1 | // Copyright 2009-2021 Intel Corporation |
2 | // SPDX-License-Identifier: Apache-2.0 |
3 | |
4 | #pragma once |
5 | |
6 | #include "../common/ray.h" |
7 | #include "../common/scene_subdiv_mesh.h" |
8 | #include "../bvh/bvh.h" |
9 | #include "../subdiv/tessellation.h" |
10 | #include "../subdiv/tessellation_cache.h" |
11 | #include "subdivpatch1.h" |
12 | |
13 | namespace embree |
14 | { |
15 | namespace isa |
16 | { |
17 | class GridSOA |
18 | { |
19 | public: |
20 | |
21 | /*! GridSOA constructor */ |
22 | GridSOA(const SubdivPatch1Base* patches, const unsigned time_steps, |
23 | const unsigned x0, const unsigned x1, const unsigned y0, const unsigned y1, const unsigned swidth, const unsigned sheight, |
24 | const SubdivMesh* const geom, const size_t totalBvhBytes, const size_t gridBytes, BBox3fa* bounds_o = nullptr); |
25 | |
26 | /*! Subgrid creation */ |
27 | template<typename Allocator> |
28 | static GridSOA* create(const SubdivPatch1Base* patches, const unsigned time_steps, |
29 | unsigned x0, unsigned x1, unsigned y0, unsigned y1, |
30 | const Scene* scene, Allocator& alloc, BBox3fa* bounds_o = nullptr) |
31 | { |
32 | const unsigned width = x1-x0+1; |
33 | const unsigned height = y1-y0+1; |
34 | const GridRange range(0,width-1,0,height-1); |
35 | size_t bvhBytes = 0; |
36 | if (time_steps == 1) |
37 | bvhBytes = getBVHBytes(range,nodeBytes: sizeof(BVH4::AABBNode),leafBytes: 0); |
38 | else { |
39 | bvhBytes = (time_steps-1)*getBVHBytes(range,nodeBytes: sizeof(BVH4::AABBNodeMB),leafBytes: 0); |
40 | bvhBytes += getTemporalBVHBytes(time_range: make_range(begin: 0,end: int(time_steps-1)),nodeBytes: sizeof(BVH4::AABBNodeMB4D)); |
41 | } |
42 | const size_t gridBytes = 4*size_t(width)*size_t(height)*sizeof(float); |
43 | size_t rootBytes = time_steps*sizeof(BVH4::NodeRef); |
44 | #if !defined(__64BIT__) |
45 | rootBytes += 4; // We read 2 elements behind the grid. As we store at least 8 root bytes after the grid we are fine in 64 bit mode. But in 32 bit mode we have to do additional padding. |
46 | #endif |
47 | void* data = alloc(offsetof(GridSOA,data)+bvhBytes+time_steps*gridBytes+rootBytes); |
48 | assert(data); |
49 | return new (data) GridSOA(patches,time_steps,x0,x1,y0,y1,patches->grid_u_res,patches->grid_v_res,scene->get<SubdivMesh>(i: patches->geomID()),bvhBytes,gridBytes,bounds_o); |
50 | } |
51 | |
52 | /*! Grid creation */ |
53 | template<typename Allocator> |
54 | static GridSOA* create(const SubdivPatch1Base* const patches, const unsigned time_steps, |
55 | const Scene* scene, const Allocator& alloc, BBox3fa* bounds_o = nullptr) |
56 | { |
57 | return create(patches,time_steps,0,patches->grid_u_res-1,0,patches->grid_v_res-1,scene,alloc,bounds_o); |
58 | } |
59 | |
60 | /*! returns reference to root */ |
61 | __forceinline BVH4::NodeRef& root(size_t t = 0) { return (BVH4::NodeRef&)data[rootOffset + t*sizeof(BVH4::NodeRef)]; } |
62 | __forceinline const BVH4::NodeRef& root(size_t t = 0) const { return (BVH4::NodeRef&)data[rootOffset + t*sizeof(BVH4::NodeRef)]; } |
63 | |
64 | /*! returns pointer to BVH array */ |
65 | __forceinline char* bvhData() { return &data[0]; } |
66 | __forceinline const char* bvhData() const { return &data[0]; } |
67 | |
68 | /*! returns pointer to Grid array */ |
69 | __forceinline float* gridData(size_t t = 0) { return (float*) &data[gridOffset + t*gridBytes]; } |
70 | __forceinline const float* gridData(size_t t = 0) const { return (float*) &data[gridOffset + t*gridBytes]; } |
71 | |
72 | __forceinline void* encodeLeaf(size_t u, size_t v) { |
73 | return (void*) (16*(v * width + u + 1)); // +1 to not create empty leaf |
74 | } |
75 | __forceinline float* decodeLeaf(size_t t, const void* ptr) { |
76 | return gridData(t) + (((size_t) (ptr) >> 4) - 1); |
77 | } |
78 | |
79 | /*! returns the size of the BVH over the grid in bytes */ |
