Add layout attribute support for LSTM operator (#2343)

Since ONNX opset version 14, layout attribute has been introduced to LSTM operator which allows two predefined layouts for input and output shapes. Add corresponding reference, onnx, and verify tests.
ROCm · Nov 15, 2023 · b2a40ea · b2a40ea
1 parent 45ccd75
commit b2a40ea
Show file tree

Hide file tree

Showing 18 changed files with 2,882 additions and 235 deletions.
diff --git a/src/onnx/parse_lstm.cpp b/src/onnx/parse_lstm.cpp
@@ -116,6 +116,37 @@ void lstm_actv_functions(op::rnn_direction dirct, std::vector<std::string>& actv
     }
 }
 
+void lstm_transpose_inputs(onnx_parser::node_info& info, std::vector<instruction_ref>& args)
+{
+    std::vector<int64_t> perm{1, 0, 2};
+    args[0] = info.add_instruction(make_op("transpose", {{"permutation", perm}}), args[0]);
+
+    if(args.size() >= 6 and not args[5]->is_undefined())
+    {
+        args[5] = info.add_instruction(make_op("transpose", {{"permutation", perm}}), args[5]);
+    }
+
+    if(args.size() >= 7 and not args[6]->is_undefined())
+    {
+        args[6] = info.add_instruction(make_op("transpose", {{"permutation", perm}}), args[6]);
+    }
+}
+
+void lstm_transpose_outputs(onnx_parser::node_info& info,
+                            instruction_ref& hidden_states,
+                            instruction_ref& last_output,
+                            instruction_ref& last_cell_output)
+{
+    std::vector<int64_t> perm_hs{2, 0, 1, 3};
+    hidden_states =
+        info.add_instruction(make_op("transpose", {{"permutation", perm_hs}}), hidden_states);
+    std::vector<int64_t> perm_last{1, 0, 2};
+    last_output =
+        info.add_instruction(make_op("transpose", {{"permutation", perm_last}}), last_output);
+    last_cell_output =
+        info.add_instruction(make_op("transpose", {{"permutation", perm_last}}), last_cell_output);
+}
+
 struct parse_lstm : op_parser<parse_lstm>
 {
     std::vector<op_desc> operators() const { return {{"LSTM"}}; }
@@ -202,13 +233,24 @@ struct parse_lstm : op_parser<parse_lstm>
             input_forget = parser.parse_value(info.attributes.at("input_forget")).at<int>();
         }
 
+        int layout = 0;
+        if(contains(info.attributes, "layout"))
+        {
+            layout = parser.parse_value(info.attributes.at("layout")).at<int>();
+        }
+
         // append undefined opeator to make 6 arguments
         if(args.size() < 8)
         {
             auto ins = info.add_instruction(make_op("undefined"));
             args.insert(args.end(), 8 - args.size(), ins);
         }
 
+        if(layout != 0)
+        {
+            lstm_transpose_inputs(info, args);
+        }
+
         // first output for concatenation of hidden states
         auto hidden_states = info.add_instruction(make_op("lstm",
                                                           {{"hidden_size", hidden_size},
@@ -224,6 +266,11 @@ struct parse_lstm : op_parser<parse_lstm>
         auto last_cell_output =
             info.add_instruction(make_op("rnn_last_cell_output"), hidden_states);
 
+        if(layout != 0)
+        {
+            lstm_transpose_outputs(info, hidden_states, last_output, last_cell_output);
+        }
+
         return {hidden_states, last_output, last_cell_output};
     }
 };

diff --git a/test/onnx/gen_onnx.py b/test/onnx/gen_onnx.py
@@ -4484,6 +4484,177 @@ def lrn_test():
     return ([node], [x], [y])
 
 
+@onnx_test()
+def lstm_bi_layout_cell_test():
+    seq = helper.make_tensor_value_info('seq', TensorProto.FLOAT, [3, 5, 10])
+    w = helper.make_tensor_value_info('w', TensorProto.FLOAT, [2, 80, 10])
+    r = helper.make_tensor_value_info('r', TensorProto.FLOAT, [2, 80, 20])
+    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [2, 160])
+    seq_len = helper.make_tensor_value_info('seq_len', TensorProto.INT32, [3])
+    h0 = helper.make_tensor_value_info('h0', TensorProto.FLOAT, [3, 2, 20])
+    c0 = helper.make_tensor_value_info('c0', TensorProto.FLOAT, [3, 2, 20])
+    pph = helper.make_tensor_value_info('pph', TensorProto.FLOAT, [2, 60])
+
+    cellout = helper.make_tensor_value_info('cellout', TensorProto.FLOAT,
+                                            [3, 2, 20])
+
+    node = onnx.helper.make_node(
+        'LSTM',
+        inputs=['seq', 'w', 'r', 'bias', 'seq_len', 'h0', 'c0', 'pph'],
+        outputs=['', '', 'cellout'],
+        activations=['sigmoid', 'tanh', 'tanh'],
+        clip=0,
+        direction='bidirectional',
+        hidden_size=20,
+        input_forget=1,
+        layout=1)
+
+    return ([node], [seq, w, r, bias, seq_len, h0, c0, pph], [cellout])
+
+
+@onnx_test()
+def lstm_bi_layout_last_test():
+    seq = helper.make_tensor_value_info('seq', TensorProto.FLOAT, [3, 5, 10])
+    w = helper.make_tensor_value_info('w', TensorProto.FLOAT, [2, 80, 10])
+    r = helper.make_tensor_value_info('r', TensorProto.FLOAT, [2, 80, 20])
+    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [2, 160])
+    seq_len = helper.make_tensor_value_info('seq_len', TensorProto.INT32, [3])
+    h0 = helper.make_tensor_value_info('h0', TensorProto.FLOAT, [3, 2, 20])
+    c0 = helper.make_tensor_value_info('c0', TensorProto.FLOAT, [3, 2, 20])
+    pph = helper.make_tensor_value_info('pph', TensorProto.FLOAT, [2, 60])
+
+    hs = helper.make_tensor_value_info('hs', TensorProto.FLOAT, [3, 5, 2, 20])
+    output = helper.make_tensor_value_info('output', TensorProto.FLOAT,
+                                           [3, 2, 20])
+
+    node = onnx.helper.make_node(
+        'LSTM',
+        inputs=['seq', 'w', 'r', 'bias', 'seq_len', 'h0', 'c0', 'pph'],
+        outputs=['hs', 'output'],
+        activations=['sigmoid', 'tanh', 'tanh'],
+        clip=0,
+        direction='bidirectional',
+        hidden_size=20,
+        input_forget=1,
+        layout=1)
+
+    return ([node], [seq, w, r, bias, seq_len, h0, c0, pph], [hs, output])
+
+
+@onnx_test()
+def lstm_f_layout_hs_test():
+    seq = helper.make_tensor_value_info('seq', TensorProto.FLOAT, [3, 5, 10])
+    w = helper.make_tensor_value_info('w', TensorProto.FLOAT, [1, 80, 10])
+    r = helper.make_tensor_value_info('r', TensorProto.FLOAT, [1, 80, 20])
+    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [1, 160])
+    seq_len = helper.make_tensor_value_info('seq_len', TensorProto.INT32, [3])
+    h0 = helper.make_tensor_value_info('h0', TensorProto.FLOAT, [3, 1, 20])
+    c0 = helper.make_tensor_value_info('c0', TensorProto.FLOAT, [3, 1, 20])
+    pph = helper.make_tensor_value_info('pph', TensorProto.FLOAT, [1, 60])
+
+    hs = helper.make_tensor_value_info('hs', TensorProto.FLOAT, [3, 5, 1, 20])
+    output = helper.make_tensor_value_info('output', TensorProto.FLOAT,
+                                           [3, 1, 20])
+
+    node = onnx.helper.make_node(
+        'LSTM',
+        inputs=['seq', 'w', 'r', 'bias', 'seq_len', 'h0', 'c0', 'pph'],
+        outputs=['hs', 'output'],
+        activations=['sigmoid', 'tanh', 'tanh'],
+        clip=0,
+        direction='forward',
+        hidden_size=20,
+        input_forget=1,
+        layout=1)
+
+    return ([node], [seq, w, r, bias, seq_len, h0, c0, pph], [hs, output])
+
+
+@onnx_test()
+def lstm_f_layout_cell_test():
+    seq = helper.make_tensor_value_info('seq', TensorProto.FLOAT, [3, 5, 10])
+    w = helper.make_tensor_value_info('w', TensorProto.FLOAT, [1, 80, 10])
+    r = helper.make_tensor_value_info('r', TensorProto.FLOAT, [1, 80, 20])
