From 88e7ec6b3f291c795e8822788fc1931afdc780bb Mon Sep 17 00:00:00 2001 From: Matthieu Gendrin Date: Wed, 30 Aug 2023 13:58:04 +0200 Subject: [PATCH] Fix rgba export / filter output based on occupancy grid Signed-off-by: Matthieu Gendrin --- src/testbed_nerf.cu | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+) diff --git a/src/testbed_nerf.cu b/src/testbed_nerf.cu index 404512d5e..15e59cb8e 100644 --- a/src/testbed_nerf.cu +++ b/src/testbed_nerf.cu @@ -3066,6 +3066,19 @@ GPUMemory Testbed::get_density_on_grid(ivec3 res3d, const BoundingBox& aa return density; } +__global__ void filter_with_occupancy(const uint32_t n_elements, float* pos, const uint32_t floats_per_coord, const uint8_t* density_grid_bitfield, float* rgbsigma) { + const uint32_t point_id = threadIdx.x + blockIdx.x * blockDim.x; + if (point_id >= n_elements) return; + const vec3 pos_vec{pos[point_id * floats_per_coord], pos[point_id * floats_per_coord + 1], pos[point_id * floats_per_coord + 2]}; + const uint32_t mip = mip_from_pos(pos_vec); + if (!density_grid_occupied_at(pos_vec, density_grid_bitfield, mip)) { + #pragma unroll + for (int i = 0; i < 4; i++) { // sigma=0 would be enough, but all to 0 compress better + rgbsigma[point_id * 4 + i] = 0.f; + } + } +} + GPUMemory Testbed::get_rgba_on_grid(ivec3 res3d, vec3 ray_dir, bool voxel_centers, float depth, bool density_as_alpha) { const uint32_t n_elements = (res3d.x*res3d.y*res3d.z); GPUMemory rgba(n_elements); @@ -3102,6 +3115,12 @@ GPUMemory Testbed::get_rgba_on_grid(ivec3 res3d, vec3 ray_dir, bool voxel_ GPUMatrix positions_matrix((float*)(positions.data() + offset * floats_per_coord), floats_per_coord, local_batch_size); GPUMatrix rgbsigma_matrix((float*)(rgba.data() + offset), 4, local_batch_size); m_network->inference(m_stream.get(), positions_matrix, rgbsigma_matrix); + linear_kernel(filter_with_occupancy, 0, m_stream.get(), + local_batch_size, + positions_matrix.data(), + floats_per_coord, + m_nerf.density_grid_bitfield.data(), + rgbsigma_matrix.data()); // convert network output to RGBA (in place) linear_kernel(compute_nerf_rgba_kernel, 0, m_stream.get(), local_batch_size, rgba.data() + offset, m_nerf.rgb_activation, m_nerf.density_activation, depth, density_as_alpha);