[ci] Reduce test parallelism for m1 (#4394)

* [ci] Reduce test parallelism for m1

* Use PLATFORM

* Remove metal from self gpu runners

.github/workflows/scripts/unix_test.sh

@@ -35,9 +35,18 @@ TI_PATH=$(python3 -c "import taichi;print(taichi.__path__[0])" | tail -1)
 TI_LIB_DIR="$TI_PATH/_lib/runtime" ./build/taichi_cpp_tests
 
 if [ -z "$GPU_TEST" ]; then
-    python3 tests/run_tests.py -vr2 -t4 -a "$TI_WANTED_ARCHS"
+    if [[ $PLATFORM == *"m1"* ]]; then
+	# Split per arch to avoid flaky test
+        python3 tests/run_tests.py -vr2 -t4 -k "not torch" -a cpu
+        # Run metal and vulkan separately so that they don't use M1 chip simultaneously.
+        python3 tests/run_tests.py -vr2 -t4 -k "not torch" -a vulkan
+        python3 tests/run_tests.py -vr2 -t2 -k "not torch" -a metal
+        python3 tests/run_tests.py -vr2 -t1 -k "torch" -a "$TI_WANTED_ARCHS"
+    else
+        python3 tests/run_tests.py -vr2 -t4 -a "$TI_WANTED_ARCHS"
+    fi
 else
-    # only split per arch for self_hosted GPU tests
+    # Split per arch to increase parallelism for linux GPU tests
     if [[ $TI_WANTED_ARCHS == *"cuda"* ]]; then
         python3 tests/run_tests.py -vr2 -t4 -k "not torch" -a cuda
     fi
@@ -50,9 +59,5 @@ else
     if [[ $TI_WANTED_ARCHS == *"opengl"* ]]; then
         python3 tests/run_tests.py -vr2 -t4 -k "not torch" -a opengl
     fi
-    # Run metal and vulkan separately so that they don't use M1 chip simultaneously.
-    if [[ $TI_WANTED_ARCHS == *"metal"* ]]; then
-        python3 tests/run_tests.py -vr2 -t4 -k "not torch" -a metal
-    fi
     python3 tests/run_tests.py -vr2 -t1 -k "torch" -a "$TI_WANTED_ARCHS"
 fi

.github/workflows/testing.yml

@@ -466,4 +466,4 @@ jobs:
         env:
           TI_WANTED_ARCHS: "metal,vulkan,cpu"
           PY: ${{ matrix.python }}
-          GPU_TEST: ON
+          PLATFORM: "m1"