manual merge

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/eliminate_allocation_test.cpp

@@ -55,7 +55,7 @@ struct allocate
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 

test/eliminate_concat_test.cpp

@@ -60,7 +60,7 @@ struct concat
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 
@@ -104,7 +104,7 @@ struct allocate
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 

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/fuse_mlir.cpp

@@ -34,7 +34,8 @@
 
 void run_pass(migraphx::program& p)
 {
-    migraphx::run_passes(p, {migraphx::gpu::fuse_mlir{}, migraphx::dead_code_elimination{}});
+    migraphx::run_passes(
+        p, {migraphx::gpu::fuse_mlir{.enable_extra = true}, migraphx::dead_code_elimination{}});
 }
 
 template <class F>
@@ -152,6 +153,8 @@ TEST_CASE(int_quant_dot_tanh_fails)
 int main(int argc, const char* argv[])
 {
     if(migraphx::gpu::mlir_enabled())
+    {
         test::run(argc, argv);
+    }
     return 0;
 }

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/gpu/jit.cpp

@@ -155,7 +155,7 @@ int main() {}
 
 migraphx::src_file make_src_file(const std::string& name, const std::string& content)
 {
-    return {name, std::make_pair(content.data(), content.data() + content.size())};
+    return {name, content};
 }
 
 TEST_CASE(simple_compile_hip)

