Update on "use-pt-pinned-commit for test-arm-{backend,reference}-delegation"

Without this, these builds don't respect the torchgen pinned commit and thus fail with #7546.

Differential Revision: [D67996459](https://our.internmc.facebook.com/intern/diff/D67996459/)

[ghstack-poisoned]

Author: swolchok

backends/cadence/aot/compiler.py

@@ -33,6 +33,7 @@
     ExecutorchProgramManager,
     to_edge,
 )
+from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.pass_base import PassResult
 from executorch.exir.passes import ToOutVarPass
 from executorch.exir.passes.sym_shape_eval_pass import HintBasedSymShapeEvalPass
@@ -186,14 +187,17 @@ def export_to_edge(
     edge_prog_manager = to_edge(
         expo_program,
         compile_config=EdgeCompileConfig(
-            _skip_dim_order=True,
             # Allow specific non-core aten ops in the IR.
             _core_aten_ops_exception_list=[
                 torch.ops.aten._native_batch_norm_legit_functional.default,
                 torch.ops.aten.linear.default,
                 torch.ops.aten.linalg_vector_norm.default,
                 torch.ops.aten.unfold.default,
                 torch.ops.aten.angle.default,
+                # cadence replaced to_dim_order_copy with _to_copy for performance
+                # skip _to_copy op to get around of dim order check
+                # We should remove this op once cadence can support dim order
+                exir_ops.edge.aten._to_copy.default,
             ],
         ),
         constant_methods=constant_methods,

backends/cadence/aot/replace_ops.py

@@ -11,6 +11,7 @@
 
 # pyre-unsafe
 
+import copy
 import math
 from operator import neg
 from typing import cast, Dict, Iterable, Sequence, Set, Tuple
@@ -35,7 +36,12 @@
 from executorch.backends.cadence.aot.utils import get_edge_overload_packet
 from executorch.exir.dialects._ops import ops as exir_ops
 from executorch.exir.dialects.edge._ops import EdgeOpOverload, EdgeOpOverloadPacket
+from executorch.exir.dim_order_utils import get_memory_format
 from executorch.exir.pass_base import ExportPass, NodeMetadata, PassResult, ProxyValue
+from executorch.exir.passes.dim_order_ops_registry import (
+    DimOrderOpsMap,
+    MemoryFormatOpsMap,
+)
 from torch._subclasses import FakeTensor
 from torch.fx.node import Argument
 
@@ -1799,6 +1805,72 @@ def call_operator(
         )
 
