Run tests in github action not circleci (#1896)

Summary: Pull Request resolved: https://github.com/facebookresearch/pytorch3d/pull/1896

Differential Revision: D65272512

Pulled By: bottler

.circleci/regenerate.py

@@ -88,6 +88,7 @@ def workflow_pair(
     upload=False,
     filter_branch,
 ):
+
     w = []
     py = python_version.replace(".", "")
     pyt = pytorch_version.replace(".", "")
@@ -126,6 +127,7 @@ def generate_base_workflow(
     btype,
     filter_branch=None,
 ):
+
     d = {
         "name": base_workflow_name,
         "python_version": python_version,

.github/workflows/build.yml

@@ -3,9 +3,6 @@ on:
   pull_request:
     branches:
       - main
-  push:
-    branches:
-      - main
 jobs:
   binary_linux_conda_cuda:
     runs-on: 4-core-ubuntu-gpu-t4
@@ -21,45 +18,3 @@ jobs:
       run: |-
         conda create --name env --yes --quiet conda-build
         conda run --no-capture-output --name env python3 ./packaging/build_conda.py --use-conda-cuda
-
-  # Build-only verification for the ROCm/HIP code paths. Runs in an AMD ROCm
-  # dev container on a CPU-only GitHub runner; we don't need an AMD GPU just
-  # to compile, and not running tests keeps the CI cost low. Catches build
-  # regressions in the ROCm code paths (USE_ROCM guards, hipify-touched sources,
-  # the pulsar HIP intrinsic replacements, etc.).
-  linux_rocm_build:
-    runs-on: ubuntu-latest
-    container:
-      # `-complete` tag bundles the full ROCm math stack (rocThrust, hipCUB,
-      # rocPRIM, ...). The plain `7.2.3` tag is HIP-runtime-only and fails to
-      # find <thrust/complex.h> when including PyTorch headers.
-      image: rocm/dev-ubuntu-22.04:7.2.3-complete
-    env:
-      PYTORCH_VERSION: "2.11.0"
-      ROCM_INDEX: "rocm7.2"
-    steps:
-    - uses: actions/checkout@v4
-    - name: Install Python and torch+rocm
-      run: |-
-        apt-get update
-        apt-get install -y --no-install-recommends python3 python3-dev python3-pip git
-        python3 -m pip install --upgrade pip
-        python3 -m pip install --index-url https://download.pytorch.org/whl/${ROCM_INDEX} torch==${PYTORCH_VERSION}
-    - name: Verify torch is ROCm-built
-      run: |-
-        python3 -c "import torch; assert torch.version.hip is not None, 'torch is not HIP-built'; print('torch.version.hip:', torch.version.hip)"
-    - name: Build pytorch3d _C extension (build only, no tests)
-      env:
-        # CPU-only runner: torch.cuda.is_available() is False, so force the
-        # CUDAExtension path. ROCM_HOME is auto-detected from /opt/rocm in
-        # the rocm/dev-ubuntu container.
-        FORCE_CUDA: "1"
-      run: |-
-        python3 -m pip install --no-build-isolation -v .
-    - name: Smoke import
-      # cd out of the checkout root so the source-tree pytorch3d/ directory
-      # (which has no _C.so since the build doesn't install in-place) doesn't
-      # shadow the site-packages install via sys.path[0] for `python -c`.
-      run: |-
-        cd /tmp
-        python3 -c "import torch; from pytorch3d import _C; print('PulsarRenderer:', hasattr(_C, 'PulsarRenderer')); print('n_symbols:', len([s for s in dir(_C) if not s.startswith('_')]))"

.gitignore

@@ -6,17 +6,6 @@ dist/
 **/.ipynb_checkpoints
 **/.ipynb_checkpoints/**
 
-# Build artifacts produced in-place by torch.utils.cpp_extension auto-hipify
-# when pytorch3d is built against a ROCm PyTorch.
-*.so
-pytorch3d/csrc/**/*.hip
-pytorch3d/csrc/**/*_hip.cpp
-pytorch3d/csrc/**/*_hip.h
-pytorch3d/csrc/**/*_hip.cuh
-
-# Debug PNG dumps written by pulsar tests when FB_TEST is unset.
-tests/pulsar/test_out/
-
 
 # Docusaurus site
 website/yarn.lock

dev/linter.sh

@@ -10,7 +10,7 @@
 DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 DIR=$(dirname "${DIR}")
 
-if [[ -f "${DIR}/BUCK" ]]
+if [[ -f "${DIR}/TARGETS" ]]
 then
   pyfmt "${DIR}"
 else
@@ -36,5 +36,5 @@ then
 
   echo "Running pyre..."
   echo "To restart/kill pyre server, run 'pyre restart' or 'pyre kill' in fbcode/"
-  ( cd ~/fbsource/fbcode; arc pyre check //vision/fair/pytorch3d/... )
+  ( cd ~/fbsource/fbcode; pyre -l vision/fair/pytorch3d/ )
 fi

docs/conf.py

@@ -19,6 +19,7 @@
 #
 import os
 import sys
+
 import unittest.mock as mock
 
 from recommonmark.parser import CommonMarkParser

docs/examples/pulsar_basic.py

@@ -10,7 +10,6 @@ This example demonstrates the most trivial, direct interface of the pulsar
 sphere renderer. It renders and saves an image with 10 random spheres.
 Output: basic.png.
 """
-
 import logging
 import math
 from os import path

docs/examples/pulsar_basic_unified.py

@@ -11,7 +11,6 @@ interface for sphere renderering. It renders and saves an image with
 10 random spheres.
 Output: basic-pt3d.png.
 """
-
 import logging
 from os import path
 

docs/examples/pulsar_cam.py

@@ -14,7 +14,6 @@ distorted. Gradient-based optimization is used to converge towards the
 original camera parameters.
 Output: cam.gif.
 """
-
 import logging
 import math
 from os import path

docs/examples/pulsar_cam_unified.py

@@ -14,7 +14,6 @@ distorted. Gradient-based optimization is used to converge towards the
 original camera parameters.
 Output: cam-pt3d.gif
 """
-
 import logging
 from os import path
 

docs/examples/pulsar_multiview.py

@@ -18,7 +18,6 @@ This example is not available yet through the 'unified' interface,
 because opacity support has not landed in PyTorch3D for general data
 structures yet.
 """
-
 import logging
 import math
 from os import path

docs/examples/pulsar_optimization.py

@@ -13,7 +13,6 @@ The scene is initialized with random spheres. Gradient-based
 optimization is used to converge towards a faithful
 scene representation.
 """
-
 import logging
 import math
 

docs/examples/pulsar_optimization_unified.py

@@ -13,7 +13,6 @@ The scene is initialized with random spheres. Gradient-based
 optimization is used to converge towards a faithful
 scene representation.
 """
-
 import logging
 import math
 

docs/modules/implicitron/datasets.rst

@@ -3,6 +3,11 @@ pytorch3d.implicitron.dataset specific datasets
 
 specific datasets
 
+.. automodule:: pytorch3d.implicitron.dataset.blender_dataset_map_provider
+    :members:
+    :undoc-members:
+    :show-inheritance:
+
 .. automodule:: pytorch3d.implicitron.dataset.json_index_dataset_map_provider
     :members:
     :undoc-members:
@@ -13,6 +18,11 @@ specific datasets
     :undoc-members:
     :show-inheritance:
 
+.. automodule:: pytorch3d.implicitron.dataset.llff_dataset_map_provider
+    :members:
+    :undoc-members:
+    :show-inheritance:
+
 .. automodule:: pytorch3d.implicitron.dataset.rendered_mesh_dataset_map_provider
     :members:
     :undoc-members:

packaging/pytorch3d/meta.yaml

@@ -32,6 +32,7 @@ requirements:
 
 build:
   string: py{{py}}_{{ environ['CU_VERSION'] }}_pyt{{ environ['PYTORCH_VERSION_NODOT']}}
+  # script: LD_LIBRARY_PATH=$PREFIX/lib:$BUILD_PREFIX/lib:$LD_LIBRARY_PATH python setup.py install --single-version-externally-managed --record=record.txt # [not win]
   script: python setup.py install --single-version-externally-managed --record=record.txt # [not win]
   script_env:
     - CUDA_HOME
@@ -56,6 +57,7 @@ test:
     - pandas
     - sqlalchemy
   commands:
+    #pytest .
     python -m unittest discover -v -s tests -t .
 
 

projects/implicitron_trainer/configs/overfit_singleseq_nerf_blender.yaml

@@ -0,0 +1,56 @@
+defaults:
+- overfit_singleseq_base
+- _self_
+exp_dir: "./data/overfit_nerf_blender_repro/${oc.env:BLENDER_SINGLESEQ_CLASS}"
+data_source_ImplicitronDataSource_args:
+  data_loader_map_provider_SequenceDataLoaderMapProvider_args:
+    dataset_length_train: 100
+  dataset_map_provider_class_type: BlenderDatasetMapProvider
+  dataset_map_provider_BlenderDatasetMapProvider_args:
+    base_dir: ${oc.env:BLENDER_DATASET_ROOT}/${oc.env:BLENDER_SINGLESEQ_CLASS}
+    n_known_frames_for_test: null
+    object_name: ${oc.env:BLENDER_SINGLESEQ_CLASS}
+    path_manager_factory_class_type: PathManagerFactory
+    path_manager_factory_PathManagerFactory_args:
+      silence_logs: true
+
+model_factory_ImplicitronModelFactory_args:
+  model_class_type: "OverfitModel"
+  model_OverfitModel_args:
+    mask_images: false
+    raysampler_class_type: AdaptiveRaySampler
+    raysampler_AdaptiveRaySampler_args:
+      n_pts_per_ray_training: 64
+      n_pts_per_ray_evaluation: 64
+      n_rays_per_image_sampled_from_mask: 4096
+      stratified_point_sampling_training: true
+      stratified_point_sampling_evaluation: false
+      scene_extent: 2.0
+      scene_center:
+      - 0.0
+      - 0.0
+      - 0.0
+    renderer_MultiPassEmissionAbsorptionRenderer_args:
+      density_noise_std_train: 0.0
+      n_pts_per_ray_fine_training: 128
+      n_pts_per_ray_fine_evaluation: 128
+      raymarcher_EmissionAbsorptionRaymarcher_args:
+        blend_output: false
+    loss_weights:
+      loss_rgb_mse: 1.0
+      loss_prev_stage_rgb_mse: 1.0
+      loss_mask_bce: 0.0
+      loss_prev_stage_mask_bce: 0.0
+      loss_autodecoder_norm: 0.00
+
+optimizer_factory_ImplicitronOptimizerFactory_args:
+  exponential_lr_step_size: 3001
+  lr_policy: LinearExponential
+  linear_exponential_lr_milestone: 200
+
+training_loop_ImplicitronTrainingLoop_args:
+  max_epochs: 6000
+  metric_print_interval: 10
+  store_checkpoints_purge: 3
+  test_when_finished: true
+  validation_interval: 100

projects/implicitron_trainer/configs/repro_singleseq_nerf_blender.yaml

@@ -0,0 +1,55 @@
+defaults:
+- repro_singleseq_base
+- _self_
+exp_dir: "./data/nerf_blender_repro/${oc.env:BLENDER_SINGLESEQ_CLASS}"
+data_source_ImplicitronDataSource_args:
+  data_loader_map_provider_SequenceDataLoaderMapProvider_args:
+    dataset_length_train: 100
+  dataset_map_provider_class_type: BlenderDatasetMapProvider
+  dataset_map_provider_BlenderDatasetMapProvider_args:
+    base_dir: ${oc.env:BLENDER_DATASET_ROOT}/${oc.env:BLENDER_SINGLESEQ_CLASS}
+    n_known_frames_for_test: null
+    object_name: ${oc.env:BLENDER_SINGLESEQ_CLASS}
+    path_manager_factory_class_type: PathManagerFactory
+    path_manager_factory_PathManagerFactory_args:
+      silence_logs: true
+
+model_factory_ImplicitronModelFactory_args:
+  model_GenericModel_args:
+    mask_images: false
+    raysampler_class_type: AdaptiveRaySampler
+    raysampler_AdaptiveRaySampler_args:
+      n_pts_per_ray_training: 64
+      n_pts_per_ray_evaluation: 64
+      n_rays_per_image_sampled_from_mask: 4096
+      stratified_point_sampling_training: true
+      stratified_point_sampling_evaluation: false
+      scene_extent: 2.0
+      scene_center:
+      - 0.0
+      - 0.0
+      - 0.0
+    renderer_MultiPassEmissionAbsorptionRenderer_args:
+      density_noise_std_train: 0.0
+      n_pts_per_ray_fine_training: 128
+      n_pts_per_ray_fine_evaluation: 128
+      raymarcher_EmissionAbsorptionRaymarcher_args:
+        blend_output: false
+    loss_weights:
+      loss_rgb_mse: 1.0
+      loss_prev_stage_rgb_mse: 1.0
+      loss_mask_bce: 0.0
+      loss_prev_stage_mask_bce: 0.0
+      loss_autodecoder_norm: 0.00
+
+optimizer_factory_ImplicitronOptimizerFactory_args:
+  exponential_lr_step_size: 3001
+  lr_policy: LinearExponential
+  linear_exponential_lr_milestone: 200
+
+training_loop_ImplicitronTrainingLoop_args:
+  max_epochs: 6000
+  metric_print_interval: 10
+  store_checkpoints_purge: 3
+  test_when_finished: true
+  validation_interval: 100

projects/implicitron_trainer/experiment.py

@@ -7,7 +7,7 @@
 
 # pyre-unsafe
 
-""" "
+""""
 This file is the entry point for launching experiments with Implicitron.
 
 Launch Training
@@ -44,22 +44,25 @@ The outputs of the experiment are saved and logged in multiple ways:
         config file.
 
 """
-
 import logging
 import os
 import warnings
+
 from dataclasses import field
 
 import hydra
+
 import torch
 from accelerate import Accelerator
 from omegaconf import DictConfig, OmegaConf
 from packaging import version
+
 from pytorch3d.implicitron.dataset.data_source import (
     DataSourceBase,
     ImplicitronDataSource,
 )
 from pytorch3d.implicitron.models.base_model import ImplicitronModelBase
+
 from pytorch3d.implicitron.models.renderer.multipass_ea import (
     MultiPassEmissionAbsorptionRenderer,
 )
@@ -117,12 +120,16 @@ class Experiment(Configurable):
             will be saved here.
     """
 
+    # pyre-fixme[13]: Attribute `data_source` is never initialized.
     data_source: DataSourceBase
     data_source_class_type: str = "ImplicitronDataSource"
+    # pyre-fixme[13]: Attribute `model_factory` is never initialized.
     model_factory: ModelFactoryBase
     model_factory_class_type: str = "ImplicitronModelFactory"
+    # pyre-fixme[13]: Attribute `optimizer_factory` is never initialized.
     optimizer_factory: OptimizerFactoryBase
     optimizer_factory_class_type: str = "ImplicitronOptimizerFactory"
+    # pyre-fixme[13]: Attribute `training_loop` is never initialized.
     training_loop: TrainingLoopBase
     training_loop_class_type: str = "ImplicitronTrainingLoop"
 
@@ -189,7 +196,6 @@ class Experiment(Configurable):
             last_epoch=start_epoch,
             model=model,
             resume=self.model_factory.resume,
-            # pyrefly: ignore [missing-attribute]
             resume_epoch=self.model_factory.resume_epoch,
         )
 
@@ -209,7 +215,6 @@ class Experiment(Configurable):
 
         # Enter the main training loop.
         self.training_loop.run(
-            # pyrefly: ignore [bad-argument-type]
             train_loader=train_loader,
             val_loader=val_loader,
             test_loader=test_loader,

projects/implicitron_trainer/impl/model_factory.py

@@ -11,6 +11,7 @@ import os
 from typing import Optional
 
 import torch.optim
+
 from accelerate import Accelerator
 from pytorch3d.implicitron.models.base_model import ImplicitronModelBase
 from pytorch3d.implicitron.tools import model_io
@@ -25,6 +26,7 @@ logger = logging.getLogger(__name__)
 
 
 class ModelFactoryBase(ReplaceableBase):
+
     resume: bool = True  # resume from the last checkpoint
 
     def __call__(self, **kwargs) -> ImplicitronModelBase:
@@ -59,6 +61,7 @@ class ImplicitronModelFactory(ModelFactoryBase):
 
     """
 
+    # pyre-fixme[13]: Attribute `model` is never initialized.
     model: ImplicitronModelBase
     model_class_type: str = "GenericModel"
     resume: bool = True
@@ -113,9 +116,7 @@ class ImplicitronModelFactory(ModelFactoryBase):
                         "cuda:%d" % 0: "cuda:%d" % accelerator.local_process_index
                     }
                 model_state_dict = torch.load(
-                    model_io.get_model_path(model_path),
-                    map_location=map_location,
-                    weights_only=True,
+                    model_io.get_model_path(model_path), map_location=map_location
                 )
 
                 try:

projects/implicitron_trainer/impl/optimizer_factory.py

@@ -14,7 +14,9 @@ from dataclasses import field
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch.optim
+
 from accelerate import Accelerator
+
 from pytorch3d.implicitron.models.base_model import ImplicitronModelBase
 from pytorch3d.implicitron.tools import model_io
 from pytorch3d.implicitron.tools.config import (
@@ -121,7 +123,6 @@ class ImplicitronOptimizerFactory(OptimizerFactoryBase):
         """
         # Get the parameters to optimize
         if hasattr(model, "_get_param_groups"):  # use the model function
-            # pyre-fixme[29]: `Union[Tensor, Module]` is not a function.
             p_groups = model._get_param_groups(self.lr, wd=self.weight_decay)
         else:
             p_groups = [
@@ -172,7 +173,6 @@ class ImplicitronOptimizerFactory(OptimizerFactoryBase):
             scheduler = torch.optim.lr_scheduler.LambdaLR(
                 optimizer,
                 lambda epoch: self.gamma ** (epoch / self.exponential_lr_step_size),
-                # pyrefly: ignore [unexpected-keyword]
                 verbose=False,
             )
         elif self.lr_policy.casefold() == "LinearExponential".casefold():
@@ -190,10 +190,7 @@ class ImplicitronOptimizerFactory(OptimizerFactoryBase):
                 return gamma
 
             scheduler = torch.optim.lr_scheduler.LambdaLR(
-                optimizer,
-                _get_lr,
-                # pyrefly: ignore [unexpected-keyword]
-                verbose=False,
+                optimizer, _get_lr, verbose=False
             )
         else:
             raise ValueError("no such lr policy %s" % self.lr_policy)
@@ -244,7 +241,7 @@ class ImplicitronOptimizerFactory(OptimizerFactoryBase):
                     map_location = {
                         "cuda:%d" % 0: "cuda:%d" % accelerator.local_process_index
                     }
-                optimizer_state = torch.load(opt_path, map_location, weights_only=True)
+                optimizer_state = torch.load(opt_path, map_location)
             else:
                 raise FileNotFoundError(f"Optimizer state {opt_path} does not exist.")
         return optimizer_state

projects/implicitron_trainer/impl/training_loop.py

@@ -36,6 +36,7 @@ class TrainingLoopBase(ReplaceableBase):
         evaluator: An EvaluatorBase instance, used to evaluate training results.
     """
 
+    # pyre-fixme[13]: Attribute `evaluator` is never initialized.
     evaluator: Optional[EvaluatorBase]
     evaluator_class_type: Optional[str] = "ImplicitronEvaluator"
 
@@ -160,6 +161,7 @@ class ImplicitronTrainingLoop(TrainingLoopBase):
         for epoch in range(start_epoch, self.max_epochs):
             # automatic new_epoch and plotting of stats at every epoch start
             with stats:
+
                 # Make sure to re-seed random generators to ensure reproducibility
                 # even after restart.
                 seed_all_random_engines(seed + epoch)
@@ -198,7 +200,6 @@ class ImplicitronTrainingLoop(TrainingLoopBase):
                     and self.test_interval > 0
                     and epoch % self.test_interval == 0
                 ):
-                    # pyrefly: ignore [missing-attribute]
                     self.evaluator.run(
                         device=device,
                         dataloader=test_loader,
@@ -215,7 +216,6 @@ class ImplicitronTrainingLoop(TrainingLoopBase):
 
         if self.test_when_finished:
             if test_loader is not None:
-                # pyrefly: ignore [missing-attribute]
                 self.evaluator.run(
                     device=device,
                     dump_to_json=True,
@@ -379,6 +379,7 @@ class ImplicitronTrainingLoop(TrainingLoopBase):
 
             # update the stats logger
             stats.update(preds, time_start=t_start, stat_set=trainmode)
+            # pyre-ignore [16]
             assert stats.it[trainmode] == it, "inconsistent stat iteration number!"
 
             # print textual status update
@@ -394,7 +395,6 @@ class ImplicitronTrainingLoop(TrainingLoopBase):
             ):
                 prefix = f"e{stats.epoch}_it{stats.it[trainmode]}"
                 if hasattr(model, "visualize"):
-                    # pyre-fixme[29]: `Union[Tensor, Module]` is not a function.
                     model.visualize(
                         viz,
                         visdom_env_imgs,

projects/implicitron_trainer/tests/experiment.yaml

@@ -13,6 +13,13 @@ hydra:
 data_source_ImplicitronDataSource_args:
   dataset_map_provider_class_type: ???
   data_loader_map_provider_class_type: SequenceDataLoaderMapProvider
+  dataset_map_provider_BlenderDatasetMapProvider_args:
+    base_dir: ???
+    object_name: ???
+    path_manager_factory_class_type: PathManagerFactory
+    n_known_frames_for_test: null
+    path_manager_factory_PathManagerFactory_args:
+      silence_logs: true
   dataset_map_provider_JsonIndexDatasetMapProvider_args:
     category: ???
     task_str: singlesequence
@@ -84,6 +91,14 @@ data_source_ImplicitronDataSource_args:
       sort_frames: false
     path_manager_factory_PathManagerFactory_args:
       silence_logs: true
+  dataset_map_provider_LlffDatasetMapProvider_args:
+    base_dir: ???
+    object_name: ???
+    path_manager_factory_class_type: PathManagerFactory
+    n_known_frames_for_test: null
+    path_manager_factory_PathManagerFactory_args:
+      silence_logs: true
+    downscale_factor: 4
   dataset_map_provider_RenderedMeshDatasetMapProvider_args:
     num_views: 40
     data_file: null

projects/implicitron_trainer/tests/test_experiment.py

@@ -12,6 +12,7 @@ import unittest
 from pathlib import Path
 
 import torch
+
 from hydra import compose, initialize_config_dir
 from omegaconf import OmegaConf
 from projects.implicitron_trainer.impl.optimizer_factory import (
@@ -52,8 +53,12 @@ class TestExperiment(unittest.TestCase):
         cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_class_type = (
             "JsonIndexDatasetMapProvider"
         )
-        dataset_args = cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
-        dataloader_args = cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_args
+        dataset_args = (
+            cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
+        )
+        dataloader_args = (
+            cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_args
+        )
         dataset_args.category = "skateboard"
         dataset_args.test_restrict_sequence_id = 0
         dataset_args.dataset_root = "manifold://co3d/tree/extracted"
@@ -89,8 +94,12 @@ class TestExperiment(unittest.TestCase):
         cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_class_type = (
             "JsonIndexDatasetMapProvider"
         )
-        dataset_args = cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
-        dataloader_args = cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_args
+        dataset_args = (
+            cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
+        )
+        dataloader_args = (
+            cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_args
+        )
         dataset_args.category = "skateboard"
         dataset_args.test_restrict_sequence_id = 0
         dataset_args.dataset_root = "manifold://co3d/tree/extracted"
@@ -102,7 +111,9 @@ class TestExperiment(unittest.TestCase):
         cfg.training_loop_ImplicitronTrainingLoop_args.max_epochs = 2
         cfg.training_loop_ImplicitronTrainingLoop_args.store_checkpoints = False
         cfg.optimizer_factory_ImplicitronOptimizerFactory_args.lr_policy = "Exponential"
-        cfg.optimizer_factory_ImplicitronOptimizerFactory_args.exponential_lr_step_size = 2
+        cfg.optimizer_factory_ImplicitronOptimizerFactory_args.exponential_lr_step_size = (
+            2
+        )
 
         if DEBUG:
             experiment.dump_cfg(cfg)

projects/implicitron_trainer/tests/test_optimizer_factory.py

@@ -81,9 +81,8 @@ class TestOptimizerFactory(unittest.TestCase):
 
     def test_param_overrides_self_param_group_assignment(self):
         pa, pb, pc = [torch.nn.Parameter(data=torch.tensor(i * 1.0)) for i in range(3)]
-        na, nb = (
-            Node(params=[pa]),
-            Node(params=[pb], param_groups={"self": "pb_self", "p1": "pb_param"}),
+        na, nb = Node(params=[pa]), Node(
+            params=[pb], param_groups={"self": "pb_self", "p1": "pb_param"}
         )
         root = Node(children=[na, nb], params=[pc], param_groups={"m1": "pb_member"})
         param_groups = self._get_param_groups(root)

projects/implicitron_trainer/visualize_reconstruction.py

@@ -135,7 +135,6 @@ def visualize_reconstruction(
                 "device": device,
                 **render_flyaround_kwargs,
             }
-            # pyrefly: ignore [bad-argument-type]
             render_flyaround(**render_kwargs)
 
 

projects/nerf/nerf/dataset.py

@@ -84,9 +84,9 @@ def get_nerf_datasets(
 
     if autodownload and any(not os.path.isfile(p) for p in (cameras_path, image_path)):
         # Automatically download the data files if missing.
-        download_data([dataset_name], data_root=data_root)
+        download_data((dataset_name,), data_root=data_root)
 
-    train_data = torch.load(cameras_path, weights_only=True)
+    train_data = torch.load(cameras_path)
     n_cameras = train_data["cameras"]["R"].shape[0]
 
     _image_max_image_pixels = Image.MAX_IMAGE_PIXELS

projects/nerf/nerf/stats.py

@@ -194,6 +194,7 @@ class Stats:
         it = self.it[stat_set]
 
         for stat in self.log_vars:
+
             if stat not in self.stats[stat_set]:
                 self.stats[stat_set][stat] = AverageMeter()
 

projects/nerf/test_nerf.py

@@ -24,6 +24,7 @@ CONFIG_DIR = os.path.join(os.path.dirname(os.path.realpath(__file__)), "configs"
 
 @hydra.main(config_path=CONFIG_DIR, config_name="lego")
 def main(cfg: DictConfig):
+
     # Device on which to run.
     if torch.cuda.is_available():
         device = "cuda"
@@ -62,7 +63,7 @@ def main(cfg: DictConfig):
         raise ValueError(f"Model checkpoint {checkpoint_path} does not exist!")
 
     print(f"Loading checkpoint {checkpoint_path}.")
-    loaded_data = torch.load(checkpoint_path, weights_only=True)
+    loaded_data = torch.load(checkpoint_path)
     # Do not load the cached xy grid.
     # - this allows setting an arbitrary evaluation image size.
     state_dict = {

projects/nerf/tests/test_raysampler.py

@@ -42,6 +42,7 @@ class TestRaysampler(unittest.TestCase):
         cameras, rays = [], []
 
         for _ in range(batch_size):
+
             R = random_rotations(1)
             T = torch.randn(1, 3)
             focal_length = torch.rand(1, 2) + 0.5

projects/nerf/train_nerf.py

@@ -25,6 +25,7 @@ CONFIG_DIR = os.path.join(os.path.dirname(os.path.realpath(__file__)), "configs"
 
 @hydra.main(config_path=CONFIG_DIR, config_name="lego")
 def main(cfg: DictConfig):
+
     # Set the relevant seeds for reproducibility.
     np.random.seed(cfg.seed)
     torch.manual_seed(cfg.seed)
@@ -76,7 +77,7 @@ def main(cfg: DictConfig):
         # Resume training if requested.
         if cfg.resume and os.path.isfile(checkpoint_path):
             print(f"Resuming from checkpoint {checkpoint_path}.")
-            loaded_data = torch.load(checkpoint_path, weights_only=True)
+            loaded_data = torch.load(checkpoint_path)
             model.load_state_dict(loaded_data["model"])
             stats = pickle.loads(loaded_data["stats"])
             print(f"   => resuming from epoch {stats.epoch}.")
@@ -218,6 +219,7 @@ def main(cfg: DictConfig):
 
         # Validation
         if epoch % cfg.validation_epoch_interval == 0 and epoch > 0:
+
             # Sample a validation camera/image.
             val_batch = next(val_dataloader.__iter__())
             val_image, val_camera, camera_idx = val_batch[0].values()

pytorch3d/__init__.py

@@ -6,4 +6,4 @@
 
 # pyre-unsafe
 
-__version__ = "0.7.9"
+__version__ = "0.7.8"

pytorch3d/common/compat.py

@@ -17,12 +17,13 @@ Some functions which depend on PyTorch or Python versions.
 
 
 def meshgrid_ij(
-    *A: Union[torch.Tensor, Sequence[torch.Tensor]],
+    *A: Union[torch.Tensor, Sequence[torch.Tensor]]
 ) -> Tuple[torch.Tensor, ...]:  # pragma: no cover
     """
     Like torch.meshgrid was before PyTorch 1.10.0, i.e. with indexing set to ij
     """
     if (
+        # pyre-fixme[16]: Callable `meshgrid` has no attribute `__kwdefaults__`.
         torch.meshgrid.__kwdefaults__ is not None
         and "indexing" in torch.meshgrid.__kwdefaults__
     ):

pytorch3d/common/workaround/symeig3x3.py

@@ -77,34 +77,17 @@ class _SymEig3x3(nn.Module):
         if inputs.shape[-2:] != (3, 3):
             raise ValueError("Only inputs of shape (..., 3, 3) are supported.")
 
-        # The thresholds and regularizers used below are absolute, so they are
-        # only meaningful for inputs of order 1: rescale each matrix to unit
-        # maximum magnitude and undo it on the eigenvalues. The clamp only
-        # matters for the all-zero matrix. The scale is detached because the
-        # decomposition is homogeneous of degree 1, so its Jacobian is
-        # homogeneous of degree 0 and holding the scale constant is exact.
-        scale = (
-            inputs.detach()
-            .abs()
-            .amax(dim=(-2, -1), keepdim=True)
-            .clamp(min=torch.finfo(inputs.dtype).tiny)
-        )
-        inputs = inputs / scale
-
         inputs_diag = inputs.diagonal(dim1=-2, dim2=-1)
         inputs_trace = inputs_diag.sum(-1)
         q = inputs_trace / 3.0
 
         # Calculate squared sum of elements outside the main diagonal / 2
-        p1 = (
-            torch.square(inputs).sum(dim=(-1, -2)) - torch.square(inputs_diag).sum(-1)
-        ) / 2
-        p2 = torch.square(inputs_diag - q[..., None]).sum(dim=-1) + 2.0 * p1.clamp(
-            self._eps
-        )
+        # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and `int`.
+        p1 = ((inputs**2).sum(dim=(-1, -2)) - (inputs_diag**2).sum(-1)) / 2
+        # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and `int`.
+        p2 = ((inputs_diag - q[..., None]) ** 2).sum(dim=-1) + 2.0 * p1.clamp(self._eps)
 
         p = torch.sqrt(p2 / 6.0)
-        # pyrefly: ignore [unsupported-operation]
         B = (inputs - q[..., None, None] * self._identity) / p[..., None, None]
 
         r = torch.det(B) / 2.0
@@ -121,9 +104,7 @@ class _SymEig3x3(nn.Module):
         # Soft dispatch between the degenerate case (diagonal A) and general.
         # diag_soft_cond -> 1.0 when p1 < 6 * eps and diag_soft_cond -> 0.0 otherwise.
         # We use 6 * eps to take into account the error accumulated during the p1 summation
-        diag_soft_cond = torch.exp(-torch.square(p1 / (6 * self._eps))).detach()[
-            ..., None
-        ]
+        diag_soft_cond = torch.exp(-((p1 / (6 * self._eps)) ** 2)).detach()[..., None]
 
         # Eigenvalues are the ordered elements of main diagonal in the degenerate case
         diag_eigenvals, _ = torch.sort(inputs_diag, dim=-1)
@@ -134,7 +115,7 @@ class _SymEig3x3(nn.Module):
         else:
             eigenvecs = None
 
-        return eigenvals * scale.squeeze(-1), eigenvecs
+        return eigenvals, eigenvecs
 
     def _construct_eigenvecs_set(
         self, inputs: torch.Tensor, eigenvals: torch.Tensor
@@ -189,10 +170,8 @@ class _SymEig3x3(nn.Module):
         """
 
         # Find the eigenvector corresponding to alpha0, its eigenvalue is distinct
-        # pyrefly: ignore [unsupported-operation]
         ev0 = self._get_ev0(inputs - alpha0[..., None, None] * self._identity)
         u, v = self._get_uv(ev0)
-        # pyrefly: ignore [unsupported-operation]
         ev1 = self._get_ev1(inputs - alpha1[..., None, None] * self._identity, u, v)
         # Third eigenvector is computed as the cross-product of the other two
         ev2 = torch.cross(ev0, ev1, dim=-1)
@@ -220,7 +199,8 @@ class _SymEig3x3(nn.Module):
             cross_products[..., :1, :]
         )
 
-        norms_sq = torch.square(cross_products).sum(dim=-1)
+        # pyre-fixme[58]: `**` is not supported for operand types `Tensor` and `int`.
+        norms_sq = (cross_products**2).sum(dim=-1)
         max_norms_index = norms_sq.argmax(dim=-1)
 
         # Pick only the cross-product with highest squared norm for each input
@@ -267,7 +247,6 @@ class _SymEig3x3(nn.Module):
         """
 
         min_idx = w.abs().argmin(dim=-1)
-        # pyrefly: ignore [bad-index]
         rotation_2d = self._rotations_3d[min_idx].to(w)
 
         u = F.normalize((rotation_2d @ w[..., None])[..., 0], dim=-1)

pytorch3d/csrc/ball_query/ball_query.cu

@@ -32,9 +32,7 @@ __global__ void BallQueryKernel(
     at::PackedTensorAccessor64<int64_t, 3, at::RestrictPtrTraits> idxs,
     at::PackedTensorAccessor64<scalar_t, 3, at::RestrictPtrTraits> dists,
     const int64_t K,
-    const float radius,
-    const float radius2,
-    const bool skip_points_outside_cube) {
+    const float radius2) {
   const int64_t N = p1.size(0);
   const int64_t chunks_per_cloud = (1 + (p1.size(1) - 1) / blockDim.x);
   const int64_t chunks_to_do = N * chunks_per_cloud;
@@ -53,19 +51,7 @@ __global__ void BallQueryKernel(
     // Iterate over points in p2 until desired count is reached or
     // all points have been considered
     for (int64_t j = 0, count = 0; j < lengths2[n] && count < K; ++j) {
-      if (skip_points_outside_cube) {
-        bool is_within_radius = true;
-        // Filter when any one coordinate is already outside the radius
-        for (int d = 0; is_within_radius && d < D; ++d) {
-          scalar_t abs_diff = fabs(p1[n][i][d] - p2[n][j][d]);
-          is_within_radius = (abs_diff <= radius);
-        }
-        if (!is_within_radius) {
-          continue;
-        }
-      }
-
-      // Else, calculate the distance between the points and compare
+      // Calculate the distance between the points
       scalar_t dist2 = 0.0;
       for (int d = 0; d < D; ++d) {
         scalar_t diff = p1[n][i][d] - p2[n][j][d];
@@ -91,8 +77,7 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
     const at::Tensor& lengths1, // (N,)
     const at::Tensor& lengths2, // (N,)
     int K,
-    float radius,
-    bool skip_points_outside_cube) {
+    float radius) {
   // Check inputs are on the same device
   at::TensorArg p1_t{p1, "p1", 1}, p2_t{p2, "p2", 2},
       lengths1_t{lengths1, "lengths1", 3}, lengths2_t{lengths2, "lengths2", 4};
@@ -135,9 +120,7 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
             idxs.packed_accessor64<int64_t, 3, at::RestrictPtrTraits>(),
             dists.packed_accessor64<float, 3, at::RestrictPtrTraits>(),
             K_64,
-            radius,
-            radius2,
-            skip_points_outside_cube);
+            radius2);
       }));
 
   AT_CUDA_CHECK(cudaGetLastError());

pytorch3d/csrc/ball_query/ball_query.h

@@ -25,9 +25,6 @@
 //      within the radius
 //    radius: the radius around each point within which the neighbors need to be
 //      located
-//    skip_points_outside_cube: If true, reduce multiplications of float values
-//      by not explicitly calculating distances to points that fall outside the
-//      D-cube with side length (2*radius) centered at each point in p1.
 //
 // Returns:
 //    p1_neighbor_idx: LongTensor of shape (N, P1, K), where
@@ -49,8 +46,7 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCpu(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     const int K,
-    const float radius,
-    const bool skip_points_outside_cube);
+    const float radius);
 
 // CUDA implementation
 std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
@@ -59,8 +55,7 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     const int K,
-    const float radius,
-    const bool skip_points_outside_cube);
+    const float radius);
 
 // Implementation which is exposed
 // Note: the backward pass reuses the KNearestNeighborBackward kernel
@@ -70,8 +65,7 @@ inline std::tuple<at::Tensor, at::Tensor> BallQuery(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     int K,
-    float radius,
-    bool skip_points_outside_cube) {
+    float radius) {
   if (p1.is_cuda() || p2.is_cuda()) {
 #ifdef WITH_CUDA
     CHECK_CUDA(p1);
@@ -82,20 +76,16 @@ inline std::tuple<at::Tensor, at::Tensor> BallQuery(
         lengths1.contiguous(),
         lengths2.contiguous(),
         K,
-        radius,
-        skip_points_outside_cube);
+        radius);
 #else
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(p1);
-  CHECK_CPU(p2);
   return BallQueryCpu(
       p1.contiguous(),
       p2.contiguous(),
       lengths1.contiguous(),
       lengths2.contiguous(),
       K,
-      radius,
-      skip_points_outside_cube);
+      radius);
 }

pytorch3d/csrc/ball_query/ball_query_cpu.cpp

@@ -6,8 +6,8 @@
  * LICENSE file in the root directory of this source tree.
  */
 
-#include <math.h>
 #include <torch/extension.h>
+#include <queue>
 #include <tuple>
 
 std::tuple<at::Tensor, at::Tensor> BallQueryCpu(
@@ -16,8 +16,7 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCpu(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     int K,
-    float radius,
-    bool skip_points_outside_cube) {
+    float radius) {
   const int N = p1.size(0);
   const int P1 = p1.size(1);
   const int D = p1.size(2);
@@ -39,16 +38,6 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCpu(
     const int64_t length2 = lengths2_a[n];
     for (int64_t i = 0; i < length1; ++i) {
       for (int64_t j = 0, count = 0; j < length2 && count < K; ++j) {
-        if (skip_points_outside_cube) {
-          bool is_within_radius = true;
-          for (int d = 0; is_within_radius && d < D; ++d) {
-            float abs_diff = fabs(p1_a[n][i][d] - p2_a[n][j][d]);
-            is_within_radius = (abs_diff <= radius);
-          }
-          if (!is_within_radius) {
-            continue;
-          }
-        }
         float dist2 = 0;
         for (int d = 0; d < D; ++d) {
           float diff = p1_a[n][i][d] - p2_a[n][j][d];

pytorch3d/csrc/blending/sigmoid_alpha_blend.h

@@ -98,11 +98,6 @@ at::Tensor SigmoidAlphaBlendBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(distances);
-  CHECK_CPU(pix_to_face);
-  CHECK_CPU(alphas);
-  CHECK_CPU(grad_alphas);
-
   return SigmoidAlphaBlendBackwardCpu(
       grad_alphas, alphas, distances, pix_to_face, sigma);
 }

pytorch3d/csrc/compositing/alpha_composite.cu

@@ -28,16 +28,17 @@ __global__ void alphaCompositeCudaForwardKernel(
     const at::PackedTensorAccessor64<float, 4, at::RestrictPtrTraits> alphas,
     const at::PackedTensorAccessor64<int64_t, 4, at::RestrictPtrTraits> points_idx) {
   // clang-format on
+  const int64_t batch_size = result.size(0);
   const int64_t C = features.size(0);
   const int64_t H = points_idx.size(2);
   const int64_t W = points_idx.size(3);
 
   // Get the batch and index
-  const auto batch = blockIdx.x;
+  const int batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const auto num_threads = gridDim.y * blockDim.x;
-  const auto tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.y * blockDim.x;
+  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
 
   // Iterate over each feature in each pixel
   for (int pid = tid; pid < num_pixels; pid += num_threads) {
@@ -78,16 +79,17 @@ __global__ void alphaCompositeCudaBackwardKernel(
     const at::PackedTensorAccessor64<float, 4, at::RestrictPtrTraits> alphas,
     const at::PackedTensorAccessor64<int64_t, 4, at::RestrictPtrTraits> points_idx) {
   // clang-format on
+  const int64_t batch_size = points_idx.size(0);
   const int64_t C = features.size(0);
   const int64_t H = points_idx.size(2);
   const int64_t W = points_idx.size(3);
 
   // Get the batch and index
-  const auto batch = blockIdx.x;
+  const int batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const auto num_threads = gridDim.y * blockDim.x;
-  const auto tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.y * blockDim.x;
+  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
 
   // Parallelize over each feature in each pixel in images of size H * W,
   // for each image in the batch of size batch_size

pytorch3d/csrc/compositing/alpha_composite.h

@@ -74,9 +74,6 @@ torch::Tensor alphaCompositeForward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(features);
-    CHECK_CPU(alphas);
-    CHECK_CPU(points_idx);
     return alphaCompositeCpuForward(features, alphas, points_idx);
   }
 }
@@ -104,11 +101,6 @@ std::tuple<torch::Tensor, torch::Tensor> alphaCompositeBackward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(grad_outputs);
-    CHECK_CPU(features);
-    CHECK_CPU(alphas);
-    CHECK_CPU(points_idx);
-
     return alphaCompositeCpuBackward(
         grad_outputs, features, alphas, points_idx);
   }

pytorch3d/csrc/compositing/norm_weighted_sum.cu

@@ -28,16 +28,17 @@ __global__ void weightedSumNormCudaForwardKernel(
     const at::PackedTensorAccessor64<float, 4, at::RestrictPtrTraits> alphas,
     const at::PackedTensorAccessor64<int64_t, 4, at::RestrictPtrTraits> points_idx) {
   // clang-format on
+  const int64_t batch_size = result.size(0);
   const int64_t C = features.size(0);
   const int64_t H = points_idx.size(2);
   const int64_t W = points_idx.size(3);
 
   // Get the batch and index
-  const auto batch = blockIdx.x;
+  const int batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const auto num_threads = gridDim.y * blockDim.x;
-  const auto tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.y * blockDim.x;
+  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
 
   // Parallelize over each feature in each pixel in images of size H * W,
   // for each image in the batch of size batch_size
@@ -91,16 +92,17 @@ __global__ void weightedSumNormCudaBackwardKernel(
     const at::PackedTensorAccessor64<float, 4, at::RestrictPtrTraits> alphas,
     const at::PackedTensorAccessor64<int64_t, 4, at::RestrictPtrTraits> points_idx) {
   // clang-format on
+  const int64_t batch_size = points_idx.size(0);
   const int64_t C = features.size(0);
   const int64_t H = points_idx.size(2);
   const int64_t W = points_idx.size(3);
 
   // Get the batch and index
-  const auto batch = blockIdx.x;
+  const int batch = blockIdx.x;
 
   const int num_pixels = C * W * H;
-  const auto num_threads = gridDim.y * blockDim.x;
-  const auto tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.y * blockDim.x;
+  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
 
   // Parallelize over each feature in each pixel in images of size H * W,
   // for each image in the batch of size batch_size

pytorch3d/csrc/compositing/norm_weighted_sum.h

@@ -73,10 +73,6 @@ torch::Tensor weightedSumNormForward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(features);
-    CHECK_CPU(alphas);
-    CHECK_CPU(points_idx);
-
     return weightedSumNormCpuForward(features, alphas, points_idx);
   }
 }
@@ -104,11 +100,6 @@ std::tuple<torch::Tensor, torch::Tensor> weightedSumNormBackward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(grad_outputs);
-    CHECK_CPU(features);
-    CHECK_CPU(alphas);
-    CHECK_CPU(points_idx);
-
     return weightedSumNormCpuBackward(
         grad_outputs, features, alphas, points_idx);
   }

pytorch3d/csrc/compositing/weighted_sum.cu

@@ -26,16 +26,17 @@ __global__ void weightedSumCudaForwardKernel(
     const at::PackedTensorAccessor64<float, 4, at::RestrictPtrTraits> alphas,
     const at::PackedTensorAccessor64<int64_t, 4, at::RestrictPtrTraits> points_idx) {
   // clang-format on
+  const int64_t batch_size = result.size(0);
   const int64_t C = features.size(0);
   const int64_t H = points_idx.size(2);
   const int64_t W = points_idx.size(3);
 
   // Get the batch and index
-  const auto batch = blockIdx.x;
+  const int batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const auto num_threads = gridDim.y * blockDim.x;
-  const auto tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.y * blockDim.x;
+  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
 
   // Parallelize over each feature in each pixel in images of size H * W,
   // for each image in the batch of size batch_size
@@ -73,16 +74,17 @@ __global__ void weightedSumCudaBackwardKernel(
     const at::PackedTensorAccessor64<float, 4, at::RestrictPtrTraits> alphas,
     const at::PackedTensorAccessor64<int64_t, 4, at::RestrictPtrTraits> points_idx) {
   // clang-format on
+  const int64_t batch_size = points_idx.size(0);
   const int64_t C = features.size(0);
   const int64_t H = points_idx.size(2);
   const int64_t W = points_idx.size(3);
 
   // Get the batch and index
-  const auto batch = blockIdx.x;
+  const int batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const auto num_threads = gridDim.y * blockDim.x;
-  const auto tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.y * blockDim.x;
+  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
 
   // Iterate over each pixel to compute the contribution to the
   // gradient for the features and weights

pytorch3d/csrc/compositing/weighted_sum.h

@@ -72,9 +72,6 @@ torch::Tensor weightedSumForward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(features);
-    CHECK_CPU(alphas);
-    CHECK_CPU(points_idx);
     return weightedSumCpuForward(features, alphas, points_idx);
   }
 }
@@ -101,11 +98,6 @@ std::tuple<torch::Tensor, torch::Tensor> weightedSumBackward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(grad_outputs);
-    CHECK_CPU(features);
-    CHECK_CPU(alphas);
-    CHECK_CPU(points_idx);
-
     return weightedSumCpuBackward(grad_outputs, features, alphas, points_idx);
   }
 }

pytorch3d/csrc/ext.cpp

@@ -8,6 +8,7 @@
 
 // clang-format off
 #include "./pulsar/global.h" // Include before <torch/extension.h>.
+#include <torch/extension.h>
 // clang-format on
 #include "./pulsar/pytorch/renderer.h"
 #include "./pulsar/pytorch/tensor_util.h"
@@ -98,22 +99,22 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   m.def("marching_cubes", &MarchingCubes);
 
   // Pulsar.
+  // Pulsar not enabled on AMD.
 #ifdef PULSAR_LOGGING_ENABLED
   c10::ShowLogInfoToStderr();
 #endif
   py::class_<
       pulsar::pytorch::Renderer,
       std::shared_ptr<pulsar::pytorch::Renderer>>(m, "PulsarRenderer")
-      .def(
-          py::init<
-              const uint&,
-              const uint&,
-              const uint&,
-              const bool&,
-              const bool&,
-              const float&,
-              const uint&,
-              const uint&>())
+      .def(py::init<
+           const uint&,
+           const uint&,
+           const uint&,
+           const bool&,
+           const bool&,
+           const float&,
+           const uint&,
+           const uint&>())
       .def(
           "__eq__",
           [](const pulsar::pytorch::Renderer& a,
@@ -148,10 +149,10 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
           py::arg("gamma"),
           py::arg("max_depth"),
           py::arg("min_depth") /* = 0.f*/,
-          py::arg("bg_col") /* = std::nullopt not exposed properly in
-                               pytorch 1.1. */
+          py::arg(
+              "bg_col") /* = at::nullopt not exposed properly in pytorch 1.1. */
           ,
-          py::arg("opacity") /* = std::nullopt ... */,
+          py::arg("opacity") /* = at::nullopt ... */,
           py::arg("percent_allowed_difference") = 0.01f,
           py::arg("max_n_hits") = MAX_UINT,
           py::arg("mode") = 0)

pytorch3d/csrc/face_areas_normals/face_areas_normals.h

@@ -60,8 +60,6 @@ std::tuple<at::Tensor, at::Tensor> FaceAreasNormalsForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(verts);
-  CHECK_CPU(faces);
   return FaceAreasNormalsForwardCpu(verts, faces);
 }
 
@@ -82,9 +80,5 @@ at::Tensor FaceAreasNormalsBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(grad_areas);
-  CHECK_CPU(grad_normals);
-  CHECK_CPU(verts);
-  CHECK_CPU(faces);
   return FaceAreasNormalsBackwardCpu(grad_areas, grad_normals, verts, faces);
 }

pytorch3d/csrc/gather_scatter/gather_scatter.cu

@@ -20,14 +20,14 @@ __global__ void GatherScatterCudaKernel(
     const size_t V,
     const size_t D,
     const size_t E) {
-  const auto tid = threadIdx.x;
+  const int tid = threadIdx.x;
 
   // Reverse the vertex order if backward.
   const int v0_idx = backward ? 1 : 0;
   const int v1_idx = backward ? 0 : 1;
 
   // Edges are split evenly across the blocks.
-  for (auto e = blockIdx.x; e < E; e += gridDim.x) {
+  for (int e = blockIdx.x; e < E; e += gridDim.x) {
     // Get indices of vertices which form the edge.
     const int64_t v0 = edges[2 * e + v0_idx];
     const int64_t v1 = edges[2 * e + v1_idx];
@@ -35,7 +35,7 @@ __global__ void GatherScatterCudaKernel(
     // Split vertex features evenly across threads.
     // This implementation will be quite wasteful when D<128 since there will be
     // a lot of threads doing nothing.
-    for (auto d = tid; d < D; d += blockDim.x) {
+    for (int d = tid; d < D; d += blockDim.x) {
       const float val = input[v1 * D + d];
       float* address = output + v0 * D + d;
       atomicAdd(address, val);

pytorch3d/csrc/gather_scatter/gather_scatter.h

@@ -53,7 +53,5 @@ at::Tensor GatherScatter(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(input);
-  CHECK_CPU(edges);
   return GatherScatterCpu(input, edges, directed, backward);
 }

pytorch3d/csrc/interp_face_attrs/interp_face_attrs.cu

@@ -20,8 +20,8 @@ __global__ void InterpFaceAttrsForwardKernel(
     const size_t P,
     const size_t F,
     const size_t D) {
-  const auto tid = threadIdx.x + blockIdx.x * blockDim.x;
-  const auto num_threads = blockDim.x * gridDim.x;
+  const int tid = threadIdx.x + blockIdx.x * blockDim.x;
+  const int num_threads = blockDim.x * gridDim.x;
   for (int pd = tid; pd < P * D; pd += num_threads) {
     const int p = pd / D;
     const int d = pd % D;
@@ -93,8 +93,8 @@ __global__ void InterpFaceAttrsBackwardKernel(
     const size_t P,
     const size_t F,
     const size_t D) {
-  const auto tid = threadIdx.x + blockIdx.x * blockDim.x;
-  const auto num_threads = blockDim.x * gridDim.x;
+  const int tid = threadIdx.x + blockIdx.x * blockDim.x;
+  const int num_threads = blockDim.x * gridDim.x;
   for (int pd = tid; pd < P * D; pd += num_threads) {
     const int p = pd / D;
     const int d = pd % D;

pytorch3d/csrc/interp_face_attrs/interp_face_attrs.h

@@ -57,8 +57,6 @@ at::Tensor InterpFaceAttrsForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(face_attrs);
-  CHECK_CPU(barycentric_coords);
   return InterpFaceAttrsForwardCpu(pix_to_face, barycentric_coords, face_attrs);
 }
 
@@ -108,9 +106,6 @@ std::tuple<at::Tensor, at::Tensor> InterpFaceAttrsBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(face_attrs);
-  CHECK_CPU(barycentric_coords);
-  CHECK_CPU(grad_pix_attrs);
   return InterpFaceAttrsBackwardCpu(
       pix_to_face, barycentric_coords, face_attrs, grad_pix_attrs);
 }

pytorch3d/csrc/iou_box3d/iou_box3d.h

@@ -44,7 +44,5 @@ inline std::tuple<at::Tensor, at::Tensor> IoUBox3D(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(boxes1);
-  CHECK_CPU(boxes2);
   return IoUBox3DCpu(boxes1.contiguous(), boxes2.contiguous());
 }

pytorch3d/csrc/iou_box3d/iou_box3d_cpu.cpp

@@ -7,6 +7,10 @@
  */
 
 #include <torch/extension.h>
+#include <torch/torch.h>
+#include <list>
+#include <numeric>
+#include <queue>
 #include <tuple>
 #include "iou_box3d/iou_utils.h"
 

pytorch3d/csrc/iou_box3d/iou_utils.cuh

@@ -461,8 +461,10 @@ __device__ inline std::tuple<float3, float3> ArgMaxVerts(
 __device__ inline bool IsCoplanarTriTri(
     const FaceVerts& tri1,
     const FaceVerts& tri2) {
+  const float3 tri1_ctr = FaceCenter({tri1.v0, tri1.v1, tri1.v2});
   const float3 tri1_n = FaceNormal({tri1.v0, tri1.v1, tri1.v2});
 
+  const float3 tri2_ctr = FaceCenter({tri2.v0, tri2.v1, tri2.v2});
   const float3 tri2_n = FaceNormal({tri2.v0, tri2.v1, tri2.v2});
 
   // Check if parallel
@@ -498,6 +500,7 @@ __device__ inline bool IsCoplanarTriPlane(
     const FaceVerts& tri,
     const FaceVerts& plane,
     const float3& normal) {
+  const float3 tri_ctr = FaceCenter({tri.v0, tri.v1, tri.v2});
   const float3 nt = FaceNormal({tri.v0, tri.v1, tri.v2});
 
   // check if parallel
@@ -725,7 +728,7 @@ __device__ inline int BoxIntersections(
       }
     }
     // Update the face_verts_out tris
-    num_tris = min(MAX_TRIS, offset);
+    num_tris = offset;
     for (int j = 0; j < num_tris; ++j) {
       face_verts_out[j] = tri_verts_updated[j];
     }

pytorch3d/csrc/knn/knn.h

@@ -74,8 +74,6 @@ std::tuple<at::Tensor, at::Tensor> KNearestNeighborIdx(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(p1);
-  CHECK_CPU(p2);
   return KNearestNeighborIdxCpu(p1, p2, lengths1, lengths2, norm, K);
 }
 
@@ -142,8 +140,6 @@ std::tuple<at::Tensor, at::Tensor> KNearestNeighborBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(p1);
-  CHECK_CPU(p2);
   return KNearestNeighborBackwardCpu(
       p1, p2, lengths1, lengths2, idxs, norm, grad_dists);
 }

pytorch3d/csrc/marching_cubes/marching_cubes.h

@@ -58,6 +58,5 @@ inline std::tuple<at::Tensor, at::Tensor, at::Tensor> MarchingCubes(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(vol);
   return MarchingCubesCpu(vol.contiguous(), isolevel);
 }

pytorch3d/csrc/marching_cubes/marching_cubes_cpu.cpp

@@ -34,18 +34,6 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> MarchingCubesCpu(
   const int H = vol.size(1);
   const int W = vol.size(2);
 
-  // The edge-id hash packs two grid-point ids as v1_id * (W*H*D) + v2_id, whose
-  // maximum value is (W*H*D)^2 - 1. Guard against signed int64 overflow so an
-  // oversized volume fails loudly instead of silently producing colliding edge
-  // ids and a corrupted mesh. floor(sqrt(INT64_MAX)) == 3037000499 (~1448^3).
-  const int64_t num_grid_points = (int64_t)W * H * D;
-  TORCH_CHECK(
-      num_grid_points <= 3037000499LL,
-      "Volume too large for CPU marching cubes: W*H*D (",
-      num_grid_points,
-      ") exceeds 3037000499 (~1448^3), which would overflow the int64 edge-id "
-      "hash and corrupt the mesh.");
-
   // Create tensor accessors
   auto vol_a = vol.accessor<float, 3>();
   // edge_id_to_v maps from an edge id to a vertex position
@@ -78,23 +66,19 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> MarchingCubesCpu(
           tri.push_back(edge);
           ps.push_back(interp_points[e]);
 
-          if (ps.size() == 3) {
-            // Check if the triangle face is degenerate. A triangle face
-            // is degenerate if any of the two vertices share the same 3D
-            // position
-            if (ps[0] != ps[1] && ps[1] != ps[2] && ps[2] != ps[0]) {
-              for (int k = 0; k < 3; k++) {
-                int64_t v = tri.at(k);
-                edge_id_to_v[v] = ps.at(k);
-                if (!uniq_edge_id.count(v)) {
-                  uniq_edge_id[v] = verts.size();
-                  verts.push_back(edge_id_to_v[v]);
-                }
-                faces.push_back(uniq_edge_id[v]);
+          // Check if the triangle face is degenerate. A triangle face
+          // is degenerate if any of the two verices share the same 3D position
+          if ((j + 1) % 3 == 0 && ps[0] != ps[1] && ps[1] != ps[2] &&
+              ps[2] != ps[0]) {
+            for (int k = 0; k < 3; k++) {
+              int64_t v = tri.at(k);
+              edge_id_to_v[v] = ps.at(k);
+              if (!uniq_edge_id.count(v)) {
+                uniq_edge_id[v] = verts.size();
+                verts.push_back(edge_id_to_v[v]);
               }
+              faces.push_back(uniq_edge_id[v]);
             }
-            // Clear unconditionally - a rejected degenerate triangle must not
-            // corrupt the buffer for the next triangle in this cube.
             tri.clear();
             ps.clear();
           } // endif

pytorch3d/csrc/marching_cubes/marching_cubes_utils.h

@@ -126,7 +126,8 @@ struct Cube {
   // edge with an integer to address floating point precision issue.
   //
   // Args:
-  //    edge: edge (ID) whose two endpoint vertices are hashed
+  //    v1_id: global id of vertex 1
+  //    v2_id: global id of vertex 2
   //    W: width of the 3d grid
   //    H: height of the 3d grid
   //    D: depth of the 3d grid
@@ -134,19 +135,11 @@ struct Cube {
   // Returns:
   //    hashing for a pair of vertex ids
   //
-  int64_t HashVpair(const int edge, int64_t W, int64_t H, int64_t D) const {
+  int64_t HashVpair(const int edge, int W, int H, int D) {
     const int v1 = _EDGE_TO_VERTICES[edge][0];
     const int v2 = _EDGE_TO_VERTICES[edge][1];
-    // p[v].x/y/z hold integral corner coordinates, so casting to int64_t is
-    // exact. We cast before multiplication to avoid float32 precision loss at
-    // grids > 256^3.
-    const int64_t v1_id =
-        (int64_t)p[v1].x + (int64_t)p[v1].y * W + (int64_t)p[v1].z * W * H;
-    const int64_t v2_id =
-        (int64_t)p[v2].x + (int64_t)p[v2].y * W + (int64_t)p[v2].z * W * H;
-    // v_id max is (W*H*D - 1). A stride of W*H*D perfectly separates v1 and v2.
-    // Note: This hash fits in int64_t safely up to grids of ~1448^3.
-    const int64_t stride = W * H * D;
-    return v1_id * stride + v2_id;
+    const int v1_id = p[v1].x + p[v1].y * W + p[v1].z * W * H;
+    const int v2_id = p[v2].x + p[v2].y * W + p[v2].z * W * H;
+    return (int64_t)v1_id * (W + W * H + W * H * D) + (int64_t)v2_id;
   }
 };

pytorch3d/csrc/packed_to_padded_tensor/packed_to_padded_tensor.h

@@ -88,8 +88,6 @@ at::Tensor PackedToPadded(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(inputs_packed);
-  CHECK_CPU(first_idxs);
   return PackedToPaddedCpu(inputs_packed, first_idxs, max_size);
 }
 
