etdump schema and etrecord changes for intermediate tensor logging (#1255)

Summary:
Pull Request resolved: #1255

All the schema changes needed for intermediate tensor logging in etdump.

Reviewed By: Olivia-liu

Differential Revision: D50900428

fbshipit-source-id: aaca12ef205f6123f32c21b7905fd19422cb1048

Author: tarun292

sdk/etdump/etdump_flatcc.cpp

@@ -131,7 +131,7 @@ void ETDumpGen::end_profiling_delegate(
   etdump_ProfileEvent_start(&builder);
   etdump_ProfileEvent_start_time_add(&builder, event_tracer_entry.start_time);
   etdump_ProfileEvent_end_time_add(&builder, end_time);
-  etdump_ProfileEvent_chain_id_add(&builder, chain_id_);
+  etdump_ProfileEvent_chain_index_add(&builder, chain_id_);
   etdump_ProfileEvent_instruction_id_add(&builder, debug_handle_);
   // Delegate debug identifier can either be of a string type or an integer
   // type. If it's a string type then it's a value of type
@@ -172,7 +172,7 @@ void ETDumpGen::log_profiling_delegate(
   etdump_ProfileEvent_start(&builder);
   etdump_ProfileEvent_start_time_add(&builder, start_time);
   etdump_ProfileEvent_end_time_add(&builder, end_time);
-  etdump_ProfileEvent_chain_id_add(&builder, chain_id_);
+  etdump_ProfileEvent_chain_index_add(&builder, chain_id_);
   etdump_ProfileEvent_instruction_id_add(&builder, debug_handle_);
   if (string_id == -1) {
     etdump_ProfileEvent_delegate_debug_id_int_add(
@@ -200,7 +200,7 @@ void ETDumpGen::end_profiling(EventTracerEntry prof_entry) {
   etdump_ProfileEvent_start(&builder);
   etdump_ProfileEvent_start_time_add(&builder, prof_entry.start_time);
   etdump_ProfileEvent_end_time_add(&builder, end_time);
-  etdump_ProfileEvent_chain_id_add(&builder, prof_entry.chain_id);
+  etdump_ProfileEvent_chain_index_add(&builder, prof_entry.chain_id);
   etdump_ProfileEvent_instruction_id_add(&builder, prof_entry.debug_handle);
   if (prof_entry.event_id != -1) {
     etdump_ProfileEvent_name_add(&builder, prof_entry.event_id);

sdk/etdump/etdump_schema_flatcc.fbs

@@ -13,7 +13,7 @@ table Tensor {
   scalar_type:executorch_flatbuffer.ScalarType;
   sizes:[int];
   strides:[int];
-  data: [ubyte];
+  offset:long;
 }
 
 table Int {
@@ -24,6 +24,10 @@ table Bool {
   bool_val:bool;
 }
 
+table Float {
+  float_val:float;
+}
+
 table Double {
   double_val:double;
 }
@@ -32,32 +36,35 @@ table String {
   string_val:string;
 }
 
-enum ValueType : byte { Null, Int, Bool, Double, Tensor, String,}
+enum ValueType : byte { Null, Int, Bool, Float, Double, Tensor, String,}
 
 // This table is only necessary because flatbuffer can't support a list of
 // union elements directly, they need to be wrapped in a table.
-struct Value {
+table Value {
     val:ValueType;
-    offset:ulong;
+    tensor:Tensor;
+    int_value:Int;
+    float_value:Float;
+    double_value:Double;
+    bool_value:Bool;
+    output:Bool;
 }
 
 // This table contains all the details that we store to debug an op executed in the
 // runtime. Each entry here reprsents an op executed in the runtime during inference.
 table DebugEvent {
   // The chain to which this instruction belongs to. For now it will always be 0,
   // as chains are not used, but left in place in case they are used in the future.
-  chain_idx:ulong;
+  chain_index:ulong;
 
-  // If this event corresponds to an operator execution then this is the debug
-  // handle that was generated by the compiler that will help us map back this
-  // operator to the source code.
-  debug_handle:ulong;
+  // Runtime instruction id to which this event corresponds to.
+  instruction_id:int = -1;
 
   // List of debug entries corresponding to this debug event. These could be
   // scalars, tensors etc.
   // The order of these entries is expected to be the same as the parameters
   // that are passed into a operator.
-  debug_entries:[Value];
+  debug_entry:Value;
 }
 
 // All the details pertaining to an allocation done in the runtime. The main
@@ -82,7 +89,7 @@ table ProfileEvent {
 
   // The chain to which this instruction belongs to. For now it will always be 0,
   // as chains are not used, but left in place in case they are used in the future.
-  chain_id:int;
+  chain_index:int;
 
   // Runtime instruction id to which this event corresponds to.
   instruction_id:int = -1;
@@ -135,6 +142,11 @@ table Allocator {
 table RunData {
   name: string;
 
+  // If bundled input was used to run this model on device then this
+  // entry will contain the index of the bundled input that was used
+  // to run this specific block.
+  bundled_input_index:int = -1;
+
   // List of allocators on which profiling was enabled in the runtime.
   allocators:[Allocator];
 

sdk/etdump/schema_flatcc.py

@@ -23,7 +23,7 @@ class Tensor:
     scalar_type: ScalarType
     sizes: List[int]
     strides: List[int]
-    data: bytes
+    offset: Optional[int]
 