+    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [1, 160])
+    seq_len = helper.make_tensor_value_info('seq_len', TensorProto.INT32, [3])
+    h0 = helper.make_tensor_value_info('h0', TensorProto.FLOAT, [3, 1, 20])
+    c0 = helper.make_tensor_value_info('c0', TensorProto.FLOAT, [3, 1, 20])
+    pph = helper.make_tensor_value_info('pph', TensorProto.FLOAT, [1, 60])
+
+    cellout = helper.make_tensor_value_info('cellout', TensorProto.FLOAT,
+                                            [3, 1, 20])
+
+    node = onnx.helper.make_node(
+        'LSTM',
+        inputs=['seq', 'w', 'r', 'bias', 'seq_len', 'h0', 'c0', 'pph'],
+        outputs=['', '', 'cellout'],
+        activations=['sigmoid', 'tanh', 'tanh'],
+        clip=0,
+        direction='forward',
+        hidden_size=20,
+        input_forget=1,
+        layout=1)
+
+    return ([node], [seq, w, r, bias, seq_len, h0, c0, pph], [cellout])
+
+
+@onnx_test()
+def lstm_r_layout_test():
+    seq = helper.make_tensor_value_info('seq', TensorProto.FLOAT, [3, 5, 10])
+    w = helper.make_tensor_value_info('w', TensorProto.FLOAT, [1, 80, 10])
+    r = helper.make_tensor_value_info('r', TensorProto.FLOAT, [1, 80, 20])
+    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [1, 160])
+    seq_len = helper.make_tensor_value_info('seq_len', TensorProto.INT32, [3])
+    h0 = helper.make_tensor_value_info('h0', TensorProto.FLOAT, [3, 1, 20])
+    c0 = helper.make_tensor_value_info('c0', TensorProto.FLOAT, [3, 1, 20])
+    pph = helper.make_tensor_value_info('pph', TensorProto.FLOAT, [1, 60])
+
+    hs = helper.make_tensor_value_info('hs', TensorProto.FLOAT, [3, 5, 1, 20])
+
+    node = onnx.helper.make_node(
+        'LSTM',
+        inputs=['seq', 'w', 'r', 'bias', 'seq_len', 'h0', 'c0', 'pph'],
+        outputs=['hs'],
+        activations=['sigmoid', 'tanh', 'tanh'],
+        clip=0,
+        direction='reverse',
+        hidden_size=20,
+        input_forget=1,
+        layout=1)
+
+    return ([node], [seq, w, r, bias, seq_len, h0, c0, pph], [hs])
+
+
+@onnx_test()
+def lstm_r_layout_hs_cell_test():
+    seq = helper.make_tensor_value_info('seq', TensorProto.FLOAT, [3, 5, 10])
+    w = helper.make_tensor_value_info('w', TensorProto.FLOAT, [1, 80, 10])
+    r = helper.make_tensor_value_info('r', TensorProto.FLOAT, [1, 80, 20])
+    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [1, 160])
+    seq_len = helper.make_tensor_value_info('seq_len', TensorProto.INT32, [3])
+    h0 = helper.make_tensor_value_info('h0', TensorProto.FLOAT, [3, 1, 20])
+    c0 = helper.make_tensor_value_info('c0', TensorProto.FLOAT, [3, 1, 20])
+    pph = helper.make_tensor_value_info('pph', TensorProto.FLOAT, [1, 60])
+
+    output = helper.make_tensor_value_info('output', TensorProto.FLOAT,
+                                           [3, 1, 20])
+    cellout = helper.make_tensor_value_info('cellout', TensorProto.FLOAT,
+                                            [3, 1, 20])
+
+    node = onnx.helper.make_node(
+        'LSTM',
+        inputs=['seq', 'w', 'r', 'bias', 'seq_len', 'h0', 'c0', 'pph'],
+        outputs=['', 'output', 'cellout'],
+        activations=['sigmoid', 'tanh', 'tanh'],
+        clip=0,
+        direction='reverse',
+        hidden_size=20,
+        input_forget=1,
+        layout=1)
+
+    return ([node], [seq, w, r, bias, seq_len, h0, c0, pph], [output, cellout])
+
+
 @onnx_test()
 def matmul_bmbm_test():
     m1 = helper.make_tensor_value_info('1', TensorProto.FLOAT, [3, 6, 7])

diff --git a/test/onnx/lstm_bi_layout_cell_test.onnx b/test/onnx/lstm_bi_layout_cell_test.onnx
diff --git a/test/onnx/lstm_bi_layout_last_test.onnx b/test/onnx/lstm_bi_layout_last_test.onnx
diff --git a/test/onnx/lstm_f_layout_cell_test.onnx b/test/onnx/lstm_f_layout_cell_test.onnx
diff --git a/test/onnx/lstm_f_layout_hs_test.onnx b/test/onnx/lstm_f_layout_hs_test.onnx
diff --git a/test/onnx/lstm_r_layout_hs_cell_test.onnx b/test/onnx/lstm_r_layout_hs_cell_test.onnx
diff --git a/test/onnx/lstm_r_layout_test.onnx b/test/onnx/lstm_r_layout_test.onnx