test/gpu/pack_int8_args.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_ref.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/gpu/target.hpp>
-#include <migraphx/gpu/allocation_model.hpp>
-#include <migraphx/apply_alpha_beta.hpp>
-#include <migraphx/adjust_allocation.hpp>
-#include <migraphx/gpu/pack_int8_args.hpp>
-#include <migraphx/gpu/rocblas.hpp>
-#include <migraphx/gpu/device_name.hpp>
-#include <migraphx/auto_contiguous.hpp>
-#include <migraphx/eliminate_contiguous.hpp>
-#include <migraphx/dead_code_elimination.hpp>
-#include <migraphx/replace_allocate.hpp>
-#include <migraphx/instruction.hpp>
-#include <migraphx/iterator_for.hpp>
-#include <migraphx/pass_manager.hpp>
-#include <migraphx/make_op.hpp>
-#include <test.hpp>
-#include "make_precompile_op.hpp"
-
-// Treat some operators as compilable to enable lowering
-MIGRAPHX_GPU_TEST_PRECOMPILE("add", "mul", "convert")
-
-void run_passes(migraphx::module& m, migraphx::gpu::context& ctx)
-{
-    migraphx::run_passes(m,
-                         {migraphx::auto_contiguous{},
-                          migraphx::gpu::lowering{&ctx, false},
-                          migraphx::eliminate_contiguous{"gpu::contiguous"},
-                          migraphx::dead_code_elimination{},
-                          migraphx::replace_allocate{migraphx::gpu::gpu_allocation_model{}},
-                          migraphx::dead_code_elimination{},
-                          migraphx::gpu::pack_int8_args{},
-                          migraphx::dead_code_elimination{}});
-}
-
-TEST_CASE(quant_dot)
-{
-    auto create_module = [] {
-        migraphx::module m("test");
-        migraphx::shape m1_shape{migraphx::shape::int8_type, {5, 8}};
-        migraphx::shape m2_shape{migraphx::shape::int8_type, {8, 7}};
-        migraphx::shape m3_shape{migraphx::shape::int32_type, {5, 7}};
-
-        auto l1 = m.add_parameter("a", m1_shape);
-        auto l2 = m.add_parameter("b", m2_shape);
-        auto l3 = m.add_parameter("c", m3_shape);
-        auto r =
-            migraphx::add_apply_alpha_beta(m, {l1, l2, l3}, migraphx::make_op("quant_dot"), 1, 1);
-        m.add_return({r});
-        return m;
-    };
-
-    auto create_optimized_int8_x4 = [](bool int8_x4) {
-        migraphx::module m("test");
-        migraphx::shape m1_shape{migraphx::shape::int8_type, {5, 8}};
-        migraphx::shape m2_shape{migraphx::shape::int8_type, {8, 7}};
-        migraphx::shape m3_shape{migraphx::shape::int32_type, {5, 7}};
-
-        auto l1         = m.add_parameter("a", m1_shape);
-        auto l2         = m.add_parameter("b", m2_shape);
-        auto l3         = m.add_parameter("c", m3_shape);
-        auto beta       = m.add_literal(1);
-        auto output     = m.add_parameter("test:#output_0", m3_shape);
-        auto gemm_alloc = m.add_instruction(
-            migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(m3_shape)}}));
-
-        auto packa = l2;
-        if(int8_x4)
-        {
-            auto alloc = m.add_instruction(
-                migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(m2_shape)}}));
-            packa = m.add_instruction(migraphx::make_op("gpu::int8_gemm_pack_a"), l2, alloc);
-        }
-        auto gemm = m.add_instruction(
-            migraphx::make_op("gpu::quant_gemm",
-                              {{"int8_x4_format", int8_x4},
-                               {"compute_fp32", migraphx::gpu::get_compute_fp32_flag()}}),
-            l1,
-            packa,
-            gemm_alloc);
-
-        auto beta_broadcast = m.add_instruction(
-            migraphx::make_op("multibroadcast", {{"out_lens", m3_shape.lens()}}), beta);
-        auto mul_alloc = m.add_instruction(
-            migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(m3_shape)}}));
-        auto m3_beta  = m.add_instruction(make_precompile_op("mul"), l3, beta_broadcast, mul_alloc);
-        auto gemm_add = m.add_instruction(make_precompile_op("add"), gemm, m3_beta, output);
-        m.add_return({gemm_add});
-
-        return m;
-    };
-
-    auto m1  = create_module();
-    auto ctx = migraphx::gpu::context{};
-    run_passes(m1, ctx);
-
-    bool int8_x4 = migraphx::gpu::get_int8_x4_format(ctx);
-    auto m2      = create_optimized_int8_x4(int8_x4);
-    EXPECT(m1 == m2);
-}
-
-TEST_CASE(quant_dot_trans)
-{
-    auto create_module = [] {
-        migraphx::module m("test");
-        migraphx::shape s1{migraphx::shape::int8_type, {3, 2, 8, 5}};
-        migraphx::shape s2{migraphx::shape::int8_type, {3, 2, 7, 8}};
-
-        auto l1 = m.add_parameter("a", s1);
-        auto tl1 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l1);
-        auto l2 = m.add_parameter("b", s2);
-        auto tl2 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l2);
-        auto r = migraphx::add_apply_alpha_beta(m, {tl1, tl2}, migraphx::make_op("quant_dot"), 3);
-        m.add_return({r});
-        return m;
-    };
-
-    auto create_optimized_int8_x4 = [](bool int8_x4) {
-        migraphx::module m("test");
-        migraphx::shape s1{migraphx::shape::int8_type, {3, 2, 8, 5}};
-        migraphx::shape s2{migraphx::shape::int8_type, {3, 2, 7, 8}};
-        migraphx::shape s3{migraphx::shape::int32_type, {3, 2, 5, 7}};
-
-        auto l1     = m.add_parameter("a", s1);
-        auto l2     = m.add_parameter("b", s2);
-        auto alpha  = m.add_literal(3);
-        auto output = m.add_parameter("test:#output_0", s3);
-
-        auto tl1 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l1);
-        migraphx::shape ts1{migraphx::shape::int8_type, {3, 2, 5, 8}};
-
-        auto tl2 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l2);
-        migraphx::shape ts2{migraphx::shape::int8_type, {3, 2, 8, 7}};
-
-        auto alpha_broadcast = m.add_instruction(
-            migraphx::make_op("multibroadcast", {{"out_lens", tl1->get_shape().lens()}}), alpha);
-        auto alpha_alloc = m.add_instruction(migraphx::make_op(
-            "hip::allocate",
-            {{"shape",
-              migraphx::to_value(migraphx::shape(migraphx::shape::int32_type, {3, 2, 5, 8}))}}));
-        // alpha = int32 and tl1 = int8, convert tl1 to int32 for multiplication and then convert
-        // back result to int8
-        auto tl1_convert =
-            m.add_instruction(make_precompile_op(migraphx::make_op(
-                                  "convert", {{"target_type", alpha->get_shape().type()}})),
-                              tl1,
-                              alpha_alloc);
-        auto mul_alloc = m.add_instruction(migraphx::make_op(
-            "hip::allocate", {{"shape", migraphx::to_value(alpha_alloc->get_shape())}}));
-        auto tl1_alpha_int32 =
-            m.add_instruction(make_precompile_op("mul"), alpha_broadcast, tl1_convert, mul_alloc);
-        // convert mul_res to int8
-        auto tl1_alpha_int8_alloc = m.add_instruction(
-            migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ts1)}}));
-        auto tl1_alpha_int8 =
-            m.add_instruction(make_precompile_op(migraphx::make_op(
-                                  "convert", {{"target_type", tl1->get_shape().type()}})),
-                              tl1_alpha_int32,
-                              tl1_alpha_int8_alloc);
-
-        auto packb = tl2;
-        if(int8_x4)
-        {
-            auto allocpb = m.add_instruction(
-                migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ts2)}}));
-            packb = m.add_instruction(migraphx::make_op("gpu::int8_gemm_pack_a"), tl2, allocpb);
-        }
-
-        auto gemm = m.add_instruction(
-            migraphx::make_op("gpu::quant_gemm",
-                              {{"int8_x4_format", int8_x4},
-                               {"compute_fp32", migraphx::gpu::get_compute_fp32_flag()}}),
-            tl1_alpha_int8,
-            packb,
-            output);
-        m.add_return({gemm});
-
-        return m;
-    };
-
-    auto m1  = create_module();
-    auto ctx = migraphx::gpu::context{};
-    run_passes(m1, ctx);
-
-    bool int8_x4 = migraphx::gpu::get_int8_x4_format(ctx);
-    auto m2      = create_optimized_int8_x4(int8_x4);
-
-    EXPECT(m1 == m2);
-}
-
-TEST_CASE(quant_dot_pad)
-{
-    auto create_module = [] {
-        migraphx::module m("test");
-        migraphx::shape s1{migraphx::shape::int8_type, {5, 6}};
-        migraphx::shape s2{migraphx::shape::int8_type, {6, 7}};
-        migraphx::shape s3{migraphx::shape::int32_type, {5, 7}};
-
-        auto l1 = m.add_parameter("a", s1);
-        auto l2 = m.add_parameter("b", s2);
-        auto l3 = m.add_parameter("c", s3);
-        auto r =
-            migraphx::add_apply_alpha_beta(m, {l1, l2, l3}, migraphx::make_op("quant_dot"), 1, 1);
-        m.add_return({r});
-        return m;
-    };
-
-    auto create_optimized_int8_x4 = [](bool int8_x4) {
-        migraphx::module m("test");
-        migraphx::shape s1{migraphx::shape::int8_type, {5, 6}};
-        migraphx::shape ps1{migraphx::shape::int8_type, {5, 8}};
-        migraphx::shape s2{migraphx::shape::int8_type, {6, 7}};
-        migraphx::shape ps2{migraphx::shape::int8_type, {8, 7}};
-        migraphx::shape s3{migraphx::shape::int32_type, {5, 7}};
-
-        auto l1     = m.add_parameter("a", s1);
-        auto l2     = m.add_parameter("b", s2);
-        auto l3     = m.add_parameter("c", s3);
-        auto beta   = m.add_literal(1);
-        auto output = m.add_parameter("test:#output_0", s3);
-
-        auto pl1   = l1;
-        auto packa = l2;
-        migraphx::instruction_ref pl2{};
-        if(int8_x4)
-        {
-            auto po1 = m.insert_instruction(
-                l1, migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ps1)}}));
-            pl1 = m.add_instruction(
-                migraphx::make_op("gpu::pad", {{"mode", 0}, {"pads", {0, 2, 0, 0}}, {"value", 0}}),
-                l1,
-                po1);
-
-            auto po2 = m.insert_instruction(
-                l2, migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ps2)}}));
-            pl2 = m.insert_instruction(
-                std::next(l2),
-                migraphx::make_op("gpu::pad", {{"mode", 0}, {"pads", {2, 0, 0, 0}}, {"value", 0}}),
-                l2,
-                po2);
-        }
-
-        auto gemm_alloc = m.add_instruction(
-            migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(s3)}}));
-
-        if(int8_x4)
-        {
-            auto alloc = m.add_instruction(
-                migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ps2)}}));
-            packa = m.add_instruction(migraphx::make_op("gpu::int8_gemm_pack_a"), pl2, alloc);
-        }
-
-        auto gemm = m.add_instruction(
-            migraphx::make_op("gpu::quant_gemm",
-                              {{"int8_x4_format", int8_x4},
-                               {"compute_fp32", migraphx::gpu::get_compute_fp32_flag()}}),
-            pl1,
-            packa,
-            gemm_alloc);
-
-        auto beta_broadcast =
-            m.add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", s3.lens()}}), beta);
-        auto mul_alloc = m.add_instruction(
-            migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(s3)}}));
-        auto m3_beta  = m.add_instruction(make_precompile_op("mul"), l3, beta_broadcast, mul_alloc);
-        auto gemm_add = m.add_instruction(make_precompile_op("add"), gemm, m3_beta, output);
-        m.add_return({gemm_add});
-        return m;
-    };
-
-    auto m1  = create_module();
-    auto ctx = migraphx::gpu::context{};
-    run_passes(m1, ctx);
-
-    bool int8_x4 = migraphx::gpu::get_int8_x4_format(ctx);
-    auto m2      = create_optimized_int8_x4(int8_x4);
-
-    EXPECT(m1 == m2);
-}
-
-TEST_CASE(quant_dot_trans_pad)
-{
-    auto create_module = [] {
-        migraphx::module m("test");
-        migraphx::shape s1{migraphx::shape::int8_type, {3, 2, 9, 5}};
-        migraphx::shape s2{migraphx::shape::int8_type, {3, 2, 7, 9}};
-
-        auto l1 = m.add_parameter("a", s1);
-        auto tl1 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l1);
-        auto l2 = m.add_parameter("b", s2);
-        auto tl2 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l2);
-        auto r = migraphx::add_apply_alpha_beta(m, {tl1, tl2}, migraphx::make_op("quant_dot"), 3);
-        m.add_return({r});
-        return m;
-    };
-
-    auto create_optimized_int8_x4 = [](bool int8_x4) {
-        migraphx::module m("test");
-        migraphx::shape s1{migraphx::shape::int8_type, {3, 2, 9, 5}};
-        migraphx::shape ps1{migraphx::shape::int8_type, {3, 2, 5, 12}};
-        migraphx::shape s2{migraphx::shape::int8_type, {3, 2, 7, 9}};
-        migraphx::shape ps2{migraphx::shape::int8_type, {3, 2, 12, 7}};
-        migraphx::shape s3{migraphx::shape::int32_type, {3, 2, 5, 7}};
-
-        auto l1     = m.add_parameter("a", s1);
-        auto l2     = m.add_parameter("b", s2);
-        auto alpha  = m.add_literal(3);
-        auto output = m.add_parameter("test:#output_0", s3);
-
-        auto tl1 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l1);
-        migraphx::shape ts1{migraphx::shape::int8_type, {3, 2, 5, 9}};
-
-        auto tl2 =
-            m.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 1, 3, 2}}}), l2);
-        migraphx::shape ts2{migraphx::shape::int8_type, {3, 2, 9, 7}};
-
-        migraphx::instruction_ref ptb{};
-        if(int8_x4)
-        {
-            ptb = m.add_instruction(
-                migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ps2)}}));
-        }
-        auto pb = tl2;
-        if(int8_x4)
-        {
-            pb = m.add_instruction(
-                migraphx::make_op("gpu::pad", {{"mode", 0}, {"pads", {0, 0, 3, 0, 0, 0, 0, 0}}}),
-                tl2,
-                ptb);
-        }
-
-        auto alpha_broadcast = m.add_instruction(
-            migraphx::make_op("multibroadcast", {{"out_lens", tl1->get_shape().lens()}}), alpha);
-        auto alpha_alloc = m.add_instruction(
-            migraphx::make_op("hip::allocate",
-                              {{"shape",
-                                migraphx::to_value(migraphx::shape(migraphx::shape::int32_type,
-                                                                   tl1->get_shape().lens()))}}));
-
-        // alpha = int32 and tl1 = int8, convert tl1 to int32 for multiplication and then convert
-        // back result to int8
-        auto tl1_convert =
-            m.add_instruction(make_precompile_op(migraphx::make_op(
-                                  "convert", {{"target_type", alpha->get_shape().type()}})),
-                              tl1,
-                              alpha_alloc);
-        auto mul_alloc = m.add_instruction(migraphx::make_op(
-            "hip::allocate", {{"shape", migraphx::to_value(alpha_alloc->get_shape())}}));
-        auto tl1_alpha_int32 =
-            m.add_instruction(make_precompile_op("mul"), alpha_broadcast, tl1_convert, mul_alloc);
-        // convert mul_res to int8
-        auto tl1_alpha_int8_alloc = m.add_instruction(
-            migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ts1)}}));
-
-        migraphx::instruction_ref pta{};
-        if(int8_x4)
-        {
-            pta = m.add_instruction(
-                migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ps1)}}));
-        }
-
-        auto tl1_alpha_int8 = m.add_instruction(
-            make_precompile_op(migraphx::make_op("convert", {{"target_type", ts1.type()}})),
-            tl1_alpha_int32,
-            tl1_alpha_int8_alloc);
-
-        auto pa = tl1_alpha_int8;
-        if(int8_x4)
-        {
-            pa = m.add_instruction(
-                migraphx::make_op("gpu::pad", {{"mode", 0}, {"pads", {0, 0, 0, 3, 0, 0, 0, 0}}}),
-                tl1_alpha_int8,
-                pta);
-        }
-
-        auto packb = pb;
-        if(int8_x4)
-        {
-            auto allocpb = m.add_instruction(
-                migraphx::make_op("hip::allocate", {{"shape", migraphx::to_value(ps2)}}));
-            packb = m.add_instruction(migraphx::make_op("gpu::int8_gemm_pack_a"), pb, allocpb);
-        }
-
-        auto gemm = m.add_instruction(
-            migraphx::make_op("gpu::quant_gemm",
-                              {{"int8_x4_format", int8_x4},
-                               {"compute_fp32", migraphx::gpu::get_compute_fp32_flag()}}),
-            pa,
-            packb,
-            output);
-        m.add_return({gemm});
-
-        return m;
-    };
-
-    auto m1  = create_module();
-    auto ctx = migraphx::gpu::context{};
-    run_passes(m1, ctx);
-
-    bool int8_x4 = migraphx::gpu::get_int8_x4_format(ctx);
-    auto m2      = create_optimized_int8_x4(int8_x4);
-
-    EXPECT(m1 == m2);
-}
-
-int main(int argc, const char* argv[]) { test::run(argc, argv); }

test/gpu/stream_sync.cpp

@@ -64,7 +64,7 @@ int main() {}
 
 migraphx::src_file make_src_file(const std::string& name, const std::string& content)
 {
-    return {name, std::make_pair(content.data(), content.data() + content.size())};
+    return {name, content};
 }
 
 hip_stream_ptr get_stream()

test/include/test.hpp

@@ -339,6 +339,8 @@ inline std::ostream& operator<<(std::ostream& os, const color& c)
     static const bool use_color = isatty(STDOUT_FILENO) != 0;
     if(use_color)
         return os << "\033[" << static_cast<std::size_t>(c) << "m";
+#else
+    (void)c;
 #endif
     return os;
 }

test/jit.cpp

@@ -48,9 +48,7 @@ compile_function(const std::string& src, const std::string& flags, const std::st
     migraphx::src_compiler compiler;
     compiler.flags  = flags + "-std=c++14 -fPIC -shared";
     compiler.output = "libsimple.so";
-    migraphx::src_file f;
-    f.path     = "main.cpp";
-    f.content  = std::make_pair(src.data(), src.data() + src.size());
+    migraphx::src_file f{"main.cpp", src};
     auto image = compiler.compile({f});
     return migraphx::dynamic_loader{image}.get_function<F>(fname);
 }

test/memory_coloring_test.cpp

@@ -55,7 +55,7 @@ struct allocate
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 

test/msgpack.cpp

@@ -97,9 +97,12 @@ TEST_CASE(test_msgpack_bool)
 
 TEST_CASE(test_msgpack_float)
 {
-    migraphx::value v = 3.0;
+    // changed all double values in this code to not end with .0 because on msgpack for Windows if
+    // input type is double and ends with .0 it could be converted to uint64_t or int64_t and the
+    // goal of these functions is to test double without conversions
+    migraphx::value v = 3.01;
     auto buffer       = migraphx::to_msgpack(v);
-    EXPECT(buffer == msgpack_buffer(3.0));
+    EXPECT(buffer == msgpack_buffer(3.01));
     EXPECT(migraphx::from_msgpack(buffer) == v);
 }
 
@@ -129,10 +132,10 @@ TEST_CASE(test_msgpack_empty_array)
 
 TEST_CASE(test_msgpack_object)
 {
-    migraphx::value v = {{"one", 1.0}, {"three", 3.0}, {"two", 2.0}};
+    migraphx::value v = {{"one", 1.01}, {"three", 3.01}, {"two", 2.01}};
     auto buffer       = migraphx::to_msgpack(v);
     EXPECT(buffer == msgpack_buffer(std::map<std::string, double>{
-                         {"one", 1.0}, {"three", 3.0}, {"two", 2.0}}));
+                         {"one", 1.01}, {"three", 3.01}, {"two", 2.01}}));
     EXPECT(migraphx::from_msgpack(buffer) == v);
 }
 
@@ -157,17 +160,17 @@ struct foo
 
 TEST_CASE(test_msgpack_object_class)
 {
-    migraphx::value v = {{"a", 1.0}, {"b", "abc"}};
+    migraphx::value v = {{"a", 1.01}, {"b", "abc"}};
     auto buffer       = migraphx::to_msgpack(v);
-    EXPECT(buffer == msgpack_buffer(foo{1.0, "abc"}));
+    EXPECT(buffer == msgpack_buffer(foo{1.01, "abc"}));
     EXPECT(migraphx::from_msgpack(buffer) == v);
 }
 
 TEST_CASE(test_msgpack_array_class)
 {
-    migraphx::value v = {{{"a", 1.0}, {"b", "abc"}}, {{"a", 3.0}, {"b", "xyz"}}};
+    migraphx::value v = {{{"a", 1.01}, {"b", "abc"}}, {{"a", 3.01}, {"b", "xyz"}}};
     auto buffer       = migraphx::to_msgpack(v);
-    EXPECT(buffer == msgpack_buffer(std::vector<foo>{foo{1.0, "abc"}, foo{3.0, "xyz"}}));
+    EXPECT(buffer == msgpack_buffer(std::vector<foo>{foo{1.01, "abc"}, foo{3.01, "xyz"}}));
     EXPECT(migraphx::from_msgpack(buffer) == v);
 }
 

test/multi_target/multitarget_test.cpp

@@ -37,7 +37,6 @@
 #include <migraphx/make_op.hpp>
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/functional.hpp>
-#include <basic_ops.hpp>
 #include <migraphx/compile_options.hpp>
 #include <migraphx/register_target.hpp>
 #include <migraphx/generate.hpp>

test/normalize_ops_test.cpp

@@ -57,7 +57,7 @@ struct normalize_test_op
                                const migraphx::shape& output_shape,
                                const std::vector<migraphx::argument>&) const
     {
-        return {output_shape};
+        return migraphx::argument{output_shape};
     }
 };
 

test/onnx/.onnxrt-commit

-6d7bc2a097a1a08541cd0d4628831c79ab8092d5
+b7b8b5b2ce80edb33990c7ae0fedac6ae3c623f4

test/onnx/gen_onnx.py

@@ -149,6 +149,21 @@ def argmax_test():
     return ([node], [x], [y])
 
 
+@onnx_test()
+def argmax_select_last_index_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4, 5, 6])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4, 6])
+
+    node = onnx.helper.make_node('ArgMax',
+                                 inputs=['x'],
+                                 outputs=['y'],
+                                 axis=2,
+                                 keepdims=0,
+                                 select_last_index=1)
+
+    return ([node], [x], [y])
+
+
 @onnx_test()
 def argmax_dyn_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [None, 4, 5, 6])
@@ -177,6 +192,21 @@ def argmin_test():
     return ([node], [x], [y])
 
 
+@onnx_test()
+def argmin_select_last_index_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4, 5, 6])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4, 5])
+
+    node = onnx.helper.make_node('ArgMin',
+                                 inputs=['x'],
+                                 outputs=['y'],
+                                 axis=3,
+                                 keepdims=0,
+                                 select_last_index=1)
+
+    return ([node], [x], [y])
+
+
 @onnx_test()
 def asin_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10])
@@ -2722,6 +2752,119 @@ def group_conv_test():
     return ([node], [x, y], [z])
 
 
+def group_norm_test(x_dims,
+                    scale_dims,
+                    bias_dims,
+                    y_dims,
+                    num_groups,
+                    eps_value=1e-5,
+                    dtype=TensorProto.FLOAT):
+    x = helper.make_tensor_value_info('x', dtype, x_dims)
+    scale = helper.make_tensor_value_info('scale', dtype, scale_dims)
+    bias = helper.make_tensor_value_info('bias', dtype, bias_dims)
+    y = helper.make_tensor_value_info('y', dtype, y_dims)
+
+    node = onnx.helper.make_node('GroupNormalization',
+                                 inputs=['x', 'scale', 'bias'],
+                                 outputs=['y'],
+                                 num_groups=num_groups,
+                                 epsilon=eps_value)
+
+    return ([node], [x, scale, bias], [y])
+
+
+@onnx_test()
+def group_norm_3d_test():
+    return group_norm_test([1, 4, 2], [2], [2], [1, 4, 2], 2)
+
+
+@onnx_test()
+def group_norm_3d_half_test():
+    return group_norm_test([1, 4, 2], [2], [2], [1, 4, 2],
+                           2,
+                           dtype=TensorProto.FLOAT16)
+
+
+@onnx_test()
+def group_norm_4d_test():
+    return group_norm_test([1, 4, 3, 3], [2], [2], [1, 4, 3, 3], 2)
+
+
+@onnx_test()
+def group_norm_4d_half_test():
+    return group_norm_test([1, 4, 3, 3], [2], [2], [1, 4, 3, 3],
+                           2,
+                           dtype=TensorProto.FLOAT16)
+
+
+@onnx_test()
+def group_norm_5d_test():
+    return group_norm_test([3, 3, 3, 3, 3], [1], [1], [3, 3, 3, 3, 3], 1)
+
+
+@onnx_test()
+def group_norm_5d_half_test():
+    return group_norm_test([3, 3, 3, 3, 3], [1], [1], [3, 3, 3, 3, 3],
+                           1,
+                           dtype=TensorProto.FLOAT16)
+
+
+@onnx_test()
+def group_norm_small_eps_half_test():
+    return group_norm_test([1, 4, 2], [2], [2], [1, 4, 2],
+                           2,
+                           eps_value=1e-12,
+                           dtype=TensorProto.FLOAT16)
+
+
+@onnx_test()
+def group_norm_invalid_num_groups_error_test():
+    return group_norm_test([1, 4, 3, 3], [2], [2], [1, 4, 3, 3], 3)
+
+
+@onnx_test()
+def group_norm_missing_attribute_error_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 4])
+    scale = helper.make_tensor_value_info('scale', TensorProto.FLOAT, [2])
+    bias = helper.make_tensor_value_info('bias', TensorProto.FLOAT, [2])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1, 4])
+
+    node = onnx.helper.make_node('GroupNormalization',
+                                 inputs=['x', 'scale', 'bias'],
+                                 outputs=['y'])
+
+    return ([node], [x, scale, bias], [y])
+
+
+@onnx_test()
+def group_norm_invalid_input_count_error_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 4, 3, 3])
+    scale = helper.make_tensor_value_info('scale', TensorProto.FLOAT, [2])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1, 4, 3, 3])
+
+    node = onnx.helper.make_node('GroupNormalization',
+                                 inputs=['x', 'scale'],
+                                 outputs=['y'],
+                                 num_groups=2)
+
+    return ([node], [x, scale], [y])
+
+
+@onnx_test()
+def group_norm_invalid_input_shape_error_test():
+    return group_norm_test([1, 4], [2], [2], [1, 4], 2)
+
+
+@onnx_test()
+def group_norm_invalid_scale_shape_test():
+    return group_norm_test([1, 4, 3, 3], [1], [2], [1, 4, 3, 3], 2)
+
+
+@onnx_test()
+def group_norm_invalid_bias_shape_test():
+    return group_norm_test([1, 4, 3, 3], [2], [3], [1, 4, 3, 3], 2)
+
+
 @onnx_test()
 def hardsigmoid_default_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 3, 4, 5])
@@ -3715,6 +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])
@@ -3804,6 +4005,110 @@ def layernorm_test():
              bias_add], [x, scale, bias], [y], [pow_tensor, epsilon_tensor])
 
 
+def make_layer_norm(shape, axis, dtype=TensorProto.FLOAT):
+    norm_axis = axis + len(shape) if axis < 0 else axis
+    x = helper.make_tensor_value_info('x', dtype, shape)
+    scale = helper.make_tensor_value_info('scale', dtype, shape[norm_axis:])
+    bias = helper.make_tensor_value_info('bias', dtype, shape[norm_axis:])
+    y = helper.make_tensor_value_info('y', dtype, shape)
+
+    node = onnx.helper.make_node('LayerNormalization',
+                                 inputs=['x', 'scale', 'bias'],
+                                 outputs=['y'],
+                                 axis=axis)
+
+    return ([node], [x, scale, bias], [y])
+
+
+@onnx_test()
+def layer_norm_invalid_shape_error_test():
+    return make_layer_norm([3], 0)
+
+
+@onnx_test()
+def layer_norm_2d_axis_zero_test():
+    return make_layer_norm([3, 4], 0)
+
+
+@onnx_test()
+def layer_norm_2d_axis_one_test():
+    return make_layer_norm([3, 4], 1)
+
+
+@onnx_test()
+def layer_norm_2d_axis_minus_one_test():
+    return make_layer_norm([3, 4], -1)
+
+
+@onnx_test()
+def layer_norm_3d_test():
+    return make_layer_norm([1, 4, 2], -1)
+
+
+@onnx_test()
+def layer_norm_3d_half_test():
+    return make_layer_norm([1, 4, 2], -1, TensorProto.FLOAT16)
+
+
+@onnx_test()
+def layer_norm_4d_test():
+    return make_layer_norm([3, 3, 3, 3], -1)
+
+
+@onnx_test()
+def layer_norm_4d_half_test():
+    return make_layer_norm([3, 3, 3, 3], -1, TensorProto.FLOAT16)
+
+
+@onnx_test()
+def layer_norm_invalid_axis_error_test():
+    return make_layer_norm([1, 4, 2], 1000)
+
+
+@onnx_test()
+def layer_norm_invalid_minus_axis_error_test():
+    return make_layer_norm([1, 4, 2], -1000)
+
+
+@onnx_test()
+def layer_norm_invalid_input_count_error_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 2])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1, 2])
+
+    node = onnx.helper.make_node('LayerNormalization',
+                                 inputs=['x'],
+                                 outputs=['y'])
+
+    return ([node], [x], [y])
+
+
+@onnx_test()
+def layer_norm_without_bias_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 2])
+    scale = helper.make_tensor_value_info('scale', TensorProto.FLOAT, [2])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1, 2])
+
+    node = onnx.helper.make_node('LayerNormalization',
+                                 inputs=['x', 'scale'],
+                                 outputs=['y'])
+
+    return ([node], [x, scale], [y])
+
+
+@onnx_test()
+def layer_norm_small_eps_half_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT16, [1, 2])
+    scale = helper.make_tensor_value_info('scale', TensorProto.FLOAT16, [2])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT16, [1, 2])
+
+    node = onnx.helper.make_node('LayerNormalization',
+                                 inputs=['x', 'scale'],
+                                 outputs=['y'],
+                                 epsilon=1e-12)
+
+    return ([node], [x, scale], [y])
+
+
 @onnx_test()
 def leaky_relu_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [3])
@@ -4029,6 +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])
@@ -4464,6 +4813,77 @@ def mean_integral_test():
     return ([node], data, [mean])
 
 
+def mvn_default_axes_test_base(dims, type=TensorProto.FLOAT):
+    data = helper.make_tensor_value_info("data", type, dims)
+    out = helper.make_tensor_value_info("out", type, dims)
+    node = helper.make_node("MeanVarianceNormalization",
+                            inputs=["data"],
+                            outputs=["out"])
+
+    return ([node], [data], [out])
+
+
+@onnx_test()
+def mvn_default_axes_test():
+    return mvn_default_axes_test_base([2, 2, 2, 2])
+
+
+@onnx_test()
+def mvn_default_axes_fp16_test():
+    return mvn_default_axes_test_base([2, 2, 2, 2], TensorProto.FLOAT16)
+
+
+@onnx_test()
+def mvn_default_axes_rank_too_small_test():
+    return mvn_default_axes_test_base([2, 2, 2])
+
+
+@onnx_test()
+def mvn_default_axes_rank_too_big_test():
+    return mvn_default_axes_test_base([2, 2, 2, 2, 2])
+
+
+def mvn_n_rank_test_base(axes, dims, type=TensorProto.FLOAT):
+    data = helper.make_tensor_value_info("data", type, dims)
+    out = helper.make_tensor_value_info("out", type, dims)
+    node = helper.make_node("MeanVarianceNormalization",
+                            inputs=["data"],
+                            outputs=["out"],
+                            axes=axes)
+
+    return ([node], [data], [out])
+
+
+@onnx_test()
+def mvn_rank_2_test():
+    return mvn_n_rank_test_base([1], [2, 2])
+
+
+@onnx_test()
+def mvn_rank_2_fp16_test():
+    return mvn_n_rank_test_base([1], [2, 2], TensorProto.FLOAT16)
+
+
+@onnx_test()
+def mvn_rank_3_test():
+    return mvn_n_rank_test_base([0, 1], [2, 2, 2])
+
+
+@onnx_test()
+def mvn_rank_3_fp16_test():
+    return mvn_n_rank_test_base([0, 1], [2, 2, 2], TensorProto.FLOAT16)
+
+
+@onnx_test()
+def mvn_axes_rank_too_small_test():
+    return mvn_n_rank_test_base([0, 1, 2], [2, 2, 2])
+
+
+@onnx_test()
+def mvn_axes_rank_too_big_test():
+    return mvn_n_rank_test_base([0], [2, 2, 2])
+
+
 @onnx_test()
 def min_test():
     a = helper.make_tensor_value_info('0', TensorProto.FLOAT, [3])
@@ -4565,9 +4985,9 @@ def mod_test_fmod_different_dtypes():
 
 @onnx_test()
 def multinomial_test():
-    sample_size = 10
-    seed = 0.0
-    input = helper.make_tensor_value_info("input", TensorProto.FLOAT, [1, 10])
+    sample_size = 13
+    seed = 0.
+    input = helper.make_tensor_value_info("input", TensorProto.FLOAT, [3, 10])
     output = helper.make_tensor_value_info("output", TensorProto.INT32,
                                            [1, 10])
 
@@ -4581,9 +5001,47 @@ def multinomial_test():
 
 
 @onnx_test()
-def multinomial_generated_seed_test():
-    sample_size = 10
-    input = helper.make_tensor_value_info("input", TensorProto.FLOAT, [1, 10])
+def multinomial_dyn_test():
+    sample_size = 100000
+    seed = 1.3
+    categories = 5
+    input = helper.make_tensor_value_info("input", TensorProto.FLOAT,
+                                          [None, categories])
+    output = helper.make_tensor_value_info("output", TensorProto.FLOAT,
+                                           [None, categories])
+
+    node = onnx.helper.make_node(
+        'Multinomial',
+        inputs=['input'],
+        sample_size=sample_size,
+        dtype=1,  # shape::float_type
+        seed=seed,
+        outputs=['output'])
+
+    return ([node], [input], [output])
+
+
+@onnx_test()
+def multinomial_autoseed_dyn_test():
+    # If seed attribute is not given, device should auto generate one at runtime
+    sample_size = 12
+    input = helper.make_tensor_value_info("input", TensorProto.FLOAT,
+                                          [None, 10])
+    output = helper.make_tensor_value_info("output", TensorProto.INT32,
+                                           [None, 10])
+
+    node = onnx.helper.make_node('Multinomial',
+                                 inputs=['input'],
+                                 sample_size=sample_size,
+                                 outputs=['output'])
+
+    return ([node], [input], [output])
+
+
+@onnx_test()
+def multinomial_generated_seed_test():
+    sample_size = 10
+    input = helper.make_tensor_value_info("input", TensorProto.FLOAT, [1, 10])
     output = helper.make_tensor_value_info("output", TensorProto.INT32,
                                            [1, 10])
 
@@ -4890,6 +5348,32 @@ def pad_test():
     return ([node], [x], [y])
 
 
+@onnx_test()
+def pad_asym_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 4, 5])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1, 6, 4, 12])
+
+    node = onnx.helper.make_node('Pad',
+                                 inputs=['0'],
+                                 pads=[0, 1, 0, 3, 0, 2, 0, 4],
+                                 outputs=['1'])
+
+    return ([node], [x], [y])
+
+
+@onnx_test()
+def pad_asym_invalid_pads_error_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 4, 5])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1, 6, 4, 12])
+
+    node = onnx.helper.make_node('Pad',
+                                 inputs=['0'],
+                                 pads=[0, 1, 0, 3, 0, 2],
+                                 outputs=['1'])
+
+    return ([node], [x], [y])
+
+
 @onnx_test()
 def pad_3arg_test():
     values = np.array([1])
@@ -4922,6 +5406,129 @@ def pad_3arg_test():
     return ([arg_val, arg_pad, node], [x], [y])
 
 
+@onnx_test()
+def pad_4arg_axes_test():
+    values = np.array([1])
+    val_tensor = helper.make_tensor(name='val',
+                                    data_type=TensorProto.FLOAT,
+                                    dims=values.reshape(()).shape,
+                                    vals=values.astype(float))
+    arg_val = onnx.helper.make_node('Constant',
+                                    inputs=[],
+                                    outputs=['arg_val'],
+                                    value=val_tensor)
+
+    sizes = np.array([1, 3, 2, 4])
+    pad_tensor = helper.make_tensor(name='pad_size',
+                                    data_type=TensorProto.INT32,
+                                    dims=sizes.shape,
+                                    vals=sizes.astype(int))
+    arg_pad = onnx.helper.make_node('Constant',
+                                    inputs=[],
+                                    outputs=['arg_pad'],
+                                    value=pad_tensor)
+
+    axes = np.array([1, 3])
+    axes_tensor = helper.make_tensor(name='pad_axes',
+                                     data_type=TensorProto.INT32,
+                                     dims=axes.shape,
+                                     vals=axes.astype(int))
+    arg_axes = onnx.helper.make_node('Constant',
+                                     inputs=[],
+                                     outputs=['arg_axes'],
+                                     value=axes_tensor)
+
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 4, 5])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1, 6, 4, 12])
+
+    node = onnx.helper.make_node(
+        'Pad', inputs=['0', 'arg_pad', 'arg_val', 'arg_axes'], outputs=['1'])
+
+    return ([arg_axes, arg_val, arg_pad, node], [x], [y])
+
+
+@onnx_test()
+def pad_4arg_invalid_axes_error_test():
+    values = np.array([1])
+    val_tensor = helper.make_tensor(name='val',
+                                    data_type=TensorProto.FLOAT,
+                                    dims=values.reshape(()).shape,
+                                    vals=values.astype(float))
+    arg_val = onnx.helper.make_node('Constant',
+                                    inputs=[],
+                                    outputs=['arg_val'],
+                                    value=val_tensor)
+
+    sizes = np.array([1, 3, 2, 4])
+    pad_tensor = helper.make_tensor(name='pad_size',
+                                    data_type=TensorProto.INT32,
+                                    dims=sizes.shape,
+                                    vals=sizes.astype(int))
+    arg_pad = onnx.helper.make_node('Constant',
+                                    inputs=[],
+                                    outputs=['arg_pad'],
+                                    value=pad_tensor)
+
+    axes = np.array([1, 2, 3])
+    axes_tensor = helper.make_tensor(name='pad_axes',
+                                     data_type=TensorProto.INT32,
+                                     dims=axes.shape,
+                                     vals=axes.astype(int))
+    arg_axes = onnx.helper.make_node('Constant',
+                                     inputs=[],
+                                     outputs=['arg_axes'],
+                                     value=axes_tensor)
+
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 4, 5])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1, 6, 4, 12])
+
+    node = onnx.helper.make_node(
+        'Pad', inputs=['0', 'arg_pad', 'arg_val', 'arg_axes'], outputs=['1'])
+
+    return ([arg_axes, arg_val, arg_pad, node], [x], [y])
+
+
+@onnx_test()
+def pad_4arg_neg_axes_test():
+    values = np.array([1])
+    val_tensor = helper.make_tensor(name='val',
+                                    data_type=TensorProto.FLOAT,
+                                    dims=values.reshape(()).shape,
+                                    vals=values.astype(float))
+    arg_val = onnx.helper.make_node('Constant',
+                                    inputs=[],
+                                    outputs=['arg_val'],
+                                    value=val_tensor)
+
+    sizes = np.array([1, 3, 2, 4])
+    pad_tensor = helper.make_tensor(name='pad_size',
+                                    data_type=TensorProto.INT32,
+                                    dims=sizes.shape,
+                                    vals=sizes.astype(int))
+    arg_pad = onnx.helper.make_node('Constant',
+                                    inputs=[],
+                                    outputs=['arg_pad'],
+                                    value=pad_tensor)
+
+    axes = np.array([-3, -1])
+    axes_tensor = helper.make_tensor(name='pad_axes',
+                                     data_type=TensorProto.INT32,
+                                     dims=axes.shape,
+                                     vals=axes.astype(int))
+    arg_axes = onnx.helper.make_node('Constant',
+                                     inputs=[],
+                                     outputs=['arg_axes'],
+                                     value=axes_tensor)
+
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 4, 5])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1, 6, 4, 12])
+
+    node = onnx.helper.make_node(
+        'Pad', inputs=['0', 'arg_pad', 'arg_val', 'arg_axes'], outputs=['1'])
+
+    return ([arg_axes, arg_val, arg_pad, node], [x], [y])
+
+
 @onnx_test()
 def pad_reflect_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [2, 2])
@@ -4945,6 +5552,39 @@ def pad_reflect_test():
     return ([arg_pad, node], [x], [y])
 
 
+@onnx_test()
+def pad_reflect_with_axes_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [2, 2])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [2, 5])
+
+    sizes = np.array([2, 1])
+    pad_tensor = helper.make_tensor(name='pad_size',
+                                    data_type=TensorProto.INT32,
+                                    dims=sizes.shape,
+                                    vals=sizes.astype(int))
+    arg_pad = onnx.helper.make_node('Constant',
+                                    inputs=[],
+                                    outputs=['arg_pad'],
+                                    value=pad_tensor)
+
+    axes = np.array([1])
+    axes_tensor = helper.make_tensor(name='pad_axes',
+                                     data_type=TensorProto.INT32,
+                                     dims=axes.shape,
+                                     vals=axes.astype(int))
+    arg_axes = onnx.helper.make_node('Constant',
+                                     inputs=[],
+                                     outputs=['arg_axes'],
+                                     value=axes_tensor)
+
+    node = onnx.helper.make_node('Pad',
+                                 mode='reflect',
+                                 inputs=['0', 'arg_pad', 'arg_axes'],
+                                 outputs=['1'])
+
+    return ([arg_axes, arg_pad, node], [x], [y])
+
+
 @onnx_test()
 def pad_reflect_multiaxis_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [2, 3])
@@ -5096,6 +5736,278 @@ def prelu_brcst_test():
     return ([node], [arg0, arg1], [arg_out])
 
 
+@onnx_test()
+def qlinearadd_test():
+    a = helper.make_tensor_value_info('A', TensorProto.UINT8, [64])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.UINT8, [], [0])
+
+    b = helper.make_tensor_value_info('B', TensorProto.UINT8, [64])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.UINT8, [],
+                                   [128])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.UINT8, [], [64])
+
+    c = helper.make_tensor_value_info('C', TensorProto.UINT8, [64])
+
+    node = onnx.helper.make_node(
+        'QLinearAdd',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
+@onnx_test()
+def qlinearadd_bcast_test():
+    a = helper.make_tensor_value_info('A', TensorProto.INT8, [64])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.INT8, [], [0])
+
+    b = helper.make_tensor_value_info('B', TensorProto.INT8, [1, 1, 64])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.INT8, [], [32])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.INT8, [], [-64])
+
+    c = helper.make_tensor_value_info('C', TensorProto.INT8, [1, 1, 64])
+
+    node = onnx.helper.make_node(
+        'QLinearAdd',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
+@onnx_test()
+def qlinearconv_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__QLinearConv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 7, 7])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [], [0.00369204697])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [132])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 1, 1], [0])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [0.00172794575])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [255])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.00162681262])
+    zero_pt_y = helper.make_tensor('7', TensorProto.UINT8, [], [123])
+
+    out = helper.make_tensor_value_info('out', TensorProto.UINT8, [1, 1, 7, 7])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_pad_1_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 3, 3],
+                            [1, 1, 1, 1, 1, 1, 1, 1, 1])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [1.0])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.UINT8, [], [0])
+
+    out = helper.make_tensor_value_info('out', TensorProto.UINT8, [1, 1, 5, 5])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[1, 1, 1, 1],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_pad_0_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor('3', TensorProto.UINT8, [1, 1, 3, 3],
+                            [1, 1, 1, 1, 1, 1, 1, 1, 1])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [], [1.0])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [], [0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.INT8, [], [-128])
+
+    out = helper.make_tensor_value_info('out', TensorProto.INT8, [1, 1, 3, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[0, 0, 0, 0],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearconv_scale_1D_test():
+    # https://xadupre.github.io/draft/onnx/onnx_doc_folder/onnx__Conv.html
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 1, 5, 5])
+    sc_x = helper.make_tensor('1', TensorProto.FLOAT, [],
+                              [0.09411764705882353])
+    zero_pt_x = helper.make_tensor('2', TensorProto.UINT8, [], [0])
+
+    wt = helper.make_tensor(
+        '3', TensorProto.UINT8, [2, 1, 3, 3],
+        [1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2])
+    sc_wt = helper.make_tensor('4', TensorProto.FLOAT, [2], [1.0, 0.5])
+    zero_pt_wt = helper.make_tensor('5', TensorProto.UINT8, [2], [0, 0])
+
+    sc_y = helper.make_tensor('6', TensorProto.FLOAT, [], [0.6352941176470588])
+    zero_pt_y = helper.make_tensor('7', TensorProto.INT8, [], [-128])
+
+    out = helper.make_tensor_value_info('out', TensorProto.INT8, [1, 2, 3, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearConv',
+        inputs=['X', '1', '2', '3', '4', '5', '6', '7'],
+        outputs=['out'],
+        pads=[0, 0, 0, 0],
+    )
+    return ([node], [x], [out],
+            [sc_x, zero_pt_x, wt, sc_wt, zero_pt_wt, sc_y, zero_pt_y])
+
+
+@onnx_test()
+def qlinearglobalavgpool_test():
+    x = helper.make_tensor_value_info('X', TensorProto.UINT8, [1, 3, 4, 4])
+
+    sc_x = helper.make_tensor('X_scale', TensorProto.FLOAT, [], [0.05])
+    z_pt_x = helper.make_tensor('X_zero_point', TensorProto.UINT8, [], [128])
+
+    y = helper.make_tensor_value_info('Y', TensorProto.UINT8, [1, 3, 1, 1])
+
+    sc_y = helper.make_tensor('Y_scale', TensorProto.FLOAT, [], [0.025])
+    z_pt_y = helper.make_tensor('Y_zero_point', TensorProto.UINT8, [], [64])
+
+    n = onnx.helper.make_node(
+        'QLinearGlobalAveragePool',
+        inputs=['X', 'X_scale', 'X_zero_point', 'Y_scale', 'Y_zero_point'],
+        outputs=['Y'],
+        channels_last=0,
+    )
+
+    return ([n], [x], [y], [sc_x, z_pt_x, sc_y, z_pt_y])
+
+
+def qlinearmatmul_1D_test():
+    a = helper.make_tensor_value_info('A', TensorProto.UINT8, [8])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.UINT8, [], [0])
+
+    b = helper.make_tensor_value_info('B', TensorProto.UINT8, [8])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.UINT8, [],
+                                   [128])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.UINT8, [], [64])
+
+    c = helper.make_tensor_value_info('C', TensorProto.UINT8, [1])
+
+    node = onnx.helper.make_node(
+        'QLinearMatMul',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
+@onnx_test()
+def qlinearmatmul_2D_test():
+    a = helper.make_tensor_value_info('A', TensorProto.UINT8, [1, 8])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.UINT8, [], [0])
+
+    b = helper.make_tensor_value_info('B', TensorProto.UINT8, [8, 1])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.UINT8, [],
+                                   [128])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.05])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.UINT8, [], [64])
+
+    c = helper.make_tensor_value_info('C', TensorProto.UINT8, [1, 1])
+
+    node = onnx.helper.make_node(
+        'QLinearMatMul',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
+@onnx_test()
+def qlinearmatmul_3D_test():
+    a = helper.make_tensor_value_info('A', TensorProto.UINT8, [2, 2, 4])
+    sc_a = helper.make_tensor('A_scale', TensorProto.FLOAT, [], [0.0066])
+    zero_pt_a = helper.make_tensor('A_zero_point', TensorProto.UINT8, [],
+                                   [113])
+
+    b = helper.make_tensor_value_info('B', TensorProto.UINT8, [2, 4, 3])
+    sc_b = helper.make_tensor('B_scale', TensorProto.FLOAT, [], [0.00705])
+    zero_pt_b = helper.make_tensor('B_zero_point', TensorProto.UINT8, [],
+                                   [114])
+
+    sc_c = helper.make_tensor('C_scale', TensorProto.FLOAT, [], [0.0107])
+    zero_pt_c = helper.make_tensor('C_zero_point', TensorProto.UINT8, [],
+                                   [118])
+
+    c = helper.make_tensor_value_info('C', TensorProto.UINT8, [2, 2, 3])
+
+    node = onnx.helper.make_node(
+        'QLinearMatMul',
+        inputs=[
+            'A', 'A_scale', 'A_zero_point', 'B', 'B_scale', 'B_zero_point',
+            'C_scale', 'C_zero_point'
+        ],
+        outputs=['C'],
+    )
+    return ([node], [a, b], [c],
+            [sc_a, zero_pt_a, sc_b, zero_pt_b, sc_c, zero_pt_c])
+
+
 @onnx_test()
 def quantizelinear_test():
     arg0 = helper.make_tensor_value_info('0', TensorProto.FLOAT, [5])
@@ -5793,6 +6705,24 @@ def reshape_non_standard_test():
     return ([trans, res], [x], [y])
 
 
+@onnx_test()
+def reshape_variable_input_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [4, 2, 3])
+    x_shape = helper.make_tensor_value_info('1', TensorProto.INT64, [2])
+    y = helper.make_tensor_value_info('2', TensorProto.FLOAT, [3, 8])
+    node = onnx.helper.make_node('Reshape', inputs=['0', '1'], outputs=['2'])
+    return ([node], [x, x_shape], [y])
+
+
+@onnx_test()
+def reshape_variable_input_dyn_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [None, 2, 3])
+    x_shape = helper.make_tensor_value_info('1', TensorProto.INT64, [2])
+    y = helper.make_tensor_value_info('2', TensorProto.FLOAT, [None, 6])
+    node = onnx.helper.make_node('Reshape', inputs=['0', '1'], outputs=['2'])
+    return ([node], [x, x_shape], [y])
+
+
 @onnx_test()
 def resize_downsample_f_test():
     scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32)
@@ -6157,6 +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])
@@ -6446,6 +7386,101 @@ def shape_gather_test():
     return ([node_const, node_shape, node_gather], [x], [z])
 
