Fix CQS signal readability-braces-around-statements in fbcode/vision/fair

Reviewed By: bottler

Differential Revision: D94068738

fbshipit-source-id: cd47c67d4269ac7461acb73da6de9e4373da9d4c

.circleci/regenerate.py

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

.github/workflows/build.yml

@@ -0,0 +1,23 @@
+name: facebookresearch/pytorch3d/build_and_test
+on:
+  pull_request:
+    branches:
+      - main
+  push:
+    branches:
+      - main
+jobs:
+  binary_linux_conda_cuda:
+    runs-on: 4-core-ubuntu-gpu-t4
+    env:
+      PYTHON_VERSION: "3.12"
+      BUILD_VERSION: "${{ github.run_number }}"
+      PYTORCH_VERSION: "2.4.1"
+      CU_VERSION: "cu121"
+      JUST_TESTRUN: 1
+    steps:
+    - uses: actions/checkout@v4
+    - name: Build and run tests
+      run: |-
+        conda create --name env --yes --quiet conda-build
+        conda run --no-capture-output --name env python3 ./packaging/build_conda.py --use-conda-cuda

dev/linter.sh

@@ -10,7 +10,7 @@
 DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 DIR=$(dirname "${DIR}")
 
-if [[ -f "${DIR}/TARGETS" ]]
+if [[ -f "${DIR}/BUCK" ]]
 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; pyre -l vision/fair/pytorch3d/ )
+  ( cd ~/fbsource/fbcode; arc pyre check //vision/fair/pytorch3d/... )
 fi

docs/conf.py

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

docs/examples/pulsar_basic.py

@@ -10,6 +10,7 @@ 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

@@ -11,6 +11,7 @@ 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

@@ -14,6 +14,7 @@ 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

@@ -14,6 +14,7 @@ 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

@@ -18,6 +18,7 @@ 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

@@ -13,6 +13,7 @@ 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

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

packaging/build_conda.py

@@ -4,10 +4,11 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import argparse
 import os.path
 import runpy
 import subprocess
-from typing import List
+from typing import List, Tuple
 
 # required env vars:
 # CU_VERSION: E.g. cu112
@@ -23,7 +24,7 @@ pytorch_major_minor = tuple(int(i) for i in PYTORCH_VERSION.split(".")[:2])
 source_root_dir = os.environ["PWD"]
 
 
-def version_constraint(version):
+def version_constraint(version) -> str:
     """
     Given version "11.3" returns " >=11.3,<11.4"
     """
@@ -32,7 +33,7 @@ def version_constraint(version):
     return f" >={version},<{upper}"
 
 
-def get_cuda_major_minor():
+def get_cuda_major_minor() -> Tuple[str, str]:
     if CU_VERSION == "cpu":
         raise ValueError("fn only for cuda builds")
     if len(CU_VERSION) != 5 or CU_VERSION[:2] != "cu":
@@ -42,11 +43,10 @@ def get_cuda_major_minor():
     return major, minor
 
 
-def setup_cuda():
+def setup_cuda(use_conda_cuda: bool) -> List[str]:
     if CU_VERSION == "cpu":
-        return
+        return []
     major, minor = get_cuda_major_minor()
-    os.environ["CUDA_HOME"] = f"/usr/local/cuda-{major}.{minor}/"
     os.environ["FORCE_CUDA"] = "1"
 
     basic_nvcc_flags = (
@@ -75,6 +75,15 @@ def setup_cuda():
 
     if os.environ.get("JUST_TESTRUN", "0") != "1":
         os.environ["NVCC_FLAGS"] = nvcc_flags
+    if use_conda_cuda:
+        os.environ["CONDA_CUDA_TOOLKIT_BUILD_CONSTRAINT1"] = "- cuda-toolkit"
+        os.environ["CONDA_CUDA_TOOLKIT_BUILD_CONSTRAINT2"] = (
+            f"- cuda-version={major}.{minor}"
+        )
+        return ["-c", f"nvidia/label/cuda-{major}.{minor}.0"]
+    else:
+        os.environ["CUDA_HOME"] = f"/usr/local/cuda-{major}.{minor}/"
+        return []
 
 
 def setup_conda_pytorch_constraint() -> List[str]:
@@ -95,7 +104,7 @@ def setup_conda_pytorch_constraint() -> List[str]:
         return ["-c", "pytorch", "-c", "nvidia"]
 
 
-def setup_conda_cudatoolkit_constraint():
+def setup_conda_cudatoolkit_constraint() -> None:
     if CU_VERSION == "cpu":
         os.environ["CONDA_CPUONLY_FEATURE"] = "- cpuonly"
         os.environ["CONDA_CUDATOOLKIT_CONSTRAINT"] = ""
@@ -116,7 +125,7 @@ def setup_conda_cudatoolkit_constraint():
     os.environ["CONDA_CUDATOOLKIT_CONSTRAINT"] = toolkit
 
