Merge branch 'develop' into resnet50_partition

src/include/migraphx/tmp_dir.hpp

@@ -34,6 +34,7 @@ struct MIGRAPHX_EXPORT tmp_dir
 {
     fs::path path;
     tmp_dir(const std::string& prefix = "");
+    tmp_dir(tmp_dir&&) = default;
 
     void execute(const std::string& exe, const std::string& args) const;
 

src/include/migraphx/type_name.hpp

@@ -34,7 +34,7 @@ template <class PrivateMigraphTypeNameProbe>
 std::string compute_type_name()
 {
     std::string name;
-#ifdef _MSC_VER
+#if defined(_MSC_VER) && !defined(__clang__)
     name = typeid(PrivateMigraphTypeNameProbe).name();
     name = name.substr(7);
 #else

src/include/migraphx/verify.hpp

@@ -29,10 +29,13 @@
 #include <functional>
 #include <iostream>
 #include <numeric>
+#include <assert.h>
 
 #include <migraphx/float_equal.hpp>
 #include <migraphx/config.hpp>
+#include <migraphx/env.hpp>
 
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_VERIFY_ENABLE_ALLCLOSE)
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace verify {
@@ -87,8 +90,7 @@ struct not_finite_fn
     template <class T>
     bool operator()(T x) const
     {
-        using std::isfinite;
-        return not isfinite(x);
+        return not std::isfinite(static_cast<double>(x));
     }
 };
 static constexpr not_finite_fn not_finite{};
@@ -98,8 +100,7 @@ struct compare_mag_fn
     template <class T, class U>
     bool operator()(T x, U y) const
     {
-        using std::fabs;
-        return fabs(x) < fabs(y);
+        return std::fabs(x) < std::fabs(y);
     }
 };
 static constexpr compare_mag_fn compare_mag{};
@@ -187,16 +188,103 @@ double rms_range(const R1& r1, const R2& r2)
         return std::numeric_limits<range_value<R1>>::max();
 }
 
+template <class R>
+double get_rms_tol(const R&, std::size_t tolerance = 80)
+{
+    double threshold = std::numeric_limits<range_value<R>>::epsilon() * tolerance;
+    return threshold;
+}
+
+/*
+C++ doesn't support named arguments, this is just wrapper that helps distinguish between actual
+results v/s expected results arguments.
+*/
+template <class T>
+struct expected
+{
+    expected() = default;
+    explicit expected(const T& input) : x(&input) {}
+    const T& data() const
+    {
+        assert(x != nullptr);
+        return *x;
+    }
+
+    private:
+    const T* x = nullptr;
+};
+
+// deduction guide for templated expected class
+template <class T>
+expected(const T&) -> expected<T>;
+
+struct tolerance
+{
+    double rms_tol = 0.001;
+    double atol    = 0.001;
+    double rtol    = 0.001;
+};
+
+/*
+MIGraphX implementation of numpy's np.allclose() which checks if elementwise absolute diff is within
+tolerance using this formula:  abs(a - b) < atol + rtol(abs(b))
+*/
+template <class R1, class R2>
+bool allclose(const R1& r1, const R2& r2, tolerance tols)
+{
+    std::size_t n = range_distance(r1);
+    if(n == range_distance(r2))
+    {
+        auto idx = mismatch_idx(r1, r2, [&](auto x, auto y) {
+            return abs_diff(double(x), double(y)) < tols.atol + tols.rtol * std::abs(double(y));
+        });
+        return idx >= range_distance(r1);
+    }
+    return false;
+}
+
 template <class R1, class R2>
-bool verify_range(const R1& r1, const R2& r2, double tolerance = 80, double* out_error = nullptr)
+bool verify_rms_range(const R1& r1,
+                      const R2& r2,
+                      std::size_t tolerance = 80,
+                      double* out_rms_error = nullptr)
 {
-    double threshold = std::numeric_limits<range_value<R1>>::epsilon() * tolerance;
+    double threshold = get_rms_tol(r1, tolerance);
     auto error       = rms_range(r1, r2);
-    if(out_error != nullptr)
-        *out_error = error;
+    if(out_rms_error != nullptr)
+        *out_rms_error = error;
     return error <= threshold;
 }
 
+template <class R1, class R2>
+bool verify_range_with_tolerance(const R1& r1,
+                                 const expected<R2>& r2,
+                                 tolerance tols        = tolerance{},
+                                 double* out_rms_error = nullptr)
+{
+    auto rms_error = rms_range(r1, r2.data());
+    // disable ewise_verify by default for now, it requires lot of tests to be fixed
+    bool ewise_verify = true;
+    if(enabled(MIGRAPHX_VERIFY_ENABLE_ALLCLOSE{}))
+    {
+        ewise_verify = allclose(r1, r2.data(), tols);
+    }
+    if(out_rms_error != nullptr)
+        *out_rms_error = rms_error;
+    return rms_error <= tols.rms_tol and ewise_verify;
+}
+
+// expected argument should be passed as second, but if it is passed as the first by mistake then
+// flip the order
+template <class R1, class R2>
+bool verify_range_with_tolerance(const expected<R1>& r1,
+                                 const R2& r2,
+                                 tolerance tols        = tolerance{},
+                                 double* out_rms_error = nullptr)
+{
+    return verify_rms_range(r2, r1, tols, out_rms_error);
+}
+
 } // namespace verify
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/verify_args.hpp

@@ -31,11 +31,15 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-MIGRAPHX_EXPORT
-bool verify_args(const std::string& name,
-                 const argument& ref_arg,
-                 const argument& target_arg,
-                 double tolerance = 80);
+MIGRAPHX_EXPORT bool verify_args(const std::string& name,
+                                 const argument& target_arg,
+                                 const verify::expected<argument>& ref_arg,
+                                 verify::tolerance);
+
+MIGRAPHX_EXPORT bool verify_args_with_tolerance(const std::string& name,
+                                                const argument& target_arg,
+                                                const verify::expected<argument>& ref_arg,
+                                                std::size_t tolerance = 80);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/onnx/broadcast_qdq.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 <migraphx/onnx/broadcast_qdq.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+// This method is to prep for quantizelinear or dequantizelinear operation for
+// either the broadcasting of weight-scale or zero-points of qlinearadd operator
+// outputs: operator op (inputs x, broadcasted: scale (float) & zero_pt (8-bit))
+instruction_ref bcast_qdq_instr(const std::string& op_name,
+                                instruction_ref x_in,
+                                instruction_ref arg_fscale,
+                                instruction_ref arg_z_pt,
+                                const onnx_parser::node_info& info)
+{
+    auto in_lens = x_in->get_shape().lens();
+
+    // prep 1: broadcast scale. it can come as a scalar or a 1-D tensor.
+    instruction_ref bcast_scale;
+    if(arg_fscale->get_shape().elements() > 1)
+        bcast_scale = info.add_instruction(
+            migraphx::make_op("broadcast", {{"axis", 0}, {"out_lens", in_lens}}), arg_fscale);
+    else
+        bcast_scale = info.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", in_lens}}), arg_fscale);
+
+    // prep 2: broadcast zero point. it can come as a scalar or a 1-D tensor.
+    instruction_ref bcast_zero_pt;
+    if(arg_z_pt->get_shape().elements() > 1)
+        bcast_zero_pt = info.add_instruction(
+            migraphx::make_op("broadcast", {{"axis", 0}, {"out_lens", in_lens}}), arg_z_pt);
+    else
+        bcast_zero_pt = info.add_instruction(
+            migraphx::make_op("multibroadcast", {{"out_lens", in_lens}}), arg_z_pt);
+
+    // op_name is either quantizelinear or dequantizelinear:
+    return info.add_instruction(migraphx::make_op(op_name), x_in, bcast_scale, bcast_zero_pt);
+}
+
+// Multibroadcast a scaler..
+instruction_ref bcast_scalar_instr(const migraphx::shape& shape_out,
+                                   instruction_ref arg_in,
+                                   const onnx_parser::node_info& info)
+{
+    auto bcast_instr_out = info.add_instruction(
+        migraphx::make_op("multibroadcast", {{"out_lens", shape_out.lens()}}), arg_in);
+    return bcast_instr_out;
+}
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/include/migraphx/onnx/broadcast_qdq.hpp

+/*
+ * 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.
+ */
+
+#ifndef MIGRAPHX_GUARD_AMDMIGRAPHX_ONNX_BROADCAST_QDQ_HPP
+#define MIGRAPHX_GUARD_AMDMIGRAPHX_ONNX_BROADCAST_QDQ_HPP
+
+#include <string>
+
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+// This method is to prep for quantizelinear or dequantizelinear operation for
+// either the broadcasting of weight-scale or zero-points of qlinearadd operator
+// outputs: operator op (inputs x, broadcasted: scale (float) & zero_pt (8-bit))
+instruction_ref bcast_qdq_instr(const std::string& op_name,
+                                instruction_ref x_in,
+                                instruction_ref arg_fscale,
+                                instruction_ref arg_z_pt,
+                                const onnx_parser::node_info& info);
+
+// Multibroadcast a scaler..
+instruction_ref bcast_scalar_instr(const migraphx::shape& shape_out,
+                                   instruction_ref arg_in,
+                                   const onnx_parser::node_info& info);
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/onnx/onnx_parser.cpp

@@ -244,7 +244,7 @@ void onnx_parser::parse_from(std::istream& is, std::string name)
     this->filename   = std::move(name);
     auto parent_path = fs::path(this->filename).parent_path();
     if(not parent_path.empty())
-        this->path = parent_path;
+        this->path = parent_path.string();
 
     onnx::ModelProto model;
     if(model.ParseFromIstream(&is))

src/onnx/parse_castlike.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 <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_castlike : op_parser<parse_castlike>
+{
+    std::vector<op_desc> operators() const { return {{"CastLike"}}; }
+
+    instruction_ref parse(const op_desc& /*opd*/,
+                          const onnx_parser& /*parser*/,
+                          const onnx_parser::node_info& info,
+                          const std::vector<instruction_ref>& args) const
+    {
+        if(not(args.size() == 2))
+        {
+            MIGRAPHX_THROW("PARSE_CASTLIKE: CastLike must have exactly 2 inputs!");
+        }
+        shape::type_t target_type = args[1]->get_shape().type();
+        return info.add_instruction(make_op("convert", {{"target_type", target_type}}), args[0]);
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_constant.cpp

@@ -25,6 +25,7 @@
 #include <migraphx/ranges.hpp>
 #include <migraphx/literal.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/stringutils.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -39,16 +40,38 @@ struct parse_constant : op_parser<parse_constant>
                           onnx_parser::node_info info,
                           const std::vector<instruction_ref>& /*args*/) const
     {
-        literal v = parser.parse_value(info.attributes.at("value"));
+        static const std::vector<std::string> attributes = {
+            "value", "value_float", "value_floats", "value_int", "value_ints"};
+
+        std::vector<std::string> present_attributes;
+        std::copy_if(attributes.begin(),
+                     attributes.end(),
+                     std::back_inserter(present_attributes),
+                     [&](const std::string& a) { return contains(info.attributes, a); });
+
+        if(present_attributes.empty())
+        {
+            MIGRAPHX_THROW("Constant node does not contain any supported attribute");
+        }
+
+        if(present_attributes.size() > 1)
+        {
+            MIGRAPHX_THROW("Constant contains multiple attributes: " +
+                           join_strings(std::move(present_attributes), ", "));
+        }
+
+        // cppcheck-suppress accessMoved
+        auto&& attr = info.attributes[present_attributes[0]];
+        literal v   = parser.parse_value(attr);
+
         // return empty literal
         if(v.get_shape().elements() == 0)
         {
             return info.add_literal(literal{v.get_shape().type()});
         }
 
-        auto dim_size = info.attributes.at("value").t().dims_size();
         // if dim_size is 0, it is a scalar
-        if(dim_size == 0)
+        if(attr.has_t() and attr.t().dims_size() == 0)
         {
             migraphx::shape scalar_shape{v.get_shape().type()};
             return info.add_literal(migraphx::literal{scalar_shape, v.data()});

src/onnx/parse_constant_of_shape.cpp

@@ -49,6 +49,8 @@ struct parse_constant_of_shape : op_parser<parse_constant_of_shape>
             {
                 MIGRAPHX_THROW("ConstantOfShape: attribute value can contain only 1 elements!");
             }
+            // convert to a scalar literal
+            l_val = literal(shape{l_val.get_shape().type(), {1}, {0}}, l_val.data());
         }
         else
         {
@@ -64,30 +66,37 @@ struct parse_constant_of_shape : op_parser<parse_constant_of_shape>
             migraphx::shape s;
             // input is empty, output is a scalar
             auto type = l_val.get_shape().type();
-            // empty input tensor, output is a scalar
-            if(args[0]->get_shape().elements() == 0)
+            migraphx::argument input = args[0]->eval();
+            if(not input.empty())
             {
-                s = migraphx::shape{type, {1}, {0}};
+                // empty input tensor, output is a scalar
+                if(args[0]->get_shape().elements() == 0)
+                {
+                    s = migraphx::shape{type, {1}, {0}};
+                }
+                else
+                {
+                    std::vector<std::size_t> dims;
+                    input.visit([&](auto ia) { dims.assign(ia.begin(), ia.end()); });
+                    s = migraphx::shape{type, dims};
+                }
+                literal l_out{};
+                l_val.visit([&](auto val) {
+                    using val_type = std::remove_cv_t<typename decltype(val)::value_type>;
+                    // l_val contains only one element
+                    std::vector<val_type> out_vec(s.elements(), val.front());
+                    l_out = literal(s, out_vec);
+                });
+                return info.add_literal(l_out);
             }
+            // has variable input (dynamic shape buffer)
             else
             {
-                migraphx::argument in = args[0]->eval();
-                check_arg_empty(in, "ConstantOfShape: dynamic shape is not supported");
-
-                std::vector<std::size_t> dims;
-                in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
-                s = migraphx::shape{type, dims};
+                auto dv_lit = info.add_literal(l_val);
+                auto alloc_ins =
+                    info.add_instruction(make_op("allocate", {{"buf_type", type}}), args[0]);
+                return info.add_instruction(make_op("fill"), dv_lit, alloc_ins);
             }
-
-            literal l_out{};
-            l_val.visit([&](auto val) {
-                using val_type = std::remove_cv_t<typename decltype(val)::value_type>;
-                // l_val contains only one element
-                std::vector<val_type> out_vec(s.elements(), val.front());
-                l_out = literal(s, out_vec);
-            });
-
-            return info.add_literal(l_out);
         }
     }
 };

src/onnx/parse_depthtospace.cpp

@@ -87,8 +87,7 @@ struct parse_depthtospace : op_parser<parse_depthtospace>
 
         auto temp1 = info.add_instruction(make_op("reshape", {{"dims", lens1}}), args[0]);
         auto temp2 = info.add_instruction(make_op("transpose", {{"permutation", perm}}), temp1);
-        return info.add_instruction(make_op("reshape", {{"dims", lens2}}),
-                                    info.make_contiguous(temp2));
+        return info.add_instruction(make_op("reshape", {{"dims", lens2}}), temp2);
     }
 };
 

src/onnx/parse_qlinearadd.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 <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/common.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/onnx/checks.hpp>
+#include <migraphx/onnx/broadcast_qdq.hpp>
+#include <migraphx/instruction.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+/*
+ *********************************************************************************
+ *  Reference: see QLinearAdd in                                                 *
+ *  https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md  *
+ *********************************************************************************
+
+  com.microsoft.QLinearAdd
+  Performs element-wise binary addition on 8 bit data types (with Numpy-style broadcasting support).
+
+  C = (A_scale * (A - A_zero_point) + B_scale * (B - B_zero_point))/C_scale + C_zero_point
+
+  Version
+  This version of the operator has been available since version 1 of the 'com.microsoft' operator
+  set.
+
+  Inputs (7 - 8)
+  A : T
+  First operand.
+
+  A_scale : tensor(float)
+  Input A's scale. It's a scalar, which means a per-tensor/layer quantization.
+
+  A_zero_point (optional) : T
+  Input A zero point. Default value is 0 if it's not specified. It's a scalar, which means a
+  per-tensor/layer quantization.
+
+  B : T
+  Second operand.
+
+  B_scale : tensor(float)
+  Input B's scale. It's a scalar, which means a per-tensor/layer quantization.
+
+  B_zero_point (optional) : T
+  Input B zero point. Default value is 0 if it's not specified. It's a scalar, which means a
+  per-tensor/layer quantization.
+
+  C_scale : tensor(float)
+  Output scale. It's a scalar, which means a per-tensor/layer quantization.
+
+  C_zero_point (optional) : T
+
+  Output zero point. Default value is 0 if it's not specified. It's a scalar, which means a
+  per-tensor/layer quantization.
+
+  Outputs
+  C : T
+  Result, has same element type as two inputs
+
+  Type Constraints
+  T : tensor(uint8), tensor(int8)
+  Constrain input and output types to 8 bit signed and unsigned tensors.
+
+*/
+
+struct parse_qlinearadd : op_parser<parse_qlinearadd>
+{
+    std::vector<op_desc> operators() const { return {{"QLinearAdd"}}; }
+
+    // basic type checking for QLinearAdd Operator
+    void check_inputs(const std::vector<instruction_ref>& args) const
+    {
+        if(args.size() < 7)
+            MIGRAPHX_THROW("QLINEARADD: missing inputs");
+
+        const auto& in_a = args[0];
+        const auto& in_b = args[3];
+
+        auto sh_a = in_a->get_shape();
+        auto sh_b = in_b->get_shape();
+
+        auto type_a = sh_a.type();
+        auto type_b = sh_b.type();
+        if(type_a != migraphx::shape::int8_type and type_a != migraphx::shape::uint8_type)
+            MIGRAPHX_THROW("QLINEARADD: unsupported input type");
+        if(type_b != migraphx::shape::int8_type and type_b != migraphx::shape::uint8_type)
+            MIGRAPHX_THROW("QLINEARADD: unsupported input type");
+        if(type_a != type_b)
+            MIGRAPHX_THROW("QLINEARADD: mismatched input types");
+    }
+
+    instruction_ref parse(const op_desc& /* opd */,
+                          const onnx_parser& /*parser*/,
+                          const onnx_parser::node_info& info,
+                          const std::vector<instruction_ref>& args) const
+    {
+        check_inputs(args);
+
+        // A
+        const auto& in_a         = args[0];
+        const auto& in_scale_a   = args[1];
+        const auto& in_zero_pt_a = args[2];
+
+        auto dquant_a = bcast_qdq_instr("dequantizelinear", in_a, in_scale_a, in_zero_pt_a, info);
+
+        // B
+        const auto& in_b         = args[3];
+        const auto& in_scale_b   = args[4];
+        const auto& in_zero_pt_b = args[5];
+        auto dquant_b = bcast_qdq_instr("dequantizelinear", in_b, in_scale_b, in_zero_pt_b, info);
+
+        // C = A + B
+        auto out_c = info.add_common_op("add", dquant_a, dquant_b);
+
+        const auto& in_scale_c = args[6];
+
+        // zero_pt for C is supplied as the last optional argument..
+        if(args.size() == 8)
+            return (bcast_qdq_instr("quantizelinear", out_c, in_scale_c, args[7], info));
+
+        // if no zero_pt: just broadcast the scale..
+        auto bcast_scale_c = bcast_scalar_instr(out_c->get_shape(), in_scale_c, info);
+        return (info.add_instruction(migraphx::make_op("quantizelinear"), out_c, bcast_scale_c));
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/parse_reshape.cpp

