* cell_errors, size_t cell_count, const unsigned int* vertex_cells, const Quadric* cell_quadrics, const Vector3* vertex_positions, size_t vertex_count)
+{
+ memset(cell_remap, -1, cell_count * sizeof(unsigned int));
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ unsigned int cell = vertex_cells[i];
+ float error = quadricError(cell_quadrics[cell], vertex_positions[i]);
+
+ if (cell_remap[cell] == ~0u || cell_errors[cell] > error)
+ {
+ cell_remap[cell] = unsigned(i);
+ cell_errors[cell] = error;
+ }
+ }
+}
+
+static size_t filterTriangles(unsigned int* destination, unsigned int* tritable, size_t tritable_size, const unsigned int* indices, size_t index_count, const unsigned int* vertex_cells, const unsigned int* cell_remap)
+{
+ TriangleHasher hasher = {destination};
+
+ memset(tritable, -1, tritable_size * sizeof(unsigned int));
+
+ size_t result = 0;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int c0 = vertex_cells[indices[i + 0]];
+ unsigned int c1 = vertex_cells[indices[i + 1]];
+ unsigned int c2 = vertex_cells[indices[i + 2]];
+
+ if (c0 != c1 && c0 != c2 && c1 != c2)
+ {
+ unsigned int a = cell_remap[c0];
+ unsigned int b = cell_remap[c1];
+ unsigned int c = cell_remap[c2];
+
+ if (b < a && b < c)
+ {
+ unsigned int t = a;
+ a = b, b = c, c = t;
+ }
+ else if (c < a && c < b)
+ {
+ unsigned int t = c;
+ c = b, b = a, a = t;
+ }
+
+ destination[result * 3 + 0] = a;
+ destination[result * 3 + 1] = b;
+ destination[result * 3 + 2] = c;
+
+ unsigned int* entry = hashLookup2(tritable, tritable_size, hasher, unsigned(result), ~0u);
+
+ if (*entry == ~0u)
+ *entry = unsigned(result++);
+ }
+ }
+
+ return result * 3;
+}
+
+static float interpolate(float y, float x0, float y0, float x1, float y1, float x2, float y2)
+{
+ // three point interpolation from "revenge of interpolation search" paper
+ float num = (y1 - y) * (x1 - x2) * (x1 - x0) * (y2 - y0);
+ float den = (y2 - y) * (x1 - x2) * (y0 - y1) + (y0 - y) * (x1 - x0) * (y1 - y2);
+ return x1 + num / den;
+}
+
+} // namespace meshopt
+
+#ifndef NDEBUG
+unsigned char* meshopt_simplifyDebugKind = 0;
+unsigned int* meshopt_simplifyDebugLoop = 0;
+unsigned int* meshopt_simplifyDebugLoopBack = 0;
+#endif
+
+// -- GODOT start --
+//size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error)
+size_t meshopt_simplify(unsigned int *destination, const unsigned int *indices, size_t index_count, const float *vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error, float *r_resulting_error)
+// -- GODOT end --
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+ assert(target_index_count <= index_count);
+
+ meshopt_Allocator allocator;
+
+ unsigned int* result = destination;
+
+ // build adjacency information
+ EdgeAdjacency adjacency = {};
+ buildEdgeAdjacency(adjacency, indices, index_count, vertex_count, allocator);
+
+ // build position remap that maps each vertex to the one with identical position
+ unsigned int* remap = allocator.allocate<unsigned int>(vertex_count);
+ unsigned int* wedge = allocator.allocate<unsigned int>(vertex_count);
+ buildPositionRemap(remap, wedge, vertex_positions_data, vertex_count, vertex_positions_stride, allocator);
+
+ // classify vertices; vertex kind determines collapse rules, see kCanCollapse
+ unsigned char* vertex_kind = allocator.allocate<unsigned char>(vertex_count);
+ unsigned int* loop = allocator.allocate<unsigned int>(vertex_count);
+ unsigned int* loopback = allocator.allocate<unsigned int>(vertex_count);
+ classifyVertices(vertex_kind, loop, loopback, vertex_count, adjacency, remap, wedge);
+
+#if TRACE
+ size_t unique_positions = 0;
+ for (size_t i = 0; i < vertex_count; ++i)
+ unique_positions += remap[i] == i;
+
+ printf("position remap: %d vertices => %d positions\n", int(vertex_count), int(unique_positions));
+
+ size_t kinds[Kind_Count] = {};
+ for (size_t i = 0; i < vertex_count; ++i)
+ kinds[vertex_kind[i]] += remap[i] == i;
+
+ printf("kinds: manifold %d, border %d, seam %d, complex %d, locked %d\n",
+ int(kinds[Kind_Manifold]), int(kinds[Kind_Border]), int(kinds[Kind_Seam]), int(kinds[Kind_Complex]), int(kinds[Kind_Locked]));
+#endif
+
+ Vector3* vertex_positions = allocator.allocate<Vector3>(vertex_count);
+// -- GODOT start --
+ //rescalePositions(vertex_positions, vertex_positions_data, vertex_count, vertex_positions_stride);
+ float extent = rescalePositions(vertex_positions, vertex_positions_data, vertex_count, vertex_positions_stride);
+// -- GODOT end --
+
+ Quadric* vertex_quadrics = allocator.allocate<Quadric>(vertex_count);
+ memset(vertex_quadrics, 0, vertex_count * sizeof(Quadric));
+
+ fillFaceQuadrics(vertex_quadrics, indices, index_count, vertex_positions, remap);
+ fillEdgeQuadrics(vertex_quadrics, indices, index_count, vertex_positions, remap, vertex_kind, loop, loopback);
+
+ if (result != indices)
+ memcpy(result, indices, index_count * sizeof(unsigned int));
+
+// -- GODOT start --
+#if TRACE
+ size_t pass_count = 0;
+ //float worst_error = 0;
+#endif
+ float worst_error = 0;
+// -- GODOT end --
+
+ Collapse* edge_collapses = allocator.allocate<Collapse>(index_count);
+ unsigned int* collapse_order = allocator.allocate<unsigned int>(index_count);
+ unsigned int* collapse_remap = allocator.allocate<unsigned int>(vertex_count);
+ unsigned char* collapse_locked = allocator.allocate<unsigned char>(vertex_count);
+
+ size_t result_count = index_count;
+
+ // target_error input is linear; we need to adjust it to match quadricError units
+ float error_limit = target_error * target_error;
+
+// -- GODOT start --
+ if (r_resulting_error) {
+ *r_resulting_error = 1.0;
+ }
+// -- GODOT end --
+
+ while (result_count > target_index_count)
+ {
+ size_t edge_collapse_count = pickEdgeCollapses(edge_collapses, result, result_count, remap, vertex_kind, loop);
+
+ // no edges can be collapsed any more due to topology restrictions
+ if (edge_collapse_count == 0)
+ break;
+
+ rankEdgeCollapses(edge_collapses, edge_collapse_count, vertex_positions, vertex_quadrics, remap);
+
+#if TRACE > 1
+ dumpEdgeCollapses(edge_collapses, edge_collapse_count, vertex_kind);
+#endif
+
+ sortEdgeCollapses(collapse_order, edge_collapses, edge_collapse_count);
+
+ // most collapses remove 2 triangles; use this to establish a bound on the pass in terms of error limit
+ // note that edge_collapse_goal is an estimate; triangle_collapse_goal will be used to actually limit collapses
+ size_t triangle_collapse_goal = (result_count - target_index_count) / 3;
+ size_t edge_collapse_goal = triangle_collapse_goal / 2;
+
+ // we limit the error in each pass based on the error of optimal last collapse; since many collapses will be locked
+ // as they will share vertices with other successfull collapses, we need to increase the acceptable error by this factor
+ const float kPassErrorBound = 1.5f;
+
+ float error_goal = edge_collapse_goal < edge_collapse_count ? edge_collapses[collapse_order[edge_collapse_goal]].error * kPassErrorBound : FLT_MAX;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ collapse_remap[i] = unsigned(i);
+
+ memset(collapse_locked, 0, vertex_count);
+
+ size_t collapses = performEdgeCollapses(collapse_remap, collapse_locked, vertex_quadrics, edge_collapses, edge_collapse_count, collapse_order, remap, wedge, vertex_kind, triangle_collapse_goal, error_goal, error_limit);
+
+ // no edges can be collapsed any more due to hitting the error limit or triangle collapse limit
+ if (collapses == 0)
+ break;
+
+ remapEdgeLoops(loop, vertex_count, collapse_remap);
+ remapEdgeLoops(loopback, vertex_count, collapse_remap);
+
+ size_t new_count = remapIndexBuffer(result, result_count, collapse_remap);
+ assert(new_count < result_count);
+
+// -- GODOT start --
+//#if TRACE
+ float pass_error = 0.f;
+ for (size_t i = 0; i < edge_collapse_count; ++i)
+ {
+ Collapse& c = edge_collapses[collapse_order[i]];
+
+ if (collapse_remap[c.v0] == c.v1)
+ pass_error = c.error;
+ }
+
+ //pass_count++;
+ worst_error = (worst_error < pass_error) ? pass_error : worst_error;
+
+#if TRACE
+ pass_count++;
+ printf("pass %d: triangles: %d -> %d, collapses: %d/%d (goal: %d), error: %e (limit %e goal %e)\n", int(pass_count), int(result_count / 3), int(new_count / 3), int(collapses), int(edge_collapse_count), int(edge_collapse_goal), pass_error, error_limit, error_goal);
+#endif
+// -- GODOT end --
+
+ result_count = new_count;
+ }
+
+// -- GODOT start --
+ if (r_resulting_error) {
+ *r_resulting_error = sqrt(worst_error) * extent;
+ }
+// -- GODOT end --
+
+#if TRACE
+ printf("passes: %d, worst error: %e\n", int(pass_count), worst_error);
+#endif
+
+#if TRACE > 1
+ dumpLockedCollapses(result, result_count, vertex_kind);
+#endif
+
+#ifndef NDEBUG
+ if (meshopt_simplifyDebugKind)
+ memcpy(meshopt_simplifyDebugKind, vertex_kind, vertex_count);
+
+ if (meshopt_simplifyDebugLoop)
+ memcpy(meshopt_simplifyDebugLoop, loop, vertex_count * sizeof(unsigned int));
+
+ if (meshopt_simplifyDebugLoopBack)
+ memcpy(meshopt_simplifyDebugLoopBack, loopback, vertex_count * sizeof(unsigned int));
+#endif
+
+ return result_count;
+}
+
+size_t meshopt_simplifySloppy(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+ assert(target_index_count <= index_count);
+
+ // we expect to get ~2 triangles/vertex in the output
+ size_t target_cell_count = target_index_count / 6;
+
+ if (target_cell_count == 0)
+ return 0;
+
+ meshopt_Allocator allocator;
+
+ Vector3* vertex_positions = allocator.allocate<Vector3>(vertex_count);
+ rescalePositions(vertex_positions, vertex_positions_data, vertex_count, vertex_positions_stride);
+
+ // find the optimal grid size using guided binary search
+#if TRACE
+ printf("source: %d vertices, %d triangles\n", int(vertex_count), int(index_count / 3));
+ printf("target: %d cells, %d triangles\n", int(target_cell_count), int(target_index_count / 3));
+#endif
+
+ unsigned int* vertex_ids = allocator.allocate<unsigned int>(vertex_count);
+
+ const int kInterpolationPasses = 5;
+
+ // invariant: # of triangles in min_grid <= target_count
+ int min_grid = 0;
+ int max_grid = 1025;
+ size_t min_triangles = 0;
+ size_t max_triangles = index_count / 3;
+
+ // instead of starting in the middle, let's guess as to what the answer might be! triangle count usually grows as a square of grid size...
+ int next_grid_size = int(sqrtf(float(target_cell_count)) + 0.5f);
+
+ for (int pass = 0; pass < 10 + kInterpolationPasses; ++pass)
+ {
+ assert(min_triangles < target_index_count / 3);
+ assert(max_grid - min_grid > 1);
+
+ // we clamp the prediction of the grid size to make sure that the search converges
+ int grid_size = next_grid_size;
+ grid_size = (grid_size <= min_grid) ? min_grid + 1 : (grid_size >= max_grid) ? max_grid - 1 : grid_size;
+
+ computeVertexIds(vertex_ids, vertex_positions, vertex_count, grid_size);
+ size_t triangles = countTriangles(vertex_ids, indices, index_count);
+
+#if TRACE
+ printf("pass %d (%s): grid size %d, triangles %d, %s\n",
+ pass, (pass == 0) ? "guess" : (pass <= kInterpolationPasses) ? "lerp" : "binary",
+ grid_size, int(triangles),
+ (triangles <= target_index_count / 3) ? "under" : "over");
+#endif
+
+ float tip = interpolate(float(target_index_count / 3), float(min_grid), float(min_triangles), float(grid_size), float(triangles), float(max_grid), float(max_triangles));
+
+ if (triangles <= target_index_count / 3)
+ {
+ min_grid = grid_size;
+ min_triangles = triangles;
+ }
+ else
+ {
+ max_grid = grid_size;
+ max_triangles = triangles;
+ }
+
+ if (triangles == target_index_count / 3 || max_grid - min_grid <= 1)
+ break;
+
+ // we start by using interpolation search - it usually converges faster
+ // however, interpolation search has a worst case of O(N) so we switch to binary search after a few iterations which converges in O(logN)
+ next_grid_size = (pass < kInterpolationPasses) ? int(tip + 0.5f) : (min_grid + max_grid) / 2;
+ }
+
+ if (min_triangles == 0)
+ return 0;
+
+ // build vertex->cell association by mapping all vertices with the same quantized position to the same cell
+ size_t table_size = hashBuckets2(vertex_count);
+ unsigned int* table = allocator.allocate<unsigned int>(table_size);
+
+ unsigned int* vertex_cells = allocator.allocate<unsigned int>(vertex_count);
+
+ computeVertexIds(vertex_ids, vertex_positions, vertex_count, min_grid);
+ size_t cell_count = fillVertexCells(table, table_size, vertex_cells, vertex_ids, vertex_count);
+
+ // build a quadric for each target cell
+ Quadric* cell_quadrics = allocator.allocate<Quadric>(cell_count);
+ memset(cell_quadrics, 0, cell_count * sizeof(Quadric));
+
+ fillCellQuadrics(cell_quadrics, indices, index_count, vertex_positions, vertex_cells);
+
+ // for each target cell, find the vertex with the minimal error
+ unsigned int* cell_remap = allocator.allocate<unsigned int>(cell_count);
+ float* cell_errors = allocator.allocate<float>(cell_count);
+
+ fillCellRemap(cell_remap, cell_errors, cell_count, vertex_cells, cell_quadrics, vertex_positions, vertex_count);
+
+ // collapse triangles!
+ // note that we need to filter out triangles that we've already output because we very frequently generate redundant triangles between cells :(
+ size_t tritable_size = hashBuckets2(min_triangles);
+ unsigned int* tritable = allocator.allocate<unsigned int>(tritable_size);
+
+ size_t write = filterTriangles(destination, tritable, tritable_size, indices, index_count, vertex_cells, cell_remap);
+ assert(write <= target_index_count);
+
+#if TRACE
+ printf("result: %d cells, %d triangles (%d unfiltered)\n", int(cell_count), int(write / 3), int(min_triangles));
+#endif
+
+ return write;
+}
+
+size_t meshopt_simplifyPoints(unsigned int* destination, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride, size_t target_vertex_count)
+{
+ using namespace meshopt;
+
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+ assert(target_vertex_count <= vertex_count);
+
+ size_t target_cell_count = target_vertex_count;
+
+ if (target_cell_count == 0)
+ return 0;
+
+ meshopt_Allocator allocator;
+
+ Vector3* vertex_positions = allocator.allocate<Vector3>(vertex_count);
+ rescalePositions(vertex_positions, vertex_positions_data, vertex_count, vertex_positions_stride);
+
+ // find the optimal grid size using guided binary search
+#if TRACE
+ printf("source: %d vertices\n", int(vertex_count));
+ printf("target: %d cells\n", int(target_cell_count));
+#endif
+
+ unsigned int* vertex_ids = allocator.allocate<unsigned int>(vertex_count);
+
+ size_t table_size = hashBuckets2(vertex_count);
+ unsigned int* table = allocator.allocate<unsigned int>(table_size);
+
+ const int kInterpolationPasses = 5;
+
+ // invariant: # of vertices in min_grid <= target_count
+ int min_grid = 0;
+ int max_grid = 1025;
+ size_t min_vertices = 0;
+ size_t max_vertices = vertex_count;
+
+ // instead of starting in the middle, let's guess as to what the answer might be! triangle count usually grows as a square of grid size...
