File size: 7,099 Bytes
186c7d4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
/*
 * Copyright (C) 2023, Inria
 * GRAPHDECO research group, https://team.inria.fr/graphdeco
 * All rights reserved.
 *
 * This software is free for non-commercial, research and evaluation use 
 * under the terms of the LICENSE.md file.
 *
 * For inquiries contact  [email protected]
 */

#include <math.h>
#include <torch/extension.h>
#include <cstdio>
#include <sstream>
#include <iostream>
#include <tuple>
#include <stdio.h>
#include <cuda_runtime_api.h>
#include <memory>
#include "cuda_rasterizer/config.h"
#include "cuda_rasterizer/rasterizer.h"
#include <fstream>
#include <string>
#include <functional>

std::function<char*(size_t N)> resizeFunctional(torch::Tensor& t) {
    auto lambda = [&t](size_t N) {
        t.resize_({(long long)N});
		return reinterpret_cast<char*>(t.contiguous().data_ptr());
    };
    return lambda;
}

std::tuple<int, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor>
RasterizeGaussiansCUDA(
	const torch::Tensor& background,
	const torch::Tensor& means3D,
    const torch::Tensor& colors,
    const torch::Tensor& opacity,
	const torch::Tensor& scales,
	const torch::Tensor& rotations,
	const float scale_modifier,
	const torch::Tensor& cov3D_precomp,
	const torch::Tensor& viewmatrix,
	const torch::Tensor& projmatrix,
	const float tan_fovx, 
	const float tan_fovy,
    const int image_height,
    const int image_width,
	const torch::Tensor& sh,
	const int degree,
	const torch::Tensor& campos,
	const bool prefiltered,
	const bool debug)
{
  if (means3D.ndimension() != 2 || means3D.size(1) != 3) {
    AT_ERROR("means3D must have dimensions (num_points, 3)");
  }
  
  const int P = means3D.size(0);
  const int H = image_height;
  const int W = image_width;

  auto int_opts = means3D.options().dtype(torch::kInt32);
  auto float_opts = means3D.options().dtype(torch::kFloat32);

  torch::Tensor out_color = torch::full({NUM_CHANNELS, H, W}, 0.0, float_opts);
  torch::Tensor out_depth = torch::full({1, H, W}, 0.0, float_opts);
  torch::Tensor radii = torch::full({P}, 0, means3D.options().dtype(torch::kInt32));
  
  torch::Device device(torch::kCUDA);
  torch::TensorOptions options(torch::kByte);
  torch::Tensor geomBuffer = torch::empty({0}, options.device(device));
  torch::Tensor binningBuffer = torch::empty({0}, options.device(device));
  torch::Tensor imgBuffer = torch::empty({0}, options.device(device));
  std::function<char*(size_t)> geomFunc = resizeFunctional(geomBuffer);
  std::function<char*(size_t)> binningFunc = resizeFunctional(binningBuffer);
  std::function<char*(size_t)> imgFunc = resizeFunctional(imgBuffer);
  
  int rendered = 0;
  if(P != 0)
  {
	  int M = 0;
	  if(sh.size(0) != 0)
	  {
		M = sh.size(1);
      }

	  rendered = CudaRasterizer::Rasterizer::forward(
	    geomFunc,
		binningFunc,
		imgFunc,
	    P, degree, M,
		background.contiguous().data<float>(),
		W, H,
		means3D.contiguous().data<float>(),
		sh.contiguous().data_ptr<float>(),
		colors.contiguous().data<float>(), 
		opacity.contiguous().data<float>(), 
		scales.contiguous().data_ptr<float>(),
		scale_modifier,
		rotations.contiguous().data_ptr<float>(),
		cov3D_precomp.contiguous().data<float>(), 
		viewmatrix.contiguous().data<float>(), 
		projmatrix.contiguous().data<float>(),
		campos.contiguous().data<float>(),
		tan_fovx,
		tan_fovy,
		prefiltered,
		out_color.contiguous().data<float>(),
		out_depth.contiguous().data<float>(),
		radii.contiguous().data<int>(),
		debug);
  }
  return std::make_tuple(rendered, out_color, out_depth, radii, geomBuffer, binningBuffer, imgBuffer);
}

std::tuple<torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor>
 RasterizeGaussiansBackwardCUDA(
 	const torch::Tensor& background,
	const torch::Tensor& means3D,
	const torch::Tensor& radii,
    const torch::Tensor& colors,
	const torch::Tensor& scales,
	const torch::Tensor& rotations,
	const float scale_modifier,
	const torch::Tensor& cov3D_precomp,
	const torch::Tensor& viewmatrix,
    const torch::Tensor& projmatrix,
	const float tan_fovx,
	const float tan_fovy,
    const torch::Tensor& dL_dout_color,
        const torch::Tensor& dL_dout_depth,
	const torch::Tensor& sh,
	const int degree,
	const torch::Tensor& campos,
	const torch::Tensor& geomBuffer,
	const int R,
	const torch::Tensor& binningBuffer,
	const torch::Tensor& imageBuffer,
	const bool debug) 
{
  const int P = means3D.size(0);
  const int H = dL_dout_color.size(1);
  const int W = dL_dout_color.size(2);
  
  int M = 0;
  if(sh.size(0) != 0)
  {	
	M = sh.size(1);
  }

  torch::Tensor dL_dmeans3D = torch::zeros({P, 3}, means3D.options());
  torch::Tensor dL_dmeans2D = torch::zeros({P, 3}, means3D.options());
  torch::Tensor dL_dcolors = torch::zeros({P, NUM_CHANNELS}, means3D.options());
  torch::Tensor dL_dconic = torch::zeros({P, 2, 2}, means3D.options());
  torch::Tensor dL_dopacity = torch::zeros({P, 1}, means3D.options());
  torch::Tensor dL_dcov3D = torch::zeros({P, 6}, means3D.options());
  torch::Tensor dL_dsh = torch::zeros({P, M, 3}, means3D.options());
  torch::Tensor dL_dscales = torch::zeros({P, 3}, means3D.options());
  torch::Tensor dL_drotations = torch::zeros({P, 4}, means3D.options());
  
  if(P != 0)
  {  
	  CudaRasterizer::Rasterizer::backward(P, degree, M, R,
	  background.contiguous().data<float>(),
	  W, H, 
	  means3D.contiguous().data<float>(),
	  sh.contiguous().data<float>(),
	  colors.contiguous().data<float>(),
	  scales.data_ptr<float>(),
	  scale_modifier,
	  rotations.data_ptr<float>(),
	  cov3D_precomp.contiguous().data<float>(),
	  viewmatrix.contiguous().data<float>(),
	  projmatrix.contiguous().data<float>(),
	  campos.contiguous().data<float>(),
	  tan_fovx,
	  tan_fovy,
	  radii.contiguous().data<int>(),
	  reinterpret_cast<char*>(geomBuffer.contiguous().data_ptr()),
	  reinterpret_cast<char*>(binningBuffer.contiguous().data_ptr()),
	  reinterpret_cast<char*>(imageBuffer.contiguous().data_ptr()),
	  dL_dout_color.contiguous().data<float>(),
	  dL_dout_depth.contiguous().data<float>(),
	  dL_dmeans2D.contiguous().data<float>(),
	  dL_dconic.contiguous().data<float>(),  
	  dL_dopacity.contiguous().data<float>(),
	  dL_dcolors.contiguous().data<float>(),
	  dL_dmeans3D.contiguous().data<float>(),
	  dL_dcov3D.contiguous().data<float>(),
	  dL_dsh.contiguous().data<float>(),
	  dL_dscales.contiguous().data<float>(),
	  dL_drotations.contiguous().data<float>(),
	  debug);
  }

  return std::make_tuple(dL_dmeans2D, dL_dcolors, dL_dopacity, dL_dmeans3D, dL_dcov3D, dL_dsh, dL_dscales, dL_drotations);
}

torch::Tensor markVisible(
		torch::Tensor& means3D,
		torch::Tensor& viewmatrix,
		torch::Tensor& projmatrix)
{ 
  const int P = means3D.size(0);
  
  torch::Tensor present = torch::full({P}, false, means3D.options().dtype(at::kBool));
 
  if(P != 0)
  {
	CudaRasterizer::Rasterizer::markVisible(P,
		means3D.contiguous().data<float>(),
		viewmatrix.contiguous().data<float>(),
		projmatrix.contiguous().data<float>(),
		present.contiguous().data<bool>());
  }
  
  return present;
}