Qualcomm AI Engine Direct - Mimi Enablement Stage 1 (#9570)

### Summary
This is the stage 1 of Mimi Enablement.
Stage 2 will consist of actual model enablement.
- Support OP:
  - exp
  - expm1
  - elu
  - transpose conv1d
  - bitwise_and
  - scalar_tensor
  - stack
  - unbind

Author: winskuo-quic

backends/qualcomm/_passes/__init__.py

@@ -1,10 +1,18 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
 from .annotate_decomposed import AnnotateDecomposed
 from .annotate_quant_attrs import AnnotateQuantAttrs
 from .constant_i64_to_i32 import ConstantI64toI32
 from .convert_bmm_to_matmul import ConvertBmmToMatmul
+from .convert_conv1d_to_conv2d import ConvertConv1dToConv2d
 from .convert_to_linear import ConvertToLinear
 from .decompose_any import DecomposeAny
 from .decompose_einsum import DecomposeEinsum
+from .decompose_expm1 import DecomposeExpM1
 from .decompose_linalg_vector_norm import DecomposeLinalgVectorNorm
 from .decompose_silu import DecomposeSilu
 from .expand_broadcast_tensor_shape import ExpandBroadcastTensorShape
@@ -19,8 +27,9 @@
 from .recompose_rms_norm import RecomposeRmsNorm
 from .reduce_dynamic_range import ReduceDynamicRange
 from .remove_redundancy import RemoveRedundancy
+from .replace_arange_args import ReplaceArangeArgs
 from .replace_index_put_input import ReplaceIndexPutInput
-from .replace_inf_buffer import ReplaceInfBuffer
+from .replace_inf_values import ReplaceInfValues
 from .tensor_i64_to_i32 import TensorI64toI32
 
 
@@ -29,10 +38,12 @@
     AnnotateQuantAttrs,
     ConstantI64toI32,
     ConvertBmmToMatmul,
+    ConvertConv1dToConv2d,
     RecomposePReLU,
     ConvertToLinear,
     DecomposeAny,
     DecomposeEinsum,
+    DecomposeExpM1,
     DecomposeLinalgVectorNorm,
     DecomposeSilu,
     ExpandBroadcastTensorShape,
@@ -46,7 +57,8 @@
     RecomposeRmsNorm,
     ReduceDynamicRange,
     RemoveRedundancy,
+    ReplaceArangeArgs,
     ReplaceIndexPutInput,
-    ReplaceInfBuffer,
+    ReplaceInfValues,
     TensorI64toI32,
 ]

backends/qualcomm/_passes/annotate_decomposed.py

@@ -17,6 +17,8 @@ class AnnotateDecomposed(ExportPass):
     generated after quantization process.
     """
 
+    decomp_ops = [torch.ops.aten.stack.default, torch.ops.aten.unbind.int]
+
     def __init__(self, edge_program: torch.export.ExportedProgram):
         super(AnnotateDecomposed, self).__init__()
         self.edge_program = edge_program
@@ -32,7 +34,7 @@ def _annotate_unbind(self, graph_module: torch.fx.GraphModule):
                         n.meta[QCOM_QUANT_ATTRS] = quant_attrs.copy()
 
     def _annotate_stack(self, graph_module: torch.fx.GraphModule):
-        partitions = get_source_partitions(graph_module.graph, [torch.stack])
+        partitions = get_source_partitions(graph_module.graph, [torch.stack, "stack"])
         for _, src_partitions in partitions.items():
             for src_partition in src_partitions:
                 output = src_partition.output_nodes[0]

backends/qualcomm/_passes/convert_conv1d_to_conv2d.py

@@ -0,0 +1,99 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+from executorch.backends.qualcomm.builders.utils import get_parameter, set_parameter
+from executorch.exir.dialects._ops import ops as exir_ops
+from executorch.exir.pass_base import ExportPass, PassResult
+
+from .utils import copy_meta
+
+
+class ConvertConv1dToConv2d(ExportPass):
+    """
+    Conv1d is not supported by QNN.
+    Change it to input -> unsqueeze -> conv2d -> squeeze -> output
+    """
+
+    def __init__(self, edge_program: torch.export.ExportedProgram):
+        super(ConvertConv1dToConv2d, self).__init__()
+        self.edge_program = edge_program
+
+    def call(self, graph_module: torch.fx.GraphModule):
+        graph = graph_module.graph
+        conv_op = exir_ops.edge.aten.convolution.default
+        for node in graph.nodes:
+            if node.target == conv_op and node.meta["val"].dim() == 3:
+
+                input_node = node.args[0]
+                with graph_module.graph.inserting_after(input_node):
+                    unsqueeze_op = exir_ops.edge.aten.unsqueeze_copy.default
+                    unsqueeze_node = graph.create_node(
+                        "call_function",
+                        unsqueeze_op,
+                        (
+                            input_node,
+                            2,
+                        ),
+                    )
+                    unsqueeze_node.meta = copy_meta(
+                        input_node.meta, lambda m: {**m, "val": m["val"].unsqueeze(2)}
+                    )
+                    with graph_module.graph.inserting_after(unsqueeze_node):
+
+                        filter_node = node.args[1]
+                        filter_node.meta["val"] = (
+                            filter_node.meta["val"].unsqueeze(2).contiguous()
+                        )
+                        filter_tensor = get_parameter(filter_node, self.edge_program)
+                        # Ensure tensor is nn.Parameter type, so program does not fail during edge_program._validate()
+                        filter_tensor = nn.Parameter(filter_tensor.unsqueeze(2))
+                        set_parameter(filter_tensor, filter_node, self.edge_program)
+
+                        bias_node = node.args[2]
+                        stride = [1] + node.args[3]
+                        padding = [0] + node.args[4]
+                        dilation = [1] + node.args[5]
+                        transpose = node.args[6]
+                        output_padding = [0] + node.args[7]
+                        groups = node.args[8]
+
+                        conv2d_node = graph.create_node(
+                            "call_function",
+                            conv_op,
+                            (
+                                unsqueeze_node,
+                                filter_node,
+                                bias_node,
+                                stride,
+                                padding,
+                                dilation,
+                                transpose,
+                                output_padding,
+                                groups,
+                            ),
+                        )
+                        conv2d_node.meta = copy_meta(
+                            node.meta, lambda m: {**m, "val": m["val"].unsqueeze(2)}
+                        )
+
+                        with graph_module.graph.inserting_after(conv2d_node):
+                            squeeze_op = exir_ops.edge.aten.squeeze_copy.dims
+                            squeeze_node = graph.create_node(
+                                "call_function",
+                                squeeze_op,
+                                (
+                                    conv2d_node,
+                                    [2],
+                                ),
+                            )
+                            squeeze_node.meta = copy_meta(node.meta)
+                for user in node.users.copy():
+                    user.replace_input_with(node, squeeze_node)
+        graph.eliminate_dead_code()
+        graph_module.recompile()
+        return PassResult(graph_module, True)

backends/qualcomm/_passes/decompose_expm1.py

@@ -0,0 +1,46 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from executorch.exir.pass_base import ExportPass, PassResult
+
+from .utils import copy_meta
+
+
+class DecomposeExpM1(ExportPass):
+    """
+    Decompose for expm1 to exponential and minus 1.
+    """
+
+    def __init__(self, quantization_capture=False) -> None:
+        super().__init__()
+
+    def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        graph = graph_module.graph
+        for node in graph.nodes:
+            if node.target == torch.ops.aten.special_expm1.default:
+                input_node = node.args[0]
+                with graph_module.graph.inserting_after(input_node):
+                    exp_op = torch.ops.aten.exp.default
+                    exp_node = graph.create_node("call_function", exp_op, (input_node,))
+                    exp_node.meta = copy_meta(node.meta)
+                    with graph_module.graph.inserting_after(exp_node):
+                        sub_op = torch.ops.aten.sub.Tensor
+                        sub_node = graph.create_node(
+                            "call_function",
+                            sub_op,
+                            (
+                                exp_node,
+                                1,
+                            ),
+                        )
+                        sub_node.meta = copy_meta(node.meta)
+                for user in node.users.copy():
+                    user.replace_input_with(node, sub_node)
+
+        graph.eliminate_dead_code()
+        graph_module.recompile()
+        return PassResult(graph_module, True)

backends/qualcomm/_passes/decompose_silu.py

@@ -3,22 +3,17 @@
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
-from typing import Dict
 
 import torch
 from executorch.exir.pass_base import ExportPass, PassResult
 
+from .utils import copy_meta
+
 
 class DecomposeSilu(ExportPass):
     def __init__(self):
         super(DecomposeSilu, self).__init__()
 
-    def _copy_meta(self, meta: Dict):
-        copied = {}
-        for k, v in meta.items():
-            copied[k] = v
-        return copied
-
     def call(self, graph_module: torch.fx.GraphModule):
         graph = graph_module.graph
         for node in graph.nodes:
@@ -34,14 +29,14 @@ def call(self, graph_module: torch.fx.GraphModule):
                         torch.ops.aten.sigmoid.default,
                         (silu_node_input,),
                     )
-                    sigmoid_node.meta = self._copy_meta(silu_node.meta)
+                    sigmoid_node.meta = copy_meta(silu_node.meta)
                     with graph_module.graph.inserting_after(sigmoid_node):
                         mul_node = graph.create_node(
                             "call_function",
                             torch.ops.aten.mul.Tensor,
                             (silu_node_input, sigmoid_node),
                         )
-                        mul_node.meta = self._copy_meta(silu_node.meta)
+                        mul_node.meta = copy_meta(silu_node.meta)
                         for user in silu_node.users.copy():
                             user.replace_input_with(silu_node, mul_node)
 

backends/qualcomm/_passes/layout_transform.py

@@ -49,12 +49,15 @@ class LayoutTransform(ExportPass):
         exir_ops.edge.aten.add.Tensor,
         exir_ops.edge.aten.bitwise_or.Tensor,
         exir_ops.edge.aten.bmm.default,
+        exir_ops.edge.aten.bitwise_and.Tensor,
         exir_ops.edge.aten.cat.default,
         exir_ops.edge.aten.ceil.default,
         exir_ops.edge.aten.clamp.default,
         exir_ops.edge.aten.constant_pad_nd.default,
         exir_ops.edge.aten.div.Tensor,
+        exir_ops.edge.aten.elu.default,
         exir_ops.edge.aten.eq.Tensor,
+        exir_ops.edge.aten.exp.default,
         exir_ops.edge.aten.full.default,
         exir_ops.edge.aten.full_like.default,
         exir_ops.edge.aten.ge.Tensor,
@@ -87,10 +90,13 @@ class LayoutTransform(ExportPass):
         exir_ops.edge.aten.sqrt.default,
         exir_ops.edge.aten.sub.Tensor,
         exir_ops.edge.aten.sum.dim_IntList,
+        exir_ops.edge.aten.stack.default,
         exir_ops.edge.aten.topk.default,
         exir_ops.edge.aten._to_copy.default,
+        exir_ops.edge.aten.unbind.int,
         exir_ops.edge.aten.where.self,
         _operator.getitem,
+        torch.ops.aten.scalar_tensor.default,
     }
 
     layout_type = {

backends/qualcomm/_passes/lift_constant_scalar_operands.py

@@ -28,24 +28,27 @@ class TensorConstant:
 class TensorOpInfo:
     target: torch._ops.OpOverload
     use_schema_args: bool
+    use_self_dtype: bool
 