 
+@register_cadence_pass(CadencePassAttribute(opt_level=0))
+class ReplaceToDimOrderCopyWithToCopyPass(ExportPass):
+    """
+    dim_order_ops::to_dim_order_copy is not supported, so this is an opt_level=0 pass.
+    If the dim order is sequential, we don't need the extra work with strides and
+    can just use to_copy.
+    """
+
+    def call_operator(
+        self,
+        op,
+        args: Tuple[Argument, ...],
+        kwargs: Dict[str, Argument],
+        meta: NodeMetadata,
+    ) -> ProxyValue:
+        if op not in DimOrderOpsMap:
+            return super().call_operator(op, args, kwargs, meta)
+
+        # new kwargs with dim_order, and no memory_format for the new op
+        nkwargs = dict(copy.deepcopy(kwargs))  # orig kwargs are immutable
+
+        ndim = None
+
+        # can always get the shape, assuming rank is specialized
+
+        # pyre-ignore[16]: `None` has no attribute `to_tensor`
+        if isinstance(args[0], ProxyValue) and args[0].is_tensor():
+            # pyre-ignore[16]: `None` has no attribute `to_tensor`
+            ndim = args[0].to_tensor().dim()
+        elif isinstance(args[0], torch.Tensor):
+            # pyre-ignore[16]: `None` has no attribute `dim`
+            ndim = args[0].dim()
+        elif isinstance(args[0], torch.fx.immutable_collections.immutable_list):
+            # pyre-ignore[6]: Incompatible parameter type
+            ndim = len(args[0])
+        else:
+            assert 0, f"Expecting a Tensor or a ProxyValue but got {type(args[0])}"
+
+        # get the "to" memory format for the EdgeOp
+        contiguous_dim_order = list(range(ndim))
+        dim_order = nkwargs.pop("dim_order", None)
+
+        # Cadence only supports contiguous memory format
+        assert (
+            dim_order is None
+            # pyre-ignore[6]: Incompatible parameter type
+            or len(dim_order) == 0
+            or dim_order == contiguous_dim_order
+        ), "Expected dim order in congituous or prevserve memory format, but got {}".format(
+            dim_order
+        )
+
+        # bring back memory format
+        # pyre-ignore[6]: Incompatible parameter type
+        nkwargs["memory_format"] = get_memory_format(dim_order)
+
+        memory_format_op = MemoryFormatOpsMap[op]
+
+        return super().call_operator(
+            memory_format_op,
+            args,
+            nkwargs,
+            meta,
+        )
+
+
 @register_cadence_pass(CadencePassAttribute(opt_level=0))
 class ReplaceFullLikeWithFullPass(ExportPass):
     """
@@ -2108,4 +2180,5 @@ class CadenceReplaceOpsInGraph:
         ReplaceSingleElementTensorArgumentsFromFullOpWithScalarPass,
         ReplaceAtenAvgPoolWithJarvisAvgPoolPass,
         ReplaceAtenLinalgVectorNormWithCadenceLinalgVectorNormPass,
+        ReplaceToDimOrderCopyWithToCopyPass,
     ]

backends/cadence/hifi/operators/op_add.cpp

@@ -16,9 +16,9 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
 
-using executorch::aten::Scalar;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
+using exec_aten::Scalar;
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::runtime::can_cast;
 using executorch::runtime::CppTypeToScalarType;
 using executorch::runtime::KernelRuntimeContext;

backends/cadence/hifi/operators/op_cat.cpp

@@ -30,7 +30,7 @@ namespace native {
 
 Tensor& cat_out(
     RuntimeContext& ctx,
-    executorch::aten::ArrayRef<Tensor> tensors,
+    exec_aten::ArrayRef<Tensor> tensors,
     int64_t dim,
     Tensor& out) {
   if (dim < 0) {

backends/cadence/hifi/operators/op_clamp.cpp

@@ -51,8 +51,8 @@ namespace native {
 Tensor& clamp_tensor_out(
     RuntimeContext& ctx,
     const Tensor& in,
-    const executorch::aten::optional<Tensor>& min_opt,
-    const executorch::aten::optional<Tensor>& max_opt,
+    const exec_aten::optional<Tensor>& min_opt,
+    const exec_aten::optional<Tensor>& max_opt,
     Tensor& out) {
   (void)ctx;
 

backends/cadence/hifi/operators/op_div.cpp

@@ -17,10 +17,10 @@
 #include <executorch/runtime/platform/assert.h>
 #include <cmath>
 
+using exec_aten::Scalar;
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::Scalar;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 using torch::executor::Error;
 
 namespace cadence {
@@ -165,7 +165,7 @@ Tensor& div_out_mode(
     RuntimeContext& ctx,
     const Tensor& a,
     const Tensor& b,
-    executorch::aten::optional<executorch::aten::string_view> mode,
+    exec_aten::optional<exec_aten::string_view> mode,
     Tensor& out) {
   ET_KERNEL_CHECK(
       ctx,

backends/cadence/hifi/operators/op_maximum.cpp

@@ -12,9 +12,9 @@
 #include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 using executorch::runtime::can_cast;
 using executorch::runtime::canCast;
 using executorch::runtime::CppTypeToScalarType;

backends/cadence/hifi/operators/op_minimum.cpp

@@ -12,9 +12,9 @@
 #include <executorch/kernels/portable/cpu/util/math_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 using executorch::runtime::can_cast;
 using executorch::runtime::canCast;
 using executorch::runtime::CppTypeToScalarType;

backends/cadence/hifi/operators/op_mul.cpp

@@ -15,10 +15,10 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
 
+using exec_aten::Scalar;
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::Scalar;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 using executorch::runtime::can_cast;
 using executorch::runtime::CppTypeToScalarType;
 using torch::executor::Error;

backends/cadence/hifi/operators/op_rsqrt.cpp

@@ -11,9 +11,9 @@
 
 #include <executorch/backends/cadence/hifi/kernels/kernels.h>
 
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 
 namespace cadence {
 namespace impl {

backends/cadence/hifi/operators/op_sigmoid.cpp

@@ -14,17 +14,17 @@
 #include <executorch/kernels/portable/cpu/util/functional_util.h>
 #include <executorch/runtime/kernel/kernel_includes.h>
 
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 using torch::executor::Error;
 
 namespace cadence {
 namespace impl {
 namespace HiFi {
 namespace native {
 
-using Tensor = executorch::aten::Tensor;
+using Tensor = exec_aten::Tensor;
 
 Tensor& sigmoid_out(RuntimeContext& ctx, const Tensor& in, Tensor& out) {
   (void)ctx;

backends/cadence/hifi/operators/op_softmax.cpp

@@ -50,7 +50,7 @@ Tensor& softmax_out(
   // Adjust for negative dim
   dim = dim < 0 ? dim + executorch::runtime::nonzero_dim(in) : dim;
 
-  const executorch::aten::optional<int64_t>& dim_t = dim;
+  const exec_aten::optional<int64_t>& dim_t = dim;
   const size_t d = ET_NORMALIZE_IX(dim_t.value(), in.dim());
   const size_t size = in.size(d);
 

backends/cadence/hifi/operators/op_sub.cpp

@@ -16,10 +16,10 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/platform/assert.h>
 
+using exec_aten::Scalar;
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::Scalar;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 using executorch::runtime::can_cast;
 using executorch::runtime::CppTypeToScalarType;
 using torch::executor::Error;

backends/cadence/hifi/operators/op_tanh.cpp

@@ -11,9 +11,9 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <cmath>
 
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
 using executorch::aten::RuntimeContext;
-using executorch::aten::ScalarType;
-using executorch::aten::Tensor;
 using torch::executor::Error;
 
 namespace cadence {

backends/cadence/hifi/operators/quantized_linear_out.cpp

@@ -219,7 +219,7 @@ void quantized_linear_out(
     int64_t out_zero_point,
     __ET_UNUSED const optional<Tensor>& offset,
     Tensor& out) {
-  if (out.scalar_type() == executorch::aten::ScalarType::Byte) {
+  if (out.scalar_type() == exec_aten::ScalarType::Byte) {
     _quantized_linear_asym8u(
         in,
         weight,
@@ -231,7 +231,7 @@ void quantized_linear_out(
         out_zero_point,
         offset,
         out);
-  } else if (out.scalar_type() == executorch::aten::ScalarType::Char) {
+  } else if (out.scalar_type() == exec_aten::ScalarType::Char) {
     _quantized_linear_asym8s(
         in,
         weight,
@@ -261,7 +261,7 @@ void quantized_linear_per_tensor_out(
     int64_t out_zero_point,
     __ET_UNUSED const optional<Tensor>& offset,
     Tensor& out) {
-  if (out.scalar_type() == executorch::aten::ScalarType::Byte) {
+  if (out.scalar_type() == exec_aten::ScalarType::Byte) {
     _quantized_linear_per_tensor_asym8u(
         in,
         weight,
@@ -273,7 +273,7 @@ void quantized_linear_per_tensor_out(
         out_zero_point,
         offset,
         out);
-  } else if (out.scalar_type() == executorch::aten::ScalarType::Char) {
+  } else if (out.scalar_type() == exec_aten::ScalarType::Char) {
     _quantized_linear_per_tensor_asym8s(
         in,
         weight,