@@ -107,7 +105,5 @@ at::Tensor PaddedToPacked(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(inputs_padded);
-  CHECK_CPU(first_idxs);
   return PaddedToPackedCpu(inputs_padded, first_idxs, num_inputs);
 }

pytorch3d/csrc/point_mesh/point_mesh_cpu.cpp

@@ -174,8 +174,8 @@ std::tuple<at::Tensor, at::Tensor> HullHullDistanceForwardCpu(
   at::Tensor idxs = at::zeros({A_N,}, as_first_idx.options());
   // clang-format on
 
-  auto as_a = as.accessor<float, H1 == 1 ? 2 : 3>();
-  auto bs_a = bs.accessor<float, H2 == 1 ? 2 : 3>();
+  auto as_a = as.accessor < float, H1 == 1 ? 2 : 3 > ();
+  auto bs_a = bs.accessor < float, H2 == 1 ? 2 : 3 > ();
   auto as_first_idx_a = as_first_idx.accessor<int64_t, 1>();
   auto bs_first_idx_a = bs_first_idx.accessor<int64_t, 1>();
   auto dists_a = dists.accessor<float, 1>();
@@ -230,10 +230,10 @@ std::tuple<at::Tensor, at::Tensor> HullHullDistanceBackwardCpu(
   at::Tensor grad_as = at::zeros_like(as);
   at::Tensor grad_bs = at::zeros_like(bs);
 
-  auto as_a = as.accessor<float, H1 == 1 ? 2 : 3>();
-  auto bs_a = bs.accessor<float, H2 == 1 ? 2 : 3>();
-  auto grad_as_a = grad_as.accessor<float, H1 == 1 ? 2 : 3>();
-  auto grad_bs_a = grad_bs.accessor<float, H2 == 1 ? 2 : 3>();
+  auto as_a = as.accessor < float, H1 == 1 ? 2 : 3 > ();
+  auto bs_a = bs.accessor < float, H2 == 1 ? 2 : 3 > ();
+  auto grad_as_a = grad_as.accessor < float, H1 == 1 ? 2 : 3 > ();
+  auto grad_bs_a = grad_bs.accessor < float, H2 == 1 ? 2 : 3 > ();
   auto idx_bs_a = idx_bs.accessor<int64_t, 1>();
   auto grad_dists_a = grad_dists.accessor<float, 1>();
 

pytorch3d/csrc/point_mesh/point_mesh_cuda.cu

@@ -110,7 +110,7 @@ __global__ void DistanceForwardKernel(
     __syncthreads();
 
     // Perform reduction in shared memory.
-    for (auto s = blockDim.x / 2; s > 32; s >>= 1) {
+    for (int s = blockDim.x / 2; s > 32; s >>= 1) {
       if (tid < s) {
         if (min_dists[tid] > min_dists[tid + s]) {
           min_dists[tid] = min_dists[tid + s];
@@ -502,8 +502,8 @@ __global__ void PointFaceArrayForwardKernel(
   const float3* tris_f3 = (float3*)tris;
 
   // Parallelize over P * S computations
-  const auto num_threads = gridDim.x * blockDim.x;
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.x * blockDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
   for (int t_i = tid; t_i < P * T; t_i += num_threads) {
     const int t = t_i / P; // segment index.
@@ -576,8 +576,8 @@ __global__ void PointFaceArrayBackwardKernel(
   const float3* tris_f3 = (float3*)tris;
 
   // Parallelize over P * S computations
-  const auto num_threads = gridDim.x * blockDim.x;
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.x * blockDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
   for (int t_i = tid; t_i < P * T; t_i += num_threads) {
     const int t = t_i / P; // triangle index.
@@ -683,8 +683,8 @@ __global__ void PointEdgeArrayForwardKernel(
   float3* segms_f3 = (float3*)segms;
 
   // Parallelize over P * S computations
-  const auto num_threads = gridDim.x * blockDim.x;
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.x * blockDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
   for (int t_i = tid; t_i < P * S; t_i += num_threads) {
     const int s = t_i / P; // segment index.
@@ -752,8 +752,8 @@ __global__ void PointEdgeArrayBackwardKernel(
   float3* segms_f3 = (float3*)segms;
 
   // Parallelize over P * S computations
-  const auto num_threads = gridDim.x * blockDim.x;
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.x * blockDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
   for (int t_i = tid; t_i < P * S; t_i += num_threads) {
     const int s = t_i / P; // segment index.

pytorch3d/csrc/point_mesh/point_mesh_cuda.h

@@ -88,10 +88,6 @@ std::tuple<torch::Tensor, torch::Tensor> PointFaceDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(points_first_idx);
-  CHECK_CPU(tris);
-  CHECK_CPU(tris_first_idx);
   return PointFaceDistanceForwardCpu(
       points, points_first_idx, tris, tris_first_idx, min_triangle_area);
 }
@@ -147,10 +143,6 @@ std::tuple<torch::Tensor, torch::Tensor> PointFaceDistanceBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(tris);
-  CHECK_CPU(idx_points);
-  CHECK_CPU(grad_dists);
   return PointFaceDistanceBackwardCpu(
       points, tris, idx_points, grad_dists, min_triangle_area);
 }
@@ -229,10 +221,6 @@ std::tuple<torch::Tensor, torch::Tensor> FacePointDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(points_first_idx);
-  CHECK_CPU(tris);
-  CHECK_CPU(tris_first_idx);
   return FacePointDistanceForwardCpu(
       points, points_first_idx, tris, tris_first_idx, min_triangle_area);
 }
@@ -289,10 +277,6 @@ std::tuple<torch::Tensor, torch::Tensor> FacePointDistanceBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(tris);
-  CHECK_CPU(idx_tris);
-  CHECK_CPU(grad_dists);
   return FacePointDistanceBackwardCpu(
       points, tris, idx_tris, grad_dists, min_triangle_area);
 }
@@ -362,10 +346,6 @@ std::tuple<torch::Tensor, torch::Tensor> PointEdgeDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(points_first_idx);
-  CHECK_CPU(segms);
-  CHECK_CPU(segms_first_idx);
   return PointEdgeDistanceForwardCpu(
       points, points_first_idx, segms, segms_first_idx, max_points);
 }
@@ -416,10 +396,6 @@ std::tuple<torch::Tensor, torch::Tensor> PointEdgeDistanceBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(segms);
-  CHECK_CPU(idx_points);
-  CHECK_CPU(grad_dists);
   return PointEdgeDistanceBackwardCpu(points, segms, idx_points, grad_dists);
 }
 
@@ -488,10 +464,6 @@ std::tuple<torch::Tensor, torch::Tensor> EdgePointDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(points_first_idx);
-  CHECK_CPU(segms);
-  CHECK_CPU(segms_first_idx);
   return EdgePointDistanceForwardCpu(
       points, points_first_idx, segms, segms_first_idx, max_segms);
 }
@@ -542,10 +514,6 @@ std::tuple<torch::Tensor, torch::Tensor> EdgePointDistanceBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(segms);
-  CHECK_CPU(idx_segms);
-  CHECK_CPU(grad_dists);
   return EdgePointDistanceBackwardCpu(points, segms, idx_segms, grad_dists);
 }
 
@@ -599,8 +567,6 @@ torch::Tensor PointFaceArrayDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(tris);
   return PointFaceArrayDistanceForwardCpu(points, tris, min_triangle_area);
 }
 
@@ -647,9 +613,6 @@ std::tuple<torch::Tensor, torch::Tensor> PointFaceArrayDistanceBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(tris);
-  CHECK_CPU(grad_dists);
   return PointFaceArrayDistanceBackwardCpu(
       points, tris, grad_dists, min_triangle_area);
 }
@@ -698,8 +661,6 @@ torch::Tensor PointEdgeArrayDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(segms);
   return PointEdgeArrayDistanceForwardCpu(points, segms);
 }
 
@@ -742,8 +703,5 @@ std::tuple<torch::Tensor, torch::Tensor> PointEdgeArrayDistanceBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(segms);
-  CHECK_CPU(grad_dists);
   return PointEdgeArrayDistanceBackwardCpu(points, segms, grad_dists);
 }

pytorch3d/csrc/points_to_volumes/points_to_volumes.h

@@ -104,12 +104,6 @@ inline void PointsToVolumesForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points_3d);
-  CHECK_CPU(points_features);
-  CHECK_CPU(volume_densities);
-  CHECK_CPU(volume_features);
-  CHECK_CPU(grid_sizes);
-  CHECK_CPU(mask);
   PointsToVolumesForwardCpu(
       points_3d,
       points_features,
@@ -189,14 +183,6 @@ inline void PointsToVolumesBackward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points_3d);
-  CHECK_CPU(points_features);
-  CHECK_CPU(grid_sizes);
-  CHECK_CPU(mask);
-  CHECK_CPU(grad_volume_densities);
-  CHECK_CPU(grad_volume_features);
-  CHECK_CPU(grad_points_3d);
-  CHECK_CPU(grad_points_features);
   PointsToVolumesBackwardCpu(
       points_3d,
       points_features,

pytorch3d/csrc/points_to_volumes/points_to_volumes_cpu.cpp

@@ -8,7 +8,9 @@
 
 #include <torch/csrc/autograd/VariableTypeUtils.h>
 #include <torch/extension.h>
+#include <algorithm>
 #include <cmath>
+#include <thread>
 #include <vector>
 
 // In the x direction, the location {0, ..., grid_size_x - 1} correspond to

pytorch3d/csrc/pulsar/global.h

@@ -15,8 +15,8 @@
 #endif
 
 #if defined(_WIN64) || defined(_WIN32)
-using uint = unsigned int;
-using ushort = unsigned short;
+#define uint unsigned int
+#define ushort unsigned short
 #endif
 
 #include "./logging.h" // <- include before torch/extension.h

pytorch3d/csrc/pulsar/gpu/commands.h

@@ -125,43 +125,29 @@ INLINE DEVICE float3 WARP_SUM_FLOAT3(
 // Floating point.
 // #define FMUL(a, b) __fmul_rn((a), (b))
 #define FMUL(a, b) ((a) * (b))
+#define FDIV(a, b) __fdiv_rn((a), (b))
 // #define FSUB(a, b) __fsub_rn((a), (b))
 #define FSUB(a, b) ((a) - (b))
-// HIP has no per-instruction rounding-mode override (round-to-nearest-even is
-// the default at the hardware level), so the CUDA *_rn intrinsics have no
-// runtime equivalent and we use plain operators. The HIP/clang compiler may
-// fuse a+b*c into a single-rounding FMA where CUDA's _rn would have prevented
-// it; if that becomes a numerical issue, add `-ffp-contract=off` to the pulsar
-// nvcc_args in setup.py.
-#if defined(USE_ROCM)
-#define FDIV(a, b) ((a) / (b))
-#define FADD(a, b) ((a) + (b))
-#define FSQRT(a) sqrtf(a)
-#define FPOW(a, b) powf((a), (b))
-#define FSATURATE(x) fmaxf(0.0f, fminf(1.0f, (x)))
-#define FMA(x, y, z) fmaf((x), (y), (z))
-#define FRCP(x) (1.0f / (x))
-#else
-#define FDIV(a, b) __fdiv_rn((a), (b))
 #define FADD(a, b) __fadd_rn((a), (b))
 #define FSQRT(a) __fsqrt_rn(a)
-#define FPOW(a, b) __powf((a), (b))
-#define FSATURATE(x) __saturatef(x)
-/** Calculates x*y+z. */
-#define FMA(x, y, z) __fmaf_rn((x), (y), (z))
-#define FRCP(x) __frcp_rn(x)
-#endif
 #define FEXP(a) fasterexp(a)
 #define FLN(a) fasterlog(a)
+#define FPOW(a, b) __powf((a), (b))
 #define FMAX(a, b) fmax((a), (b))
 #define FMIN(a, b) fmin((a), (b))
 #define FCEIL(a) ceilf(a)
 #define FFLOOR(a) floorf(a)
 #define FROUND(x) nearbyintf(x)
+#define FSATURATE(x) __saturatef(x)
 #define FABS(a) abs(a)
+#define IASF(a, loc) (loc) = __int_as_float(a)
+#define FASI(a, loc) (loc) = __float_as_int(a)
 #define FABSLEQAS(a, b, c) \
   ((a) <= (b) ? FSUB((b), (a)) <= (c) : FSUB((a), (b)) < (c))
+/** Calculates x*y+z. */
+#define FMA(x, y, z) __fmaf_rn((x), (y), (z))
 #define I2F(a) __int2float_rn(a)
+#define FRCP(x) __frcp_rn(x)
 #if !defined(USE_ROCM)
 __device__ static float atomicMax(float* address, float val) {
   int* address_as_i = (int*)address;
@@ -215,16 +201,8 @@ __device__ static float atomicMin(float* address, float val) {
   ATOMICADD(&((PTR)->x), VAL.x); \
   ATOMICADD(&((PTR)->y), VAL.y); \
   ATOMICADD(&((PTR)->z), VAL.z);
-#if !defined(USE_ROCM) && (CUDART_VERSION >= 10000) && (__CUDA_ARCH__ >= 600)
+#if (CUDART_VERSION >= 10000) && (__CUDA_ARCH__ >= 600)
 #define ATOMICADD_B(PTR, VAL) atomicAdd_block((PTR), (VAL))
-#elif defined(USE_ROCM)
-// HIP has no atomicAdd_block, but the semantic equivalent is a relaxed
-// fetch_add scoped to the workgroup (HIP's name for a CUDA thread block).
-// This avoids device-wide L2-coherent atomics for what are block-local
-// counters in pulsar's inner sphere-loading loop.
-#define ATOMICADD_B(PTR, VAL) \
-  __hip_atomic_fetch_add(     \
-      (PTR), (VAL), __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_WORKGROUP)
 #else
 #define ATOMICADD_B(PTR, VAL) ATOMICADD(PTR, VAL)
 #endif
@@ -439,7 +417,7 @@ __device__ static float atomicMin(float* address, float val) {
       (OUT_PTR),              \
       (NUM_SELECTED_PTR),     \
       (NUM_ITEMS),            \
-      (STREAM));
+      stream = (STREAM));
 
 #define COPY_HOST_DEV(PTR_D, PTR_H, TYPE, SIZE) \
   HANDLECUDA(cudaMemcpy(                        \

pytorch3d/csrc/pulsar/host/commands.h

@@ -114,6 +114,8 @@ INLINE float atomicMin(float* address, float val) {
 #define FMA(x, y, z) ((x) * (y) + (z))
 #define I2F(a) static_cast<float>(a)
 #define FRCP(x) (1.f / (x))
+#define IASF(x, loc) memcpy(&(loc), &(x), sizeof(x))
+#define FASI(x, loc) memcpy(&(loc), &(x), sizeof(x))
 #define DMAX(a, b) std::max((a), (b))
 #define DMIN(a, b) std::min((a), (b))
 #define DSATURATE(a) DMIN(1., DMAX(0., (a)))
@@ -355,11 +357,11 @@ void MAX_WS(
 //
 //
 #define END_PARALLEL() \
-  end_parallel:;       \
+  end_parallel :;      \
   }
 #define END_PARALLEL_NORET() }
 #define END_PARALLEL_2D() \
-  end_parallel:;          \
+  end_parallel :;         \
   }                       \
   }
 #define END_PARALLEL_2D_NORET() \

pytorch3d/csrc/pulsar/include/camera.h

@@ -70,6 +70,11 @@ struct CamGradInfo {
   float3 pixel_dir_y;
 };
 
+// TODO: remove once https://github.com/NVlabs/cub/issues/172 is resolved.
+struct IntWrapper {
+  int val;
+};
+
 } // namespace pulsar
 
 #endif

pytorch3d/csrc/pulsar/include/commands.h

@@ -9,11 +9,7 @@
 #ifndef PULSAR_NATIVE_COMMANDS_ROUTING_H_
 #define PULSAR_NATIVE_COMMANDS_ROUTING_H_
 
-#include <cstring>
 #include "../global.h"
-#if __cplusplus >= 202002L
-#include <bit>
-#endif
 
 // Commands available everywhere.
 #define MALLOC_HOST(VAR, TYPE, SIZE) \
@@ -36,37 +32,4 @@
 #include "../host/commands.h"
 #endif
 
-// Host + device int <-> float bit punning. Replaces IASF / FASI macros.
-IHD float pulsar_int_as_float(int v) noexcept {
-#ifdef __CUDA_ARCH__
-  return __int_as_float(v);
-#elif defined(__HIP_DEVICE_COMPILE__)
-  return __int_as_float(v);
-#else
-#if __cpp_lib_bit_cast >= 201806L
-  return std::bit_cast<float>(v);
-#else
-  float f;
-  std::memcpy(&f, &v, sizeof(f));
-  return f;
-#endif
-#endif
-}
-
-IHD int pulsar_float_as_int(float v) noexcept {
-#ifdef __CUDA_ARCH__
-  return __float_as_int(v);
-#elif defined(__HIP_DEVICE_COMPILE__)
-  return __float_as_int(v);
-#else
-#if __cpp_lib_bit_cast >= 201806L
-  return std::bit_cast<int>(v);
-#else
-  int i;
-  std::memcpy(&i, &v, sizeof(i));
-  return i;
-#endif
-#endif
-}
-
 #endif

pytorch3d/csrc/pulsar/include/math.h

@@ -149,6 +149,11 @@ IHD CamGradInfo operator*(const CamGradInfo& a, const float& b) {
   return res;
 }
 
+IHD IntWrapper operator+(const IntWrapper& a, const IntWrapper& b) {
+  IntWrapper res;
+  res.val = a.val + b.val;
+  return res;
+}
 } // namespace pulsar
 
 #endif

pytorch3d/csrc/pulsar/include/renderer.backward.device.h

@@ -155,8 +155,8 @@ void backward(
         stream);
     CHECKLAUNCH();
     SUM_WS(
-        self->ids_sorted_d,
-        self->n_grad_contributions_d,
+        (IntWrapper*)(self->ids_sorted_d),
+        (IntWrapper*)(self->n_grad_contributions_d),
         static_cast<int>(num_balls),
         self->workspace_d,
         self->workspace_size,

pytorch3d/csrc/pulsar/include/renderer.calc_gradients.device.h

@@ -121,8 +121,8 @@ GLOBAL void calc_gradients(
       grad_im_l[2]);
   // Start processing.
   for (int grad_idx = 0; grad_idx < n_track; ++grad_idx) {
-    int sphere_idx =
-        pulsar_float_as_int(forw_info_d[fwi_loc + 3 + 2 * grad_idx]);
+    int sphere_idx;
+    FASI(forw_info_d[fwi_loc + 3 + 2 * grad_idx], sphere_idx);
     PASSERT(
         sphere_idx == -1 ||
         sphere_idx >= 0 && static_cast<uint>(sphere_idx) < num_balls);

pytorch3d/csrc/pulsar/include/renderer.construct.device.h

@@ -52,7 +52,7 @@ HOST void construct(
   self->cam.film_width = width;
   self->cam.film_height = height;
   self->max_num_balls = max_num_balls;
-  MALLOC(self->result_d, float, width * height * n_channels);
+  MALLOC(self->result_d, float, width* height* n_channels);
   self->cam.orthogonal_projection = orthogonal_projection;
   self->cam.right_handed = right_handed_system;
   self->cam.background_normalization_depth = background_normalization_depth;
@@ -93,7 +93,7 @@ HOST void construct(
   MALLOC(self->di_sorted_d, DrawInfo, max_num_balls);
   MALLOC(self->region_flags_d, char, max_num_balls);
   MALLOC(self->num_selected_d, size_t, 1);
-  MALLOC(self->forw_info_d, float, width * height * (3 + 2 * n_track));
+  MALLOC(self->forw_info_d, float, width* height * (3 + 2 * n_track));
   MALLOC(self->min_max_pixels_d, IntersectInfo, 1);
   MALLOC(self->grad_pos_d, float3, max_num_balls);
   MALLOC(self->grad_col_d, float, max_num_balls* n_channels);

pytorch3d/csrc/pulsar/include/renderer.destruct.device.h

@@ -18,89 +18,68 @@ namespace Renderer {
 
 template <bool DEV>
 HOST void destruct(Renderer* self) {
-  if (self->result_d != NULL) {
+  if (self->result_d != NULL)
     FREE(self->result_d);
-  }
   self->result_d = NULL;
-  if (self->min_depth_d != NULL) {
+  if (self->min_depth_d != NULL)
     FREE(self->min_depth_d);
-  }
   self->min_depth_d = NULL;
-  if (self->min_depth_sorted_d != NULL) {
+  if (self->min_depth_sorted_d != NULL)
     FREE(self->min_depth_sorted_d);
-  }
   self->min_depth_sorted_d = NULL;
-  if (self->ii_d != NULL) {
+  if (self->ii_d != NULL)
     FREE(self->ii_d);
-  }
   self->ii_d = NULL;
-  if (self->ii_sorted_d != NULL) {
+  if (self->ii_sorted_d != NULL)
     FREE(self->ii_sorted_d);
-  }
   self->ii_sorted_d = NULL;
-  if (self->ids_d != NULL) {
+  if (self->ids_d != NULL)
     FREE(self->ids_d);
-  }
   self->ids_d = NULL;
-  if (self->ids_sorted_d != NULL) {
+  if (self->ids_sorted_d != NULL)
     FREE(self->ids_sorted_d);
-  }
   self->ids_sorted_d = NULL;
-  if (self->workspace_d != NULL) {
+  if (self->workspace_d != NULL)
     FREE(self->workspace_d);
-  }
   self->workspace_d = NULL;
-  if (self->di_d != NULL) {
+  if (self->di_d != NULL)
     FREE(self->di_d);
-  }
   self->di_d = NULL;
-  if (self->di_sorted_d != NULL) {
+  if (self->di_sorted_d != NULL)
     FREE(self->di_sorted_d);
-  }
   self->di_sorted_d = NULL;
-  if (self->region_flags_d != NULL) {
+  if (self->region_flags_d != NULL)
     FREE(self->region_flags_d);
-  }
   self->region_flags_d = NULL;
-  if (self->num_selected_d != NULL) {
+  if (self->num_selected_d != NULL)
     FREE(self->num_selected_d);
-  }
   self->num_selected_d = NULL;
-  if (self->forw_info_d != NULL) {
+  if (self->forw_info_d != NULL)
     FREE(self->forw_info_d);
-  }
   self->forw_info_d = NULL;
-  if (self->min_max_pixels_d != NULL) {
+  if (self->min_max_pixels_d != NULL)
     FREE(self->min_max_pixels_d);
-  }
   self->min_max_pixels_d = NULL;
-  if (self->grad_pos_d != NULL) {
+  if (self->grad_pos_d != NULL)
     FREE(self->grad_pos_d);
-  }
   self->grad_pos_d = NULL;
-  if (self->grad_col_d != NULL) {
+  if (self->grad_col_d != NULL)
     FREE(self->grad_col_d);
-  }
   self->grad_col_d = NULL;
-  if (self->grad_rad_d != NULL) {
+  if (self->grad_rad_d != NULL)
     FREE(self->grad_rad_d);
-  }
   self->grad_rad_d = NULL;
-  if (self->grad_cam_d != NULL) {
+  if (self->grad_cam_d != NULL)
     FREE(self->grad_cam_d);
-  }
   self->grad_cam_d = NULL;
-  if (self->grad_cam_buf_d != NULL) {
+  if (self->grad_cam_buf_d != NULL)
     FREE(self->grad_cam_buf_d);
-  }
   self->grad_cam_buf_d = NULL;
-  if (self->grad_opy_d != NULL) {
+  if (self->grad_opy_d != NULL)
     FREE(self->grad_opy_d);
-  }
   self->grad_opy_d = NULL;
-  if (self->n_grad_contributions_d != NULL) {
+  if (self->n_grad_contributions_d != NULL)
     FREE(self->n_grad_contributions_d);
-  }
   self->n_grad_contributions_d = NULL;
 }
 

pytorch3d/csrc/pulsar/include/renderer.fill_bg.device.h

@@ -41,8 +41,7 @@ GLOBAL void fill_bg(
     // sphere IDs and intersection depths.
     for (int i = 0; i < renderer.n_track; ++i) {
       int sphere_id = -1;
-      renderer.forw_info_d[write_loc + 3 + i * 2] =
-          pulsar_int_as_float(sphere_id);
+      IASF(sphere_id, renderer.forw_info_d[write_loc + 3 + i * 2]);
       renderer.forw_info_d[write_loc + 3 + i * 2 + 1] = -1.f;
     }
     if (mode == 0) {

pytorch3d/csrc/pulsar/include/renderer.h

@@ -255,7 +255,7 @@ GLOBAL void calc_signature(
  * for every iteration through the loading loop every thread could add a
  * 'hit' to the buffer.
  */
-#define RENDER_BUFFER_SIZE RENDER_BLOCK_SIZE * RENDER_BLOCK_SIZE * 2
+#define RENDER_BUFFER_SIZE RENDER_BLOCK_SIZE* RENDER_BLOCK_SIZE * 2
 /**
  * The threshold after which the spheres that are in the render buffer
  * are rendered and the buffer is flushed.

pytorch3d/csrc/pulsar/include/renderer.norm_sphere_gradients.device.h

@@ -64,9 +64,8 @@ GLOBAL void norm_sphere_gradients(Renderer renderer, const int num_balls) {
   // The sphere only contributes to the camera gradients if it is
   // large enough in screen space.
   if (renderer.ids_sorted_d[idx] > 0 && ii.max.x >= ii.min.x + 3 &&
-      ii.max.y >= ii.min.y + 3) {
+      ii.max.y >= ii.min.y + 3)
     renderer.ids_sorted_d[idx] = 1;
-  }
   END_PARALLEL_NORET();
 };
 

pytorch3d/csrc/pulsar/include/renderer.render.device.h

@@ -139,9 +139,8 @@ GLOBAL void render(
       coord_y < cam_norm.film_border_top + cam_norm.film_height) {
     // Initialize the result.
     if (mode == 0u) {
-      for (uint c_id = 0; c_id < cam_norm.n_channels; ++c_id) {
+      for (uint c_id = 0; c_id < cam_norm.n_channels; ++c_id)
         result[c_id] = bg_col[c_id];
-      }
     } else {
       result[0] = 0.f;
     }
@@ -191,22 +190,20 @@ GLOBAL void render(
             "render|found intersection with sphere %u.\n",
             sphere_id_l[write_idx]);
       }
-      if (ii.min.x == MAX_USHORT) {
+      if (ii.min.x == MAX_USHORT)
         // This is an invalid sphere (out of image). These spheres have
         // maximum depth. Since we ordered the spheres by earliest possible
         // intersection depth we re certain that there will no other sphere
         // that is relevant after this one.
         loading_done = true;
-      }
     }
     // Reset n_pixels_done.
     n_pixels_done = 0;
     thread_block.sync(); // Make sure n_loaded is updated.
     if (n_loaded > RENDER_BUFFER_LOAD_THRESH) {
       // The load buffer is full enough. Draw.
-      if (thread_block.thread_rank() == 0) {
+      if (thread_block.thread_rank() == 0)
         n_balls_loaded += n_loaded;
-      }
       max_closest_possible_intersection = 0.f;
       // This excludes threads outside of the image boundary. Also, it reduces
       // block artifacts.
@@ -293,9 +290,8 @@ GLOBAL void render(
       uint warp_done = thread_warp.ballot(done);
       int warp_done_bit_cnt = POPC(warp_done);
 #endif //__CUDACC__ && __HIP_PLATFORM_AMD__
-      if (thread_warp.thread_rank() == 0) {
+      if (thread_warp.thread_rank() == 0)
         ATOMICADD_B(&n_pixels_done, warp_done_bit_cnt);
-      }
       // This sync is necessary to keep n_loaded until all threads are done with
       // painting.
       thread_block.sync();
@@ -303,9 +299,8 @@ GLOBAL void render(
     }
     thread_block.sync();
   }
-  if (thread_block.thread_rank() == 0) {
+  if (thread_block.thread_rank() == 0)
     n_balls_loaded += n_loaded;
-  }
   PULSAR_LOG_DEV_PIX(
       PULSAR_LOG_RENDER_PIX,
       "render|loaded %d balls in total.\n",
@@ -391,9 +386,8 @@ GLOBAL void render(
             static_cast<float>(tracker.get_n_hits());
   } else {
     float sm_d_normfac = FRCP(FMAX(sm_d, FEPS));
-    for (uint c_id = 0; c_id < cam_norm.n_channels; ++c_id) {
+    for (uint c_id = 0; c_id < cam_norm.n_channels; ++c_id)
       result[c_id] *= sm_d_normfac;
-    }
     int write_loc = (coord_y - cam_norm.film_border_top) * cam_norm.film_width *
             (3 + 2 * n_track) +
         (coord_x - cam_norm.film_border_left) * (3 + 2 * n_track);
@@ -406,7 +400,7 @@ GLOBAL void render(
         IMIN(n_track, tracker.get_n_hits()));
     for (int i = 0; i < n_track; ++i) {
       int sphere_id = tracker.get_closest_sphere_id(i);
-      forw_info_d[write_loc + 3 + i * 2] = pulsar_int_as_float(sphere_id);
+      IASF(sphere_id, forw_info_d[write_loc + 3 + i * 2]);
       forw_info_d[write_loc + 3 + i * 2 + 1] =
           tracker.get_closest_sphere_depth(i) == MAX_FLOAT
           ? -1.f

pytorch3d/csrc/pulsar/pytorch/renderer.cpp

@@ -7,11 +7,13 @@
  */
 
 #include "./renderer.h"
+#include "../include/commands.h"
 #include "./camera.h"
 #include "./util.h"
 
+#include <ATen/ATen.h>
 #ifdef WITH_CUDA
-#include <c10/cuda/CUDAException.h>
+#include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 #endif
 
