Merge branch 'develop' into simplify_1_mul_div_ops

src/include/migraphx/operators.hpp

@@ -57,6 +57,7 @@
 #include <migraphx/op/exp.hpp>
 #include <migraphx/op/flatten.hpp>
 #include <migraphx/op/floor.hpp>
+#include <migraphx/op/fmod.hpp>
 #include <migraphx/op/gather.hpp>
 #include <migraphx/op/gathernd.hpp>
 #include <migraphx/op/get_tuple_elem.hpp>
@@ -79,6 +80,7 @@
 #include <migraphx/op/lstm.hpp>
 #include <migraphx/op/max.hpp>
 #include <migraphx/op/min.hpp>
+#include <migraphx/op/mod.hpp>
 #include <migraphx/op/mul.hpp>
 #include <migraphx/op/multibroadcast.hpp>
 #include <migraphx/op/neg.hpp>

src/include/migraphx/pad_calc.hpp

@@ -24,38 +24,36 @@
 #ifndef MIGRAPHX_GUARD_OPERATORS_PAD_CALC_HPP
 #define MIGRAPHX_GUARD_OPERATORS_PAD_CALC_HPP
 
-#include <utility>
+#include <migraphx/config.hpp>
 #include <cstdint>
 #include <vector>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-inline void calculate_padding(int64_t idx,
-                              std::vector<int64_t>& pads,
-                              int64_t input_dim,
-                              int64_t stride,
-                              int64_t dilation,
-                              int64_t weight_dim,
-                              bool is_same_upper = true)
-{
-    int64_t output_dim     = (input_dim + stride - 1) / stride; // round up result
-    int64_t new_weight_dim = weight_dim + (weight_dim - 1) * (dilation - 1);
-    int64_t pad =
-        std::max(static_cast<int64_t>(0), (output_dim - 1) * stride + new_weight_dim - input_dim);
-    auto pad_ndims = pads.size() / 2;
+void calculate_padding(int64_t idx,
+                       std::vector<int64_t>& pads,
+                       int64_t input_dim,
+                       int64_t stride,
+                       int64_t dilation,
+                       int64_t weight_dim,
+                       bool is_same_upper = true);
 
-    if(is_same_upper)
-    {
-        pads[idx]             = pad / 2;
-        pads[idx + pad_ndims] = pad - pad / 2;
-    }
-    else
-    {
-        pads[idx + pad_ndims] = pad / 2;
-        pads[idx]             = pad - pad / 2;
-    }
-}
+/*!
+ * Calculate the padding for auto_padding. Used for dynamic shapes
+ * where the padding calculation must be done at evaluation time.
+ * \param tensor_lens input tensor image shape
+ * \param k_lens weights kernel shape
+ * \param strides strides for the kernel
+ * \param dilations dilations for the kernel
+ * \param use_upper put odd padding on upper or lower side
+ * \return padding in the form of {x0_begin, x1_begin, ... x0_end , x1_end, ...}
+ */
+std::vector<std::size_t> calc_dyn_auto_pad(std::vector<std::size_t> tensor_lens,
+                                           std::vector<std::size_t> k_lens,
+                                           std::vector<std::size_t> strides,
+                                           std::vector<std::size_t> dilations,
+                                           bool use_upper = true);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/sqlite.hpp

+/*
+ * 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.
+ */
+#ifndef MIGRAPHX_GUARD_MIGRAPHX_SQLITE_HPP
+#define MIGRAPHX_GUARD_MIGRAPHX_SQLITE_HPP
+
+#include <migraphx/config.hpp>
+#include <migraphx/filesystem.hpp>
+#include <memory>
+#include <unordered_map>
+#include <vector>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct sqlite_impl;
+
+struct sqlite
+{
+    sqlite() = default;
+    static sqlite read(const fs::path& p);
+    static sqlite write(const fs::path& p);
+    std::vector<std::unordered_map<std::string, std::string>> execute(const std::string& s);
+
+    private:
+    std::shared_ptr<sqlite_impl> impl;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_MIGRAPHX_SQLITE_HPP

