Merge branch 'develop' into enable_navi_32_ci

src/targets/gpu/include/migraphx/gpu/gemm_impl.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 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
@@ -24,26 +24,64 @@
 #ifndef MIGRAPHX_GUARD_RTGLIB_GEMM_IMPL_HPP
 #define MIGRAPHX_GUARD_RTGLIB_GEMM_IMPL_HPP
 
+#include <iterator>
 #include <migraphx/shape.hpp>
 #include <migraphx/argument.hpp>
 #include <migraphx/gpu/context.hpp>
 
+// Set this environment variable to "true" to perform GEMM tuning even when the
+// --exhaustive-tune option isn't set.  Can be used to skip slow convolution tuning.
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_ENABLE_GEMM_TUNING);
+
+using milliseconds = std::chrono::duration<double, std::milli>;
+using microseconds = std::chrono::duration<double, std::micro>;
+
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace gpu {
 
-void gemm(context& ctx,
-          const shape& output_shape,
-          const std::vector<argument>& args,
-          float alpha,
-          float beta,
-          bool compute_fp32);
-void gemm(context& ctx,
-          const shape& output_shape,
-          const std::vector<argument>& args,
-          int32_t alpha,
-          int32_t beta,
-          bool compute_fp32);
+/**
+ * @brief Templated implementations of the compute() and finalize() methods of the Gemm operator.
+ *        For each function there are overloads using either float or int32_t for the arguments
+ * alpha and beta.
+ *
+ * @param ctx .
+ * @param output_shape .
+ * @param args .
+ * @param alpha .
+ * @param beta .
+ * @param compute_fp32 .
+ */
+void gemm_compute(context& ctx,
+                  const shape& output_shape,
+                  const std::vector<argument>& args,
+                  float alpha,
+                  float beta,
+                  bool compute_fp32,
+                  int32_t solution_idx);
+
+void gemm_compute(context& ctx,
+                  const shape& output_shape,
+                  const std::vector<argument>& args,
+                  int32_t alpha,
+                  int32_t beta,
+                  bool compute_fp32,
+                  int32_t solution_idx);
+
+int32_t gemm_finalize(context& ctx,
+                      const shape& output_shape,
+                      const std::vector<shape>& input_shapes,
+                      float alpha,
+                      float beta,
+                      bool compute_fp32);
+
+int32_t gemm_finalize(context& ctx,
+                      const shape& output_shape,
+                      const std::vector<shape>& input_shapes,
+                      int32_t alpha,
+                      int32_t beta,
+                      bool compute_fp32,
+                      int32_t solution_idx);
 
 } // namespace gpu
 } // namespace MIGRAPHX_INLINE_NS

src/targets/gpu/include/migraphx/gpu/rocblas.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 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

src/targets/gpu/kernels/include/migraphx/kernels/math.hpp

@@ -101,7 +101,9 @@ MIGRAPHX_DEVICE_MATH(erf, ::erf)
 MIGRAPHX_DEVICE_MATH(exp, ::exp)
 MIGRAPHX_DEVICE_MATH(floor, ::floor)
 MIGRAPHX_DEVICE_MATH(isnan, ::isnan)
+MIGRAPHX_DEVICE_MATH(isinf, ::isinf)
 MIGRAPHX_DEVICE_MATH(log, ::log)
+MIGRAPHX_DEVICE_MATH(nearbyint, ::nearbyint)
 MIGRAPHX_DEVICE_MATH(pow, ::pow)
 MIGRAPHX_DEVICE_MATH(remainder, ::remainder)
 MIGRAPHX_DEVICE_MATH(round, ::round)
@@ -135,6 +137,7 @@ MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, ceil, ::hceil)
 MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, cos, ::hcos)
 MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, exp, ::hexp)
 MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, floor, ::hfloor)
+MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, isinf, ::__hisinf)
 MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, isnan, ::__hisnan)
 MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, log, ::hlog)
 MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, rsqrt, ::hrsqrt)
@@ -150,6 +153,7 @@ MIGRAPHX_DEVICE_MATH_HALF(atan, ::atan)
 MIGRAPHX_DEVICE_MATH_HALF(atanh, ::atanh)
 MIGRAPHX_DEVICE_MATH_HALF(cosh, ::cosh)
 MIGRAPHX_DEVICE_MATH_HALF(erf, ::erf)
+MIGRAPHX_DEVICE_MATH_HALF(nearbyint, ::nearbyint)
 MIGRAPHX_DEVICE_MATH_HALF(pow, ::pow)
 MIGRAPHX_DEVICE_MATH_HALF(remainder, ::remainder)
 MIGRAPHX_DEVICE_MATH_HALF(round, ::round)
@@ -229,10 +233,12 @@ MIGRAPHX_DEVICE_MATH_VEC(erf)
 MIGRAPHX_DEVICE_MATH_VEC(exp)
 MIGRAPHX_DEVICE_MATH_VEC(floor)
 MIGRAPHX_DEVICE_MATH_VEC(fmod)
+MIGRAPHX_DEVICE_MATH_VEC(isinf)
 MIGRAPHX_DEVICE_MATH_VEC(isnan)
 MIGRAPHX_DEVICE_MATH_VEC(log)
 MIGRAPHX_DEVICE_MATH_VEC(max)
 MIGRAPHX_DEVICE_MATH_VEC(min)
+MIGRAPHX_DEVICE_MATH_VEC(nearbyint)
 MIGRAPHX_DEVICE_MATH_VEC(pow)
 MIGRAPHX_DEVICE_MATH_VEC(remainder)
 MIGRAPHX_DEVICE_MATH_VEC(round)

