Add eltwise types resolving. Support big int constants. (openvinotool…

…kit#15415) * Add eltwise types resolving. Support big int constants. * Update src/bindings/python/src/openvino/frontend/pytorch/decoder.py * Small fix * Fix some cases * Add tests for add in different types * Add tests for mul * Add tests for sub and div * Small fixes * Return list handling (needed for empty lists) * Add test for empty list * Update src/frontends/pytorch/src/op/mul.cpp Co-authored-by: Roman Kazantsev <[email protected]> * Use refs instead of ptrs * Apply suggestions from code review Co-authored-by: Roman Kazantsev <[email protected]> * Apply code review suggestions * Fix code style * Add more eltwise ops --------- Co-authored-by: Roman Kazantsev <[email protected]>
zhangYiIntel · Feb 2, 2023 · 9264910 · 9264910
1 parent 8051c2d
commit 9264910
Show file tree

Hide file tree

Showing 15 changed files with 1,009 additions and 132 deletions.
diff --git a/src/bindings/python/src/openvino/frontend/pytorch/decoder.py b/src/bindings/python/src/openvino/frontend/pytorch/decoder.py
@@ -16,7 +16,10 @@ def get_type_from_py_type(value):
     if isinstance(value, float):
         return OVType.f32
     if isinstance(value, int):
-        return OVType.i32
+        # Python int is 64 bit, but we will convert it to int32 except cases when it can't fit in 32 bits
+        if torch.iinfo(torch.int).min <= value <= torch.iinfo(torch.int).max:
+            return OVType.i32
+        return OVType.i64
     if isinstance(value, bool):
         return OVType.boolean
     return OVType.dynamic
@@ -27,13 +30,13 @@ def ivalue_to_constant(ivalue):
     if ov_type.is_static():
         return op.Constant(ov_type, Shape([]), [ivalue]).outputs()
 
-    if isinstance(ivalue, list):
+    if isinstance(ivalue, (list, tuple)):
         assert len(ivalue) > 0, "Can't deduce type for empty list"
         ov_type = get_type_from_py_type(ivalue[0])
         assert ov_type.is_static(), "Can't deduce type for list"
         return op.Constant(ov_type, Shape([len(ivalue)]), ivalue).outputs()
 
-    if ivalue.type() in pt_to_ov_type_map:
+    if isinstance(ivalue, torch.Tensor) and ivalue.type() in pt_to_ov_type_map:
         try:
             ovshape = PartialShape(ivalue.size())
             ovtype = pt_to_ov_type_map[ivalue.type()]
@@ -46,6 +49,7 @@ def ivalue_to_constant(ivalue):
             ovshape = PartialShape(nvalues.shape)
             ov_const = op.Constant(ovtype, ovshape.get_shape(), nvalues.flatten().tolist())
         return ov_const.outputs()
+    return None
 
 
 def get_value_from_getattr(getattr_node, self_module):
@@ -69,25 +73,22 @@ def get_value_from_getattr(getattr_node, self_module):
 pt_to_ov_type_map = {
     "float": OVType.f32,
     "int": OVType.i32,
+    "bool": OVType.boolean,
+    "torch.float16": OVType.f16,
     "torch.float32": OVType.f32,
+    "torch.float64": OVType.f64,
+    "torch.uint8": OVType.u8,
+    "torch.int8": OVType.i8,
     "torch.int32": OVType.i32,
-    "torch.bool": OVType.boolean,
     "torch.int64": OVType.i64,
+    "torch.bool": OVType.boolean,
+    "torch.DoubleTensor": OVType.f64,
     "torch.FloatTensor": OVType.f32,
     "torch.IntTensor": OVType.i32,
     "torch.LongTensor": OVType.i64,
     "torch.BoolTensor": OVType.boolean,
 }
 
-pt_to_py_type_map = {
-    "float": "float",
-    "int": "int",
-    "torch.float32": "float",
-    "torch.int32": "int",
-    "torch.int64": "int",
-    "torch.bool": "bool",
-}
-
 np_to_ov_type_map = {
     "float32": OVType.f32,
     "int32": OVType.i32,
@@ -106,7 +107,7 @@ def __init__(self, pt_module, graph_element=None):
             self.graph_element = graph_element
         self.pt_module = pt_module
 
-    def inputs(self):
+    def inputs(self) -> list:
         return [x.unique() for x in self.graph_element.inputs()]
 
     def get_input(self, index: int):
@@ -150,7 +151,7 @@ def _get_known_type_for_value(self, pt_type):
             # Not yet recognized
             return OVAny(OVType.dynamic)
 
-    def get_shape_for_value(self, value):
+    def get_shape_for_value(self, value: torch.Value):
         if value.isCompleteTensor():
             ps = PartialShape(value.type().sizes())
             return ps
@@ -161,7 +162,7 @@ def get_shape_for_value(self, value):
             pass
         return PartialShape.dynamic()
 
-    def get_type_for_value(self, value):
+    def get_type_for_value(self, value: torch.Value):
         full_type = self._get_known_type_for_value(value.type())
         return full_type
 
@@ -184,46 +185,46 @@ def get_output_transpose_order(self, index: int) -> list:
     def get_subgraph_size(self) -> int:
         return len(self.get_subgraphs()) if hasattr(self.graph_element, "blocks") else 1
 
-    def visit_subgraph(self, node_visitor):
+    def visit_subgraph(self, node_visitor) -> None:
         # make sure topological order is satisfied
         for node in self.graph_element.nodes():
             decoder = TorchScriptPythonDecoder(self.pt_module, node)
             self.m_decoders.append(decoder)
             node_visitor(decoder)
 
-    def get_subgraphs(self):
+    def get_subgraphs(self) -> list:
         return list(self.graph_element.blocks())
 
-    def get_subgraph_decoder(self, index):
+    def get_subgraph_decoder(self, index: int):
         decoder = TorchScriptPythonDecoder(self.pt_module, self.get_subgraphs()[index])
         self.m_decoders.append(decoder)
         return decoder
 
-    def get_op_type(self):
+    def get_op_type(self) -> str:
         return self.graph_element.kind()
 
-    def get_schema(self):
+    def get_schema(self) -> str:
         return self.graph_element.schema()
 
-    def outputs(self):
+    def outputs(self) -> list:
         return [x.unique() for x in self.graph_element.outputs()]
 
-    def _raw_outputs(self):
+    def _raw_outputs(self) -> list:
         return list(self.graph_element.outputs())
 
-    def _raw_output(self, index):
+    def _raw_output(self, index: int):
         return self._raw_outputs()[index]
 
-    def _raw_inputs(self):
+    def _raw_inputs(self) -> list:
         return list(self.graph_element.inputs())
 
-    def _raw_input(self, index):
+    def _raw_input(self, index: int):
         return self._raw_inputs()[index]
 
     def num_of_outputs(self):
         return len(self.outputs())
 
-    def output(self, index):
+    def output(self, index: int):
         return self.outputs()[index]
 
     def mark_node(self, node):
@@ -232,7 +233,7 @@ def mark_node(self, node):
     def try_decode_get_attr(self):
         pt_value = get_value_from_getattr(self.graph_element, self.pt_module)
         assert pt_value is not None, "Couldn't retrieve value from prim::GetAttr"
-        if not isinstance(pt_value, torch.jit.ScriptModule) or isinstance(pt_value, torch.jit.TracedModule):
+        if not isinstance(pt_value, (torch.jit.ScriptModule, torch.jit.TracedModule)):
             return ivalue_to_constant(pt_value)
         else:
             return []
@@ -244,17 +245,10 @@ def as_constant(self):
 
         pt_type = pt_value.type()
         if isinstance(pt_type, torch.TensorType):
-            return self.as_constant_tensor(pt_value)
+            return self._as_constant_tensor(pt_value)
         if isinstance(pt_type, torch.ListType):
-            return self.as_constant_list(pt_value)
-        if str(pt_type) in ["torch.int32", "int"]:
-            return op.Constant(OVType.i32, Shape([]), [pt_value.toIValue()]).outputs()
-        if str(pt_type) in ["torch.float", "torch.FloatType", "float"]:
-            return op.Constant(OVType.f32, Shape([]), [pt_value.toIValue()]).outputs()
-        if str(pt_type) in ["torch.bool", "bool"]:
-            return op.Constant(OVType.boolean, Shape([]), [pt_value.toIValue()]).outputs()
-
-        return None
+            return self._as_constant_list(pt_value)
+        return ivalue_to_constant(pt_value.toIValue())
 
     def as_string(self):
         if not self.get_op_type() == "prim::Constant":
@@ -265,7 +259,8 @@ def as_string(self):
             return pt_value.toIValue()
         return None
 
-    def as_constant_tensor(self, pt_value):
+    @staticmethod
+    def _as_constant_tensor(pt_value: torch.Value):
         ivalue = pt_value.toIValue()
         if pt_value.isCompleteTensor():
             try:
@@ -295,7 +290,8 @@ def as_constant_tensor(self, pt_value):
             return ivalue_to_constant(ivalue)
         return None
 
-    def as_constant_list(self, pt_value):
+    @staticmethod
+    def _as_constant_list(pt_value: torch.Value):
         # For now it is treat a list as a 1D tensor; it is required by converters to avoid need to massively
         # rewrite them in that part where constant attributes are queried
         pt_element_type = str(pt_value.type().getElementType())
@@ -308,7 +304,7 @@ def as_constant_list(self, pt_value):
             ov_const = op.Constant(ovtype, ovshape.get_shape(), ivalue)
             return ov_const.outputs()
 
-    def input_is_none(self, index):
+    def input_is_none(self, index: int) -> bool:
         if index >= len(self.inputs()) or self._raw_input(index) is None:
             return True
         else:

diff --git a/src/frontends/pytorch/src/op/add.cpp b/src/frontends/pytorch/src/op/add.cpp
@@ -2,8 +2,11 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 
+#include "openvino/op/add.hpp"
+
 #include "openvino/frontend/pytorch/node_context.hpp"
-#include "openvino/opsets/opset10.hpp"
+#include "openvino/op/convert_like.hpp"
+#include "openvino/op/multiply.hpp"
 #include "utils.hpp"
 
 namespace ov {
@@ -12,12 +15,14 @@ namespace pytorch {
 namespace op {
 
 OutputVector translate_add(NodeContext& context) {
+    auto lhs = context.get_input(0);
     auto rhs = context.get_input(1);
+    align_eltwise_input_types(context, lhs, rhs);
     if (!context.input_is_none(2)) {
-        auto converted_alpha = std::make_shared<opset10::ConvertLike>(context.get_input(2), rhs);
-        rhs = std::make_shared<opset10::Multiply>(converted_alpha, rhs);
+        auto converted_alpha = context.mark_node(std::make_shared<ov::op::v1::ConvertLike>(context.get_input(2), rhs));
+        rhs = context.mark_node(std::make_shared<ov::op::v1::Multiply>(converted_alpha, rhs));
     }
-    return {context.mark_node(std::make_shared<opset10::Add>(context.get_input(0), rhs))};
+    return {context.mark_node(std::make_shared<ov::op::v1::Add>(lhs, rhs))};
 };
 
 }  // namespace op

diff --git a/src/frontends/pytorch/src/op/div.cpp b/src/frontends/pytorch/src/op/div.cpp
@@ -3,9 +3,14 @@
 //
 
 #include "openvino/frontend/pytorch/node_context.hpp"
-#include "openvino/opsets/opset10.hpp"
+#include "openvino/op/convert.hpp"
+#include "openvino/op/convert_like.hpp"
+#include "openvino/op/divide.hpp"
+#include "openvino/op/floor.hpp"
 #include "utils.hpp"
 
