Question for render feature?

[SYSUykLin]

Hello:
I previously attempted to render features with 256 dimensions, but CUDA indicated insufficient shared memory, allowing for a maximum of only 40 dimensions to be rendered. May I ask what changes you made to enable it to render 256 dimensions?

[41xu]

As far as I understand, in the rasterization process, they use shared memory for calculating the collected features/colors and for gradient calculation. The shared memory is limited by specific GPU. In this paper, they dynamically allocate a cuda array as a cache for the collected features to avoid using shared memory (of course it's the tradeoff between the need for dimension and shared memory issue). You can see the implementation here:

feature-3dgs/submodules/diff-gaussian-rasterization/cuda_rasterizer/rasterizer_impl.cu

Line 398 in 9e714ff

cudaMalloc((void**)&collected_semantic_feature, NUM_SEMANTIC_CHANNELS * BLOCK_SIZE * sizeof(float));

If I misunderstand, please point me out.

[JrMeng0312]

graphdeco-inria/gaussian-splatting#41 (comment) you can try this: adding "-Xcompiler -fno-gnu-unique" option in submodules/diff-gaussian-rasterization/setup.py: line 29 resolves the illegal memory access error in training.

extra_compile_args={"nvcc": ["-Xcompiler", "-fno-gnu-unique","-I" + os.path.join(os.path.dirname(os.path.abspath(file)), "third_party/glm/")]})

[SYSUykLin]

As far as I understand, in the rasterization process, they use shared memory for calculating the collected features/colors and for gradient calculation. The shared memory is limited by specific GPU. In this paper, they dynamically allocate a cuda array as a cache for the collected features to avoid using shared memory (of course it's the tradeoff between the need for dimension and shared memory issue). You can see the implementation here:

feature-3dgs/submodules/diff-gaussian-rasterization/cuda_rasterizer/rasterizer_impl.cu

Line 398 in 9e714ff

cudaMalloc((void**)&collected_semantic_feature, NUM_SEMANTIC_CHANNELS * BLOCK_SIZE * sizeof(float));

If I misunderstand, please point me out.

Thanks very very very much.

[SYSUykLin]

As far as I understand, in the rasterization process, they use shared memory for calculating the collected features/colors and for gradient calculation. The shared memory is limited by specific GPU. In this paper, they dynamically allocate a cuda array as a cache for the collected features to avoid using shared memory (of course it's the tradeoff between the need for dimension and shared memory issue). You can see the implementation here:

feature-3dgs/submodules/diff-gaussian-rasterization/cuda_rasterizer/rasterizer_impl.cu

Line 398 in 9e714ff

cudaMalloc((void**)&collected_semantic_feature, NUM_SEMANTIC_CHANNELS * BLOCK_SIZE * sizeof(float));

If I misunderstand, please point me out.

Thanks very very very much.

Thanks