test/api/test_cpu.cpp

@@ -198,4 +198,29 @@ TEST_CASE(set_loop_default_iter_num)
     EXPECT(out_shapes[1].lengths() == out_lens1);
 }
 
+TEST_CASE(set_loop_limit_iterations)
+{
+    migraphx::onnx_options option;
+    option.set_default_loop_iterations(15);
+    option.set_limit_loop_iterations(10);
+    auto p                             = migraphx::parse_onnx("loop_default_test.onnx", option);
+    auto out_shapes                    = p.get_output_shapes();
+    std::vector<std::size_t> out_lens0 = {1};
+    EXPECT(out_shapes[0].lengths() == out_lens0);
+    std::vector<std::size_t> out_lens1 = {10, 1};
+    EXPECT(out_shapes[1].lengths() == out_lens1);
+}
+
+TEST_CASE(set_loop_limit_iterations2)
+{
+    migraphx::onnx_options option;
+    option.set_limit_loop_iterations(10);
+    auto p          = migraphx::parse_onnx("loop_test_implicit_tripcnt.onnx", option);
+    auto out_shapes = p.get_output_shapes();
+    std::vector<std::size_t> out_lens0 = {1};
+    EXPECT(out_shapes[0].lengths() == out_lens0);
+    std::vector<std::size_t> out_lens1 = {10, 1};
+    EXPECT(out_shapes[1].lengths() == out_lens1);
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/api/test_gpu.cpp

@@ -317,4 +317,59 @@ TEST_CASE(loop_test)
     }
 }
 
+TEST_CASE(loop_test_limit_max_iter)
+{
+    auto run_prog = [&](int64_t limit_max_iterations) {
+        migraphx::onnx_options parse_options;
+        parse_options.set_limit_loop_iterations(limit_max_iterations);
+        auto p             = migraphx::parse_onnx("loop_test_implicit_tripcnt.onnx", parse_options);
+        auto shapes_before = p.get_output_shapes();
+        migraphx::compile_options options;
+        options.set_offload_copy();
+        p.compile(migraphx::target("gpu"), options);
+        auto shapes_after = p.get_output_shapes();
+        CHECK(shapes_before.size() == 2);
+        CHECK(bool{shapes_before.front() == shapes_after.front()});
+
+        migraphx::program_parameters pp;
+        auto param_shapes     = p.get_parameter_shapes();
+        auto aas              = param_shapes["a"];
+        std::vector<float> xd = {1.0f};
+        pp.add("a", migraphx::argument(aas, xd.data()));
+        auto bbs              = param_shapes["b"];
+        std::vector<float> yd = {2.0};
+        pp.add("b", migraphx::argument(bbs, yd.data()));
+
+        auto cs   = param_shapes["keep_going_cond"];
+        bool cond = true;
+        pp.add("keep_going_cond", migraphx::argument(cs, &cond));
+
+        auto outputs = p.eval(pp);
+        auto output  = outputs[0];
+        std::vector<std::vector<float>> ret;
+        ret.push_back(output.as_vector<float>());
+
+        output = outputs[1];
+        ret.push_back(output.as_vector<float>());
+
+        return ret;
+    };
+
+    {
+        auto result_vector       = run_prog(5);
+        std::vector<float> gold0 = {2.0f};
+        EXPECT(result_vector.at(0) == gold0);
+        std::vector<float> gold1 = {-2, 4, 0, 0, 0};
+        EXPECT(result_vector.at(1) == gold1);
+    }
+
+    {
+        auto result_vector       = run_prog(20);
+        std::vector<float> gold0 = {2.0f};
+        EXPECT(result_vector.at(0) == gold0);
+        std::vector<float> gold1 = {-2, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+        EXPECT(result_vector.at(1) == gold1);
+    }
+}
+
 int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/codegen_literal.cpp

