pnnx fix onnx clip conversion, add onnx clamp test, always reserve on…

…nx split outputs, convert onnx mod (#5834)

tools/pnnx/src/pass_level2/torch_clamp.cpp

@@ -47,7 +47,7 @@ class torch_clamp_onnx : public GraphRewriterPass
         return R"PNNXIR(7767517
 3 2
 pnnx.Input              input       0 1 input
-aten::clamp             op_0        1 1 input out min=%min max=%max
+Clip                    op_0        1 1 input out min=%min max=%max
 pnnx.Output             output      1 0 out
 )PNNXIR";
     }
@@ -68,7 +68,7 @@ class torch_clamp_onnx_1 : public GraphRewriterPass
         return R"PNNXIR(7767517
 3 2
 pnnx.Input              input       0 1 input
-aten::clamp             op_0        1 1 input out max=%max
+Clip                    op_0        1 1 input out max=%max
 pnnx.Output             output      1 0 out
 )PNNXIR";
     }
@@ -80,11 +80,38 @@ pnnx.Output             output      1 0 out
 
     void write(Operator* op, const std::map<std::string, Parameter>& captured_params) const
     {
-        op->params["min"] = 0.f;
+        op->params["min"] = Parameter();
         op->params["max"] = captured_params.at("max");
     }
 };
 
 REGISTER_GLOBAL_PNNX_GRAPH_REWRITER_PASS(torch_clamp_onnx_1, 40)
 
+class torch_clamp_onnx_2 : public GraphRewriterPass
+{
+public:
+    const char* match_pattern_graph() const
+    {
+        return R"PNNXIR(7767517
+3 2
+pnnx.Input              input       0 1 input
+Clip                    op_0        1 1 input out min=%min
+pnnx.Output             output      1 0 out
+)PNNXIR";
+    }
+
+    const char* type_str() const
+    {
+        return "torch.clamp";
+    }
+
+    void write(Operator* op, const std::map<std::string, Parameter>& captured_params) const
+    {
+        op->params["max"] = Parameter();
+        op->params["min"] = captured_params.at("min");
+    }
+};
+
+REGISTER_GLOBAL_PNNX_GRAPH_REWRITER_PASS(torch_clamp_onnx_2, 40)
+
 } // namespace pnnx

tools/pnnx/src/pass_onnx.cpp

@@ -739,9 +739,9 @@ void pass_onnx(const onnx::ModelProto& model, Graph& pnnx_graph)
             if (op_type == "And") sim_op_type = "aten::__and__";
             if (op_type == "Or") sim_op_type = "aten::__or__";
             if (op_type == "Xor") sim_op_type = "aten::__xor__";
+            if (op_type == "Mod" && onnx2pnnx::OnnxNodeProxy(node).attribute("fmod").value_i() == 1) sim_op_type = "aten::fmod";
 
             // trinaryop
-            if (op_type == "Clip") sim_op_type = "aten::clamp";
             if (op_type == "Where") sim_op_type = "aten::where";
         }
         else if (string_starts_with(op_type, "aten_"))

tools/pnnx/src/pass_onnx/dead_code_elimination.cpp

@@ -44,16 +44,25 @@ static void collect_dead_nodes(const onnx::GraphProto& graph, std::vector<std::s
 
         if (is_outputs_live)
         {
-            for (int j = node.output_size() - 1; j >= 0; j--)
+            bool is_outputs_optional = true;
+
+            // some operator output be discarded even if not used
+            const std::string& op_type = node.op_type();
+            if (op_type == "Split") is_outputs_optional = false;
+
+            if (is_outputs_optional)
             {
-                if (live_inputs.find(node.output(j)) == live_inputs.end())
-                {
-                    dead_outputs.push_back(node.output(j));
-                }
-                else
+                for (int j = node.output_size() - 1; j >= 0; j--)
                 {
-                    // leading outputs cannot be optional
-                    break;
+                    if (live_inputs.find(node.output(j)) == live_inputs.end())
+                    {
+                        dead_outputs.push_back(node.output(j));
+                    }
+                    else
+                    {
+                        // leading outputs cannot be optional
+                        break;
+                    }
                 }
             }
 
@@ -65,6 +74,11 @@ static void collect_dead_nodes(const onnx::GraphProto& graph, std::vector<std::s
         else
         {
             dead_node_indexes.push_back(i);
+
+            for (int j = node.output_size() - 1; j >= 0; j--)
+            {
+                dead_outputs.push_back(node.output(j));
+            }
         }
 
         if (is_outputs_live)

tools/pnnx/src/pass_onnx/fuse_constant_as_attribute.cpp

@@ -32,6 +32,8 @@ struct constant_as_attribute
 };
 
 static constant_as_attribute caas[] = {
+    {"Clip", 1, "min"},
+    {"Clip", 2, "max"},
     {"Expand", 1, "shape"},
     {"Gather", 1, "indices"},
     {"If", 0, "cond"},

tools/pnnx/tests/onnx/CMakeLists.txt

@@ -137,6 +137,7 @@ pnnx_onnx_add_test(Tensor_view)
 pnnx_onnx_add_test(torch_cat)
 pnnx_onnx_add_test(torch_ceil)
 pnnx_onnx_add_test(torch_chunk)
+pnnx_onnx_add_test(torch_clamp)
 pnnx_onnx_add_test(torch_flatten)
 pnnx_onnx_add_test(torch_floor)
 pnnx_onnx_add_test(torch_max)

tools/pnnx/tests/onnx/test_torch_clamp.py

@@ -0,0 +1,60 @@
+# Tencent is pleased to support the open source community by making ncnn available.
+#
+# Copyright (C) 2024 THL A29 Limited, a Tencent company. All rights reserved.
+#
+# Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+# in compliance with the License. You may obtain a copy of the License at
+#
+# https://opensource.org/licenses/BSD-3-Clause
+#
+# Unless required by applicable law or agreed to in writing, software distributed
+# under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+# CONDITIONS OF ANY KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations under the License.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class Model(nn.Module):
+    def __init__(self):
+        super(Model, self).__init__()
+
+    def forward(self, x, y, z):
+        x = torch.clamp(x, max=2)
+        y = torch.clamp(y, min=0)
+        z = torch.clamp(z, min=-1, max=1)
+        return x, y, z
+
+def test():
+    net = Model()
+    net.eval()
+
+    torch.manual_seed(0)
+    x = torch.rand(1, 3, 16)
+    y = torch.rand(1, 5, 9, 11)
+    z = torch.rand(14, 8, 5, 9, 10)
+
+    a = net(x, y, z)
+
+    # export onnx
+    torch.onnx.export(net, (x, y, z), "test_torch_clamp.onnx")
+
+    # onnx to pnnx
+    import os
+    os.system("../../src/pnnx test_torch_clamp.onnx inputshape=[1,3,16],[1,5,9,11],[14,8,5,9,10]")
+
+    # pnnx inference
+    import test_torch_clamp_pnnx
+    b = test_torch_clamp_pnnx.test_inference()
+
+    for a0, b0 in zip(a, b):
+        if not torch.equal(a0, b0):
+            return False
+    return True
+
+if __name__ == "__main__":
+    if test():
+        exit(0)
+    else:
+        exit(1)