 
 @dataclass
@@ -46,6 +46,11 @@ class Double:
     double_val: float
 
 
+@dataclass
+class Float:
+    float_val: float
+
+
 @dataclass
 class String:
     string_val: str
@@ -62,6 +67,7 @@ class ValueType(Enum):
     NULL = "Null"
     INT = "Int"
     BOOL = "Bool"
+    FLOAT = "Float"
     DOUBLE = "Double"
     TENSOR = "Tensor"
     STRING = "String"
@@ -70,14 +76,19 @@ class ValueType(Enum):
 @dataclass
 class Value:
     val: str  # Member of ValueType
-    offset: int
+    tensor: Optional[Tensor]
+    int_value: Optional[Int]
+    float_value: Optional[Float]
+    double_value: Optional[Double]
+    bool_value: Optional[Bool]
+    output: Optional[Bool]
 
 
 @dataclass
 class DebugEvent:
-    chain_idx: int
-    debug_handle: int
-    debug_entries: List[Value]
+    chain_index: int
+    instruction_id: int
+    debug_entry: Value
 
 
 # Note the differing value style is a result of ETDump string
@@ -91,7 +102,7 @@ class PROFILE_EVENT_ENUM(Enum):
 @dataclass
 class ProfileEvent:
     name: Optional[str]
-    chain_id: int
+    chain_index: int
     instruction_id: int
     delegate_debug_id_int: Optional[int]
     delegate_debug_id_str: Optional[str]
@@ -122,6 +133,7 @@ class Event:
 @dataclass
 class RunData:
     name: str
+    bundled_input_index: Optional[int]
     allocators: Optional[List[Allocator]]
     events: Optional[List[Event]]
 

sdk/etdump/tests/etdump_test.cpp

@@ -251,7 +251,7 @@ TEST_F(ProfilerETDumpTest, VerifyData) {
           etdump_ProfileEvent_name(single_prof_event),
           strlen(etdump_ProfileEvent_name(single_prof_event))),
       "test_event");
-  EXPECT_EQ(etdump_ProfileEvent_chain_id(single_prof_event), 0);
+  EXPECT_EQ(etdump_ProfileEvent_chain_index(single_prof_event), 0);
 
   flatbuffers_string_t allocator_name =
       etdump_Allocator_name(etdump_Allocator_vec_at(allocator_vec, 0));

sdk/etdump/tests/serialize_test.py

@@ -36,6 +36,7 @@ def get_sample_etdump_flatcc() -> flatcc.ETDumpFlatCC:
         run_data=[
             flatcc.RunData(
                 name="test_block",
+                bundled_input_index=-1,
                 allocators=[
                     flatcc.Allocator(
                         name="test_allocator",
@@ -45,7 +46,7 @@ def get_sample_etdump_flatcc() -> flatcc.ETDumpFlatCC:
                     flatcc.Event(
                         profile_event=flatcc.ProfileEvent(
                             name="test_profile_event",
-                            chain_id=1,
+                            chain_index=1,
                             instruction_id=1,
                             delegate_debug_id_str="",
                             delegate_debug_id_int=-1,
@@ -59,7 +60,7 @@ def get_sample_etdump_flatcc() -> flatcc.ETDumpFlatCC:
                     flatcc.Event(
                         profile_event=flatcc.ProfileEvent(
                             name="test_profile_event_delegated",
-                            chain_id=1,
+                            chain_index=1,
                             instruction_id=1,
                             delegate_debug_id_str="",
                             delegate_debug_id_int=13,
@@ -82,14 +83,22 @@ def get_sample_etdump_flatcc() -> flatcc.ETDumpFlatCC:
                         profile_event=None,
                         allocation_event=None,
                         debug_event=flatcc.DebugEvent(
-                            chain_idx=1,
-                            debug_handle=0,
-                            debug_entries=[
-                                flatcc.Value(
-                                    val=flatcc.ValueType.TENSOR.value,
+                            chain_index=1,
+                            instruction_id=0,
+                            debug_entry=flatcc.Value(
+                                val=flatcc.ValueType.TENSOR.value,
+                                tensor=flatcc.Tensor(
+                                    scalar_type=flatcc.ScalarType.INT,
+                                    sizes=[1],
+                                    strides=[1],
                                     offset=12345,
-                                )
-                            ],
+                                ),
+                                int_value=flatcc.Int(1),
+                                float_value=flatcc.Float(1.0),
+                                double_value=flatcc.Double(1.0),
+                                bool_value=flatcc.Bool(False),
+                                output=flatcc.Bool(True),
+                            ),
                         ),
                     ),
                 ],

sdk/inspector/tests/event_blocks_test.py

@@ -63,6 +63,7 @@ def _get_sample_etdump_flatcc() -> flatcc.ETDumpFlatCC:
         )
         run_data_1 = flatcc.RunData(
             name="run_data_1",
+            bundled_input_index=-1,
             allocators=[],
             events=[
                 flatcc.Event(
@@ -77,6 +78,7 @@ def _get_sample_etdump_flatcc() -> flatcc.ETDumpFlatCC:
         )
         run_data_2 = flatcc.RunData(
             name="run_data_2",
+            bundled_input_index=-1,
             allocators=[],
             events=[
                 flatcc.Event(
@@ -95,6 +97,7 @@ def _get_sample_etdump_flatcc() -> flatcc.ETDumpFlatCC:
         )
         run_data_3 = flatcc.RunData(
             name="run_data_3",
+            bundled_input_index=-1,
             allocators=[],
             events=[
                 flatcc.Event(