@@ -53,8 +53,7 @@ struct parse_reshape : op_parser<parse_reshape>
             s.visit([&](auto v) { copy(v, std::back_inserter(dims)); });
         }
 
-        auto cont = info.add_instruction(make_op("contiguous"), args[0]);
-        return info.add_instruction(make_op("reshape", {{"dims", dims}}), cont);
+        return info.add_instruction(make_op("reshape", {{"dims", dims}}), args[0]);
     }
 };
 

src/onnx/parse_resize.cpp

@@ -320,7 +320,8 @@ struct parse_resize : op_parser<parse_resize>
 
             // get the number of dimensions
             std::size_t n_dim = out_lens.size();
-            auto vvv_ind = std::vector(n_dim, std::vector(2, std::vector<size_t>(out_elements)));
+            std::vector<std::vector<std::size_t>> vv_ind(2, std::vector<std::size_t>(out_elements));
+            std::vector<std::vector<std::vector<std::size_t>>> vvv_ind(n_dim, vv_ind);
             std::vector<std::vector<float>> delta(n_dim, std::vector<float>(out_elements));
 
             shape_for_each(out_s, [&](const auto& out_idx_v, size_t out_idx) {

src/onnx/parse_roialign.cpp

@@ -37,15 +37,18 @@ struct parse_roialign : op_parser<parse_roialign>
     std::vector<op_desc> operators() const { return {{"RoiAlign"}}; }
 
     instruction_ref parse(const op_desc& /*opd*/,
-                          const onnx_parser& /*parser*/,
+                          const onnx_parser& parser,
                           onnx_parser::node_info info,
                           const std::vector<instruction_ref>& args) const
     {
-        std::string coord_trans_mode = "half_pixel";
-        if(contains(info.attributes, "coordinate_transformation_mode"))
+        std::string coord_trans_mode =
+            parser.opset_version >= 16 ? "half_pixel" : "output_half_pixel";
+
+        if(const auto* a = "coordinate_transformation_mode"; contains(info.attributes, a))
         {
-            coord_trans_mode = info.attributes.at("coordinate_transformation_mode").s();
+            coord_trans_mode = info.attributes.at(a).s();
         }
+
         if(not contains({"half_pixel", "output_half_pixel"}, coord_trans_mode))
         {
             MIGRAPHX_THROW("coordinate_transformation_mode \"" + coord_trans_mode +

src/onnx/parse_spacetodepth.cpp

@@ -73,8 +73,7 @@ struct parse_spacetodepth : op_parser<parse_spacetodepth>
         std::vector<int64_t> perm = {0, 3, 5, 1, 2, 4};
         auto temp1 = info.add_instruction(make_op("reshape", {{"dims", trans_lens}}), args[0]);
         auto temp2 = info.add_instruction(make_op("transpose", {{"permutation", perm}}), temp1);
-        return info.add_instruction(make_op("reshape", {{"dims", res_lens}}),
-                                    info.make_contiguous(temp2));
+        return info.add_instruction(make_op("reshape", {{"dims", res_lens}}), temp2);
     }
 };
 

src/optimize_module.cpp

 /*
  * 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
@@ -36,8 +36,12 @@ void optimize_module::apply(module_pass_manager& mpm) const
 {
     for(int i = 0; i < 2; i++)
     {
-        mpm.run_pass(simplify_reshapes{});
-        mpm.run_pass(simplify_algebra{});
+        // loop to further optimize after initial transformations
+        for(int j = 0; j < 2; j++)
+        {
+            mpm.run_pass(simplify_reshapes{});
+            mpm.run_pass(simplify_algebra{});
+        }
         mpm.run_pass(eliminate_common_subexpression{});
         mpm.run_pass(dead_code_elimination{});
         mpm.run_pass(propagate_constant{});

src/process.cpp

@@ -26,13 +26,23 @@
 #include <migraphx/env.hpp>
 #include <functional>
 #include <iostream>
+#include <optional>
+
+#ifdef _WIN32
+// cppcheck-suppress definePrefix
+#define WIN32_LEAN_AND_MEAN
+#include <Windows.h>
+#else
 #include <unistd.h>
+#endif
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_CMD_EXECUTE)
 
+#ifndef _WIN32
+
 std::function<void(const char*)> redirect_to(std::ostream& os)
 {
     return [&](const char* x) { os << x; };
@@ -74,6 +84,155 @@ int exec(const std::string& cmd, std::function<void(process::writer)> std_in)
     });
 }
 
+#else
+
+constexpr std::size_t MIGRAPHX_PROCESS_BUFSIZE = 4096;
+
+class pipe
+{
+    public:
+    explicit pipe(bool inherit_handle = true)
+    {
+        SECURITY_ATTRIBUTES attrs;
+        attrs.nLength              = sizeof(SECURITY_ATTRIBUTES);
+        attrs.bInheritHandle       = inherit_handle ? TRUE : FALSE;
+        attrs.lpSecurityDescriptor = nullptr;
+
+        if(CreatePipe(&m_read, &m_write, &attrs, 0) == FALSE)
+            throw GetLastError();
+
+        if(SetHandleInformation(&m_read, HANDLE_FLAG_INHERIT, 0) == FALSE)
+            throw GetLastError();
+    }
+
+    pipe(const pipe&)            = delete;
+    pipe& operator=(const pipe&) = delete;
+
+    pipe(pipe&&) = default;
+
+    ~pipe()
+    {
+        CloseHandle(m_read);
+        m_read = nullptr;
+        CloseHandle(m_write);
+        m_write = nullptr;
+    }
+
+    std::optional<std::pair<bool, DWORD>> read(LPVOID buffer, DWORD length) const
+    {
+        DWORD bytes_read;
+        if(ReadFile(m_read, buffer, length, &bytes_read, nullptr) == FALSE)
+        {
+            DWORD error{GetLastError()};
+            if(error != ERROR_MORE_DATA)
+            {
+                return std::nullopt;
+            }
+            return {{true, bytes_read}};
+        }
+        return {{false, bytes_read}};
+    }
+
+    HANDLE get_read_handle() const { return m_read; }
+
+    bool write(LPCVOID buffer, DWORD length) const
+    {
+        DWORD bytes_written;
+        return WriteFile(m_write, buffer, length, &bytes_written, nullptr) == TRUE;
+    }
+
+    HANDLE get_write_handle() const { return m_write; }
+
+    private:
+    HANDLE m_write = nullptr, m_read = nullptr;
+};
+
+template <typename F>
+int exec(const std::string& cmd, F f)
+{
+    try
+    {
+        if(enabled(MIGRAPHX_TRACE_CMD_EXECUTE{}))
+            std::cout << cmd << std::endl;
+
+        STARTUPINFO info;
+        PROCESS_INFORMATION process_info;
+
+        pipe in{}, out{};
+
+        ZeroMemory(&info, sizeof(STARTUPINFO));
+        info.cb         = sizeof(STARTUPINFO);
+        info.hStdError  = out.get_write_handle();
+        info.hStdOutput = out.get_write_handle();
+        info.hStdInput  = in.get_read_handle();
+        info.dwFlags |= STARTF_USESTDHANDLES;
+
+        ZeroMemory(&process_info, sizeof(process_info));
+
+        if(CreateProcess(nullptr,
+                         const_cast<LPSTR>(cmd.c_str()),
+                         nullptr,
+                         nullptr,
+                         TRUE,
+                         0,
+                         nullptr,
+                         nullptr,
+                         &info,
+                         &process_info) == FALSE)
+        {
+            return GetLastError();
+        }
+
+        f(in, out);
+
+        WaitForSingleObject(process_info.hProcess, INFINITE);
+
+        DWORD status{};
+        GetExitCodeProcess(process_info.hProcess, &status);
+
+        CloseHandle(process_info.hProcess);
+        CloseHandle(process_info.hThread);
+
+        return static_cast<int>(status);
+    }
+    // cppcheck-suppress catchExceptionByValue
+    catch(DWORD last_error)
+    {
+        return last_error;
+    }
+}
+
+int exec(const std::string& cmd)
+{
+    TCHAR buffer[MIGRAPHX_PROCESS_BUFSIZE];
+    HANDLE std_out{GetStdHandle(STD_OUTPUT_HANDLE)};
+    return (std_out == nullptr or std_out == INVALID_HANDLE_VALUE)
+               ? GetLastError()
+               : exec(cmd, [&](const pipe&, const pipe& out) {
+                     for(;;)
+                     {
+                         if(auto result = out.read(buffer, MIGRAPHX_PROCESS_BUFSIZE))
+                         {
+                             auto [more_data, bytes_read] = *result;
+                             if(not more_data or bytes_read == 0)
+                                 break;
+                             DWORD written;
+                             if(WriteFile(std_out, buffer, bytes_read, &written, nullptr) == FALSE)
+                                 break;
+                         }
+                     }
+                 });
+}
+
+int exec(const std::string& cmd, std::function<void(process::writer)> std_in)
+{
+    return exec(cmd, [&](const pipe& in, const pipe&) {
+        std_in([&](const char* buffer, std::size_t n) { in.write(buffer, n); });
+    });
+}
+
+#endif
+
 struct process_impl
 {
     std::string command{};
@@ -119,7 +278,14 @@ process& process::cwd(const fs::path& p)
     return *this;
 }
 
-void process::exec() { impl->check_exec(impl->get_command(), redirect_to(std::cout)); }
+void process::exec()
+{
+#ifndef _WIN32
+    impl->check_exec(impl->get_command(), redirect_to(std::cout));
+#else
+    impl->check_exec(impl->get_command());
+#endif
+}
 
 void process::write(std::function<void(process::writer)> pipe_in)
 {

src/program.cpp

@@ -352,7 +352,7 @@ void program::finalize()
 template <class T>
 std::string classify(T x)
 {
-    switch(std::fpclassify(x))
+    switch(std::fpclassify(static_cast<double>(x)))
     {
     case FP_INFINITE: return "inf";
     case FP_NAN: return "nan";

src/propagate_constant.cpp

 /*
  * 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
@@ -35,10 +35,10 @@ inline namespace MIGRAPHX_INLINE_NS {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_PROPAGATE_CONSTANT)
 
-bool skip_propogate(instruction_ref ins)
+bool skip_propagate(instruction_ref ins)
 {
     if(ins->name() == "contiguous")
-        return skip_propogate(ins->inputs().front());
+        return skip_propagate(ins->inputs().front());
     auto&& s = ins->get_shape();
     if(s.broadcasted() and not s.scalar())
         return true;
@@ -47,7 +47,7 @@ bool skip_propogate(instruction_ref ins)
     return false;
 }
 
-bool is_const_ins(instruction_ref ins) { return ins->can_eval() and not skip_propogate(ins); }
+bool is_const_ins(instruction_ref ins) { return ins->can_eval() and not skip_propagate(ins); }
 
 void propagate_constant::apply(module& m) const
 {