Merge branch 'develop' into add_parity_check_ci

src/include/migraphx/dynamic_loader.hpp

@@ -26,6 +26,7 @@
 
 #include <migraphx/config.hpp>
 #include <migraphx/filesystem.hpp>
+#include <migraphx/optional.hpp>
 #include <functional>
 #include <memory>
 #include <vector>
@@ -43,6 +44,9 @@ struct MIGRAPHX_EXPORT dynamic_loader
         return path(reinterpret_cast<void*>(address));
     }
     static fs::path path(void* address);
+
+    static optional<dynamic_loader> try_load(const fs::path& p);
+
     dynamic_loader() = default;
 
     dynamic_loader(const fs::path& p);

src/include/migraphx/marker.hpp

@@ -80,7 +80,7 @@ struct marker
     {
         using std::swap;
         auto* derived = this->any_cast<PrivateDetailTypeErasedT>();
-        if(derived and private_detail_te_handle_mem_var.unique())
+        if(derived and private_detail_te_handle_mem_var.use_count() == 1)
         {
             *derived = std::forward<PrivateDetailTypeErasedT>(value);
         }
@@ -233,7 +233,7 @@ struct marker
     private_detail_te_handle_base_type& private_detail_te_get_handle()
     {
         assert(private_detail_te_handle_mem_var != nullptr);
-        if(not private_detail_te_handle_mem_var.unique())
+        if(private_detail_te_handle_mem_var.use_count() > 1)
             private_detail_te_handle_mem_var = private_detail_te_handle_mem_var->clone();
         return *private_detail_te_handle_mem_var;
     }

src/include/migraphx/matcher.hpp

@@ -381,22 +381,24 @@ void find_matches_for(source_location location, Mod& mod, instruction_ref ins, M
     const int trace         = value_of(MIGRAPHX_TRACE_MATCHES{});
     const bool validate     = enabled(MIGRAPHX_VALIDATE_MATCHES{});
     const auto trace_filter = string_value_of(MIGRAPHX_TRACE_MATCHES_FOR{});
-    const bool trace_for    = not trace_filter.empty() and
-                           (contains(std::string{location.file_name()}, trace_filter) or
-                            contains(std::string{location.function_name()}, trace_filter));
-    bool match = false;
+    bool match              = false;
     each_args(
         [&](auto&& m) {
+            const auto& matcher_name = get_type_name(m);
+            const bool trace_for     = not trace_filter.empty() and
+                                   (contains(std::string{location.file_name()}, trace_filter) or
+                                    contains(std::string{location.function_name()}, trace_filter) or
+                                    contains(matcher_name, trace_filter));
             if(match)
                 return;
-            if(trace > 1 or trace_for)
-                std::cout << "Match: " << get_type_name(m) << std::endl;
+            if(trace > 1 and trace_for)
+                std::cout << "Match: " << matcher_name << std::endl;
             auto r = match_instruction(get_module(mod), ins, m.matcher());
             if(r.result == get_module(mod).end())
                 return;
             if(trace > 0 or trace_for)
             {
-                std::cout << "Matched by " << get_type_name(m) << std::endl;
+                std::cout << "Matched by " << matcher_name << std::endl;
                 get_module(mod).debug_print(ins);
             }
             // If its already invalid dont validate it again
@@ -407,7 +409,7 @@ void find_matches_for(source_location location, Mod& mod, instruction_ref ins, M
                 auto invalid = get_module(mod).validate();
                 if(invalid != get_module(mod).end())
                 {
-                    std::cout << "Invalid program from match: " << get_type_name(m) << std::endl;
+                    std::cout << "Invalid program from match: " << matcher_name << std::endl;
                     std::cout << "Invalid instructions: " << std::endl;
                     get_module(mod).debug_print(invalid->inputs());
                     get_module(mod).debug_print(invalid);
@@ -621,6 +623,8 @@ MIGRAPHX_PRED_MATCHER(broadcast, instruction_ref ins)
 template <class... Ms>
 auto skip(Ms... ms)
 {
+    static_assert(((not std::is_convertible<Ms, std::string>{}) and ...),
+                  "Use a matcher not a string for skip.");
     auto m = any_of(ms...);
     return make_basic_fun_matcher([=](matcher_context& ctx, instruction_ref start) {
         return fix<optional<instruction_ref>>(

src/include/migraphx/module.hpp

@@ -222,7 +222,17 @@ struct MIGRAPHX_EXPORT module
     void annotate(std::ostream& os, std::function<void(instruction_ref)> a) const;
 
     std::vector<module_ref> get_sub_modules(bool shallow = false) const;
+    /* sorts the module in topological order aka reverse-post order (RPO) DFS order
+       it takes last instruction or @return as the root and walks back the graph and moves inputs
+       of the each instruction such that it appears before the instruction itself.
+    */
     module& sort();
+    /* Any instruction "X" can have module arguments and those modules inside them can use any other
+     * instruction "Y" from predecessor modules of the instruction "X". Such instruction "Y" inside
+     * module args are not listed as input instructions to "X". But those instructions "Y" must be
+     * evaluted before the instruction "X" can. Therefore such "Y" instructions are considered
+     * implicit dependency to "X".
+     */
     ins_dep_map calc_implicit_deps() const;
 
     MIGRAPHX_EXPORT friend std::ostream& operator<<(std::ostream& os, const module& m);

src/include/migraphx/normalize_attributes.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -28,6 +28,7 @@
 #include <migraphx/shape.hpp>
 #include <cstring>
 #include <vector>
+#include <migraphx/op/normalize_attribute.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -42,6 +43,36 @@ struct select_dependent_type
 template <class T, class... Ts>
 using dependent_type = typename select_dependent_type<T, Ts...>::type;
 
+/**
+ * Used to normalize variable input axes at model runtime.
+ * Example: the axes inputs of the slice operator.
+ *
+ * \param axes the axes to normalize
+ * \param input_shape shape of the input tensor
+ * \param attr_val the normalize_axes attributes from the operator
+ * \param prefix error message prefix
+ */
+std::vector<int64_t> normalize_axes(const std::vector<int64_t>& axes,
+                                    const shape& input_shape,
+                                    const value& attr_val,
+                                    const std::string& prefix = "");
+
+/**
+ * Used to normalize variable input axes at model runtime.
+ * Example: the starts and ends inputs of the slice operator.
+ *
+ * \param indices the indices to normalize
+ * \param axes which axes the indices apply over
+ * \param input_shape shape of the input tensor
+ * \param attr_val the normalize_axes attributes from the operator
+ * \param prefix error message 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 = "");
+
 MIGRAPHX_EXPORT
 bool normalize_attributes(operation& op, const shape& input_shape);
 

src/include/migraphx/op/allocate.hpp

@@ -36,20 +36,48 @@ namespace op {
 struct allocate
 {
     shape s{};
+    // for dynamic allocate to set the buffer type
+    shape::type_t buf_type = shape::half_type;
+
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.s, "shape"));
+        return pack(f(self.s, "shape"), f(self.buf_type, "buf_type"));
     }
+
     std::string name() const { return "allocate"; }
+
     shape compute_shape(const std::vector<shape>& inputs) const
     {
-        migraphx::check_shapes{inputs, *this, true}.has(0);
-        return s;
+        migraphx::check_shapes{inputs, *this, true}.has(0, 1);
+        // check if shape attribute is not default
+        if(s != shape())
+        {
+            return s;
+        }
+        else
+        {
+            const auto& out_dims = inputs.at(0);
+            assert(not out_dims.dynamic());
+            assert(out_dims.ndim() == 1);
+            std::size_t max_val = std::numeric_limits<std::size_t>::max();
+            std::vector<shape::dynamic_dimension> dyn_dims(out_dims.lens().at(0),
+                                                           shape::dynamic_dimension{0, max_val});
+            return {buf_type, dyn_dims};
+        }
     }
-    argument compute(const shape& output_shape, const std::vector<argument>&) const
+    argument compute(const shape& output_shape, const std::vector<argument>& args) const
     {
-        return {output_shape};
+        if(args.empty())
+        {
+            return {output_shape};
+        }
+        else
+        {
+            std::vector<std::size_t> output_dims(output_shape.ndim());
+            args.at(0).visit([&](auto a) { output_dims.assign(a.begin(), a.end()); });
+            return {shape{buf_type, output_dims}};
+        }
     }
 };
 

src/include/migraphx/op/common.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -33,8 +33,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+// Specifies where to add the "extra" cell of padding if the
+// calculated padding is an odd number.
 // Padding mode is default_ for fixed shape padding.
-// same_lower and same_upper used for dynamic padding.
+// same_lower and same_upper specify dynamic padding.
+// The odd cell goes at the beginning of the dimension
+// (same_lower) or end (same_upper).
 enum padding_mode_t
 {
     default_, // NOLINT

src/include/migraphx/op/contiguous.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal

src/include/migraphx/op/convolution.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -82,7 +82,7 @@ struct convolution
         const auto input_ndim   = inputs[0].ndim();
         const auto padding_size = padding.size();
 