 
 SCALAR_OPS = {
-    aten.eq.Scalar: TensorOpInfo(aten.eq.Tensor, False),
-    aten.ge.Scalar: TensorOpInfo(aten.ge.Tensor, False),
-    aten.gt.Scalar: TensorOpInfo(aten.gt.Tensor, False),
-    aten.le.Scalar: TensorOpInfo(aten.le.Tensor, False),
-    aten.lt.Scalar: TensorOpInfo(aten.lt.Tensor, False),
-    aten.ne.Scalar: TensorOpInfo(aten.ne.Tensor, False),
-    aten.add.Scalar: TensorOpInfo(aten.add.Tensor, False),
-    aten.add_.Scalar: TensorOpInfo(aten.add_.Tensor, False),
-    aten.div.Scalar: TensorOpInfo(aten.div.Tensor, False),
-    aten.mul.Scalar: TensorOpInfo(aten.mul.Tensor, False),
-    aten.rsub.Scalar: TensorOpInfo(aten.rsub.Tensor, False),
-    aten.sub.Scalar: TensorOpInfo(aten.sub.Tensor, False),
-    aten.pow.Tensor_Scalar: TensorOpInfo(aten.pow.Tensor_Tensor, False),
+    aten.eq.Scalar: TensorOpInfo(aten.eq.Tensor, False, False),
+    aten.ge.Scalar: TensorOpInfo(aten.ge.Tensor, False, False),
+    aten.gt.Scalar: TensorOpInfo(aten.gt.Tensor, False, False),
+    aten.le.Scalar: TensorOpInfo(aten.le.Tensor, False, False),
+    aten.lt.Scalar: TensorOpInfo(aten.lt.Tensor, False, False),
+    aten.ne.Scalar: TensorOpInfo(aten.ne.Tensor, False, False),
+    aten.add.Scalar: TensorOpInfo(aten.add.Tensor, False, False),
+    aten.add_.Scalar: TensorOpInfo(aten.add_.Tensor, False, False),
+    aten.div.Scalar: TensorOpInfo(aten.div.Tensor, False, False),
+    aten.mul.Scalar: TensorOpInfo(aten.mul.Tensor, False, False),
+    aten.rsub.Scalar: TensorOpInfo(aten.rsub.Tensor, False, False),
+    aten.sub.Scalar: TensorOpInfo(aten.sub.Tensor, False, False),
+    aten.pow.Tensor_Scalar: TensorOpInfo(aten.pow.Tensor_Tensor, False, False),
     # The scalar number arg[1] is missing when using default. Result in a corner case to deal
-    aten.leaky_relu.default: TensorOpInfo(aten.prelu.default, True),
+    aten.leaky_relu.default: TensorOpInfo(aten.prelu.default, True, False),
+    aten.where.ScalarOther: TensorOpInfo(aten.where.self, False, True),
+    aten.where.Scalar: TensorOpInfo(aten.where.self, False, True),
 }
 
 
@@ -63,11 +66,14 @@ def __init__(self):
     def _build_tensor_constant(
         self, gm: torch.fx.GraphModule, node: fx.Node, const_val
     ) -> TensorConstant:
+        # For dtype, in some cases, we cannot use node.args[0] as scalar dtype.
+        # Ex: Where op args[0] can be bool, however, we probably want args[1] and args[2] to be dtype same as node.meta["val"] instead of bool type
         tensor = torch.tensor(
             [const_val],
             dtype=(
                 node.args[0].meta["val"].dtype
                 if not is_float_tensor(node)
+                and not SCALAR_OPS.get(node.target).use_self_dtype
                 else node.meta["val"].dtype
             ),
             device=node.meta["val"].device,

backends/qualcomm/_passes/replace_arange_args.py

@@ -0,0 +1,48 @@
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from executorch.exir.pass_base import ExportPass, PassResult
+
+from .utils import copy_meta
+
+
+class ReplaceArangeArgs(ExportPass):
+    """
+    During annotation, kwargs for arange will be removed due to restrictions by quantizer.
+    This causes arange to have no dtype, which means FP nodes might be inferred as INT nodes during calibration.
+    This can cause calibration to fail since QDQ can only be applied on FP nodes but not INT nodes.
+    To hint the dtype, we provide step size as 1.0 instead of 1, which makes the node a FP node.
+    """
+
+    def __init__(self, quantization_capture=False) -> None:
+        super().__init__()
+        self.quantization_capture = quantization_capture
+
+    def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        graph = graph_module.graph
+        for node in graph.nodes:
+            if node.target == torch.ops.aten.arange.default:
+                if torch.is_floating_point(node.meta["val"]) and len(node.args) == 1:
+                    with graph_module.graph.inserting_after(node):
+                        step_arange_op = torch.torch.ops.aten.arange.start_step
+                        step_arange_node = graph.create_node(
+                            "call_function",
+                            step_arange_op,
+                            (
+                                0,
+                                node.args[0],
+                                1.0,
+                            ),
+                        )
+                        step_arange_node.meta = copy_meta(node.meta)
+
+                        for user in node.users.copy():
+                            user.replace_input_with(node, step_arange_node)
+
+        graph.eliminate_dead_code()
+        graph_module.recompile()
+        return PassResult(graph_module, True)