Qualcomm AI Engine Direct - Suite Operator Test Support (Part1) (pytorch#14618)

### Summary

- Support Add/Sub with alpha values
- Support Conv3d
- Support TransposeConv3d

### Test plan
UT added

Author: winskuo-quic

backends/qualcomm/_passes/__init__.py

@@ -13,6 +13,7 @@
 from .convert_linear_to_conv2d import ConvertLinearToConv2d
 from .convert_square_to_pow import ConvertSquareToPow
 from .decompose_any import DecomposeAny
+from .decompose_binary_alpha import DecomposeBinaryAlpha
 from .decompose_cdist import DecomposeCDist
 from .decompose_col_im import DecomposeColIm
 from .decompose_einsum import DecomposeEinsum
@@ -54,6 +55,7 @@
     ConvertLinearToConv2d,
     ConvertSquareToPow,
     DecomposeAny,
+    DecomposeBinaryAlpha,
     DecomposeCDist,
     DecomposeColIm,
     DecomposeEinsum,

backends/qualcomm/_passes/canonicalize_conv.py

@@ -34,6 +34,7 @@ def __init__(self, edge_program: torch.export.ExportedProgram):
         self.transpose_conv_set = {
             torch.ops.aten.conv_transpose1d.default,
             torch.ops.aten.conv_transpose2d.input,
+            torch.ops.aten.conv_transpose3d.input,
         }
 
     def dilate(self, tensor, dilation):

backends/qualcomm/_passes/decompose_binary_alpha.py

@@ -0,0 +1,61 @@
+# 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
+
+decomp_set = {torch.ops.aten.add.Tensor, torch.ops.aten.sub.Tensor}
+
+
+class DecomposeBinaryAlpha(ExportPass):
+    """
+    QNN does not support alpha parameter for add/sub.
+    Decompose to mul + add / mul + sub
+    """
+
+    def __init__(self) -> None:
+        super().__init__()
+
+    def call(self, graph_module: torch.fx.GraphModule) -> PassResult:
+        graph = graph_module.graph
+        for node in graph.nodes:
+            if (
+                node.target in decomp_set
+                and "alpha" in node.kwargs
+                and node.kwargs["alpha"] != 1
+            ):
+                alpha = node.kwargs["alpha"]
+                # Remove alpha from immutable dict
+                node.kwargs = {k: v for k, v in node.kwargs.items() if k != "alpha"}
+                input2_node = node.args[1]
+                # If input2 is constant, we can just multiply the value for optimization
+                if isinstance(input2_node, (int, float)):
+                    arg_list = list(node.args)
+                    arg_list[1] = input2_node * alpha
+                    node.args = tuple(arg_list)
+                    continue
+                with graph.inserting_before(node):
+                    mul_op = torch.ops.aten.mul.Scalar
+                    mul_node = graph.create_node(
+                        "call_function",
+                        mul_op,
+                        (
+                            input2_node,
+                            alpha,
+                        ),
+                    )
+                    mul_node.meta = copy_meta(node.meta)
+                    node.replace_input_with(input2_node, mul_node)
+                    node.args = (
+                        node.args[0],
+                        mul_node,
+                    )
+
+        graph.eliminate_dead_code()
+        graph_module.recompile()
+        return PassResult(graph_module, True)

backends/qualcomm/_passes/qnn_pass_manager.py

@@ -18,6 +18,7 @@
     ConvertLinearToConv2d,
     ConvertSquareToPow,
     DecomposeAny,
+    DecomposeBinaryAlpha,
     DecomposeCDist,
     DecomposeColIm,
     DecomposeEinsum,
@@ -194,6 +195,7 @@ def transform_for_annotation_pipeline(self, graph_module: GraphModule):
         self.add_pass(RecomposePixelUnshuffle(quantization_capture=True))
         self.add_pass(RecomposeRmsNorm(quantization_capture=True))
         self.add_pass(ReplaceArangeArgs())
+        self.add_pass(DecomposeBinaryAlpha())
         self.add_pass(DecomposeCDist())
         self.add_pass(DecomposeScaledDotProductAttention())
         self.add_pass(DecomposeRoll())
@@ -210,6 +212,7 @@ def transform_for_annotation_pipeline(self, graph_module: GraphModule):
     def transform_for_export_pipeline(
         self, exported_program: ExportedProgram, convert_linear_to_conv2d: bool = False
     ):
+        self.add_pass(DecomposeBinaryAlpha())
         self.add_pass(DecomposeCDist())
         self.add_pass(DecomposeScaledDotProductAttention())
         self.add_pass(DecomposeRoll())

backends/qualcomm/builders/README.md

@@ -365,7 +365,7 @@ Please help update following table if you are contributing new operators:
 + 🚫 = Deprecated, supported with other QNN Ops
 
 
-| Operators | HTP - 90/116 Enabled |
+| Operators | HTP - 92/116 Enabled |
 |-----------|---------|
 | Argmax | ✓ |
 | Argmin | ✓ |
@@ -375,7 +375,7 @@ Please help update following table if you are contributing new operators:
 | ChannelShuffle | ✗ |
 | Concat | ✓ |
 | Conv2d | ✓ |
-| Conv3d | ✗ |
+| Conv3d | ✓ |
 | Convert | ✓ |
 | CreateSparse | ✗ |
 | CumulativeSum | ✓ |
@@ -481,7 +481,7 @@ Please help update following table if you are contributing new operators:
 | TopK | ✓ |
 | TransPose | ✓ |
 | TransPoseConv2d | ✓ |
-| TransPoseConv3d | ✗ |
+| TransPoseConv3d | ✓ |
 | Unpack | ✓ |
 
 ## Issues

backends/qualcomm/builders/__init__.py

@@ -24,7 +24,7 @@
     op_cat,
     op_ceil,
     op_clamp,
-    op_conv2d,
+    op_conv,
     op_copy,
     op_cos,
     op_cum_sum,
@@ -129,7 +129,7 @@
     op_cat,
     op_ceil,
     op_clamp,
-    op_conv2d,
+    op_conv,
     op_copy,
     op_cos,
     op_cum_sum,

backends/qualcomm/builders/op_conv2d.py renamed to backends/qualcomm/builders/op_conv.py

@@ -7,7 +7,6 @@
 from typing import cast, Dict, List
 
 import executorch.backends.qualcomm.python.PyQnnWrapperAdaptor as PyQnnWrapper
-
 import numpy as np
 import torch
 from executorch.backends.qualcomm.utils.constants import QCOM_DATA
@@ -16,8 +15,10 @@
 from .node_visitor_manager import register_node_visitor
 from .qnn_constants import (
     OpConv2d,
+    OpConv3d,
     OpDepthWiseConv2d,
     OpTransposeConv2d,
+    OpTransposeConv3d,
     QNN_OP_PACKAGE_NAME_QTI_AISW,
 )
 from .utils import get_parameter
@@ -66,7 +67,7 @@ def _add_conv_op_parameter(
             len(padding_shape),
             padding_shape,
             np.array(
-                [[padding[0], padding[0]], [padding[1], padding[1]]],
+                padding,
                 dtype=np.uint32,
             ),
             True,
@@ -108,8 +109,14 @@ def define_node(
         input_node = self.get_node(node.args[0])
         input_tensor = self.get_tensor(input_node, node)
         assert (
-            input_tensor.dim() == 4
+            input_tensor.dim() != 3
         ), "All Conv1D should be converted to Conv2D in CanonicalizeConv,"
+        assert input_tensor.dim() in {
+            4,
+            5,
+        }, "Only Conv2d and Conv3d is supported in conv builder,"
+
+        is_conv2d = input_tensor.dim() == 4
         input_tensor_wrapper = self.define_tensor(
             input_node,
             node,
@@ -120,9 +127,15 @@ def define_node(
 