+ int next_grid_size = int(sqrtf(float(target_cell_count)) + 0.5f);
+
+ for (int pass = 0; pass < 10 + kInterpolationPasses; ++pass)
+ {
+ assert(min_vertices < target_vertex_count);
+ assert(max_grid - min_grid > 1);
+
+ // we clamp the prediction of the grid size to make sure that the search converges
+ int grid_size = next_grid_size;
+ grid_size = (grid_size <= min_grid) ? min_grid + 1 : (grid_size >= max_grid) ? max_grid - 1 : grid_size;
+
+ computeVertexIds(vertex_ids, vertex_positions, vertex_count, grid_size);
+ size_t vertices = countVertexCells(table, table_size, vertex_ids, vertex_count);
+
+#if TRACE
+ printf("pass %d (%s): grid size %d, vertices %d, %s\n",
+ pass, (pass == 0) ? "guess" : (pass <= kInterpolationPasses) ? "lerp" : "binary",
+ grid_size, int(vertices),
+ (vertices <= target_vertex_count) ? "under" : "over");
+#endif
+
+ float tip = interpolate(float(target_vertex_count), float(min_grid), float(min_vertices), float(grid_size), float(vertices), float(max_grid), float(max_vertices));
+
+ if (vertices <= target_vertex_count)
+ {
+ min_grid = grid_size;
+ min_vertices = vertices;
+ }
+ else
+ {
+ max_grid = grid_size;
+ max_vertices = vertices;
+ }
+
+ if (vertices == target_vertex_count || max_grid - min_grid <= 1)
+ break;
+
+ // we start by using interpolation search - it usually converges faster
+ // however, interpolation search has a worst case of O(N) so we switch to binary search after a few iterations which converges in O(logN)
+ next_grid_size = (pass < kInterpolationPasses) ? int(tip + 0.5f) : (min_grid + max_grid) / 2;
+ }
+
+ if (min_vertices == 0)
+ return 0;
+
+ // build vertex->cell association by mapping all vertices with the same quantized position to the same cell
+ unsigned int* vertex_cells = allocator.allocate<unsigned int>(vertex_count);
+
+ computeVertexIds(vertex_ids, vertex_positions, vertex_count, min_grid);
+ size_t cell_count = fillVertexCells(table, table_size, vertex_cells, vertex_ids, vertex_count);
+
+ // build a quadric for each target cell
+ Quadric* cell_quadrics = allocator.allocate<Quadric>(cell_count);
+ memset(cell_quadrics, 0, cell_count * sizeof(Quadric));
+
+ fillCellQuadrics(cell_quadrics, vertex_positions, vertex_count, vertex_cells);
+
+ // for each target cell, find the vertex with the minimal error
+ unsigned int* cell_remap = allocator.allocate<unsigned int>(cell_count);
+ float* cell_errors = allocator.allocate<float>(cell_count);
+
+ fillCellRemap(cell_remap, cell_errors, cell_count, vertex_cells, cell_quadrics, vertex_positions, vertex_count);
+
+ // copy results to the output
+ assert(cell_count <= target_vertex_count);
+ memcpy(destination, cell_remap, sizeof(unsigned int) * cell_count);
+
+#if TRACE
+ printf("result: %d cells\n", int(cell_count));
+#endif
+
+ return cell_count;
+}
diff --git a/thirdparty/meshoptimizer/spatialorder.cpp b/thirdparty/meshoptimizer/spatialorder.cpp
new file mode 100644
index 0000000000..b09f80ac6f
--- /dev/null
+++ b/thirdparty/meshoptimizer/spatialorder.cpp
@@ -0,0 +1,194 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <float.h>
+#include <string.h>
+
+// This work is based on:
+// Fabian Giesen. Decoding Morton codes. 2009
+namespace meshopt
+{
+
+// "Insert" two 0 bits after each of the 10 low bits of x
+inline unsigned int part1By2(unsigned int x)
+{
+ x &= 0x000003ff; // x = ---- ---- ---- ---- ---- --98 7654 3210
+ x = (x ^ (x << 16)) & 0xff0000ff; // x = ---- --98 ---- ---- ---- ---- 7654 3210
+ x = (x ^ (x << 8)) & 0x0300f00f; // x = ---- --98 ---- ---- 7654 ---- ---- 3210
+ x = (x ^ (x << 4)) & 0x030c30c3; // x = ---- --98 ---- 76-- --54 ---- 32-- --10
+ x = (x ^ (x << 2)) & 0x09249249; // x = ---- 9--8 --7- -6-- 5--4 --3- -2-- 1--0
+ return x;
+}
+
+static void computeOrder(unsigned int* result, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride)
+{
+ size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
+
+ float minv[3] = {FLT_MAX, FLT_MAX, FLT_MAX};
+ float maxv[3] = {-FLT_MAX, -FLT_MAX, -FLT_MAX};
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ const float* v = vertex_positions_data + i * vertex_stride_float;
+
+ for (int j = 0; j < 3; ++j)
+ {
+ float vj = v[j];
+
+ minv[j] = minv[j] > vj ? vj : minv[j];
+ maxv[j] = maxv[j] < vj ? vj : maxv[j];
+ }
+ }
+
+ float extent = 0.f;
+
+ extent = (maxv[0] - minv[0]) < extent ? extent : (maxv[0] - minv[0]);
+ extent = (maxv[1] - minv[1]) < extent ? extent : (maxv[1] - minv[1]);
+ extent = (maxv[2] - minv[2]) < extent ? extent : (maxv[2] - minv[2]);
+
+ float scale = extent == 0 ? 0.f : 1.f / extent;
+
+ // generate Morton order based on the position inside a unit cube
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ const float* v = vertex_positions_data + i * vertex_stride_float;
+
+ int x = int((v[0] - minv[0]) * scale * 1023.f + 0.5f);
+ int y = int((v[1] - minv[1]) * scale * 1023.f + 0.5f);
+ int z = int((v[2] - minv[2]) * scale * 1023.f + 0.5f);
+
+ result[i] = part1By2(x) | (part1By2(y) << 1) | (part1By2(z) << 2);
+ }
+}
+
+static void computeHistogram(unsigned int (&hist)[1024][3], const unsigned int* data, size_t count)
+{
+ memset(hist, 0, sizeof(hist));
+
+ // compute 3 10-bit histograms in parallel
+ for (size_t i = 0; i < count; ++i)
+ {
+ unsigned int id = data[i];
+
+ hist[(id >> 0) & 1023][0]++;
+ hist[(id >> 10) & 1023][1]++;
+ hist[(id >> 20) & 1023][2]++;
+ }
+
+ unsigned int sumx = 0, sumy = 0, sumz = 0;
+
+ // replace histogram data with prefix histogram sums in-place
+ for (int i = 0; i < 1024; ++i)
+ {
+ unsigned int hx = hist[i][0], hy = hist[i][1], hz = hist[i][2];
+
+ hist[i][0] = sumx;
+ hist[i][1] = sumy;
+ hist[i][2] = sumz;
+
+ sumx += hx;
+ sumy += hy;
+ sumz += hz;
+ }
+
+ assert(sumx == count && sumy == count && sumz == count);
+}
+
+static void radixPass(unsigned int* destination, const unsigned int* source, const unsigned int* keys, size_t count, unsigned int (&hist)[1024][3], int pass)
+{
+ int bitoff = pass * 10;
+
+ for (size_t i = 0; i < count; ++i)
+ {
+ unsigned int id = (keys[source[i]] >> bitoff) & 1023;
+
+ destination[hist[id][pass]++] = source[i];
+ }
+}
+
+} // namespace meshopt
+
+void meshopt_spatialSortRemap(unsigned int* destination, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ using namespace meshopt;
+
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+
+ meshopt_Allocator allocator;
+
+ unsigned int* keys = allocator.allocate<unsigned int>(vertex_count);
+ computeOrder(keys, vertex_positions, vertex_count, vertex_positions_stride);
+
+ unsigned int hist[1024][3];
+ computeHistogram(hist, keys, vertex_count);
+
+ unsigned int* scratch = allocator.allocate<unsigned int>(vertex_count);
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ destination[i] = unsigned(i);
+
+ // 3-pass radix sort computes the resulting order into scratch
+ radixPass(scratch, destination, keys, vertex_count, hist, 0);
+ radixPass(destination, scratch, keys, vertex_count, hist, 1);
+ radixPass(scratch, destination, keys, vertex_count, hist, 2);
+
+ // since our remap table is mapping old=>new, we need to reverse it
+ for (size_t i = 0; i < vertex_count; ++i)
+ destination[scratch[i]] = unsigned(i);
+}
+
+void meshopt_spatialSortTriangles(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
+ assert(vertex_positions_stride % sizeof(float) == 0);
+
+ (void)vertex_count;
+
+ size_t face_count = index_count / 3;
+ size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
+
+ meshopt_Allocator allocator;
+
+ float* centroids = allocator.allocate<float>(face_count * 3);
+
+ for (size_t i = 0; i < face_count; ++i)
+ {
+ unsigned int a = indices[i * 3 + 0], b = indices[i * 3 + 1], c = indices[i * 3 + 2];
+ assert(a < vertex_count && b < vertex_count && c < vertex_count);
+
+ const float* va = vertex_positions + a * vertex_stride_float;
+ const float* vb = vertex_positions + b * vertex_stride_float;
+ const float* vc = vertex_positions + c * vertex_stride_float;
+
+ centroids[i * 3 + 0] = (va[0] + vb[0] + vc[0]) / 3.f;
+ centroids[i * 3 + 1] = (va[1] + vb[1] + vc[1]) / 3.f;
+ centroids[i * 3 + 2] = (va[2] + vb[2] + vc[2]) / 3.f;
+ }
+
+ unsigned int* remap = allocator.allocate<unsigned int>(face_count);
+
+ meshopt_spatialSortRemap(remap, centroids, face_count, sizeof(float) * 3);
+
+ // support in-order remap
+ if (destination == indices)
+ {
+ unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
+ memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
+ indices = indices_copy;
+ }
+
+ for (size_t i = 0; i < face_count; ++i)
+ {
+ unsigned int a = indices[i * 3 + 0], b = indices[i * 3 + 1], c = indices[i * 3 + 2];
+ unsigned int r = remap[i];
+
+ destination[r * 3 + 0] = a;
+ destination[r * 3 + 1] = b;
+ destination[r * 3 + 2] = c;
+ }
+}
diff --git a/thirdparty/meshoptimizer/stripifier.cpp b/thirdparty/meshoptimizer/stripifier.cpp
new file mode 100644
index 0000000000..8ce17ef3dc
--- /dev/null
+++ b/thirdparty/meshoptimizer/stripifier.cpp
@@ -0,0 +1,295 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <string.h>
+
+// This work is based on:
+// Francine Evans, Steven Skiena and Amitabh Varshney. Optimizing Triangle Strips for Fast Rendering. 1996
+namespace meshopt
+{
+
+static unsigned int findStripFirst(const unsigned int buffer[][3], unsigned int buffer_size, const unsigned int* valence)
+{
+ unsigned int index = 0;
+ unsigned int iv = ~0u;
+
+ for (size_t i = 0; i < buffer_size; ++i)
+ {
+ unsigned int va = valence[buffer[i][0]], vb = valence[buffer[i][1]], vc = valence[buffer[i][2]];
+ unsigned int v = (va < vb && va < vc) ? va : (vb < vc) ? vb : vc;
+
+ if (v < iv)
+ {
+ index = unsigned(i);
+ iv = v;
+ }
+ }
+
+ return index;
+}
+
+static int findStripNext(const unsigned int buffer[][3], unsigned int buffer_size, unsigned int e0, unsigned int e1)
+{
+ for (size_t i = 0; i < buffer_size; ++i)
+ {
+ unsigned int a = buffer[i][0], b = buffer[i][1], c = buffer[i][2];
+
+ if (e0 == a && e1 == b)
+ return (int(i) << 2) | 2;
+ else if (e0 == b && e1 == c)
+ return (int(i) << 2) | 0;
+ else if (e0 == c && e1 == a)
+ return (int(i) << 2) | 1;
+ }
+
+ return -1;
+}
+
+} // namespace meshopt
+
+size_t meshopt_stripify(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int restart_index)
+{
+ assert(destination != indices);
+ assert(index_count % 3 == 0);
+
+ using namespace meshopt;
+
+ meshopt_Allocator allocator;
+
+ const size_t buffer_capacity = 8;
+
+ unsigned int buffer[buffer_capacity][3] = {};
+ unsigned int buffer_size = 0;
+
+ size_t index_offset = 0;
+
+ unsigned int strip[2] = {};
+ unsigned int parity = 0;
+
+ size_t strip_size = 0;
+
+ // compute vertex valence; this is used to prioritize starting triangle for strips
+ unsigned int* valence = allocator.allocate<unsigned int>(vertex_count);
+ memset(valence, 0, vertex_count * sizeof(unsigned int));
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices[i];
+ assert(index < vertex_count);
+
+ valence[index]++;
+ }
+
+ int next = -1;
+
+ while (buffer_size > 0 || index_offset < index_count)
+ {
+ assert(next < 0 || (size_t(next >> 2) < buffer_size && (next & 3) < 3));
+
+ // fill triangle buffer
+ while (buffer_size < buffer_capacity && index_offset < index_count)
+ {
+ buffer[buffer_size][0] = indices[index_offset + 0];
+ buffer[buffer_size][1] = indices[index_offset + 1];
+ buffer[buffer_size][2] = indices[index_offset + 2];
+
+ buffer_size++;
+ index_offset += 3;
+ }
+
+ assert(buffer_size > 0);
+
+ if (next >= 0)
+ {
+ unsigned int i = next >> 2;
+ unsigned int a = buffer[i][0], b = buffer[i][1], c = buffer[i][2];
+ unsigned int v = buffer[i][next & 3];
+
+ // ordered removal from the buffer
+ memmove(buffer[i], buffer[i + 1], (buffer_size - i - 1) * sizeof(buffer[0]));
+ buffer_size--;
+
+ // update vertex valences for strip start heuristic
+ valence[a]--;
+ valence[b]--;
+ valence[c]--;
+
+ // find next triangle (note that edge order flips on every iteration)
+ // in some cases we need to perform a swap to pick a different outgoing triangle edge
+ // for [a b c], the default strip edge is [b c], but we might want to use [a c]
+ int cont = findStripNext(buffer, buffer_size, parity ? strip[1] : v, parity ? v : strip[1]);
+ int swap = cont < 0 ? findStripNext(buffer, buffer_size, parity ? v : strip[0], parity ? strip[0] : v) : -1;
+
+ if (cont < 0 && swap >= 0)
+ {
+ // [a b c] => [a b a c]
+ destination[strip_size++] = strip[0];
+ destination[strip_size++] = v;
+
+ // next strip has same winding
+ // ? a b => b a v
+ strip[1] = v;
+
+ next = swap;
+ }
+ else
+ {
+ // emit the next vertex in the strip
+ destination[strip_size++] = v;
+
+ // next strip has flipped winding
+ strip[0] = strip[1];
+ strip[1] = v;
+ parity ^= 1;
+
+ next = cont;
+ }
+ }
+ else
+ {
+ // if we didn't find anything, we need to find the next new triangle
+ // we use a heuristic to maximize the strip length
+ unsigned int i = findStripFirst(buffer, buffer_size, &valence[0]);
+ unsigned int a = buffer[i][0], b = buffer[i][1], c = buffer[i][2];
+
+ // ordered removal from the buffer
+ memmove(buffer[i], buffer[i + 1], (buffer_size - i - 1) * sizeof(buffer[0]));
+ buffer_size--;
+
+ // update vertex valences for strip start heuristic
+ valence[a]--;
+ valence[b]--;
+ valence[c]--;
+
+ // we need to pre-rotate the triangle so that we will find a match in the existing buffer on the next iteration
+ int ea = findStripNext(buffer, buffer_size, c, b);
+ int eb = findStripNext(buffer, buffer_size, a, c);
+ int ec = findStripNext(buffer, buffer_size, b, a);
+
+ // in some cases we can have several matching edges; since we can pick any edge, we pick the one with the smallest
+ // triangle index in the buffer. this reduces the effect of stripification on ACMR and additionally - for unclear
+ // reasons - slightly improves the stripification efficiency
+ int mine = INT_MAX;
+ mine = (ea >= 0 && mine > ea) ? ea : mine;
+ mine = (eb >= 0 && mine > eb) ? eb : mine;
+ mine = (ec >= 0 && mine > ec) ? ec : mine;
+
+ if (ea == mine)
+ {
+ // keep abc
+ next = ea;
+ }
+ else if (eb == mine)
+ {
+ // abc -> bca
+ unsigned int t = a;
+ a = b, b = c, c = t;
+
+ next = eb;
+ }
+ else if (ec == mine)
+ {
+ // abc -> cab
+ unsigned int t = c;
+ c = b, b = a, a = t;
+
+ next = ec;
+ }
+
+ if (restart_index)
+ {
+ if (strip_size)
+ destination[strip_size++] = restart_index;
+
+ destination[strip_size++] = a;
+ destination[strip_size++] = b;
+ destination[strip_size++] = c;
+
+ // new strip always starts with the same edge winding
+ strip[0] = b;
+ strip[1] = c;
+ parity = 1;
+ }
+ else
+ {
+ if (strip_size)
+ {
+ // connect last strip using degenerate triangles
+ destination[strip_size++] = strip[1];
+ destination[strip_size++] = a;
+ }
+
+ // note that we may need to flip the emitted triangle based on parity
+ // we always end up with outgoing edge "cb" in the end
+ unsigned int e0 = parity ? c : b;
+ unsigned int e1 = parity ? b : c;
+
+ destination[strip_size++] = a;
+ destination[strip_size++] = e0;
+ destination[strip_size++] = e1;
+
+ strip[0] = e0;
+ strip[1] = e1;
+ parity ^= 1;
+ }
+ }
+ }
+
+ return strip_size;
+}
+
+size_t meshopt_stripifyBound(size_t index_count)
+{
+ assert(index_count % 3 == 0);
+
+ // worst case without restarts is 2 degenerate indices and 3 indices per triangle
+ // worst case with restarts is 1 restart index and 3 indices per triangle
+ return (index_count / 3) * 5;
+}
+
+size_t meshopt_unstripify(unsigned int* destination, const unsigned int* indices, size_t index_count, unsigned int restart_index)
+{
+ assert(destination != indices);
+
+ size_t offset = 0;
+ size_t start = 0;
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ if (restart_index && indices[i] == restart_index)
+ {
+ start = i + 1;
+ }
+ else if (i - start >= 2)
+ {
+ unsigned int a = indices[i - 2], b = indices[i - 1], c = indices[i];
+
+ // flip winding for odd triangles
+ if ((i - start) & 1)
+ {
+ unsigned int t = a;
+ a = b, b = t;
+ }
+
+ // although we use restart indices, strip swaps still produce degenerate triangles, so skip them
+ if (a != b && a != c && b != c)
+ {
+ destination[offset + 0] = a;
+ destination[offset + 1] = b;
+ destination[offset + 2] = c;
+ offset += 3;
+ }
+ }
+ }
+
+ return offset;
+}
+
+size_t meshopt_unstripifyBound(size_t index_count)
+{
+ assert(index_count == 0 || index_count >= 3);
+
+ return (index_count == 0) ? 0 : (index_count - 2) * 3;
+}
diff --git a/thirdparty/meshoptimizer/vcacheanalyzer.cpp b/thirdparty/meshoptimizer/vcacheanalyzer.cpp
new file mode 100644
index 0000000000..3682743820
--- /dev/null
+++ b/thirdparty/meshoptimizer/vcacheanalyzer.cpp
@@ -0,0 +1,73 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <string.h>
+
+meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int primgroup_size)
+{
+ assert(index_count % 3 == 0);
+ assert(cache_size >= 3);
+ assert(warp_size == 0 || warp_size >= 3);
+
+ meshopt_Allocator allocator;
+
+ meshopt_VertexCacheStatistics result = {};
+
+ unsigned int warp_offset = 0;
+ unsigned int primgroup_offset = 0;
+
+ unsigned int* cache_timestamps = allocator.allocate<unsigned int>(vertex_count);
+ memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
+
+ unsigned int timestamp = cache_size + 1;
+
+ for (size_t i = 0; i < index_count; i += 3)
+ {
+ unsigned int a = indices[i + 0], b = indices[i + 1], c = indices[i + 2];
+ assert(a < vertex_count && b < vertex_count && c < vertex_count);
+
+ bool ac = (timestamp - cache_timestamps[a]) > cache_size;
+ bool bc = (timestamp - cache_timestamps[b]) > cache_size;
+ bool cc = (timestamp - cache_timestamps[c]) > cache_size;
+
+ // flush cache if triangle doesn't fit into warp or into the primitive buffer
+ if ((primgroup_size && primgroup_offset == primgroup_size) || (warp_size && warp_offset + ac + bc + cc > warp_size))
+ {
+ result.warps_executed += warp_offset > 0;
+
+ warp_offset = 0;
+ primgroup_offset = 0;
+
+ // reset cache
+ timestamp += cache_size + 1;
+ }
+
+ // update cache and add vertices to warp
+ for (int j = 0; j < 3; ++j)
+ {
+ unsigned int index = indices[i + j];
+
+ if (timestamp - cache_timestamps[index] > cache_size)
+ {
+ cache_timestamps[index] = timestamp++;
+ result.vertices_transformed++;
+ warp_offset++;
+ }
+ }
+
+ primgroup_offset++;
+ }
+
+ size_t unique_vertex_count = 0;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ unique_vertex_count += cache_timestamps[i] > 0;
+
+ result.warps_executed += warp_offset > 0;
+
+ result.acmr = index_count == 0 ? 0 : float(result.vertices_transformed) / float(index_count / 3);
+ result.atvr = unique_vertex_count == 0 ? 0 : float(result.vertices_transformed) / float(unique_vertex_count);
+
+ return result;
+}
diff --git a/thirdparty/meshoptimizer/vcacheoptimizer.cpp b/thirdparty/meshoptimizer/vcacheoptimizer.cpp
new file mode 100644
index 0000000000..fb8ade4b77
--- /dev/null
+++ b/thirdparty/meshoptimizer/vcacheoptimizer.cpp
@@ -0,0 +1,473 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <string.h>
+
+// This work is based on:
+// Tom Forsyth. Linear-Speed Vertex Cache Optimisation. 2006
+// Pedro Sander, Diego Nehab and Joshua Barczak. Fast Triangle Reordering for Vertex Locality and Reduced Overdraw. 2007
+namespace meshopt
+{
+
+const size_t kCacheSizeMax = 16;
+const size_t kValenceMax = 8;
+
+struct VertexScoreTable
+{
+ float cache[1 + kCacheSizeMax];
+ float live[1 + kValenceMax];
+};
+
+// Tuned to minimize the ACMR of a GPU that has a cache profile similar to NVidia and AMD
+static const VertexScoreTable kVertexScoreTable = {
+ {0.f, 0.779f, 0.791f, 0.789f, 0.981f, 0.843f, 0.726f, 0.847f, 0.882f, 0.867f, 0.799f, 0.642f, 0.613f, 0.600f, 0.568f, 0.372f, 0.234f},
+ {0.f, 0.995f, 0.713f, 0.450f, 0.404f, 0.059f, 0.005f, 0.147f, 0.006f},
+};
+
+// Tuned to minimize the encoded index buffer size
+static const VertexScoreTable kVertexScoreTableStrip = {
+ {0.f, 1.000f, 1.000f, 1.000f, 0.453f, 0.561f, 0.490f, 0.459f, 0.179f, 0.526f, 0.000f, 0.227f, 0.184f, 0.490f, 0.112f, 0.050f, 0.131f},
+ {0.f, 0.956f, 0.786f, 0.577f, 0.558f, 0.618f, 0.549f, 0.499f, 0.489f},
+};
+
+struct TriangleAdjacency
+{
+ unsigned int* counts;
+ unsigned int* offsets;
+ unsigned int* data;
+};
+
+static void buildTriangleAdjacency(TriangleAdjacency& adjacency, const unsigned int* indices, size_t index_count, size_t vertex_count, meshopt_Allocator& allocator)
+{
+ size_t face_count = index_count / 3;
+
+ // allocate arrays
+ adjacency.counts = allocator.allocate<unsigned int>(vertex_count);
+ adjacency.offsets = allocator.allocate<unsigned int>(vertex_count);
+ adjacency.data = allocator.allocate<unsigned int>(index_count);
+
+ // fill triangle counts
+ memset(adjacency.counts, 0, vertex_count * sizeof(unsigned int));
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ assert(indices[i] < vertex_count);
+
+ adjacency.counts[indices[i]]++;
+ }
+
+ // fill offset table
+ unsigned int offset = 0;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ adjacency.offsets[i] = offset;
+ offset += adjacency.counts[i];
+ }
+
+ assert(offset == index_count);
+
+ // fill triangle data
+ for (size_t i = 0; i < face_count; ++i)
+ {
+ unsigned int a = indices[i * 3 + 0], b = indices[i * 3 + 1], c = indices[i * 3 + 2];
+
+ adjacency.data[adjacency.offsets[a]++] = unsigned(i);
+ adjacency.data[adjacency.offsets[b]++] = unsigned(i);
+ adjacency.data[adjacency.offsets[c]++] = unsigned(i);
+ }
+
+ // fix offsets that have been disturbed by the previous pass
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ assert(adjacency.offsets[i] >= adjacency.counts[i]);
+
+ adjacency.offsets[i] -= adjacency.counts[i];
+ }
+}
+
+static unsigned int getNextVertexDeadEnd(const unsigned int* dead_end, unsigned int& dead_end_top, unsigned int& input_cursor, const unsigned int* live_triangles, size_t vertex_count)
+{
+ // check dead-end stack
+ while (dead_end_top)
+ {
+ unsigned int vertex = dead_end[--dead_end_top];
+
+ if (live_triangles[vertex] > 0)
+ return vertex;
+ }
+
+ // input order
+ while (input_cursor < vertex_count)
+ {
+ if (live_triangles[input_cursor] > 0)
+ return input_cursor;
+
+ ++input_cursor;
+ }
+
+ return ~0u;
+}
+
+static unsigned int getNextVertexNeighbour(const unsigned int* next_candidates_begin, const unsigned int* next_candidates_end, const unsigned int* live_triangles, const unsigned int* cache_timestamps, unsigned int timestamp, unsigned int cache_size)
+{
+ unsigned int best_candidate = ~0u;
+ int best_priority = -1;
+
+ for (const unsigned int* next_candidate = next_candidates_begin; next_candidate != next_candidates_end; ++next_candidate)
+ {
+ unsigned int vertex = *next_candidate;
+
+ // otherwise we don't need to process it
+ if (live_triangles[vertex] > 0)
+ {
+ int priority = 0;
+
+ // will it be in cache after fanning?
+ if (2 * live_triangles[vertex] + timestamp - cache_timestamps[vertex] <= cache_size)
+ {
+ priority = timestamp - cache_timestamps[vertex]; // position in cache
+ }
+
+ if (priority > best_priority)
+ {
+ best_candidate = vertex;
+ best_priority = priority;
+ }
+ }
+ }
+
+ return best_candidate;
+}
+
+static float vertexScore(const VertexScoreTable* table, int cache_position, unsigned int live_triangles)
+{
+ assert(cache_position >= -1 && cache_position < int(kCacheSizeMax));
+
+ unsigned int live_triangles_clamped = live_triangles < kValenceMax ? live_triangles : kValenceMax;
+
+ return table->cache[1 + cache_position] + table->live[live_triangles_clamped];
+}
+
+static unsigned int getNextTriangleDeadEnd(unsigned int& input_cursor, const unsigned char* emitted_flags, size_t face_count)
+{
+ // input order
+ while (input_cursor < face_count)
+ {
+ if (!emitted_flags[input_cursor])
+ return input_cursor;
+
+ ++input_cursor;
+ }
+
+ return ~0u;
+}
+
+} // namespace meshopt
+
+void meshopt_optimizeVertexCacheTable(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const meshopt::VertexScoreTable* table)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+
+ meshopt_Allocator allocator;
+
+ // guard for empty meshes
+ if (index_count == 0 || vertex_count == 0)
+ return;
+
+ // support in-place optimization
+ if (destination == indices)
+ {
+ unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
+ memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
+ indices = indices_copy;
+ }
+
+ unsigned int cache_size = 16;
+ assert(cache_size <= kCacheSizeMax);
+
+ size_t face_count = index_count / 3;
+
+ // build adjacency information
+ TriangleAdjacency adjacency = {};
+ buildTriangleAdjacency(adjacency, indices, index_count, vertex_count, allocator);
+
+ // live triangle counts
+ unsigned int* live_triangles = allocator.allocate<unsigned int>(vertex_count);
+ memcpy(live_triangles, adjacency.counts, vertex_count * sizeof(unsigned int));
+
+ // emitted flags
+ unsigned char* emitted_flags = allocator.allocate<unsigned char>(face_count);
+ memset(emitted_flags, 0, face_count);
+
+ // compute initial vertex scores
+ float* vertex_scores = allocator.allocate<float>(vertex_count);
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ vertex_scores[i] = vertexScore(table, -1, live_triangles[i]);
+
+ // compute triangle scores
+ float* triangle_scores = allocator.allocate<float>(face_count);
+
+ for (size_t i = 0; i < face_count; ++i)
+ {
+ unsigned int a = indices[i * 3 + 0];
+ unsigned int b = indices[i * 3 + 1];
+ unsigned int c = indices[i * 3 + 2];
+
+ triangle_scores[i] = vertex_scores[a] + vertex_scores[b] + vertex_scores[c];
+ }
+
+ unsigned int cache_holder[2 * (kCacheSizeMax + 3)];
+ unsigned int* cache = cache_holder;
+ unsigned int* cache_new = cache_holder + kCacheSizeMax + 3;
+ size_t cache_count = 0;
+
+ unsigned int current_triangle = 0;
+ unsigned int input_cursor = 1;
+
+ unsigned int output_triangle = 0;
+
+ while (current_triangle != ~0u)
+ {
+ assert(output_triangle < face_count);
+
+ unsigned int a = indices[current_triangle * 3 + 0];
+ unsigned int b = indices[current_triangle * 3 + 1];
+ unsigned int c = indices[current_triangle * 3 + 2];
+
+ // output indices
+ destination[output_triangle * 3 + 0] = a;
+ destination[output_triangle * 3 + 1] = b;
+ destination[output_triangle * 3 + 2] = c;
+ output_triangle++;
+
+ // update emitted flags
+ emitted_flags[current_triangle] = true;
+ triangle_scores[current_triangle] = 0;
+
+ // new triangle
+ size_t cache_write = 0;
+ cache_new[cache_write++] = a;
+ cache_new[cache_write++] = b;
+ cache_new[cache_write++] = c;
+
+ // old triangles
+ for (size_t i = 0; i < cache_count; ++i)
+ {
+ unsigned int index = cache[i];
+
+ if (index != a && index != b && index != c)
+ {
+ cache_new[cache_write++] = index;
+ }
+ }
+
+ unsigned int* cache_temp = cache;
+ cache = cache_new, cache_new = cache_temp;
+ cache_count = cache_write > cache_size ? cache_size : cache_write;
+
+ // update live triangle counts
+ live_triangles[a]--;
+ live_triangles[b]--;
+ live_triangles[c]--;
+
+ // remove emitted triangle from adjacency data
+ // this makes sure that we spend less time traversing these lists on subsequent iterations
+ for (size_t k = 0; k < 3; ++k)
+ {
+ unsigned int index = indices[current_triangle * 3 + k];
+
+ unsigned int* neighbours = &adjacency.data[0] + adjacency.offsets[index];
+ size_t neighbours_size = adjacency.counts[index];
+
+ for (size_t i = 0; i < neighbours_size; ++i)
+ {
+ unsigned int tri = neighbours[i];
+
+ if (tri == current_triangle)
+ {
+ neighbours[i] = neighbours[neighbours_size - 1];
+ adjacency.counts[index]--;
+ break;
+ }
+ }
+ }
+
+ unsigned int best_triangle = ~0u;
+ float best_score = 0;
+
+ // update cache positions, vertex scores and triangle scores, and find next best triangle
+ for (size_t i = 0; i < cache_write; ++i)
+ {
+ unsigned int index = cache[i];
+
+ int cache_position = i >= cache_size ? -1 : int(i);
+
+ // update vertex score
+ float score = vertexScore(table, cache_position, live_triangles[index]);
+ float score_diff = score - vertex_scores[index];
+
+ vertex_scores[index] = score;
+
+ // update scores of vertex triangles
+ const unsigned int* neighbours_begin = &adjacency.data[0] + adjacency.offsets[index];
+ const unsigned int* neighbours_end = neighbours_begin + adjacency.counts[index];
+
+ for (const unsigned int* it = neighbours_begin; it != neighbours_end; ++it)
+ {
+ unsigned int tri = *it;
+ assert(!emitted_flags[tri]);
+
+ float tri_score = triangle_scores[tri] + score_diff;
+ assert(tri_score > 0);
+
+ if (best_score < tri_score)
+ {
+ best_triangle = tri;
+ best_score = tri_score;
+ }
+
+ triangle_scores[tri] = tri_score;
+ }
+ }
+
+ // step through input triangles in order if we hit a dead-end
+ current_triangle = best_triangle;
+
+ if (current_triangle == ~0u)
+ {
+ current_triangle = getNextTriangleDeadEnd(input_cursor, &emitted_flags[0], face_count);
+ }
+ }
+
+ assert(input_cursor == face_count);
+ assert(output_triangle == face_count);
+}
+
+void meshopt_optimizeVertexCache(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count)
+{
+ meshopt_optimizeVertexCacheTable(destination, indices, index_count, vertex_count, &meshopt::kVertexScoreTable);
+}
+
+void meshopt_optimizeVertexCacheStrip(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count)
+{
+ meshopt_optimizeVertexCacheTable(destination, indices, index_count, vertex_count, &meshopt::kVertexScoreTableStrip);
+}
+
+void meshopt_optimizeVertexCacheFifo(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size)
+{
+ using namespace meshopt;
+
+ assert(index_count % 3 == 0);
+ assert(cache_size >= 3);
+
+ meshopt_Allocator allocator;
+
+ // guard for empty meshes
+ if (index_count == 0 || vertex_count == 0)
+ return;
+
+ // support in-place optimization
+ if (destination == indices)
+ {
+ unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
+ memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
+ indices = indices_copy;
+ }
+
+ size_t face_count = index_count / 3;
+
+ // build adjacency information
+ TriangleAdjacency adjacency = {};
+ buildTriangleAdjacency(adjacency, indices, index_count, vertex_count, allocator);
+
+ // live triangle counts
+ unsigned int* live_triangles = allocator.allocate<unsigned int>(vertex_count);
+ memcpy(live_triangles, adjacency.counts, vertex_count * sizeof(unsigned int));
+
+ // cache time stamps
+ unsigned int* cache_timestamps = allocator.allocate<unsigned int>(vertex_count);
+ memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
+
+ // dead-end stack
+ unsigned int* dead_end = allocator.allocate<unsigned int>(index_count);
+ unsigned int dead_end_top = 0;
+
+ // emitted flags
+ unsigned char* emitted_flags = allocator.allocate<unsigned char>(face_count);
+ memset(emitted_flags, 0, face_count);
+
+ unsigned int current_vertex = 0;
+
+ unsigned int timestamp = cache_size + 1;
+ unsigned int input_cursor = 1; // vertex to restart from in case of dead-end
+
+ unsigned int output_triangle = 0;
+
+ while (current_vertex != ~0u)
+ {
+ const unsigned int* next_candidates_begin = &dead_end[0] + dead_end_top;
+
+ // emit all vertex neighbours
+ const unsigned int* neighbours_begin = &adjacency.data[0] + adjacency.offsets[current_vertex];
+ const unsigned int* neighbours_end = neighbours_begin + adjacency.counts[current_vertex];
+
+ for (const unsigned int* it = neighbours_begin; it != neighbours_end; ++it)
+ {
+ unsigned int triangle = *it;
+
+ if (!emitted_flags[triangle])
+ {
+ unsigned int a = indices[triangle * 3 + 0], b = indices[triangle * 3 + 1], c = indices[triangle * 3 + 2];
+
+ // output indices
+ destination[output_triangle * 3 + 0] = a;
+ destination[output_triangle * 3 + 1] = b;
+ destination[output_triangle * 3 + 2] = c;
+ output_triangle++;
+
+ // update dead-end stack
+ dead_end[dead_end_top + 0] = a;
+ dead_end[dead_end_top + 1] = b;
+ dead_end[dead_end_top + 2] = c;
+ dead_end_top += 3;
+
+ // update live triangle counts
+ live_triangles[a]--;
+ live_triangles[b]--;
+ live_triangles[c]--;
+
+ // update cache info
+ // if vertex is not in cache, put it in cache
+ if (timestamp - cache_timestamps[a] > cache_size)
+ cache_timestamps[a] = timestamp++;
+
+ if (timestamp - cache_timestamps[b] > cache_size)
+ cache_timestamps[b] = timestamp++;
+
+ if (timestamp - cache_timestamps[c] > cache_size)
+ cache_timestamps[c] = timestamp++;
+
+ // update emitted flags
+ emitted_flags[triangle] = true;
+ }
+ }
+
+ // next candidates are the ones we pushed to dead-end stack just now
+ const unsigned int* next_candidates_end = &dead_end[0] + dead_end_top;
+
+ // get next vertex
+ current_vertex = getNextVertexNeighbour(next_candidates_begin, next_candidates_end, &live_triangles[0], &cache_timestamps[0], timestamp, cache_size);
+
+ if (current_vertex == ~0u)
+ {
+ current_vertex = getNextVertexDeadEnd(&dead_end[0], dead_end_top, input_cursor, &live_triangles[0], vertex_count);
+ }
+ }
+
+ assert(output_triangle == face_count);
+}
diff --git a/thirdparty/meshoptimizer/vertexcodec.cpp b/thirdparty/meshoptimizer/vertexcodec.cpp
new file mode 100644
index 0000000000..784c9a13db
--- /dev/null
+++ b/thirdparty/meshoptimizer/vertexcodec.cpp
@@ -0,0 +1,1265 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <string.h>
+
+// The block below auto-detects SIMD ISA that can be used on the target platform
+#ifndef MESHOPTIMIZER_NO_SIMD
+
+// The SIMD implementation requires SSSE3, which can be enabled unconditionally through compiler settings
+#if defined(__AVX__) || defined(__SSSE3__)
+#define SIMD_SSE
+#endif
+
+// An experimental implementation using AVX512 instructions; it's only enabled when AVX512 is enabled through compiler settings
+#if defined(__AVX512VBMI2__) && defined(__AVX512VBMI__) && defined(__AVX512VL__) && defined(__POPCNT__)
+#undef SIMD_SSE
+#define SIMD_AVX
+#endif
+
+// MSVC supports compiling SSSE3 code regardless of compile options; we use a cpuid-based scalar fallback
+#if !defined(SIMD_SSE) && !defined(SIMD_AVX) && defined(_MSC_VER) && !defined(__clang__) && (defined(_M_IX86) || defined(_M_X64))
+#define SIMD_SSE
+#define SIMD_FALLBACK
+#endif
+
+// GCC 4.9+ and clang 3.8+ support targeting SIMD ISA from individual functions; we use a cpuid-based scalar fallback
+#if !defined(SIMD_SSE) && !defined(SIMD_AVX) && ((defined(__clang__) && __clang_major__ * 100 + __clang_minor__ >= 308) || (defined(__GNUC__) && __GNUC__ * 100 + __GNUC_MINOR__ >= 409)) && (defined(__i386__) || defined(__x86_64__))
+#define SIMD_SSE
+#define SIMD_FALLBACK
+#define SIMD_TARGET __attribute__((target("ssse3")))
+#endif
+
+// GCC/clang define these when NEON support is available
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#define SIMD_NEON
+#endif
+
+// On MSVC, we assume that ARM builds always target NEON-capable devices
+#if !defined(SIMD_NEON) && defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+#define SIMD_NEON
+#endif
+
+// When targeting Wasm SIMD we can't use runtime cpuid checks so we unconditionally enable SIMD
+#if defined(__wasm_simd128__)
+#define SIMD_WASM
+#endif
+
+#ifndef SIMD_TARGET
+#define SIMD_TARGET
+#endif
+
+#endif // !MESHOPTIMIZER_NO_SIMD
+
+#ifdef SIMD_SSE
+#include <tmmintrin.h>
+#endif
+
+#if defined(SIMD_SSE) && defined(SIMD_FALLBACK)
+#ifdef _MSC_VER
+#include <intrin.h> // __cpuid
+#else
+#include <cpuid.h> // __cpuid
+#endif
+#endif
+
+#ifdef SIMD_AVX
+#include <immintrin.h>
+#endif
+
+#ifdef SIMD_NEON
+#if defined(_MSC_VER) && defined(_M_ARM64)
+#include <arm64_neon.h>
+#else
+#include <arm_neon.h>
+#endif
+#endif
+
+#ifdef SIMD_WASM
+#include <wasm_simd128.h>
+#endif
+
+#ifndef TRACE
+#define TRACE 0
+#endif
+
+#if TRACE
+#include <stdio.h>
+#endif
+
+#ifdef SIMD_WASM
+#define wasmx_splat_v32x4(v, i) wasm_v32x4_shuffle(v, v, i, i, i, i)
+#define wasmx_unpacklo_v8x16(a, b) wasm_v8x16_shuffle(a, b, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23)
+#define wasmx_unpackhi_v8x16(a, b) wasm_v8x16_shuffle(a, b, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31)
+#define wasmx_unpacklo_v16x8(a, b) wasm_v16x8_shuffle(a, b, 0, 8, 1, 9, 2, 10, 3, 11)
+#define wasmx_unpackhi_v16x8(a, b) wasm_v16x8_shuffle(a, b, 4, 12, 5, 13, 6, 14, 7, 15)
+#define wasmx_unpacklo_v64x2(a, b) wasm_v64x2_shuffle(a, b, 0, 2)
+#define wasmx_unpackhi_v64x2(a, b) wasm_v64x2_shuffle(a, b, 1, 3)
+#endif
+
+namespace meshopt
+{
+
+const unsigned char kVertexHeader = 0xa0;
+
+static int gEncodeVertexVersion = 0;
+
+const size_t kVertexBlockSizeBytes = 8192;
+const size_t kVertexBlockMaxSize = 256;
+const size_t kByteGroupSize = 16;
+const size_t kByteGroupDecodeLimit = 24;
+const size_t kTailMaxSize = 32;
+
+static size_t getVertexBlockSize(size_t vertex_size)
+{
+ // make sure the entire block fits into the scratch buffer
+ size_t result = kVertexBlockSizeBytes / vertex_size;
+
+ // align to byte group size; we encode each byte as a byte group
+ // if vertex block is misaligned, it results in wasted bytes, so just truncate the block size
+ result &= ~(kByteGroupSize - 1);
+
+ return (result < kVertexBlockMaxSize) ? result : kVertexBlockMaxSize;
+}
+
+inline unsigned char zigzag8(unsigned char v)
+{
+ return ((signed char)(v) >> 7) ^ (v << 1);
+}
+
+inline unsigned char unzigzag8(unsigned char v)
+{
+ return -(v & 1) ^ (v >> 1);
+}
+
+#if TRACE
+struct Stats
+{
+ size_t size;
+ size_t header;
+ size_t bitg[4];
+ size_t bitb[4];
+};
+
+Stats* bytestats;
+Stats vertexstats[256];
+#endif
+
+static bool encodeBytesGroupZero(const unsigned char* buffer)
+{
+ for (size_t i = 0; i < kByteGroupSize; ++i)
+ if (buffer[i])
+ return false;
+
+ return true;
+}
+
+static size_t encodeBytesGroupMeasure(const unsigned char* buffer, int bits)
+{
+ assert(bits >= 1 && bits <= 8);
+
+ if (bits == 1)
+ return encodeBytesGroupZero(buffer) ? 0 : size_t(-1);
+
+ if (bits == 8)
+ return kByteGroupSize;
+
+ size_t result = kByteGroupSize * bits / 8;
+
+ unsigned char sentinel = (1 << bits) - 1;
+
+ for (size_t i = 0; i < kByteGroupSize; ++i)
+ result += buffer[i] >= sentinel;
+
+ return result;
+}
+
+static unsigned char* encodeBytesGroup(unsigned char* data, const unsigned char* buffer, int bits)
+{
+ assert(bits >= 1 && bits <= 8);
+
+ if (bits == 1)
+ return data;
+
+ if (bits == 8)
+ {
+ memcpy(data, buffer, kByteGroupSize);
+ return data + kByteGroupSize;
+ }
+
+ size_t byte_size = 8 / bits;
+ assert(kByteGroupSize % byte_size == 0);
+
+ // fixed portion: bits bits for each value
+ // variable portion: full byte for each out-of-range value (using 1...1 as sentinel)
+ unsigned char sentinel = (1 << bits) - 1;
+
+ for (size_t i = 0; i < kByteGroupSize; i += byte_size)
+ {
+ unsigned char byte = 0;
+
+ for (size_t k = 0; k < byte_size; ++k)
+ {
+ unsigned char enc = (buffer[i + k] >= sentinel) ? sentinel : buffer[i + k];
+
+ byte <<= bits;
+ byte |= enc;
+ }
+
+ *data++ = byte;
+ }
+
+ for (size_t i = 0; i < kByteGroupSize; ++i)
+ {
+ if (buffer[i] >= sentinel)
+ {
+ *data++ = buffer[i];
+ }
+ }
+
+ return data;
+}
+
+static unsigned char* encodeBytes(unsigned char* data, unsigned char* data_end, const unsigned char* buffer, size_t buffer_size)
+{
+ assert(buffer_size % kByteGroupSize == 0);
+
+ unsigned char* header = data;
+
+ // round number of groups to 4 to get number of header bytes
+ size_t header_size = (buffer_size / kByteGroupSize + 3) / 4;
+
+ if (size_t(data_end - data) < header_size)
+ return 0;
+
+ data += header_size;
+
+ memset(header, 0, header_size);
+
+ for (size_t i = 0; i < buffer_size; i += kByteGroupSize)
+ {
+ if (size_t(data_end - data) < kByteGroupDecodeLimit)
+ return 0;
+
+ int best_bits = 8;
+ size_t best_size = encodeBytesGroupMeasure(buffer + i, 8);
+
+ for (int bits = 1; bits < 8; bits *= 2)
+ {
+ size_t size = encodeBytesGroupMeasure(buffer + i, bits);
+
+ if (size < best_size)
+ {
+ best_bits = bits;
+ best_size = size;
+ }
+ }
+
+ int bitslog2 = (best_bits == 1) ? 0 : (best_bits == 2) ? 1 : (best_bits == 4) ? 2 : 3;
+ assert((1 << bitslog2) == best_bits);
+
+ size_t header_offset = i / kByteGroupSize;
+
+ header[header_offset / 4] |= bitslog2 << ((header_offset % 4) * 2);
+
+ unsigned char* next = encodeBytesGroup(data, buffer + i, best_bits);
+
+ assert(data + best_size == next);
+ data = next;
+
+#if TRACE > 1
+ bytestats->bitg[bitslog2]++;
+ bytestats->bitb[bitslog2] += best_size;
+#endif
+ }
+
+#if TRACE > 1
+ bytestats->header += header_size;
+#endif
+
+ return data;
+}
+
+static unsigned char* encodeVertexBlock(unsigned char* data, unsigned char* data_end, const unsigned char* vertex_data, size_t vertex_count, size_t vertex_size, unsigned char last_vertex[256])
+{
+ assert(vertex_count > 0 && vertex_count <= kVertexBlockMaxSize);
+
+ unsigned char buffer[kVertexBlockMaxSize];
+ assert(sizeof(buffer) % kByteGroupSize == 0);
+
+ // we sometimes encode elements we didn't fill when rounding to kByteGroupSize
+ memset(buffer, 0, sizeof(buffer));
+
+ for (size_t k = 0; k < vertex_size; ++k)
+ {
+ size_t vertex_offset = k;
+
+ unsigned char p = last_vertex[k];
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ buffer[i] = zigzag8(vertex_data[vertex_offset] - p);
+
+ p = vertex_data[vertex_offset];
+
+ vertex_offset += vertex_size;
+ }
+
+#if TRACE
+ const unsigned char* olddata = data;
+ bytestats = &vertexstats[k];
+#endif
+
+ data = encodeBytes(data, data_end, buffer, (vertex_count + kByteGroupSize - 1) & ~(kByteGroupSize - 1));
+ if (!data)
+ return 0;
+
+#if TRACE
+ bytestats = 0;
+ vertexstats[k].size += data - olddata;
+#endif
+ }
+
+ memcpy(last_vertex, &vertex_data[vertex_size * (vertex_count - 1)], vertex_size);
+
+ return data;
+}
+
+#if defined(SIMD_FALLBACK) || (!defined(SIMD_SSE) && !defined(SIMD_NEON) && !defined(SIMD_AVX))
+static const unsigned char* decodeBytesGroup(const unsigned char* data, unsigned char* buffer, int bitslog2)
+{
+#define READ() byte = *data++
+#define NEXT(bits) enc = byte >> (8 - bits), byte <<= bits, encv = *data_var, *buffer++ = (enc == (1 << bits) - 1) ? encv : enc, data_var += (enc == (1 << bits) - 1)
+
+ unsigned char byte, enc, encv;
+ const unsigned char* data_var;
+
+ switch (bitslog2)
+ {
+ case 0:
+ memset(buffer, 0, kByteGroupSize);
+ return data;
+ case 1:
+ data_var = data + 4;
+
+ // 4 groups with 4 2-bit values in each byte
+ READ(), NEXT(2), NEXT(2), NEXT(2), NEXT(2);
+ READ(), NEXT(2), NEXT(2), NEXT(2), NEXT(2);
+ READ(), NEXT(2), NEXT(2), NEXT(2), NEXT(2);
+ READ(), NEXT(2), NEXT(2), NEXT(2), NEXT(2);
+
+ return data_var;
+ case 2:
+ data_var = data + 8;
+
+ // 8 groups with 2 4-bit values in each byte
+ READ(), NEXT(4), NEXT(4);
+ READ(), NEXT(4), NEXT(4);
+ READ(), NEXT(4), NEXT(4);
+ READ(), NEXT(4), NEXT(4);
+ READ(), NEXT(4), NEXT(4);
+ READ(), NEXT(4), NEXT(4);
+ READ(), NEXT(4), NEXT(4);
+ READ(), NEXT(4), NEXT(4);
+
+ return data_var;
+ case 3:
+ memcpy(buffer, data, kByteGroupSize);
+ return data + kByteGroupSize;
+ default:
+ assert(!"Unexpected bit length"); // unreachable since bitslog2 is a 2-bit value
+ return data;
+ }
+
+#undef READ
+#undef NEXT
+}
+
+static const unsigned char* decodeBytes(const unsigned char* data, const unsigned char* data_end, unsigned char* buffer, size_t buffer_size)
+{
+ assert(buffer_size % kByteGroupSize == 0);
+
+ const unsigned char* header = data;
+
+ // round number of groups to 4 to get number of header bytes
+ size_t header_size = (buffer_size / kByteGroupSize + 3) / 4;
+
+ if (size_t(data_end - data) < header_size)
+ return 0;
+
+ data += header_size;
+
+ for (size_t i = 0; i < buffer_size; i += kByteGroupSize)
+ {
+ if (size_t(data_end - data) < kByteGroupDecodeLimit)
+ return 0;
+
+ size_t header_offset = i / kByteGroupSize;
+
+ int bitslog2 = (header[header_offset / 4] >> ((header_offset % 4) * 2)) & 3;
+
+ data = decodeBytesGroup(data, buffer + i, bitslog2);
+ }
+
+ return data;
+}
+
+static const unsigned char* decodeVertexBlock(const unsigned char* data, const unsigned char* data_end, unsigned char* vertex_data, size_t vertex_count, size_t vertex_size, unsigned char last_vertex[256])
+{
+ assert(vertex_count > 0 && vertex_count <= kVertexBlockMaxSize);
+
+ unsigned char buffer[kVertexBlockMaxSize];
+ unsigned char transposed[kVertexBlockSizeBytes];
+
+ size_t vertex_count_aligned = (vertex_count + kByteGroupSize - 1) & ~(kByteGroupSize - 1);
+
+ for (size_t k = 0; k < vertex_size; ++k)
+ {
+ data = decodeBytes(data, data_end, buffer, vertex_count_aligned);
+ if (!data)
+ return 0;
+
+ size_t vertex_offset = k;
+
+ unsigned char p = last_vertex[k];
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ {
+ unsigned char v = unzigzag8(buffer[i]) + p;
+
+ transposed[vertex_offset] = v;
+ p = v;
+
+ vertex_offset += vertex_size;
+ }
+ }
+
+ memcpy(vertex_data, transposed, vertex_count * vertex_size);
+
+ memcpy(last_vertex, &transposed[vertex_size * (vertex_count - 1)], vertex_size);
+
+ return data;
+}
+#endif
+
+#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
+static unsigned char kDecodeBytesGroupShuffle[256][8];
+static unsigned char kDecodeBytesGroupCount[256];
+
+#ifdef __wasm__
+__attribute__((cold)) // this saves 500 bytes in the output binary - we don't need to vectorize this loop!
+#endif
+static bool
+decodeBytesGroupBuildTables()
+{
+ for (int mask = 0; mask < 256; ++mask)
+ {
+ unsigned char shuffle[8];
+ unsigned char count = 0;
+
+ for (int i = 0; i < 8; ++i)
+ {
+ int maski = (mask >> i) & 1;
+ shuffle[i] = maski ? count : 0x80;
+ count += (unsigned char)(maski);
+ }
+
+ memcpy(kDecodeBytesGroupShuffle[mask], shuffle, 8);
+ kDecodeBytesGroupCount[mask] = count;
+ }
+
+ return true;
+}
+
+static bool gDecodeBytesGroupInitialized = decodeBytesGroupBuildTables();
+#endif
+
+#ifdef SIMD_SSE
+SIMD_TARGET
+static __m128i decodeShuffleMask(unsigned char mask0, unsigned char mask1)
+{
+ __m128i sm0 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(&kDecodeBytesGroupShuffle[mask0]));
+ __m128i sm1 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(&kDecodeBytesGroupShuffle[mask1]));
+ __m128i sm1off = _mm_set1_epi8(kDecodeBytesGroupCount[mask0]);
+
+ __m128i sm1r = _mm_add_epi8(sm1, sm1off);
+
+ return _mm_unpacklo_epi64(sm0, sm1r);
+}
+
+SIMD_TARGET
+static const unsigned char* decodeBytesGroupSimd(const unsigned char* data, unsigned char* buffer, int bitslog2)
+{
+ switch (bitslog2)
+ {
+ case 0:
+ {
+ __m128i result = _mm_setzero_si128();
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
+
+ return data;
+ }
+
+ case 1:
+ {
+#ifdef __GNUC__
+ typedef int __attribute__((aligned(1))) unaligned_int;
+#else
+ typedef int unaligned_int;
+#endif
+
+ __m128i sel2 = _mm_cvtsi32_si128(*reinterpret_cast<const unaligned_int*>(data));
+ __m128i rest = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + 4));
+
+ __m128i sel22 = _mm_unpacklo_epi8(_mm_srli_epi16(sel2, 4), sel2);
+ __m128i sel2222 = _mm_unpacklo_epi8(_mm_srli_epi16(sel22, 2), sel22);
+ __m128i sel = _mm_and_si128(sel2222, _mm_set1_epi8(3));
+
+ __m128i mask = _mm_cmpeq_epi8(sel, _mm_set1_epi8(3));
+ int mask16 = _mm_movemask_epi8(mask);
+ unsigned char mask0 = (unsigned char)(mask16 & 255);
+ unsigned char mask1 = (unsigned char)(mask16 >> 8);
+
+ __m128i shuf = decodeShuffleMask(mask0, mask1);
+
+ __m128i result = _mm_or_si128(_mm_shuffle_epi8(rest, shuf), _mm_andnot_si128(mask, sel));
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
+
+ return data + 4 + kDecodeBytesGroupCount[mask0] + kDecodeBytesGroupCount[mask1];
+ }
+
+ case 2:
+ {
+ __m128i sel4 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(data));
+ __m128i rest = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + 8));
+
+ __m128i sel44 = _mm_unpacklo_epi8(_mm_srli_epi16(sel4, 4), sel4);
+ __m128i sel = _mm_and_si128(sel44, _mm_set1_epi8(15));
+
+ __m128i mask = _mm_cmpeq_epi8(sel, _mm_set1_epi8(15));
+ int mask16 = _mm_movemask_epi8(mask);
+ unsigned char mask0 = (unsigned char)(mask16 & 255);
+ unsigned char mask1 = (unsigned char)(mask16 >> 8);
+
+ __m128i shuf = decodeShuffleMask(mask0, mask1);
+
+ __m128i result = _mm_or_si128(_mm_shuffle_epi8(rest, shuf), _mm_andnot_si128(mask, sel));
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
+
+ return data + 8 + kDecodeBytesGroupCount[mask0] + kDecodeBytesGroupCount[mask1];
+ }
+
+ case 3:
+ {
+ __m128i result = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data));
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
+
+ return data + 16;
+ }
+
+ default:
+ assert(!"Unexpected bit length"); // unreachable since bitslog2 is a 2-bit value
+ return data;
+ }
+}
+#endif
+
+#ifdef SIMD_AVX
+static const __m128i decodeBytesGroupConfig[] = {
+ _mm_set1_epi8(3),
+ _mm_set1_epi8(15),
+ _mm_setr_epi8(6, 4, 2, 0, 14, 12, 10, 8, 22, 20, 18, 16, 30, 28, 26, 24),
+ _mm_setr_epi8(4, 0, 12, 8, 20, 16, 28, 24, 36, 32, 44, 40, 52, 48, 60, 56),
+};
+
+static const unsigned char* decodeBytesGroupSimd(const unsigned char* data, unsigned char* buffer, int bitslog2)
+{
+ switch (bitslog2)
+ {
+ case 0:
+ {
+ __m128i result = _mm_setzero_si128();
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
+
+ return data;
+ }
+
+ case 1:
+ case 2:
+ {
+ const unsigned char* skip = data + (bitslog2 << 2);
+
+ __m128i selb = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(data));
+ __m128i rest = _mm_loadu_si128(reinterpret_cast<const __m128i*>(skip));
+
+ __m128i sent = decodeBytesGroupConfig[bitslog2 - 1];
+ __m128i ctrl = decodeBytesGroupConfig[bitslog2 + 1];
+
+ __m128i selw = _mm_shuffle_epi32(selb, 0x44);
+ __m128i sel = _mm_and_si128(sent, _mm_multishift_epi64_epi8(ctrl, selw));
+ __mmask16 mask16 = _mm_cmp_epi8_mask(sel, sent, _MM_CMPINT_EQ);
+
+ __m128i result = _mm_mask_expand_epi8(sel, mask16, rest);
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
+
+ return skip + _mm_popcnt_u32(mask16);
+ }
+
+ case 3:
+ {
+ __m128i result = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data));
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(buffer), result);
+
+ return data + 16;
+ }
+
+ default:
+ assert(!"Unexpected bit length"); // unreachable since bitslog2 is a 2-bit value
+ return data;
+ }
+}
+#endif
+
+#ifdef SIMD_NEON
+static uint8x16_t shuffleBytes(unsigned char mask0, unsigned char mask1, uint8x8_t rest0, uint8x8_t rest1)
+{
+ uint8x8_t sm0 = vld1_u8(kDecodeBytesGroupShuffle[mask0]);
+ uint8x8_t sm1 = vld1_u8(kDecodeBytesGroupShuffle[mask1]);
+
+ uint8x8_t r0 = vtbl1_u8(rest0, sm0);
+ uint8x8_t r1 = vtbl1_u8(rest1, sm1);
+
+ return vcombine_u8(r0, r1);
+}
+
+static void neonMoveMask(uint8x16_t mask, unsigned char& mask0, unsigned char& mask1)
+{
+ static const unsigned char byte_mask_data[16] = {1, 2, 4, 8, 16, 32, 64, 128, 1, 2, 4, 8, 16, 32, 64, 128};
+
+ uint8x16_t byte_mask = vld1q_u8(byte_mask_data);
+ uint8x16_t masked = vandq_u8(mask, byte_mask);
+
+#ifdef __aarch64__
+ // aarch64 has horizontal sums; MSVC doesn't expose this via arm64_neon.h so this path is exclusive to clang/gcc
+ mask0 = vaddv_u8(vget_low_u8(masked));
+ mask1 = vaddv_u8(vget_high_u8(masked));
+#else
+ // we need horizontal sums of each half of masked, which can be done in 3 steps (yielding sums of sizes 2, 4, 8)
+ uint8x8_t sum1 = vpadd_u8(vget_low_u8(masked), vget_high_u8(masked));
+ uint8x8_t sum2 = vpadd_u8(sum1, sum1);
+ uint8x8_t sum3 = vpadd_u8(sum2, sum2);
+
+ mask0 = vget_lane_u8(sum3, 0);
+ mask1 = vget_lane_u8(sum3, 1);
+#endif
+}
+
+static const unsigned char* decodeBytesGroupSimd(const unsigned char* data, unsigned char* buffer, int bitslog2)
+{
+ switch (bitslog2)
+ {
+ case 0:
+ {
+ uint8x16_t result = vdupq_n_u8(0);
+
+ vst1q_u8(buffer, result);
+
+ return data;
+ }
+
+ case 1:
+ {
+ uint8x8_t sel2 = vld1_u8(data);
+ uint8x8_t sel22 = vzip_u8(vshr_n_u8(sel2, 4), sel2).val[0];
+ uint8x8x2_t sel2222 = vzip_u8(vshr_n_u8(sel22, 2), sel22);
+ uint8x16_t sel = vandq_u8(vcombine_u8(sel2222.val[0], sel2222.val[1]), vdupq_n_u8(3));
+
+ uint8x16_t mask = vceqq_u8(sel, vdupq_n_u8(3));
+ unsigned char mask0, mask1;
+ neonMoveMask(mask, mask0, mask1);
+
+ uint8x8_t rest0 = vld1_u8(data + 4);
+ uint8x8_t rest1 = vld1_u8(data + 4 + kDecodeBytesGroupCount[mask0]);
+
+ uint8x16_t result = vbslq_u8(mask, shuffleBytes(mask0, mask1, rest0, rest1), sel);
+
+ vst1q_u8(buffer, result);
+
+ return data + 4 + kDecodeBytesGroupCount[mask0] + kDecodeBytesGroupCount[mask1];
+ }
+
+ case 2:
+ {
+ uint8x8_t sel4 = vld1_u8(data);
+ uint8x8x2_t sel44 = vzip_u8(vshr_n_u8(sel4, 4), vand_u8(sel4, vdup_n_u8(15)));
+ uint8x16_t sel = vcombine_u8(sel44.val[0], sel44.val[1]);
+
+ uint8x16_t mask = vceqq_u8(sel, vdupq_n_u8(15));
+ unsigned char mask0, mask1;
+ neonMoveMask(mask, mask0, mask1);
+
+ uint8x8_t rest0 = vld1_u8(data + 8);
+ uint8x8_t rest1 = vld1_u8(data + 8 + kDecodeBytesGroupCount[mask0]);
+
+ uint8x16_t result = vbslq_u8(mask, shuffleBytes(mask0, mask1, rest0, rest1), sel);
+
+ vst1q_u8(buffer, result);
+
+ return data + 8 + kDecodeBytesGroupCount[mask0] + kDecodeBytesGroupCount[mask1];
+ }
+
+ case 3:
+ {
+ uint8x16_t result = vld1q_u8(data);
+
+ vst1q_u8(buffer, result);
+
+ return data + 16;
+ }
+
+ default:
+ assert(!"Unexpected bit length"); // unreachable since bitslog2 is a 2-bit value
+ return data;
+ }
+}
+#endif
+
+#ifdef SIMD_WASM
+SIMD_TARGET
+static v128_t decodeShuffleMask(unsigned char mask0, unsigned char mask1)
+{
+ v128_t sm0 = wasm_v128_load(&kDecodeBytesGroupShuffle[mask0]);
+ v128_t sm1 = wasm_v128_load(&kDecodeBytesGroupShuffle[mask1]);
+
+ v128_t sm1off = wasm_v128_load(&kDecodeBytesGroupCount[mask0]);
+ sm1off = wasm_v8x16_shuffle(sm1off, sm1off, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
+
+ v128_t sm1r = wasm_i8x16_add(sm1, sm1off);
+
+ return wasmx_unpacklo_v64x2(sm0, sm1r);
+}
+
+SIMD_TARGET
+static void wasmMoveMask(v128_t mask, unsigned char& mask0, unsigned char& mask1)
+{
+ v128_t mask_0 = wasm_v32x4_shuffle(mask, mask, 0, 2, 1, 3);
+
+ uint64_t mask_1a = wasm_i64x2_extract_lane(mask_0, 0) & 0x0804020108040201ull;
+ uint64_t mask_1b = wasm_i64x2_extract_lane(mask_0, 1) & 0x8040201080402010ull;
+
+ // TODO: This can use v8x16_bitmask in the future
+ uint64_t mask_2 = mask_1a | mask_1b;
+ uint64_t mask_4 = mask_2 | (mask_2 >> 16);
+ uint64_t mask_8 = mask_4 | (mask_4 >> 8);
+
+ mask0 = uint8_t(mask_8);
+ mask1 = uint8_t(mask_8 >> 32);
+}
+
+SIMD_TARGET
+static const unsigned char* decodeBytesGroupSimd(const unsigned char* data, unsigned char* buffer, int bitslog2)
+{
+ unsigned char byte, enc, encv;
+ const unsigned char* data_var;
+
+ switch (bitslog2)
+ {
+ case 0:
+ {
+ v128_t result = wasm_i8x16_splat(0);
+
+ wasm_v128_store(buffer, result);
+
+ return data;
+ }
+
+ case 1:
+ {
+ v128_t sel2 = wasm_v128_load(data);
+ v128_t rest = wasm_v128_load(data + 4);
+
+ v128_t sel22 = wasmx_unpacklo_v8x16(wasm_i16x8_shr(sel2, 4), sel2);
+ v128_t sel2222 = wasmx_unpacklo_v8x16(wasm_i16x8_shr(sel22, 2), sel22);
+ v128_t sel = wasm_v128_and(sel2222, wasm_i8x16_splat(3));
+
+ v128_t mask = wasm_i8x16_eq(sel, wasm_i8x16_splat(3));
+
+ unsigned char mask0, mask1;
+ wasmMoveMask(mask, mask0, mask1);
+
+ v128_t shuf = decodeShuffleMask(mask0, mask1);
+
+ v128_t result = wasm_v128_bitselect(wasm_v8x16_swizzle(rest, shuf), sel, mask);
+
+ wasm_v128_store(buffer, result);
+
+ return data + 4 + kDecodeBytesGroupCount[mask0] + kDecodeBytesGroupCount[mask1];
+ }
+
+ case 2:
+ {
+ v128_t sel4 = wasm_v128_load(data);
+ v128_t rest = wasm_v128_load(data + 8);
+
+ v128_t sel44 = wasmx_unpacklo_v8x16(wasm_i16x8_shr(sel4, 4), sel4);
+ v128_t sel = wasm_v128_and(sel44, wasm_i8x16_splat(15));
+
+ v128_t mask = wasm_i8x16_eq(sel, wasm_i8x16_splat(15));
+
+ unsigned char mask0, mask1;
+ wasmMoveMask(mask, mask0, mask1);
+
+ v128_t shuf = decodeShuffleMask(mask0, mask1);
+
+ v128_t result = wasm_v128_bitselect(wasm_v8x16_swizzle(rest, shuf), sel, mask);
+
+ wasm_v128_store(buffer, result);
+
+ return data + 8 + kDecodeBytesGroupCount[mask0] + kDecodeBytesGroupCount[mask1];
+ }
+
+ case 3:
+ {
+ v128_t result = wasm_v128_load(data);
+
+ wasm_v128_store(buffer, result);
+
+ return data + 16;
+ }
+
+ default:
+ assert(!"Unexpected bit length"); // unreachable since bitslog2 is a 2-bit value
+ return data;
+ }
+}
+#endif
+
+#if defined(SIMD_SSE) || defined(SIMD_AVX)
+SIMD_TARGET
+static void transpose8(__m128i& x0, __m128i& x1, __m128i& x2, __m128i& x3)
+{
+ __m128i t0 = _mm_unpacklo_epi8(x0, x1);
+ __m128i t1 = _mm_unpackhi_epi8(x0, x1);
+ __m128i t2 = _mm_unpacklo_epi8(x2, x3);
+ __m128i t3 = _mm_unpackhi_epi8(x2, x3);
+
+ x0 = _mm_unpacklo_epi16(t0, t2);
+ x1 = _mm_unpackhi_epi16(t0, t2);
+ x2 = _mm_unpacklo_epi16(t1, t3);
+ x3 = _mm_unpackhi_epi16(t1, t3);
+}
+
+SIMD_TARGET
+static __m128i unzigzag8(__m128i v)
+{
+ __m128i xl = _mm_sub_epi8(_mm_setzero_si128(), _mm_and_si128(v, _mm_set1_epi8(1)));
+ __m128i xr = _mm_and_si128(_mm_srli_epi16(v, 1), _mm_set1_epi8(127));
+
+ return _mm_xor_si128(xl, xr);
+}
+#endif
+
+#ifdef SIMD_NEON
+static void transpose8(uint8x16_t& x0, uint8x16_t& x1, uint8x16_t& x2, uint8x16_t& x3)
+{
+ uint8x16x2_t t01 = vzipq_u8(x0, x1);
+ uint8x16x2_t t23 = vzipq_u8(x2, x3);
+
+ uint16x8x2_t x01 = vzipq_u16(vreinterpretq_u16_u8(t01.val[0]), vreinterpretq_u16_u8(t23.val[0]));
+ uint16x8x2_t x23 = vzipq_u16(vreinterpretq_u16_u8(t01.val[1]), vreinterpretq_u16_u8(t23.val[1]));
+
+ x0 = vreinterpretq_u8_u16(x01.val[0]);
+ x1 = vreinterpretq_u8_u16(x01.val[1]);
+ x2 = vreinterpretq_u8_u16(x23.val[0]);
+ x3 = vreinterpretq_u8_u16(x23.val[1]);
+}
+
+static uint8x16_t unzigzag8(uint8x16_t v)
+{
+ uint8x16_t xl = vreinterpretq_u8_s8(vnegq_s8(vreinterpretq_s8_u8(vandq_u8(v, vdupq_n_u8(1)))));
+ uint8x16_t xr = vshrq_n_u8(v, 1);
+
+ return veorq_u8(xl, xr);
+}
+#endif
+
+#ifdef SIMD_WASM
+SIMD_TARGET
+static void transpose8(v128_t& x0, v128_t& x1, v128_t& x2, v128_t& x3)
+{
+ v128_t t0 = wasmx_unpacklo_v8x16(x0, x1);
+ v128_t t1 = wasmx_unpackhi_v8x16(x0, x1);
+ v128_t t2 = wasmx_unpacklo_v8x16(x2, x3);
+ v128_t t3 = wasmx_unpackhi_v8x16(x2, x3);
+
+ x0 = wasmx_unpacklo_v16x8(t0, t2);
+ x1 = wasmx_unpackhi_v16x8(t0, t2);
+ x2 = wasmx_unpacklo_v16x8(t1, t3);
+ x3 = wasmx_unpackhi_v16x8(t1, t3);
+}
+
+SIMD_TARGET
+static v128_t unzigzag8(v128_t v)
+{
+ v128_t xl = wasm_i8x16_neg(wasm_v128_and(v, wasm_i8x16_splat(1)));
+ v128_t xr = wasm_u8x16_shr(v, 1);
+
+ return wasm_v128_xor(xl, xr);
+}
+#endif
+
+#if defined(SIMD_SSE) || defined(SIMD_AVX) || defined(SIMD_NEON) || defined(SIMD_WASM)
+SIMD_TARGET
+static const unsigned char* decodeBytesSimd(const unsigned char* data, const unsigned char* data_end, unsigned char* buffer, size_t buffer_size)
+{
+ assert(buffer_size % kByteGroupSize == 0);
+ assert(kByteGroupSize == 16);
+
+ const unsigned char* header = data;
+
+ // round number of groups to 4 to get number of header bytes
+ size_t header_size = (buffer_size / kByteGroupSize + 3) / 4;
+
+ if (size_t(data_end - data) < header_size)
+ return 0;
+
+ data += header_size;
+
+ size_t i = 0;
+
+ // fast-path: process 4 groups at a time, do a shared bounds check - each group reads <=24b
+ for (; i + kByteGroupSize * 4 <= buffer_size && size_t(data_end - data) >= kByteGroupDecodeLimit * 4; i += kByteGroupSize * 4)
+ {
+ size_t header_offset = i / kByteGroupSize;
+ unsigned char header_byte = header[header_offset / 4];
+
+ data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 0, (header_byte >> 0) & 3);
+ data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 1, (header_byte >> 2) & 3);
+ data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 2, (header_byte >> 4) & 3);
+ data = decodeBytesGroupSimd(data, buffer + i + kByteGroupSize * 3, (header_byte >> 6) & 3);
+ }
+
+ // slow-path: process remaining groups
+ for (; i < buffer_size; i += kByteGroupSize)
+ {
+ if (size_t(data_end - data) < kByteGroupDecodeLimit)
+ return 0;
+
+ size_t header_offset = i / kByteGroupSize;
+
+ int bitslog2 = (header[header_offset / 4] >> ((header_offset % 4) * 2)) & 3;
+
+ data = decodeBytesGroupSimd(data, buffer + i, bitslog2);
+ }
+
+ return data;
+}
+
+SIMD_TARGET
+static const unsigned char* decodeVertexBlockSimd(const unsigned char* data, const unsigned char* data_end, unsigned char* vertex_data, size_t vertex_count, size_t vertex_size, unsigned char last_vertex[256])
+{
+ assert(vertex_count > 0 && vertex_count <= kVertexBlockMaxSize);
+
+ unsigned char buffer[kVertexBlockMaxSize * 4];
+ unsigned char transposed[kVertexBlockSizeBytes];
+
+ size_t vertex_count_aligned = (vertex_count + kByteGroupSize - 1) & ~(kByteGroupSize - 1);
+
+ for (size_t k = 0; k < vertex_size; k += 4)
+ {
+ for (size_t j = 0; j < 4; ++j)
+ {
+ data = decodeBytesSimd(data, data_end, buffer + j * vertex_count_aligned, vertex_count_aligned);
+ if (!data)
+ return 0;
+ }
+
+#if defined(SIMD_SSE) || defined(SIMD_AVX)
+#define TEMP __m128i
+#define PREP() __m128i pi = _mm_cvtsi32_si128(*reinterpret_cast<const int*>(last_vertex + k))
+#define LOAD(i) __m128i r##i = _mm_loadu_si128(reinterpret_cast<const __m128i*>(buffer + j + i * vertex_count_aligned))
+#define GRP4(i) t0 = _mm_shuffle_epi32(r##i, 0), t1 = _mm_shuffle_epi32(r##i, 1), t2 = _mm_shuffle_epi32(r##i, 2), t3 = _mm_shuffle_epi32(r##i, 3)
+#define FIXD(i) t##i = pi = _mm_add_epi8(pi, t##i)
+#define SAVE(i) *reinterpret_cast<int*>(savep) = _mm_cvtsi128_si32(t##i), savep += vertex_size
+#endif
+
+#ifdef SIMD_NEON
+#define TEMP uint8x8_t
+#define PREP() uint8x8_t pi = vreinterpret_u8_u32(vld1_lane_u32(reinterpret_cast<uint32_t*>(last_vertex + k), vdup_n_u32(0), 0))
+#define LOAD(i) uint8x16_t r##i = vld1q_u8(buffer + j + i * vertex_count_aligned)
+#define GRP4(i) t0 = vget_low_u8(r##i), t1 = vreinterpret_u8_u32(vdup_lane_u32(vreinterpret_u32_u8(t0), 1)), t2 = vget_high_u8(r##i), t3 = vreinterpret_u8_u32(vdup_lane_u32(vreinterpret_u32_u8(t2), 1))
+#define FIXD(i) t##i = pi = vadd_u8(pi, t##i)
+#define SAVE(i) vst1_lane_u32(reinterpret_cast<uint32_t*>(savep), vreinterpret_u32_u8(t##i), 0), savep += vertex_size
+#endif
+
+#ifdef SIMD_WASM
+#define TEMP v128_t
+#define PREP() v128_t pi = wasm_v128_load(last_vertex + k)
+#define LOAD(i) v128_t r##i = wasm_v128_load(buffer + j + i * vertex_count_aligned)
+#define GRP4(i) t0 = wasmx_splat_v32x4(r##i, 0), t1 = wasmx_splat_v32x4(r##i, 1), t2 = wasmx_splat_v32x4(r##i, 2), t3 = wasmx_splat_v32x4(r##i, 3)
+#define FIXD(i) t##i = pi = wasm_i8x16_add(pi, t##i)
+#define SAVE(i) *reinterpret_cast<int*>(savep) = wasm_i32x4_extract_lane(t##i, 0), savep += vertex_size
+#endif
+
+ PREP();
+
+ unsigned char* savep = transposed + k;
+
+ for (size_t j = 0; j < vertex_count_aligned; j += 16)
+ {
+ LOAD(0);
+ LOAD(1);
+ LOAD(2);
+ LOAD(3);
+
+ r0 = unzigzag8(r0);
+ r1 = unzigzag8(r1);
+ r2 = unzigzag8(r2);
+ r3 = unzigzag8(r3);
+
+ transpose8(r0, r1, r2, r3);
+
+ TEMP t0, t1, t2, t3;
+
+ GRP4(0);
+ FIXD(0), FIXD(1), FIXD(2), FIXD(3);
+ SAVE(0), SAVE(1), SAVE(2), SAVE(3);
+
+ GRP4(1);
+ FIXD(0), FIXD(1), FIXD(2), FIXD(3);
+ SAVE(0), SAVE(1), SAVE(2), SAVE(3);
+
+ GRP4(2);
+ FIXD(0), FIXD(1), FIXD(2), FIXD(3);
+ SAVE(0), SAVE(1), SAVE(2), SAVE(3);
+
+ GRP4(3);
+ FIXD(0), FIXD(1), FIXD(2), FIXD(3);
+ SAVE(0), SAVE(1), SAVE(2), SAVE(3);
+
+#undef TEMP
+#undef PREP
+#undef LOAD
+#undef GRP4
+#undef FIXD
+#undef SAVE
+ }
+ }
+
+ memcpy(vertex_data, transposed, vertex_count * vertex_size);
+
+ memcpy(last_vertex, &transposed[vertex_size * (vertex_count - 1)], vertex_size);
+
+ return data;
+}
+#endif
+
+#if defined(SIMD_SSE) && defined(SIMD_FALLBACK)
+static unsigned int getCpuFeatures()
+{
+ int cpuinfo[4] = {};
+#ifdef _MSC_VER
+ __cpuid(cpuinfo, 1);
+#else
+ __cpuid(1, cpuinfo[0], cpuinfo[1], cpuinfo[2], cpuinfo[3]);
+#endif
+ return cpuinfo[2];
+}
+
+unsigned int cpuid = getCpuFeatures();
+#endif
+
+} // namespace meshopt
+
+size_t meshopt_encodeVertexBuffer(unsigned char* buffer, size_t buffer_size, const void* vertices, size_t vertex_count, size_t vertex_size)
+{
+ using namespace meshopt;
+
+ assert(vertex_size > 0 && vertex_size <= 256);
+ assert(vertex_size % 4 == 0);
+
+#if TRACE
+ memset(vertexstats, 0, sizeof(vertexstats));
+#endif
+
+ const unsigned char* vertex_data = static_cast<const unsigned char*>(vertices);
+
+ unsigned char* data = buffer;
+ unsigned char* data_end = buffer + buffer_size;
+
+ if (size_t(data_end - data) < 1 + vertex_size)
+ return 0;
+
+ int version = gEncodeVertexVersion;
+
+ *data++ = (unsigned char)(kVertexHeader | version);
+
+ unsigned char first_vertex[256] = {};
+ if (vertex_count > 0)
+ memcpy(first_vertex, vertex_data, vertex_size);
+
+ unsigned char last_vertex[256] = {};
+ memcpy(last_vertex, first_vertex, vertex_size);
+
+ size_t vertex_block_size = getVertexBlockSize(vertex_size);
+
+ size_t vertex_offset = 0;
+
+ while (vertex_offset < vertex_count)
+ {
+ size_t block_size = (vertex_offset + vertex_block_size < vertex_count) ? vertex_block_size : vertex_count - vertex_offset;
+
+ data = encodeVertexBlock(data, data_end, vertex_data + vertex_offset * vertex_size, block_size, vertex_size, last_vertex);
+ if (!data)
+ return 0;
+
+ vertex_offset += block_size;
+ }
+
+ size_t tail_size = vertex_size < kTailMaxSize ? kTailMaxSize : vertex_size;
+
+ if (size_t(data_end - data) < tail_size)
+ return 0;
+
+ // write first vertex to the end of the stream and pad it to 32 bytes; this is important to simplify bounds checks in decoder
+ if (vertex_size < kTailMaxSize)
+ {
+ memset(data, 0, kTailMaxSize - vertex_size);
+ data += kTailMaxSize - vertex_size;
+ }
+
+ memcpy(data, first_vertex, vertex_size);
+ data += vertex_size;
+
+ assert(data >= buffer + tail_size);
+ assert(data <= buffer + buffer_size);
+
+#if TRACE
+ size_t total_size = data - buffer;
+
+ for (size_t k = 0; k < vertex_size; ++k)
+ {
+ const Stats& vsk = vertexstats[k];
+
+ printf("%2d: %d bytes\t%.1f%%\t%.1f bpv", int(k), int(vsk.size), double(vsk.size) / double(total_size) * 100, double(vsk.size) / double(vertex_count) * 8);
+
+#if TRACE > 1
+ printf("\t\thdr %d bytes\tbit0 %d (%d bytes)\tbit1 %d (%d bytes)\tbit2 %d (%d bytes)\tbit3 %d (%d bytes)",
+ int(vsk.header),
+ int(vsk.bitg[0]), int(vsk.bitb[0]),
+ int(vsk.bitg[1]), int(vsk.bitb[1]),
+ int(vsk.bitg[2]), int(vsk.bitb[2]),
+ int(vsk.bitg[3]), int(vsk.bitb[3]));
+#endif
+
+ printf("\n");
+ }
+#endif
+
+ return data - buffer;
+}
+
+size_t meshopt_encodeVertexBufferBound(size_t vertex_count, size_t vertex_size)
+{
+ using namespace meshopt;
+
+ assert(vertex_size > 0 && vertex_size <= 256);
+ assert(vertex_size % 4 == 0);
+
+ size_t vertex_block_size = getVertexBlockSize(vertex_size);
+ size_t vertex_block_count = (vertex_count + vertex_block_size - 1) / vertex_block_size;
+
+ size_t vertex_block_header_size = (vertex_block_size / kByteGroupSize + 3) / 4;
+ size_t vertex_block_data_size = vertex_block_size;
+
+ size_t tail_size = vertex_size < kTailMaxSize ? kTailMaxSize : vertex_size;
+
+ return 1 + vertex_block_count * vertex_size * (vertex_block_header_size + vertex_block_data_size) + tail_size;
+}
+
+void meshopt_encodeVertexVersion(int version)
+{
+ assert(unsigned(version) <= 0);
+
+ meshopt::gEncodeVertexVersion = version;
+}
+
+int meshopt_decodeVertexBuffer(void* destination, size_t vertex_count, size_t vertex_size, const unsigned char* buffer, size_t buffer_size)
+{
+ using namespace meshopt;
+
+ assert(vertex_size > 0 && vertex_size <= 256);
+ assert(vertex_size % 4 == 0);
+
+ const unsigned char* (*decode)(const unsigned char*, const unsigned char*, unsigned char*, size_t, size_t, unsigned char[256]) = 0;
+
+#if defined(SIMD_SSE) && defined(SIMD_FALLBACK)
+ decode = (cpuid & (1 << 9)) ? decodeVertexBlockSimd : decodeVertexBlock;
+#elif defined(SIMD_SSE) || defined(SIMD_AVX) || defined(SIMD_NEON) || defined(SIMD_WASM)
+ decode = decodeVertexBlockSimd;
+#else
+ decode = decodeVertexBlock;
+#endif
+
+#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
+ assert(gDecodeBytesGroupInitialized);
+ (void)gDecodeBytesGroupInitialized;
+#endif
+
+ unsigned char* vertex_data = static_cast<unsigned char*>(destination);
+
+ const unsigned char* data = buffer;
+ const unsigned char* data_end = buffer + buffer_size;
+
+ if (size_t(data_end - data) < 1 + vertex_size)
+ return -2;
+
+ unsigned char data_header = *data++;
+
+ if ((data_header & 0xf0) != kVertexHeader)
+ return -1;
+
+ int version = data_header & 0x0f;
+ if (version > 0)
+ return -1;
+
+ unsigned char last_vertex[256];
+ memcpy(last_vertex, data_end - vertex_size, vertex_size);
+
+ size_t vertex_block_size = getVertexBlockSize(vertex_size);
+
+ size_t vertex_offset = 0;
+
+ while (vertex_offset < vertex_count)
+ {
+ size_t block_size = (vertex_offset + vertex_block_size < vertex_count) ? vertex_block_size : vertex_count - vertex_offset;
+
+ data = decode(data, data_end, vertex_data + vertex_offset * vertex_size, block_size, vertex_size, last_vertex);
+ if (!data)
+ return -2;
+
+ vertex_offset += block_size;
+ }
+
+ size_t tail_size = vertex_size < kTailMaxSize ? kTailMaxSize : vertex_size;
+
+ if (size_t(data_end - data) != tail_size)
+ return -3;
+
+ return 0;
+}
+
+#undef SIMD_NEON
+#undef SIMD_SSE
+#undef SIMD_AVX
+#undef SIMD_WASM
+#undef SIMD_FALLBACK
+#undef SIMD_TARGET
diff --git a/thirdparty/meshoptimizer/vertexfilter.cpp b/thirdparty/meshoptimizer/vertexfilter.cpp
new file mode 100644
index 0000000000..e7ad2c9d39
--- /dev/null
+++ b/thirdparty/meshoptimizer/vertexfilter.cpp
@@ -0,0 +1,825 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <math.h>
+
+// The block below auto-detects SIMD ISA that can be used on the target platform
+#ifndef MESHOPTIMIZER_NO_SIMD
+
+// The SIMD implementation requires SSE2, which can be enabled unconditionally through compiler settings
+#if defined(__SSE2__)
+#define SIMD_SSE
+#endif
+
+// MSVC supports compiling SSE2 code regardless of compile options; we assume all 32-bit CPUs support SSE2
+#if !defined(SIMD_SSE) && defined(_MSC_VER) && !defined(__clang__) && (defined(_M_IX86) || defined(_M_X64))
+#define SIMD_SSE
+#endif
+
+// GCC/clang define these when NEON support is available
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#define SIMD_NEON
+#endif
+
+// On MSVC, we assume that ARM builds always target NEON-capable devices
+#if !defined(SIMD_NEON) && defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+#define SIMD_NEON
+#endif
+
+// When targeting Wasm SIMD we can't use runtime cpuid checks so we unconditionally enable SIMD
+#if defined(__wasm_simd128__)
+#define SIMD_WASM
+#endif
+
+#endif // !MESHOPTIMIZER_NO_SIMD
+
+#ifdef SIMD_SSE
+#include <emmintrin.h>
+#include <stdint.h>
+#endif
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#ifdef SIMD_NEON
+#if defined(_MSC_VER) && defined(_M_ARM64)
+#include <arm64_neon.h>
+#else
+#include <arm_neon.h>
+#endif
+#endif
+
+#ifdef SIMD_WASM
+#include <wasm_simd128.h>
+#endif
+
+#ifdef SIMD_WASM
+#define wasmx_unpacklo_v16x8(a, b) wasm_v16x8_shuffle(a, b, 0, 8, 1, 9, 2, 10, 3, 11)
+#define wasmx_unpackhi_v16x8(a, b) wasm_v16x8_shuffle(a, b, 4, 12, 5, 13, 6, 14, 7, 15)
+#define wasmx_unziplo_v32x4(a, b) wasm_v32x4_shuffle(a, b, 0, 2, 4, 6)
+#define wasmx_unziphi_v32x4(a, b) wasm_v32x4_shuffle(a, b, 1, 3, 5, 7)
+#endif
+
+namespace meshopt
+{
+
+#if !defined(SIMD_SSE) && !defined(SIMD_NEON) && !defined(SIMD_WASM)
+template <typename T>
+static void decodeFilterOct(T* data, size_t count)
+{
+ const float max = float((1 << (sizeof(T) * 8 - 1)) - 1);
+
+ for (size_t i = 0; i < count; ++i)
+ {
+ // convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
+ float x = float(data[i * 4 + 0]);
+ float y = float(data[i * 4 + 1]);
+ float z = float(data[i * 4 + 2]) - fabsf(x) - fabsf(y);
+
+ // fixup octahedral coordinates for z<0
+ float t = (z >= 0.f) ? 0.f : z;
+
+ x += (x >= 0.f) ? t : -t;
+ y += (y >= 0.f) ? t : -t;
+
+ // compute normal length & scale
+ float l = sqrtf(x * x + y * y + z * z);
+ float s = max / l;
+
+ // rounded signed float->int
+ int xf = int(x * s + (x >= 0.f ? 0.5f : -0.5f));
+ int yf = int(y * s + (y >= 0.f ? 0.5f : -0.5f));
+ int zf = int(z * s + (z >= 0.f ? 0.5f : -0.5f));
+
+ data[i * 4 + 0] = T(xf);
+ data[i * 4 + 1] = T(yf);
+ data[i * 4 + 2] = T(zf);
+ }
+}
+
+static void decodeFilterQuat(short* data, size_t count)
+{
+ const float scale = 1.f / sqrtf(2.f);
+
+ for (size_t i = 0; i < count; ++i)
+ {
+ // recover scale from the high byte of the component
+ int sf = data[i * 4 + 3] | 3;
+ float ss = scale / float(sf);
+
+ // convert x/y/z to [-1..1] (scaled...)
