Add opt-in flag for debug mode in unit tests.

Author: duburcqa

.github/workflows/generic.yml

@@ -32,6 +32,7 @@ jobs:
       TI_ENABLE_METAL: "0"
       TI_ENABLE_OPENGL: "0"
       TI_ENABLE_VULKAN: "0"
+      TI_DEBUG: "0"
 
     runs-on: ${{ matrix.OS }}
     if: github.event_name != 'release'
@@ -112,7 +113,7 @@ jobs:
 
       - name: Run unit tests
         run: |
-          pytest -v --logical --forked -m required ./tests
+          pytest -v --logical --dev --forked -m required ./tests
 
       - name: Save Updated Taichi Kernel Cache
         if: always()

.github/workflows/production.yml

@@ -67,7 +67,7 @@ jobs:
             bash -c "
               pip install --extra-index-url https://pypi.nvidia.com/ omniverse-kit && \
               pip install -e '.[dev,render,usd]' && \
-              pytest -v --forked ./tests
+              pytest -v --dev --forked ./tests
             "
 
       - name: Run benchmarks

genesis/__init__.py

@@ -213,7 +213,7 @@ def init(
             arch=ti_arch,
             # Debug mode was buggy and causing segfault prior to v1.7.4
             # Add a (hidden) mechanism to forceable disable taichi debug mode as it is still a bit experimental
-            debug=debug and ti.__version__ >= (1, 7, 4) and (os.environ.get("TI_DEBUG", 1) == "0"),
+            debug=debug and ti.__version__ >= (1, 7, 4) and (os.environ.get("TI_DEBUG") != "0"),
             check_out_of_bound=debug,
             # force_scalarize_matrix=True for speeding up kernel compilation
             # Turning off 'force_scalarize_matrix' is causing numerical instabilities ('nan') on MacOS

tests/conftest.py

@@ -148,6 +148,7 @@ def pytest_addoption(parser):
         "--logical", action="store_true", default=False, help="Consider logical cores in default number of workers."
     )
     parser.addoption("--vis", action="store_true", default=False, help="Enable interactive viewer.")
+    parser.addoption("--dev", action="store_true", default=False, help="Enable genesis debug mode.")
 
 
 @pytest.fixture(scope="session")
@@ -276,13 +277,8 @@ def initialize_genesis(request, backend, taichi_offline_cache):
     from genesis.utils.misc import ALLOCATE_TENSOR_WARNING
 
     logging_level = request.config.getoption("--log-cli-level")
-    if backend == gs.cpu:
-        precision = "64"
-        debug = True
-    else:
-        # Disabling debug mode on GPU because we are likely needing maximum performance, notably for benchmarks
-        precision = "32"
-        debug = False
+    debug = request.config.getoption("--dev")
+    precision = "64" if backend == gs.cpu else "32"
 
     try:
         if not taichi_offline_cache: