diff --git a/.github/workflows/presubmit.yml b/.github/workflows/presubmit.yml
index de74f1605..0b48dbe58 100644
--- a/.github/workflows/presubmit.yml
+++ b/.github/workflows/presubmit.yml
@@ -1,9 +1,12 @@
+
+  
 name: Presubmit Checks
 on:
   pull_request:
     types: [opened, synchronize, reopened]
 
 jobs:
+
   title_format:
     name: Check PR Title
     if: ${{ github.event.pull_request }}
@@ -40,7 +43,6 @@ jobs:
           git remote add upstream https://github.com/taichi-dev/taichi.git
           git fetch upstream master
           sudo apt install clang-format-10
-
       - name: Cache PIP
         uses: actions/cache@v2
         with:
@@ -50,7 +52,6 @@ jobs:
       - name: Install requirements
         run: |
           python3 -m pip install --user -r requirements_dev.txt
-
       - name: Check code format
         run: |
           python3 misc/code_format.py
@@ -58,232 +59,32 @@ jobs:
           git commit -am "enforce code format" || true
           # exit with 1 if there were differences:
           git diff _fake_squash _enforced_format --exit-code
-
-      - name: Pylint
-        run: |
-          # Make sure pylint doesn't regress
-          pylint python/taichi/ --disable=all --enable=C0121,C0415
-          if [ $? -eq 0 ]
-          then
-            echo "PASSED: pylint is happy"
-            exit 0
-          else
-            echo "FAILED: please run the pylint command above and make sure it passes"
-            exit 1
-          fi
-
-  build_and_test_cpu_required:
-    # This job will be required to pass before merging to master branch.
-    name: Required Build and Test (CPU)
-    needs: check_code_format
+  
+  performance_monitoring:
+    name: Performance monitoring (NVGPU)
     timeout-minutes: 60
-    strategy:
-      matrix:
-        include:
-          - os: ubuntu-latest
-            python: 3.6
-    runs-on: ${{ matrix.os }}
+    runs-on: [self-hosted, x64, cuda, linux, benchmark]
     steps:
       - uses: actions/checkout@v2
         with:
-          submodules: "recursive"
-
-      - uses: actions/setup-python@v2
-        with:
-          python-version: ${{ matrix.python }}
-
-      - name: Download Pre-Built LLVM 10.0.0
-        run: python misc/ci_download.py
-        env:
-          CI_PLATFORM: ${{ matrix.os }}
+          submodules: 'recursive'
 
       - name: Build & Install
-        run: .github/workflows/scripts/unix_build.sh
-        env:
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CC:BOOL=ON -DTI_WITH_VULKAN:BOOL=OFF -DTI_BUILD_TESTS:BOOL=ON
-          CXX: clang++
-
-      - name: Test
-        run: .github/workflows/scripts/unix_test.sh
-        env:
-          RUN_CPP_TESTS: ON
-
-  build_and_test_cpu:
-    name: Build and Test (CPU)
-    needs: build_and_test_cpu_required
-    timeout-minutes: 60
-    strategy:
-      matrix:
-        include:
-          - os: macos-latest
-            python: 3.7
-            with_cc: OFF
-            with_cpp_tests: ON
-          - os: ubuntu-latest
-            python: 3.9
-            with_cc: OFF
-            with_cpp_tests: OFF
-          - os: ubuntu-latest
-            python: 3.8
-            with_cc: ON
-            with_cpp_tests: OFF
-    runs-on: ${{ matrix.os }}
-    steps:
-      - uses: actions/checkout@v2
-        with:
-          submodules: "recursive"
-
-      - uses: actions/setup-python@v2
-        with:
-          python-version: ${{ matrix.python }}
-
-      - name: Download Pre-Built LLVM 10.0.0
-        run: python misc/ci_download.py
-        env:
-          CI_PLATFORM: ${{ matrix.os }}
-
-      - name: Build & Install
-        run: .github/workflows/scripts/unix_build.sh
-        env:
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CC:BOOL=${{ matrix.with_cc }} -DTI_WITH_VULKAN:BOOL=OFF -DTI_BUILD_TESTS:BOOL=${{ matrix.with_cpp_tests }}
-          CXX: clang++
-      # [DEBUG] Copy this step around to enable debugging inside Github Action instances.
-      #- name: Setup tmate session
-      #  uses: mxschmitt/action-tmate@v3
-      #  with:
-      #    limit-access-to-actor: true
-
-      - name: Test
-        run: .github/workflows/scripts/unix_test.sh
-        env:
-          RUN_CPP_TESTS: ${{ matrix.with_cpp_tests }}
-
-  build_and_test_gpu_linux:
-    name: Build and Test (GPU)
-    needs: check_code_format
-    runs-on: [self-hosted, cuda, vulkan, cn]
-    timeout-minutes: 60
-    steps:
-      - uses: actions/checkout@v2
-        with:
-          submodules: "recursive"
-
-      - name: Build
         run: |
           export PATH=$PATH:/usr/local/cuda/bin
           .github/workflows/scripts/unix_build.sh
         env:
           LLVM_LIB_ROOT_DIR: /opt/taichi-llvm-10.0.0
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=ON -DTI_WITH_CC:BOOL=OFF -DTI_WITH_VULKAN:BOOL=ON -DTI_BUILD_TESTS:BOOL=ON
-          BUILD_NUM_THREADS: 8
           LLVM_PATH: /opt/taichi-llvm-10.0.0/bin
           LLVM_DIR: /opt/taichi-llvm-10.0.0/lib/cmake/llvm
-          CXX: clang++-8
-
-      - name: Test
-        run: .github/workflows/scripts/unix_test.sh
-        env:
-          DISPLAY: :1
-          GPU_TEST: ON
-          RUN_CPP_TESTS: ON
-
-  build_and_test_windows:
-    name: Build and Test (Windows)
-    needs: check_code_format
-    runs-on: windows-latest
-    timeout-minutes: 90
-    steps:
-      - name: Install 7Zip PowerShell
-        shell: powershell
-        run: Install-Module 7Zip4PowerShell -Force -Verbose
-
-      - uses: actions/checkout@v2
-        with:
-          submodules: "recursive"
-
-      - uses: actions/setup-python@v2
-        with:
-          python-version: 3.7
-
-      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v1.0.2
-
-      - name: Download And Install Vulkan
-        shell: powershell
-        run: |
-          Invoke-WebRequest -Uri "https://sdk.lunarg.com/sdk/download/1.2.189.0/windows/VulkanSDK-1.2.189.0-Installer.exe" -OutFile VulkanSDK.exe
-          $installer = Start-Process -FilePath VulkanSDK.exe -Wait -PassThru -ArgumentList @("/S");
-          $installer.WaitForExit();
-
-      - name: Build
-        shell: powershell
-        run: |
-          $env:Path += ";C:/VulkanSDK/1.2.189.0/Bin"
-          cd C:\
-          Remove-item alias:curl
-          curl --retry 10 --retry-delay 5 https://github.com/taichi-dev/taichi_assets/releases/download/llvm10/taichi-llvm-10.0.0-msvc2019.zip -LO
-          7z x taichi-llvm-10.0.0-msvc2019.zip -otaichi_llvm
-          curl --retry 10 --retry-delay 5 https://github.com/taichi-dev/taichi_assets/releases/download/llvm10/clang-10.0.0-win.zip -LO
-          7z x clang-10.0.0-win.zip -otaichi_clang
-          $env:PATH = ";C:\taichi_llvm\bin;C:\taichi_clang\bin;" + $env:PATH
-          clang --version
-          cd D:\a\taichi\taichi
-          python -m pip install -r requirements_dev.txt
-          cd python
-          git fetch origin master
-          $env:TAICHI_CMAKE_ARGS = $env:CI_SETUP_CMAKE_ARGS
-          python build.py build
-          cd ..\dist
-          $env:WHL = $(dir *.whl)
-          python -m pip install $env:WHL
-        env:
-          PYTHON: C:\hostedtoolcache\windows\Python\3.7.9\x64\python.exe
-          CI_SETUP_CMAKE_ARGS: -G "Visual Studio 16 2019" -A x64 -DLLVM_DIR=C:\taichi_llvm\lib\cmake\llvm -DTI_WITH_VULKAN:BOOL=ON
-          VULKAN_SDK: C:/VulkanSDK/1.2.189.0
-
-      - name: Test
-        shell: powershell
-        run: |
-          $env:PATH = ";C:\taichi_llvm\bin;C:\taichi_clang\bin;" + $env:PATH
-          python -c "import taichi"
-          python examples/algorithm/laplace.py
-          python bin/taichi diagnose
-          python bin/taichi changelog
-          python bin/taichi test -vr2 -t2
-        env:
-          PYTHON: C:\hostedtoolcache\windows\Python\3.7.9\x64\python.exe
-
-  build_and_test_m1:
-    name: Build and Test (Apple M1)
-    needs: check_code_format
-    timeout-minutes: 60
-    strategy:
-      matrix:
-        include:
-          - os: macos-latest
-            python: 3.8
-    defaults:
-      run:
-        # https://github.com/actions/runner/issues/805#issuecomment-844426478
-        shell: "/usr/bin/arch -arch arm64e /bin/bash --noprofile --norc -eo pipefail {0}"
-    runs-on: [self-hosted, m1]
-    steps:
-      - uses: actions/checkout@v2
-        with:
-          submodules: "recursive"
+          CUDA_TOOLKIT_ROOT_DIR: /usr/local/cuda/
+          CI_SETUP_CMAKE_ARGS: -DTI_WITH_CUDA_TOOLKIT:BOOL=ON
+          BUILD_NUM_THREADS: 8
+          CXX: clang++-10
 