@@ -64,7 +64,7 @@ TEST_CASE(mul_literal_round_test)
     auto l1 = mm->add_literal(1 / 0.00787402f);
 
     auto mul   = mm->add_instruction(migraphx::make_op("mul"), l0, l1);
-    auto round = mm->add_instruction(migraphx::make_op("round"), mul);
+    auto round = mm->add_instruction(migraphx::make_op("nearbyint"), mul);
 
     mm->add_return({round});
 

test/gpu/gemm_tune.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 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 <iostream>
+#include <vector>
+#include <migraphx/gpu/gemm.hpp>
+#include <hip/hip_runtime_api.h>
+#include <migraphx/gpu/target.hpp>
+#include <migraphx/verify.hpp>
+#include <test.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/iterator_for.hpp>
+
+// includes needed for run_lowering
+#include <migraphx/gpu/lowering.hpp>
+#include <migraphx/auto_contiguous.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/pass_manager.hpp>
+
+// Abbreviated lowering; we don't need the usual cleanup passes for this test
+void run_lowering(migraphx::program& p, bool offload_copy = false)
+{
+    auto ctx = migraphx::gpu::context{};
+    migraphx::run_passes(
+        *p.get_main_module(),
+        {migraphx::auto_contiguous{}, migraphx::gpu::lowering{&ctx, offload_copy}});
+}
+
+/**
+ * Tests the automatic GEMM tuning feature.  In the finalize() method of the gemm op,
+ * rocBLAS API functions are called to quickly benchmark all the GEMM solutions
+ * available in the currently installed rocBLAS library and choose the index of the fastest.
+ */
+TEST_CASE(gemm_tune_with_rocblas)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    migraphx::shape sa{migraphx::shape::float_type, {4, 2}};
+    migraphx::shape sb{migraphx::shape::float_type, {2, 3}};
+    auto a = mm->add_parameter("a", sa);
+    auto b = mm->add_parameter("b", sb);
+
+    migraphx::operation dot_op = migraphx::make_op("dot");
+    mm->add_instruction(dot_op, a, b);
+
+    // lowering adds gemm implementation for dot operator
+    run_lowering(p);
+
+    migraphx::target gpu_t = migraphx::gpu::target{};
+    migraphx::compile_options options;
+    options.exhaustive_tune = true;
+    p.compile(gpu_t, options);
+
+    migraphx::value solution_idx(0);
+    for(auto ins : iterator_for(*p.get_main_module()))
+    {
+        if(ins->name() == "gpu::gemm")
+        {
+            auto gemm_op = migraphx::get_operation(ins);
+
+            // tuned solution index is not deterministic, but anything other than 0
+            // (default, invalid, or not available) is good.
+            // gemm_op.to_value().debug_print();
+            solution_idx = gemm_op.to_value()["solution_idx"];
+            break;
+        }
+    }
+#ifdef MIGRAPHX_USE_ROCBLAS_TUNING_API
+    EXPECT(0 != solution_idx.to<std::size_t>());
+#else
+    EXPECT(0 == solution_idx.to<std::size_t>());
+#endif
+}
+
+// GEMM tuning of a strided-batch matrix; invokes rocblas_gemm_strided_batched_ex
+TEST_CASE(gemm_tune_strided)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    migraphx::shape sa{migraphx::shape::float_type, {4, 2, 2}};
+    migraphx::shape sb{migraphx::shape::float_type, {4, 2, 2}};
+    migraphx::shape s_output{migraphx::shape::float_type, {4, 2, 2}};
+    auto a      = mm->add_parameter("a", sa);
+    auto b      = mm->add_parameter("b", sb);
+    auto output = mm->add_parameter("out", s_output);
+
+    auto gemm_oper = migraphx::make_op("gpu::gemm", {{"beta", 2}});
+    mm->add_instruction(gemm_oper, a, b, output);
+
+    migraphx::target gpu_t = migraphx::gpu::target{};
+    migraphx::compile_options options;
+    options.exhaustive_tune = true;
+    p.compile(gpu_t, options);
+
+    migraphx::value solution_idx(0);
+    for(auto ins : iterator_for(*p.get_main_module()))
+    {
+        if(ins->name() == "gpu::gemm")
+        {
+            auto gemm_op = migraphx::get_operation(ins);
+            auto gemmv   = gemm_op.to_value();
+
+            // tuned solution index is not deterministic, but anything other than 0
+            // (default, invalid, or not available) is good.
+            solution_idx = gemm_op.to_value()["solution_idx"];
+            break;
+        }
+    }
+#ifdef MIGRAPHX_USE_ROCBLAS_TUNING_API
+    EXPECT(0 != solution_idx.to<std::size_t>());
+#else
+    EXPECT(0 == solution_idx.to<std::size_t>());
+#endif
+}
+
+// GEMM tuning of a strided-batch matrix; created by lowering
+TEST_CASE(gemm_tune_strided_lowered)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    // At time of writing this test, gemm_impl considers a shape is strided if it has
+    // at least three dimensions and the 3rd-to-last is nonzero, invoking
+    // rocblas_gemm_strided_batched_ex. Also, DOT operator requires all dimensions except the last
+    // two to be equal.
+    migraphx::shape sa{migraphx::shape::float_type, {4, 2, 5}};
+    migraphx::shape sb{migraphx::shape::float_type, {4, 5, 3}};
+    auto a = mm->add_parameter("a", sa);
+    auto b = mm->add_parameter("b", sb);
+
+    migraphx::operation dot_op = migraphx::make_op("dot");
+    mm->add_instruction(dot_op, a, b);
+
+    // lowering adds gemm implementation for dot operator
+    run_lowering(p);
+
+    migraphx::target gpu_t = migraphx::gpu::target{};
+    migraphx::compile_options options;
+    options.exhaustive_tune = true;
+    p.compile(gpu_t, options);
+
+    migraphx::value solution_idx(0);
+    for(auto ins : iterator_for(*p.get_main_module()))
+    {
+        if(ins->name() == "gpu::gemm")
+        {
+            auto gemm_op = migraphx::get_operation(ins);
+
+            // tuned solution index is not deterministic, but anything other than 0
+            // (default, invalid, or not available) is good.
+            solution_idx = gemm_op.to_value()["solution_idx"];
+            break;
+        }
+    }
+#ifdef MIGRAPHX_USE_ROCBLAS_TUNING_API
+    EXPECT(0 != solution_idx.to<std::size_t>());
+#else
+    EXPECT(0 == solution_idx.to<std::size_t>());
+#endif
+}
+
+TEST_CASE(gemm_tune_invalid_sol_index)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    migraphx::shape sa{migraphx::shape::float_type, {4, 2}};
+    migraphx::shape sb{migraphx::shape::float_type, {2, 3}};
+    migraphx::shape s_output{migraphx::shape::float_type, {4, 3}};
+    auto a      = mm->add_parameter("a", sa);
+    auto b      = mm->add_parameter("b", sb);
+    auto output = mm->add_parameter("out", s_output);
+
+    auto gemm_oper = migraphx::make_op("gpu::gemm", {{"solution_idx", 987654321}});
+    mm->add_instruction(gemm_oper, a, b, output);
+
+    migraphx::target gpu_t = migraphx::gpu::target{};
+    migraphx::compile_options options;
+    options.exhaustive_tune = true;
+    p.compile(gpu_t, options);
+
+    migraphx::value solution_idx(0);
+    for(auto ins : iterator_for(*p.get_main_module()))
+    {
+        if(ins->name() == "gpu::gemm")
+        {
+            auto gemm_op = migraphx::get_operation(ins);
+            auto gemmv   = gemm_op.to_value();
+
+            // given invalid starting index, should return default 0
+            solution_idx = gemm_op.to_value()["solution_idx"];
+            break;
+        }
+    }
+#ifdef MIGRAPHX_USE_ROCBLAS_TUNING_API
+    EXPECT(0 == solution_idx.to<std::size_t>());
+#else
+    EXPECT(0 != solution_idx.to<std::size_t>());
+#endif
+}
+
+int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/onnx/.onnxrt-commit

-2eeafc37bca21dc8bf337dda7020b486543162d7
+b7b8b5b2ce80edb33990c7ae0fedac6ae3c623f4

test/onnx/gen_onnx.py

@@ -3858,6 +3858,64 @@ def instance_norm_val_3d_test():
     return ([node], [], [y], [x_tensor, scale_tensor, bias_tensor])
 
 
+@onnx_test()
+def isinf_half_test():
+    t1 = helper.make_tensor_value_info('t1', TensorProto.FLOAT16, [2, 3])
+    t2 = helper.make_tensor_value_info('t2', TensorProto.BOOL, [2, 3])
+
+    node = onnx.helper.make_node(
+        'IsInf',
+        inputs=['t1'],
+        outputs=['t2'],
+    )
+    return ([node], [t1], [t2])
+
+
+@onnx_test()
+def isinf_neg_test():
+    t1 = helper.make_tensor_value_info('t1', TensorProto.FLOAT, [2, 3])
+    t2 = helper.make_tensor_value_info('t2', TensorProto.BOOL, [2, 3])
+
+    node = onnx.helper.make_node(
+        'IsInf',
+        detect_negative=[1],
+        detect_positive=[0],
+        inputs=['t1'],
+        outputs=['t2'],
+    )
+    return ([node], [t1], [t2])
+
+
+@onnx_test()
+def isinf_double_pos_test():
+    t1 = helper.make_tensor_value_info('t1', TensorProto.DOUBLE, [2, 3])
+    t2 = helper.make_tensor_value_info('t2', TensorProto.BOOL, [2, 3])
+
+    node = onnx.helper.make_node(
+        'IsInf',
+        detect_negative=[0],
+        detect_positive=[1],
+        inputs=['t1'],
+        outputs=['t2'],
+    )
+    return ([node], [t1], [t2])
+
+
+@onnx_test()
+def isinf_no_detect_test():
+    t1 = helper.make_tensor_value_info('t1', TensorProto.FLOAT, [2, 3])
+    t2 = helper.make_tensor_value_info('t2', TensorProto.BOOL, [2, 3])
+
+    node = onnx.helper.make_node(
+        'IsInf',
+        detect_negative=[0],
+        detect_positive=[0],
+        inputs=['t1'],
+        outputs=['t2'],
+    )
+    return ([node], [t1], [t2])
+
+
 @onnx_test()
 def isnan_float_test():
     t1 = helper.make_tensor_value_info('t1', TensorProto.FLOAT, [2, 3])
@@ -4276,6 +4334,50 @@ def loop_test():
     return ([node], [iter, cond, a, b], [b_loop, uout])
 
 
+@onnx_test()
+def loop_test_implicit_tripcnt():
+    body = helper.make_graph([
+        helper.make_node("Add", ["a", "b_in"], ["my_local"]),
+        helper.make_node("Sub", ["a", "b_in"], ["a_sub_b_in"]),
+        helper.make_node("Greater", ["my_local", "a_sub_b_in"],
+                         ["keep_going"]),
+        helper.make_node("Add", ["a_sub_b_in", "a_sub_b_in"],
+                         ["user_defined_vals"]),
+    ], "body", [
+        helper.make_tensor_value_info('iteration_num', TensorProto.INT64, [1]),
+        helper.make_tensor_value_info('keep_going_inp', TensorProto.BOOL, [1]),
+        helper.make_tensor_value_info('b_in', TensorProto.FLOAT, [1])
+    ], [
+        helper.make_tensor_value_info('keep_going', TensorProto.BOOL, [1]),
+        helper.make_tensor_value_info('a_sub_b_in', TensorProto.FLOAT, [1]),
+        helper.make_tensor_value_info('my_local', TensorProto.FLOAT, [1]),
+        helper.make_tensor_value_info('user_defined_vals', TensorProto.FLOAT,
+                                      [1]),
+    ])
+
+    iter = helper.make_tensor(name='max_trip_count',
+                              data_type=TensorProto.INT64,
+                              dims=[1],
+                              vals=[15])
+
+    node = helper.make_node(
+        "Loop",
+        inputs=["max_trip_count", "keep_going_cond", "b"],
+        outputs=["b_loop", "my_local_loop", "user_defined_vals_loop"],
+        body=body)
+
+    a = helper.make_tensor_value_info('a', TensorProto.FLOAT, [1])
+    b = helper.make_tensor_value_info('b', TensorProto.FLOAT, [1])
+    cond = helper.make_tensor_value_info('keep_going_cond', TensorProto.BOOL,
+                                         [1])
+
+    b_loop = helper.make_tensor_value_info('b_loop', TensorProto.FLOAT, [1])
+    uout = helper.make_tensor_value_info('user_defined_vals_loop',
+                                         TensorProto.FLOAT, [2, 1])
+
+    return ([node], [cond, a, b], [b_loop, uout], [iter])
+
+
 @onnx_test()
 def lpnormalization_axis_error_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [2, 3])
@@ -6985,6 +7087,16 @@ def roialign_test():
     return ([node], [x, roi, bi], [y])
 
 
+@onnx_test()
+def round_half_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT16, [4, 4])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT16, [4, 4])
+
+    node = onnx.helper.make_node('Round', inputs=['x'], outputs=['y'])
+
+    return ([node], [x], [y])
+
+
 @onnx_test()
 def scatter_add_test():
     x = helper.make_tensor_value_info('data', TensorProto.FLOAT, [3, 4, 5, 6])
@@ -7904,6 +8016,32 @@ def slice_var_input_dyn1():
     return ([node], [data, starts, ends, axes], [output])
 
 
+@onnx_test()
+def slice_var_input_default_steps():
+    step = np.array([1, 1])
+    step_tensor = helper.make_tensor(name="step",
+                                     data_type=TensorProto.INT64,
+                                     dims=step.shape,
+                                     vals=step.astype(int))
+    arg_step = helper.make_node("Constant",
+                                inputs=[],
+                                outputs=['arg_step'],
+                                value=step_tensor)
+
+    data = helper.make_tensor_value_info('data', TensorProto.FLOAT, [None, 2])
+    starts = helper.make_tensor_value_info('starts', TensorProto.INT64, [2])
+    ends = helper.make_tensor_value_info('ends', TensorProto.INT64, [2])
+    axes = helper.make_tensor_value_info('axes', TensorProto.INT64, [2])
+    output = helper.make_tensor_value_info('output', TensorProto.FLOAT, [1, 2])
+
+    node = onnx.helper.make_node(
+        'Slice',
+        inputs=['data', 'starts', 'ends', 'axes', 'arg_step'],
+        outputs=['output'])
+
+    return ([arg_step, node], [data, starts, ends, axes], [output])
+
+
 @onnx_test()
 def slice_var_input_steps_error():
     step = np.array([2, 1])
@@ -7917,9 +8055,9 @@ def slice_var_input_steps_error():
                                 value=step_tensor)
 
     data = helper.make_tensor_value_info('data', TensorProto.FLOAT, [3, 2])
-    starts = helper.make_tensor_value_info('starts', TensorProto.FLOAT, [2])
-    ends = helper.make_tensor_value_info('ends', TensorProto.FLOAT, [2])
-    axes = helper.make_tensor_value_info('axes', TensorProto.FLOAT, [2])
+    starts = helper.make_tensor_value_info('starts', TensorProto.INT64, [2])
+    ends = helper.make_tensor_value_info('ends', TensorProto.INT64, [2])
+    axes = helper.make_tensor_value_info('axes', TensorProto.INT64, [2])
     output = helper.make_tensor_value_info('output', TensorProto.FLOAT, [1, 2])
 
     node = onnx.helper.make_node(
@@ -8929,6 +9067,20 @@ def upsample_test():
     return ([node], [X], [Y], [scale_tensor])
 
 
+@onnx_test()
+def upsample_ver7_test():
+    X = helper.make_tensor_value_info('X', TensorProto.FLOAT, [1, 1, 2, 2])
+    Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [1, 1, 4, 6])
+
+    node = onnx.helper.make_node('Upsample',
+                                 inputs=['X'],
+                                 outputs=['Y'],
+                                 mode='nearest',
+                                 scales=[1.0, 1.0, 2.0, 3.0])
+
+    return ([node], [X], [Y])
+
+
 @onnx_test()
 def variable_batch_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT,

test/onnx/isinf_half_test.onnx

+	isinf_half_test:N
+
+t1t2"IsInfisinf_half_testZ
+t1
+
+
+
+b
+t2
+	
+
+B
\ No newline at end of file

test/onnx/loop_test_implicit_tripcnt.onnx

+	loop_test_implicit_tripcnt:
+
+max_trip_count
+keep_going_cond
+
bb_loop
my_local_loopuser_defined_vals_loop"Loop*
+body2
+
+
a
+b_inmy_local"Add
+
+
a
+b_in
+a_sub_b_in"Sub
++
+my_local
+
+a_sub_b_in
+keep_going"Greater
+0
+
+a_sub_b_in
+
+a_sub_b_inuser_defined_vals"AddbodyZ
+iteration_num
+
+
+
Z
+keep_going_inp
+
+	
+
Z
+b_in
+
+