80 | static size_t getBVHBytes(const GridRange& range, const size_t nodeBytes, const size_t leafBytes); |
81 | |
82 | /*! returns the size of the temporal BVH over the time range BVHs */ |
83 | static size_t getTemporalBVHBytes(const range<int> time_range, const size_t nodeBytes); |
84 | |
85 | /*! calculates bounding box of grid range */ |
86 | __forceinline BBox3fa calculateBounds(size_t time, const GridRange& range) const |
87 | { |
88 | const float* const grid_array = gridData(t: time); |
89 | const float* const grid_x_array = grid_array + 0 * dim_offset; |
90 | const float* const grid_y_array = grid_array + 1 * dim_offset; |
91 | const float* const grid_z_array = grid_array + 2 * dim_offset; |
92 | |
93 | /* compute the bounds just for the range! */ |
94 | BBox3fa bounds( empty ); |
95 | for (unsigned v = range.v_start; v<=range.v_end; v++) |
96 | { |
97 | for (unsigned u = range.u_start; u<=range.u_end; u++) |
98 | { |
99 | const float x = grid_x_array[ v * width + u]; |
100 | const float y = grid_y_array[ v * width + u]; |
101 | const float z = grid_z_array[ v * width + u]; |
102 | bounds.extend( other: Vec3fa(x,y,z) ); |
103 | } |
104 | } |
105 | assert(is_finite(bounds)); |
106 | return bounds; |
107 | } |
108 | |
109 | /*! Evaluates grid over patch and builds BVH4 tree over the grid. */ |
110 | std::pair<BVH4::NodeRef,BBox3fa> buildBVH(BBox3fa* bounds_o); |
111 | |
112 | /*! Create BVH4 tree over grid. */ |
113 | std::pair<BVH4::NodeRef,BBox3fa> buildBVH(const GridRange& range, size_t& allocator); |
114 | |
115 | /*! Evaluates grid over patch and builds MSMBlur BVH4 tree over the grid. */ |
116 | std::pair<BVH4::NodeRef,LBBox3fa> buildMSMBlurBVH(const range<int> time_range, BBox3fa* bounds_o); |
117 | |
118 | /*! Create MBlur BVH4 tree over grid. */ |
119 | std::pair<BVH4::NodeRef,LBBox3fa> buildMBlurBVH(size_t time, const GridRange& range, size_t& allocator); |
120 | |
121 | /*! Create MSMBlur BVH4 tree over grid. */ |
122 | std::pair<BVH4::NodeRef,LBBox3fa> buildMSMBlurBVH(const range<int> time_range, size_t& allocator, BBox3fa* bounds_o); |
123 | |
124 | template<typename Loader> |
125 | struct MapUV |
126 | { |
127 | typedef typename Loader::vfloat vfloat; |
128 | const float* const grid_uv; |
129 | size_t line_offset; |
130 | size_t lines; |
131 | |
132 | __forceinline MapUV(const float* const grid_uv, size_t line_offset, const size_t lines) |
133 | : grid_uv(grid_uv), line_offset(line_offset), lines(lines) {} |
134 | |
135 | __forceinline void operator() (vfloat& u, vfloat& v, Vec3<vfloat>& Ng) const { |
136 | const Vec3<vfloat> tri_v012_uv = Loader::gather(grid_uv,line_offset,lines); |
137 | const Vec2<vfloat> uv0 = GridSOA::decodeUV(tri_v012_uv[0]); |
138 | const Vec2<vfloat> uv1 = GridSOA::decodeUV(tri_v012_uv[1]); |
139 | const Vec2<vfloat> uv2 = GridSOA::decodeUV(tri_v012_uv[2]); |
140 | const Vec2<vfloat> uv = u * uv1 + v * uv2 + (1.0f-u-v) * uv0; |
141 | u = uv[0];v = uv[1]; |
142 | } |
143 | }; |
144 | |
145 | struct Gather2x3 |
146 | { |
147 | enum { M = 4 }; |
148 | typedef vbool4 vbool; |
149 | typedef vint4 vint; |
150 | typedef vfloat4 vfloat; |
151 | |
152 | static __forceinline const Vec3vf4 gather(const float* const grid, const size_t line_offset, const size_t lines) |
153 | { |
154 | vfloat4 r0 = vfloat4::loadu(a: grid + 0*line_offset); |
155 | vfloat4 r1 = vfloat4::loadu(a: grid + 1*line_offset); // this accesses 2 elements too much in case of 2x2 grid, but this is ok as we ensure enough padding after the grid |
156 | if (unlikely(line_offset == 2)) |
157 | { |
158 | r0 = shuffle<0,1,1,1>(v: r0); |
159 | r1 = shuffle<0,1,1,1>(v: r1); |
160 | } |
161 | return Vec3vf4(unpacklo(a: r0,b: r1), // r00, r10, r01, r11 |
162 | shuffle<1,1,2,2>(v: r0), // r01, r01, r02, r02 |
163 | shuffle<0,1,1,2>(v: r1)); // r10, r11, r11, r12 |
