Merge branch 'main' into reject-casting-to-fp-in-bi-profile

Author: zingo

backends/arm/test/ops/test_conv_combos.py

@@ -430,16 +430,8 @@ def test_convolution_2d_tosa_MI_avgpool2d(test_data: torch.Tensor):
     pipeline.run()
 
 
-x_fails = {
-    "combo_conv_avgpool_20_x_4d": "AssertionError: Output 0 does not match reference output.",
-    "combo_conv_avgpool_4d": "AssertionError: Output 0 does not match reference output.",
-    "combo_conv_avgpool_5_x_4d_randn": "AssertionError: Output 0 does not match reference output.",
-    "combo_conv_avgpool_2_x_4d": "AssertionError: Output 0 does not match reference output.",
-}
-
-
 @pytest.mark.flaky(reruns=5)  # TODO: Investigate flakyness (MLTORCH-307)
-@common.parametrize("test_data", ComboConvAvgPool2d.test_data, x_fails)
+@common.parametrize("test_data", ComboConvAvgPool2d.test_data)
 def test_convolution_2d_tosa_BI_avgpool2d(test_data: torch.Tensor):
     model = ComboConvAvgPool2d()
     pipeline = TosaPipelineBI[input_t1](

backends/arm/test/tester/test_pipeline.py

@@ -282,7 +282,7 @@ def __init__(
         custom_path: str = None,
         atol: float = 1e-03,
         rtol: float = 1e-03,
-        qtol: int = 0,
+        qtol: int = 1,
     ):
         tosa_profiles = {
             "0.80": TosaSpecification.create_from_string("TOSA-0.80+BI"),

backends/mediatek/preprocess.py

@@ -20,6 +20,8 @@
 from executorch.exir.backend.compile_spec_schema import CompileSpec
 
 SKIP_COMPILE_SPEC_KEYS = {"ImportForever"}
+REQUIRED_COMPILE_SPEC_KEYS = {"platform-config"}
+SUPPORTED_PLATFORM_CONFIGS = {"mt6989", "mt6991"}
 
 
 def assert_default_dim_order(edge_graph_module: torch.fx.GraphModule) -> None:
@@ -47,6 +49,28 @@ def preprocess(
         cls, edge_program: ExportedProgram, module_compile_spec: List[CompileSpec]
     ) -> PreprocessResult:
 
+        # Validate CompileSpec settings
+        compile_spec_keys = [spec.key for spec in module_compile_spec]
+        if len(compile_spec_keys) != len(set(compile_spec_keys)):
+            raise RuntimeError(
+                "Unsupported duplicated keys in the CompileSpec settings."
+            )
+        if not REQUIRED_COMPILE_SPEC_KEYS.issubset(set(compile_spec_keys)):
+            raise RuntimeError(
+                "Following keys are required in the CompileSpec settings: {}."
+                "".format(REQUIRED_COMPILE_SPEC_KEYS)
+            )
+        platform = [
+            spec.value.decode("utf-8")
+            for spec in module_compile_spec
+            if spec.key == "platform-config"
+        ][0]
+        if platform not in SUPPORTED_PLATFORM_CONFIGS:
+            raise ValueError(
+                "Unsupported value of platform-config CompileSpec. Given {} but expected to be one "
+                "of {}.".format(platform, SUPPORTED_PLATFORM_CONFIGS)
+            )
+
         # Make sure all inputs are contiguous_format or NCHW or default dim order
         assert_default_dim_order(edge_program.graph_module)
 
@@ -64,8 +88,7 @@ def preprocess(
             if name_to_node_mappings[name].meta["val"].dtype == torch.float32
         ]
 
-        # This default compile options are only for mt6989 SOC
-        compile_options = ["--arch=mdla5.1,edpa1.0", "--relax-fp32", "--opt=3"]
+        compile_options = ["--relax-fp32", "--opt=3"]
         for spec in module_compile_spec:
             if spec.key in SKIP_COMPILE_SPEC_KEYS:
                 continue

examples/mediatek/aot_utils/oss_utils/utils.py

@@ -14,6 +14,7 @@
     NeuropilotQuantizer,
     Precision,
 )
+from executorch.exir.backend.backend_details import CompileSpec
 from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 
 
@@ -48,7 +49,7 @@ def build_executorch_binary(
     edge_compile_config = exir.EdgeCompileConfig(_check_ir_validity=False)
     # skipped op names are used for deeplabV3 model
     neuro_partitioner = NeuropilotPartitioner(
-        [],
+        [CompileSpec("platform-config", b"mt6989")],
         op_names_to_skip={
             "aten_convolution_default_106",
             "aten_convolution_default_107",

examples/mediatek/model_export_scripts/llama.py

@@ -350,10 +350,11 @@ def export_to_et_ir(
 
         print("Delegating Edge Program to Neuropilot Backend")
         compile_spec = [
-            CompileSpec("gno", struct.pack("3s", b"LTS")),
-            CompileSpec("gno-exp", struct.pack("0s", b"")),
-            CompileSpec("gno-non-4d-tiling", struct.pack("0s", b"")),
+            CompileSpec("gno", b"LTS"),
+            CompileSpec("gno-exp", b""),
+            CompileSpec("gno-non-4d-tiling", b""),
             CompileSpec("ImportForever", struct.pack("?", True)),
+            CompileSpec("platform-config", b"mt6989"),
         ]
         partitioner = NeuropilotPartitioner(compile_spec)
         delegated_program = edge_program.to_backend(partitioner)