Fix packaging (#2121)

Summary:
- Disable problematic tests on OSX.
- Ensure compiler compatibility with CUDA builds.
- Fix path for Python extension libraries.
- Use CentOS for CUDA packaging.
- Update CUDA versions in CI (10.2 and 11.3).

Pull Request resolved: #2121

Reviewed By: mdouze

Differential Revision: D32921117

Pulled By: beauby

fbshipit-source-id: 588c18add8084b8228ff5abc651eaa4567919cc6

Author: beauby

.circleci/config.yml

@@ -221,7 +221,7 @@ jobs:
   build_linux_gpu:
     machine:
       resource_class: gpu.nvidia.medium
-      image: ubuntu-1604-cuda-10.1:201909-23
+      image: ubuntu-2004-cuda-11.4:202110-01
       docker_layer_caching: true
     steps:
       - checkout
@@ -266,6 +266,8 @@ jobs:
         type: string
       cuda_archs:
         type: string
+      compiler_version:
+        type: string
     machine:
       resource_class: gpu.nvidia.medium
       image: ubuntu-1604-cuda-10.1:201909-23
@@ -281,7 +283,7 @@ jobs:
               -e CUDA_ARCHS="<<parameters.cuda_archs>>" \
               -e ANACONDA_API_TOKEN=$ANACONDA_API_TOKEN \
               faiss \
-              conda build faiss-gpu --variants '{ "cudatoolkit": "<<parameters.cuda>>" }' \
+              conda build faiss-gpu --variants '{ "cudatoolkit": "<<parameters.cuda>>", "c_compiler_version": "<<parameters.compiler_version>>", "cxx_compiler_version": "<<parameters.compiler_version>>" }' \
                 --user pytorch --label <<parameters.label>>
           no_output_timeout: 60m
 
@@ -368,28 +370,11 @@ workflows:
               only: /^v.*/
             branches:
               ignore: /.*/
-      - deploy_linux_gpu:
-          name: Linux GPU packages (CUDA 10.1)
-          cuda: "10.1"
-          cuda_archs: "35;52;60;61;70;72;75"
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
       - deploy_linux_gpu:
           name: Linux GPU packages (CUDA 10.2)
           cuda: "10.2"
           cuda_archs: "35;52;60;61;70;72;75"
-          filters:
-            tags:
-              only: /^v.*/
-            branches:
-              ignore: /.*/
-      - deploy_linux_gpu:
-          name: Linux GPU packages (CUDA 11.0)
-          cuda: "11.0"
-          cuda_archs: "60;61;70;72;75;80"
+          compiler_version: "8.4"
           filters:
             tags:
               only: /^v.*/
@@ -399,6 +384,7 @@ workflows:
           name: Linux GPU packages (CUDA 11.3)
           cuda: "11.3"
           cuda_archs: "60;61;70;72;75;80;86"
+          compiler_version: "9.3"
           filters:
             tags:
               only: /^v.*/
@@ -431,25 +417,17 @@ workflows:
       - deploy_linux:
           name: Linux nightlies
           label: nightly
-      - deploy_linux_gpu:
-          name: Linux GPU nightlies (CUDA 10.1)
-          cuda: "10.1"
-          cuda_archs: "35;52;60;61;70;72;75"
-          label: nightly
       - deploy_linux_gpu:
           name: Linux GPU nightlies (CUDA 10.2)
           cuda: "10.2"
           cuda_archs: "35;52;60;61;70;72;75"
-          label: nightly
-      - deploy_linux_gpu:
-          name: Linux GPU nightlies (CUDA 11.0)
-          cuda: "11.0"
-          cuda_archs: "60;61;70;72;75;80"
+          compiler_version: "8.4"
           label: nightly
       - deploy_linux_gpu:
           name: Linux GPU nightlies (CUDA 11.3)
           cuda: "11.3"
           cuda_archs: "60;61;70;72;75;80;86"
+          compiler_version: "9.3"
           label: nightly
       - deploy_windows:
           name: Windows nightlies

conda/Dockerfile.cuda10.1

@@ -3,9 +3,9 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
-FROM nvidia/cuda:10.2-devel-ubuntu18.04
+FROM nvidia/cuda:10.2-devel-centos8
 
-RUN apt-get update && apt-get install -y wget git libcublas-dev
+RUN yum install -y wget git libcublas-devel-10-2
 
 RUN wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
         bash Miniconda3-latest-Linux-x86_64.sh -b -p ~/miniconda3

conda/Dockerfile.cuda10.2

@@ -3,9 +3,9 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
-FROM nvidia/cuda:11.3.0-devel-ubuntu18.04
+FROM nvidia/cuda:11.3.1-devel-centos8
 
-RUN apt-get update && apt-get install -y wget git libcublas-dev-11-3
+RUN yum install -y wget git libcublas-devel-11-3
 
 RUN wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh && \
         bash Miniconda3-latest-Linux-x86_64.sh -b -p ~/miniconda3

conda/Dockerfile.cuda11.0

@@ -7,7 +7,7 @@
 set -e
 
 
-# Build libfaiss_avx2.so.
+# Build libfaiss.so/libfaiss_avx2.so.
 cmake -B _build \
       -DBUILD_SHARED_LIBS=ON \
       -DBUILD_TESTING=OFF \
@@ -16,10 +16,10 @@ cmake -B _build \
       -DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHS}" \
       -DFAISS_ENABLE_PYTHON=OFF \
       -DBLA_VENDOR=Intel10_64lp \
+      -DCMAKE_INSTALL_LIBDIR=lib \
       -DCMAKE_BUILD_TYPE=Release .
 
-make -C _build -j $CPU_COUNT faiss
-make -C _build -j $CPU_COUNT faiss_avx2
+make -C _build -j $CPU_COUNT faiss faiss_avx2
 
 cmake --install _build --prefix $PREFIX
 cmake --install _build --prefix _libfaiss_stage/

conda/Dockerfile.cuda11.3

@@ -4,8 +4,8 @@
 # LICENSE file in the root directory of this source tree.
 
 """this is a basic test script for simple indices work"""
-from __future__ import absolute_import, division, print_function, unicode_literals
 
+import os
 import sys
 import numpy as np
 import unittest
@@ -340,6 +340,8 @@ def test_hnsw(self):
 
 class TestReplicasAndShards(unittest.TestCase):
 
+    @unittest.skipIf(os.name == "posix" and os.uname().sysname == "Darwin",
+                     "There is a bug in the OpenMP implementation on OSX.")
     def test_replicas(self):
         d = 32
         nq = 100
@@ -352,13 +354,6 @@ def test_replicas(self):
 
         Dref, Iref = index_ref.search(xq, 10)
 
-        # there is a OpenMP bug in this configuration, so disable threading
-        if sys.platform == "darwin" and "Clang 12" in sys.version:
-            nthreads = faiss.omp_get_max_threads()
-            faiss.omp_set_num_threads(1)
-        else:
-            nthreads = None
-
         nrep = 5
         index = faiss.IndexBinaryReplicas()
         for _i in range(nrep):
@@ -379,9 +374,6 @@ def test_replicas(self):
         index2.add(xb)
         D2, I2 = index2.search(xq, 10)
 
-        if nthreads is not None:
-            faiss.omp_set_num_threads(nthreads)
-
         self.assertTrue((Dref == D2).all())
         self.assertTrue((Iref == I2).all())
 

conda/faiss-gpu/build-lib.sh

@@ -3,10 +3,7 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
-from __future__ import absolute_import, division, print_function, unicode_literals
-
-# translation of test_meta_index.lua
-
+import os
 import sys
 import numpy as np
 import faiss
@@ -78,6 +75,8 @@ def test_id_remap_ivf(self):
 
 class Shards(unittest.TestCase):
 
+    @unittest.skipIf(os.name == "posix" and os.uname().sysname == "Darwin",
+                     "There is a bug in the OpenMP implementation on OSX.")
     def test_shards(self):
         k = 32
         ref_index = faiss.IndexFlatL2(d)
@@ -87,13 +86,6 @@ def test_shards(self):
         _Dref, Iref = ref_index.search(xq, k)
         print(Iref[:5, :6])
 
-        # there is a OpenMP bug in this configuration, so disable threading
-        if sys.platform == "darwin" and "Clang 12" in sys.version:
-            nthreads = faiss.omp_get_max_threads()
-            faiss.omp_set_num_threads(1)
-        else:
-            nthreads = None
-
         shard_index = faiss.IndexShards(d)
         shard_index_2 = faiss.IndexShards(d, True, False)
 
@@ -139,8 +131,6 @@ def test_shards(self):
             print('%d / %d differences' % (ndiff, nq * k))
             assert(ndiff < nq * k / 1000.)
 
-        if nthreads is not None:
-            faiss.omp_set_num_threads(nthreads)
 
 class Merge(unittest.TestCase):