fix issues with compiling lstm ops in fp16 mode (#1450)

Currently, quantizing a program with rnn layers to fp16 results in segmentation faults due to a "convert" operation being applied to an "undefined" instruction.

The following changes are implemented to fix this issue:

Added is_undefined method to the instruction class that returns true if all inputs to the instruction are from an undefined op.
Updated rewrite_rnn pass to use the new is_undefined method rather than checking ins->name()
Updated the dead_code_elimination pass to also use this new method rather than only checking the instruction name

src/dead_code_elimination.cpp

@@ -51,8 +51,8 @@ void dead_code_elimination::apply(module& m) const
         // Skip instruction with empty shape as output unless its [dynamic, builtin, undefined,
         // identity, allocate]
         if((not i->get_shape().dynamic() and i->get_shape().elements() == 0) and
-           i->name().front() != '@' and
-           not contains({"undefined", "identity", "allocate"}, i->name()))
+           not(i->name().front() == '@') and not contains({"identity", "allocate"}, i->name()) and
+           not i->is_undefined())
             continue;
         assert(std::distance(m.begin(), i) <= std::distance(m.begin(), last));
         std::unordered_set<instruction_ref> visited;

src/include/migraphx/instruction.hpp

@@ -121,6 +121,8 @@ struct instruction
 
     bool can_eval() const;
 
+    bool is_undefined() const;
+
     argument eval(bool check_eval = true) const;
 
     void finalize(context& ctx);

src/instruction.cpp

@@ -302,6 +302,24 @@ void instruction::replace_mod_argument(module_ref old, module_ref new_mod)
     std::replace(module_args.begin(), module_args.end(), old, new_mod);
 }
 
+bool instruction::is_undefined() const
+{
+    if(op.name() == "undefined")
+    {
+        return true;
+    }
+    else if(this->inputs().empty())
+    {
+        return false;
+    }
+    else
+    {
+        return std::all_of(this->inputs().begin(), this->inputs().end(), [](auto arg) {
+            return arg->is_undefined();
+        });
+    }
+}
+
 bool instruction::can_eval() const
 {
     if(op.name() == "@literal")

src/rewrite_rnn.cpp

@@ -92,7 +92,7 @@ void rewrite_rnn::apply_vanilla_rnn(module& m, instruction_ref ins) const
 
     // process sequence length
     instruction_ref seq_lens = m.end();
-    if((args.size() >= 5) && args[4]->name() != "undefined")
+    if((args.size() >= 5) and not args[4]->is_undefined())
     {
         seq_lens = args[4];
     }
@@ -117,7 +117,7 @@ void rewrite_rnn::apply_vanilla_rnn(module& m, instruction_ref ins) const
         // process bias
         instruction_ref bias_forward = m.end();
         instruction_ref bias_reverse = m.end();
-        if(args.size() >= 4 && args[3]->name() != "undefined")
+        if(args.size() >= 4 and not args[3]->is_undefined())
         {
             bias_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[3]);
@@ -129,7 +129,7 @@ void rewrite_rnn::apply_vanilla_rnn(module& m, instruction_ref ins) const
         // or the 5th one (if the sequence len argument is ignored)
         instruction_ref ih_forward{};
         instruction_ref ih_reverse{};
-        if(args.size() == 6 && args[5]->name() != "undefined")
+        if(args.size() == 6 and not args[5]->is_undefined())
         {
             ih_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[5]);
@@ -195,14 +195,14 @@ void rewrite_rnn::apply_vanilla_rnn(module& m, instruction_ref ins) const
 
         // process bias and initial hidden state
         instruction_ref bias = m.end();
-        if(args.size() >= 4 && args[3]->name() != "undefined")
+        if(args.size() >= 4 and not args[3]->is_undefined())
         {
             bias = args[3];
         }
 
         // process intial hidden state
         instruction_ref ih;
-        if(args.size() == 6 && args[5]->name() != "undefined")
+        if(args.size() == 6 and not args[5]->is_undefined())
         {
             ih = args[5];
         }
@@ -398,7 +398,7 @@ void rewrite_rnn::apply_gru(module& m, instruction_ref ins) const
 
     // process sequence length
     instruction_ref seq_lens = m.end();
-    if((args.size() >= 5) && args[4]->name() != "undefined")
+    if((args.size() >= 5) and not args[4]->is_undefined())
     {
         seq_lens = args[4];
     }
@@ -423,7 +423,7 @@ void rewrite_rnn::apply_gru(module& m, instruction_ref ins) const
         // bias
         instruction_ref bias_forward = m.end();
         instruction_ref bias_reverse = m.end();
-        if(args.size() >= 4 && args[3]->name() != "undefined")
+        if(args.size() >= 4 and not args[3]->is_undefined())
         {
             bias_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[3]);
@@ -434,7 +434,7 @@ void rewrite_rnn::apply_gru(module& m, instruction_ref ins) const
         // intial hidden state
         instruction_ref ih_forward{};
         instruction_ref ih_reverse{};
-        if(args.size() == 6 && args[5]->name() != "undefined")
+        if(args.size() == 6 and not args[5]->is_undefined())
         {
             ih_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[5]);
@@ -501,14 +501,14 @@ void rewrite_rnn::apply_gru(module& m, instruction_ref ins) const
 
         // bias
         instruction_ref bias = m.end();
-        if(args.size() >= 4 && args[3]->name() != "undefined")
+        if(args.size() >= 4 and not args[3]->is_undefined())
         {
             bias = args[3];
         }
 
         // intial hidden state
         instruction_ref ih{};
-        if(args.size() == 6 && args[5]->name() != "undefined")
+        if(args.size() == 6 and not args[5]->is_undefined())
         {
             ih = args[5];
         }
@@ -784,7 +784,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
 
     // process sequence length
     instruction_ref seq_lens = m.end();
-    if((args.size() >= 5) && args[4]->name() != "undefined")
+    if((args.size() >= 5) and not args[4]->is_undefined())
     {
         seq_lens = args[4];
     }
@@ -813,7 +813,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
         // process bias
         instruction_ref bias_forward = m.end();
         instruction_ref bias_reverse = m.end();
-        if(args.size() >= 4 && args[3]->name() != "undefined")
+        if(args.size() >= 4 and not args[3]->is_undefined())
         {
             bias_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[3]);
@@ -824,7 +824,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
         // process intial hidden state, it is the 6th argument
         instruction_ref ih_forward{};
         instruction_ref ih_reverse{};
-        if(args.size() >= 6 && args[5]->name() != "undefined")
+        if(args.size() >= 6 and not args[5]->is_undefined())
         {
             ih_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[5]);
@@ -840,7 +840,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
         // process initial cell value
         instruction_ref ic_forward{};
         instruction_ref ic_reverse{};
-        if(args.size() >= 7 && args[6]->name() != "undefined")
+        if(args.size() >= 7 and not args[6]->is_undefined())
         {
             ic_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[6]);
@@ -856,7 +856,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
         // process weight of the peephole
         instruction_ref pph_forward = m.end();
         instruction_ref pph_reverse = m.end();
-        if(args.size() == 8 && args[7]->name() != "undefined")
+        if(args.size() == 8 and not args[7]->is_undefined())
         {
             pph_forward = m.insert_instruction(
                 ins, make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {1}}}), args[7]);
@@ -940,14 +940,14 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
 
         // bias
         instruction_ref bias = m.end();
-        if(args.size() >= 4 && args[3]->name() != "undefined")
+        if(args.size() >= 4 and not args[3]->is_undefined())
         {
             bias = args[3];
         }
 
         // initial hidden state
         instruction_ref ih{};
-        if(args.size() >= 6 && args[5]->name() != "undefined")
+        if(args.size() >= 6 and not args[5]->is_undefined())
         {
             ih = args[5];
         }
@@ -958,7 +958,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
 
         // initial cell value
         instruction_ref ic{};
-        if(args.size() >= 7 && args[6]->name() != "undefined")
+        if(args.size() >= 7 and not args[6]->is_undefined())
         {
             ic = args[6];
         }
@@ -969,7 +969,7 @@ void rewrite_rnn::apply_lstm(module& m, instruction_ref ins) const
 
         // process weight of the peephole
         instruction_ref pph = m.end();
-        if(args.size() == 8 && args[7]->name() != "undefined")
+        if(args.size() == 8 and not args[7]->is_undefined())
         {
             pph = args[7];
         }

test/instruction.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include <migraphx/instruction.hpp>
+#include <migraphx/program.hpp>
+#include <migraphx/make_op.hpp>
+#include "test.hpp"
+
+TEST_CASE(check_undefined)
+{
+    migraphx::module m;
+    auto und = m.add_instruction(migraphx::make_op("undefined"));
+    auto cov = m.add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), und);
+    auto abs = m.add_instruction(migraphx::make_op("abs"), cov);
+
+    migraphx::shape xs{migraphx::shape::float_type, {2, 3}};
+    std::vector<float> datax = {1, 2, 3, 4, 5, 6};
+
+    auto lit = m.add_literal(migraphx::literal(xs, datax));
+    auto mul = m.add_instruction(migraphx::make_op("mul"), lit, lit);
+
+    EXPECT(und->is_undefined());
+    EXPECT(cov->is_undefined());
+    EXPECT(abs->is_undefined());
+    EXPECT(not lit->is_undefined());
+    EXPECT(not mul->is_undefined());
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/ref_rnn_ops_test.cpp

@@ -31,6 +31,7 @@
 #include <migraphx/onnx.hpp>
 #include <migraphx/make_op.hpp>
 
+#include <migraphx/quantization.hpp>
 #include <migraphx/serialize.hpp>
 
 #include "test.hpp"
@@ -932,6 +933,90 @@ TEST_CASE(rnn_bidirectional)
     }
 }
 
+TEST_CASE(rnn_fp16)
+{
+    std::size_t batch_size  = 2;
+    std::size_t seq_len     = 2;
+    std::size_t hidden_size = 4;
+    std::size_t input_size  = 3;
+    std::size_t num_dirct   = 1;
+    std::vector<float> w_data{0.4691,
+                              0.3185,
+                              -0.2227,
+                              0.4423,
+                              -0.0609,
+                              -0.2803,
+                              0.1744,
+                              0.3146,
+                              0.4049,
+                              -0.3973,
+                              -0.0890,
+                              -0.1636};
+
+    std::vector<float> r_data{-0.0456,
+                              0.1061,
+                              0.1574,
+                              -0.4928,
+                              -0.4300,
+                              -0.1909,
+                              -0.0225,
+                              -0.2668,
+                              0.1840,
+                              -0.4453,
+                              -0.4896,
+                              0.1302,
+                              -0.0929,
+                              0.3545,
+                              -0.4981,
+                              0.0616};
+
+    std::vector<float> bias_data{
+        -0.4938, 0.4355, -0.3186, 0.2094, 0.1037, -0.1071, 0.4504, -0.3990};
+    std::vector<float> ih_data(num_dirct * batch_size * hidden_size, 0);
+    std::vector<float> input(seq_len * batch_size * input_size, 0);
+    input[0] = input[1] = 1.0;
+    migraphx::shape in_shape{migraphx::shape::float_type, {seq_len, batch_size, input_size}};
+    migraphx::shape ih_shape{migraphx::shape::float_type, {num_dirct, batch_size, hidden_size}};
+    migraphx::shape w_shape{migraphx::shape::float_type, {num_dirct, hidden_size, input_size}};
+    migraphx::shape r_shape{migraphx::shape::float_type, {num_dirct, hidden_size, hidden_size}};
+    migraphx::shape b_shape{migraphx::shape::float_type, {num_dirct, 2 * hidden_size}};
+    float clip = 0.0f;
+
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto seq = mm->add_literal(migraphx::literal{in_shape, input});
+    auto w   = mm->add_literal(migraphx::literal{w_shape, w_data});
+    auto r   = mm->add_literal(migraphx::literal{r_shape, r_data});
+
+    auto seq_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), seq);
+    auto w_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), w);
+    auto r_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), r);
+
+    auto out_hs = mm->add_instruction(
+        migraphx::make_op("rnn",
+                          {{"hidden_size", hidden_size},
+                           {"actv_func", {}},
+                           {"direction", migraphx::to_value(migraphx::op::rnn_direction::forward)},
+                           {"clip", clip}}),
+        seq_half,
+        w_half,
+        r_half);
+    mm->add_instruction(migraphx::make_op("rnn_last_hs_output"), out_hs);
+
+    p.compile(migraphx::ref::target{});
+
+    auto last_output = p.eval({}).back();
+    std::vector<float> last_output_data;
+    last_output.visit([&](auto out) { last_output_data.assign(out.begin(), out.end()); });
+
+    std::vector<float> last_output_data_gold{
+        0.2935145, -0.23719997, -0.31123261, -0.18357255, 0., 0., 0., 0.};
+    EXPECT(migraphx::verify_range(last_output_data, last_output_data_gold, 5e4));
+}
+
 TEST_CASE(gru_forward)
 {
     std::size_t batch_size  = 2;
@@ -2797,6 +2882,116 @@ TEST_CASE(gru_bidirectional_seq_1)
     EXPECT(migraphx::verify_range(hs_data, hs_data_gold));
 }
 
+TEST_CASE(gru_fp16)
+{
+    std::size_t batch_size  = 2;
+    std::size_t seq_len     = 3;
+    std::size_t hidden_size = 5;
+    std::size_t input_size  = 3;
+    std::size_t num_dirct   = 1;
+    migraphx::shape w_shape{migraphx::shape::float_type, {num_dirct, 3 * hidden_size, input_size}};
+    std::vector<float> w_data{
+        0.3485,  -0.0378, -0.1782, 0.1416,  -0.3096, -0.2212, -0.3883, 0.1983,  -0.2418,
+        0.1480,  -0.3255, 0.1359,  -0.3551, -0.3605, -0.3482, -0.1424, -0.0495, -0.1640,
+        -0.1979, -0.2577, -0.4097, -0.1211, -0.0412, 0.1801,  0.1721,  -0.4327, -0.0498,
+        0.2628,  -0.1573, -0.1577, 0.2759,  -0.2023, -0.1185, -0.2136, 0.1294,  -0.2331,
+        0.0701,  0.4316,  0.0480,  0.0247,  -0.0166, -0.2729, 0.1712,  -0.3984, -0.3905};
+
+    migraphx::shape r_shape{migraphx::shape::float_type, {num_dirct, 3 * hidden_size, hidden_size}};
+    std::vector<float> r_data{
+        0.2848,  -0.2851, -0.3466, -0.1718, -0.1492, -0.0082, 0.2452,  -0.0401, 0.3399,  0.2529,
+        -0.0953, -0.0903, -0.1518, -0.1373, 0.3848,  -0.0130, -0.4339, 0.0406,  -0.1926, -0.1131,
+        0.4285,  -0.0013, 0.2243,  0.2752,  0.1776,  -0.1720, 0.0822,  -0.0295, 0.1062,  -0.2721,
+        -0.2736, -0.1826, 0.3541,  -0.4259, 0.2188,  0.0706,  0.3650,  0.3947,  0.2522,  0.2179,
+        -0.0744, 0.2122,  -0.4346, 0.2760,  0.4076,  0.1183,  -0.1500, -0.1704, 0.3090,  -0.0706,
+        -0.2442, 0.3021,  0.1680,  0.0783,  -0.3754, -0.3469, -0.2972, -0.0170, 0.4143,  0.3801,
+        0.3852,  -0.1170, -0.2937, 0.2979,  -0.1357, 0.4257,  0.3884,  -0.2916, 0.1071,  0.0934,
+        0.3645,  -0.4310, -0.3480, 0.0702,  -0.1558};
+
+    migraphx::shape b_shape{migraphx::shape::float_type, {num_dirct, 6 * hidden_size}};
+    std::vector<float> bias_data{
+        0.0560,  0.0310, -0.1669, -0.0781, 0.1793, -0.1758, 0.3173,  -0.1650, -0.3732, 0.2946,
+        -0.0912, 0.3118, 0.1391,  0.2755,  0.2695, -0.1059, -0.2357, 0.3629,  -0.2534, -0.0494,
+        0.0556,  0.0881, -0.2592, -0.2213, 0.2310, -0.4044, 0.1801,  0.1438,  0.3108,  -0.3607};
+
+    migraphx::shape in_shape{migraphx::shape::float_type, {seq_len, batch_size, input_size}};
+    std::vector<float> input{-0.8432,
+                             -0.9887,
+                             1.3041,
+                             -2.6430,
+                             -0.3306,
+                             -0.8504,
+                             -0.3933,
+                             0.5151,
+                             -0.2951,
+                             0.0093,
+                             -1.1948,
+                             -0.1239,
+                             0.0373,
+                             1.3211,
+                             0.7854,
+                             -0.4838,
+                             -1.0536,
+                             -0.2529};
+
+    migraphx::shape ih_shape{migraphx::shape::float_type, {num_dirct, batch_size, hidden_size}};
+    std::vector<float> ih_data{
+        -0.0468, 0.5691, -0.0882, 0.8340, 0.1483, -0.3902, -0.5348, 0.4178, 1.0175, 0.9212};
+    float clip = 0.0f;
+
+    migraphx::program p;
+    auto* mm  = p.get_main_module();
+    auto seq  = mm->add_literal(migraphx::literal{in_shape, input});
+    auto w    = mm->add_literal(migraphx::literal{w_shape, w_data});
+    auto r    = mm->add_literal(migraphx::literal{r_shape, r_data});
+    auto bias = mm->add_literal(migraphx::literal{b_shape, bias_data});
+    auto und  = mm->add_instruction(migraphx::make_op("undefined"));
+    auto ih   = mm->add_literal(migraphx::literal{ih_shape, ih_data});
+
+    auto seq_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), seq);
+    auto w_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), w);
+    auto r_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), r);
+    auto bias_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), bias);
+    auto und_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), und);
+    auto ih_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), ih);
+
+    mm->add_instruction(
+        migraphx::make_op("gru",
+                          {{"hidden_size", hidden_size},
+                           {"actv_func",
+                            migraphx::to_value(std::vector<migraphx::operation>{
+                                migraphx::make_op("sigmoid"), migraphx::make_op("tanh")})},
+                           {"direction", migraphx::to_value(migraphx::op::rnn_direction::forward)},
+                           {"clip", clip},
+                           {"linear_before_reset", 1}}),
+        seq_half,
+        w_half,
+        r_half,
+        bias_half,
+        und_half,
+        ih_half);
+
+    p.compile(migraphx::ref::target{});
+    auto hs_concat = p.eval({}).back();
+    std::vector<float> hs_data;
+    hs_concat.visit([&](auto output) { hs_data.assign(output.begin(), output.end()); });
+
+    std::vector<float> hs_data_gold{-0.27298412, 0.42363745, -0.09368783, 0.4823072,  -0.02183238,
+                                    -0.6873896,  0.16144305, 0.31932795,  0.6104771,  0.79759157,
+                                    -0.31791314, 0.5249062,  0.08800987,  0.46404213, -0.11872687,
+                                    -0.26210734, 0.34448293, -0.0176422,  0.48523626, 0.60002893,
+                                    -0.3969709,  0.43360898, 0.35775262,  0.23280787, -0.52179873,
+                                    -0.21944991, 0.4535257,  -0.13735442, 0.51757574, 0.50380427};
+
+    EXPECT(migraphx::verify_range(hs_data, hs_data_gold, 5e4));
+}
+
 TEST_CASE(lstm_forward)
 {
     std::size_t batch_size  = 3;
@@ -4651,4 +4846,148 @@ TEST_CASE(lstm_bidirectional_actv_func)
     }
 }
 
+TEST_CASE(lstm_fp16)
+{
+    std::size_t batch_size  = 3;
+    std::size_t seq_len     = 4;
+    std::size_t hidden_size = 4;
+    std::size_t input_size  = 3;
+    std::size_t num_dirct   = 1;
+    std::vector<float> w_data{
+        0.1236,  -0.3942, 0.4149,  0.0795,  0.4934,  -0.2858, 0.2602,  -0.3098, 0.0567,  0.3344,
+        0.3607,  -0.0551, 0.4952,  0.3799,  0.0630,  -0.3532, 0.0023,  -0.0592, 0.4267,  0.2382,
+        -0.0784, -0.0032, -0.2476, -0.0206, -0.4963, 0.4837,  0.0827,  0.0123,  -0.1203, -0.0279,
+        -0.0049, 0.4721,  -0.3564, -0.1286, 0.4090,  -0.0504, 0.0575,  -0.2138, 0.1071,  0.1976,
+        -0.0758, 0.0139,  -0.0761, 0.3991,  -0.2965, -0.4845, -0.1496, 0.3285};
+
+    std::vector<float> r_data{
+        0.1237,  0.1229,  -0.0766, -0.1144, -0.1186, 0.2922,  0.2478,  0.3159,  -0.0522, 0.1685,
+        -0.4621, 0.1728,  0.0670,  -0.2458, -0.3835, -0.4589, -0.3109, 0.4908,  -0.0133, -0.1858,
+        -0.0590, -0.0347, -0.2353, -0.0671, -0.3812, -0.0004, -0.1432, 0.2406,  0.1033,  -0.0265,
+        -0.3902, 0.0755,  0.3733,  0.4383,  -0.3140, 0.2537,  -0.1818, -0.4127, 0.3506,  0.2562,
+        0.2926,  0.1620,  -0.4849, -0.4861, 0.4426,  0.2106,  -0.0005, 0.4418,  -0.2926, -0.3100,
+        0.1500,  -0.0362, -0.3801, -0.0065, -0.0631, 0.1277,  0.2315,  0.4087,  -0.3963, -0.4161,
+        -0.2169, -0.1344, 0.3468,  -0.2260};
+
+    std::vector<float> bias_data{0.0088,  0.1183,  0.1642,  -0.2631, -0.1330, -0.4008, 0.3881,
+                                 -0.4407, -0.2760, 0.1274,  -0.0083, -0.2885, 0.3949,  -0.0182,
+                                 0.4445,  0.3477,  0.2266,  0.3423,  -0.0674, -0.4067, 0.0807,
+                                 0.1109,  -0.2036, 0.1782,  -0.2467, -0.0730, -0.4216, 0.0316,
+                                 -0.3025, 0.3637,  -0.3181, -0.4655};
+
+    std::vector<float> input_data{
+        -0.5516, 0.2391, -1.6951, -0.4313, -0.9730, -0.2005, 2.3930,  -0.5221, -0.1331,
+        -0.0910, 1.2122, -0.1952, 0.4661,  0.6494,  2.1332,  -1.0972, 0.9816,  0.1122,
+        0.3577,  1.3508, -0.5366, 1.7449,  0.5483,  -0.0701, -0.4100, -2.2344, 0.3685,
+        0.4583,  2.3794, 1.0372,  -0.8887, 0.7892,  -0.4012, -0.2818, -2.3374, 1.5310};
+
+    std::vector<float> ih_data{1.9104,
+                               -1.9004,
+                               0.3337,
+                               0.5741,
+                               0.5671,
+                               0.0458,
+                               0.4514,
+                               -0.8968,
+                               -0.9201,
+                               0.1962,
+                               0.5771,
+                               -0.5332};
+
+    std::vector<float> ic_data{0.9569,
+                               -0.5981,
+                               1.1312,
+                               1.0945,
+                               1.1055,
+                               -0.1212,
+                               -0.9097,
+                               0.7831,
+                               -1.6991,
+                               -1.9498,
+                               -1.2567,
+                               -0.4114};
+
+    std::vector<float> pph_data{1.84369764,
+                                0.68413646,
+                                -0.44892886,
+                                -1.50904413,
+                                0.3860796,
+                                -0.52186625,
+                                1.08474445,
+                                -1.80867321,
+                                1.32594529,
+                                0.4336262,
+                                -0.83699064,
+                                0.49162736};
+
+    float clip = 0.0f;
+    migraphx::shape in_shape{migraphx::shape::float_type, {seq_len, batch_size, input_size}};
+    migraphx::shape ih_shape{migraphx::shape::float_type, {num_dirct, batch_size, hidden_size}};
+    migraphx::shape ic_shape{migraphx::shape::float_type, {num_dirct, batch_size, hidden_size}};
+    migraphx::shape w_shape{migraphx::shape::float_type, {num_dirct, 4 * hidden_size, input_size}};
+    migraphx::shape r_shape{migraphx::shape::float_type, {num_dirct, 4 * hidden_size, hidden_size}};
+    migraphx::shape b_shape{migraphx::shape::float_type, {num_dirct, 8 * hidden_size}};
+    migraphx::shape pph_shape{migraphx::shape::float_type, {num_dirct, 3 * hidden_size}};
+
+    migraphx::program p;
+    auto* mm  = p.get_main_module();
+    auto seq  = mm->add_literal(migraphx::literal{in_shape, input_data});
+    auto w    = mm->add_literal(migraphx::literal{w_shape, w_data});
+    auto r    = mm->add_literal(migraphx::literal{r_shape, r_data});
+    auto bias = mm->add_literal(migraphx::literal{b_shape, bias_data});
+    auto ih   = mm->add_literal(migraphx::literal{ih_shape, ih_data});
+    auto ic   = mm->add_literal(migraphx::literal{ic_shape, ic_data});
+    auto und  = mm->add_instruction(migraphx::make_op("undefined"));
+
+    auto seq_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), seq);
+    auto w_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), w);
+    auto r_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), r);
+    auto bias_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), bias);
+    auto ih_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), ih);
+    auto ic_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), ic);
+    auto und_half = mm->add_instruction(
+        migraphx::make_op("convert", {{"target_type", migraphx::shape::half_type}}), und);
+
+    mm->add_instruction(
+        migraphx::make_op(
+            "lstm",
+            {{"hidden_size", hidden_size},
+             {"actv_func",
+              migraphx::to_value(std::vector<migraphx::operation>{migraphx::make_op("sigmoid"),
+                                                                  migraphx::make_op("tanh"),
+                                                                  migraphx::make_op("tanh")})},
+             {"direction", migraphx::to_value(migraphx::op::rnn_direction::forward)},
+             {"clip", clip},
+             {"input_forget", 0}}),
+        seq_half,
+        w_half,
+        r_half,
+        bias_half,
+        und_half,
+        ih_half,
+        ic_half,
+        und_half);
+    p.compile(migraphx::ref::target{});
+
+    auto hs_concat = p.eval({}).back();
+    std::vector<float> hs_data;
+    hs_concat.visit([&](auto output) { hs_data.assign(output.begin(), output.end()); });
+
+    std::vector<float> hs_data_gold{
+        0.0417273, -0.272355,  0.206765,   0.223879,    0.138193,   -0.0322939, -0.0891815,
+        0.15773,   0.19139,    -0.127708,  -0.409371,   -0.136186,  0.0742487,  -0.0800085,
+        0.259897,  0.0670196,  0.184266,   0.0610048,   -0.138041,  0.0963885,  0.0213755,
+        -0.146027, -0.0324509, -0.0620429, -0.00532985, 0.0440265,  0.29654,    -0.0463156,
+        0.0498799, 0.125772,   0.0533032,  -0.131413,   0.0988431,  -0.018085,  -0.159434,
+        0.030266,  -0.0847427, 0.0874114,  0.304256,    -0.0585745, -0.0223018, 0.131113,
+        0.135643,  -0.0566208, 0.142701,   0.0342236,   -0.198664,  0.0702607};
+    EXPECT(migraphx::verify_range(hs_data, hs_data_gold, 5e4));
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }