Revert "Search graph for quantization nodes (#6452)"

This reverts commit 63017e4.

Author: digantdesai

backends/arm/_passes/annotate_channels_last_dim_order_pass.py

@@ -14,7 +14,7 @@
     get_first_fake_tensor,
     insert_q_dq_pair,
 )
-from executorch.backends.arm.tosa_quant_utils import dq_op, q_op, register_passable_op
+from executorch.backends.arm.tosa_quant_utils import dq_op, q_op
 from executorch.backends.arm.tosa_utils import is_consumer_node_depthwise_conv2d
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass, PassResult
@@ -42,9 +42,6 @@ def _transpose_impl(*args, **kwargs):
     return args[0]
 
 
-register_passable_op(torch.ops.passthrough_to_tosa._transpose)
-
-
 class AnnotateChannelsLastDimOrder(ExportPass):
     """
     Annotates each node with a tosa_dim_order. tosa_dim_order can be seen as a channels-last dim-order

backends/arm/_passes/insert_squeeze_after_sum_pass.py

@@ -8,7 +8,9 @@
 
 import torch
 import torch.fx
-from executorch.backends.arm._passes.arm_pass_utils import create_node
+from executorch.backends.arm._passes.arm_pass_utils import create_node, insert_q_dq_pair
+
+from executorch.backends.arm.tosa_quant_utils import get_quant_node_args, is_quant_node
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass, PassResult
 
@@ -26,6 +28,8 @@ class InsertSqueezeAfterSumPass(ExportPass):
         sum(dims, keep_dim = False)
     After pass:
         sum(dims, keep_dim = True)
+        (q)
+        (dq)
         squeeze(dim = dims)
     """
 
@@ -41,6 +45,12 @@ def call(self, graph_module: torch.fx.GraphModule):
                 continue
 
             dim_list = cast(list[int], sum_node.args[1])
+            quantized = is_quant_node(sum_node)
+            if quantized:
+                qparams = get_quant_node_args(sum_node.all_input_nodes[0])
+                qparams = qparams + (torch.int8,)
+            else:
+                qparams = None
 
             # Add keep_dim = True arg to sum node.
             sum_node.args = sum_node.args[0:2] + (True,)
@@ -51,6 +61,8 @@ def call(self, graph_module: torch.fx.GraphModule):
                 )
                 sum_node.replace_all_uses_with(squeeze_node)
                 squeeze_node.args = (sum_node, dim_list)
+                if quantized:
+                    sum_node = insert_q_dq_pair(graph_module.graph, sum_node, qparams)
         graph_module.graph.eliminate_dead_code()
         graph_module.recompile()
         graph_module = super().call(graph_module).graph_module

backends/arm/_passes/size_adjust_conv2d_pass.py

@@ -9,7 +9,7 @@
 from typing import cast, Optional
 
 import torch.fx
-from executorch.backends.arm.tosa_quant_utils import is_node_quantized
+from executorch.backends.arm.tosa_quant_utils import is_quant_node
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import ExportPass, PassResult
 from torch._ops import OpOverload
@@ -113,7 +113,7 @@ def call(self, graph_module: torch.fx.GraphModule):
                     slice_node = graph.create_node(
                         "call_function", self.slice_op, (last_node,) + args
                     )
-                    if is_node_quantized(last_node):
+                    if is_quant_node(last_node):
                         q_params = last_node.args[1:]
                         dq_node = insert_q_dq_pair(
                             graph_module.graph, slice_node, q_params

backends/arm/operators/op_addmm.py

@@ -14,13 +14,10 @@
     register_node_visitor,
 )
 from executorch.backends.arm.tosa_mapping import TosaArg
-from executorch.backends.arm.tosa_quant_utils import (
-    build_rescale,
-    search_quant_arg_downstream,
-    search_quant_arg_upstream,
-)
+from executorch.backends.arm.tosa_quant_utils import build_rescale, get_quant_node_args
 
 from executorch.backends.arm.tosa_utils import build_reshape
+from executorch.exir.dialects._ops import ops as exir_ops
 from serializer.tosa_serializer import TosaOp
 