backends/vulkan/runtime/VulkanBackend.cpp

@@ -417,10 +417,10 @@ bool maybe_update_scalar_tensor(
     executorch::aten::Tensor& scalar_tensor_src) {
   const int32_t cur_val = graph->read_symint(ref);
   int32_t scalar_tensor_val = 0;
-  executorch::aten::ScalarType dtype = scalar_tensor_src.scalar_type();
-  if (dtype == executorch::aten::ScalarType::Int) {
+  exec_aten::ScalarType dtype = scalar_tensor_src.scalar_type();
+  if (dtype == exec_aten::ScalarType::Int) {
     scalar_tensor_val = *scalar_tensor_src.const_data_ptr<int32_t>();
-  } else if (dtype == executorch::aten::ScalarType::Long) {
+  } else if (dtype == exec_aten::ScalarType::Long) {
     scalar_tensor_val = int32_t(*scalar_tensor_src.const_data_ptr<int64_t>());
   }
   bool was_updated = false;

codegen/tools/gen_selected_op_variants.py

@@ -17,7 +17,7 @@
 from torchgen.code_template import CodeTemplate
 
 
-ops_and_dtypes_template_str = """((executorch::aten::string_view(operator_name).compare("$operator_name") == 0)\n        && ($dtype_checks))"""
+ops_and_dtypes_template_str = """((exec_aten::string_view(operator_name).compare("$operator_name") == 0)\n        && ($dtype_checks))"""
 ops_and_dtypes_template = CodeTemplate(ops_and_dtypes_template_str)
 
 selected_kernel_dtypes_h_template_str = """#pragma once
@@ -27,7 +27,7 @@
 
 inline constexpr bool should_include_kernel_dtype(
   const char *operator_name,
-  executorch::aten::ScalarType scalar_type
+  exec_aten::ScalarType scalar_type
 ) {
   return $body;
 }
@@ -91,8 +91,7 @@ def write_selected_op_variants(yaml_file_path: str, output_dir: str) -> None:
                 dtype_set = set([x.split(";")[0] for x in tensor_meta])
                 dtype_list = sorted([dtype_enum_to_type[x] for x in dtype_set])
                 conditions = [
-                    "scalar_type == executorch::aten::ScalarType::" + x
-                    for x in dtype_list
+                    "scalar_type == exec_aten::ScalarType::" + x for x in dtype_list
                 ]
             body_parts.append(
                 ops_and_dtypes_template.substitute(

codegen/tools/test/test_gen_selected_op_variants.py

@@ -71,13 +71,13 @@ def test_generates_correct_header(self) -> None:
 
 inline constexpr bool should_include_kernel_dtype(
   const char *operator_name,
-  executorch::aten::ScalarType scalar_type
+  exec_aten::ScalarType scalar_type
 ) {
-  return ((executorch::aten::string_view(operator_name).compare("add.out") == 0)
-        && (scalar_type == executorch::aten::ScalarType::Float || scalar_type == executorch::aten::ScalarType::Int))
- || ((executorch::aten::string_view(operator_name).compare("mul.out") == 0)
-        && (scalar_type == executorch::aten::ScalarType::Float))
- || ((executorch::aten::string_view(operator_name).compare("sub.out") == 0)
+  return ((exec_aten::string_view(operator_name).compare("add.out") == 0)
+        && (scalar_type == exec_aten::ScalarType::Float || scalar_type == exec_aten::ScalarType::Int))
+ || ((exec_aten::string_view(operator_name).compare("mul.out") == 0)
+        && (scalar_type == exec_aten::ScalarType::Float))
+ || ((exec_aten::string_view(operator_name).compare("sub.out") == 0)
         && (true));
 }
 """,
@@ -124,7 +124,7 @@ def test_generates_correct_header(self) -> None:
 
 inline constexpr bool should_include_kernel_dtype(
   const char *operator_name,
-  executorch::aten::ScalarType scalar_type
+  exec_aten::ScalarType scalar_type
 ) {
   return true;
 }