 
-def do_build(start_args: List[str]):
+def do_build(start_args: List[str]) -> None:
     args = start_args.copy()
 
     test_flag = os.environ.get("TEST_FLAG")
@@ -132,8 +141,16 @@ def do_build(start_args: List[str]):
 
 
 if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Build the conda package.")
+    parser.add_argument(
+        "--use-conda-cuda",
+        action="store_true",
+        help="get cuda from conda ignoring local cuda",
+    )
+    our_args = parser.parse_args()
+
     args = ["conda", "build"]
-    setup_cuda()
+    args += setup_cuda(use_conda_cuda=our_args.use_conda_cuda)
 
     init_path = source_root_dir + "/pytorch3d/__init__.py"
     build_version = runpy.run_path(init_path)["__version__"]

packaging/pytorch3d/meta.yaml

@@ -8,10 +8,13 @@ source:
 requirements:
   build:
     - {{ compiler('c') }} # [win]
+    {{ environ.get('CONDA_CUDA_TOOLKIT_BUILD_CONSTRAINT1', '') }}
+    {{ environ.get('CONDA_CUDA_TOOLKIT_BUILD_CONSTRAINT2', '') }}
     {{ environ.get('CONDA_CUB_CONSTRAINT') }}
 
   host:
     - python
+    - mkl =2023  # [x86_64]
     {{ environ.get('SETUPTOOLS_CONSTRAINT') }}
     {{ environ.get('CONDA_PYTORCH_BUILD_CONSTRAINT') }}
     {{ environ.get('CONDA_PYTORCH_MKL_CONSTRAINT') }}
@@ -22,6 +25,7 @@ requirements:
     - python
     - numpy >=1.11
     - torchvision >=0.5
+    - mkl =2023  # [x86_64]
     - iopath
     {{ environ.get('CONDA_PYTORCH_CONSTRAINT') }}
     {{ environ.get('CONDA_CUDATOOLKIT_CONSTRAINT') }}
@@ -47,8 +51,11 @@ test:
     - imageio
     - hydra-core
     - accelerate
+    - matplotlib
+    - tabulate
+    - pandas
+    - sqlalchemy
   commands:
-    #pytest .
     python -m unittest discover -v -s tests -t .
 
 

projects/implicitron_trainer/experiment.py

@@ -7,7 +7,7 @@
 
 # pyre-unsafe
 
-""""
+""" "
 This file is the entry point for launching experiments with Implicitron.
 
 Launch Training
@@ -44,25 +44,22 @@ 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,
 )

projects/implicitron_trainer/impl/model_factory.py

@@ -11,7 +11,6 @@ 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
@@ -26,7 +25,6 @@ logger = logging.getLogger(__name__)
 
 
 class ModelFactoryBase(ReplaceableBase):
-
     resume: bool = True  # resume from the last checkpoint
 
     def __call__(self, **kwargs) -> ImplicitronModelBase:
@@ -116,7 +114,9 @@ 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
+                    model_io.get_model_path(model_path),
+                    map_location=map_location,
+                    weights_only=True,
                 )
 
                 try:

projects/implicitron_trainer/impl/optimizer_factory.py

@@ -14,9 +14,7 @@ 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 (
@@ -123,6 +121,7 @@ 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 = [
@@ -241,7 +240,7 @@ class ImplicitronOptimizerFactory(OptimizerFactoryBase):
                     map_location = {
                         "cuda:%d" % 0: "cuda:%d" % accelerator.local_process_index
                     }
-                optimizer_state = torch.load(opt_path, map_location)
+                optimizer_state = torch.load(opt_path, map_location, weights_only=True)
             else:
                 raise FileNotFoundError(f"Optimizer state {opt_path} does not exist.")
         return optimizer_state

projects/implicitron_trainer/impl/training_loop.py

@@ -161,7 +161,6 @@ 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)
@@ -395,6 +394,7 @@ 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/test_experiment.py

@@ -12,7 +12,6 @@ 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 (
@@ -53,12 +52,8 @@ class TestExperiment(unittest.TestCase):
         cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_class_type = (
             "JsonIndexDatasetMapProvider"
         )
-        dataset_args = (
+        dataset_args = cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
-            cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
+        dataloader_args = cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_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"
@@ -94,12 +89,8 @@ class TestExperiment(unittest.TestCase):
         cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_class_type = (
             "JsonIndexDatasetMapProvider"
         )
-        dataset_args = (
+        dataset_args = cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
-            cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
+        dataloader_args = cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_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"
@@ -111,9 +102,7 @@ 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 = (
+        cfg.optimizer_factory_ImplicitronOptimizerFactory_args.exponential_lr_step_size = 2
-            2
-        )
 
         if DEBUG:
             experiment.dump_cfg(cfg)

projects/implicitron_trainer/tests/test_optimizer_factory.py

@@ -81,8 +81,9 @@ 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(
+        na, nb = (
-            params=[pb], param_groups={"self": "pb_self", "p1": "pb_param"}
+            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/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)
+    train_data = torch.load(cameras_path, weights_only=True)
     n_cameras = train_data["cameras"]["R"].shape[0]
 