 
@@ -70,7 +67,12 @@ def define_node(
         input_zp = 0
         if is_quant_node:
             input_node = node.all_input_nodes[1]
-            input_zp = search_quant_arg_upstream(input_node).zp
+            # rank > 2 linear layer
+            if input_node.target == exir_ops.edge.aten.view_copy.default:
+                quant_node = input_node.all_input_nodes[0]
+            else:
+                quant_node = input_node
+            input_zp = get_quant_node_args(quant_node).zp
         attr.ConvAttribute(
             pad=pad_attr,
             stride=stride_attr,
@@ -105,16 +107,24 @@ def define_node(
             # Read inputs' parent nodes
             _, input_node, weight_node = node.all_input_nodes
 
-            qargs = search_quant_arg_upstream(input_node)
-            input_scale = qargs.scale
-            consumer_node = list(node.users)[0]
-            quant_args = search_quant_arg_downstream(consumer_node)
-
-            consumer_node_scale = quant_args.scale
-            consumer_node_node_zp = quant_args.zp
+            # rank > 2 linear layer
+            if input_node.target == exir_ops.edge.aten.view_copy.default:
+                quant_node = input_node.all_input_nodes[0]
+                input_scale = get_quant_node_args(quant_node).scale
+                consumer_node = list(node.users)[0]
+                consumer_consumer_node = list(consumer_node.users)[0]
+                quant_args = get_quant_node_args(consumer_consumer_node)
+                consumer_node_scale = quant_args.scale
+                consumer_node_node_zp = quant_args.zp
+            else:
+                input_scale = get_quant_node_args(input_node).scale
+                consumer_node = list(node.users)[0]
+                quant_args = get_quant_node_args(consumer_node)
+                consumer_node_scale = quant_args.scale
+                consumer_node_node_zp = quant_args.zp
 
             weight_node_q_node = weight_node.all_input_nodes[0]
-            weight_scale = search_quant_arg_upstream(weight_node_q_node).scale
+            weight_scale = get_quant_node_args(weight_node_q_node).scale
 
             output_rescale_scale = (input_scale * weight_scale) / consumer_node_scale
 

backends/arm/operators/op_bmm.py

@@ -14,11 +14,7 @@
     register_node_visitor,
 )
 from executorch.backends.arm.tosa_mapping import TosaArg
-from executorch.backends.arm.tosa_quant_utils import (
-    build_rescale,
-    search_quant_arg_downstream,
-    search_quant_arg_upstream,
-)
+from executorch.backends.arm.tosa_quant_utils import build_rescale, get_quant_node_args
 from executorch.backends.arm.tosa_utils import get_two_inputs
 from serializer.tosa_serializer import TosaOp
 
@@ -46,10 +42,8 @@ def define_node(
         # For INT8, we need to get the zero points and add an intermediate tensor
         # for a later rescale.
         if is_quant_node:
-            input0_q_params = search_quant_arg_upstream(input0)
-            input1_q_params = search_quant_arg_upstream(input1)
-            input0_zp = input0_q_params.zp
-            input1_zp = input1_q_params.zp
+            input0_zp = get_quant_node_args(input0).zp
+            input1_zp = get_quant_node_args(input1).zp
             bmm_result = tosa_graph.addIntermediate(output.shape, ts.DType.INT32)
             bmm_output_name = bmm_result.name
         else:
@@ -69,7 +63,9 @@ def define_node(
 
         # As INT8 accumulates into INT32, we need to rescale it back to INT8
         if is_quant_node:
-            output_q_params = search_quant_arg_downstream(list(node.users)[0])
+            input0_q_params = get_quant_node_args(input0)
+            input1_q_params = get_quant_node_args(input1)
+            output_q_params = get_quant_node_args(list(node.users)[0])
 
             final_output_scale = (
                 input0_q_params.scale * input1_q_params.scale

backends/arm/operators/op_conv2d.py

@@ -4,7 +4,7 @@
 # LICENSE file in the root directory of this source tree.
 
 # pyre-unsafe
-from typing import List
+from typing import cast, List
 
 import serializer.tosa_serializer as ts
 import torch
@@ -15,10 +15,9 @@
 from executorch.backends.arm.tosa_mapping import TosaArg
 from executorch.backends.arm.tosa_quant_utils import (
     build_rescale_conv_output,
-    search_quant_arg_downstream,
-    search_quant_arg_upstream,
+    get_quant_node_args,
 )
-from executorch.backends.arm.tosa_utils import build_reshape, tosa_shape
+from executorch.backends.arm.tosa_utils import build_reshape, getNodeArgs, tosa_shape
 
 from serializer.tosa_serializer import TosaOp
 
@@ -83,9 +82,7 @@ def define_node(
         )
 
         input_zp = (
-            search_quant_arg_upstream(node.all_input_nodes[0]).zp
-            if is_quant_node
-            else 0
+            get_quant_node_args(node.all_input_nodes[0]).zp if is_quant_node else 0
         )
 
         attr.ConvAttribute(
@@ -161,10 +158,9 @@ def define_node(
         # integer value domain of the next op. Otherwise return float32 output.
         if is_quant_node:
             # Get scale_factor from input, weight, and output.
-            input_scale = search_quant_arg_upstream(node.all_input_nodes[0]).scale
-            weight_scale = search_quant_arg_upstream(node.all_input_nodes[1]).scale
-            output_qargs = search_quant_arg_downstream(list(node.users)[0])
-
+            _, input_scale, _, _, _, _ = getNodeArgs(cast(torch.fx.Node, node.args[0]))
+            _, weight_scale, _, _, _, _ = getNodeArgs(cast(torch.fx.Node, node.args[1]))
+            _, output_scale, output_zp, _, _, _ = getNodeArgs(list(node.users)[0])
             build_rescale_conv_output(
                 tosa_graph,
                 # pyre-fixme[61]: Uninitialized local [61]: Local variable `conv2d_res` is undefined, or not always defined.
@@ -173,6 +169,6 @@ def define_node(
                 actual_out_type,
                 input_scale,
                 weight_scale,
-                output_qargs.scale,
-                output_qargs.zp,
+                output_scale,
+                output_zp,
             )

backends/arm/operators/op_exp.py

@@ -17,10 +17,9 @@
 
 from executorch.backends.arm.tosa_quant_utils import (
     dequantize_value,
+    get_quant_node_args,
     QuantArgs,
     quantize_value,
-    search_quant_arg_downstream,
-    search_quant_arg_upstream,
 )
 from serializer.tosa_serializer import TosaOp
 from torch.fx import Node
@@ -49,9 +48,9 @@ def define_node(
 
             # Create attribute for 8 bit table lookup.
             input_node = node.all_input_nodes[0]
-            in_quantargs = search_quant_arg_upstream(input_node)
+            in_quantargs = get_quant_node_args(input_node)
             output_node = list(node.users)[0]
-            out_quantargs = search_quant_arg_downstream(output_node)
+            out_quantargs = get_quant_node_args(output_node)
 
             table = exp_table_8bit(in_quantargs, out_quantargs)
             table_attr = ts.TosaSerializerAttribute()

backends/arm/operators/op_full.py

@@ -14,10 +14,7 @@
     register_node_visitor,
 )
 from executorch.backends.arm.tosa_mapping import TosaArg
-from executorch.backends.arm.tosa_quant_utils import (
-    quantize_value,
-    search_quant_arg_downstream,
-)
+from executorch.backends.arm.tosa_quant_utils import get_quant_node_args
 from executorch.backends.arm.tosa_utils import tosa_shape
 from torch.fx import Node
 
@@ -42,8 +39,10 @@ def define_node(
 
         value = inputs[1].number
         if is_quant_node:
-            qargs = search_quant_arg_downstream(list(node.users)[0])
-            qvalue = quantize_value(value, qargs)
+            qargs = get_quant_node_args(list(node.users)[0])
+            qvalue = np.clip(
+                np.round(value / qargs.scale) + qargs.zp, qargs.qmin, qargs.qmax
+            )
             dtype = ts.DType.INT8
             data = np.full(shape, qvalue, dtype=np.int8)
         else:

backends/arm/operators/op_hardtanh.py

@@ -14,10 +14,7 @@
 )
 from executorch.backends.arm.tosa_mapping import TosaArg
 
-from executorch.backends.arm.tosa_quant_utils import (
-    quantize_value,
-    search_quant_arg_upstream,
-)
+from executorch.backends.arm.tosa_quant_utils import get_quant_node_args
 from serializer.tosa_serializer import TosaOp
 
 
@@ -40,10 +37,12 @@ def define_node(
 
         if is_quant_node:
             # Get quant parameters
-            qargs = search_quant_arg_upstream(node.all_input_nodes[0])
+            scale, zp, qmin, qmax = get_quant_node_args(node.all_input_nodes[0])
             # Convert to quantized representation
-            clamp_min_qs = quantize_value(inputs[1].number, qargs)
-            clamp_max_qs = quantize_value(inputs[2].number, qargs)
+            clamp_min_qs = round((inputs[1].number / scale) + zp)
+            clamp_min_qs = max(clamp_min_qs, qmin)
+            clamp_max_qs = round((inputs[2].number / scale) + zp)
+            clamp_max_qs = min(clamp_max_qs, qmax)
             # Set fp values to 0.0 since they are not used
             clamp_min_fp = 0.0
             clamp_max_fp = 0.0

backends/arm/operators/op_log.py

@@ -17,10 +17,9 @@
 
 from executorch.backends.arm.tosa_quant_utils import (
     dequantize_value,
+    get_quant_node_args,
     QuantArgs,
     quantize_value,
-    search_quant_arg_downstream,
-    search_quant_arg_upstream,
 )
 from serializer.tosa_serializer import TosaOp
 from torch.fx import Node
@@ -50,9 +49,9 @@ def define_node(
 
             # Create attribute for 8 bit table lookup.
             input_node = node.all_input_nodes[0]
-            in_quantargs = search_quant_arg_upstream(input_node)
+            in_quantargs = get_quant_node_args(input_node)
             output_node = list(node.users)[0]
-            out_quantargs = search_quant_arg_downstream(output_node)
+            out_quantargs = get_quant_node_args(output_node)
 
             table = log_table_8bit(in_quantargs, out_quantargs)
             table_attr = ts.TosaSerializerAttribute()

backends/arm/operators/op_mm.py

@@ -14,11 +14,7 @@
     register_node_visitor,
 )
 from executorch.backends.arm.tosa_mapping import TosaArg
-from executorch.backends.arm.tosa_quant_utils import (
-    build_rescale,
-    search_quant_arg_downstream,
-    search_quant_arg_upstream,
-)
+from executorch.backends.arm.tosa_quant_utils import build_rescale, get_quant_node_args
 from executorch.backends.arm.tosa_utils import (
     build_reshape,
     expand_dims,
@@ -58,8 +54,8 @@ def define_node(
         # For INT8, we need to get the zero point, otherwise it is 0
         input0_zp, input1_zp = 0, 0
         if is_quant_node:
-            input0_zp = search_quant_arg_upstream(input0).zp
-            input1_zp = search_quant_arg_upstream(input1).zp
+            input0_zp = get_quant_node_args(input0).zp
+            input1_zp = get_quant_node_args(input1).zp
 
         mat_mul_result = tosa_graph.addIntermediate(
             output_new_shape, ts.DType.INT32 if is_quant_node else output.dtype
@@ -90,9 +86,9 @@ def define_node(
 
         # As INT8 accumulates into INT32, we need to rescale it back to INT8
         if is_quant_node:
-            input0_q_params = search_quant_arg_upstream(input0)
-            input1_q_params = search_quant_arg_upstream(input1)
-            output_q_params = search_quant_arg_downstream(list(node.users)[0])
+            input0_q_params = get_quant_node_args(input0)
+            input1_q_params = get_quant_node_args(input1)
+            output_q_params = get_quant_node_args(list(node.users)[0])
 
             final_output_scale = (
                 input0_q_params.scale * input1_q_params.scale

backends/arm/operators/op_mul.py

@@ -37,10 +37,10 @@ def define_node(
         if is_quant_node:
             input_A = inputs[0]
             input_B = inputs[1]
-            input_A_qargs = tqutils.search_quant_arg_upstream(
+            input_A_qargs = tqutils.get_quant_node_args(
                 cast(torch.fx.Node, node.args[0])
             )
-            input_B_qargs = tqutils.search_quant_arg_upstream(
+            input_B_qargs = tqutils.get_quant_node_args(
                 cast(torch.fx.Node, node.args[1])
             )