+
b
+
+keep_going
+
+	
+
b
+
+a_sub_b_in
+
+

+
b
+my_local
+
+

+
b
+user_defined_vals
+
+

+

loop_test_implicit_tripcnt*
:
Bmax_trip_countZ
+keep_going_cond
+
+	
+
Z
+
a
+
+

+
Z
+
b
+
+

+
b
+b_loop
+
+

+
b(
+user_defined_vals_loop
+

+
+
B
\ No newline at end of file

test/onnx/onnx_test.cpp

@@ -3413,6 +3413,82 @@ TEST_CASE(if_tuple_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(isinf_half_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s{migraphx::shape::half_type, {2, 3}};
+    auto t1  = mm->add_parameter("t1", s);
+    auto ret = mm->add_instruction(migraphx::make_op("isinf"), t1);
+    mm->add_return({ret});
+
+    auto prog = migraphx::parse_onnx("isinf_half_test.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(isinf_neg_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+    auto t1     = mm->add_parameter("t1", s);
+    auto is_inf = mm->add_instruction(migraphx::make_op("isinf"), t1);
+    auto zero_l = mm->add_literal(migraphx::literal{migraphx::shape::float_type, {0}});
+    auto mb_zero =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", s.lens()}}), zero_l);
+
+    auto is_neg = mm->add_instruction(migraphx::make_op("less"), t1, mb_zero);
+    if(is_neg->get_shape().type() != migraphx::shape::bool_type)
+    {
+        is_neg = mm->add_instruction(
+            migraphx::make_op("convert", {{"target_type", migraphx::shape::bool_type}}), is_neg);
+    }
+    auto ret = mm->add_instruction(migraphx::make_op("logical_and"), is_inf, is_neg);
+    mm->add_return({ret});
+
+    auto prog = migraphx::parse_onnx("isinf_neg_test.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(isinf_double_pos_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s{migraphx::shape::double_type, {2, 3}};
+    auto t1     = mm->add_parameter("t1", s);
+    auto is_inf = mm->add_instruction(migraphx::make_op("isinf"), t1);
+    auto zero_l = mm->add_literal(migraphx::literal{migraphx::shape::double_type, {0}});
+    auto mb_zero =
+        mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", s.lens()}}), zero_l);
+
+    auto is_neg = mm->add_instruction(migraphx::make_op("greater"), t1, mb_zero);
+    if(is_neg->get_shape().type() != migraphx::shape::bool_type)
+    {
+        is_neg = mm->add_instruction(
+            migraphx::make_op("convert", {{"target_type", migraphx::shape::bool_type}}), is_neg);
+    }
+    auto ret = mm->add_instruction(migraphx::make_op("logical_and"), is_inf, is_neg);
+    mm->add_return({ret});
+
+    auto prog = migraphx::parse_onnx("isinf_double_pos_test.onnx");
+    EXPECT(p == prog);
+}
+
+TEST_CASE(isinf_no_detect_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    migraphx::shape s{migraphx::shape::float_type, {2, 3}};
+    mm->add_parameter("t1", s);
+    auto ret = mm->add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", s.lens()}}),
+        mm->add_literal(migraphx::literal{migraphx::shape{migraphx::shape::bool_type}, {false}}));
+    mm->add_return({ret});
+
+    auto prog = migraphx::parse_onnx("isinf_no_detect_test.onnx");
+    EXPECT(p == prog);
+}
+
 TEST_CASE(isnan_float_test)
 {
     migraphx::program p;
@@ -5712,9 +5788,9 @@ TEST_CASE(quantizelinear_test)
     auto l1_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {5}}}), l1);
     auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
-    auto round = mm->add_instruction(migraphx::make_op("round"), div);
-    auto s     = round->get_shape();
-    auto clip  = insert_quantizelinear_clip(*mm, div, round, s, 0, 255);
+    auto nearbyint = mm->add_instruction(migraphx::make_op("nearbyint"), div);
+    auto s         = nearbyint->get_shape();
+    auto clip      = insert_quantizelinear_clip(*mm, div, nearbyint, s, 0, 255);
     mm->add_instruction(
         migraphx::make_op("convert",
                           {{"target_type", migraphx::to_value(migraphx::shape::uint8_type)}}),
@@ -5737,9 +5813,9 @@ TEST_CASE(quantizelinear_int32_test)
                           {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
         l0);
     auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
-    auto round = mm->add_instruction(migraphx::make_op("round"), div);
-    auto s     = round->get_shape();
-    auto clip  = insert_quantizelinear_clip(*mm, div, round, s, 0, 255);
+    auto nearbyint = mm->add_instruction(migraphx::make_op("nearbyint"), div);
+    auto s         = nearbyint->get_shape();
+    auto clip      = insert_quantizelinear_clip(*mm, div, nearbyint, s, 0, 255);
     mm->add_instruction(
         migraphx::make_op("convert",
                           {{"target_type", migraphx::to_value(migraphx::shape::uint8_type)}}),
@@ -5759,7 +5835,7 @@ TEST_CASE(quantizelinear_zero_point_test)
     auto l1_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {5}}}), l1);
     auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
-    auto round = mm->add_instruction(migraphx::make_op("round"), div);
+    auto round = mm->add_instruction(migraphx::make_op("nearbyint"), div);
     auto l2_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {5}}}), l2);
     l2_mbcast = mm->add_instruction(
@@ -5792,7 +5868,7 @@ migraphx::program make_quantizelinear_axis_prog()
         migraphx::make_op("broadcast", {{"axis", axis}, {"out_lens", input_lens}}), l1);
 
     auto div      = mm->add_instruction(migraphx::make_op("div"), l0, l1_bcast);
-    auto round    = mm->add_instruction(migraphx::make_op("round"), div);
+    auto round    = mm->add_instruction(migraphx::make_op("nearbyint"), div);
     auto l2_bcast = mm->add_instruction(
         migraphx::make_op("broadcast", {{"axis", axis}, {"out_lens", input_lens}}), l2);
     l2_bcast = mm->add_instruction(
@@ -6481,9 +6557,8 @@ TEST_CASE(resize_nonstd_input_test)
     auto tx =
         mm->add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), inx);
     mm->add_instruction(migraphx::make_op("undefined"));
-    auto tx_cont = mm->add_instruction(migraphx::make_op("contiguous"), tx);
 
-    auto lrsp = mm->add_instruction(migraphx::make_op("reshape", {{"dims", {8}}}), tx_cont);
+    auto lrsp = mm->add_instruction(migraphx::make_op("reshape", {{"dims", {8}}}), tx);
     auto r    = mm->add_instruction(migraphx::make_op("gather", {{"axis", 0}}), lrsp, li);
     mm->add_return({r});
 
@@ -6922,7 +6997,7 @@ TEST_CASE(round_test)
     migraphx::program p;
     auto* mm   = p.get_main_module();
     auto input = mm->add_parameter("x", migraphx::shape{migraphx::shape::double_type, {10, 5}});
-    mm->add_instruction(migraphx::make_op("round"), input);
+    mm->add_instruction(migraphx::make_op("nearbyint"), input);
 
     auto prog = optimize_onnx("round_test.onnx");
     EXPECT(p == prog);
@@ -7578,6 +7653,25 @@ TEST_CASE(slice_var_input_dyn1)
     EXPECT(p == prog);
 }
 
+TEST_CASE(slice_var_input_default_steps)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto data =
+        mm->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {{3, 8}, {2, 2}}});
+    auto starts = mm->add_parameter("starts", migraphx::shape{migraphx::shape::int64_type, {2}});
+    auto ends   = mm->add_parameter("ends", migraphx::shape{migraphx::shape::int64_type, {2}});
+    auto axes   = mm->add_parameter("axes", migraphx::shape{migraphx::shape::int64_type, {2}});
+    mm->add_literal({{migraphx::shape::int64_type, {2}}, {1, 1}});
+    auto ret = mm->add_instruction(migraphx::make_op("slice"), data, starts, ends, axes);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {3, 8};
+    auto prog                     = parse_onnx("slice_var_input_default_steps.onnx", options);
+    EXPECT(p == prog);
+}
+
 TEST_CASE(slice_var_input_steps_error)
 {
     EXPECT(test::throws([&] { migraphx::parse_onnx("slice_var_input_steps_error.onnx"); }));
@@ -8342,6 +8436,27 @@ TEST_CASE(upsample_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(upsample_ver7_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+
+    migraphx::shape sx{migraphx::shape::float_type, {1, 1, 2, 2}};
+    auto ix = mm->add_parameter("X", sx);
+
+    migraphx::shape si{migraphx::shape::int32_type, {1, 1, 4, 6}};
+    std::vector<int> ind = {0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 2, 2, 2, 3, 3, 3};
+
+    auto li  = mm->add_literal(migraphx::literal(si, ind));
+    auto rsp = mm->add_instruction(migraphx::make_op("reshape", {{"dims", {4}}}), ix);
+    auto r   = mm->add_instruction(migraphx::make_op("gather", {{"axis", 0}}), rsp, li);
+    mm->add_return({r});
+
+    auto prog = migraphx::parse_onnx("upsample_ver7_test.onnx");
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(unknown_test_throw_print_error)
 {
     migraphx::onnx_options options;

test/onnx/reshape_variable_input_test0.onnx

+reshape_variable_input_test0:q
+
+
0
+
1
2"Reshapereshape_variable_input_test0Z
+
0
+

+
+
+Z
+
1
+
+
+b
+
2
+

+
+B
\ No newline at end of file

test/onnx/round_half_test.onnx

+	round_half_test:J
+
+
x
y"Roundround_half_testZ
+
x
+
+
+
+b
+
y
+
+
+
+B
\ No newline at end of file