     _image_max_image_pixels = Image.MAX_IMAGE_PIXELS

projects/nerf/nerf/stats.py

@@ -194,7 +194,6 @@ 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

@@ -24,7 +24,6 @@ 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"
@@ -63,7 +62,7 @@ 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)
+    loaded_data = torch.load(checkpoint_path, weights_only=True)
     # Do not load the cached xy grid.
     # - this allows setting an arbitrary evaluation image size.
     state_dict = {

projects/nerf/tests/test_raysampler.py

@@ -42,7 +42,6 @@ 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

@@ -25,7 +25,6 @@ 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)
@@ -77,7 +76,7 @@ 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)
+            loaded_data = torch.load(checkpoint_path, weights_only=True)
             model.load_state_dict(loaded_data["model"])
             stats = pickle.loads(loaded_data["stats"])
             print(f"   => resuming from epoch {stats.epoch}.")
@@ -219,7 +218,6 @@ 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__ = "0.7.8"
+__version__ = "0.7.9"

pytorch3d/common/compat.py

@@ -17,7 +17,7 @@ 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

pytorch3d/csrc/ball_query/ball_query.cu

@@ -32,7 +32,9 @@ __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 radius2) {
+    const float radius,
+    const float radius2,
+    const bool skip_points_outside_cube) {
   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;
@@ -51,7 +53,19 @@ __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) {
-      // Calculate the distance between the points
+      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
       scalar_t dist2 = 0.0;
       for (int d = 0; d < D; ++d) {
         scalar_t diff = p1[n][i][d] - p2[n][j][d];
@@ -77,7 +91,8 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
     const at::Tensor& lengths1, // (N,)
     const at::Tensor& lengths2, // (N,)
     int K,
-    float radius) {
+    float radius,
+    bool skip_points_outside_cube) {
   // 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};
@@ -120,7 +135,9 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
             idxs.packed_accessor64<int64_t, 3, at::RestrictPtrTraits>(),
             dists.packed_accessor64<float, 3, at::RestrictPtrTraits>(),
             K_64,
-            radius2);
+            radius,
+            radius2,
+            skip_points_outside_cube);
       }));
 
   AT_CUDA_CHECK(cudaGetLastError());

pytorch3d/csrc/ball_query/ball_query.h

@@ -25,6 +25,9 @@
 //      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
@@ -46,7 +49,8 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCpu(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     const int K,
-    const float radius);
+    const float radius,
+    const bool skip_points_outside_cube);
 
 // CUDA implementation
 std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
@@ -55,7 +59,8 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCuda(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     const int K,
-    const float radius);
+    const float radius,
+    const bool skip_points_outside_cube);
 
 // Implementation which is exposed
 // Note: the backward pass reuses the KNearestNeighborBackward kernel
@@ -65,7 +70,8 @@ inline std::tuple<at::Tensor, at::Tensor> BallQuery(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     int K,
-    float radius) {
+    float radius,
+    bool skip_points_outside_cube) {
   if (p1.is_cuda() || p2.is_cuda()) {
 #ifdef WITH_CUDA
     CHECK_CUDA(p1);
@@ -76,16 +82,20 @@ inline std::tuple<at::Tensor, at::Tensor> BallQuery(
         lengths1.contiguous(),
         lengths2.contiguous(),
         K,
-        radius);
+        radius,
+        skip_points_outside_cube);
 #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);
+      radius,
+      skip_points_outside_cube);
 }

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(
@@ -16,7 +16,8 @@ std::tuple<at::Tensor, at::Tensor> BallQueryCpu(
     const at::Tensor& lengths1,
     const at::Tensor& lengths2,
     int K,
-    float radius) {
+    float radius,
+    bool skip_points_outside_cube) {
   const int N = p1.size(0);
   const int P1 = p1.size(1);
   const int D = p1.size(2);
@@ -38,6 +39,16 @@ 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

@@ -98,6 +98,11 @@ 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

@@ -28,17 +28,16 @@ __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 int batch = blockIdx.x;
+  const auto batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const int num_threads = gridDim.y * blockDim.x;
+  const auto num_threads = gridDim.y * blockDim.x;
-  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const auto tid = blockIdx.y * blockDim.x + threadIdx.x;
 
   // Iterate over each feature in each pixel
   for (int pid = tid; pid < num_pixels; pid += num_threads) {
@@ -79,17 +78,16 @@ __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 int batch = blockIdx.x;
+  const auto batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const int num_threads = gridDim.y * blockDim.x;
+  const auto num_threads = gridDim.y * blockDim.x;
-  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const auto 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

@@ -74,6 +74,9 @@ 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);
   }
 }
@@ -101,6 +104,11 @@ 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

@@ -28,17 +28,16 @@ __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 int batch = blockIdx.x;
+  const auto batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const int num_threads = gridDim.y * blockDim.x;
+  const auto num_threads = gridDim.y * blockDim.x;
-  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const auto 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
@@ -92,17 +91,16 @@ __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 int batch = blockIdx.x;
+  const auto batch = blockIdx.x;
 
   const int num_pixels = C * W * H;
-  const int num_threads = gridDim.y * blockDim.x;
+  const auto num_threads = gridDim.y * blockDim.x;
-  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const auto 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

@@ -73,6 +73,10 @@ 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);
   }
 }
@@ -100,6 +104,11 @@ 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

@@ -26,17 +26,16 @@ __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 int batch = blockIdx.x;
+  const auto batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const int num_threads = gridDim.y * blockDim.x;
+  const auto num_threads = gridDim.y * blockDim.x;
-  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const auto 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
@@ -74,17 +73,16 @@ __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 int batch = blockIdx.x;
+  const auto batch = blockIdx.x;
 
   const int num_pixels = C * H * W;
-  const int num_threads = gridDim.y * blockDim.x;
+  const auto num_threads = gridDim.y * blockDim.x;
-  const int tid = blockIdx.y * blockDim.x + threadIdx.x;
+  const auto 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

@@ -72,6 +72,9 @@ 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);
   }
 }
@@ -98,6 +101,11 @@ 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

@@ -7,15 +7,10 @@
  */
 
 // clang-format off
-#if !defined(USE_ROCM)
 #include "./pulsar/global.h" // Include before <torch/extension.h>.
-#endif
-#include <torch/extension.h>
 // clang-format on
-#if !defined(USE_ROCM)
 #include "./pulsar/pytorch/renderer.h"
 #include "./pulsar/pytorch/tensor_util.h"
-#endif
 #include "ball_query/ball_query.h"
 #include "blending/sigmoid_alpha_blend.h"
 #include "compositing/alpha_composite.h"
@@ -104,22 +99,22 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
 
   // Pulsar.
   // Pulsar not enabled on AMD.
-#if !defined(USE_ROCM)
 #ifdef PULSAR_LOGGING_ENABLED
   c10::ShowLogInfoToStderr();
 #endif
   py::class_<
       pulsar::pytorch::Renderer,
       std::shared_ptr<pulsar::pytorch::Renderer>>(m, "PulsarRenderer")
-      .def(py::init<
+      .def(
-           const uint&,
+          py::init<
-           const uint&,
+              const uint&,
-           const uint&,
+              const uint&,
-           const bool&,
+              const uint&,
-           const bool&,
+              const bool&,
-           const float&,
+              const bool&,
-           const uint&,
+              const float&,
-           const uint&>())
+              const uint&,
+              const uint&>())
       .def(
           "__eq__",
           [](const pulsar::pytorch::Renderer& a,
@@ -154,10 +149,10 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
           py::arg("gamma"),
           py::arg("max_depth"),
           py::arg("min_depth") /* = 0.f*/,
-          py::arg(
+          py::arg("bg_col") /* = std::nullopt not exposed properly in
-              "bg_col") /* = at::nullopt not exposed properly in pytorch 1.1. */
+                               pytorch 1.1. */
           ,
-          py::arg("opacity") /* = at::nullopt ... */,
+          py::arg("opacity") /* = std::nullopt ... */,
           py::arg("percent_allowed_difference") = 0.01f,
           py::arg("max_n_hits") = MAX_UINT,
           py::arg("mode") = 0)
@@ -189,5 +184,4 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   m.attr("MAX_UINT") = py::int_(MAX_UINT);
   m.attr("MAX_USHORT") = py::int_(MAX_USHORT);
   m.attr("PULSAR_MAX_GRAD_SPHERES") = py::int_(MAX_GRAD_SPHERES);
-#endif
 }

pytorch3d/csrc/face_areas_normals/face_areas_normals.h

@@ -60,6 +60,8 @@ 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);
 }
 
@@ -80,5 +82,9 @@ 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 int tid = threadIdx.x;
+  const auto 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 (int e = blockIdx.x; e < E; e += gridDim.x) {
+  for (auto 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 (int d = tid; d < D; d += blockDim.x) {
+    for (auto 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

@@ -53,5 +53,7 @@ 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 int tid = threadIdx.x + blockIdx.x * blockDim.x;
+  const auto tid = threadIdx.x + blockIdx.x * blockDim.x;
-  const int num_threads = blockDim.x * gridDim.x;
+  const auto 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 int tid = threadIdx.x + blockIdx.x * blockDim.x;
+  const auto tid = threadIdx.x + blockIdx.x * blockDim.x;
-  const int num_threads = blockDim.x * gridDim.x;
+  const auto 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

@@ -57,6 +57,8 @@ 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);
 }
 
@@ -106,6 +108,9 @@ 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

@@ -44,5 +44,7 @@ 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

@@ -7,10 +7,7 @@
  */
 
 #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

@@ -461,10 +461,8 @@ __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
@@ -500,7 +498,6 @@ __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
@@ -728,7 +725,7 @@ __device__ inline int BoxIntersections(
       }
     }
     // Update the face_verts_out tris
-    num_tris = offset;
+    num_tris = min(MAX_TRIS, offset);
     for (int j = 0; j < num_tris; ++j) {
       face_verts_out[j] = tri_verts_updated[j];
     }

pytorch3d/csrc/knn/knn.h

@@ -74,6 +74,8 @@ 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);
 }
 
@@ -140,6 +142,8 @@ 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

@@ -58,5 +58,6 @@ 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/packed_to_padded_tensor/packed_to_padded_tensor.h

@@ -88,6 +88,8 @@ 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);
 }
 
@@ -105,5 +107,7 @@ 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 as_a = as.accessor<float, H1 == 1 ? 2 : 3>();
-  auto bs_a = bs.accessor < float, H2 == 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 as_a = as.accessor<float, H1 == 1 ? 2 : 3>();
-  auto bs_a = bs.accessor < float, H2 == 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_as_a = grad_as.accessor<float, H1 == 1 ? 2 : 3>();
-  auto grad_bs_a = grad_bs.accessor < float, H2 == 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 (int s = blockDim.x / 2; s > 32; s >>= 1) {
+    for (auto 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 int num_threads = gridDim.x * blockDim.x;
+  const auto num_threads = gridDim.x * blockDim.x;
-  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto 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 int num_threads = gridDim.x * blockDim.x;
+  const auto num_threads = gridDim.x * blockDim.x;
-  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto 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 int num_threads = gridDim.x * blockDim.x;
+  const auto num_threads = gridDim.x * blockDim.x;
-  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto 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 int num_threads = gridDim.x * blockDim.x;
+  const auto num_threads = gridDim.x * blockDim.x;
-  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto 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

@@ -88,6 +88,10 @@ 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);
 }
@@ -143,6 +147,10 @@ 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);
 }
@@ -221,6 +229,10 @@ 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);
 }
@@ -277,6 +289,10 @@ 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);
 }
@@ -346,6 +362,10 @@ 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);
 }
@@ -396,6 +416,10 @@ 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);
 }
 
@@ -464,6 +488,10 @@ 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);
 }
@@ -514,6 +542,10 @@ 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);
 }
 
@@ -567,6 +599,8 @@ torch::Tensor PointFaceArrayDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
+  CHECK_CPU(points);
+  CHECK_CPU(tris);
   return PointFaceArrayDistanceForwardCpu(points, tris, min_triangle_area);
 }
 
@@ -613,6 +647,9 @@ 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);
 }
@@ -661,6 +698,8 @@ torch::Tensor PointEdgeArrayDistanceForward(
     AT_ERROR("Not compiled with GPU support.");
 #endif
   }
+  CHECK_CPU(points);
+  CHECK_CPU(segms);
   return PointEdgeArrayDistanceForwardCpu(points, segms);
 }
 
@@ -703,5 +742,8 @@ 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

@@ -104,6 +104,12 @@ 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,
@@ -183,6 +189,14 @@ 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

@@ -8,9 +8,7 @@
 
 #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)
-#define uint unsigned int
+using uint = unsigned int;
-#define ushort unsigned short
+using ushort = unsigned short;
 #endif
 
 #include "./logging.h" // <- include before torch/extension.h
@@ -36,11 +36,13 @@
 #pragma nv_diag_suppress 2951
 #pragma nv_diag_suppress 2967
 #else
+#if !defined(USE_ROCM)
 #pragma diag_suppress = attribute_not_allowed
 #pragma diag_suppress = 1866
 #pragma diag_suppress = 2941
 #pragma diag_suppress = 2951
 #pragma diag_suppress = 2967
+#endif //! USE_ROCM
 #endif
 #else // __CUDACC__
 #define INLINE inline
@@ -56,7 +58,9 @@
 #pragma clang diagnostic pop
 #ifdef WITH_CUDA
 #include <ATen/cuda/CUDAContext.h>
+#if !defined(USE_ROCM)
 #include <vector_functions.h>
+#endif //! USE_ROCM
 #else
 #ifndef cudaStream_t
 typedef void* cudaStream_t;

pytorch3d/csrc/pulsar/gpu/commands.h

@@ -59,6 +59,11 @@ getLastCudaError(const char* errorMessage, const char* file, const int line) {
 #define SHARED __shared__
 #define ACTIVEMASK() __activemask()
 #define BALLOT(mask, val) __ballot_sync((mask), val)
+
+/* TODO (ROCM-6.2): None of the WARP_* are used anywhere and ROCM-6.2 natively
+ * supports __shfl_*. Disabling until the move to ROCM-6.2.
+ */
+#if !defined(USE_ROCM)
 /**
  * Find the cumulative sum within a warp up to the current
  * thread lane, with each mask thread contributing base.
@@ -115,6 +120,7 @@ INLINE DEVICE float3 WARP_SUM_FLOAT3(
   ret.z = WARP_SUM(group, mask, base.z);
   return ret;
 }
+#endif //! USE_ROCM
 
 // Floating point.
 // #define FMUL(a, b) __fmul_rn((a), (b))
@@ -142,6 +148,7 @@ INLINE DEVICE float3 WARP_SUM_FLOAT3(
 #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;
   int old = *address_as_i, assumed;
@@ -166,6 +173,7 @@ __device__ static float atomicMin(float* address, float val) {
   } while (assumed != old);
   return __int_as_float(old);
 }
+#endif //! USE_ROCM
 #define DMAX(a, b) FMAX(a, b)
 #define DMIN(a, b) FMIN(a, b)
 #define DSQRT(a) sqrt(a)
@@ -409,7 +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

@@ -357,11 +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.device.h

@@ -14,7 +14,7 @@
 #include "./commands.h"
 
 namespace pulsar {
-IHD CamGradInfo::CamGradInfo() {
+IHD CamGradInfo::CamGradInfo(int x) {
   cam_pos = make_float3(0.f, 0.f, 0.f);
   pixel_0_0_center = make_float3(0.f, 0.f, 0.f);
   pixel_dir_x = make_float3(0.f, 0.f, 0.f);

pytorch3d/csrc/pulsar/include/camera.h

@@ -63,18 +63,13 @@ inline bool operator==(const CamInfo& a, const CamInfo& b) {
 };
 
 struct CamGradInfo {
-  HOST DEVICE CamGradInfo();
+  HOST DEVICE CamGradInfo(int = 0);
   float3 cam_pos;
   float3 pixel_0_0_center;
   float3 pixel_dir_x;
   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

@@ -24,7 +24,7 @@
 // #pragma diag_suppress = 68
 #include <ATen/cuda/CUDAContext.h>
 // #pragma pop
-#include "../cuda/commands.h"
+#include "../gpu/commands.h"
 #else
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Weverything"

pytorch3d/csrc/pulsar/include/math.h

@@ -46,6 +46,7 @@ IHD float3 outer_product_sum(const float3& a) {
 }
 
 // TODO: put intrinsics here.
+#if !defined(USE_ROCM)
 IHD float3 operator+(const float3& a, const float3& b) {
   return make_float3(a.x + b.x, a.y + b.y, a.z + b.z);
 }
@@ -93,6 +94,7 @@ IHD float3 operator*(const float3& a, const float3& b) {
 IHD float3 operator*(const float& a, const float3& b) {
   return b * a;
 }
+#endif //! USE_ROCM
 
 INLINE DEVICE float length(const float3& v) {
   // TODO: benchmark what's faster.
@@ -147,11 +149,6 @@ 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(
-        (IntWrapper*)(self->ids_sorted_d),
+        self->ids_sorted_d,
-        (IntWrapper*)(self->n_grad_contributions_d),
+        self->n_grad_contributions_d,
         static_cast<int>(num_balls),
         self->workspace_d,
         self->workspace_size,

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

@@ -18,68 +18,89 @@ 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.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

@@ -64,8 +64,9 @@ 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

@@ -139,8 +139,9 @@ 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;
     }
@@ -190,20 +191,22 @@ 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.
@@ -283,9 +286,16 @@ GLOBAL void render(
           (percent_allowed_difference > 0.f &&
            max_closest_possible_intersection > depth_threshold) ||
           tracker.get_n_hits() >= max_n_hits;
+#if defined(__CUDACC__) && defined(__HIP_PLATFORM_AMD__)
+      unsigned long long warp_done = __ballot(done);
+      int warp_done_bit_cnt = __popcll(warp_done);
+#else
       uint warp_done = thread_warp.ballot(done);
-      if (thread_warp.thread_rank() == 0)
+      int warp_done_bit_cnt = POPC(warp_done);
-        ATOMICADD_B(&n_pixels_done, POPC(warp_done));
+#endif //__CUDACC__ && __HIP_PLATFORM_AMD__
+      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();
@@ -293,8 +303,9 @@ 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",
@@ -380,8 +391,9 @@ 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);

pytorch3d/csrc/pulsar/pytorch/renderer.cpp

@@ -213,8 +213,8 @@ std::tuple<size_t, size_t, bool, torch::Tensor> Renderer::arg_check(
     const float& gamma,
     const float& max_depth,
     float& min_depth,
-    const c10::optional<torch::Tensor>& bg_col,
+    const std::optional<torch::Tensor>& bg_col,
-    const c10::optional<torch::Tensor>& opacity,
+    const std::optional<torch::Tensor>& opacity,
     const float& percent_allowed_difference,
     const uint& max_n_hits,
     const uint& mode) {
@@ -668,8 +668,8 @@ std::tuple<torch::Tensor, torch::Tensor> Renderer::forward(
     const float& gamma,
     const float& max_depth,
     float min_depth,
-    const c10::optional<torch::Tensor>& bg_col,
+    const std::optional<torch::Tensor>& bg_col,
-    const c10::optional<torch::Tensor>& opacity,
+    const std::optional<torch::Tensor>& opacity,
     const float& percent_allowed_difference,
     const uint& max_n_hits,
     const uint& mode) {
@@ -860,8 +860,9 @@ 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,
@@ -888,14 +889,14 @@ std::tuple<torch::Tensor, torch::Tensor> Renderer::forward(
 };
 
 std::tuple<
-    at::optional<torch::Tensor>,
+    std::optional<torch::Tensor>,
-    at::optional<torch::Tensor>,
+    std::optional<torch::Tensor>,
-    at::optional<torch::Tensor>,
+    std::optional<torch::Tensor>,
-    at::optional<torch::Tensor>,
+    std::optional<torch::Tensor>,
-    at::optional<torch::Tensor>,
+    std::optional<torch::Tensor>,
-    at::optional<torch::Tensor>,
+    std::optional<torch::Tensor>,
-    at::optional<torch::Tensor>,
+    std::optional<torch::Tensor>,
-    at::optional<torch::Tensor>>
+    std::optional<torch::Tensor>>
 Renderer::backward(
     const torch::Tensor& grad_im,
     const torch::Tensor& image,
@@ -912,8 +913,8 @@ Renderer::backward(
     const float& gamma,
     const float& max_depth,
     float min_depth,
-    const c10::optional<torch::Tensor>& bg_col,
+    const std::optional<torch::Tensor>& bg_col,
-    const c10::optional<torch::Tensor>& opacity,
+    const std::optional<torch::Tensor>& opacity,
     const float& percent_allowed_difference,
     const uint& max_n_hits,
     const uint& mode,
@@ -922,7 +923,7 @@ Renderer::backward(
     const bool& dif_rad,
     const bool& dif_cam,
     const bool& dif_opy,
-    const at::optional<std::pair<uint, uint>>& dbg_pos) {
+    const std::optional<std::pair<uint, uint>>& dbg_pos) {
   this->ensure_on_device(this->device_tracker.device());
   size_t batch_size;
   size_t n_points;
@@ -1045,14 +1046,14 @@ Renderer::backward(
   }
   // Prepare the return value.
   std::tuple<
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>>
+      std::optional<torch::Tensor>>
       ret;
   if (mode == 1 || (!dif_pos && !dif_col && !dif_rad && !dif_cam && !dif_opy)) {
     return ret;

pytorch3d/csrc/pulsar/pytorch/renderer.h

@@ -44,21 +44,21 @@ struct Renderer {
       const float& gamma,
       const float& max_depth,
       float min_depth,
-      const c10::optional<torch::Tensor>& bg_col,
+      const std::optional<torch::Tensor>& bg_col,
-      const c10::optional<torch::Tensor>& opacity,
+      const std::optional<torch::Tensor>& opacity,
       const float& percent_allowed_difference,
       const uint& max_n_hits,
       const uint& mode);
 
   std::tuple<
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>,
+      std::optional<torch::Tensor>,
-      at::optional<torch::Tensor>>
+      std::optional<torch::Tensor>>
   backward(
       const torch::Tensor& grad_im,
       const torch::Tensor& image,
@@ -75,8 +75,8 @@ struct Renderer {
       const float& gamma,
       const float& max_depth,
       float min_depth,
-      const c10::optional<torch::Tensor>& bg_col,
+      const std::optional<torch::Tensor>& bg_col,
-      const c10::optional<torch::Tensor>& opacity,
+      const std::optional<torch::Tensor>& opacity,
       const float& percent_allowed_difference,
       const uint& max_n_hits,
       const uint& mode,
@@ -85,7 +85,7 @@ struct Renderer {
       const bool& dif_rad,
       const bool& dif_cam,
       const bool& dif_opy,
-      const at::optional<std::pair<uint, uint>>& dbg_pos);
+      const std::optional<std::pair<uint, uint>>& dbg_pos);
 
   // Infrastructure.
   /**
@@ -115,8 +115,8 @@ struct Renderer {
       const float& gamma,
       const float& max_depth,
       float& min_depth,
-      const c10::optional<torch::Tensor>& bg_col,
+      const std::optional<torch::Tensor>& bg_col,
-      const c10::optional<torch::Tensor>& opacity,
+      const std::optional<torch::Tensor>& opacity,
       const float& percent_allowed_difference,
       const uint& max_n_hits,
       const uint& mode);
@@ -128,8 +128,9 @@ 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

@@ -8,6 +8,7 @@
 
 #ifdef WITH_CUDA
 #include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAException.h>
 #include <cuda_runtime_api.h>
 #endif
 #include <torch/extension.h>
@@ -33,13 +34,13 @@ torch::Tensor sphere_ids_from_result_info_nograd(
           .contiguous();
   if (forw_info.device().type() == c10::DeviceType::CUDA) {
 #ifdef WITH_CUDA
-    cudaMemcpyAsync(
+    C10_CUDA_CHECK(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

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

pytorch3d/csrc/pulsar/warnings.cpp

@@ -6,9 +6,6 @@
  * 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 (int i = threadIdx.x; i < H * W * D; i += blockDim.x) {
+    for (auto 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 int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
-  const int num_threads = blockDim.x * gridDim.x;
+  const auto 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 int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
-  const int num_threads = blockDim.x * gridDim.x;
+  const auto 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 (int chunk = blockIdx.x; chunk < num_chunks; chunk += gridDim.x) {
+  for (auto 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 (int e = threadIdx.x; e < chunk_size; e += blockDim.x) {
+    for (auto 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 (int byx = threadIdx.x; byx < num_bins_y * num_bins_x;
+    for (auto 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
-  int num_threads = gridDim.x * blockDim.x;
+  auto num_threads = gridDim.x * blockDim.x;
-  int tid = blockDim.x * blockIdx.x + threadIdx.x;
+  auto 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 int num_threads = gridDim.x * blockDim.x;
+  const auto num_threads = gridDim.x * blockDim.x;
-  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto 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;
-  int num_threads = gridDim.x * blockDim.x;
+  auto num_threads = gridDim.x * blockDim.x;
-  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  auto 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

@@ -138,6 +138,9 @@ 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,
@@ -232,6 +235,11 @@ 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,
@@ -306,6 +314,9 @@ 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,
@@ -423,6 +434,8 @@ 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

@@ -9,7 +9,6 @@
 #include <torch/extension.h>
 #include <algorithm>
 #include <list>
-#include <queue>
 #include <thread>
 #include <tuple>
 #include "ATen/core/TensorAccessor.h"
@@ -107,6 +106,8 @@ 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 {
@@ -119,7 +120,6 @@ 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 int num_threads = gridDim.x * blockDim.x;
+  const auto num_threads = gridDim.x * blockDim.x;
-  const int tid = blockDim.x * blockIdx.x + threadIdx.x;
+  const auto 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 int num_threads = gridDim.x * blockDim.x;
+  const auto num_threads = gridDim.x * blockDim.x;
-  const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const auto 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.
-  int num_threads = gridDim.x * blockDim.x;
+  auto num_threads = gridDim.x * blockDim.x;
-  int tid = blockIdx.x * blockDim.x + threadIdx.x;
+  auto 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

@@ -91,6 +91,10 @@ 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,
@@ -166,6 +170,10 @@ 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,
@@ -232,6 +240,8 @@ 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");
   }
 }
@@ -284,6 +294,10 @@ 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

@@ -35,8 +35,6 @@ __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
@@ -109,7 +107,8 @@ 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 at::Tensor& start_idxs,
+    const int64_t max_K_known = -1) {
   // 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};
@@ -131,7 +130,12 @@ at::Tensor FarthestPointSamplingCuda(
 
   const int64_t N = points.size(0);
   const int64_t P = points.size(1);
-  const int64_t max_K = at::max(K).item<int64_t>();
+  int64_t max_K;
+  if (max_K_known > 0) {
+    max_K = max_K_known;
+  } else {
+    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

@@ -43,7 +43,8 @@ at::Tensor FarthestPointSamplingCuda(
     const at::Tensor& points,
     const at::Tensor& lengths,
     const at::Tensor& K,
-    const at::Tensor& start_idxs);
+    const at::Tensor& start_idxs,
+    const int64_t max_K_known = -1);
 
 at::Tensor FarthestPointSamplingCpu(
     const at::Tensor& points,
@@ -56,17 +57,23 @@ at::Tensor FarthestPointSampling(
     const at::Tensor& points,
     const at::Tensor& lengths,
     const at::Tensor& K,
-    const at::Tensor& start_idxs) {
+    const at::Tensor& start_idxs,
+    const int64_t max_K_known = -1) {
   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);
+    return FarthestPointSamplingCuda(
+        points, lengths, K, start_idxs, max_K_known);
 #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

@@ -71,6 +71,8 @@ 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

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

pytorch3d/csrc/utils/geometry_utils.cuh

@@ -376,8 +376,6 @@ 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

@@ -15,3 +15,7 @@
 #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/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/utils.py

@@ -36,7 +36,6 @@ def collate_batched_meshes(batch: List[Dict]):  # pragma: no cover
 
     collated_dict["mesh"] = None
     if {"verts", "faces"}.issubset(collated_dict.keys()):
-
         textures = None
         if "textures" in collated_dict:
             textures = TexturesAtlas(atlas=collated_dict["textures"])

pytorch3d/implicitron/dataset/dataset_base.py

@@ -21,7 +21,6 @@ from typing import (
 )
 
 import torch
-
 from pytorch3d.implicitron.dataset.frame_data import FrameData
 from pytorch3d.implicitron.dataset.utils import GenericWorkaround