Qualcomm AI Engine Direct - LLAMA2 Infrastructure (#2020)

Summary:
1. OPs
- Add pow_tensor_scalar op
- Add rsqrt op
- Add sigmoid op
- Refine axis handling of cat op
- Refine parameters related functions
2. Passes
- Add AnnotateDecomposed for unbind and stak op
- Add DecomposeSilu for quantizer
- Add ReplaceInfBuffer for quantizer
- Change pass name ConvertAddmmmmWithLinear to ConvertToLinear
- Change pass name ConvertScaledDotProductAttention to DecomposeScaledDotProductAttention
- Support more args for sdpa op in DecomposeScaledDotProductAttention
- Support mm case for ConvertToLinear
- Move q_ops and dq_ops to pass/utils.py
3. Tests
- Add dummy llama2 test script
- Add single op test cases
4. Others
- Fix error of popping missing buffer
- Reorder the order of test models
- Reorder the order of op in qnn_constant

Pull Request resolved: #2020

Reviewed By: kirklandsign

Differential Revision: D54010593

Pulled By: cccclai

fbshipit-source-id: 657994dc223cb9bd88a263bfc2479295384fcb4d

Author: Joey Tsai

backends/qualcomm/builders/__init__.py

@@ -30,10 +30,13 @@
     op_mean_dim,
     op_mul,
     op_pad,
+    op_pow,
     op_quantize,
     op_relu,
     op_reshape,
+    op_rsqrt,
     op_select_copy,
+    op_sigmoid,
     op_skip_ops,
     op_slice_copy,
     op_softmax,
@@ -70,10 +73,13 @@
     op_mean_dim,
     op_mul,
     op_pad,
+    op_pow,
     op_quantize,
     op_relu,
     op_reshape,
+    op_rsqrt,
     op_select_copy,
+    op_sigmoid,
     op_skip_ops,
     op_slice_copy,
     op_softmax,

backends/qualcomm/builders/op_cat.py

@@ -54,7 +54,14 @@ def define_node(
             nodes_to_wrappers,
         )
 
-        axis = cast(int, node.args[1])
+        # node args[1] might not exist
+        axis = 0
+        if len(node.args) == 2:
+            axis = cast(int, node.args[1])
+
+        if axis < 0:
+            axis += node.meta["val"].dim()
+
         if "axis_order" in node.meta:
             axis = node.meta["axis_order"].index(axis)
 

backends/qualcomm/builders/op_pow.py

@@ -0,0 +1,92 @@
+# 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 typing import Dict
+
+import executorch.backends.qualcomm.python.PyQnnWrapperAdaptor as PyQnnWrapper
+
+import torch
+from executorch.exir.dialects._ops import ops as exir_ops
+
+from .node_visitor import NodeVisitor, register_node_visitor
+from .qnn_constants import OpElementWisePower, QNN_OP_PACKAGE_NAME_QTI_AISW
+
+
+# TODO Add more class Like PowTensorTensor if needed
+@register_node_visitor
+class PowTensorScalar(NodeVisitor):
+    target = "aten.pow.Tensor_Scalar"
+
+    def __init__(self, *args) -> None:
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        nodes_to_wrappers: Dict[torch.fx.Node, PyQnnWrapper.TensorWrapper],
+    ) -> PyQnnWrapper.PyQnnOpWrapper:
+        out_tensor = self.get_tensor(node, node)
+        output_tensor_wrapper = self.define_tensor(
+            node,
+            out_tensor,
+            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
+            nodes_to_wrappers,
+        )
+        pow_output_tensors = [output_tensor_wrapper]
+
+        # tensor input
+        input_node = node.args[0]
+        input_tensor = self.get_tensor(input_node, node)
+
+        tensor_type = PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE
+
+        input_tensor_wrapper = self.define_tensor(
+            input_node,
+            input_tensor,
+            tensor_type,
+            nodes_to_wrappers,
+        )
+
+        # scalar input
+        scalar = node.args[1]
+        scalar_tensor = torch.full(input_tensor.size(), scalar).to(torch.float32)
+
+        # 'graph', 'name', 'op', 'target', 'args', and 'kwargs'
+        scalar_node = torch.fx.Node(
+            node.graph,
+            node.name + "_runtime_scalar",
+            "call_function",
+            exir_ops.edge.aten.full.default,
+            (),  # args
+            {},  # kwargs
+        )
+
+        if pow_quant_attrs := node.meta.get("quant_attrs"):
+            quant_attrs = pow_quant_attrs.copy()
+            quant_range = quant_attrs["quant_max"] - quant_attrs["quant_min"]
+            quant_attrs["zero_point"] = 0 if scalar >= 0 else quant_attrs["quant_max"]
+            quant_attrs["scale"] = (
+                scalar / quant_range if scalar >= 0 else -scalar / quant_range
+            )
+            scalar_node.meta["quant_attrs"] = quant_attrs
+
+        scalar_tensor_wrapper = self.define_tensor(
+            scalar_node,
+            scalar_tensor,
+            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_STATIC,
+            nodes_to_wrappers,
+        )
+
+        pow_input_tensors = [input_tensor_wrapper, scalar_tensor_wrapper]
+
+        pow_op = PyQnnWrapper.PyQnnOpWrapper(
+            node.name,
+            QNN_OP_PACKAGE_NAME_QTI_AISW,
+            OpElementWisePower.op_name,
+        )
+        pow_op.AddInputTensors(pow_input_tensors)
+        pow_op.AddOutputTensors(pow_output_tensors)
+
+        return pow_op

backends/qualcomm/builders/op_rsqrt.py

@@ -0,0 +1,55 @@
+# 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 typing import Dict
+
+import executorch.backends.qualcomm.python.PyQnnWrapperAdaptor as PyQnnWrapper
+
+import torch
+
+from .node_visitor import NodeVisitor, register_node_visitor
+from .qnn_constants import OpElementWiseRsqrt, QNN_OP_PACKAGE_NAME_QTI_AISW
+
+
+@register_node_visitor
+class Rsqrt(NodeVisitor):
+    target = "aten.rsqrt.default"
+
+    def __init__(self, *args) -> None:
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        nodes_to_wrappers: Dict[torch.fx.Node, PyQnnWrapper.TensorWrapper],
+    ) -> PyQnnWrapper.PyQnnOpWrapper:
+        input_node = node.args[0]
+        input_tensor = self.get_tensor(input_node, node)
+        rsqrt_inp_tensor_wrapper = self.define_tensor(
+            input_node,
+            input_tensor,
+            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
+            nodes_to_wrappers,
+        )
+        rsqrt_input_tensors = [rsqrt_inp_tensor_wrapper]
+
+        output_tensor = self.get_tensor(node, node)
+        output_tensor_wrapper = self.define_tensor(
+            node,
+            output_tensor,
+            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
+            nodes_to_wrappers,
+        )
+        rsqrt_output_tensors = [output_tensor_wrapper]
+
+        rsqrt_op = PyQnnWrapper.PyQnnOpWrapper(
+            node.name,
+            QNN_OP_PACKAGE_NAME_QTI_AISW,
+            OpElementWiseRsqrt.op_name,
+        )
+        rsqrt_op.AddInputTensors(rsqrt_input_tensors)
+        rsqrt_op.AddOutputTensors(rsqrt_output_tensors)
+
+        return rsqrt_op

backends/qualcomm/builders/op_sigmoid.py

@@ -0,0 +1,55 @@
+# 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 typing import Dict
+
+import executorch.backends.qualcomm.python.PyQnnWrapperAdaptor as PyQnnWrapper
+
+import torch
+
+from .node_visitor import NodeVisitor, register_node_visitor
+from .qnn_constants import OpSigmoid, QNN_OP_PACKAGE_NAME_QTI_AISW
+
+
+@register_node_visitor
+class Sigmoid(NodeVisitor):
+    target = "aten.sigmoid.default"
+
+    def __init__(self, *args) -> None:
+        super().__init__(*args)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+        nodes_to_wrappers: Dict[torch.fx.Node, PyQnnWrapper.TensorWrapper],
+    ) -> PyQnnWrapper.PyQnnOpWrapper:
+        input_node = node.args[0]
+        input_tensor = self.get_tensor(input_node, node)
+        sigmoid_inp_tensor_wrapper = self.define_tensor(
+            input_node,
+            input_tensor,
+            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
+            nodes_to_wrappers,
+        )
+        sigmoid_input_tensors = [sigmoid_inp_tensor_wrapper]
+
+        output_tensor = self.get_tensor(node, node)
+        output_tensor_wrapper = self.define_tensor(
+            node,
+            output_tensor,
+            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
+            nodes_to_wrappers,
+        )
+        sigmoid_output_tensors = [output_tensor_wrapper]
+
+        sigmoid_op = PyQnnWrapper.PyQnnOpWrapper(
+            node.name,
+            QNN_OP_PACKAGE_NAME_QTI_AISW,
+            OpSigmoid.op_name,
+        )
+        sigmoid_op.AddInputTensors(sigmoid_input_tensors)
+        sigmoid_op.AddOutputTensors(sigmoid_output_tensors)
+
+        return sigmoid_op