-      - name: Build
-        run: |
-          rm -rf $HOME/Library/Python/3.8/lib/python/site-packages/taichi
-          .github/workflows/scripts/unix_build.sh
+      - name: Dump GitHub context
         env:
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CUDA:BOOL=OFF -DTI_WITH_CC:BOOL=OFF -DTI_WITH_VULKAN:BOOL=OFF -DTI_BUILD_TESTS:BOOL=ON
-          CXX: clang++
-
-      - name: Test
+          GITHUB_CONTEXT: ${{ toJson(github) }}
         run: |
-          export PATH=$PATH:$HOME/Library/Python/3.8/bin
-          python3 examples/algorithm/laplace.py
-          TI_LIB_DIR=`python3 -c "import taichi;print(taichi.__path__[0])" | tail -1`
-          TI_LIB_DIR="$TI_LIB_DIR/lib" ./build/taichi_cpp_tests
-          ti test -vr2 -t4 -x
+          python3 benchmarks/misc/test.py
+          python3 benchmarks/misc/print.py
\ No newline at end of file
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 33a82fea0..03956a923 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -20,9 +20,9 @@ jobs:
       - id: set-matrix
         run: |
           # For nightly release, we only run on python 3.8
-          [ -z "${{ github.event.action }}" ] && matrix="[{\"name\":\"taichi-nightly\",\"python\":\"3.8\"}]"
+          [ -z "${{ github.event.action }}" ] && matrix="[{\"name\":\"taichi-nightly\",\"python\":\"3.8\",\"conda_python\":\"py38\"}]"
           # For production release, we run on four python versions.
-          [ -z "${{ github.event.action }}" ] || matrix="[{\"name\":\"taichi\",\"python\":\"3.6\"},{\"name\":\"taichi\",\"python\":\"3.7\"},{\"name\":\"taichi\",\"python\":\"3.8\"},{\"name\":\"taichi\",\"python\":\"3.9\"}]"
+          [ -z "${{ github.event.action }}" ] || matrix="[{\"name\":\"taichi\",\"python\":\"3.6\",\"conda_python\":\"py36\"},{\"name\":\"taichi\",\"python\":\"3.7\",\"conda_python\":\"py37\"},{\"name\":\"taichi\",\"python\":\"3.8\",\"conda_python\":\"py38\"},{\"name\":\"taichi\",\"python\":\"3.9\",\"conda_python\":\"py39\"}]"
           echo ::set-output name=matrix::{\"include\":$(echo $matrix)}\"
           # M1 only supports py38 and py39(conda), so change matrix.
           [ -z "${{ github.event.action }}" ] && matrix_osx="[{\"name\":\"taichi-nightly\",\"python\":\"3.8\"}]"
@@ -41,56 +41,37 @@ jobs:
         with:
           submodules: "recursive"
 
-      - name: Create Python Wheel
+      - name: Build
         run: |
-          # We hacked here because conda activate in CI won't update python PATH
-          # automatically. So we don't activate and use desired python version
-          # directly.
-          export PATH=/home/buildbot/miniconda3/envs/$PYTHON/bin:$PATH
-          TAICHI_REPO_DIR=`pwd`
-          export PATH=$LLVM_LIB_ROOT_DIR/bin:/usr/local/cuda/bin:$PATH
-          export LLVM_DIR=$LLVM_LIB_ROOT_DIR/lib/cmake/llvm
-          export CXX=clang++-8
-          python3 -m pip uninstall taichi taichi-nightly -y
-          python3 -m pip install -r requirements_dev.txt
-          python3 -m pip install twine
-          cd python
-          git fetch origin master
-          export TAICHI_CMAKE_ARGS=$CI_SETUP_CMAKE_ARGS
-          python3 build.py build --project_name $PROJECT_NAME
-          cd ..
-          NUM_WHL=`ls dist/*.whl | wc -l`
-          if [ $NUM_WHL -ne 1 ]; then echo 'ERROR: created more than 1 whl.' && exit 1; fi
-          python3 -m pip install dist/*.whl
+          mkdir -m777 wheel
+
+          docker create -v `pwd`/wheel:/wheel --user dev --name taichi_build --gpus all -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix registry.taichigraphics.com/taichidev-ubuntu18.04:v0.1.1 /home/dev/test_actions/.github/workflows/scripts/unix_docker_build.sh $PY $GPU_BUILD $PROJECT_NAME "$CI_SETUP_CMAKE_ARGS"
+          tar -cf - ../${{ github.event.repository.name }} --mode u=+rwx,g=+rwx,o=+rwx --owner dev --group dev | docker cp - taichi_build:/home/dev/
+          docker start -a taichi_build
         env:
-          LLVM_LIB_ROOT_DIR: /opt/taichi-llvm-10.0.0
-          BUILD_NUM_THREADS: 8
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_VULKAN:BOOL=ON -DTI_WITH_OPENGL:BOOL=ON -DTI_WITH_CC:BOOL=OFF -DTI_BUILD_TESTS:BOOL=${{ matrix.with_cpp_tests }}
+          PY: ${{ matrix.conda_python }}
+          GPU_BUILD: ON
+          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=ON -DTI_WITH_CC:BOOL=OFF -DTI_WITH_VULKAN:BOOL=ON -DTI_BUILD_TESTS:BOOL=ON
           PROJECT_NAME: ${{ matrix.name }}
-          PYTHON: ${{ matrix.python }}
+          DISPLAY: :1
 
       - name: Archive Wheel Artifacts
         uses: actions/upload-artifact@v2
         with:
           name: ${{ matrix.name }}-py${{ matrix.python }}-linux.whl
-          path: dist/*.whl
+          path: wheel/*.whl
           retention-days: 20
 
       - name: Test
         run: |
-          export PATH=/home/buildbot/miniconda3/envs/$PYTHON/bin:$PATH
-          python3 examples/algorithm/laplace.py
-          export DISPLAY=:1
-          hash -r
-          glewinfo
-          ti diagnose
-          ti changelog
-          ti test -vr2 -t2 -k "not ndarray and not torch"
-          # ndarray test might OOM if run with -t2.
-          # FIXME: unify this with presubmit.yml to avoid further divergence
-          ti test -vr2 -t1 -k "ndarray or torch"
+          docker create --user dev --name taichi_test --gpus all -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix registry.taichigraphics.com/taichidev-ubuntu18.04:v0.1.1 /home/dev/unix_docker_test.sh $PY $GPU_TEST
+          docker cp .github/workflows/scripts/unix_docker_test.sh taichi_test:/home/dev/unix_docker_test.sh
+          docker cp wheel/*.whl taichi_test:/home/dev/
+          docker start -a taichi_test
         env:
-          PYTHON: ${{ matrix.python }}
+          PY: ${{ matrix.conda_python }}
+          GPU_TEST: ON
+          DISPLAY: :1
 
       - name: Upload PyPI
         env:
@@ -100,11 +81,17 @@ jobs:
           PROJECT_NAME: ${{ matrix.name }}
           PYTHON: ${{ matrix.python }}
         run: |
-          export PATH=/home/buildbot/miniconda3/envs/$PYTHON/bin:$PATH
+          mkdir dist
+          cp wheel/*.whl dist/
           cd python
           if [ $PROJECT_NAME == "taichi-nightly" ]; then export PYPI_PWD="$NIGHT_PWD" && python3 build.py upload --skip_build --testpypi --project_name $PROJECT_NAME
           elif [ $PROJECT_NAME == "taichi" ]; then export PYPI_PWD="$PROD_PWD" && python3 build.py upload --skip_build; fi
 
+      - name: clean docker container
+        if: always()
+        run: |
+          docker rm taichi_build taichi_test -f
+
   build_and_upload_mac:
     name: Build and Upload (macOS only)
     needs: matrix_prep
@@ -132,6 +119,7 @@ jobs:
           export PATH=$TAICHI_REPO_DIR/taichi-llvm/bin/:$PATH
           export CXX=clang++
           python -m pip install -r requirements_dev.txt
+          python -m pip install -r requirements_test.txt
           cd python
           git fetch origin master
           export TAICHI_CMAKE_ARGS=$CI_SETUP_CMAKE_ARGS
@@ -141,7 +129,7 @@ jobs:
           if [ $NUM_WHL -ne 1 ]; then echo 'ERROR: created more than 1 whl.' && exit 1; fi
           pip install dist/*.whl
         env:
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_VULKAN:BOOL=OFF -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CC:BOOL=OFF -DTI_BUILD_TESTS:BOOL=${{ matrix.with_cpp_tests }}
+          CI_SETUP_CMAKE_ARGS: -DTI_WITH_VULKAN:BOOL=ON -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CC:BOOL=OFF -DTI_BUILD_TESTS:BOOL=${{ matrix.with_cpp_tests }}
           PROJECT_NAME: ${{ matrix.name }}
 