 
+@onnx_test()
+def shrink_hard_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [5])
+
+    node = onnx.helper.make_node(
+        "Shrink",
+        inputs=["x"],
+        outputs=["y"],
+        lambd=1.5,
+    )
+
+    return ([node], [x], [y])
+
+
+@onnx_test()
+def shrink_soft_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [5])
+
+    node = onnx.helper.make_node(
+        "Shrink",
+        inputs=["x"],
+        outputs=["y"],
+        lambd=1.5,
+        bias=1.5,
+    )
+
+    return ([node], [x], [y])
+
+
+@onnx_test()
+def shrink_verify_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT16, [5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT16, [5])
+
+    node = onnx.helper.make_node(
+        "Shrink",
+        inputs=["x"],
+        outputs=["y"],
+        lambd=-5.0,
+        bias=1.0,
+    )
+
+    return ([node], [x], [y])
+
+
+@onnx_test()
+def shrink_verify2_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT16, [5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT16, [5])
+
+    node = onnx.helper.make_node(
+        "Shrink",
+        inputs=["x"],
+        outputs=["y"],
+        lambd=-6.0,
+        bias=5.0,
+    )
+
+    return ([node], [x], [y])
+
+
+@onnx_test()
+def shrink_int8_test():
+    x = helper.make_tensor_value_info('x', TensorProto.INT8, [3, 3])
+    y = helper.make_tensor_value_info('y', TensorProto.INT8, [3, 3])
+
+    node = onnx.helper.make_node(
+        "Shrink",
+        inputs=["x"],
+        outputs=["y"],
+        lambd=1.5,
+        bias=1.5,
+    )
+
+    return ([node], [x], [y])
+
+
+@onnx_test()
+def shrink_uint8_test():
+    x = helper.make_tensor_value_info('x', TensorProto.UINT8, [3, 3])
+    y = helper.make_tensor_value_info('y', TensorProto.UINT8, [3, 3])
+
+    node = onnx.helper.make_node(
+        "Shrink",
+        inputs=["x"],
+        outputs=["y"],
+        lambd=5.0,
+        bias=-4.5,
+    )
+
+    return ([node], [x], [y])
+
+
 @onnx_test()
 def sign_test():
     x = helper.make_tensor_value_info('x', TensorProto.DOUBLE, [10, 5])
@@ -6981,6 +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])
@@ -6994,9 +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(
@@ -7157,6 +8218,42 @@ def split_test_no_attribute():
     return ([const_node, node], [x], [y1, y2, y3, y4])
 
 
+@onnx_test()
+def split_test_uneven():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [12, 15])
+    y1 = helper.make_tensor_value_info('y1', TensorProto.FLOAT, [3, 15])
+    y2 = helper.make_tensor_value_info('y2', TensorProto.FLOAT, [3, 15])
+    y3 = helper.make_tensor_value_info('y3', TensorProto.FLOAT, [3, 15])
+    y4 = helper.make_tensor_value_info('y4', TensorProto.FLOAT, [3, 15])
+    y5 = helper.make_tensor_value_info('y5', TensorProto.FLOAT, [0, 15])
+
+    node = onnx.helper.make_node(
+        'Split',
+        inputs=['x'],
+        outputs=['y1', 'y2', 'y3', 'y4', 'y5'],
+    )
+
+    return ([node], [x], [y1, y2, y3, y4, y5])
+
+
+@onnx_test()
+def split_test_uneven_num_outputs():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [11, 15])
+    y1 = helper.make_tensor_value_info('y1', TensorProto.FLOAT, [3, 15])
+    y2 = helper.make_tensor_value_info('y2', TensorProto.FLOAT, [3, 15])
+    y3 = helper.make_tensor_value_info('y3', TensorProto.FLOAT, [3, 15])
+    y4 = helper.make_tensor_value_info('y4', TensorProto.FLOAT, [2, 15])
+
+    node = onnx.helper.make_node(
+        'Split',
+        inputs=['x'],
+        outputs=['y1', 'y2', 'y3', 'y4'],
+        num_outputs=4,
+    )
+
+    return ([node], [x], [y1, y2, y3, y4])
+
+
 @onnx_test()
 def split_test_no_attribute_invalid_split():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [300, 15])
@@ -7216,6 +8313,24 @@ def split_test_no_attribute_invalid_input_split():
     return ([node], [x], [y1, y2, y3])
 
 
+@onnx_test()
+def split_test_invalid_num_outputs():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [11, 15])
+    y1 = helper.make_tensor_value_info('y1', TensorProto.FLOAT, [3, 15])
+    y2 = helper.make_tensor_value_info('y2', TensorProto.FLOAT, [3, 15])
+    y3 = helper.make_tensor_value_info('y3', TensorProto.FLOAT, [3, 15])
+    y4 = helper.make_tensor_value_info('y4', TensorProto.FLOAT, [2, 15])
+
+    node = onnx.helper.make_node(
+        'Split',
+        inputs=['x'],
+        outputs=['y1', 'y2', 'y3', 'y4'],
+        num_outputs=5,
+    )
+
+    return ([node], [x], [y1, y2, y3, y4])
+
+
 @onnx_test()
 def sqrt_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [10, 15])
@@ -7718,7 +8833,7 @@ def transpose_gather_test():
 
 
 @onnx_test()
-def trilu_test():
+def triu_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4])
     y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4])
 
@@ -7731,7 +8846,7 @@ def trilu_test():
 
 
 @onnx_test()
-def trilu_batch_diff_k_test():
+def triu_batch_diff_k_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [2, 2, 3])
     k = np.array([2])
     y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 2, 3])
@@ -7749,7 +8864,24 @@ def trilu_batch_diff_k_test():
 
 
 @onnx_test()
-def trilu_lower_test():
+def tril_batch_diff_k_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [2, 2, 3])
+    k = np.array([2])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [2, 2, 3])
+    k_tensor = helper.make_tensor(name='k',
+                                  data_type=TensorProto.INT64,
+                                  dims=k.shape,
+                                  vals=k.astype(np.int64))
+
+    node = onnx.helper.make_node('Trilu',
+                                 inputs=['x', 'k'],
+                                 outputs=['y'],
+                                 upper=0)
+    return ([node], [x], [y], [k_tensor])
+
+
+@onnx_test()
+def tril_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4])
     y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4])
 
@@ -7758,7 +8890,7 @@ def trilu_lower_test():
 
 
 @onnx_test()
-def trilu_neg_k_test():
+def triu_neg_k_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4])
     k = np.array([-1])
     y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4])
@@ -7772,7 +8904,23 @@ def trilu_neg_k_test():
 
 
 @onnx_test()
-def trilu_out_k_test():
+def tril_neg_k_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4])
+    k = np.array([-1])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4])
+    k_tensor = helper.make_tensor(name='k',
+                                  data_type=TensorProto.INT64,
+                                  dims=k.shape,
+                                  vals=k.astype(np.int64))
+    node = onnx.helper.make_node('Trilu',
+                                 inputs=['x', 'k'],
+                                 outputs=['y'],
+                                 upper=0)
+    return ([node], [x], [y], [k_tensor])
+
+
+@onnx_test()
+def triu_out_k_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4])
     k = np.array([5])
     y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4])
@@ -7786,7 +8934,23 @@ def trilu_out_k_test():
 
 
 @onnx_test()
-def trilu_row_one_test():
+def tril_out_k_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [3, 4])
+    k = np.array([5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [3, 4])
+    k_tensor = helper.make_tensor(name='k',
+                                  data_type=TensorProto.INT64,
+                                  dims=k.shape,
+                                  vals=k.astype(np.int64))
+    node = onnx.helper.make_node('Trilu',
+                                 inputs=['x', 'k'],
+                                 outputs=['y'],
+                                 upper=0)
+    return ([node], [x], [y], [k_tensor])
+
+
+@onnx_test()
+def triu_row_one_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 4])
     k = np.array([1])
     y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1, 4])
@@ -7803,6 +8967,23 @@ def trilu_row_one_test():
     return ([node], [x], [y], [k_tensor])
 
 
+@onnx_test()
+def tril_row_one_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 4])
+    k = np.array([1])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [1, 4])
+    k_tensor = helper.make_tensor(name='k',
+                                  data_type=TensorProto.INT64,
+                                  dims=k.shape,
+                                  vals=k.astype(np.int64))
+
+    node = onnx.helper.make_node('Trilu',
+                                 inputs=['x', 'k'],
+                                 outputs=['y'],
+                                 upper=0)
+    return ([node], [x], [y], [k_tensor])
+
+
 @onnx_test()
 def undefined_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [2, 3, 4, 5])
@@ -7886,6 +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/group_norm_3d_half_test.onnx

+group_norm_3d_half_test:

+M
+
x
+scale
+bias
y"GroupNormalization*
+epsilon'7

*
+
+num_groups
group_norm_3d_half_testZ
+
x
+
+
+

+
+Z
+scale
+
+
+
+Z
+bias
+
+
+
+b
+
y
+
+
+

+
+B
\ No newline at end of file