Merge branch 'develop' into ubuntu-22.04-default

src/module.cpp

@@ -873,12 +873,11 @@ module::print_py(std::ostream& os,
             if(ins->name() == "@literal")
             {
                 os << mname << ".add_literal(";
-                bool use_abs = false;
-                ins->get_literal().visit([&](auto v) {
-                    use_abs = std::none_of(v.begin(), v.end(), [](auto x) { return x < 0; });
-                });
+                const bool use_abs = false;
                 // Disable abs for now
-                use_abs = false;
+                // ins->get_literal().visit([&](auto v) {
+                //     use_abs = std::none_of(v.begin(), v.end(), [](auto x) { return x < 0; });
+                // });
                 if(use_abs)
                     os << "migraphx.abs_literal(";
                 os << "migraphx.generate_argument(";
@@ -1011,9 +1010,17 @@ std::vector<module_ref> module::get_sub_modules(bool shallow) const
 
 module& module::sort()
 {
+    auto implicit_deps = calc_implicit_deps();
     fix([&](auto self, auto ins) {
         this->move_instruction(ins, this->begin());
-        for(auto child : ins->inputs())
+        auto ins_inputs = ins->inputs();
+        if(implicit_deps.find(ins) != implicit_deps.end())
+        {
+            auto ins_implict_inputs = implicit_deps.at(ins);
+            ins_inputs.insert(
+                ins_inputs.end(), ins_implict_inputs.begin(), ins_implict_inputs.end());
+        }
+        for(auto child : ins_inputs)
         {
             if(not contains(this->impl->instructions, child))
             {

src/normalize_attributes.cpp

@@ -26,7 +26,7 @@
 #include <migraphx/normalize_attributes.hpp>
 #include <migraphx/stringutils.hpp>
 #include <migraphx/op/normalize_attribute.hpp>
-
+#include <migraphx/op/common.hpp>
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
@@ -49,6 +49,10 @@ auto tune_attribute(const std::vector<int64_t>& vec,
                     Message m)
 {
     std::vector<int64_t> result(vec);
+    if(result.empty())
+    {
+        return result;
+    };
     int64_t n_rank                                 = input_shape.ndim();
     std::vector<op::normalize_attribute> vec_attrs = val.to_vector<op::normalize_attribute>();
     if(contains(vec_attrs, op::normalize_attribute::use_output))
@@ -187,15 +191,21 @@ bool normalize_attributes(operation& op, const shape& input_shape)
     auto attrs = op.attributes();
     auto val   = op.to_value();
     if(attrs.contains("normalize_padding"))
+    {
+        bool use_auto_padding =
+            (val.contains("padding_mode") and
+             (val.at("padding_mode").to<int>() != migraphx::op::padding_mode_t::default_));
+        if(not use_auto_padding)
         {
             auto padding       = val.at(attrs.at("normalize_padding").to<std::string>());
             auto padding_size  = padding.size();
             auto padding_start = 2;
-
             if(padding_size == 2 * (input_shape.ndim() - padding_start))
                 tuned = true;
             else if(padding_size != (input_shape.ndim() - padding_start))
-            MIGRAPHX_THROW("inconsistent padding size");
+            {
+                MIGRAPHX_THROW("normalize_attributes: inconsistent padding vector size ");
+            }
             else
             {
                 auto result    = tune_pad_attribute(padding);
@@ -204,6 +214,7 @@ bool normalize_attributes(operation& op, const shape& input_shape)
                 tuned = true;
             }
         }
+    }
     if(not attrs.contains("normalize_axes"))
     {
         return tuned;
@@ -251,5 +262,22 @@ bool normalize_attributes(operation& op, const shape& input_shape)
     return tuned;
 }
 
+std::vector<int64_t> normalize_axes(const std::vector<int64_t>& axes,
+                                    const shape& input_shape,
+                                    const value& attr_val,
+                                    const std::string& prefix)
+{
+    return tune_attribute(axes, {}, attr_val, input_shape, [&] { return prefix; });
+}
+
+std::vector<int64_t> normalize_indices(const std::vector<int64_t>& indices,
+                                       const std::vector<int64_t>& axes,
+                                       const shape& input_shape,
+                                       const value& attr_val,
+                                       const std::string& prefix)
+{
+    return tune_attribute(indices, axes, attr_val, input_shape, [&] { return prefix; });
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/onnx/include/migraphx/onnx/onnx_parser.hpp

@@ -117,6 +117,7 @@ struct onnx_parser
     parse_graph(module* mod, const onnx::GraphProto& graph, bool inlining = false);
     literal parse_value(const onnx::AttributeProto& attr) const;
     literal parse_tensor(const onnx::TensorProto& t) const;
+    shape parse_type(const onnx::TypeProto& t) const;
     shape parse_type(const onnx::TypeProto& t, const std::vector<std::size_t>& input_dims) const;
 };
 

src/onnx/onnx_parser.cpp

@@ -357,10 +357,9 @@ parse_inputs(const onnx_parser& parser,
             }
 
             shape s;
-            std::vector<std::size_t> dims;
             if(parser.map_input_dims.count(name) > 0)
             {
-                dims = parser.map_input_dims.at(name);
+                std::vector<std::size_t> dims = parser.map_input_dims.at(name);
                 s    = parser.parse_type(input.type(), dims);
             }
             else if(parser.map_dyn_input_dims.count(name) > 0)
@@ -370,7 +369,7 @@ parse_inputs(const onnx_parser& parser,
             }
             else
             {
-                s = parser.parse_type(input.type(), dims);
+                s = parser.parse_type(input.type());
             }
             mod_insts[name] = mod->add_parameter(name, s);
         }
@@ -553,14 +552,9 @@ literal onnx_parser::parse_tensor(const onnx::TensorProto& t) const
     }
     MIGRAPHX_THROW("PARSE_TENSOR: Invalid tensor type");
 }
-shape onnx_parser::parse_type(const onnx::TypeProto& t,
-                              const std::vector<std::size_t>& input_dims) const
+shape onnx_parser::parse_type(const onnx::TypeProto& t) const
 {
     shape::type_t shape_type = get_type(t.tensor_type().elem_type());
-    if(not input_dims.empty())
-    {
-        return {shape_type, input_dims};
-    }
 
     std::vector<shape::dynamic_dimension> dynamic_dims;
     auto&& tensor_dims = t.tensor_type().shape().dim();
@@ -590,6 +584,15 @@ shape onnx_parser::parse_type(const onnx::TypeProto& t,
     return shape_from_dyn_dims(shape_type, dynamic_dims);
 }
 
+shape onnx_parser::parse_type(const onnx::TypeProto& t,
+                              const std::vector<std::size_t>& input_dims) const
+{
+    shape::type_t shape_type = get_type(t.tensor_type().elem_type());
+    if(input_dims.empty())
+        return {shape_type};
+    return {shape_type, input_dims};
+}
+
 shape::type_t get_type(int dtype)
 {
     switch(dtype)

src/onnx/parse_constant_of_shape.cpp

@@ -55,9 +55,6 @@ struct parse_constant_of_shape : op_parser<parse_constant_of_shape>
             l_val = literal({shape::float_type, {1}, {0}}, {0.0f});
         }
 
-        // input is empty, output is a scalar
-        auto type = l_val.get_shape().type();
-
         if(args.empty())
         {
             MIGRAPHX_THROW("ConstantOfShape : must have 1 input!");
@@ -65,6 +62,8 @@ struct parse_constant_of_shape : op_parser<parse_constant_of_shape>
         else
         {
             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)
             {

src/onnx/parse_pooling.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
@@ -151,26 +151,6 @@ struct parse_pooling : op_parser<parse_pooling>
                 kdims, paddings.size() / 2, "PARSE_POOLING: inconsistent explicit paddings");
         }
 