@@ -107,10 +109,10 @@ void Renderer::ensure_on_device(torch::Device device, bool /*non_blocking*/) {
     LOG_IF(INFO, PULSAR_LOG_INIT)
         << "Transferring render buffers between devices.";
     int prev_active;
-    C10_CUDA_CHECK(cudaGetDevice(&prev_active));
+    cudaGetDevice(&prev_active);
     if (this->device_type == c10::DeviceType::CUDA) {
       LOG_IF(INFO, PULSAR_LOG_INIT) << "  Destructing on CUDA.";
-      C10_CUDA_CHECK(cudaSetDevice(this->device_index));
+      cudaSetDevice(this->device_index);
       for (auto& nrend : this->renderer_vec) {
         PRE::destruct<true>(&nrend);
       }
@@ -122,7 +124,7 @@ void Renderer::ensure_on_device(torch::Device device, bool /*non_blocking*/) {
     }
     if (device.type() == c10::DeviceType::CUDA) {
       LOG_IF(INFO, PULSAR_LOG_INIT) << "  Constructing on CUDA.";
-      C10_CUDA_CHECK(cudaSetDevice(device.index()));
+      cudaSetDevice(device.index());
       for (auto& nrend : this->renderer_vec) {
         PRE::construct<true>(
             &nrend,
@@ -150,7 +152,7 @@ void Renderer::ensure_on_device(torch::Device device, bool /*non_blocking*/) {
             this->n_track());
       }
     }
-    C10_CUDA_CHECK(cudaSetDevice(prev_active));
+    cudaSetDevice(prev_active);
     this->device_type = device.type();
     this->device_index = device.index();
 #else
@@ -744,8 +746,8 @@ std::tuple<torch::Tensor, torch::Tensor> Renderer::forward(
 // moved to a CUDA device).
 #ifdef WITH_CUDA
     int prev_active;
-    C10_CUDA_CHECK(cudaGetDevice(&prev_active));
-    C10_CUDA_CHECK(cudaSetDevice(this->device_index));
+    cudaGetDevice(&prev_active);
+    cudaSetDevice(this->device_index);
 #ifdef PULSAR_TIMINGS_BATCHED_ENABLED
     START_TIME_CU(batch_forward);
 #endif
@@ -790,7 +792,7 @@ std::tuple<torch::Tensor, torch::Tensor> Renderer::forward(
     std::cout << "Forward render batched time per example: "
               << time_ms / static_cast<float>(batch_size) << "ms" << std::endl;
 #endif
-    C10_CUDA_CHECK(cudaSetDevice(prev_active));
+    cudaSetDevice(prev_active);
 #endif
   } else {
 #ifdef PULSAR_TIMINGS_BATCHED_ENABLED
@@ -858,9 +860,8 @@ std::tuple<torch::Tensor, torch::Tensor> Renderer::forward(
             ? (cudaStream_t) nullptr
 #endif
             : (cudaStream_t) nullptr);
-    if (mode == 1) {
+    if (mode == 1)
       results[batch_i] = results[batch_i].slice(2, 0, 1, 1);
-    }
     forw_infos[batch_i] = from_blob(
         this->renderer_vec[batch_i].forw_info_d,
         {this->renderer_vec[0].cam.film_height,
@@ -1101,8 +1102,8 @@ Renderer::backward(
 // the renderer to a CUDA device if not built with CUDA.
 #ifdef WITH_CUDA
     int prev_active;
-    C10_CUDA_CHECK(cudaGetDevice(&prev_active));
-    C10_CUDA_CHECK(cudaSetDevice(this->device_index));
+    cudaGetDevice(&prev_active);
+    cudaSetDevice(this->device_index);
 #ifdef PULSAR_TIMINGS_BATCHED_ENABLED
     START_TIME_CU(batch_backward);
 #endif
@@ -1206,7 +1207,7 @@ Renderer::backward(
             at::cuda::getCurrentCUDAStream());
       }
     }
-    C10_CUDA_CHECK(cudaSetDevice(prev_active));
+    cudaSetDevice(prev_active);
 #ifdef PULSAR_TIMINGS_BATCHED_ENABLED
     STOP_TIME_CU(batch_backward);
     float time_ms;

pytorch3d/csrc/pulsar/pytorch/renderer.h

@@ -128,9 +128,8 @@ struct Renderer {
     stream << "pulsar::Renderer[";
     // Device info.
     stream << self.device_type;
-    if (self.device_index != -1) {
+    if (self.device_index != -1)
       stream << ", ID " << self.device_index;
-    }
     stream << "]";
     return stream;
   }

pytorch3d/csrc/pulsar/pytorch/tensor_util.cpp

@@ -8,7 +8,6 @@
 
 #ifdef WITH_CUDA
 #include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAException.h>
 #include <cuda_runtime_api.h>
 #endif
 #include <torch/extension.h>
@@ -34,13 +33,13 @@ torch::Tensor sphere_ids_from_result_info_nograd(
           .contiguous();
   if (forw_info.device().type() == c10::DeviceType::CUDA) {
 #ifdef WITH_CUDA
-    C10_CUDA_CHECK(cudaMemcpyAsync(
+    cudaMemcpyAsync(
         result.data_ptr(),
         tmp.data_ptr(),
         sizeof(uint32_t) * tmp.size(0) * tmp.size(1) * tmp.size(2) *
             tmp.size(3),
         cudaMemcpyDeviceToDevice,
-        at::cuda::getCurrentCUDAStream()));
+        at::cuda::getCurrentCUDAStream());
 #else
     throw std::runtime_error(
         "Copy on CUDA device initiated but built "

pytorch3d/csrc/pulsar/pytorch/util.cpp

@@ -7,7 +7,6 @@
  */
 
 #ifdef WITH_CUDA
-#include <c10/cuda/CUDAException.h>
 #include <cuda_runtime_api.h>
 
 namespace pulsar {
@@ -18,8 +17,7 @@ void cudaDevToDev(
     const void* src,
     const int& size,
     const cudaStream_t& stream) {
-  C10_CUDA_CHECK(
-      cudaMemcpyAsync(trg, src, size, cudaMemcpyDeviceToDevice, stream));
+  cudaMemcpyAsync(trg, src, size, cudaMemcpyDeviceToDevice, stream);
 }
 
 void cudaDevToHost(
@@ -27,8 +25,7 @@ void cudaDevToHost(
     const void* src,
     const int& size,
     const cudaStream_t& stream) {
-  C10_CUDA_CHECK(
-      cudaMemcpyAsync(trg, src, size, cudaMemcpyDeviceToHost, stream));
+  cudaMemcpyAsync(trg, src, size, cudaMemcpyDeviceToHost, stream);
 }
 
 } // namespace pytorch

pytorch3d/csrc/pulsar/warnings.cpp

@@ -6,6 +6,9 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#include "./global.h"
+#include "./logging.h"
+
 /**
  * A compilation unit to provide warnings about the code and avoid
  * repeated messages.

pytorch3d/csrc/rasterize_coarse/bitmask.cuh

@@ -25,7 +25,7 @@ class BitMask {
 
   // Use all threads in the current block to clear all bits of this BitMask
   __device__ void block_clear() {
-    for (auto i = threadIdx.x; i < H * W * D; i += blockDim.x) {
+    for (int i = threadIdx.x; i < H * W * D; i += blockDim.x) {
       data[i] = 0;
     }
     __syncthreads();

pytorch3d/csrc/rasterize_coarse/rasterize_coarse.cu

@@ -23,8 +23,8 @@ __global__ void TriangleBoundingBoxKernel(
     const float blur_radius,
     float* bboxes, // (4, F)
     bool* skip_face) { // (F,)
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
-  const auto num_threads = blockDim.x * gridDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = blockDim.x * gridDim.x;
   const float sqrt_radius = sqrt(blur_radius);
   for (int f = tid; f < F; f += num_threads) {
     const float v0x = face_verts[f * 9 + 0 * 3 + 0];
@@ -56,8 +56,8 @@ __global__ void PointBoundingBoxKernel(
     const int P,
     float* bboxes, // (4, P)
     bool* skip_points) {
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
-  const auto num_threads = blockDim.x * gridDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = blockDim.x * gridDim.x;
   for (int p = tid; p < P; p += num_threads) {
     const float x = points[p * 3 + 0];
     const float y = points[p * 3 + 1];
@@ -113,7 +113,7 @@ __global__ void RasterizeCoarseCudaKernel(
   const int chunks_per_batch = 1 + (E - 1) / chunk_size;
   const int num_chunks = N * chunks_per_batch;
 
-  for (auto chunk = blockIdx.x; chunk < num_chunks; chunk += gridDim.x) {
+  for (int chunk = blockIdx.x; chunk < num_chunks; chunk += gridDim.x) {
     const int batch_idx = chunk / chunks_per_batch; // batch index
     const int chunk_idx = chunk % chunks_per_batch;
     const int elem_chunk_start_idx = chunk_idx * chunk_size;
@@ -123,7 +123,7 @@ __global__ void RasterizeCoarseCudaKernel(
     const int64_t elem_stop_idx = elem_start_idx + elems_per_batch[batch_idx];
 
     // Have each thread handle a different face within the chunk
-    for (auto e = threadIdx.x; e < chunk_size; e += blockDim.x) {
+    for (int e = threadIdx.x; e < chunk_size; e += blockDim.x) {
       const int e_idx = elem_chunk_start_idx + e;
 
       // Check that we are still within the same element of the batch
@@ -170,7 +170,7 @@ __global__ void RasterizeCoarseCudaKernel(
     // Now we have processed every elem in the current chunk. We need to
     // count the number of elems in each bin so we can write the indices
     // out to global memory. We have each thread handle a different bin.
-    for (auto byx = threadIdx.x; byx < num_bins_y * num_bins_x;
+    for (int byx = threadIdx.x; byx < num_bins_y * num_bins_x;
          byx += blockDim.x) {
       const int by = byx / num_bins_x;
       const int bx = byx % num_bins_x;

pytorch3d/csrc/rasterize_meshes/rasterize_meshes.cu

@@ -260,8 +260,8 @@ __global__ void RasterizeMeshesNaiveCudaKernel(
     float* pix_dists,
     float* bary) {
   // Simple version: One thread per output pixel
-  auto num_threads = gridDim.x * blockDim.x;
-  auto tid = blockDim.x * blockIdx.x + threadIdx.x;
+  int num_threads = gridDim.x * blockDim.x;
+  int tid = blockDim.x * blockIdx.x + threadIdx.x;
 
   for (int i = tid; i < N * H * W; i += num_threads) {
     // Convert linear index to 3D index
@@ -446,8 +446,8 @@ __global__ void RasterizeMeshesBackwardCudaKernel(
 
   // Parallelize over each pixel in images of
   // size H * W, for each image in the batch of size N.
-  const auto num_threads = gridDim.x * blockDim.x;
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.x * blockDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
   for (int t_i = tid; t_i < N * H * W; t_i += num_threads) {
     // Convert linear index to 3D index
@@ -650,8 +650,8 @@ __global__ void RasterizeMeshesFineCudaKernel(
 ) {
   // This can be more than H * W if H or W are not divisible by bin_size.
   int num_pixels = N * BH * BW * bin_size * bin_size;
-  auto num_threads = gridDim.x * blockDim.x;
-  auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  int num_threads = gridDim.x * blockDim.x;
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
   for (int pid = tid; pid < num_pixels; pid += num_threads) {
     // Convert linear index into bin and pixel indices. We make the within

pytorch3d/csrc/rasterize_meshes/rasterize_meshes.h

@@ -138,9 +138,6 @@ RasterizeMeshesNaive(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(face_verts);
-    CHECK_CPU(mesh_to_face_first_idx);
-    CHECK_CPU(num_faces_per_mesh);
     return RasterizeMeshesNaiveCpu(
         face_verts,
         mesh_to_face_first_idx,
@@ -235,11 +232,6 @@ torch::Tensor RasterizeMeshesBackward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(face_verts);
-    CHECK_CPU(pix_to_face);
-    CHECK_CPU(grad_zbuf);
-    CHECK_CPU(grad_bary);
-    CHECK_CPU(grad_dists);
     return RasterizeMeshesBackwardCpu(
         face_verts,
         pix_to_face,
@@ -314,9 +306,6 @@ torch::Tensor RasterizeMeshesCoarse(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(face_verts);
-    CHECK_CPU(mesh_to_face_first_idx);
-    CHECK_CPU(num_faces_per_mesh);
     return RasterizeMeshesCoarseCpu(
         face_verts,
         mesh_to_face_first_idx,
@@ -434,8 +423,6 @@ RasterizeMeshesFine(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(face_verts);
-    CHECK_CPU(bin_faces);
     AT_ERROR("NOT IMPLEMENTED");
   }
 }

pytorch3d/csrc/rasterize_meshes/rasterize_meshes_cpu.cpp

@@ -8,6 +8,8 @@
 
 #include <torch/extension.h>
 #include <algorithm>
+#include <list>
+#include <queue>
 #include <thread>
 #include <tuple>
 #include "ATen/core/TensorAccessor.h"
@@ -105,8 +107,6 @@ auto ComputeFaceAreas(const torch::Tensor& face_verts) {
   return face_areas;
 }
 
-namespace {
-
 // Helper function to use with std::find_if to find the index of any
 // values in the top k struct which match a given idx.
 struct IsNeighbor {
@@ -119,6 +119,7 @@ struct IsNeighbor {
   int neighbor_idx;
 };
 
+namespace {
 void RasterizeMeshesNaiveCpu_worker(
     const int start_yi,
     const int end_yi,

pytorch3d/csrc/rasterize_points/rasterize_points.cu

@@ -97,8 +97,8 @@ __global__ void RasterizePointsNaiveCudaKernel(
     float* zbuf, // (N, H, W, K)
     float* pix_dists) { // (N, H, W, K)
   // Simple version: One thread per output pixel
-  const auto num_threads = gridDim.x * blockDim.x;
-  const auto tid = blockDim.x * blockIdx.x + threadIdx.x;
+  const int num_threads = gridDim.x * blockDim.x;
+  const int tid = blockDim.x * blockIdx.x + threadIdx.x;
   for (int i = tid; i < N * H * W; i += num_threads) {
     // Convert linear index to 3D index
     const int n = i / (H * W); // Batch index
@@ -237,8 +237,8 @@ __global__ void RasterizePointsFineCudaKernel(
     float* pix_dists) { // (N, H, W, K)
   // This can be more than H * W if H or W are not divisible by bin_size.
   const int num_pixels = N * BH * BW * bin_size * bin_size;
-  const auto num_threads = gridDim.x * blockDim.x;
-  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const int num_threads = gridDim.x * blockDim.x;
+  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
 
   for (int pid = tid; pid < num_pixels; pid += num_threads) {
     // Convert linear index into bin and pixel indices. We make the within
@@ -376,8 +376,8 @@ __global__ void RasterizePointsBackwardCudaKernel(
     float* grad_points) { // (P, 3)
   // Parallelized over each of K points per pixel, for each pixel in images of
   // size H * W, for each image in the batch of size N.
-  auto num_threads = gridDim.x * blockDim.x;
-  auto tid = blockIdx.x * blockDim.x + threadIdx.x;
+  int num_threads = gridDim.x * blockDim.x;
+  int tid = blockIdx.x * blockDim.x + threadIdx.x;
   for (int i = tid; i < N * H * W * K; i += num_threads) {
     // const int n = i / (H * W * K); // batch index (not needed).
     const int yxk = i % (H * W * K);

pytorch3d/csrc/rasterize_points/rasterize_points.h

@@ -91,10 +91,6 @@ std::tuple<torch::Tensor, torch::Tensor, torch::Tensor> RasterizePointsNaive(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(points);
-    CHECK_CPU(cloud_to_packed_first_idx);
-    CHECK_CPU(num_points_per_cloud);
-    CHECK_CPU(radius);
     return RasterizePointsNaiveCpu(
         points,
         cloud_to_packed_first_idx,
@@ -170,10 +166,6 @@ torch::Tensor RasterizePointsCoarse(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(points);
-    CHECK_CPU(cloud_to_packed_first_idx);
-    CHECK_CPU(num_points_per_cloud);
-    CHECK_CPU(radius);
     return RasterizePointsCoarseCpu(
         points,
         cloud_to_packed_first_idx,
@@ -240,8 +232,6 @@ std::tuple<torch::Tensor, torch::Tensor, torch::Tensor> RasterizePointsFine(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(points);
-    CHECK_CPU(bin_points);
     AT_ERROR("NOT IMPLEMENTED");
   }
 }
@@ -294,10 +284,6 @@ torch::Tensor RasterizePointsBackward(
     AT_ERROR("Not compiled with GPU support");
 #endif
   } else {
-    CHECK_CPU(points);
-    CHECK_CPU(idxs);
-    CHECK_CPU(grad_zbuf);
-    CHECK_CPU(grad_dists);
     return RasterizePointsBackwardCpu(points, idxs, grad_zbuf, grad_dists);
   }
 }

pytorch3d/csrc/sample_farthest_points/sample_farthest_points.cu

@@ -35,6 +35,8 @@ __global__ void FarthestPointSamplingKernel(
   __shared__ int64_t selected_store;
 