-        if(input_ndim != padding_size / 2 + 2 && input_ndim != padding_size + 2)
+        if(input_ndim != padding_size / 2 + 2 and input_ndim != padding_size + 2)
         {
             MIGRAPHX_THROW("CONVOLUTION: input and attribute size mismatch!");
         }
@@ -206,6 +206,7 @@ struct convolution
         std::vector<std::size_t> new_padding;
         if(padding_mode != op::padding_mode_t::default_)
         {
+            // auto-Calculate the padding sizes with calc_dyn_auto_pad
             auto input_lens   = args[0].get_shape().lens();
             auto weights_lens = args[1].get_shape().lens();
             new_padding =
@@ -217,6 +218,7 @@ struct convolution
         }
         else
         {
+            // Use the padding that was given
             new_padding = padding;
             if(output_shape.dynamic())
             {

src/include/migraphx/op/convolution_backwards.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -164,7 +164,7 @@ struct convolution_backwards
             shape win_shape{dyn_out.computed_shape.type(), win_size};
 
             par_dfor(in_n, wei_c)([&](int o, int k) {
-                shape_for_each(win_shape, [&](auto idx_win) {
+                shape_for_each(win_shape, [&](const auto& idx_win) {
                     const int w = idx_win[0];
 
                     auto input_dims_start = idx_win.begin() + 1;

src/include/migraphx/op/fill.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_OPERATORS_FILL_HPP
+#define MIGRAPHX_GUARD_OPERATORS_FILL_HPP
+
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/dyn_output.hpp>
+#include <migraphx/par_for.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+/**
+ * fill(default_value, output_buffer)
+ * Fill an output buffer with the given default_value.
+ * Note that if the default_value is a literal and the output_buffer
+ * has a static shape this operator can be replaced with a literal.
+ */
+struct fill
+{
+    std::string name() const { return "fill"; }
+
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs, *this, true}.has(2).same_type();
+        if(inputs.at(0).dynamic() or inputs.at(0).elements() != 1)
+        {
+            MIGRAPHX_THROW("FILL: default_value is dynamic or more than one element");
+        }
+        return inputs.back();
+    }
+
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
+    {
+        visit_all(args[0], args[1])([&](auto value, auto output) {
+            par_for(dyn_out.computed_shape.elements(), [&](auto i) { output[i] = value.front(); });
+        });
+        return args[1];
+    }
+
+    std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 1; }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/op/gather.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -125,13 +125,12 @@ struct gather
                     auto out_lens  = data.get_shape().lens();
                     out_lens[axis] = indices.get_shape().elements();
                     migraphx::shape out_comp_shape{data.get_shape().type(), out_lens};
-                    shape_for_each(out_comp_shape, [&](const auto& out_idx) {
-                        auto data_idx  = out_idx;
-                        auto in_index  = indices[data_idx[axis]];
-                        in_index       = (in_index < 0) ? in_index + axis_dim_size : in_index;
-                        data_idx[axis] = in_index;
-                        output[out_comp_shape.index(out_idx.begin(), out_idx.end())] =
-                            data(data_idx.begin(), data_idx.end());
+                    shape_for_each(out_comp_shape, [&](const auto& out_idx_v, size_t out_idx) {
+                        auto data_idx   = out_idx_v;
+                        auto in_index   = indices[data_idx[axis]];
+                        in_index        = (in_index < 0) ? in_index + axis_dim_size : in_index;
+                        data_idx[axis]  = in_index;
+                        output[out_idx] = data(data_idx.begin(), data_idx.end());
                     });
                 }
             });

src/include/migraphx/op/if_op.hpp

@@ -71,7 +71,7 @@ struct if_op
         std::unordered_map<std::string, argument> params;
 
         std::set<std::string> pnames;
-        for(const auto& smod : mods)
+        for(const_module_ref smod : mods)
         {
             auto names = smod->get_parameter_names();
             pnames.insert(names.begin(), names.end());

src/include/migraphx/op/loop.hpp

@@ -59,9 +59,9 @@ struct loop
             MIGRAPHX_THROW("LOOP: operator should have one submodule.");
         }
 
-        const auto& mod     = mods.front();
-        auto mod_out_shapes = mod->get_output_shapes();
-        auto dep_param_num  = inputs.size() - 2;
+        const_module_ref mod = mods.front();
+        auto mod_out_shapes  = mod->get_output_shapes();
+        auto dep_param_num   = inputs.size() - 2;
 
         // first item of the mod output shapes is condition used in loop,
         // which is not needed to compute output shape

src/include/migraphx/op/nonmaxsuppression.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -258,7 +258,7 @@ struct nonmaxsuppression
         selected_boxes_inside_class.reserve(max_output_shape.elements());
         // iterate over batches and classes
         shape comp_s{shape::double_type, {num_batches, num_classes}};
-        shape_for_each(comp_s, [&](auto idx) {
+        shape_for_each(comp_s, [&](const auto& idx) {
             auto batch_idx = idx[0];
             auto class_idx = idx[1];
             // index offset for this class

src/include/migraphx/op/nonzero.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -56,10 +56,10 @@ struct nonzero
         std::vector<std::vector<std::size_t>> vec_idx;
         auto s = args.front().get_shape();
         args.front().visit([&](auto v) {
-            shape_for_each(s, [&](auto idx) {
-                if(not float_equal(v[s.index(idx)], 0))
+            shape_for_each(s, [&](const auto& idx_v, size_t idx) {
+                if(not float_equal(v[idx], 0))
                 {
-                    vec_idx.push_back(idx);
+                    vec_idx.push_back(idx_v);
                 }
             });
         });

src/include/migraphx/op/pooling.hpp

@@ -29,6 +29,7 @@
 #include <migraphx/config.hpp>
 #include <migraphx/value.hpp>
 #include <migraphx/argument.hpp>
+#include <migraphx/pad_calc.hpp>
 #include <migraphx/par_for.hpp>
 #include <migraphx/shape_for_each.hpp>
 #include <migraphx/dyn_output.hpp>
@@ -40,10 +41,20 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+// The Pooling operator mostly follows the specifications for the Onnx pooling op.
+// It assumes an NCHW layout, extended to support any number of spatial dimensions
+// from 1 on up; dimensions are <batch index, channels, spatial dimensions...>
+//
 struct pooling
 {
+    //  Class members mode, ceil_mode, padding_mode have similar names but refer to separate
+    //  concepts.
     pooling_mode mode = {pooling_mode::average};
 
+    // If the input has rank other than 4 then padding, stride, lengths must all be specified
+    // since the defaults have 2-dimensions.  Exception: padding not required if
+    // padding_mode != default_
+
     // Padding along each spatial input dimension
     // Can be ndim or 2*ndim values where ndim is size of lengths
     // ndim values means pad the same before and after each dimension
@@ -63,13 +74,14 @@ struct pooling
 
     // ceiling mode is a flag affecting output size
     // or equivalently, placements of the pooling kernel.
-    // When true, round the size upwards, possibly
-    // including partial placements where the kernel extends beyond the edge
-    // of input and even padding.  When false, round down so that all
+    // When true, round the size upwards.  When false, round down so that all
     // kernel placements fit but some input values may be dropped.
     bool ceil_mode = false;
     int lp_order   = 2;
 
+    // Mode for auto padding.  default_ indicates no auto padding.
+    padding_mode_t padding_mode = padding_mode_t::default_;
+
     // Global pooling with dynamic shape input
     bool dyn_global = false;
 