-        if(contains(info.attributes, "auto_pad"))
-        {
-            if(in_shape.dynamic())
-            {
-                MIGRAPHX_THROW(
-                    "PARSE_POOLING: Auto padding pooling with dynamic input shape not supported");
-            }
-            else
-            {
-                values["padding"].clear();
-                // return paddings could be empty, then setting to 0 for no padding
-                cal_auto_padding_size(info,
-                                      values,
-                                      values["lengths"].to_vector<std::size_t>(),
-                                      {1, 1},
-                                      in_shape.lens(),
-                                      paddings);
-            }
-        }
-
         if(paddings.size() != 2 * kdims)
         {
             paddings.resize(kdims * 2);
@@ -192,6 +172,36 @@ struct parse_pooling : op_parser<parse_pooling>
         // used to calculate the supposed output shape
         std::vector<int64_t> orig_padding = paddings;
 
+        // TODO:  add parsing for dilations
+        if(contains(info.attributes, "auto_pad") and
+           to_upper(info.attributes["auto_pad"].s()) != "NOTSET")
+        {
+            auto auto_pad = to_upper(info.attributes["auto_pad"].s());
+            // don't use the given padding sizes, if any
+            // values["padding"].clear();
+            if(in_shape.dynamic())
+            {
+                // set padding_mode to trigger auto padding at runtime
+                bool is_same_upper     = (auto_pad.find("SAME_UPPER") != std::string::npos);
+                values["padding_mode"] = is_same_upper ? to_value(op::padding_mode_t::same_upper)
+                                                       : to_value(op::padding_mode_t::same_lower);
+            }
+            else
+            {
+                // Calculate auto padding
+                // dilations (argument 4) not supported; default to all 1's
+                cal_auto_padding_size(info,
+                                      values,
+                                      values["lengths"].to_vector<std::size_t>(),
+                                      std::vector<size_t>(in_shape.ndim() - 2, 1),
+                                      in_shape.lens(),
+                                      paddings);
+                values["padding"] = paddings;
+                // default padding_mode indicates that padding sizes are not calculated dynamically
+                values["padding_mode"] = migraphx::op::padding_mode_t::default_;
+            }
+        }
+
         std::vector<int64_t> slice_start;
         std::vector<int64_t> slice_end;
         tune_padding_size(values, paddings, count_include_pad, slice_start);
@@ -208,8 +218,9 @@ struct parse_pooling : op_parser<parse_pooling>
             orig_padding.insert(orig_padding.begin(), 2, 0);
             op::pad pad{orig_padding, 0.0f};
             shape padded_shape = pad.compute_shape({l0->get_shape()});
-            auto out_lens      = make_op("pooling", values).compute_shape({padded_shape}).lens();
 
+            // make an op just to get its output shape
+            auto out_lens = make_op("pooling", values).compute_shape({padded_shape}).lens();
             // compute slice_end information
             slice_end.resize(slice_start.size());
             std::transform(out_lens.begin() + 2,

src/onnx/parse_randomuniform_ops.cpp

@@ -96,7 +96,7 @@ struct parse_randomuniform_ops : op_parser<parse_randomuniform_ops>
         if(contains(info.attributes, "seed"))
             gen.seed(info.attributes.at("seed").f());
 
-        std::uniform_real_distribution<> d(high, low);
+        std::uniform_real_distribution<> d(low, high);
         std::vector<double> rand_vals(out_shape.elements());
         std::generate(rand_vals.begin(), rand_vals.end(), [&]() { return d(gen); });
 

src/onnx/parse_slice.cpp

@@ -34,16 +34,65 @@ namespace onnx {
 
 struct parse_slice : op_parser<parse_slice>
 {
+
     std::vector<op_desc> operators() const { return {{"Slice"}}; }
 
+    struct slice_desc
+    {
+        op::slice op;
+        std::vector<instruction_ref> op_args;
+        std::vector<int64_t> steps;
+        std::vector<int64_t> raxes;
+
+        void always_insert(instruction_ref arg) { op_args.insert(op_args.begin(), arg); }
+
+        std::vector<int64_t> insert(instruction_ref arg)
+        {
+            std::vector<int64_t> result;
+            migraphx::argument arg_value = arg->eval();
+            if(arg_value.empty())
+            {
+                op_args.insert(op_args.begin(), arg);
+            }
+            else
+            {
+                arg_value.visit([&](auto s) { result.assign(s.begin(), s.end()); });
+            }
+            return result;
+        }
+    };
+
     instruction_ref parse(const op_desc& /*opd*/,
                           const onnx_parser& parser,
+                          const onnx_parser::node_info& info,
+                          const std::vector<instruction_ref>& args) const
+    {
+        auto sd  = construct_slice_desc(parser, info, args);
+        auto ins = info.add_instruction(sd.op, sd.op_args);
+        if(not sd.raxes.empty())
+        {
+            ins = info.add_instruction(make_op("reverse", {{"axes", sd.raxes}}), ins);
+        }
+        // If any steps are other than default 1, add a "steps" op
+        if(std::any_of(sd.steps.begin(), sd.steps.end(), [](auto s) { return std::abs(s) != 1; }))
+        {
+            std::vector<int64_t> nsteps;
+            std::transform(sd.steps.begin(),
+                           sd.steps.end(),
+                           std::back_inserter(nsteps),
+                           [](auto s) { return std::abs(s); });
+            return ins = info.add_instruction(
+                       make_op("step", {{"axes", sd.op.axes}, {"steps", nsteps}}), ins);
+        }
+        else
+            return ins;
+    }
+
+    slice_desc construct_slice_desc(const onnx_parser& parser,
                                     onnx_parser::node_info info,
                                     std::vector<instruction_ref> args) const
     {
-        op::slice op;
-
-        std::vector<int64_t> steps;
+        slice_desc sd;
 
         // slice can have up to 5 inputs, we first check the 5th one
         // to decide whether MIGRAPHX can handle this slice.
@@ -51,89 +100,73 @@ struct parse_slice : op_parser<parse_slice>
         {
             migraphx::argument step_arg = args.back()->eval();
             check_arg_empty(step_arg, "PARSE_SLICE: cannot handle variable steps for slice");
-            step_arg.visit([&](auto s) { steps.assign(s.begin(), s.end()); });
+            step_arg.visit([&](auto s) { sd.steps.assign(s.begin(), s.end()); });
         }
 
         if(args.size() >= 4)
         {
-            migraphx::argument axes_arg = args.at(3)->eval();
-            check_arg_empty(axes_arg, "PARSE_SLICE: cannot handle variable axes for slice");
-            axes_arg.visit([&](auto s) { op.axes.assign(s.begin(), s.end()); });
+            sd.op.axes = sd.insert(args.at(3));
         }
         else if(contains(info.attributes, "axes"))
         {
             literal s = parser.parse_value(info.attributes.at("axes"));
-            s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
+            s.visit([&](auto v) { copy(v, std::back_inserter(sd.op.axes)); });
         }
 
         if(args.size() >= 3)
         {
-            migraphx::argument end_arg = args.at(2)->eval();
-            check_arg_empty(end_arg, "PARSE_SLICE: cannot handle variable ends for slice");
-            end_arg.visit([&](auto s) { op.ends.assign(s.begin(), s.end()); });
+            sd.op.ends = sd.insert(args.at(2));
         }
         else if(contains(info.attributes, "ends"))
         {
             literal s = parser.parse_value(info.attributes.at("ends"));
-            s.visit([&](auto v) { copy(v, std::back_inserter(op.ends)); });
+            s.visit([&](auto v) { copy(v, std::back_inserter(sd.op.ends)); });
         }
 
         if(args.size() >= 2)
         {
-            migraphx::argument start_arg = args.at(1)->eval();
-            check_arg_empty(start_arg, "PARSE_SLICE: cannot handle variable starts for slice");
-            start_arg.visit([&](auto s) { op.starts.assign(s.begin(), s.end()); });
+            sd.op.starts = sd.insert(args.at(1));
         }
         else if(contains(info.attributes, "starts"))
         {
             literal s = parser.parse_value(info.attributes.at("starts"));
-            s.visit([&](auto v) { copy(v, std::back_inserter(op.starts)); });
+            s.visit([&](auto v) { copy(v, std::back_inserter(sd.op.starts)); });
         }
 
+        // data input argument
+        sd.always_insert(args.at(0));
+
         // If axes arg is not given, the default is all of them.
-        if(op.axes.empty())
+        if(sd.op.axes.empty() and sd.op_args.size() < 3)
         {
             std::vector<int64_t> axes(args[0]->get_shape().ndim());
             std::iota(axes.begin(), axes.end(), int64_t{0});
-            op.axes = axes;
+            sd.op.axes = axes;
         }
 
-        std::vector<int64_t> raxes;
+        if(not sd.steps.empty())
+        {
+            if(sd.op.starts.empty() or sd.op.ends.empty())
+                MIGRAPHX_THROW("PARSE_SLICE: steps and variable starts and ends is not supported");
+            if(sd.op.axes.empty())
+                MIGRAPHX_THROW("PARSE_SLICE: steps and variable axes is not supported");
+        }
 
-        assert(steps.empty() or steps.size() == op.axes.size());
-        assert(op.axes.size() == op.starts.size());
-        assert(op.axes.size() == op.ends.size());
+        assert(sd.steps.empty() or sd.steps.size() == sd.op.axes.size());
 
         // If any axes have negative step, prepare to add a "reverse" op
-        for(auto i : range(steps.size()))
+        for(auto i : range(sd.steps.size()))
         {
-            if(steps[i] >= 0)
+            if(sd.steps[i] >= 0)
                 continue;
-            op.starts[i] += 1;
-            if(op.starts[i] == 0)
-                op.starts[i] = INT_MAX;
-            op.ends[i] += 1;
-            raxes.push_back(op.axes[i]);
-            std::swap(op.starts[i], op.ends[i]);
-        }
-
-        auto ins = info.add_instruction(op, args[0]);
-        if(not raxes.empty())
-        {
-            ins = info.add_instruction(make_op("reverse", {{"axes", raxes}}), ins);
+            sd.op.starts[i] += 1;
+            if(sd.op.starts[i] == 0)
+                sd.op.starts[i] = INT_MAX;
+            sd.op.ends[i] += 1;
+            sd.raxes.push_back(sd.op.axes[i]);
+            std::swap(sd.op.starts[i], sd.op.ends[i]);
         }
-        // If any steps are other than default 1, add a "steps" op
-        if(std::any_of(steps.begin(), steps.end(), [](auto s) { return std::abs(s) != 1; }))
-        {
-            std::vector<int64_t> nsteps;
-            std::transform(steps.begin(), steps.end(), std::back_inserter(nsteps), [](auto s) {
-                return std::abs(s);
-            });
-            return ins = info.add_instruction(
-                       make_op("step", {{"axes", op.axes}, {"steps", nsteps}}), ins);
-        }
-        else
-            return ins;
+        return sd;
     }
 };
 

src/pad_calc.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
@@ -52,6 +52,11 @@ void calculate_padding(int64_t idx,
     }
 }
 
+/**
+ * Given the input array dimensions; kernel (wei_lens); strides; and dilations,
+ * calculate the padding value in each dimension.
+ *
+ */
 std::vector<std::size_t> calc_dyn_auto_pad(const std::vector<std::size_t>& input_lens,
                                            const std::vector<std::size_t>& wei_lens,
                                            const std::vector<std::size_t>& strides,
@@ -60,6 +65,7 @@ std::vector<std::size_t> calc_dyn_auto_pad(const std::vector<std::size_t>& input
 {
     std::vector<std::size_t> padding;
     assert(input_lens.size() >= 3);
+    assert(input_lens.size() == wei_lens.size());
     std::size_t num_spatial_dims = input_lens.size() - 2;
     padding.resize(2 * num_spatial_dims);
     for(std::size_t i = 0; i < num_spatial_dims; i++)
@@ -88,6 +94,11 @@ std::vector<std::size_t> calc_dyn_auto_pad(const std::vector<std::size_t>& input
     return padding;
 }
 
+/**
+ *   Calculate the correct output shape for a convolution with
+ *   a given input size and other parameters.
+ *
+ */
 shape compute_padded_shape(const shape& input,
                            const shape& weights,
                            const std::vector<std::size_t>& padding,
@@ -111,5 +122,33 @@ shape compute_padded_shape(const shape& input,
     return input.with_lens(output_lens);
 }
 
+/**
+ *   Calculate the correct output shape for a pooling with
+ *   a given input size and other parameters.  This uses
+ *   the same formula for pooling that compute_padded_shape() uses
+ *   for convolutions, but takes slightly different inputs.
+ *
+ */
+shape compute_padded_pool_shape(const shape& input,
+                                const shape& kernel,
+                                const std::vector<std::size_t>& padding,
+                                const std::vector<std::size_t>& stride,
+                                const std::vector<std::size_t>& dilation)
+{
+    const size_t num_spatial_dims = input.lens().size() - 2;
+
+    std::vector<size_t> output_lens{input.lens()[0], input.lens()[1]};
+    // calculate the output shape of the pooling: ((W - K + 2P) / S) + 1
+    for(size_t i = 0; i < num_spatial_dims; ++i)
+    {
+        auto padding_factor = padding[i] + padding[i + num_spatial_dims];
+        output_lens.push_back(std::size_t(std::max<std::ptrdiff_t>(
+            1,
+            (input.lens()[i + 2] - (1 + dilation[i] * (kernel.lens()[i] - 1)) + padding_factor) /
+                    stride[i] +
+                1)));
+    }
+    return input.with_lens(output_lens);
+}
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/program.cpp

@@ -40,13 +40,14 @@
 #include <migraphx/make_op.hpp>
 #include <migraphx/marker.hpp>
 #include <migraphx/supported_segments.hpp>
+
 #include <iostream>
+#include <queue>
 #include <sstream>
 #include <algorithm>
 #include <set>
 #include <unordered_map>
 #include <utility>
-
 #include <unordered_set>
 #include <map>
 #include <cassert>
@@ -222,7 +223,7 @@ void program::compile(const std::vector<target>& targets, std::vector<compile_op
     // Gather all the target roots
     std::unordered_multimap<std::size_t, module_ref> roots;
     auto mods = this->get_modules();
-    for(auto* mod : mods)
+    for(const auto* mod : mods)
     {
         for(const auto& ins : *mod)
         {
@@ -547,7 +548,7 @@ std::vector<argument> program::eval(parameter_map params, execution_environment
             ins_out[x] = ss.str();
         });
         ret = generic_eval(*this, contexts, std::move(params), [&](instruction_ref ins, auto f) {
-            auto& ctx = contexts[ins->get_target_id()];
+            const auto& ctx = contexts[ins->get_target_id()];
             ctx.finish();
             std::cout << "Run instruction: " << ins_out.at(ins) << std::endl;
             timer t{};
@@ -727,7 +728,7 @@ static void mod_from_val(module_ref mod,
                                std::back_inserter(module_inputs),
                                [&](const value& i) { return map_mods.at(i.to<std::string>()); });
 
-                for(auto& smod : module_inputs)
+                for(const auto& smod : module_inputs)
                 {
                     mod_from_val(smod, v, instructions, map_mods);
                 }
@@ -1185,17 +1186,25 @@ void program::remove_unused_modules()
     std::vector<module*> unused;
     generic_get_unused_modules(
         impl->modules, generic_get_modules(this->get_main_module()), std::back_inserter(unused));
-    for(auto* m : unused)
+    for(const auto* m : unused)
         this->remove_module(m->name());
 }
 
 program& program::sort()
 {
-    for(auto& pp : this->impl->modules)
+    std::queue<migraphx::module_ref> mqueue;
+    mqueue.push(get_main_module());
+    while(not mqueue.empty())
     {
-        pp.second.sort();
+        module_ref current_mod = mqueue.front();
+        current_mod->sort();
+        mqueue.pop();
+        auto child_mods = current_mod->get_sub_modules(true);
+        for(auto& sub_mod : child_mods)
+        {
+            mqueue.push(sub_mod);
+        }
     }
-
     return *this;
 }
 

src/py/CMakeLists.txt

@@ -24,6 +24,7 @@
 
 option(MIGRAPHX_ENABLE_PYTHON "Enable python bindings" ON)
 add_library(migraphx_py py_loader.cpp)
+migraphx_generate_export_header(migraphx_py)
 target_include_directories(migraphx_py PRIVATE include)
 target_link_libraries(migraphx_py PUBLIC migraphx)
 rocm_install_targets(TARGETS migraphx_py INCLUDE include)

src/py/include/migraphx/py.hpp

@@ -26,11 +26,12 @@
 
 #include <migraphx/config.hpp>
 #include <migraphx/program.hpp>
+#include <migraphx/py/export.h>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-program load_py(const std::string& filename);
+MIGRAPHX_PY_EXPORT program load_py(const std::string& filename);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/py/py_loader.cpp

@@ -35,9 +35,9 @@ static std::vector<fs::path> find_available_python_versions()
     auto path = dynamic_loader::path(&load_py).parent_path();
     for(const auto& entry : fs::directory_iterator{path})
     {
-        if(not entry.is_regular_file())
-            continue;
         auto p = entry.path();
+        if(not fs::is_regular_file(p))
+            continue;
         if(not contains(p.stem().string(), "migraphx_py_"))
             continue;
         result.push_back(p);
@@ -64,7 +64,7 @@ static dynamic_loader py_lib()
     return lib;
 }
 
-program load_py(const std::string& filename)
+MIGRAPHX_PY_EXPORT program load_py(const std::string& filename)
 {
     static auto f = py_lib().get_function<program(const std::string&)>("migraphx_load_py");
     return f(filename);

src/rewrite_quantization.cpp

@@ -28,6 +28,7 @@
 #include <migraphx/tune_axis.hpp>
 #include <migraphx/program.hpp>
 #include <migraphx/shape.hpp>
+#include <migraphx/common.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -62,12 +63,9 @@ void apply_quantizelinear(module& m, instruction_ref ins)
         min_quant = qt.min();
     });
     auto s        = add_zero_point->get_shape();
-    std::vector<int> min_data(s.elements(), min_quant);
-    std::vector<int> max_data(s.elements(), max_quant);
-    auto min_arg = m.add_literal(literal(s, min_data));
-    auto max_arg = m.add_literal(literal(s, max_data));
-
-    auto saturate = m.insert_instruction(ins, make_op("clip"), add_zero_point, min_arg, max_arg);
+    auto min_arg  = m.add_literal(literal{shape{s.type()}, {min_quant}});
+    auto max_arg  = m.add_literal(literal{shape{s.type()}, {max_quant}});
+    auto saturate = insert_common_op(m, ins, make_op("clip"), {add_zero_point, min_arg, max_arg});
     m.replace_instruction(
         ins, make_op("convert", {{"target_type", ins->get_shape().type()}}), saturate);
 }

src/simplify_algebra.cpp

@@ -1095,8 +1095,9 @@ MIGRAPHX_PRED_MATCHER(horiz_conv_dot, instruction_ref ins)
         };
     };
     auto dots  = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("dot"));
+    auto qdots = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("quant_dot"));
     auto convs = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("convolution"));
-    return (dots >= 2 or convs >= 2);
+    return (dots >= 2 or convs >= 2 or qdots >= 2);
 }
 
 struct find_conv_dot_horiz_fusion
@@ -1110,7 +1111,7 @@ struct find_conv_dot_horiz_fusion
         auto pred = [](auto i, auto j) {
             if(i->get_operator() != j->get_operator())
                 return false;
-            if(not contains({"dot", "convolution"}, i->name()))
+            if(not contains({"quant_dot", "dot", "convolution"}, i->name()))
                 return true;
             auto x = i->inputs()[1]->get_shape().lens();
             auto y = j->inputs()[1]->get_shape().lens();
@@ -1118,7 +1119,7 @@ struct find_conv_dot_horiz_fusion
                 return false;
             // Check that non-axes match
             int axis = 1;
-            if(i->name() == "dot")
+            if(i->name() == "dot" or i->name() == "quant_dot")
             {
                 axis = x.size() - 1;
             }
@@ -1129,7 +1130,7 @@ struct find_conv_dot_horiz_fusion
             if(std::distance(start, last) < 2)
                 return;
             auto&& name = (*start)->name();
-            if(not contains({"dot", "convolution"}, name))
+            if(not contains({"quant_dot", "dot", "convolution"}, name))
                 return;
             auto op   = (*start)->get_operator();
             int group = 1;
@@ -1144,7 +1145,7 @@ struct find_conv_dot_horiz_fusion
                 start, last, std::back_inserter(args), [&](auto x) { return x->inputs().at(1); });
             int axis        = 1;
             int concat_axis = 0;
-            if(name == "dot")
+            if(name == "dot" or name == "quant_dot")
             {
                 axis        = int(args.front()->get_shape().lens().size() - 1);
                 concat_axis = axis;

src/sqlite.cpp

@@ -48,6 +48,7 @@ struct sqlite_impl
     template <class F>
     void exec(const char* sql, F f)
     {
+        // cppcheck-suppress constParameterPointer
         auto callback = [](void* obj, auto... xs) -> int {
             try
             {

src/targets/cpu/gemm.cpp

@@ -43,7 +43,11 @@ struct dnnl_gemm : dnnl_extend_op<dnnl_gemm, dnnl::matmul, op::dot>
                 MIGRAPHX_DNNL_PREFIX(ARG_BIAS)};
     }
 
-    void required(const check_shapes& cs) const { cs.not_broadcasted(); }
+    template <class T>
+    void required(const check_shapes<T>& cs) const
+    {
+        cs.not_broadcasted();
+    }
 
     dnnl::matmul::desc get_desc(const std::unordered_map<int, dnnl::memory::desc>& m) const
     {

src/targets/cpu/include/migraphx/cpu/dnnl.hpp

@@ -400,7 +400,11 @@ struct dnnl_extend_op : dnnl_op<Derived, Primitive>
     }
 
     // dnnl has some issues with non-packed inputs
-    void required(const check_shapes& cs) const { cs.packed_or_broadcasted(); }
+    template <class T>
+    void required(const check_shapes<T>& cs) const
+    {
+        cs.packed_or_broadcasted();
+    }
 
     std::string name() const { return "dnnl::" + op.name(); }
     shape compute_shape(std::vector<shape> inputs) const

src/targets/cpu/target.cpp

@@ -61,7 +61,7 @@ namespace cpu {
 
 std::string target::name() const { return "cpu"; }
 
-// cppcheck-suppress constParameter
+// cppcheck-suppress constParameterReference
 std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_options&) const
 {
     auto& ctx = any_cast<context>(gctx);

src/targets/gpu/CMakeLists.txt

@@ -123,6 +123,7 @@ add_library(migraphx_gpu
     lrn.cpp
     mlir.cpp
     multinomial.cpp
+    no_device.cpp
     nonzero.cpp
     pack_args.cpp
     pack_int8_args.cpp