+using namespace ov::op;
+
 namespace ov {
 namespace frontend {
 namespace pytorch {
@@ -14,21 +19,27 @@ namespace op {
 OutputVector translate_div(NodeContext& context) {
     auto x = context.get_input(0);
     auto y = context.get_input(1);
-    auto res = context.mark_node(std::make_shared<opset10::Divide>(x, y, true));
+    std::string rounding_mode = "";
     if (!context.input_is_none(2)) {
-        auto rounding_mode = context.const_input<std::string>(2);
-        if (rounding_mode == "floor") {
-            res = context.mark_node(std::make_shared<opset10::Floor>(res));
-        } else if (rounding_mode == "trunc") {
-            const auto convert = context.mark_node(std::make_shared<opset10::Convert>(res, element::i64));
-            res = context.mark_node(std::make_shared<opset10::ConvertLike>(convert, x));
-        } else {
-            FRONT_END_OP_CONVERSION_CHECK(false,
-                                          "Openvino Pytorch Frontend doesn't support rounding mode ",
-                                          rounding_mode,
-                                          " for aten::div");
+        rounding_mode = context.const_input<std::string>(2);
+    }
+    if (rounding_mode.empty()) {
+        // if no rounding mode and both inputs are ints cast BOTH to fp32
+        const auto x_dtype = x.get_element_type();
+        const auto y_dtype = y.get_element_type();
+        if (x_dtype.is_static() && x_dtype.is_integral() && y_dtype.is_static() && y_dtype.is_integral()) {
+            x = context.mark_node(std::make_shared<v0::Convert>(x, element::f32));
+            y = context.mark_node(std::make_shared<v0::Convert>(y, element::f32));
         }
     }
+    align_eltwise_input_types(context, x, y, true);
+    auto res = context.mark_node(std::make_shared<v1::Divide>(x, y, true));
+    if (rounding_mode == "floor") {
+        res = context.mark_node(std::make_shared<v0::Floor>(res));
+    } else if (rounding_mode == "trunc") {
+        const auto convert = context.mark_node(std::make_shared<v0::Convert>(res, element::i64));
+        res = context.mark_node(std::make_shared<v1::ConvertLike>(convert, x));
+    }
     return {res};
 };
 

diff --git a/src/frontends/pytorch/src/op/pow.cpp b/src/frontends/pytorch/src/op/pow.cpp
@@ -0,0 +1,25 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "openvino/frontend/pytorch/node_context.hpp"
+#include "openvino/op/power.hpp"
+#include "utils.hpp"
+
+namespace ov {
+namespace frontend {
+namespace pytorch {
+namespace op {
+
+OutputVector translate_pow(NodeContext& context) {
+    num_inputs_check(context, 1, 2);
+    auto lhs = context.get_input(0);
+    auto rhs = context.get_input(1);
+    align_eltwise_input_types(context, lhs, rhs, true);
+    return {context.mark_node(std::make_shared<ov::op::v1::Power>(lhs, rhs))};
+}
+
+}  // namespace op
+}  // namespace pytorch
+}  // namespace frontend
+}  // namespace ov
diff --git a/src/frontends/pytorch/src/op/sub.cpp b/src/frontends/pytorch/src/op/sub.cpp
@@ -3,9 +3,13 @@
 //
 
 #include "openvino/frontend/pytorch/node_context.hpp"
-#include "openvino/opsets/opset10.hpp"
+#include "openvino/op/convert_like.hpp"
+#include "openvino/op/multiply.hpp"
+#include "openvino/op/subtract.hpp"
 #include "utils.hpp"
 
+using namespace ov::op;
+
 namespace ov {
 namespace frontend {
 namespace pytorch {
@@ -14,13 +18,14 @@ namespace op {
 OutputVector translate_sub(NodeContext& context) {
     auto x = context.get_input(0);
     auto y = context.get_input(1);
+    align_eltwise_input_types(context, x, y);
     // default alpha is 1 so no need to multiply if alpha is not provided
     if (!context.input_is_none(2)) {
         auto alpha = context.get_input(2);
-        auto casted_alpha = context.mark_node(std::make_shared<opset10::ConvertLike>(alpha, y));
-        y = context.mark_node(std::make_shared<opset10::Multiply>(casted_alpha, y));
+        auto casted_alpha = context.mark_node(std::make_shared<v1::ConvertLike>(alpha, y));
+        y = context.mark_node(std::make_shared<v1::Multiply>(casted_alpha, y));
     }
-    return {context.mark_node(std::make_shared<opset10::Subtract>(x, y))};
+    return {context.mark_node(std::make_shared<v1::Subtract>(x, y))};
 };
 
 }  // namespace op