   // Get constants
+  const int64_t N = points.size(0);
+  const int64_t P = points.size(1);
   const int64_t D = points.size(2);
 
   // Get batch index and thread index
@@ -107,8 +109,7 @@ at::Tensor FarthestPointSamplingCuda(
     const at::Tensor& points, // (N, P, 3)
     const at::Tensor& lengths, // (N,)
     const at::Tensor& K, // (N,)
-    const at::Tensor& start_idxs,
-    const int64_t max_K_known = -1) {
+    const at::Tensor& start_idxs) {
   // Check inputs are on the same device
   at::TensorArg p_t{points, "points", 1}, lengths_t{lengths, "lengths", 2},
       k_t{K, "K", 3}, start_idxs_t{start_idxs, "start_idxs", 4};
@@ -130,12 +131,7 @@ at::Tensor FarthestPointSamplingCuda(
 
   const int64_t N = points.size(0);
   const int64_t P = points.size(1);
-  int64_t max_K;
-  if (max_K_known > 0) {
-    max_K = max_K_known;
-  } else {
-    max_K = at::max(K).item<int64_t>();
-  }
+  const int64_t max_K = at::max(K).item<int64_t>();
 
   // Initialize the output tensor with the sampled indices
   auto idxs = at::full({N, max_K}, -1, lengths.options());

pytorch3d/csrc/sample_farthest_points/sample_farthest_points.h

@@ -43,8 +43,7 @@ at::Tensor FarthestPointSamplingCuda(
     const at::Tensor& points,
     const at::Tensor& lengths,
     const at::Tensor& K,
-    const at::Tensor& start_idxs,
-    const int64_t max_K_known = -1);
+    const at::Tensor& start_idxs);
 
 at::Tensor FarthestPointSamplingCpu(
     const at::Tensor& points,
@@ -57,23 +56,17 @@ at::Tensor FarthestPointSampling(
     const at::Tensor& points,
     const at::Tensor& lengths,
     const at::Tensor& K,
-    const at::Tensor& start_idxs,
-    const int64_t max_K_known = -1) {
+    const at::Tensor& start_idxs) {
   if (points.is_cuda() || lengths.is_cuda() || K.is_cuda()) {
 #ifdef WITH_CUDA
     CHECK_CUDA(points);
     CHECK_CUDA(lengths);
     CHECK_CUDA(K);
     CHECK_CUDA(start_idxs);
-    return FarthestPointSamplingCuda(
-        points, lengths, K, start_idxs, max_K_known);
+    return FarthestPointSamplingCuda(points, lengths, K, start_idxs);
 #else
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(points);
-  CHECK_CPU(lengths);
-  CHECK_CPU(K);
-  CHECK_CPU(start_idxs);
   return FarthestPointSamplingCpu(points, lengths, K, start_idxs);
 }

pytorch3d/csrc/sample_pdf/sample_pdf.h

@@ -71,8 +71,6 @@ inline void SamplePdf(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
-  CHECK_CPU(weights);
-  CHECK_CPU(outputs);
   CHECK_CONTIGUOUS(outputs);
   SamplePdfCpu(bins, weights, outputs, eps);
 }

pytorch3d/csrc/utils/dispatch.cuh

@@ -99,7 +99,8 @@ namespace {
 // and increment it via template recursion until it is equal to the run-time
 // argument N.
 template <
-    template <typename, int64_t> class Kernel,
+    template <typename, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,
@@ -123,7 +124,8 @@ struct DispatchKernelHelper1D {
 // 1D dispatch: Specialization when curN == maxN
 // We need this base case to avoid infinite template recursion.
 template <
-    template <typename, int64_t> class Kernel,
+    template <typename, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,
@@ -143,7 +145,8 @@ struct DispatchKernelHelper1D<Kernel, T, minN, maxN, maxN, Args...> {
 // the run-time values of N and M, at which point we dispatch to the run
 // method of the kernel.
 template <
-    template <typename, int64_t, int64_t> class Kernel,
+    template <typename, int64_t, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,
@@ -200,7 +203,8 @@ struct DispatchKernelHelper2D {
 
 // 2D dispatch, specialization for curN == maxN
 template <
-    template <typename, int64_t, int64_t> class Kernel,
+    template <typename, int64_t, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,
@@ -239,7 +243,8 @@ struct DispatchKernelHelper2D<
 
 // 2D dispatch, specialization for curM == maxM
 template <
-    template <typename, int64_t, int64_t> class Kernel,
+    template <typename, int64_t, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,
@@ -278,7 +283,8 @@ struct DispatchKernelHelper2D<
 
 // 2D dispatch, specialization for curN == maxN, curM == maxM
 template <
-    template <typename, int64_t, int64_t> class Kernel,
+    template <typename, int64_t, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,
@@ -307,7 +313,8 @@ struct DispatchKernelHelper2D<
 
 // This is the function we expect users to call to dispatch to 1D functions
 template <
-    template <typename, int64_t> class Kernel,
+    template <typename, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,
@@ -323,7 +330,8 @@ void DispatchKernel1D(const int64_t N, Args... args) {
 
 // This is the function we expect users to call to dispatch to 2D functions
 template <
-    template <typename, int64_t, int64_t> class Kernel,
+    template <typename, int64_t, int64_t>
+    class Kernel,
     typename T,
     int64_t minN,
     int64_t maxN,

pytorch3d/csrc/utils/geometry_utils.cuh

@@ -376,6 +376,8 @@ PointLineDistanceBackward(
   float tt = t_top / t_bot;
   tt = __saturatef(tt);
   const float2 p_proj = (1.0f - tt) * v0 + tt * v1;
+  const float2 d = p - p_proj;
+  const float dist = sqrt(dot(d, d));
 
   const float2 grad_p = -1.0f * grad_dist * 2.0f * (p_proj - p);
   const float2 grad_v0 = grad_dist * (1.0f - tt) * 2.0f * (p_proj - p);

pytorch3d/csrc/utils/pytorch3d_cutils.h

@@ -15,7 +15,3 @@
 #define CHECK_CONTIGUOUS_CUDA(x) \
   CHECK_CUDA(x);                 \
   CHECK_CONTIGUOUS(x)
-#define CHECK_CPU(x)                    \
-  TORCH_CHECK(                          \
-      x.device().type() == torch::kCPU, \
-      "Cannot use CPU implementation: " #x " not on CPU.")

pytorch3d/csrc/utils/vec2.h

@@ -19,7 +19,7 @@ template <
         std::is_same<T, double>::value || std::is_same<T, float>::value>>
 struct vec2 {
   T x, y;
-  using scalar_t = T;
+  typedef T scalar_t;
   vec2(T x, T y) : x(x), y(y) {}
 };
 

pytorch3d/csrc/utils/vec3.h

@@ -18,7 +18,7 @@ template <
         std::is_same<T, double>::value || std::is_same<T, float>::value>>
 struct vec3 {
   T x, y, z;
-  using scalar_t = T;
+  typedef T scalar_t;
   vec3(T x, T y, T z) : x(x), y(y), z(z) {}
 };
 

pytorch3d/csrc/utils/warp_reduce.cuh

@@ -10,20 +10,10 @@
 #include <math.h>
 #include <cstdio>
 
-// Helper functions WarpReduceMin and WarpReduceMax used in .cu files.
-// Starting in Volta, instructions are no longer synchronous within a warp,
-// so on CUDA __syncwarp() is required between dependent shared-memory
-// accesses in the unrolled tail reduction.
-//
-// On AMD/HIP no __syncwarp() is needed here: all wavefront lanes execute
-// in lockstep (AMD has no equivalent of NVIDIA's Independent Thread
-// Scheduling), and the AMDGPU memory model guarantees that LDS operations
-// issued by the same wavefront are observed in program order without an
-// explicit s_waitcnt — see the LLVM AMDGPU backend memory-model rules
-// (https://llvm.org/docs/AMDGPUUsage.html) and the HIP hardware-
-// implementation docs. This holds for both wave32 (RDNA, gfx10xx/11xx/
-// 12xx) and wave64 (CDNA, gfx9xx), so the USE_ROCM skip is
-// architecture-independent.
+// Helper functions WarpReduceMin and WarpReduceMax used in .cu files
+// Starting in Volta, instructions are no longer synchronous within a warp.
+// We need to call __syncwarp() to sync the 32 threads in the warp
+// instead of all the threads in the block.
 
 template <typename scalar_t>
 __device__ void
@@ -33,6 +23,8 @@ WarpReduceMin(scalar_t* min_dists, int64_t* min_idxs, const size_t tid) {
     min_idxs[tid] = min_idxs[tid + 32];
     min_dists[tid] = min_dists[tid + 32];
   }
+// AMD does not use explicit syncwarp and instead automatically inserts memory
+// fences during compilation.
 #if !defined(USE_ROCM)
   __syncwarp();
 #endif

pytorch3d/datasets/r2n2/r2n2.py

@@ -200,7 +200,6 @@ class R2N2(ShapeNetBase):  # pragma: no cover
             ) % (shapenet_dir, ", ".join(synset_not_present))
             warnings.warn(msg)
 
-    # pyrefly: ignore [bad-override]
     def __getitem__(self, model_idx, view_idxs: Optional[List[int]] = None) -> Dict:
         """
         Read a model by the given index.
@@ -371,7 +370,6 @@ class R2N2(ShapeNetBase):  # pragma: no cover
         T = RT[3, :3]
         return R, T
 
-    # pyrefly: ignore [bad-override]
     def render(
         self,
         model_ids: Optional[List[str]] = None,

pytorch3d/datasets/r2n2/utils.py

@@ -62,10 +62,8 @@ def collate_batched_R2N2(batch: List[Dict]):  # pragma: no cover
     # all models have the same number of views V, stack the batches of
     # views of each model into a new batch of shape (N, V, H, W, 3).
     # Otherwise leave it as a list.
-    # pyrefly: ignore [not-iterable]
     if "images" in collated_dict:
         try:
-            # pyrefly: ignore [unsupported-operation]
             collated_dict["images"] = torch.stack(collated_dict["images"])
         except RuntimeError:
             print(
@@ -77,14 +75,10 @@ def collate_batched_R2N2(batch: List[Dict]):  # pragma: no cover
     # matrices and that all models have the same number of views V, stack each
     # type of matrices into a new batch of shape (N, V, ...).
     # Otherwise leave them as lists.
-    # pyrefly: ignore [not-iterable]
     if all(x in collated_dict for x in ["R", "T", "K"]):
         try:
-            # pyrefly: ignore [unsupported-operation]
             collated_dict["R"] = torch.stack(collated_dict["R"])  # (N, V, 3, 3)
-            # pyrefly: ignore [unsupported-operation]
             collated_dict["T"] = torch.stack(collated_dict["T"])  # (N, V, 3)
-            # pyrefly: ignore [unsupported-operation]
             collated_dict["K"] = torch.stack(collated_dict["K"])  # (N, V, 4, 4)
         except RuntimeError:
             print(
@@ -95,10 +89,8 @@ def collate_batched_R2N2(batch: List[Dict]):  # pragma: no cover
     # If collate_batched_meshes receives voxels and all models have the same
     # number of views V, stack the batches of voxels into a new batch of shape
     # (N, V, S, S, S), where S is the voxel size.
-    # pyrefly: ignore [not-iterable]
     if "voxels" in collated_dict:
         try:
-            # pyrefly: ignore [unsupported-operation]
             collated_dict["voxels"] = torch.stack(collated_dict["voxels"])
         except RuntimeError:
             print(
@@ -466,8 +458,6 @@ class BlenderCamera(CamerasBase):  # pragma: no cover
 
     def get_projection_transform(self, **kwargs) -> Transform3d:
         transform = Transform3d(device=self.device)
-        # pyre-fixme[29]: `Union[(self: TensorBase, dim0: int, dim1: int) -> Tensor,
-        #  Tensor, Module]` is not a function.
         transform._matrix = self.K.transpose(1, 2).contiguous()
         return transform
 

pytorch3d/datasets/shapenet/shapenet_core.py

@@ -83,7 +83,7 @@ class ShapeNetCore(ShapeNetBase):  # pragma: no cover
                 ):
                     synset_set.add(synset)
                 elif (synset in self.synset_inv.keys()) and (
-                    path.isdir(path.join(data_dir, self.synset_inv[synset]))
+                    (path.isdir(path.join(data_dir, self.synset_inv[synset])))
                 ):
                     synset_set.add(self.synset_inv[synset])
                 else:

pytorch3d/datasets/shapenet_base.py

@@ -52,7 +52,6 @@ class ShapeNetBase(torch.utils.data.Dataset):  # pragma: no cover
         """
         return len(self.model_ids)
 
-    # pyrefly: ignore [bad-override-param-name]
     def __getitem__(self, idx) -> Dict:
         """
         Read a model by the given index. Need to be implemented for every child class
@@ -148,17 +147,12 @@ class ShapeNetBase(torch.utils.data.Dataset):  # pragma: no cover
         idxs = self._handle_render_inputs(model_ids, categories, sample_nums, idxs)
         # Use the getitem method which loads mesh + texture
         models = [self[idx] for idx in idxs]
-        # pyrefly: ignore [unsupported-operation]
         meshes = collate_batched_meshes(models)["mesh"]
-        # pyrefly: ignore [missing-attribute]
         if meshes.textures is None:
-            # pyrefly: ignore [missing-attribute]
             meshes.textures = TexturesVertex(
-                # pyrefly: ignore [missing-attribute]
                 verts_features=torch.ones_like(meshes.verts_padded(), device=device)
             )
 
-        # pyrefly: ignore [missing-attribute]
         meshes = meshes.to(device)
         cameras = kwargs.get("cameras", FoVPerspectiveCameras()).to(device)
         if len(cameras) != 1 and len(cameras) % len(meshes) != 0: