pulsar build and CI changes

Summary:
Changes to CI and some minor fixes now that pulsar is part of pytorch3d. Most significantly, add CUB to CI builds.

Make CUB_HOME override the CUB already in cudatoolkit (important for cuda11.0 which uses cub 1.9.9 which pulsar doesn't work well with.
Make imageio available for testing.
Lint fixes.
Fix some test verbosity.
Avoid use of atomicAdd_block on older GPUs.

Reviewed By: nikhilaravi, classner

Differential Revision: D24773716

fbshipit-source-id: 2428356bb2e62735f2bc0c15cbe4cff35b1b24b8

.circleci/config.in.yml

@@ -68,14 +68,26 @@ jobs:
     steps:
       - checkout
       - <<: *setupcuda
-      - run: pip3 install --progress-bar off wheel matplotlib 'pillow<7'
+      - run: pip3 install --progress-bar off imageio wheel matplotlib 'pillow<7'
       - run: pip3 install --progress-bar off torch torchvision
       # - run: conda create -p ~/conda_env python=3.7 numpy
       # - run: conda activate ~/conda_env
       # - run: conda install -c pytorch pytorch torchvision
 
       - run: pip3 install --progress-bar off 'git+https://github.com/facebookresearch/fvcore'
-      - run: LD_LIBRARY_PATH=$LD_LIBARY_PATH:/usr/local/cuda-10.2/lib64 python3 setup.py build_ext --inplace
+      - run:
+          name: get cub
+          command: |
+            cd ..
+            wget --no-verbose https://github.com/NVIDIA/cub/archive/1.10.0.tar.gz
+            tar xzf 1.10.0.tar.gz
+            # This expands to a directory called cub-1.10.0
+      - run:
+          name: build
+          command: |
+            export LD_LIBRARY_PATH=$LD_LIBARY_PATH:/usr/local/cuda-10.2/lib64
+            export CUB_HOME=$(realpath ../cub-1.10.0)
+            python3 setup.py build_ext --inplace
       - run: LD_LIBRARY_PATH=$LD_LIBARY_PATH:/usr/local/cuda-10.2/lib64 python -m unittest discover -v -s tests
       - run: python3 setup.py bdist_wheel
 
@@ -89,7 +101,7 @@ jobs:
     resource_class: 2xlarge+
     steps:
       - checkout
-      - run: packaging/build_wheel.sh
+      - run: MAX_JOBS=15 packaging/build_wheel.sh
       - store_artifacts:
           path: dist
       - persist_to_workspace:
@@ -109,7 +121,7 @@ jobs:
       - checkout
       # This is building with cuda but no gpu present,
       # so we aren't running the tests.
-      - run: TEST_FLAG=--no-test packaging/build_conda.sh
+      - run: MAX_JOBS=15 TEST_FLAG=--no-test packaging/build_conda.sh
       - store_artifacts:
           path: /opt/conda/conda-bld/linux-64
       - persist_to_workspace:
@@ -215,9 +227,9 @@ workflows:
           context: DOCKERHUB_TOKEN
       {{workflows()}}
       - binary_linux_conda_cuda:
-          name: testrun_conda_cuda_py37_cu101_pyt14
+          name: testrun_conda_cuda_py36_cu101_pyt14
           context: DOCKERHUB_TOKEN
-          python_version: "3.7"
+          python_version: "3.6"
           pytorch_version: "1.4"
           cu_version: "cu101"
       - binary_linux_conda_cuda:

.circleci/config.yml

@@ -68,14 +68,26 @@ jobs:
     steps:
       - checkout
       - <<: *setupcuda
-      - run: pip3 install --progress-bar off wheel matplotlib 'pillow<7'
+      - run: pip3 install --progress-bar off imageio wheel matplotlib 'pillow<7'
       - run: pip3 install --progress-bar off torch torchvision
       # - run: conda create -p ~/conda_env python=3.7 numpy
       # - run: conda activate ~/conda_env
       # - run: conda install -c pytorch pytorch torchvision
 
       - run: pip3 install --progress-bar off 'git+https://github.com/facebookresearch/fvcore'
-      - run: LD_LIBRARY_PATH=$LD_LIBARY_PATH:/usr/local/cuda-10.2/lib64 python3 setup.py build_ext --inplace
+      - run:
+          name: get cub
+          command: |
+            cd ..
+            wget --no-verbose https://github.com/NVIDIA/cub/archive/1.10.0.tar.gz
+            tar xzf 1.10.0.tar.gz
+            # This expands to a directory called cub-1.10.0
+      - run:
+          name: build
+          command: |
+            export LD_LIBRARY_PATH=$LD_LIBARY_PATH:/usr/local/cuda-10.2/lib64
+            export CUB_HOME=$(realpath ../cub-1.10.0)
+            python3 setup.py build_ext --inplace
       - run: LD_LIBRARY_PATH=$LD_LIBARY_PATH:/usr/local/cuda-10.2/lib64 python -m unittest discover -v -s tests
       - run: python3 setup.py bdist_wheel
 
@@ -89,7 +101,7 @@ jobs:
     resource_class: 2xlarge+
     steps:
       - checkout
-      - run: packaging/build_wheel.sh
+      - run: MAX_JOBS=15 packaging/build_wheel.sh
       - store_artifacts:
           path: dist
       - persist_to_workspace:
@@ -109,7 +121,7 @@ jobs:
       - checkout
       # This is building with cuda but no gpu present,
       # so we aren't running the tests.
-      - run: TEST_FLAG=--no-test packaging/build_conda.sh
+      - run: MAX_JOBS=15 TEST_FLAG=--no-test packaging/build_conda.sh
       - store_artifacts:
           path: /opt/conda/conda-bld/linux-64
       - persist_to_workspace:
@@ -489,6 +501,12 @@ workflows:
           python_version: "3.6"
           pytorch_version: "1.4"
           cu_version: "cu101"
+      - binary_linux_conda_cuda:
+          name: testrun_conda_cuda_py37_cu102_pyt160
+          context: DOCKERHUB_TOKEN
+          python_version: "3.7"
+          pytorch_version: '1.6.0'
+          cu_version: "cu102"
       - binary_linux_conda_cuda:
           name: testrun_conda_cuda_py37_cu110_pyt170
           context: DOCKERHUB_TOKEN

docs/examples/pulsar_basic.py

@@ -5,9 +5,9 @@ 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
-import logging
 
 import imageio
 import torch

docs/examples/pulsar_cam_unified.py

@@ -9,8 +9,8 @@ distorted. Gradient-based optimization is used to converge towards the
 original camera parameters.
 Output: cam-pt3d.gif
 """
-from os import path
 import logging
+from os import path
 
 import cv2
 import imageio

docs/examples/pulsar_multiview.py

@@ -13,9 +13,9 @@ 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
-import logging
 
 import cv2
 import imageio

docs/examples/pulsar_optimization.py

@@ -8,8 +8,8 @@ The scene is initialized with random spheres. Gradient-based
 optimization is used to converge towards a faithful
 scene representation.
 """
-import math
 import logging
+import math
 
 import cv2
 import imageio

docs/examples/pulsar_optimization_unified.py

@@ -8,8 +8,8 @@ The scene is initialized with random spheres. Gradient-based
 optimization is used to converge towards a faithful
 scene representation.
 """
-import math
 import logging
+import math
 
 import cv2
 import imageio

packaging/build_conda.sh

@@ -17,4 +17,4 @@ setup_conda_pytorch_constraint
 setup_conda_cudatoolkit_constraint
 setup_visual_studio_constraint
 # shellcheck disable=SC2086
-conda build $CONDA_CHANNEL_FLAGS ${TEST_FLAG:-} -c defaults -c conda-forge --no-anaconda-upload -c fvcore --python "$PYTHON_VERSION" packaging/pytorch3d
+conda build $CONDA_CHANNEL_FLAGS ${TEST_FLAG:-} -c bottler -c defaults -c conda-forge --no-anaconda-upload -c fvcore --python "$PYTHON_VERSION" packaging/pytorch3d

packaging/build_wheel.sh

@@ -12,5 +12,6 @@ setup_env "$VERSION"
 setup_wheel_python
 pip_install numpy
 setup_pip_pytorch_version
+download_nvidiacub_if_needed
 python setup.py clean
 IS_WHEEL=1 python setup.py bdist_wheel

packaging/pkg_helpers.bash

@@ -251,24 +251,32 @@ setup_conda_pytorch_constraint() {
 # Translate CUDA_VERSION into CUDA_CUDATOOLKIT_CONSTRAINT
 setup_conda_cudatoolkit_constraint() {
   export CONDA_CPUONLY_FEATURE=""
+  export CONDA_CUB_CONSTRAINT=""
   if [[ "$(uname)" == Darwin ]]; then
     export CONDA_CUDATOOLKIT_CONSTRAINT=""
   else
     case "$CU_VERSION" in
       cu110)
         export CONDA_CUDATOOLKIT_CONSTRAINT="- cudatoolkit >=11.0,<11.1 # [not osx]"
+        # Even though cudatoolkit 11.0 provides CUB we need our own, to control the
+        # version, because the built-in 1.9.9 in the cudatoolkit causes problems.
+        export CONDA_CUB_CONSTRAINT="- nvidiacub"
         ;;
       cu102)
         export CONDA_CUDATOOLKIT_CONSTRAINT="- cudatoolkit >=10.2,<10.3 # [not osx]"
+        export CONDA_CUB_CONSTRAINT="- nvidiacub"
         ;;
       cu101)
         export CONDA_CUDATOOLKIT_CONSTRAINT="- cudatoolkit >=10.1,<10.2 # [not osx]"
+        export CONDA_CUB_CONSTRAINT="- nvidiacub"
         ;;
       cu100)
         export CONDA_CUDATOOLKIT_CONSTRAINT="- cudatoolkit >=10.0,<10.1 # [not osx]"
+        export CONDA_CUB_CONSTRAINT="- nvidiacub"
         ;;
       cu92)
         export CONDA_CUDATOOLKIT_CONSTRAINT="- cudatoolkit >=9.2,<9.3 # [not osx]"
+        export CONDA_CUB_CONSTRAINT="- nvidiacub"
         ;;
       cpu)
         export CONDA_CUDATOOLKIT_CONSTRAINT=""
@@ -292,3 +300,17 @@ setup_visual_studio_constraint() {
       cp packaging/$VSTOOLCHAIN_PACKAGE/conda_build_config.yaml packaging/pytorch3d/conda_build_config.yaml
   fi
 }
+
+download_nvidiacub_if_needed() {
+  case "$CU_VERSION" in
+    cu110|cu102|cu101|cu100|cu92)
+      echo "Downloading cub"
+      wget --no-verbose https://github.com/NVIDIA/cub/archive/1.10.0.tar.gz
+      tar xzf 1.10.0.tar.gz
+      CUB_HOME=$(realpath ./cub-1.10.0)
+      export CUB_HOME
+      echo "CUB_HOME is now $CUB_HOME"
+      ;;
+  esac
+  # We don't need CUB for a cpu build or if cuda is 11.1 or higher
+}

packaging/pytorch3d/meta.yaml

@@ -8,6 +8,7 @@ source:
 requirements:
   build:
     - {{ compiler('c') }} # [win]
+    {{ environ.get('CONDA_CUB_CONSTRAINT') }}
 
   host:
     - python
@@ -31,6 +32,7 @@ build:
     - CUDA_HOME
     - FORCE_CUDA
     - NVCC_FLAGS
+    - MAX_JOBS
   features:
     {{ environ.get('CONDA_CPUONLY_FEATURE') }}
 
@@ -41,7 +43,7 @@ test:
     - tests
     - docs
   requires:
-    - ca-certificates
+    - imageio
   commands:
     #pytest .
     python -m unittest discover -v -s tests

pytorch3d/csrc/pulsar/cuda/commands.h

@@ -186,7 +186,7 @@ __device__ static float atomicMin(float* address, float val) {
   ATOMICADD(&((PTR)->x), VAL.x); \
   ATOMICADD(&((PTR)->y), VAL.y); \
   ATOMICADD(&((PTR)->z), VAL.z);
-#if (CUDART_VERSION >= 10000)
+#if (CUDART_VERSION >= 10000) && (__CUDA_ARCH__ >= 600)
 #define ATOMICADD_B(PTR, VAL) atomicAdd_block((PTR), (VAL))
 #else
 #define ATOMICADD_B(PTR, VAL) ATOMICADD(PTR, VAL)

setup.py

@@ -20,12 +20,18 @@ def get_extensions():
 
     extra_compile_args = {"cxx": ["-std=c++14"]}
     define_macros = []
+    include_dirs = [extensions_dir]
 
     force_cuda = os.getenv("FORCE_CUDA", "0") == "1"
     if (torch.cuda.is_available() and CUDA_HOME is not None) or force_cuda:
         extension = CUDAExtension
         sources += source_cuda
         define_macros += [("WITH_CUDA", None)]
+        # Thrust is only used for its tuple objects.
+        # With CUDA 11.0 we can't use the cudatoolkit's version of cub.
+        # We take the risk that CUB and Thrust are incompatible, because
+        # we aren't using parts of Thrust which actually use CUB.
+        define_macros += [("THRUST_IGNORE_CUB_VERSION_CHECK", None)]
         cub_home = os.environ.get("CUB_HOME", None)
         nvcc_args = [
             "-std=c++14",
@@ -34,6 +40,11 @@ def get_extensions():
             "-D__CUDA_NO_HALF_CONVERSIONS__",
             "-D__CUDA_NO_HALF2_OPERATORS__",
         ]
+        if cub_home is None:
+            prefix = os.environ.get("CONDA_PREFIX", None)
+            if prefix is not None and os.path.isdir(prefix + "/include/cub"):
+                cub_home = prefix + "/include"
+
         if cub_home is None:
             warnings.warn(
                 "The environment variable `CUB_HOME` was not found. "
@@ -43,14 +54,13 @@ def get_extensions():
                 "`CUB_HOME` to the folder containing the `CMakeListst.txt` file."
             )
         else:
-            nvcc_args.insert(
+            include_dirs.append(os.path.realpath(cub_home).replace("\\ ", " "))
-                0, "-I%s" % (os.path.realpath(cub_home).replace("\\ ", " "))
-            )
         nvcc_flags_env = os.getenv("NVCC_FLAGS", "")
         if nvcc_flags_env != "":
             nvcc_args.extend(nvcc_flags_env.split(" "))
 
-        # It's better if pytorch can do this by default ..
+        # This is needed for pytorch 1.6 and earlier. See e.g.
+        # https://github.com/facebookresearch/pytorch3d/issues/436
         CC = os.environ.get("CC", None)
         if CC is not None:
             CC_arg = "-ccbin={}".format(CC)
@@ -63,8 +73,6 @@ def get_extensions():
 
     sources = [os.path.join(extensions_dir, s) for s in sources]
 
-    include_dirs = [extensions_dir]
-
     ext_modules = [
         extension(
             "pytorch3d._C",
@@ -100,7 +108,7 @@ setup(
     url="https://github.com/facebookresearch/pytorch3d",
     description="PyTorch3D is FAIR's library of reusable components "
     "for deep Learning with 3D data.",
-    packages=find_packages(exclude=("configs", "tests")),
+    packages=find_packages(exclude=("configs", "tests", "tests.*")),
     install_requires=["torchvision>=0.4", "fvcore"],
     extras_require={
         "all": ["matplotlib", "tqdm>4.29.0", "imageio", "ipywidgets"],

tests/test_build.py

@@ -41,7 +41,6 @@ class TestBuild(unittest.TestCase):
 
         for extension in extensions:
             for i in root_dir.glob(f"**/*.{extension}"):
-                print(i)
                 if str(i).endswith(
                     "pytorch3d/transforms/external/kornia_angle_axis_to_rotation_matrix.py"
                 ):