src/include/migraphx/stringutils.hpp

@@ -174,27 +174,27 @@ inline std::string interpolate_string(const std::string& input,
 }
 
 template <class Iterator>
-inline std::string to_string_range(Iterator start, Iterator last)
+inline std::string to_string_range(Iterator start, Iterator last, const char* delim = ", ")
 {
     std::stringstream ss;
     if(start != last)
     {
         ss << *start;
-        std::for_each(std::next(start), last, [&](auto&& x) { ss << ", " << x; });
+        std::for_each(std::next(start), last, [&](auto&& x) { ss << delim << x; });
     }
     return ss.str();
 }
 
 template <class Range>
-inline std::string to_string_range(const Range& r)
+inline std::string to_string_range(const Range& r, const char* delim = ", ")
 {
-    return to_string_range(r.begin(), r.end());
+    return to_string_range(r.begin(), r.end(), delim);
 }
 
 template <class T>
-inline std::string to_string_range(const std::initializer_list<T>& r)
+inline std::string to_string_range(const std::initializer_list<T>& r, const char* delim = ", ")
 {
-    return to_string_range(r.begin(), r.end());
+    return to_string_range(r.begin(), r.end(), delim);
 }
 
 template <class T>

src/insert_pad.cpp

@@ -40,6 +40,12 @@ static void update_op(const instruction_ref& input, const instruction_ref& ins,
     auto val        = op.to_value();
     auto op_padding = val.at("padding").to_vector<size_t>();
 
+    // skip if shape is dynamic
+    if(input->get_shape().dynamic())
+    {
+        return;
+    }
+
     auto kdims = input->get_shape().lens().size() - 2;
     if(std::equal(op_padding.begin(),
                   op_padding.begin() + kdims,

src/instruction.cpp

@@ -445,8 +445,8 @@ operation instruction::normalized_operator() const
     operation o = this->get_operator();
     if(this->need_normalization())
     {
-        auto lens = this->inputs().front()->get_shape().lens();
-        if(!normalize_attributes(o, lens))
+        auto s = this->inputs().front()->get_shape();
+        if(!normalize_attributes(o, s.max_lens()))
             return this->get_operator();
     }
     return o;

src/normalize_ops.cpp

@@ -43,9 +43,9 @@ void normalize_ops::apply(module& m) const
         if(inputs.empty())
             continue;
 
-        auto lens                    = inputs[0]->get_shape().lens();
+        auto s                       = inputs[0]->get_shape();
         migraphx::operation tuned_op = ins->get_operator();
-        if(normalize_attributes(tuned_op, lens))
+        if(normalize_attributes(tuned_op, s.max_lens()))
         {
             m.replace_instruction(ins, tuned_op, inputs);
             ins->set_normalized();

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

@@ -119,6 +119,7 @@ struct onnx_parser
 };
 
 shape::type_t get_type(int dtype);
+bool is_type_float(shape::type_t dtype);
 
 } // namespace onnx
 } // namespace MIGRAPHX_INLINE_NS

src/onnx/onnx_parser.cpp

@@ -28,7 +28,6 @@
 #include <migraphx/stringutils.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/instruction.hpp>
-#include <migraphx/pad_calc.hpp>
 #include <migraphx/common.hpp>
 #include <migraphx/type_traits.hpp>
 #include <migraphx/float_equal.hpp>
@@ -60,7 +59,7 @@ create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, const
         std::accumulate(dims.begin(), dims.end(), std::size_t(1), std::multiplies<std::size_t>());
     if(elem_num == 0)
     {
-        return {};
+        return literal{shape_type};
     }
 
     // in case of scalar constants in onnx file, use dims=1 to fill initializer data
@@ -77,7 +76,7 @@ static literal create_literal(shape::type_t shape_type, const std::vector<size_t
         std::accumulate(dims.begin(), dims.end(), std::size_t(1), std::multiplies<std::size_t>());
     if(elem_num == 0)
     {
-        return {};
+        return literal{shape_type};
     }
 
     // scalar input
@@ -514,6 +513,16 @@ shape::type_t get_type(int dtype)
     }
 }
 
+bool is_type_float(shape::type_t dtype)
+{
+    bool r = false;
+    if(dtype == shape::float_type || dtype == shape::double_type || dtype == shape::half_type)
+    {
+        r = true;
+    }
+    return r;
+}
+
 } // namespace onnx
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/onnx/parse_constant.cpp

@@ -43,7 +43,7 @@ struct parse_constant : op_parser<parse_constant>
         // return empty literal
         if(v.get_shape().elements() == 0)
         {
-            return info.add_literal(literal{});
+            return info.add_literal(literal{v.get_shape().type()});
         }
 
         auto dim_size = info.attributes.at("value").t().dims_size();

src/onnx/parse_convolution.cpp

@@ -47,15 +47,17 @@ struct parse_convolution : op_parser<parse_convolution>
                           onnx_parser::node_info info,
                           std::vector<instruction_ref> args) const
     {
-        auto op      = make_op(opd.op_name);
-        auto values  = op.to_value();
-        auto l0      = args[0];
-        auto weights = args[1];
-        auto in_lens = l0->get_shape().lens();
+        auto op       = make_op(opd.op_name);
+        auto values   = op.to_value();
+        auto l0       = args[0];
+        auto weights  = args[1];
+        auto l0_shape = l0->get_shape();
+        auto w_shape  = weights->get_shape();
+        auto in_lens  = l0_shape.max_lens();
         assert(in_lens.size() > 2);
         auto kdims = in_lens.size() - 2;
 
-        // ensure pads availabe only when auto_pad is "NOT_SET"
+        // ensure pads available only when auto_pad is "NOT_SET"
         check_padding_mode(info, "CONV");
 
         if(contains(info.attributes, "strides"))
@@ -79,21 +81,65 @@ struct parse_convolution : op_parser<parse_convolution>
             copy(info.attributes["pads"].ints(), std::back_inserter(padding));
             check_attr_sizes(kdims, padding.size() / 2, "PARSE_CONV: inconsistent paddings");
         }
-
         if(contains(info.attributes, "auto_pad"))
         {
-            auto weight_lens = weights->get_shape().lens();
-            std::vector<std::size_t> k_lens(weight_lens.begin() + 2, weight_lens.end());
-            cal_auto_padding_size(info,
-                                  values,
-                                  k_lens,
-                                  values["dilation"].to_vector<std::size_t>(),
-                                  in_lens,
-                                  padding);
-            auto auto_pad = info.attributes["auto_pad"].s();
+            bool is_same_padding = false;
+            auto auto_pad        = info.attributes["auto_pad"].s();
             if(auto_pad.find("SAME") != std::string::npos)
             {
-                values["padding_mode"] = to_value(op::padding_mode_t::same);
+                is_same_padding = true;
+            }
+
+            // check if image shape is dynamic
+            bool image_shape_dynamic = false;
+            if(l0_shape.dynamic())
+            {
+                auto dyn_dims = l0_shape.dyn_dims();
+                std::for_each(dyn_dims.begin() + 2, dyn_dims.end(), [&](auto dyn_dim) {
+                    if(not dyn_dim.is_fixed())
+                    {
+                        image_shape_dynamic = true;
+                    }
+                });
+            }
+
+            // check if kernel shape is dynamic
+            bool kernel_shape_dynamic = false;
+            if(w_shape.dynamic())
+            {
+                auto dyn_dims = w_shape.dyn_dims();
+                std::for_each(dyn_dims.begin() + 2, dyn_dims.end(), [&](auto dyn_dim) {
+                    if(not dyn_dim.is_fixed())
+                    {
+                        kernel_shape_dynamic = true;
+                    }
+                });
+            }
+
+            if(is_same_padding)
+            {
+                if(image_shape_dynamic or kernel_shape_dynamic)
+                {
+                    // must calculate "same" padding with input shape data
+                    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);
+                    values["use_dynamic_same_auto_pad"] = true;
+                }
+                else
+                {
+                    values["padding_mode"] = to_value(op::padding_mode_t::same);
+                    // kernel shape will be fixed, so max_lens() == min_len() for kernel lengths
+                    auto weight_lens = weights->get_shape().max_lens();
+                    std::vector<std::size_t> k_lens(weight_lens.begin() + 2, weight_lens.end());
+                    cal_auto_padding_size(info,
+                                          values,
+                                          k_lens,
+                                          values["dilation"].to_vector<std::size_t>(),
+                                          in_lens,
+                                          padding);
+                }
             }
         }
         values["padding"] = std::vector<size_t>(padding.begin(), padding.end());

src/onnx/parse_if.cpp

@@ -47,7 +47,8 @@ struct parse_if : op_parser<parse_if>
 
         if(args.front()->get_shape().elements() != 1)
         {
-            MIGRAPHX_THROW("PARSE_IF: condition input can have only one element!");
+            MIGRAPHX_THROW("PARSE_IF: " + info.name +
+                           " condition input can have only one element!");
         }
 
         std::string then_name = info.name + "_if";
@@ -69,7 +70,8 @@ struct parse_if : op_parser<parse_if>
                           else_out_shapes.begin(),
                           else_out_shapes.end()))
         {
-            MIGRAPHX_THROW("PARSE_IF: then and else sub_grahps must have same output shapes!");
+            MIGRAPHX_THROW("PARSE_IF: " + info.name +
+                           " then and else sub_grahps must have same output shapes!");
         }
 
         auto if_ret = info.add_instruction(make_op("if"), args, {then_mdl, else_mdl});

src/onnx/parse_instancenorm.cpp

@@ -32,9 +32,12 @@ namespace onnx {
 
 struct parse_instancenorm : op_parser<parse_instancenorm>
 {
+    const std::set<shape::type_t> valid_types = {
+        shape::float_type, shape::half_type, shape::double_type};
+
     std::vector<op_desc> operators() const { return {{"InstanceNormalization"}}; }
 
-    instruction_ref parse(const op_desc& /*opd*/,
+    instruction_ref parse(const op_desc& opd,
                           const onnx_parser& parser,
                           onnx_parser::node_info info,
                           std::vector<instruction_ref> args) const
@@ -52,6 +55,11 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
         auto scale = args[1];
         auto bias  = args[2];
         auto dims  = x->get_shape().lens();
+        auto dtype = x->get_shape().type();
+        if(not contains(valid_types, dtype))
+            MIGRAPHX_THROW(opd.op_name + ": invalid output type: " + std::to_string(dtype) +
+                           ". Valid types are 1 (float), 10 (half), and 11 (double).");
+
         auto ndims = dims.size();
         assert(ndims >= 2);
         auto kdims = ndims - 2;
@@ -65,7 +73,7 @@ struct parse_instancenorm : op_parser<parse_instancenorm>
         auto l0              = info.add_instruction(make_op("sqdiff"), x, mean_bcast);
         auto variance        = info.add_instruction(make_op("reduce_mean", {{"axes", axes}}), l0);
         auto l1              = info.add_instruction(make_op("sub"), x, mean_bcast);
-        auto epsilon_literal = info.add_literal(epsilon);
+        auto epsilon_literal = info.add_literal(literal{shape{dtype}, {epsilon}});
         auto epsilon_bcast =
             info.add_instruction(make_op("multibroadcast", {{"out_lens", dims}}), epsilon_literal);
         auto variance_bcast =

src/onnx/parse_mod.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/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/make_op.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+struct parse_mod : op_parser<parse_mod>
+{
+    std::vector<op_desc> operators() const { return {{"Mod"}}; }
+
+    instruction_ref parse(const op_desc& /*opd*/,
+                          const onnx_parser& parser,
+                          onnx_parser::node_info info,
+                          std::vector<instruction_ref> args) const
+    {
+        std::string mod = "mod";
+        if(is_type_float(args[0]->get_shape().type()) || is_type_float(args[1]->get_shape().type()))
+        {
+            if(!contains(info.attributes, "fmod"))
+            {
+                MIGRAPHX_THROW("Mod operator with float args and fmod=0 invalid");
+            }
+        }
+
+        if(contains(info.attributes, "fmod"))
+        {
+            if(parser.parse_value(info.attributes.at("fmod")).at<int>() == 1)
+            {
+                mod = "fmod";
+            }
+        }
+        return info.add_common_op(mod, args[0], args[1]);
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/pad_calc.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/pad_calc.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+void calculate_padding(int64_t idx,
+                       std::vector<int64_t>& pads,
+                       int64_t input_dim,
+                       int64_t stride,
+                       int64_t dilation,
+                       int64_t weight_dim,
+                       bool is_same_upper)
+{
+    int64_t output_dim     = (input_dim + stride - 1) / stride; // round up result
+    int64_t new_weight_dim = weight_dim + (weight_dim - 1) * (dilation - 1);
+    int64_t pad =
+        std::max(static_cast<int64_t>(0), (output_dim - 1) * stride + new_weight_dim - input_dim);
+    auto pad_ndims = pads.size() / 2;
+
+    if(is_same_upper)
+    {
+        pads[idx]             = pad / 2;
+        pads[idx + pad_ndims] = pad - pad / 2;
+    }
+    else
+    {
+        pads[idx + pad_ndims] = pad / 2;
+        pads[idx]             = pad - pad / 2;
+    }
+}
+
+std::vector<std::size_t> calc_dyn_auto_pad(std::vector<std::size_t> tensor_lens,
+                                           std::vector<std::size_t> k_lens,
+                                           std::vector<std::size_t> strides,
+                                           std::vector<std::size_t> dilations,
+                                           bool use_upper)
+{
+    std::vector<std::size_t> padding;
+    padding.resize(2 * k_lens.size());
+    for(size_t i = 0; i < padding.size() / 2; i++)
+    {
+        std::ptrdiff_t input_dim      = tensor_lens[i];
+        std::ptrdiff_t stride         = strides[i];
+        std::ptrdiff_t weight_dim     = k_lens[i];
+        std::ptrdiff_t dilation       = dilations[i];
+        std::ptrdiff_t output_dim     = (input_dim + stride - 1) / stride; // round up result
+        std::ptrdiff_t new_weight_dim = weight_dim + (weight_dim - 1) * (dilation - 1);
+        std::size_t pad               = std::max(static_cast<std::ptrdiff_t>(0),
+                                   (output_dim - 1) * stride + new_weight_dim - input_dim);
+        auto pad_ndims                = padding.size() / 2;
+
+        if(use_upper)
+        {
+            padding[i]             = pad / 2;
+            padding[i + pad_ndims] = pad - pad / 2;
+        }
+        else
+        {
+            padding[i + pad_ndims] = pad / 2;
+            padding[i]             = pad - pad / 2;
+        }
+    }
+    return padding;
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/program.cpp

@@ -307,9 +307,12 @@ std::vector<argument> generic_eval(const module* mod,
                     if(not contains(params, param_name))
                         MIGRAPHX_THROW("Parameter not found: " + param_name);
                     auto param = params[param_name];
-                    if(param.get_shape() != ins->get_shape())
+                    // TODO: may want to check correct number of dimensions and/or was within bounds
+                    if(not ins->get_shape().dynamic() and param.get_shape() != ins->get_shape())
+                    {
                         MIGRAPHX_THROW("Incorrect shape {" + to_string(param.get_shape()) +
                                        "} for parameter: " + param_name);
+                    }
                     return param;
                 }));
         }
@@ -352,7 +355,10 @@ std::vector<argument> generic_eval(const module* mod,
                             }));
         }
         assert(results.find(ins) != results.end());
-        assert(results.at(ins).get_shape() == ins->get_shape());
+        if(not ins->get_shape().dynamic())
+        {
+            assert(results.at(ins).get_shape() == ins->get_shape());
+        }
     }
     return {results.at(std::prev(mod->end()))};
 }

src/sqlite.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/sqlite.hpp>
+#include <migraphx/manage_ptr.hpp>
+#include <migraphx/errors.hpp>
+#include <sqlite3.h>
+#include <algorithm>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+using sqlite3_ptr = MIGRAPHX_MANAGE_PTR(sqlite3*, sqlite3_close);
+
+struct sqlite_impl
+{
+    sqlite3* get() const { return ptr.get(); }
+    void open(const fs::path& p, int flags)
+    {
+        sqlite3* ptr_tmp = nullptr;
+        int rc           = sqlite3_open_v2(p.string().c_str(), &ptr_tmp, flags, nullptr);
+        ptr              = sqlite3_ptr{ptr_tmp};
+        if(rc != 0)
+            MIGRAPHX_THROW("error opening " + p.string() + ": " + error_message());
+    }
+
+    template <class F>
+    void exec(const char* sql, F f)
+    {
+        auto callback = [](void* obj, auto... xs) -> int {
+            try
+            {
+                const auto* g = static_cast<const F*>(obj);
+                (*g)(xs...);
+                return 0;
+            }
+            catch(...)
+            {
+                return -1;
+            }
+        };
+        int rc = sqlite3_exec(get(), sql, callback, &f, nullptr);
+        if(rc != 0)
+            MIGRAPHX_THROW(error_message());
+    }
+
+    std::string error_message() const
+    {
+        std::string msg = "sqlite3: ";
+        return msg + sqlite3_errmsg(get());
+    }
+    sqlite3_ptr ptr;
+};
+
+sqlite sqlite::read(const fs::path& p)
+{
+    sqlite r;
+    r.impl = std::make_shared<sqlite_impl>();
+    r.impl->open(p, SQLITE_OPEN_READONLY);
+    return r;
+}
+
+sqlite sqlite::write(const fs::path& p)
+{
+    sqlite r;
+    r.impl = std::make_shared<sqlite_impl>();
+    // Using '+' instead of bitwise '|' to avoid compilation warning
+    r.impl->open(p, SQLITE_OPEN_READWRITE + SQLITE_OPEN_CREATE);
+    return r;
+}
+
+std::vector<std::unordered_map<std::string, std::string>> sqlite::execute(const std::string& s)
+{
+    std::vector<std::unordered_map<std::string, std::string>> result;
+    impl->exec(s.c_str(), [&](int n, char** texts, char** names) {
+        std::unordered_map<std::string, std::string> row;
+        row.reserve(n);
+        std::transform(
+            names,
+            names + n,
+            texts,
+            std::inserter(row, row.begin()),
+            [&](const char* name, const char* text) { return std::make_pair(name, text); });
+        result.push_back(row);
+    });
+    return result;
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/CMakeLists.txt

@@ -184,6 +184,7 @@ add_library(migraphx_gpu
     pack_int8_args.cpp
     prefuse_ops.cpp
     pad.cpp
+    perfdb.cpp
     pooling.cpp
     quant_convolution.cpp
     reverse.cpp

src/targets/gpu/hip.cpp

@@ -23,13 +23,13 @@
  */
 
 #include <migraphx/gpu/hip.hpp>
-
 #include <migraphx/manage_ptr.hpp>
 #include <migraphx/register_op.hpp>
 #include <migraphx/gpu/context.hpp>
 #include <migraphx/gpu/device/contiguous.hpp>
 #include <miopen/miopen.h>
-
+#include <memory>
+#include <mutex>
 #include <vector>
 
 namespace migraphx {
@@ -77,12 +77,38 @@ void* get_device_ptr(void* hptr)
     return result;
 }
 
-hip_ptr allocate_gpu(std::size_t sz, bool host = false)
+struct host_ptr_cache
+{
+    std::unordered_map<void*, std::weak_ptr<void>> cache;
+    std::mutex m;
+    std::shared_ptr<void> get(void* ptr)
+    {
+        std::lock_guard<std::mutex> lock(m);
+        auto it = cache.find(ptr);
+        if(it != cache.end())
+            return it->second.lock();
+        return nullptr;
+    }
+
+    void put(const std::shared_ptr<void>& p)
+    {
+        std::lock_guard<std::mutex> lock(m);
+        cache[p.get()] = p;
+    }
+};
+
+static host_ptr_cache& get_host_ptr_cache()
+{
+    static host_ptr_cache cache;
+    return cache;
+}
+
+std::shared_ptr<void> allocate_gpu(std::size_t sz, bool host = false)
 {
     if(sz > get_available_gpu_memory())
         MIGRAPHX_THROW("Memory not available to allocate buffer: " + std::to_string(sz));
-    void* result = nullptr;
-    auto status  = host ? hipHostMalloc(&result, sz) : hipMalloc(&result, sz);
+    void* alloc_ptr = nullptr;
+    auto status     = host ? hipHostMalloc(&alloc_ptr, sz) : hipMalloc(&alloc_ptr, sz);
     if(status != hipSuccess)
     {
         if(host)
@@ -90,16 +116,28 @@ hip_ptr allocate_gpu(std::size_t sz, bool host = false)
         else
             return allocate_gpu(sz, true);
     }
-    assert(result != nullptr);
-    return hip_ptr{result};
+    assert(alloc_ptr != nullptr);
+    std::shared_ptr<void> result = share(hip_ptr{alloc_ptr});
+    if(host)
+    {
+        get_host_ptr_cache().put(result);
+    }
+    return result;
 }
 
-hip_host_ptr register_on_gpu(void* ptr, std::size_t sz)
+std::shared_ptr<void> register_on_gpu(void* ptr, std::size_t sz)
 {
+    std::shared_ptr<void> result = get_host_ptr_cache().get(ptr);
+    if(result)
+    {
+        return result;
+    }
     auto status = hipHostRegister(ptr, sz, hipHostRegisterMapped);
     if(status != hipSuccess)
         MIGRAPHX_THROW("Gpu register failed: " + hip_error(status));
-    return hip_host_ptr{ptr};
+    result = share(hip_host_ptr{ptr});
+    get_host_ptr_cache().put(result);
+    return result;
 }
 
 template <class T>
@@ -115,7 +153,7 @@ std::vector<T> read_from_gpu(const void* x, std::size_t sz)
     return result;
 }
 
-hip_ptr write_to_gpu(const void* x, std::size_t sz, bool host = false)
+std::shared_ptr<void> write_to_gpu(const void* x, std::size_t sz, bool host = false)
 {
     gpu_sync();
     auto result = allocate_gpu(sz, host);
@@ -137,22 +175,21 @@ hip_ptr write_to_gpu(const T& x)
 
 argument allocate_gpu(const shape& s, bool host)
 {
-    auto p = share(allocate_gpu(s.bytes() + 1, host));
+    auto p = allocate_gpu(s.bytes() + 1, host);
     return {s, [p]() mutable { return reinterpret_cast<char*>(p.get()); }};
 }
 
 argument register_on_gpu(const argument& arg)
 {
     auto arg_shared = arg.share();
-    auto p          = share(register_on_gpu(arg_shared.data(), arg_shared.get_shape().bytes()));
-    return {arg_shared.get_shape(), [p, a = std::move(arg_shared)]() mutable {
-                return get_device_ptr(p.get());
-            }}; // namespace gpu
-} // namespace MIGRAPHX_INLINE_NS
+    auto p          = register_on_gpu(arg_shared.data(), arg_shared.get_shape().bytes());
+    return {arg_shared.get_shape(),
+            [p, a = std::move(arg_shared)]() mutable { return get_device_ptr(p.get()); }};
+}
 
 argument to_gpu(const argument& arg, bool host)
 {
-    auto p = share(write_to_gpu(arg.data(), arg.get_shape().bytes(), host));
+    auto p = write_to_gpu(arg.data(), arg.get_shape().bytes(), host);
     return {arg.get_shape(), p};
 }
 

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

+/*
+ * 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.
+ */
+#ifndef MIGRAPHX_GUARD_GPU_PERFDB_HPP
+#define MIGRAPHX_GUARD_GPU_PERFDB_HPP
+
+#include <migraphx/config.hpp>
+#include <migraphx/shape.hpp>
+#include <migraphx/operation.hpp>
+#include <string>
+#include <vector>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace gpu {
+
+struct problem_params
+{
+    operation op;
+    std::vector<shape> inputs;
+    shape output;
+};
+
+std::string get_mlir_perf_for_conv(const problem_params& pp);
+
+} // namespace gpu
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_GPU_PERFDB_HPP