Batch kernels for backward pass of Preprocessing#3
Conversation
…-gaussian-rasterization into mlsys/batched_preprocess
…of batched_raster_settings
…in preprocess functions
…ion_backward_tests.py
…o/diff-gaussian-rasterization into mlsys/batched_preprocess
prapti19
left a comment
prapti19
left a comment
There was a problem hiding this comment.
Tested everything, works great!
sandeepnmenon
changed the title
| // Backward pass for conversion of spherical harmonics to RGB for | ||
| // each Gaussian. | ||
| __device__ void computeColorFromSH(int idx, int deg, int max_coeffs, const glm::vec3* means, glm::vec3 campos, const float* shs, const bool* clamped, const glm::vec3* dL_dcolor, glm::vec3* dL_dmeans, glm::vec3* dL_dshs) | ||
| __device__ void computeColorFromSH(int point_idx, int result_idx, int deg, int max_coeffs, const glm::vec3* means, glm::vec3 campos, const float* shs, const bool* clamped, const glm::vec3* dL_dcolor, glm::vec3* dL_dmeans, glm::vec3* dL_dshs) |
There was a problem hiding this comment.
We should make sure your two pull request will not conflict with each other when we merge both of them. Maybe we could do this together in tomorrow meeting.
There was a problem hiding this comment.
They are separate PRs. We'll resolve merge conflict after the forward pass PR is merged in.
| // TODO: remove it later by modifying FORWARD::preprocess when we deprecate sep_rendering==False case | ||
| uint32_t* tiles_touched_temp_buffer; | ||
| CHECK_CUDA(cudaMalloc(&tiles_touched_temp_buffer, P * sizeof(uint32_t)), debug); | ||
| CHECK_CUDA(cudaMemset(tiles_touched_temp_buffer, 0, P * sizeof(uint32_t)), debug); |
There was a problem hiding this comment.
Same as comments in the other PR. We could delete these memory allocation.
| torch.cuda.synchronize() | ||
| start_backward_event.record() | ||
|
|
||
| loss = compute_dummy_loss(means3D, scales, rotations, shs, opacity) |
There was a problem hiding this comment.
This might be wrong. To test the correctness of your batched implementation, we should indeed execute the backward kernels. However, the dummy loss only use original parameters as input. Therefore, we won't call the backward kernel of preprocess when you do loss.backward(). You should change to
loss = compute_dummy_loss( batched_means2D, batched_conic_opacity, batched_rgb, batched_depths, batched_radii)
3 participants
This is a follow up for PR #2 and should be merged after.
Changes
preprocess_gaussians_backward_batchedResults
Tests ran on V100
Time taken by run_batched_gaussian_rasterizer: 95.2484 ms
Time taken by run_batched_gaussian_rasterizer BACKWARD: 84.6223 ms
Time taken by run_batched_gaussian_rasterizer_batch_processing: 28.9444 ms
Time taken by run_batched_gaussian_rasterizer_batch_processing BACKWARD: 4.9838 ms