Hello, may I ask why the sorting step still seems to exist in the forward function? Has it not been removed?

[hwlys]

CHECK_CUDA(cub::DeviceScan::InclusiveSum(geomState.scanning_space, geomState.scan_size, geomState.tiles_touched, geomState.point_offsets, P), debug)

// Retrieve total number of Gaussian instances to launch and resize aux buffers

CHECK_CUDA(cudaMemcpy(&num_rendered_gs, geomState.point_offsets + P - 1, sizeof(int), cudaMemcpyDeviceToHost), debug);
// printf("num rendered: %d\n", num_rendered);

size_t gs_binning_chunk_size = required<GS_BinningState>(num_rendered_gs);
char* gs_binning_chunkptr = gs_binningBuffer(gs_binning_chunk_size);
gs_binningState = GS_BinningState::fromChunk(gs_binning_chunkptr, num_rendered_gs);

// For each instance to be rendered, produce adequate [ tile | depth ] key 
// and corresponding dublicated Gaussian indices to be sorted
duplicateWithKeys << <(P + 255) / 256, 256 >> > (
	P,
	geomState.means2D,
	geomState.depths,
	geomState.point_offsets,
	gs_binningState.point_list_keys_unsorted,
	gs_binningState.point_list_unsorted,
	radii,
	tile_grid)
CHECK_CUDA(, debug)

int bit = getHigherMsb(tile_grid.x * tile_grid.y);

// Sort complete list of (duplicated) Gaussian indices by keys
CHECK_CUDA(cub::DeviceRadixSort::SortPairs(
	gs_binningState.list_sorting_space,
	gs_binningState.sorting_size,
	gs_binningState.point_list_keys_unsorted, gs_binningState.point_list_keys,
	gs_binningState.point_list_unsorted, gs_binningState.point_list,
	num_rendered_gs, 0, bit), debug)

CHECK_CUDA(cudaMemset(imgState.ranges, 0, tile_grid.x * tile_grid.y * sizeof(uint2)), debug);

// Identify start and end of per-tile workloads in sorted list
if (num_rendered_gs > 0)
	identifyTileRanges << <(num_rendered_gs + 255) / 256, 256 >> > (
		num_rendered_gs,
		gs_binningState.point_list_keys,
		imgState.ranges);
CHECK_CUDA(, debug)

// Let each tile blend its range of Gaussians independently in parallel
const float* feature_ptr = colors_precomp != nullptr ? colors_precomp : geomState.rgb;

CHECK_CUDA(FORWARD::render(
	weight_background,
	tile_grid, block,
	imgState.ranges,
	gs_binningState.point_list,
	width, height,
	geomState.means2D,
	feature_ptr,
	geomState.conic_opacity,
	geomState.alpha_weight,
	imgState.weight_sum,
	background,
	out_color
	), debug)

return num_rendered_gs;

}

[LiYukeee]

Our implementation of sort-free Gaussian Splatting still uses the tile-based rendering approach from 3DGS. This means we need to assign the Gaussians (GS) to different tiles. Think of the ‘sort’ part here as replacing grouping with sorting instead.

You can see this explained in the README of our GitHub repo https://github.com/LiYukeee/sort-free-diff-gaussian-rasterization:
The code still includes a Radix Sort, but we have removed the GS depth component from the keys required by the Radix Sort. This means that sorting is now based solely on the tile ID to which the GS belongs. This approach is a form of lazy approach, using sorting as a substitute for classification. While this change may impact performance, it will not affect rendering quality.

[hwlys]

Thank you!