164 | } |
165 | |
166 | static __forceinline void gather(const float* const grid_x, |
167 | const float* const grid_y, |
168 | const float* const grid_z, |
169 | const size_t line_offset, |
170 | const size_t lines, |
171 | Vec3vf4& v0_o, |
172 | Vec3vf4& v1_o, |
173 | Vec3vf4& v2_o) |
174 | { |
175 | const Vec3vf4 tri_v012_x = gather(grid: grid_x,line_offset,lines); |
176 | const Vec3vf4 tri_v012_y = gather(grid: grid_y,line_offset,lines); |
177 | const Vec3vf4 tri_v012_z = gather(grid: grid_z,line_offset,lines); |
178 | v0_o = Vec3vf4(tri_v012_x[0],tri_v012_y[0],tri_v012_z[0]); |
179 | v1_o = Vec3vf4(tri_v012_x[1],tri_v012_y[1],tri_v012_z[1]); |
180 | v2_o = Vec3vf4(tri_v012_x[2],tri_v012_y[2],tri_v012_z[2]); |
181 | } |
182 | }; |
183 | |
184 | #if defined (__AVX__) |
185 | struct Gather3x3 |
186 | { |
187 | enum { M = 8 }; |
188 | typedef vbool8 vbool; |
189 | typedef vint8 vint; |
190 | typedef vfloat8 vfloat; |
191 | |
192 | static __forceinline const Vec3vf8 gather(const float* const grid, const size_t line_offset, const size_t lines) |
193 | { |
194 | vfloat4 ra = vfloat4::loadu(grid + 0*line_offset); |
195 | vfloat4 rb = vfloat4::loadu(grid + 1*line_offset); // this accesses 2 elements too much in case of 2x2 grid, but this is ok as we ensure enough padding after the grid |
196 | vfloat4 rc; |
197 | if (likely(lines > 2)) |
198 | rc = vfloat4::loadu(grid + 2*line_offset); |
199 | else |
200 | rc = rb; |
201 | |
202 | if (unlikely(line_offset == 2)) |
203 | { |
204 | ra = shuffle<0,1,1,1>(ra); |
205 | rb = shuffle<0,1,1,1>(rb); |
206 | rc = shuffle<0,1,1,1>(rc); |
207 | } |
208 | |
209 | const vfloat8 r0 = vfloat8(ra,rb); |
210 | const vfloat8 r1 = vfloat8(rb,rc); |
211 | return Vec3vf8(unpacklo(r0,r1), // r00, r10, r01, r11, r10, r20, r11, r21 |
212 | shuffle<1,1,2,2>(r0), // r01, r01, r02, r02, r11, r11, r12, r12 |
213 | shuffle<0,1,1,2>(r1)); // r10, r11, r11, r12, r20, r21, r21, r22 |
214 | } |
215 | |
216 | static __forceinline void gather(const float* const grid_x, |
217 | const float* const grid_y, |
218 | const float* const grid_z, |
219 | const size_t line_offset, |
220 | const size_t lines, |
221 | Vec3vf8& v0_o, |
222 | Vec3vf8& v1_o, |
223 | Vec3vf8& v2_o) |
224 | { |
225 | const Vec3vf8 tri_v012_x = gather(grid_x,line_offset,lines); |
226 | const Vec3vf8 tri_v012_y = gather(grid_y,line_offset,lines); |
227 | const Vec3vf8 tri_v012_z = gather(grid_z,line_offset,lines); |
228 | v0_o = Vec3vf8(tri_v012_x[0],tri_v012_y[0],tri_v012_z[0]); |
229 | v1_o = Vec3vf8(tri_v012_x[1],tri_v012_y[1],tri_v012_z[1]); |
230 | v2_o = Vec3vf8(tri_v012_x[2],tri_v012_y[2],tri_v012_z[2]); |
231 | } |
232 | }; |
233 | #endif |
234 | |
235 | template<typename vfloat> |
236 | static __forceinline Vec2<vfloat> decodeUV(const vfloat& uv) |
237 | { |
238 | typedef typename vfloat::Int vint; |
239 | const vint iu = asInt(uv) & 0xffff; |
240 | const vint iv = srl(asInt(uv),16); |
241 | const vfloat u = (vfloat)iu * vfloat(8.0f/0x10000); |
242 | const vfloat v = (vfloat)iv * vfloat(8.0f/0x10000); |
243 | return Vec2<vfloat>(u,v); |
244 | } |
245 | |
246 | __forceinline unsigned int geomID() const { |
247 | return _geomID; |
248 | } |
249 | |
250 | __forceinline unsigned int primID() const { |
251 | return _primID; |
252 | } |
253 | |
254 | public: |
255 | BVH4::NodeRef troot; |
256 | #if !defined(__64BIT__) |
257 | unsigned align1; |
258 | #endif |
259 | unsigned time_steps; |
260 | unsigned width; |
261 | |
262 | unsigned height; |
263 | unsigned dim_offset; |
264 | unsigned _geomID; |
265 | unsigned _primID; |
266 | |
267 | unsigned align2; |
268 | unsigned gridOffset; |
269 | unsigned gridBytes; |
270 | unsigned rootOffset; |
271 | |
272 | char data[1]; //!< after the struct we first store the BVH, then the grid, and finally the roots |
273 | }; |
274 | } |
275 | } |
276 | |