       - name: Archive Wheel Artifacts
@@ -155,7 +143,7 @@ jobs:
         run: |
           python examples/algorithm/laplace.py
           ti diagnose
-          ti test -vr2 -t2
+          ti test -vr2 -t2 -a cpu,metal
 
       - name: Upload PyPI
         env:
@@ -193,6 +181,7 @@ jobs:
           git --version
           export CXX=clang++
           python3 -m pip install -r requirements_dev.txt
+          python3 -m pip install -r requirements_test.txt
           cd python
           git fetch origin master
           export TAICHI_CMAKE_ARGS=$CI_SETUP_CMAKE_ARGS
@@ -202,7 +191,7 @@ jobs:
           if [ $NUM_WHL -ne 1 ]; then echo 'ERROR: created more than 1 whl.' && exit 1; fi
           python3 -m pip install dist/*.whl
         env:
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CUDA:BOOL=OFF -DTI_WITH_CC:BOOL=OFF -DTI_WITH_VULKAN:BOOL=OFF -DTI_WITH_TESTS:BOOL=ON
+          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CUDA:BOOL=OFF -DTI_WITH_CC:BOOL=OFF -DTI_WITH_VULKAN:BOOL=ON -DTI_WITH_TESTS:BOOL=ON
           PROJECT_NAME: ${{ matrix.name }}
           PYTHON: ${{ matrix.python }}
 
@@ -250,6 +239,7 @@ jobs:
           git --version
           export CXX=clang++
           python3 -m pip install -r requirements_dev.txt
+          python3 -m pip install -r requirements_test.txt
           cd python
           git fetch origin master
           export TAICHI_CMAKE_ARGS=$CI_SETUP_CMAKE_ARGS
@@ -259,7 +249,7 @@ jobs:
           if [ $NUM_WHL -ne 1 ]; then echo 'ERROR: created more than 1 whl.' && exit 1; fi
           python3 -m pip install dist/*.whl
         env:
-          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CUDA:BOOL=OFF -DTI_WITH_CC:BOOL=OFF -DTI_WITH_VULKAN:BOOL=OFF -DTI_WITH_TESTS:BOOL=ON
+          CI_SETUP_CMAKE_ARGS: -DTI_WITH_OPENGL:BOOL=OFF -DTI_WITH_CUDA:BOOL=OFF -DTI_WITH_CC:BOOL=OFF -DTI_WITH_VULKAN:BOOL=ON -DTI_WITH_TESTS:BOOL=ON
           PROJECT_NAME: ${{ matrix.name }}
           PYTHON: ${{ matrix.python }}
 
@@ -336,6 +326,7 @@ jobs:
           clang --version
           cd D:\a\taichi\taichi
           python -m pip install -r requirements_dev.txt
+          python -m pip install -r requirements_test.txt
           cd python
           git fetch origin master
           $env:TAICHI_CMAKE_ARGS = $env:CI_SETUP_CMAKE_ARGS
diff --git a/.github/workflows/scripts/unix_docker_build.sh b/.github/workflows/scripts/unix_docker_build.sh
new file mode 100755
index 000000000..e69de29bb
diff --git a/.github/workflows/scripts/unix_docker_test.sh b/.github/workflows/scripts/unix_docker_test.sh
new file mode 100755
index 000000000..d3df0c2b5
--- /dev/null
+++ b/.github/workflows/scripts/unix_docker_test.sh
@@ -0,0 +1,29 @@
+#!/bin/bash
+
+set -ex
+
+# Parse ARGs
+PY=$1
+GPU_TEST=$2
+
+source /home/dev/miniconda/etc/profile.d/conda.sh
+conda activate $PY
+
+python3 -m pip install ./*.whl
+
+if [[ $GPU_TEST == "OFF" ]]
+then
+    python3 -m pip install -r requirements_test.txt
+fi
+
+export TI_IN_DOCKER=true
+ti diagnose
+ti changelog
+
+if [[ $GPU_TEST == "OFF" ]]
+then
+    ti test -vr2 -t2
+else
+    ti test -vr2 -t2 -k "not ndarray and not torch"
+    ti test -vr2 -t1 -k "ndarray or torch"
+fi
diff --git a/.github/workflows/scripts/unix_test.sh b/.github/workflows/scripts/unix_test.sh
index 907c52160..53bf5d07e 100755
--- a/.github/workflows/scripts/unix_test.sh
+++ b/.github/workflows/scripts/unix_test.sh
@@ -1,5 +1,6 @@
 set -ex
 TAICHI_REPO_DIR=`pwd`
+python3 -m pip install -r requirements_test.txt
 TI_LIB_DIR=`python3 -c "import taichi;print(taichi.__path__[0])" | tail -1`
 [[ $RUN_CPP_TESTS == "ON" ]] && TI_LIB_DIR="$TI_LIB_DIR/lib" ./build/taichi_cpp_tests
 export PATH=$TAICHI_REPO_DIR/taichi-llvm/bin/:$PATH
diff --git a/benchmarks/misc/membound.py b/benchmarks/misc/membound.py
index 6f1952857..8b1378917 100644
--- a/benchmarks/misc/membound.py
+++ b/benchmarks/misc/membound.py
@@ -1,78 +1 @@
-import time
 
-from membound_cases import memory_bound_cases_list
-from utils import (arch_name, dtype2str, geometric_mean, kibibyte,
-                   md_table_header, size2str)
-
-import taichi as ti
-
-
-class MemoryBound:
-    suite_name = 'memorybound'
-    supported_archs = [ti.cpu, ti.cuda]
-    test_cases = memory_bound_cases_list
-    test_dtype_list = [ti.i32, ti.i64, ti.f32, ti.f64]
-    test_dsize_list = [(4**i) * kibibyte
-                       for i in range(1, 10)]  #[4KB,16KB...256MB]
-    basic_repeat_times = 10
-    evaluator = [geometric_mean]
-
-    def __init__(self, arch):
-        self.arch = arch
-        self.cases_impl = []
-        for case in self.test_cases:
-            for dtype in self.test_dtype_list:
-                impl = CaseImpl(case, arch, dtype, self.test_dsize_list,
-                                self.evaluator)
-                self.cases_impl.append(impl)
-
-    def run(self):
-        for case in self.cases_impl:
-            case.run()
-
-    def get_markdown_lines(self):
-        lines = []
-        lines += md_table_header(self.suite_name, self.arch,
-                                 self.test_dsize_list, self.basic_repeat_times,
-                                 self.evaluator)
-
-        result_header = '|kernel elapsed time(ms)' + ''.join(
-            '|' for i in range(
-                len(self.test_dsize_list) + len(MemoryBound.evaluator)))
-        lines += [result_header]
-        for case in self.cases_impl:
-            lines += case.get_markdown_lines()
-        lines.append('')
-        return lines
-
-
-class CaseImpl:
-    def __init__(self, func, arch, test_dtype, test_dsize_list, evaluator):
-        self.func = func
-        self.name = func.__name__
-        self.arch = arch
-        self.test_dtype = test_dtype
-        self.test_dsize_list = test_dsize_list
-        self.min_time_in_us = []  #test results
-        self.evaluator = evaluator
-
-    def run(self):
-        ti.init(kernel_profiler=True, arch=self.arch)
-        print("TestCase[%s.%s.%s]" % (self.func.__name__, arch_name(
-            self.arch), dtype2str[self.test_dtype]))
-        for test_dsize in self.test_dsize_list:
-            print("test_dsize = %s" % (size2str(test_dsize)))
-            self.min_time_in_us.append(
-                self.func(self.arch, self.test_dtype, test_dsize,
-                          MemoryBound.basic_repeat_times))
-            time.sleep(0.2)
-        ti.reset()
-
-    def get_markdown_lines(self):
-        string = '|' + self.name + '.' + dtype2str[self.test_dtype] + '|'
-        string += ''.join(
-            str(round(time, 4)) + '|' for time in self.min_time_in_us)
-        string += ''.join(
-            str(round(item(self.min_time_in_us), 4)) + '|'
-            for item in self.evaluator)
-        return [string]
diff --git a/benchmarks/misc/membound_cases.py b/benchmarks/misc/membound_cases.py
index 5792a85a2..8b1378917 100644
--- a/benchmarks/misc/membound_cases.py
+++ b/benchmarks/misc/membound_cases.py
@@ -1,79 +1 @@
-from utils import dtype_size, scaled_repeat_times
 