@@ -84,6 +96,7 @@ struct pooling
     {
         return pack(f(self.mode, "mode"),
                     f(self.padding, "padding"),
+                    f(self.padding_mode, "padding_mode"),
                     f(self.stride, "stride"),
                     f(self.lengths, "lengths"),
                     f(self.ceil_mode, "ceil_mode"),
@@ -97,7 +110,8 @@ struct pooling
     {
         if(dyn_global)
             return;
-        if((padding.size() != stride.size() and (padding.size()) != stride.size() * 2) or
+        if((padding_mode != default_ and padding.size() != stride.size() and
+            (padding.size()) != stride.size() * 2) or
            stride.size() != lengths.size())
         {
             MIGRAPHX_THROW("POOLING: inconsistent attribute sizes");
@@ -137,8 +151,19 @@ struct pooling
             std::size_t padding_factor = 2 * padding[i];
             if(padding.size() == 2 * kdims)
                 padding_factor = padding[i] + padding[i + kdims];
-            assert(input_lens[i + 2] + padding_factor >= lengths[i]);
-            std::size_t dim_size = input_lens[i + 2] + padding_factor - lengths[i];
+            std::size_t dim_size;
+            if(input_lens[i + 2] + padding_factor < lengths[i])
+            {
+                if(padding_mode == default_)
+                    MIGRAPHX_THROW("POOLING: not enough padding for the given kernel size");
+                // lengths can be legitimately larger only if we're doing auto padding
+                // with a dynamic shape, in which case given padding is ignored.  Set a dummy value.
+                dim_size = 2;
+            }
+            else
+            {
+                dim_size = input_lens[i + 2] + padding_factor - lengths[i];
+            }
             std::size_t len =
                 (ceil_mode)
                     ? dim_size / stride[i] +
@@ -151,17 +176,13 @@ struct pooling
 
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this, true}.has(1);
+        check_shapes{inputs, *this, true}.has(1).min_ndims(3);
         check_attribute_size();
 
         const shape& input = inputs.at(0);
-        auto padding_size  = padding.size();
+        auto stride_size   = stride.size();
         size_t kdims       = input.ndim() - 2;
-        if(input.ndim() < 3)
-        {
-            MIGRAPHX_THROW("POOLING: input must have 3 or more dimensions and be nonempty");
-        }
-        if(input.ndim() * 2 != padding_size + 4 and input.ndim() != padding_size + 2)
+        if(input.ndim() != stride_size + 2)
         {
             MIGRAPHX_THROW("POOLING: input and attribute size mismatch!");
         }
@@ -179,6 +200,28 @@ struct pooling
                 }
                 return {input.type(), output_dyn_dims};
             }
+            else if(padding_mode != default_)
+            {
+                const size_t num_spatial_dims = inputs[0].ndim() - 2;
+                const shape& x_shape          = inputs[0];
+                // same as convolution::dynamic_compute_shape()
+
+                for(std::size_t i = 0; i < num_spatial_dims; ++i)
+                {
+                    auto ceil_div = [](std::size_t x, std::size_t y) { return (x + y - 1) / y; };
+                    auto s        = stride[i];
+
+                    auto x = x_shape.dyn_dims()[i + 2];
+                    std::set<std::size_t> optimals{};
+                    std::transform(x.optimals.begin(),
+                                   x.optimals.end(),
+                                   std::inserter(optimals, optimals.begin()),
+                                   [&](auto o) { return ceil_div(o, s); });
+                    output_dyn_dims.push_back(
+                        shape::dynamic_dimension{ceil_div(x.min, s), ceil_div(x.max, s), optimals});
+                }
+                return {input.type(), output_dyn_dims};
+            }
             else
             {
                 // does not compute optimals
@@ -267,6 +310,7 @@ struct pooling
                       Out& output,
                       const In& input,
                       const std::vector<std::size_t>& kernel_dims,
+                      const std::vector<std::size_t>& padding_vals,
                       Op op) const
     {
         auto in_s    = input.get_shape();
@@ -283,9 +327,9 @@ struct pooling
             // For each spatial dimension, find starting and ending index of pooling kernel
             for(std::size_t dim = 2; dim < n_dim; ++dim)
             {
-                auto d_2 = dim - 2;
-                int start =
-                    static_cast<int>(idx_o[dim] * stride[d_2]) - static_cast<int>(padding[d_2]);
+                auto d_2  = dim - 2;
+                int start = static_cast<int>(idx_o[dim] * stride[d_2]) -
+                            static_cast<int>(padding_vals[d_2]);
                 int end;
                 // NOLINT
                 if(count_include_pad and ceil_mode and (mode != pooling_mode::max))
@@ -297,7 +341,7 @@ struct pooling
 
                     // Check if this kernel extends beyond the padding at end of dimension
                     end = std::min(start + kernel_dims[d_2],
-                                   in_lens[dim] + static_cast<int>(padding[d_2]));
+                                   in_lens[dim] + static_cast<int>(padding_vals[d_2]));
                 }
                 else
                 {
@@ -316,11 +360,12 @@ struct pooling
             }
 
             shape win_shape{output_shape.type(), win_size};
+
             auto pool_size    = win_shape.elements();
             double output_val = op.template init<Type>();
 
             // for each element in the window...
-            shape_for_each(win_shape, [&](auto idx_w) {
+            shape_for_each(win_shape, [&](const auto& idx_w) {
                 // the coordinates of this element
                 auto idx = idx_o;
 
@@ -354,30 +399,65 @@ struct pooling
 
     argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
-        argument result{dyn_out.computed_shape};
+        argument result;
         auto input_lens = args[0].get_shape().lens();
         std::vector<std::size_t> kernel_dims;
+        shape output_shape;
+        // If we have to auto-calculate padding, it will be passed to calc_pooling() as an argument
+        // instead of the member variable padding.
+        std::vector<std::size_t> temp_padding(padding);
         if(dyn_global)
         {
+            // for dynamic GlobalPooling, there's no padding
             kernel_dims.insert(kernel_dims.end(), input_lens.begin() + 2, input_lens.end());
+            output_shape = dyn_out.computed_shape;
+            result       = dyn_out.computed_shape;
         }
-        else
+        else if((padding_mode != op::padding_mode_t::default_))
         {
+            // if padding_mode is set, input was a dynamic size.  Calculate padded size now.
+
+            // kernel_lens is the same as kernel_dims, but prepended with the 2 non-
+            // spatial dimensions.  For size computations, it's used like the weights
+            // tensor for convolutions.
+            std::vector<std::size_t> kernel_lens;
+            kernel_lens.insert(kernel_lens.end(), input_lens.begin(), input_lens.begin() + 2);
+            kernel_lens.insert(kernel_lens.end(), lengths.begin(), lengths.end());
             kernel_dims = this->lengths;
+
+            auto type = args[0].get_shape().type();
+            // dilation not currently supported for pooling, so default to all 1's
+            temp_padding = calc_dyn_auto_pad(
+                input_lens, kernel_lens, stride, {1, 1}, bool(padding_mode == op::same_upper));
+
+            output_shape = compute_padded_pool_shape(
+                args[0].get_shape(), shape(type, kernel_dims), temp_padding, stride, {1, 1});
+
+            result = argument(output_shape);
+        }
+        else // fixed/static input
+        {
+            kernel_dims  = this->lengths;
+            output_shape = dyn_out.computed_shape;
+            result       = dyn_out.computed_shape;
         }
+
+        // Perform the computation and populate result
         visit_all(result, args[0])([&](auto output, auto input) {
             using type = typename decltype(output)::value_type;
             switch(mode)
             {
             case migraphx::op::pooling_mode::average:
-                calc_pooling<type>(dyn_out.computed_shape, output, input, kernel_dims, avg_pool{});
+                calc_pooling<type>(
+                    output_shape, output, input, kernel_dims, temp_padding, avg_pool{});
                 break;
             case migraphx::op::pooling_mode::max:
-                calc_pooling<type>(dyn_out.computed_shape, output, input, kernel_dims, max_pool{});
+                calc_pooling<type>(
+                    output_shape, output, input, kernel_dims, temp_padding, max_pool{});
                 break;
             case migraphx::op::pooling_mode::lpnorm:
                 calc_pooling<type>(
-                    dyn_out.computed_shape, output, input, kernel_dims, lpnorm_pool{lp_order});
+                    output_shape, output, input, kernel_dims, temp_padding, lpnorm_pool{lp_order});
                 break;
             }
         });

src/include/migraphx/op/prefix_scan_op.hpp

 /*
  * The MIT License (MIT)
  *
- * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a copy
  * of this software and associated documentation files (the "Software"), to deal
@@ -21,6 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
+
 #ifndef MIGRAPHX_GUARD_OPERATORS_SCAN_OP_HPP
 #define MIGRAPHX_GUARD_OPERATORS_SCAN_OP_HPP
 
@@ -37,6 +38,12 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+/**
+ * Parent struct for prefix scan operations.  A prefix scan is equivalent to the C++
+ * std::exclusive_scan or std::inclusive_scan.  Given a list of numbers, a prefix scan
+ * sum op returns an equal size list of running totals of the values.  Other operations
+ * besides addition can be supported by their own child ops.
+ */
 template <class Derived>
 struct prefix_scan_op : op_name<Derived>
 {
@@ -60,9 +67,13 @@ struct prefix_scan_op : op_name<Derived>
 