+ float x = float(data[i * 4 + 0]) * ss;
+ float y = float(data[i * 4 + 1]) * ss;
+ float z = float(data[i * 4 + 2]) * ss;
+
+ // reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
+ float ww = 1.f - x * x - y * y - z * z;
+ float w = sqrtf(ww >= 0.f ? ww : 0.f);
+
+ // rounded signed float->int
+ int xf = int(x * 32767.f + (x >= 0.f ? 0.5f : -0.5f));
+ int yf = int(y * 32767.f + (y >= 0.f ? 0.5f : -0.5f));
+ int zf = int(z * 32767.f + (z >= 0.f ? 0.5f : -0.5f));
+ int wf = int(w * 32767.f + 0.5f);
+
+ int qc = data[i * 4 + 3] & 3;
+
+ // output order is dictated by input index
+ data[i * 4 + ((qc + 1) & 3)] = short(xf);
+ data[i * 4 + ((qc + 2) & 3)] = short(yf);
+ data[i * 4 + ((qc + 3) & 3)] = short(zf);
+ data[i * 4 + ((qc + 0) & 3)] = short(wf);
+ }
+}
+
+static void decodeFilterExp(unsigned int* data, size_t count)
+{
+ for (size_t i = 0; i < count; ++i)
+ {
+ unsigned int v = data[i];
+
+ // decode mantissa and exponent
+ int m = int(v << 8) >> 8;
+ int e = int(v) >> 24;
+
+ union
+ {
+ float f;
+ unsigned int ui;
+ } u;
+
+ // optimized version of ldexp(float(m), e)
+ u.ui = unsigned(e + 127) << 23;
+ u.f = u.f * float(m);
+
+ data[i] = u.ui;
+ }
+}
+#endif
+
+#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
+inline uint64_t rotateleft64(uint64_t v, int x)
+{
+#if defined(_MSC_VER) && !defined(__clang__)
+ return _rotl64(v, x);
+// Apple's Clang 8 is actually vanilla Clang 3.9, there we need to look for
+// version 11 instead: https://en.wikipedia.org/wiki/Xcode#Toolchain_versions
+#elif defined(__clang__) && ((!defined(__apple_build_version__) && __clang_major__ >= 8) || __clang_major__ >= 11)
+ return __builtin_rotateleft64(v, x);
+#else
+ return (v << (x & 63)) | (v >> ((64 - x) & 63));
+#endif
+}
+#endif
+
+#ifdef SIMD_SSE
+static void decodeFilterOctSimd(signed char* data, size_t count)
+{
+ const __m128 sign = _mm_set1_ps(-0.f);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ __m128i n4 = _mm_loadu_si128(reinterpret_cast<__m128i*>(&data[i * 4]));
+
+ // sign-extends each of x,y in [x y ? ?] with arithmetic shifts
+ __m128i xf = _mm_srai_epi32(_mm_slli_epi32(n4, 24), 24);
+ __m128i yf = _mm_srai_epi32(_mm_slli_epi32(n4, 16), 24);
+
+ // unpack z; note that z is unsigned so we technically don't need to sign extend it
+ __m128i zf = _mm_srai_epi32(_mm_slli_epi32(n4, 8), 24);
+
+ // convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
+ __m128 x = _mm_cvtepi32_ps(xf);
+ __m128 y = _mm_cvtepi32_ps(yf);
+ __m128 z = _mm_sub_ps(_mm_cvtepi32_ps(zf), _mm_add_ps(_mm_andnot_ps(sign, x), _mm_andnot_ps(sign, y)));
+
+ // fixup octahedral coordinates for z<0
+ __m128 t = _mm_min_ps(z, _mm_setzero_ps());
+
+ x = _mm_add_ps(x, _mm_xor_ps(t, _mm_and_ps(x, sign)));
+ y = _mm_add_ps(y, _mm_xor_ps(t, _mm_and_ps(y, sign)));
+
+ // compute normal length & scale
+ __m128 ll = _mm_add_ps(_mm_mul_ps(x, x), _mm_add_ps(_mm_mul_ps(y, y), _mm_mul_ps(z, z)));
+ __m128 s = _mm_mul_ps(_mm_set1_ps(127.f), _mm_rsqrt_ps(ll));
+
+ // rounded signed float->int
+ __m128i xr = _mm_cvtps_epi32(_mm_mul_ps(x, s));
+ __m128i yr = _mm_cvtps_epi32(_mm_mul_ps(y, s));
+ __m128i zr = _mm_cvtps_epi32(_mm_mul_ps(z, s));
+
+ // combine xr/yr/zr into final value
+ __m128i res = _mm_and_si128(n4, _mm_set1_epi32(0xff000000));
+ res = _mm_or_si128(res, _mm_and_si128(xr, _mm_set1_epi32(0xff)));
+ res = _mm_or_si128(res, _mm_slli_epi32(_mm_and_si128(yr, _mm_set1_epi32(0xff)), 8));
+ res = _mm_or_si128(res, _mm_slli_epi32(_mm_and_si128(zr, _mm_set1_epi32(0xff)), 16));
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(&data[i * 4]), res);
+ }
+}
+
+static void decodeFilterOctSimd(short* data, size_t count)
+{
+ const __m128 sign = _mm_set1_ps(-0.f);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ __m128 n4_0 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 0) * 4]));
+ __m128 n4_1 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 2) * 4]));
+
+ // gather both x/y 16-bit pairs in each 32-bit lane
+ __m128i n4 = _mm_castps_si128(_mm_shuffle_ps(n4_0, n4_1, _MM_SHUFFLE(2, 0, 2, 0)));
+
+ // sign-extends each of x,y in [x y] with arithmetic shifts
+ __m128i xf = _mm_srai_epi32(_mm_slli_epi32(n4, 16), 16);
+ __m128i yf = _mm_srai_epi32(n4, 16);
+
+ // unpack z; note that z is unsigned so we don't need to sign extend it
+ __m128i z4 = _mm_castps_si128(_mm_shuffle_ps(n4_0, n4_1, _MM_SHUFFLE(3, 1, 3, 1)));
+ __m128i zf = _mm_and_si128(z4, _mm_set1_epi32(0x7fff));
+
+ // convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
+ __m128 x = _mm_cvtepi32_ps(xf);
+ __m128 y = _mm_cvtepi32_ps(yf);
+ __m128 z = _mm_sub_ps(_mm_cvtepi32_ps(zf), _mm_add_ps(_mm_andnot_ps(sign, x), _mm_andnot_ps(sign, y)));
+
+ // fixup octahedral coordinates for z<0
+ __m128 t = _mm_min_ps(z, _mm_setzero_ps());
+
+ x = _mm_add_ps(x, _mm_xor_ps(t, _mm_and_ps(x, sign)));
+ y = _mm_add_ps(y, _mm_xor_ps(t, _mm_and_ps(y, sign)));
+
+ // compute normal length & scale
+ __m128 ll = _mm_add_ps(_mm_mul_ps(x, x), _mm_add_ps(_mm_mul_ps(y, y), _mm_mul_ps(z, z)));
+ __m128 s = _mm_div_ps(_mm_set1_ps(32767.f), _mm_sqrt_ps(ll));
+
+ // rounded signed float->int
+ __m128i xr = _mm_cvtps_epi32(_mm_mul_ps(x, s));
+ __m128i yr = _mm_cvtps_epi32(_mm_mul_ps(y, s));
+ __m128i zr = _mm_cvtps_epi32(_mm_mul_ps(z, s));
+
+ // mix x/z and y/0 to make 16-bit unpack easier
+ __m128i xzr = _mm_or_si128(_mm_and_si128(xr, _mm_set1_epi32(0xffff)), _mm_slli_epi32(zr, 16));
+ __m128i y0r = _mm_and_si128(yr, _mm_set1_epi32(0xffff));
+
+ // pack x/y/z using 16-bit unpacks; note that this has 0 where we should have .w
+ __m128i res_0 = _mm_unpacklo_epi16(xzr, y0r);
+ __m128i res_1 = _mm_unpackhi_epi16(xzr, y0r);
+
+ // patch in .w
+ res_0 = _mm_or_si128(res_0, _mm_and_si128(_mm_castps_si128(n4_0), _mm_set1_epi64x(0xffff000000000000)));
+ res_1 = _mm_or_si128(res_1, _mm_and_si128(_mm_castps_si128(n4_1), _mm_set1_epi64x(0xffff000000000000)));
+
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(&data[(i + 0) * 4]), res_0);
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(&data[(i + 2) * 4]), res_1);
+ }
+}
+
+static void decodeFilterQuatSimd(short* data, size_t count)
+{
+ const float scale = 1.f / sqrtf(2.f);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ __m128 q4_0 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 0) * 4]));
+ __m128 q4_1 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 2) * 4]));
+
+ // gather both x/y 16-bit pairs in each 32-bit lane
+ __m128i q4_xy = _mm_castps_si128(_mm_shuffle_ps(q4_0, q4_1, _MM_SHUFFLE(2, 0, 2, 0)));
+ __m128i q4_zc = _mm_castps_si128(_mm_shuffle_ps(q4_0, q4_1, _MM_SHUFFLE(3, 1, 3, 1)));
+
+ // sign-extends each of x,y in [x y] with arithmetic shifts
+ __m128i xf = _mm_srai_epi32(_mm_slli_epi32(q4_xy, 16), 16);
+ __m128i yf = _mm_srai_epi32(q4_xy, 16);
+ __m128i zf = _mm_srai_epi32(_mm_slli_epi32(q4_zc, 16), 16);
+ __m128i cf = _mm_srai_epi32(q4_zc, 16);
+
+ // get a floating-point scaler using zc with bottom 2 bits set to 1 (which represents 1.f)
+ __m128i sf = _mm_or_si128(cf, _mm_set1_epi32(3));
+ __m128 ss = _mm_div_ps(_mm_set1_ps(scale), _mm_cvtepi32_ps(sf));
+
+ // convert x/y/z to [-1..1] (scaled...)
+ __m128 x = _mm_mul_ps(_mm_cvtepi32_ps(xf), ss);
+ __m128 y = _mm_mul_ps(_mm_cvtepi32_ps(yf), ss);
+ __m128 z = _mm_mul_ps(_mm_cvtepi32_ps(zf), ss);
+
+ // reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
+ __m128 ww = _mm_sub_ps(_mm_set1_ps(1.f), _mm_add_ps(_mm_mul_ps(x, x), _mm_add_ps(_mm_mul_ps(y, y), _mm_mul_ps(z, z))));
+ __m128 w = _mm_sqrt_ps(_mm_max_ps(ww, _mm_setzero_ps()));
+
+ __m128 s = _mm_set1_ps(32767.f);
+
+ // rounded signed float->int
+ __m128i xr = _mm_cvtps_epi32(_mm_mul_ps(x, s));
+ __m128i yr = _mm_cvtps_epi32(_mm_mul_ps(y, s));
+ __m128i zr = _mm_cvtps_epi32(_mm_mul_ps(z, s));
+ __m128i wr = _mm_cvtps_epi32(_mm_mul_ps(w, s));
+
+ // mix x/z and w/y to make 16-bit unpack easier
+ __m128i xzr = _mm_or_si128(_mm_and_si128(xr, _mm_set1_epi32(0xffff)), _mm_slli_epi32(zr, 16));
+ __m128i wyr = _mm_or_si128(_mm_and_si128(wr, _mm_set1_epi32(0xffff)), _mm_slli_epi32(yr, 16));
+
+ // pack x/y/z/w using 16-bit unpacks; we pack wxyz by default (for qc=0)
+ __m128i res_0 = _mm_unpacklo_epi16(wyr, xzr);
+ __m128i res_1 = _mm_unpackhi_epi16(wyr, xzr);
+
+ // store results to stack so that we can rotate using scalar instructions
+ uint64_t res[4];
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(&res[0]), res_0);
+ _mm_storeu_si128(reinterpret_cast<__m128i*>(&res[2]), res_1);
+
+ // rotate and store
+ uint64_t* out = reinterpret_cast<uint64_t*>(&data[i * 4]);
+
+ out[0] = rotateleft64(res[0], data[(i + 0) * 4 + 3] << 4);
+ out[1] = rotateleft64(res[1], data[(i + 1) * 4 + 3] << 4);
+ out[2] = rotateleft64(res[2], data[(i + 2) * 4 + 3] << 4);
+ out[3] = rotateleft64(res[3], data[(i + 3) * 4 + 3] << 4);
+ }
+}
+
+static void decodeFilterExpSimd(unsigned int* data, size_t count)
+{
+ for (size_t i = 0; i < count; i += 4)
+ {
+ __m128i v = _mm_loadu_si128(reinterpret_cast<__m128i*>(&data[i]));
+
+ // decode exponent into 2^x directly
+ __m128i ef = _mm_srai_epi32(v, 24);
+ __m128i es = _mm_slli_epi32(_mm_add_epi32(ef, _mm_set1_epi32(127)), 23);
+
+ // decode 24-bit mantissa into floating-point value
+ __m128i mf = _mm_srai_epi32(_mm_slli_epi32(v, 8), 8);
+ __m128 m = _mm_cvtepi32_ps(mf);
+
+ __m128 r = _mm_mul_ps(_mm_castsi128_ps(es), m);
+
+ _mm_storeu_ps(reinterpret_cast<float*>(&data[i]), r);
+ }
+}
+#endif
+
+#if defined(SIMD_NEON) && !defined(__aarch64__) && !defined(_M_ARM64)
+inline float32x4_t vsqrtq_f32(float32x4_t x)
+{
+ float32x4_t r = vrsqrteq_f32(x);
+ r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(r, x), r)); // refine rsqrt estimate
+ return vmulq_f32(r, x);
+}
+
+inline float32x4_t vdivq_f32(float32x4_t x, float32x4_t y)
+{
+ float32x4_t r = vrecpeq_f32(y);
+ r = vmulq_f32(r, vrecpsq_f32(y, r)); // refine rcp estimate
+ return vmulq_f32(x, r);
+}
+#endif
+
+#ifdef SIMD_NEON
+static void decodeFilterOctSimd(signed char* data, size_t count)
+{
+ const int32x4_t sign = vdupq_n_s32(0x80000000);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ int32x4_t n4 = vld1q_s32(reinterpret_cast<int32_t*>(&data[i * 4]));
+
+ // sign-extends each of x,y in [x y ? ?] with arithmetic shifts
+ int32x4_t xf = vshrq_n_s32(vshlq_n_s32(n4, 24), 24);
+ int32x4_t yf = vshrq_n_s32(vshlq_n_s32(n4, 16), 24);
+
+ // unpack z; note that z is unsigned so we technically don't need to sign extend it
+ int32x4_t zf = vshrq_n_s32(vshlq_n_s32(n4, 8), 24);
+
+ // convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
+ float32x4_t x = vcvtq_f32_s32(xf);
+ float32x4_t y = vcvtq_f32_s32(yf);
+ float32x4_t z = vsubq_f32(vcvtq_f32_s32(zf), vaddq_f32(vabsq_f32(x), vabsq_f32(y)));
+
+ // fixup octahedral coordinates for z<0
+ float32x4_t t = vminq_f32(z, vdupq_n_f32(0.f));
+
+ x = vaddq_f32(x, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(x), sign))));
+ y = vaddq_f32(y, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(y), sign))));
+
+ // compute normal length & scale
+ float32x4_t ll = vaddq_f32(vmulq_f32(x, x), vaddq_f32(vmulq_f32(y, y), vmulq_f32(z, z)));
+ float32x4_t rl = vrsqrteq_f32(ll);
+ float32x4_t s = vmulq_f32(vdupq_n_f32(127.f), rl);
+
+ // fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
+ // note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
+ const float32x4_t fsnap = vdupq_n_f32(3 << 22);
+
+ int32x4_t xr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(x, s), fsnap));
+ int32x4_t yr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(y, s), fsnap));
+ int32x4_t zr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(z, s), fsnap));
+
+ // combine xr/yr/zr into final value
+ int32x4_t res = vandq_s32(n4, vdupq_n_s32(0xff000000));
+ res = vorrq_s32(res, vandq_s32(xr, vdupq_n_s32(0xff)));
+ res = vorrq_s32(res, vshlq_n_s32(vandq_s32(yr, vdupq_n_s32(0xff)), 8));
+ res = vorrq_s32(res, vshlq_n_s32(vandq_s32(zr, vdupq_n_s32(0xff)), 16));
+
+ vst1q_s32(reinterpret_cast<int32_t*>(&data[i * 4]), res);
+ }
+}
+
+static void decodeFilterOctSimd(short* data, size_t count)
+{
+ const int32x4_t sign = vdupq_n_s32(0x80000000);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ int32x4_t n4_0 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 0) * 4]));
+ int32x4_t n4_1 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 2) * 4]));
+
+ // gather both x/y 16-bit pairs in each 32-bit lane
+ int32x4_t n4 = vuzpq_s32(n4_0, n4_1).val[0];
+
+ // sign-extends each of x,y in [x y] with arithmetic shifts
+ int32x4_t xf = vshrq_n_s32(vshlq_n_s32(n4, 16), 16);
+ int32x4_t yf = vshrq_n_s32(n4, 16);
+
+ // unpack z; note that z is unsigned so we don't need to sign extend it
+ int32x4_t z4 = vuzpq_s32(n4_0, n4_1).val[1];
+ int32x4_t zf = vandq_s32(z4, vdupq_n_s32(0x7fff));
+
+ // convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
+ float32x4_t x = vcvtq_f32_s32(xf);
+ float32x4_t y = vcvtq_f32_s32(yf);
+ float32x4_t z = vsubq_f32(vcvtq_f32_s32(zf), vaddq_f32(vabsq_f32(x), vabsq_f32(y)));
+
+ // fixup octahedral coordinates for z<0
+ float32x4_t t = vminq_f32(z, vdupq_n_f32(0.f));
+
+ x = vaddq_f32(x, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(x), sign))));
+ y = vaddq_f32(y, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(y), sign))));
+
+ // compute normal length & scale
+ float32x4_t ll = vaddq_f32(vmulq_f32(x, x), vaddq_f32(vmulq_f32(y, y), vmulq_f32(z, z)));
+ float32x4_t rl = vrsqrteq_f32(ll);
+ rl = vmulq_f32(rl, vrsqrtsq_f32(vmulq_f32(rl, ll), rl)); // refine rsqrt estimate
+ float32x4_t s = vmulq_f32(vdupq_n_f32(32767.f), rl);
+
+ // fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
+ // note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
+ const float32x4_t fsnap = vdupq_n_f32(3 << 22);
+
+ int32x4_t xr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(x, s), fsnap));
+ int32x4_t yr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(y, s), fsnap));
+ int32x4_t zr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(z, s), fsnap));
+
+ // mix x/z and y/0 to make 16-bit unpack easier
+ int32x4_t xzr = vorrq_s32(vandq_s32(xr, vdupq_n_s32(0xffff)), vshlq_n_s32(zr, 16));
+ int32x4_t y0r = vandq_s32(yr, vdupq_n_s32(0xffff));
+
+ // pack x/y/z using 16-bit unpacks; note that this has 0 where we should have .w
+ int32x4_t res_0 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(xzr), vreinterpretq_s16_s32(y0r)).val[0]);
+ int32x4_t res_1 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(xzr), vreinterpretq_s16_s32(y0r)).val[1]);
+
+ // patch in .w
+ res_0 = vbslq_s32(vreinterpretq_u32_u64(vdupq_n_u64(0xffff000000000000)), n4_0, res_0);
+ res_1 = vbslq_s32(vreinterpretq_u32_u64(vdupq_n_u64(0xffff000000000000)), n4_1, res_1);
+
+ vst1q_s32(reinterpret_cast<int32_t*>(&data[(i + 0) * 4]), res_0);
+ vst1q_s32(reinterpret_cast<int32_t*>(&data[(i + 2) * 4]), res_1);
+ }
+}
+
+static void decodeFilterQuatSimd(short* data, size_t count)
+{
+ const float scale = 1.f / sqrtf(2.f);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ int32x4_t q4_0 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 0) * 4]));
+ int32x4_t q4_1 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 2) * 4]));
+
+ // gather both x/y 16-bit pairs in each 32-bit lane
+ int32x4_t q4_xy = vuzpq_s32(q4_0, q4_1).val[0];
+ int32x4_t q4_zc = vuzpq_s32(q4_0, q4_1).val[1];
+
+ // sign-extends each of x,y in [x y] with arithmetic shifts
+ int32x4_t xf = vshrq_n_s32(vshlq_n_s32(q4_xy, 16), 16);
+ int32x4_t yf = vshrq_n_s32(q4_xy, 16);
+ int32x4_t zf = vshrq_n_s32(vshlq_n_s32(q4_zc, 16), 16);
+ int32x4_t cf = vshrq_n_s32(q4_zc, 16);
+
+ // get a floating-point scaler using zc with bottom 2 bits set to 1 (which represents 1.f)
+ int32x4_t sf = vorrq_s32(cf, vdupq_n_s32(3));
+ float32x4_t ss = vdivq_f32(vdupq_n_f32(scale), vcvtq_f32_s32(sf));
+
+ // convert x/y/z to [-1..1] (scaled...)
+ float32x4_t x = vmulq_f32(vcvtq_f32_s32(xf), ss);
+ float32x4_t y = vmulq_f32(vcvtq_f32_s32(yf), ss);
+ float32x4_t z = vmulq_f32(vcvtq_f32_s32(zf), ss);
+
+ // reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
+ float32x4_t ww = vsubq_f32(vdupq_n_f32(1.f), vaddq_f32(vmulq_f32(x, x), vaddq_f32(vmulq_f32(y, y), vmulq_f32(z, z))));
+ float32x4_t w = vsqrtq_f32(vmaxq_f32(ww, vdupq_n_f32(0.f)));
+
+ float32x4_t s = vdupq_n_f32(32767.f);
+
+ // fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
+ // note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
+ const float32x4_t fsnap = vdupq_n_f32(3 << 22);
+
+ int32x4_t xr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(x, s), fsnap));
+ int32x4_t yr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(y, s), fsnap));
+ int32x4_t zr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(z, s), fsnap));
+ int32x4_t wr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(w, s), fsnap));
+
+ // mix x/z and w/y to make 16-bit unpack easier
+ int32x4_t xzr = vorrq_s32(vandq_s32(xr, vdupq_n_s32(0xffff)), vshlq_n_s32(zr, 16));
+ int32x4_t wyr = vorrq_s32(vandq_s32(wr, vdupq_n_s32(0xffff)), vshlq_n_s32(yr, 16));
+
+ // pack x/y/z/w using 16-bit unpacks; we pack wxyz by default (for qc=0)
+ int32x4_t res_0 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(wyr), vreinterpretq_s16_s32(xzr)).val[0]);
+ int32x4_t res_1 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(wyr), vreinterpretq_s16_s32(xzr)).val[1]);
+
+ // rotate and store
+ uint64_t* out = (uint64_t*)&data[i * 4];
+
+ out[0] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_0), 0), vgetq_lane_s32(cf, 0) << 4);
+ out[1] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_0), 1), vgetq_lane_s32(cf, 1) << 4);
+ out[2] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_1), 0), vgetq_lane_s32(cf, 2) << 4);
+ out[3] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_1), 1), vgetq_lane_s32(cf, 3) << 4);
+ }
+}
+
+static void decodeFilterExpSimd(unsigned int* data, size_t count)
+{
+ for (size_t i = 0; i < count; i += 4)
+ {
+ int32x4_t v = vld1q_s32(reinterpret_cast<int32_t*>(&data[i]));
+
+ // decode exponent into 2^x directly
+ int32x4_t ef = vshrq_n_s32(v, 24);
+ int32x4_t es = vshlq_n_s32(vaddq_s32(ef, vdupq_n_s32(127)), 23);
+
+ // decode 24-bit mantissa into floating-point value
+ int32x4_t mf = vshrq_n_s32(vshlq_n_s32(v, 8), 8);
+ float32x4_t m = vcvtq_f32_s32(mf);
+
+ float32x4_t r = vmulq_f32(vreinterpretq_f32_s32(es), m);
+
+ vst1q_f32(reinterpret_cast<float*>(&data[i]), r);
+ }
+}
+#endif
+
+#ifdef SIMD_WASM
+static void decodeFilterOctSimd(signed char* data, size_t count)
+{
+ const v128_t sign = wasm_f32x4_splat(-0.f);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ v128_t n4 = wasm_v128_load(&data[i * 4]);
+
+ // sign-extends each of x,y in [x y ? ?] with arithmetic shifts
+ v128_t xf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 24), 24);
+ v128_t yf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 16), 24);
+
+ // unpack z; note that z is unsigned so we technically don't need to sign extend it
+ v128_t zf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 8), 24);
+
+ // convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
+ v128_t x = wasm_f32x4_convert_i32x4(xf);
+ v128_t y = wasm_f32x4_convert_i32x4(yf);
+ v128_t z = wasm_f32x4_sub(wasm_f32x4_convert_i32x4(zf), wasm_f32x4_add(wasm_f32x4_abs(x), wasm_f32x4_abs(y)));
+
+ // fixup octahedral coordinates for z<0
+ // note: i32x4_min with 0 is equvalent to f32x4_min
+ v128_t t = wasm_i32x4_min(z, wasm_i32x4_splat(0));
+
+ x = wasm_f32x4_add(x, wasm_v128_xor(t, wasm_v128_and(x, sign)));
+ y = wasm_f32x4_add(y, wasm_v128_xor(t, wasm_v128_and(y, sign)));
+
+ // compute normal length & scale
+ v128_t ll = wasm_f32x4_add(wasm_f32x4_mul(x, x), wasm_f32x4_add(wasm_f32x4_mul(y, y), wasm_f32x4_mul(z, z)));
+ v128_t s = wasm_f32x4_div(wasm_f32x4_splat(127.f), wasm_f32x4_sqrt(ll));
+
+ // fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
+ // note: the result is offset by 0x4B40_0000, but we only need the low 8 bits so we can omit the subtraction
+ const v128_t fsnap = wasm_f32x4_splat(3 << 22);
+
+ v128_t xr = wasm_f32x4_add(wasm_f32x4_mul(x, s), fsnap);
+ v128_t yr = wasm_f32x4_add(wasm_f32x4_mul(y, s), fsnap);
+ v128_t zr = wasm_f32x4_add(wasm_f32x4_mul(z, s), fsnap);
+
+ // combine xr/yr/zr into final value
+ v128_t res = wasm_v128_and(n4, wasm_i32x4_splat(0xff000000));
+ res = wasm_v128_or(res, wasm_v128_and(xr, wasm_i32x4_splat(0xff)));
+ res = wasm_v128_or(res, wasm_i32x4_shl(wasm_v128_and(yr, wasm_i32x4_splat(0xff)), 8));
+ res = wasm_v128_or(res, wasm_i32x4_shl(wasm_v128_and(zr, wasm_i32x4_splat(0xff)), 16));
+
+ wasm_v128_store(&data[i * 4], res);
+ }
+}
+
+static void decodeFilterOctSimd(short* data, size_t count)
+{
+ const v128_t sign = wasm_f32x4_splat(-0.f);
+ const v128_t zmask = wasm_i32x4_splat(0x7fff);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ v128_t n4_0 = wasm_v128_load(&data[(i + 0) * 4]);
+ v128_t n4_1 = wasm_v128_load(&data[(i + 2) * 4]);
+
+ // gather both x/y 16-bit pairs in each 32-bit lane
+ v128_t n4 = wasmx_unziplo_v32x4(n4_0, n4_1);
+
+ // sign-extends each of x,y in [x y] with arithmetic shifts
+ v128_t xf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 16), 16);
+ v128_t yf = wasm_i32x4_shr(n4, 16);
+
+ // unpack z; note that z is unsigned so we don't need to sign extend it
+ v128_t z4 = wasmx_unziphi_v32x4(n4_0, n4_1);
+ v128_t zf = wasm_v128_and(z4, zmask);
+
+ // convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
+ v128_t x = wasm_f32x4_convert_i32x4(xf);
+ v128_t y = wasm_f32x4_convert_i32x4(yf);
+ v128_t z = wasm_f32x4_sub(wasm_f32x4_convert_i32x4(zf), wasm_f32x4_add(wasm_f32x4_abs(x), wasm_f32x4_abs(y)));
+
+ // fixup octahedral coordinates for z<0
+ // note: i32x4_min with 0 is equvalent to f32x4_min
+ v128_t t = wasm_i32x4_min(z, wasm_i32x4_splat(0));
+
+ x = wasm_f32x4_add(x, wasm_v128_xor(t, wasm_v128_and(x, sign)));
+ y = wasm_f32x4_add(y, wasm_v128_xor(t, wasm_v128_and(y, sign)));
+
+ // compute normal length & scale
+ v128_t ll = wasm_f32x4_add(wasm_f32x4_mul(x, x), wasm_f32x4_add(wasm_f32x4_mul(y, y), wasm_f32x4_mul(z, z)));
+ v128_t s = wasm_f32x4_div(wasm_f32x4_splat(32767.f), wasm_f32x4_sqrt(ll));
+
+ // fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
+ // note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
+ const v128_t fsnap = wasm_f32x4_splat(3 << 22);
+
+ v128_t xr = wasm_f32x4_add(wasm_f32x4_mul(x, s), fsnap);
+ v128_t yr = wasm_f32x4_add(wasm_f32x4_mul(y, s), fsnap);
+ v128_t zr = wasm_f32x4_add(wasm_f32x4_mul(z, s), fsnap);
+
+ // mix x/z and y/0 to make 16-bit unpack easier
+ v128_t xzr = wasm_v128_or(wasm_v128_and(xr, wasm_i32x4_splat(0xffff)), wasm_i32x4_shl(zr, 16));
+ v128_t y0r = wasm_v128_and(yr, wasm_i32x4_splat(0xffff));
+
+ // pack x/y/z using 16-bit unpacks; note that this has 0 where we should have .w
+ v128_t res_0 = wasmx_unpacklo_v16x8(xzr, y0r);
+ v128_t res_1 = wasmx_unpackhi_v16x8(xzr, y0r);
+
+ // patch in .w
+ res_0 = wasm_v128_or(res_0, wasm_v128_and(n4_0, wasm_i64x2_splat(0xffff000000000000)));
+ res_1 = wasm_v128_or(res_1, wasm_v128_and(n4_1, wasm_i64x2_splat(0xffff000000000000)));
+
+ wasm_v128_store(&data[(i + 0) * 4], res_0);
+ wasm_v128_store(&data[(i + 2) * 4], res_1);
+ }
+}
+
+static void decodeFilterQuatSimd(short* data, size_t count)
+{
+ const float scale = 1.f / sqrtf(2.f);
+
+ for (size_t i = 0; i < count; i += 4)
+ {
+ v128_t q4_0 = wasm_v128_load(&data[(i + 0) * 4]);
+ v128_t q4_1 = wasm_v128_load(&data[(i + 2) * 4]);
+
+ // gather both x/y 16-bit pairs in each 32-bit lane
+ v128_t q4_xy = wasmx_unziplo_v32x4(q4_0, q4_1);
+ v128_t q4_zc = wasmx_unziphi_v32x4(q4_0, q4_1);
+
+ // sign-extends each of x,y in [x y] with arithmetic shifts
+ v128_t xf = wasm_i32x4_shr(wasm_i32x4_shl(q4_xy, 16), 16);
+ v128_t yf = wasm_i32x4_shr(q4_xy, 16);
+ v128_t zf = wasm_i32x4_shr(wasm_i32x4_shl(q4_zc, 16), 16);
+ v128_t cf = wasm_i32x4_shr(q4_zc, 16);
+
+ // get a floating-point scaler using zc with bottom 2 bits set to 1 (which represents 1.f)
+ v128_t sf = wasm_v128_or(cf, wasm_i32x4_splat(3));
+ v128_t ss = wasm_f32x4_div(wasm_f32x4_splat(scale), wasm_f32x4_convert_i32x4(sf));
+
+ // convert x/y/z to [-1..1] (scaled...)
+ v128_t x = wasm_f32x4_mul(wasm_f32x4_convert_i32x4(xf), ss);
+ v128_t y = wasm_f32x4_mul(wasm_f32x4_convert_i32x4(yf), ss);
+ v128_t z = wasm_f32x4_mul(wasm_f32x4_convert_i32x4(zf), ss);
+
+ // reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
+ // note: i32x4_max with 0 is equivalent to f32x4_max
+ v128_t ww = wasm_f32x4_sub(wasm_f32x4_splat(1.f), wasm_f32x4_add(wasm_f32x4_mul(x, x), wasm_f32x4_add(wasm_f32x4_mul(y, y), wasm_f32x4_mul(z, z))));
+ v128_t w = wasm_f32x4_sqrt(wasm_i32x4_max(ww, wasm_i32x4_splat(0)));
+
+ v128_t s = wasm_f32x4_splat(32767.f);
+
+ // fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
+ // note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
+ const v128_t fsnap = wasm_f32x4_splat(3 << 22);
+
+ v128_t xr = wasm_f32x4_add(wasm_f32x4_mul(x, s), fsnap);
+ v128_t yr = wasm_f32x4_add(wasm_f32x4_mul(y, s), fsnap);
+ v128_t zr = wasm_f32x4_add(wasm_f32x4_mul(z, s), fsnap);
+ v128_t wr = wasm_f32x4_add(wasm_f32x4_mul(w, s), fsnap);
+
+ // mix x/z and w/y to make 16-bit unpack easier
+ v128_t xzr = wasm_v128_or(wasm_v128_and(xr, wasm_i32x4_splat(0xffff)), wasm_i32x4_shl(zr, 16));
+ v128_t wyr = wasm_v128_or(wasm_v128_and(wr, wasm_i32x4_splat(0xffff)), wasm_i32x4_shl(yr, 16));
+
+ // pack x/y/z/w using 16-bit unpacks; we pack wxyz by default (for qc=0)
+ v128_t res_0 = wasmx_unpacklo_v16x8(wyr, xzr);
+ v128_t res_1 = wasmx_unpackhi_v16x8(wyr, xzr);
+
+ // compute component index shifted left by 4 (and moved into i32x4 slot)
+ // TODO: volatile here works around LLVM mis-optimizing code; https://github.com/emscripten-core/emscripten/issues/11449
+ volatile v128_t cm = wasm_i32x4_shl(cf, 4);
+
+ // rotate and store
+ uint64_t* out = reinterpret_cast<uint64_t*>(&data[i * 4]);
+
+ out[0] = rotateleft64(wasm_i64x2_extract_lane(res_0, 0), wasm_i32x4_extract_lane(cm, 0));
+ out[1] = rotateleft64(wasm_i64x2_extract_lane(res_0, 1), wasm_i32x4_extract_lane(cm, 1));
+ out[2] = rotateleft64(wasm_i64x2_extract_lane(res_1, 0), wasm_i32x4_extract_lane(cm, 2));
+ out[3] = rotateleft64(wasm_i64x2_extract_lane(res_1, 1), wasm_i32x4_extract_lane(cm, 3));
+ }
+}
+
+static void decodeFilterExpSimd(unsigned int* data, size_t count)
+{
+ for (size_t i = 0; i < count; i += 4)
+ {
+ v128_t v = wasm_v128_load(&data[i]);
+
+ // decode exponent into 2^x directly
+ v128_t ef = wasm_i32x4_shr(v, 24);
+ v128_t es = wasm_i32x4_shl(wasm_i32x4_add(ef, wasm_i32x4_splat(127)), 23);
+
+ // decode 24-bit mantissa into floating-point value
+ v128_t mf = wasm_i32x4_shr(wasm_i32x4_shl(v, 8), 8);
+ v128_t m = wasm_f32x4_convert_i32x4(mf);
+
+ v128_t r = wasm_f32x4_mul(es, m);
+
+ wasm_v128_store(&data[i], r);
+ }
+}
+#endif
+
+} // namespace meshopt
+
+void meshopt_decodeFilterOct(void* buffer, size_t vertex_count, size_t vertex_size)
+{
+ using namespace meshopt;
+
+ assert(vertex_count % 4 == 0);
+ assert(vertex_size == 4 || vertex_size == 8);
+
+#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
+ if (vertex_size == 4)
+ decodeFilterOctSimd(static_cast<signed char*>(buffer), vertex_count);
+ else
+ decodeFilterOctSimd(static_cast<short*>(buffer), vertex_count);
+#else
+ if (vertex_size == 4)
+ decodeFilterOct(static_cast<signed char*>(buffer), vertex_count);
+ else
+ decodeFilterOct(static_cast<short*>(buffer), vertex_count);
+#endif
+}
+
+void meshopt_decodeFilterQuat(void* buffer, size_t vertex_count, size_t vertex_size)
+{
+ using namespace meshopt;
+
+ assert(vertex_count % 4 == 0);
+ assert(vertex_size == 8);
+ (void)vertex_size;
+
+#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
+ decodeFilterQuatSimd(static_cast<short*>(buffer), vertex_count);
+#else
+ decodeFilterQuat(static_cast<short*>(buffer), vertex_count);
+#endif
+}
+
+void meshopt_decodeFilterExp(void* buffer, size_t vertex_count, size_t vertex_size)
+{
+ using namespace meshopt;
+
+ assert(vertex_count % 4 == 0);
+ assert(vertex_size % 4 == 0);
+
+#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
+ decodeFilterExpSimd(static_cast<unsigned int*>(buffer), vertex_count * (vertex_size / 4));
+#else
+ decodeFilterExp(static_cast<unsigned int*>(buffer), vertex_count * (vertex_size / 4));
+#endif
+}
+
+#undef SIMD_SSE
+#undef SIMD_NEON
+#undef SIMD_WASM
diff --git a/thirdparty/meshoptimizer/vfetchanalyzer.cpp b/thirdparty/meshoptimizer/vfetchanalyzer.cpp
new file mode 100644
index 0000000000..51dca873f8
--- /dev/null
+++ b/thirdparty/meshoptimizer/vfetchanalyzer.cpp
@@ -0,0 +1,58 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <string.h>
+
+meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const unsigned int* indices, size_t index_count, size_t vertex_count, size_t vertex_size)
+{
+ assert(index_count % 3 == 0);
+ assert(vertex_size > 0 && vertex_size <= 256);
+
+ meshopt_Allocator allocator;
+
+ meshopt_VertexFetchStatistics result = {};
+
+ unsigned char* vertex_visited = allocator.allocate<unsigned char>(vertex_count);
+ memset(vertex_visited, 0, vertex_count);
+
+ const size_t kCacheLine = 64;
+ const size_t kCacheSize = 128 * 1024;
+
+ // simple direct mapped cache; on typical mesh data this is close to 4-way cache, and this model is a gross approximation anyway
+ size_t cache[kCacheSize / kCacheLine] = {};
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices[i];
+ assert(index < vertex_count);
+
+ vertex_visited[index] = 1;
+
+ size_t start_address = index * vertex_size;
+ size_t end_address = start_address + vertex_size;
+
+ size_t start_tag = start_address / kCacheLine;
+ size_t end_tag = (end_address + kCacheLine - 1) / kCacheLine;
+
+ assert(start_tag < end_tag);
+
+ for (size_t tag = start_tag; tag < end_tag; ++tag)
+ {
+ size_t line = tag % (sizeof(cache) / sizeof(cache[0]));
+
+ // we store +1 since cache is filled with 0 by default
+ result.bytes_fetched += (cache[line] != tag + 1) * kCacheLine;
+ cache[line] = tag + 1;
+ }
+ }
+
+ size_t unique_vertex_count = 0;
+
+ for (size_t i = 0; i < vertex_count; ++i)
+ unique_vertex_count += vertex_visited[i];
+
+ result.overfetch = unique_vertex_count == 0 ? 0 : float(result.bytes_fetched) / float(unique_vertex_count * vertex_size);
+
+ return result;
+}
diff --git a/thirdparty/meshoptimizer/vfetchoptimizer.cpp b/thirdparty/meshoptimizer/vfetchoptimizer.cpp
new file mode 100644
index 0000000000..465d6df5ca
--- /dev/null
+++ b/thirdparty/meshoptimizer/vfetchoptimizer.cpp
@@ -0,0 +1,74 @@
+// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
+#include "meshoptimizer.h"
+
+#include <assert.h>
+#include <string.h>
+
+size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count)
+{
+ assert(index_count % 3 == 0);
+
+ memset(destination, -1, vertex_count * sizeof(unsigned int));
+
+ unsigned int next_vertex = 0;
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices[i];
+ assert(index < vertex_count);
+
+ if (destination[index] == ~0u)
+ {
+ destination[index] = next_vertex++;
+ }
+ }
+
+ assert(next_vertex <= vertex_count);
+
+ return next_vertex;
+}
+
+size_t meshopt_optimizeVertexFetch(void* destination, unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
+{
+ assert(index_count % 3 == 0);
+ assert(vertex_size > 0 && vertex_size <= 256);
+
+ meshopt_Allocator allocator;
+
+ // support in-place optimization
+ if (destination == vertices)
+ {
+ unsigned char* vertices_copy = allocator.allocate<unsigned char>(vertex_count * vertex_size);
+ memcpy(vertices_copy, vertices, vertex_count * vertex_size);
+ vertices = vertices_copy;
+ }
+
+ // build vertex remap table
+ unsigned int* vertex_remap = allocator.allocate<unsigned int>(vertex_count);
+ memset(vertex_remap, -1, vertex_count * sizeof(unsigned int));
+
+ unsigned int next_vertex = 0;
+
+ for (size_t i = 0; i < index_count; ++i)
+ {
+ unsigned int index = indices[i];
+ assert(index < vertex_count);
+
+ unsigned int& remap = vertex_remap[index];
+
+ if (remap == ~0u) // vertex was not added to destination VB
+ {
+ // add vertex
+ memcpy(static_cast<unsigned char*>(destination) + next_vertex * vertex_size, static_cast<const unsigned char*>(vertices) + index * vertex_size, vertex_size);
+
+ remap = next_vertex++;
+ }
+
+ // modify indices in place
+ indices[i] = remap;
+ }
+
+ assert(next_vertex <= vertex_count);
+
+ return next_vertex;
+}