-import taichi as ti
-
-
-def init_const(x, dtype, num_elements):
-    @ti.kernel
-    def init_const(x: ti.template(), n: ti.i32):
-        for i in range(n):
-            x[i] = ti.cast(0.7, dtype)
-
-    init_const(x, num_elements)
-
-
-def membound_benchmark(func, num_elements, repeat):
-    # compile the kernel first
-    func(num_elements)
-    ti.clear_kernel_profile_info()
-    for i in range(repeat):
-        func(num_elements)
-    kernelname = func.__name__
-    quering_result = ti.query_kernel_profile_info(kernelname)
-    return quering_result.min
-
-
-def fill(arch, dtype, dsize, repeat=10):
-
-    repeat = scaled_repeat_times(arch, dsize, repeat)
-    num_elements = dsize // dtype_size[dtype]
-
-    x = ti.field(dtype, shape=num_elements)
-
-    @ti.kernel
-    def fill_const(n: ti.i32):
-        for i in range(n):
-            x[i] = ti.cast(0.7, dtype)
-
-    return membound_benchmark(fill_const, num_elements, repeat)
-
-
-def saxpy(arch, dtype, dsize, repeat=10):
-
-    repeat = scaled_repeat_times(arch, dsize, repeat)
-    num_elements = dsize // dtype_size[dtype] // 3  #z=x+y
-
-    x = ti.field(dtype, shape=num_elements)
-    y = ti.field(dtype, shape=num_elements)
-    z = ti.field(dtype, shape=num_elements)
-
-    @ti.kernel
-    def saxpy(n: ti.i32):
-        for i in range(n):
-            z[i] = 17 * x[i] + y[i]
-
-    init_const(x, dtype, num_elements)
-    init_const(y, dtype, num_elements)
-
-    return membound_benchmark(saxpy, num_elements, repeat)
-
-
-def reduction(arch, dtype, dsize, repeat=10):
-
-    repeat = scaled_repeat_times(arch, dsize, repeat)
-    num_elements = dsize // dtype_size[dtype]
-
-    x = ti.field(dtype, shape=num_elements)
-    y = ti.field(dtype, shape=())
-    y[None] = 0
-
-    @ti.kernel
-    def reduction(n: ti.i32):
-        for i in range(n):
-            y[None] += x[i]
-
-    init_const(x, dtype, num_elements)
-    return membound_benchmark(reduction, num_elements, repeat)
-
-
-memory_bound_cases_list = [fill, saxpy, reduction]
diff --git a/benchmarks/misc/print.py b/benchmarks/misc/print.py
new file mode 100644
index 000000000..8b5c01809
--- /dev/null
+++ b/benchmarks/misc/print.py
@@ -0,0 +1,13 @@
+import json
+import os
+import re
+import sys
+
+
+def main():
+    num = os.environ.get('PULL_REQUEST_NUMBER')
+    print(f'PULL_REQUEST_NUMBER = {num}')
+
+
+if __name__ == '__main__':
+    main()
diff --git a/benchmarks/misc/run.py b/benchmarks/misc/run.py
index 4922e043d..8b1378917 100644
--- a/benchmarks/misc/run.py
+++ b/benchmarks/misc/run.py
@@ -1,78 +1 @@
-import os
-import warnings
 
-from membound import MemoryBound
-from taichi.core import ti_core as _ti_core
-from utils import arch_name, datatime_with_format
-
-import taichi as ti
-
-benchmark_suites = [MemoryBound]
-benchmark_archs = [ti.cpu, ti.cuda]
-
-
-class CommitInfo:
-    def __init__(self, pull_request_id, commit_hash):
-        self.pull_request_id = pull_request_id
-        self.commit_hash = commit_hash  #str
-        self.archs = []  #['x64','cuda','vulkan', ...]
-        self.datetime = []  #[start, end]
-
-
-class BenchmarkSuites:
-    def __init__(self, arch):
-        self.suites = []
-        self.arch = arch
-        for suite in benchmark_suites:
-            if self.check_supported(arch, suite):
-                self.suites.append(suite(arch))
-
-    def check_supported(self, arch, suite):
-        if arch in suite.supported_archs:
-            return True
-        else:
-            warnings.warn(
-                f'Arch [{arch_name(arch)}] does not exist in {suite.__name__}.supported_archs.',
-                UserWarning,
-                stacklevel=2)
-            return False
-
-    def run(self):
-        print(f'Arch [{arch_name(self.arch)}] Running...')
-        for suite in self.suites:
-            suite.run()
-
-    def save_to_markdown(self, arch_dir='./'):
-        current_time = datatime_with_format()
-        commit_hash = _ti_core.get_commit_hash()  #[:8]
-        for suite in self.suites:
-            file_name = f'{suite.suite_name}.md'
-            path = os.path.join(arch_dir, file_name)
-            with open(path, 'w') as f:
-                lines = [
-                    f'commit_hash: {commit_hash}\n',
-                    f'datatime: {current_time}\n'
-                ]
-                lines += suite.get_markdown_lines()
-                for line in lines:
-                    print(line, file=f)
-
-
-def main():
-
-    benchmark_dir = os.path.join(os.getcwd(), 'results')
-    os.makedirs(benchmark_dir)
-
-    for arch in benchmark_archs:
-        #make dir
-        arch_dir = os.path.join(benchmark_dir, arch_name(arch))
-        os.makedirs(arch_dir)
-        #init & run
-        suites = BenchmarkSuites(arch)
-        suites.run()
-        #save result
-        suites.save_to_markdown(arch_dir)
-
-
-if __name__ == '__main__':
-    main()
diff --git a/benchmarks/misc/test.py b/benchmarks/misc/test.py
new file mode 100644
index 000000000..7f550176b
--- /dev/null
+++ b/benchmarks/misc/test.py
@@ -0,0 +1,17 @@
+import json
+import os
+import re
+import sys
+
+
+def main():
+    github_context = os.environ.get('GITHUB_CONTEXT')
+    github_info_dict = {}
+    github_info_dict = json.loads(github_context)
+    pr_number = re.sub("\D", "", github_info_dict["ref"])
+    os.environ['PULL_REQUEST_NUMBER'] = pr_number
+    print(os.environ.get('PULL_REQUEST_NUMBER'))
+
+
+if __name__ == '__main__':
+    main()
diff --git a/benchmarks/misc/utils.py b/benchmarks/misc/utils.py
index 3548151ec..8b1378917 100644
--- a/benchmarks/misc/utils.py
+++ b/benchmarks/misc/utils.py
@@ -1,68 +1 @@
-import datetime
 
-import taichi as ti
-
-kibibyte = 1024
-
-bls2str = {False: "BLS_off", True: "BLS_on"}
-dense2str = {False: "Struct_for", True: "Range_for"}
-
-dtype2str = {ti.i32: "i32", ti.i64: "i64", ti.f32: "f32", ti.f64: "f64"}
-dtype_size = {ti.i32: 4, ti.i64: 8, ti.f32: 4, ti.f64: 8}
-
-# for output string
-size_subsection = [(0.0, 'B'), (1024.0, 'KB'), (1048576.0, 'MB'),
-                   (1073741824.0, 'GB'), (float('inf'), 'INF')]  #B KB MB GB
-
-
-def arch_name(arch):
-    return str(arch).replace('Arch.', '')
-
-
-def datatime_with_format():
-    return datetime.datetime.now().isoformat()
-
-
-def size2str(size_in_byte):
-    for dsize, units in reversed(size_subsection):
-        if size_in_byte >= dsize:
-            return str(round(size_in_byte / dsize, 4)) + units
-
-
-def geometric_mean(data_array):
-    product = 1
-    for data in data_array:
-        product *= data
-    return pow(product, 1.0 / len(data_array))
-
-
-def scaled_repeat_times(arch, datasize, repeat=1):
-    if (arch == ti.gpu) | (arch == ti.opengl) | (arch == ti.cuda):
-        repeat *= 10
-    if datasize <= 4 * 1024 * 1024:
-        repeat *= 10
-    return repeat
-
-
-def md_table_header(suite_name, arch, test_dsize, test_repeat,
-                    results_evaluation):
-    header = '|' + suite_name + '.' + ti.core.arch_name(arch) + '|'
-    header += ''.join('|' for i in range(len(test_dsize)))
-    header += ''.join(item.__name__ + '|' for item in results_evaluation)
-
-    layout = '|:--:|'
-    layout += ''.join(
-        ':--:|' for i in range(len(test_dsize) + len(results_evaluation)))
-
-    size = '|**data size**|'
-    size += ''.join(size2str(size) + '|' for size in test_dsize)
-    size += ''.join('|' for i in range(len(results_evaluation)))
-
-    repeat = '|**repeat**|'
-    repeat += ''.join(
-        str(scaled_repeat_times(arch, size, test_repeat)) + '|'
-        for size in test_dsize)
-    repeat += ''.join('|' for i in range(len(results_evaluation)))
-
-    lines = [header, layout, size, repeat]
-    return lines