Add interface for depth in both forward rendering and backward propagation

cuda_rasterizer/forward.cu

@@ -266,11 +266,13 @@ renderCUDA(
 	int W, int H,
 	const float2* __restrict__ points_xy_image,
 	const float* __restrict__ features,
+	const float* __restrict__ depths,
 	const float4* __restrict__ conic_opacity,
 	float* __restrict__ final_T,
 	uint32_t* __restrict__ n_contrib,
 	const float* __restrict__ bg_color,
-	float* __restrict__ out_color)
+	float* __restrict__ out_color,
+	float* __restrict__ out_depth)
 {
 	// Identify current tile and associated min/max pixel range.
 	auto block = cg::this_thread_block();
@@ -301,6 +303,7 @@ renderCUDA(
 	uint32_t contributor = 0;
 	uint32_t last_contributor = 0;
 	float C[CHANNELS] = { 0 };
+	float D = { 0 };
 
 	// Iterate over batches until all done or range is complete
 	for (int i = 0; i < rounds; i++, toDo -= BLOCK_SIZE)
@@ -353,6 +356,7 @@ renderCUDA(
 			// Eq. (3) from 3D Gaussian splatting paper.
 			for (int ch = 0; ch < CHANNELS; ch++)
 				C[ch] += features[collected_id[j] * CHANNELS + ch] * alpha * T;
+			D += depths[collected_id[j]] * alpha * T;
 
 			T = test_T;
 
@@ -370,6 +374,7 @@ renderCUDA(
 		n_contrib[pix_id] = last_contributor;
 		for (int ch = 0; ch < CHANNELS; ch++)
 			out_color[ch * H * W + pix_id] = C[ch] + T * bg_color[ch];
+		out_depth[pix_id] = D;
 	}
 }
 
@@ -380,23 +385,27 @@ void FORWARD::render(
 	int W, int H,
 	const float2* means2D,
 	const float* colors,
+	const float* depths,
 	const float4* conic_opacity,
 	float* final_T,
 	uint32_t* n_contrib,
 	const float* bg_color,
-	float* out_color)
+	float* out_color,
+	float* out_depth)
 {
 	renderCUDA<NUM_CHANNELS> << <grid, block >> > (
 		ranges,
 		point_list,
 		W, H,
 		means2D,
 		colors,
+		depths,
 		conic_opacity,
 		final_T,
 		n_contrib,
 		bg_color,
-		out_color);
+		out_color,
+		out_depth);
 }
 
 void FORWARD::preprocess(int P, int D, int M,
@@ -452,4 +461,4 @@ void FORWARD::preprocess(int P, int D, int M,
 		tiles_touched,
 		prefiltered
 		);
-}
+}

cuda_rasterizer/forward.h

@@ -55,11 +55,13 @@ namespace FORWARD
 		int W, int H,
 		const float2* points_xy_image,
 		const float* features,
+		const float* depths,
 		const float4* conic_opacity,
 		float* final_T,
 		uint32_t* n_contrib,
 		const float* bg_color,
-		float* out_color);
+		float* out_color,
+		float* out_depth);
 }
 
 

cuda_rasterizer/rasterizer.h

@@ -49,6 +49,7 @@ namespace CudaRasterizer
 			const float tan_fovx, float tan_fovy,
 			const bool prefiltered,
 			float* out_color,
+			float* out_depth,
 			int* radii = nullptr);
 
 		static void backward(

cuda_rasterizer/rasterizer_impl.cu

@@ -216,6 +216,7 @@ int CudaRasterizer::Rasterizer::forward(
 	const float tan_fovx, float tan_fovy,
 	const bool prefiltered,
 	float* out_color,
+	float* out_depth,
 	int* radii)
 {
 	const float focal_y = height / (2.0f * tan_fovy);
@@ -326,11 +327,13 @@ int CudaRasterizer::Rasterizer::forward(
 		width, height,
 		geomState.means2D,
 		feature_ptr,
+		geomState.depths,
 		geomState.conic_opacity,
 		imgState.accum_alpha,
 		imgState.n_contrib,
 		background,
-		out_color);
+		out_color,
+		out_depth);
 
 	return num_rendered;
 }
@@ -430,4 +433,4 @@ void CudaRasterizer::Rasterizer::backward(
 		dL_dsh,
 		(glm::vec3*)dL_dscale,
 		(glm::vec4*)dL_drot);
-}
+}

diff_gaussian_rasterization/__init__.py

@@ -75,16 +75,16 @@ def forward(
         )
 
         # Invoke C++/CUDA rasterizer
-        num_rendered, color, radii, geomBuffer, binningBuffer, imgBuffer = _C.rasterize_gaussians(*args)
+        num_rendered, color, depth, radii, geomBuffer, binningBuffer, imgBuffer = _C.rasterize_gaussians(*args)
 
         # Keep relevant tensors for backward
         ctx.raster_settings = raster_settings
         ctx.num_rendered = num_rendered
         ctx.save_for_backward(colors_precomp, means3D, scales, rotations, cov3Ds_precomp, radii, sh, geomBuffer, binningBuffer, imgBuffer)
-        return color, radii
+        return color, radii, depth
 
     @staticmethod
-    def backward(ctx, grad_out_color, _):
+    def backward(ctx, grad_out_color, grad_radii, grad_depth):
 
         # Restore necessary values from context
         num_rendered = ctx.num_rendered

rasterize_points.cu

@@ -32,7 +32,7 @@ std::function<char*(size_t N)> resizeFunctional(torch::Tensor& t) {
     return lambda;
 }
 
-std::tuple<int, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor>
+std::tuple<int, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor>
 RasterizeGaussiansCUDA(
 	const torch::Tensor& background,
 	const torch::Tensor& means3D,
@@ -65,6 +65,7 @@ RasterizeGaussiansCUDA(
   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);
@@ -107,9 +108,10 @@ RasterizeGaussiansCUDA(
 		tan_fovy,
 		prefiltered,
 		out_color.contiguous().data<float>(),
+		out_depth.contiguous().data<float>(),
 		radii.contiguous().data<int>());
   }
-  return std::make_tuple(rendered, out_color, radii, geomBuffer, binningBuffer, imgBuffer);
+  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>

rasterize_points.h

@@ -15,7 +15,7 @@
 #include <tuple>
 #include <string>
 
-std::tuple<int, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor>
+std::tuple<int, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor>
 RasterizeGaussiansCUDA(
 	const torch::Tensor& background,
 	const torch::Tensor& means3D,