         filter_node = self.get_node(node.args[1])
         filter_tensor = get_parameter(filter_node, self.edge_program)
-        # weight of pytorch OIHW(conv2d) | IOHW(conv_transpose2d), yet QNN is HWIO
+        # weight of pytorch OIHW(conv2d) / OIDHW(conv3d) or IOHW(conv_transpose2d) / IODHW(conv_transpose3d),
+        # yet QNN is HWIO or DHWIO
         is_transpose_conv = cast(bool, node.args[6])
-        filter_axis_order = (2, 3, 0, 1) if is_transpose_conv else (2, 3, 1, 0)
+        if is_conv2d:
+            filter_axis_order = (2, 3, 0, 1) if is_transpose_conv else (2, 3, 1, 0)
+        else:
+            filter_axis_order = (
+                (2, 3, 4, 0, 1) if is_transpose_conv else (2, 3, 4, 1, 0)
+            )
         filter_tensor = filter_tensor.permute(dims=filter_axis_order).contiguous()
         filter_tensor_wrapper = self.define_tensor(
             filter_node,
@@ -132,7 +145,6 @@ def define_node(
             nodes_to_wrappers,
         )
         conv_input_tensors = [input_tensor_wrapper, filter_tensor_wrapper]
-
         if node.args[2] is not None:
             bias_node = self.get_node(node.args[2])
             bias_tensor = get_parameter(bias_node, self.edge_program)
@@ -159,11 +171,10 @@ def define_node(
         padding = cast(List[int], node.args[4])
         dilation = cast(List[int], node.args[5])
         output_padding = cast(List[int], node.args[7])
-
         groups = cast(int, node.args[8])
-        # Qnn filter tensor is (H, W, Cin, Cout)
-        group_input_channels = filter_tensor.shape[2]
-        group_output_channels = int(filter_tensor.shape[3] / groups)
+        # Qnn filter tensor is (H, W, Cin, Cout) or (D, H, W, Cin, Cout)
+        group_input_channels = filter_tensor.shape[-2]
+        group_output_channels = int(filter_tensor.shape[-1] / groups)
         # 1) groups = input_channels (i.e. group_input_channels = 1)
         # 2) output_channels is a positive integer multiple of input channels
         # TODO: Currently, negative results will be zero with Depthwise conv2d when input_channel == groups == 1
@@ -175,18 +186,23 @@ def define_node(
         )
         if len(padding) == 1:
             padding = padding + padding
+        padding = [[x, x] for x in padding]
 
         stride_shape = [len(stride)]
-        padding_shape = [2, 2]
+        padding_shape = [len(padding), len(padding[0])]
         dilation_shape = [len(dilation)]
         output_padding_shape = [len(output_padding)]
 
-        if is_depthwise_conv:
+        if is_transpose_conv:
+            assert all(
+                val == 1 for val in dilation
+            ), "CanonicalizeConv pass should perform dilate for transpose_conv."
+            op_class = OpTransposeConv2d if is_conv2d else OpTransposeConv3d
+        elif is_depthwise_conv:
+            assert is_conv2d, "DepthWise only supports Conv2d"
             op_class = OpDepthWiseConv2d
-        elif is_transpose_conv:
-            op_class = OpTransposeConv2d
         else:
-            op_class = OpConv2d
+            op_class = OpConv2d if is_conv2d else OpConv3d
 
         conv_op = PyQnnWrapper.PyQnnOpWrapper(
             node.name,

backends/qualcomm/builders/qnn_constants.py

@@ -59,6 +59,15 @@ class OpConv2d:
     param_dilation: str = "dilation"
 
 
+@dataclass(init=False, frozen=True)
+class OpConv3d:
+    op_name: str = "Conv3d"
+    param_stride: str = "stride"
+    param_pad_amount: str = "pad_amount"
+    param_group: str = "group"
+    param_dilation: str = "dilation"
+
+
 @dataclass(init=False, frozen=True)
 class OpConvert:
     op_name: str = "Convert"
@@ -573,6 +582,15 @@ class OpTransposeConv2d:
     param_output_padding: str = "output_padding"
 
 
+@dataclass(init=False, frozen=True)
+class OpTransposeConv3d:
+    op_name: str = "TransposeConv3d"
+    param_stride: str = "stride"
+    param_pad_amount: str = "pad_amount"
+    param_group: str = "group"
+    param_output_padding: str = "output_padding"
+
+
 @dataclass(init=False, frozen=True)
 class OpUnpack:
     op_name: str = "UnPack"

backends/qualcomm/quantizer/annotators.py

@@ -1094,11 +1094,13 @@ def annotate_cdist(node: Node, quantization_config: QuantizationConfig) -> None:
 
 @register_annotator(
     [
+        torch.ops.aten.conv1d.default,
         torch.ops.aten.conv2d.default,
         torch.ops.aten.conv2d.padding,
-        torch.ops.aten.conv1d.default,
-        torch.ops.aten.conv_transpose2d.input,
+        torch.ops.aten.conv3d.default,
         torch.ops.aten.conv_transpose1d.default,
+        torch.ops.aten.conv_transpose2d.input,
+        torch.ops.aten.conv_transpose3d.input,
         torch.ops.aten.convolution.default,
     ]
 )