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this, true}.has(1);
         auto s = inputs.front();
-        if(s.broadcasted())
+        if(s.dynamic())
+        {
+            return s;
+        }
+        else if(s.broadcasted())
         {
             return {s.type(), s.lens()};
         }
@@ -72,8 +83,9 @@ struct prefix_scan_op : op_name<Derived>
         }
     }
 
-    argument compute(const shape& output_shape, std::vector<argument> args) const
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
+        shape output_shape(dyn_out.computed_shape);
         argument result{output_shape};
         auto s = args[0].get_shape();
         if(s == output_shape)

src/include/migraphx/op/random_seed.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_OPERATORS_RANDOM_SEED_HPP
+#define MIGRAPHX_GUARD_OPERATORS_RANDOM_SEED_HPP
+
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/argument.hpp>
+#include <random>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+/**
+ *    Generates a random seed for the use of random number generators.  Generating the seed
+ * at runtime guarantees there will be a different random sequence on every execution.
+ * This operation has no inputs or attributes, and outputs an unsigned integer tensor with
+ * a single value.
+ */
+struct random_seed
+{
+    shape::type_t dtype = shape::type_t::uint64_type;
+
+    template <class Self, class F>
+    static auto reflect(Self& self, F f)
+    {
+        return pack(f(self.dtype, "dtype"));
+    }
+
+    std::string name() const { return "random_seed"; }
+    shape compute_shape(const std::vector<shape>& inputs) const
+    {
+        check_shapes{inputs, *this}.has(0);
+        return shape{dtype};
+    }
+
+    argument compute(const shape& output_shape, const std::vector<argument>&) const
+    {
+        argument result(output_shape);
+
+        result.visit([&](auto output) { output.front() = std::random_device{}(); });
+        return result;
+    }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/op/random_uniform.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.
+ */
+
+/**
+ * Random Uniform distribution operator.  Given a shape, populate it with random
+ * values.  Calls to random_uniform using the same randomization seed as a
+ * literal input will
+ * always generate the same pseudo-random sequence.
+ *
+ *      Inputs:   (1) randomization seed (any type is allowed)
+ *                (2) output buffer argument to be populated.
+ *
+ *      Attributes:  none
+ *
+ *      Output:   Returns the buffer from input #2.
+ *
+ */
+#ifndef MIGRAPHX_GUARD_OPERATORS_RANDOM_UNIFORM_HPP
+#define MIGRAPHX_GUARD_OPERATORS_RANDOM_UNIFORM_HPP
+
+#include <migraphx/check_shapes.hpp>
+#include <migraphx/argument.hpp>
+#include <random>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace op {
+
+/**
+ * random_uniform populates the passed shape with random numbers, in a uniform
+ * distribution.  Range for floating-point data types is (0, 1);
+ * for integer types it is [0, <max value for the type>]
+ */
+struct random_uniform
+{
+    // The random_uniform operation needs the random number generator seed
+    // to be passed as a runtime input.
+
+    std::string name() const { return "random_uniform"; }
+    shape compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs, *this, true}.has(2);
+
+        return inputs.at(1);
+    }
+
+    argument compute(const shape&, std::vector<argument> args) const
+    {
+        // Output goes into the passed buffer, not the shape output.
+        auto result = args[1];
+
+        uint64_t local_seed = args[0].at<uint64_t>(0);
+        std::mt19937 gen(local_seed);
+
+        result.visit([&](auto output) {
+            using type = typename decltype(output)::value_type;
+            if constexpr(std::is_integral<type>{})
+            {
+                // default range for all integer types is
+                // (0, std::uniform_int_distribution<type>::max()).
+                // Todo:  enable different ranges
+                std::uniform_int_distribution<type> dis;
+                std::generate(output.begin(), output.end(), [&] { return dis(gen); });
+            }
+            else
+            {
+                // default real distribution type is double with range (0, 1);
+                std::uniform_real_distribution<> dis;
+                std::generate(output.begin(), output.end(), [&] { return dis(gen); });
+            }
+        });
+        return result;
+    }
+
+    std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 1; }
+};
+
+} // namespace op
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif