Merge branch 'develop' into ck-gemm-fused

examples/nlp/python_bert_squad/requirements_bertsquad.txt

@@ -21,6 +21,6 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 # THE SOFTWARE.
 #####################################################################################
-tensorflow==2.7.2
+tensorflow==2.9.3
 onnxruntime
 tokenizers
\ No newline at end of file

src/include/migraphx/int_divide.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_RTGLIB_INT_DIVIDE_HPP
-#define MIGRAPHX_GUARD_RTGLIB_INT_DIVIDE_HPP
-
-#include <migraphx/config.hpp>
-#include <cmath>
-
-namespace migraphx {
-inline namespace MIGRAPHX_INLINE_NS {
-
-template <class R, class T, class U>
-R floor_divide(T x, U y)
-{
-    return R(std::floor(double(x) / double(y)));
-}
-
-template <class R, class T, class U>
-R ceil_divide(T x, U y)
-{
-    return R(std::ceil(double(x) / double(y)));
-}
-
-} // namespace MIGRAPHX_INLINE_NS
-} // namespace migraphx
-
-#endif

src/include/migraphx/literal.hpp

@@ -80,6 +80,7 @@ struct literal : raw_data<literal>
         fill(start, end);
     }
 
+    // Directly copies buffer of x
     template <class T, MIGRAPHX_REQUIRES(sizeof(T) == 1)>
     literal(const shape& s, T* x) : buffer(make_shared_array<char>(s.bytes())), m_shape(s)
     {
@@ -107,25 +108,15 @@ struct literal : raw_data<literal>
     std::shared_ptr<char> buffer;
     shape m_shape;
 
+    // Keeps the same data ordering as the given container
     template <class Iterator>
     void fill(Iterator start, Iterator end)
     {
         assert(std::distance(start, end) == m_shape.elements());
-        if(m_shape.standard())
-        {
-            m_shape.visit_type([&](auto as) { std::copy(start, end, as.from(buffer.get())); });
-        }
-        else
-        {
-            auto it = start;
-            m_shape.visit_type([&](auto as) {
-                auto output = make_view(m_shape, as.from(buffer.get()));
-                shape_for_each(output.get_shape(), [&](const auto& idx) {
-                    output(idx.begin(), idx.end()) = *it; // NOLINT(bugprone-signed-char-misuse)
-                    it++;
-                });
-            });
-        }
+        m_shape.visit_type([&](auto as) {
+            auto output = make_view(m_shape, as.from(buffer.get()));
+            std::copy(start, end, output.begin());
+        });
     }
 };
 

src/include/migraphx/op/pooling.hpp

@@ -31,7 +31,7 @@
 #include <migraphx/argument.hpp>
 #include <migraphx/par_for.hpp>
 #include <migraphx/shape_for_each.hpp>
-#include <migraphx/int_divide.hpp>
+#include <migraphx/dyn_output.hpp>
 #include <cmath>
 #include <utility>
 
@@ -49,6 +49,9 @@ struct pooling
     bool ceil_mode                   = false;
     int lp_order                     = 2;
 
+    // Global pooling with dynamic shape input
+    bool dyn_global = false;
+
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
@@ -57,7 +60,8 @@ struct pooling
                     f(self.stride, "stride"),
                     f(self.lengths, "lengths"),
                     f(self.ceil_mode, "ceil_mode"),
-                    f(self.lp_order, "lp_order"));
+                    f(self.lp_order, "lp_order"),
+                    f(self.dyn_global, "dyn_global"));
     }
 
     std::string name() const { return "pooling"; }
@@ -65,51 +69,111 @@ struct pooling
     void check_attribute_size() const
     {
         if((padding.size() != stride.size() and (padding.size() / 2) != stride.size()) or
-           stride.size() != lengths.size())
+           (not dyn_global and stride.size() != lengths.size()))
         {
             MIGRAPHX_THROW("POOLING: inconsistent attribute sizes");
         }
     }
 
+    size_t kdims() const
+    {
+        check_attribute_size();
+        return stride.size();
+    }
+
     value attributes() const { return {{"normalize_padding", "padding"}}; }
 
+    std::vector<std::size_t> calc_spatial_dim_out(const std::vector<std::size_t>& input_lens,
+                                                  std::size_t kdims) const
+    {
+        std::vector<std::size_t> output_lens{};
+        for(size_t i = 0; i < kdims; ++i)
+        {
+            if(input_lens[i + 2] == 0)
+            {
+                // handle opt = 0
+                output_lens.push_back(0);
+            }
+            else
+            {
+                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 len =
+                    (ceil_mode)
+                        ? dim_size / stride[i] + static_cast<std::size_t>((dim_size % stride[i] !=
+                                                                           0)) // ceil uint divide
+                        : dim_size / stride[i];                                // floor divide
+                output_lens.push_back(len + 1);
+            }
+        }
+        return output_lens;
+    }
+
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this, true}.has(1);
+        check_attribute_size();
 
         const shape& input = inputs.at(0);
-
-        auto input_lens   = input.lens();
-        size_t kdims      = input_lens.size() - 2;
-        auto input_size   = inputs[0].lens().size();
-        auto padding_size = padding.size();
-        if(input_size != padding_size / 2 + 2 and input_size != padding_size + 2)
+        auto padding_size  = padding.size();
+        size_t kdims       = input.ndim() - 2;
+        if(input.ndim() != padding_size / 2 + 2 and input.ndim() != padding_size + 2)
         {
             MIGRAPHX_THROW("POOLING: input and attribute size mismatch!");
         }
 
-        std::vector<std::size_t> output_lens(input_lens.begin(), input_lens.begin() + 2);
-
-        for(size_t i = 0; i < kdims; i++)
+        if(input.dynamic())
         {
-            std::ptrdiff_t dim_size;
-            auto padding_factor = 2 * padding[i];
-            if(padding_size == 2 * kdims)
-                padding_factor = padding[i] + padding[i + kdims];
-            dim_size = input_lens[i + 2] + padding_factor - lengths[i];
-            assert(dim_size >= 0);
-            std::size_t len = (ceil_mode) ? ceil_divide<std::ptrdiff_t>(dim_size, stride[i])
-                                          : floor_divide<std::ptrdiff_t>(dim_size, stride[i]);
-
-            output_lens.push_back(std::size_t(std::max<std::ptrdiff_t>(1, len + 1)));
+            auto input_dyn_dims = input.dyn_dims();
+            std::vector<shape::dynamic_dimension> output_dyn_dims(input_dyn_dims.begin(),
+                                                                  input_dyn_dims.begin() + 2);
+            if(dyn_global)
+            {
+                for(size_t i = 0; i < kdims; ++i)
+                {
+                    output_dyn_dims.push_back(shape::dynamic_dimension{1, 1, 1});
+                }
+                return {input.type(), output_dyn_dims};
+            }
+            else
+            {
+                auto min_spatial_dims = calc_spatial_dim_out(input.min_lens(), kdims);
+                auto max_spatial_dims = calc_spatial_dim_out(input.max_lens(), kdims);
+                auto opt_spatial_dims = calc_spatial_dim_out(input.opt_lens(), kdims);
+                for(size_t i = 0; i < kdims; ++i)
+                {
+                    output_dyn_dims.push_back(shape::dynamic_dimension{
+                        min_spatial_dims[i], max_spatial_dims[i], opt_spatial_dims[i]});
+                }
+                return {input.type(), output_dyn_dims};
+            }
         }
-        return inputs[0].with_lens(output_lens);
-    }
+        else
+        {
+            auto input_lens = input.lens();
 
-    size_t kdims() const
-    {
-        check_attribute_size();
-        return stride.size();
+            std::vector<std::size_t> output_lens(input_lens.begin(), input_lens.begin() + 2);
+            // Used for when normalize_compute_shape() is called again at model eval time
+            // for an originally dynamic shape. Since kernel shape is not used with dyn_global.
+            if(dyn_global)
+            {
+                for(size_t i = 0; i < kdims; ++i)
+                {
+                    output_lens.push_back(1);
+                }
+                return {input.type(), output_lens};
+            }
+            else
+            {
+                auto output_spatial_lens = calc_spatial_dim_out(input_lens, kdims);
+                output_lens.insert(
+                    output_lens.end(), output_spatial_lens.begin(), output_spatial_lens.end());
+                return inputs[0].with_lens(output_lens);
+            }
+        }
     }
 
     struct lpnorm_pool
@@ -158,7 +222,11 @@ struct pooling
     };
 
     template <class Type, class Out, class In, class Op>
-    void calc_pooling(const shape& output_shape, Out& output, const In& input, Op op) const
+    void calc_pooling(const shape& output_shape,
+                      Out& output,
+                      const In& input,
+                      const std::vector<std::size_t>& kernel_dims,
+                      Op op) const
     {
         auto in_s    = input.get_shape();
         auto in_lens = in_s.lens();
@@ -172,7 +240,7 @@ struct pooling
                 auto d_2 = dim - 2;
                 int start =
                     static_cast<int>(idx_o[dim] * stride[d_2]) - static_cast<int>(padding[d_2]);
-                int end = std::min(start + lengths[d_2], in_lens[dim]);
+                int end = std::min(start + kernel_dims[d_2], in_lens[dim]);
                 start   = std::max(start, 0);
                 win_start.push_back(start);
                 win_size.push_back(end - start);
@@ -198,21 +266,32 @@ struct pooling
         });
     }
 
-    argument compute(const shape& output_shape, std::vector<argument> args) const
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
-        argument result{output_shape};
+        argument result{dyn_out.computed_shape};
+        auto input_lens = args[0].get_shape().lens();
+        std::vector<std::size_t> kernel_dims;
+        if(dyn_global)
+        {
+            kernel_dims.insert(kernel_dims.end(), input_lens.begin() + 2, input_lens.end());
+        }
+        else
+        {
+            kernel_dims = this->lengths;
+        }
         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>(output_shape, output, input, avg_pool{});
+                calc_pooling<type>(dyn_out.computed_shape, output, input, kernel_dims, avg_pool{});
                 break;
             case migraphx::op::pooling_mode::max:
-                calc_pooling<type>(output_shape, output, input, max_pool{});
+                calc_pooling<type>(dyn_out.computed_shape, output, input, kernel_dims, max_pool{});
                 break;
             case migraphx::op::pooling_mode::lpnorm:
-                calc_pooling<type>(output_shape, output, input, lpnorm_pool{lp_order});
+                calc_pooling<type>(
+                    dyn_out.computed_shape, output, input, kernel_dims, lpnorm_pool{lp_order});
                 break;
             }
         });

src/include/migraphx/op/transpose.hpp

@@ -29,6 +29,7 @@
 #include <migraphx/config.hpp>
 #include <migraphx/value.hpp>
 #include <migraphx/op/normalize_attribute.hpp>
+#include <migraphx/dyn_output.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -45,17 +46,15 @@ struct transpose
     }
 
     std::string name() const { return "transpose"; }
+
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
-        auto input         = inputs.at(0);
-        auto input_lens    = input.lens();
-        auto input_strides = input.strides();
-        auto t             = input.type();
+        check_shapes{inputs, *this, true}.has(1);
+        auto input = inputs.at(0);
 
-        if(dims.size() != input_lens.size())
+        if(dims.size() != input.ndim())
         {
-            MIGRAPHX_THROW("Permutation has wrong number of axes");
+            MIGRAPHX_THROW("TRANSPOSE: Permutation has wrong number of axes");
         }
         std::vector<int64_t> axes(dims.size());
         std::iota(axes.begin(), axes.end(), 0);
@@ -63,19 +62,36 @@ struct transpose
         {
             MIGRAPHX_THROW("TRANSPOSE: Invalid permutation");
         }
-        std::vector<size_t> output_lens(input_lens.size());
-        std::vector<size_t> output_strides(input_lens.size());
-        for(std::size_t i = 0; i < output_lens.size(); i++)
+
+        if(input.dynamic())
         {
-            output_lens[i]    = input_lens[dims[i]];
-            output_strides[i] = input_strides[dims[i]];
+            std::vector<shape::dynamic_dimension> output_dyn_dims(input.ndim());
+            std::transform(dims.cbegin(), dims.cend(), output_dyn_dims.begin(), [&](auto dim) {
+                return input.dyn_dims()[dim];
+            });
+            return {input.type(), output_dyn_dims};
+        }
+        else
+        {
+            auto input_lens    = input.lens();
+            auto input_strides = input.strides();
+
+            std::vector<size_t> output_lens(input.ndim());
+            std::vector<size_t> output_strides(input.ndim());
+            for(std::size_t i = 0; i < input.ndim(); i++)
+            {
+                output_lens[i]    = input_lens[dims[i]];
+                output_strides[i] = input_strides[dims[i]];
+            }
+            return {input.type(), output_lens, output_strides};
         }
-        return {t, output_lens, output_strides};
     }
-    argument compute(shape output_shape, std::vector<argument> args) const
+
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
-        return args[0].reshape(output_shape);
+        return args[0].reshape(dyn_out.computed_shape);
     }
+
     std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 0; }
 };
 

src/include/migraphx/shape_for_each.hpp

@@ -31,6 +31,9 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
+/**
+ * Iterates the given function over the indices from the shape in order.
+ */
 template <class F>
 void shape_for_each(const migraphx::shape& s, F f)
 {
@@ -51,7 +54,6 @@ void shape_for_each(const migraphx::shape& s, F f)
         call(indices);
     }
 }
-
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx
 

src/insert_pad.cpp

@@ -77,14 +77,14 @@ static void update_pooling(const instruction_ref& input, const instruction_ref&
     {
         return;
     }
-    auto kdims = input->get_shape().lens().size() - 2;
+    auto kdims = input->get_shape().ndim() - 2;
     if(std::equal(op.padding.begin(),
                   op.padding.begin() + kdims,
                   op.padding.begin() + kdims,
                   op.padding.end()))
         return;
 
-    std::vector<int64_t> padding(input->get_shape().lens().size() * 2, 0);
+    std::vector<int64_t> padding(input->get_shape().ndim() * 2, 0);
     std::vector<size_t> pads_l(op.padding.begin(), op.padding.begin() + kdims);
     std::vector<size_t> pads_r(op.padding.begin() + kdims, op.padding.end());
     op.padding = std::vector<size_t>(kdims * 2, 0);

src/onnx/parse_pooling.cpp

@@ -47,52 +47,42 @@ struct parse_pooling : op_parser<parse_pooling>
                 {"GlobalLpPool", "lpnorm"}};
     }
 
-    instruction_ref parse(const op_desc& opd,
-                          const onnx_parser& /*parser*/,
-                          onnx_parser::node_info info,
-                          std::vector<instruction_ref> args) const
+    value handle_values(const op_desc& opd,
+                        onnx_parser::node_info info,
+                        const shape& in_shape,
+                        value values) const
     {
-        const std::unordered_map<std::string, op::pooling_mode> mode_map = {
-            {"max", op::pooling_mode::max},
-            {"average", op::pooling_mode::average},
-            {"lpnorm", op::pooling_mode::lpnorm}};
-        std::string mode = opd.op_name;
-        if(not contains(mode_map, mode))
-        {
-            MIGRAPHX_THROW("onnx pooling mode must be [\"max\", \"average\", \"lpnorm\"]");
-        }
-        operation op = make_op("pooling", {{"mode", mode_map.at(mode)}});
-        value values = op.to_value();
-        auto l0      = args[0];
-        auto in_lens = l0->get_shape().lens();
-        assert(in_lens.size() > 2);
-        auto kdims = in_lens.size() - 2;
-
+        auto kdims = in_shape.ndim() - 2;
         if(starts_with(opd.onnx_name, "Global"))
         {
-            values["lengths"] = std::vector<size_t>(in_lens.begin() + 2, in_lens.end());
+            // if spatial dimensions are dynamic use dyn_global flag
+            if(in_shape.dynamic() and std::any_of(in_shape.dyn_dims().cbegin() + 2,
+                                                  in_shape.dyn_dims().cend(),
+                                                  [](auto dd) { return not dd.is_fixed(); }))
+            {
+                values["dyn_global"] = true;
+                values["lengths"]    = std::vector<size_t>();
+            }
+            else
+            {
+                // works with static and fixed dynamic shape
+                auto m_lens       = in_shape.max_lens();
+                values["lengths"] = std::vector<size_t>(m_lens.begin() + 2, m_lens.end());
+            }
         }
 
-        // does not support ceil_mode
         if(contains(info.attributes, "ceil_mode"))
         {
             values["ceil_mode"] = static_cast<bool>(info.attributes.at("ceil_mode").i());
         }
 
-        // count include padding, if count include pad is 1, we always use
-        // explicit pad
-        int count_include_pad = 0;
-        if(contains(info.attributes, "count_include_pad"))
-        {
-            count_include_pad = info.attributes.at("count_include_pad").i();
-        }
-
         if(contains(info.attributes, "strides"))
         {
             values["stride"].clear();
             copy(info.attributes["strides"].ints(), std::back_inserter(values["stride"]));
             check_attr_sizes(kdims, values["stride"].size(), "PARSE_POOLING: inconsistent strides");
         }
+
         if(contains(info.attributes, "kernel_shape"))
         {
             values["lengths"].clear();
@@ -110,6 +100,46 @@ struct parse_pooling : op_parser<parse_pooling>
         // ensure pads availabe only when auto_pad is "NOT_SET"
         check_padding_mode(info, "POOLING");
 
+        return values;
+    }
+
+    instruction_ref parse(const op_desc& opd,
+                          const onnx_parser& /*parser*/,
+                          onnx_parser::node_info info,
+                          std::vector<instruction_ref> args) const
+    {
+        std::string mode                                                 = opd.op_name;
+        const std::unordered_map<std::string, op::pooling_mode> mode_map = {
+            {"max", op::pooling_mode::max},
+            {"average", op::pooling_mode::average},
+            {"lpnorm", op::pooling_mode::lpnorm}};
+        if(not contains(mode_map, mode))
+        {
+            MIGRAPHX_THROW(
+                "PARSE_POOLING: onnx pooling mode must be [\"max\", \"average\", \"lpnorm\"]");
+        }
+        operation op  = make_op("pooling", {{"mode", mode_map.at(mode)}});
+        value values  = op.to_value();
+        auto l0       = args[0];
+        auto in_shape = l0->get_shape();
+        assert(in_shape.ndim() > 2);
+        auto kdims = in_shape.ndim() - 2;
+
+        values = handle_values(opd, info, in_shape, values);
+
+        // count include padding, if count include pad is 1, we always use
+        // explicit pad
+        int count_include_pad = 0;
+        if(contains(info.attributes, "count_include_pad"))
+        {
+            if(in_shape.dynamic())
+            {
+                MIGRAPHX_THROW("PARSE_POOLING: count_include_pad attribute is not supported for "
+                               "dynamic input shape");
+            }
+            count_include_pad = info.attributes.at("count_include_pad").i();
+        }
+
         std::vector<int64_t> paddings;
         float pad_val = ((mode == "max") ? std::numeric_limits<float>::lowest() : 0.0f);
 
@@ -123,14 +153,22 @@ struct parse_pooling : op_parser<parse_pooling>
 
         if(contains(info.attributes, "auto_pad"))
         {
-            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_lens,
-                                  paddings);
+            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)
@@ -150,6 +188,7 @@ struct parse_pooling : op_parser<parse_pooling>
             values["stride"].resize(kdims);
             std::fill_n(values["stride"].begin(), kdims, 1);
         }
+
         // used to calculate the supposed output shape
         std::vector<int64_t> orig_padding = paddings;
 
@@ -159,6 +198,11 @@ struct parse_pooling : op_parser<parse_pooling>
 
         if(not slice_start.empty())
         {
+            if(in_shape.dynamic())
+            {
+                MIGRAPHX_THROW(
+                    "PARSE_POOLING: asymmetric padding not supported for dynamic input shape");
+            }
             // calculate expected output shape
             orig_padding.insert(orig_padding.begin() + kdims, 2, 0);
             orig_padding.insert(orig_padding.begin(), 2, 0);

src/onnx/parse_transpose.cpp

@@ -47,7 +47,7 @@ struct parse_transpose : op_parser<parse_transpose>
         }
 
         // if perm is empty, use the default value
-        auto n_dim = args.front()->get_shape().lens().size();
+        auto n_dim = args.front()->get_shape().ndim();
         if(perm.empty())
         {
             perm.resize(n_dim);

src/targets/gpu/target.cpp

@@ -149,8 +149,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         pack_int8_args{},
         dead_code_elimination{},
-        adjust_allocation{gpu_allocation_model{}},
-        dead_code_elimination{},
         fuse_ops{&ctx, options.fast_math},
         dead_code_elimination{},
         replace_allocate{gpu_allocation_model{}, options.offload_copy},

test/literal_test.cpp

@@ -49,6 +49,25 @@ TEST_CASE(literal_test)
     EXPECT(l4.empty());
 }
 
+TEST_CASE(literal_nstd_shape_vector)
+{
+    migraphx::shape nstd_shape{migraphx::shape::float_type, {1, 3, 2, 2}, {12, 1, 6, 3}};
+    std::vector<float> data(12);
+    std::iota(data.begin(), data.end(), 0);
+    auto l0 = migraphx::literal{nstd_shape, data};
+
+    // check data buffer is read in correctly
+    std::vector<float> expected_buffer = {0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11};
+    const auto* start                  = reinterpret_cast<const float*>(l0.data());
+    std::vector<float> l0_data{start, start + 12};
+    EXPECT(l0_data == expected_buffer);
+
+    // check that using visit() (that uses a tensor view) gives data in correct order
+    std::vector<float> results_vector(12);
+    l0.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
+    EXPECT(results_vector == data);
+}
+
 TEST_CASE(literal_os1)
 {
     migraphx::literal l{1};

test/onnx/gen_onnx.py

@@ -237,6 +237,64 @@ def averagepool_3d_test():
     return ([node], [x], [out])
 
 
+@onnx_test
+def averagepool_dyn_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT,
+                                      [None, 3, 5, 5, 5])
+    out = helper.make_tensor_value_info('1', TensorProto.FLOAT,
+                                        [None, 3, 3, 3, 3])
+
+    node = onnx.helper.make_node('AveragePool',
+                                 inputs=['0'],
+                                 outputs=['1'],
+                                 kernel_shape=[3, 3, 3])
+
+    return ([node], [x], [out])
+
+
+@onnx_test
+def averagepool_dyn_autopad_error_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [None, 1, 5, 5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [None, 1, 5, 5])
+
+    node = onnx.helper.make_node('AveragePool',
+                                 inputs=['x'],
+                                 outputs=['y'],
+                                 kernel_shape=[2, 2],
+                                 auto_pad='SAME_LOWER')
+
+    return ([node], [x], [y])
+
+
+@onnx_test
+def averagepool_dyn_asym_padding_error_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [None, 1, 5, 5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [None, 1, 3, 3])
+
+    node = onnx.helper.make_node('AveragePool',
+                                 inputs=['x'],
+                                 outputs=['y'],
+                                 kernel_shape=[2, 2],
+                                 strides=[2, 2],
+                                 pads=[0, 0, 1, 1])
+
+    return ([node], [x], [y])
+
+
+@onnx_test
+def averagepool_dyn_cip_error_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [None, 1, 5, 5])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [None, 1, 1, 1])
+
+    node = onnx.helper.make_node('AveragePool',
+                                 inputs=['x'],
+                                 outputs=['y'],
+                                 kernel_shape=[2, 2],
+                                 count_include_pad=1)
+
+    return ([node], [x], [y])
+
+
 @onnx_test
 def averagepool_notset_test():
     x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [1, 1, 5, 5])
@@ -2069,6 +2127,21 @@ def globalavgpool_test():
     return ([node], [x], [y])
 
 
+@onnx_test
+def globalavgpool_dyn_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT,
+                                      [None, 3, 16, 16])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [None, 3, 1, 1])
+
+    node = onnx.helper.make_node(
+        'GlobalAveragePool',
+        inputs=['0'],
+        outputs=['1'],
+    )
+
+    return ([node], [x], [y])
+
+
 @onnx_test
 def globallppool_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 16, 16])
@@ -2083,6 +2156,21 @@ def globallppool_test():
     return ([node], [x], [y])
 
 
+@onnx_test
+def globallppool_dyn_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT,
+                                      [1, 3, None, None])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [1, 3, 1, 1])
+
+    node = onnx.helper.make_node(
+        'GlobalLpPool',
+        inputs=['0'],
+        outputs=['1'],
+    )
+
+    return ([node], [x], [y])
+
+
 @onnx_test
 def globalmaxpool_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [1, 3, 16, 16])
@@ -2097,6 +2185,21 @@ def globalmaxpool_test():
     return ([node], [x], [y])
 
 
+@onnx_test
+def globalmaxpool_dyn_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT,
+                                      [None, 3, 32, 32])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [None, 3, 1, 1])
+
+    node = onnx.helper.make_node(
+        'GlobalMaxPool',
+        inputs=['0'],
+        outputs=['1'],
+    )
+
+    return ([node], [x], [y])
+
+
 @onnx_test
 def greater_test():
     ax1 = np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0])
@@ -6277,6 +6380,21 @@ def transpose_test():
     return ([node], [x], [y])
 
 
+@onnx_test
+def transpose_dyn_test():
+    x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [None, 2, 2, 3])
+    y = helper.make_tensor_value_info('1', TensorProto.FLOAT, [None, 3, 2, 2])
+
+    node = onnx.helper.make_node(
+        'Transpose',
+        perm=[0, 3, 1, 2],
+        inputs=['0'],
+        outputs=['1'],
+    )
+
+    return ([node], [x], [y])
+
+
 @onnx_test
 def transpose_gather_test():
     x = helper.make_tensor_value_info('data', TensorProto.FLOAT, [3, 5, 4, 6])

test/onnx/onnx_test.cpp

@@ -273,6 +273,51 @@ TEST_CASE(averagepool_3d_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(averagepool_dyn_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto l0  = mm->add_parameter(
+        "0",
+        {migraphx::shape::float_type, {{1, 4, 0}, {3, 3, 0}, {5, 5, 0}, {5, 5, 0}, {5, 5, 0}}});
+    auto ret = mm->add_instruction(migraphx::make_op("pooling",
+                                                     {{"mode", migraphx::op::pooling_mode::average},
+                                                      {"padding", {0, 0, 0, 0, 0, 0}},
+                                                      {"stride", {1, 1, 1}},
+                                                      {"lengths", {3, 3, 3}}}),
+                                   l0);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    auto prog                     = migraphx::parse_onnx("averagepool_dyn_test.onnx", options);
+    EXPECT(p == prog);
+}
+
+TEST_CASE(averagepool_dyn_autopad_error_test)
+{
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    EXPECT(test::throws(
+        [&] { migraphx::parse_onnx("averagepool_dyn_autopad_error_test.onnx", options); }));
+}
+
+TEST_CASE(averagepool_dyn_asym_padding_error_test)
+{
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    EXPECT(test::throws(
+        [&] { migraphx::parse_onnx("averagepool_dyn_asym_padding_error_test.onnx", options); }));
+}
+
+TEST_CASE(averagepool_dyn_cip_error_test)
+{
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    EXPECT(test::throws(
+        [&] { migraphx::parse_onnx("averagepool_dyn_cip_error_test.onnx", options); }));
+}
+
 TEST_CASE(averagepool_notset_test)
 {
     migraphx::program p;
@@ -2144,6 +2189,28 @@ TEST_CASE(globalavgpool_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(globalavgpool_dyn_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto input =
+        mm->add_parameter("0",
+                          migraphx::shape{migraphx::shape::float_type,
+                                          {{1, 4, 0}, {3, 3, 0}, {16, 16, 0}, {16, 16, 0}}});
+    auto ret = mm->add_instruction(migraphx::make_op("pooling",
+                                                     {{"mode", migraphx::op::pooling_mode::average},
+                                                      {"lengths", {16, 16}},
+                                                      {"padding", {0, 0, 0, 0}}}),
+                                   input);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    auto prog                     = parse_onnx("globalavgpool_dyn_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(globallppool_test)
 {
     migraphx::program p;
@@ -2161,6 +2228,29 @@ TEST_CASE(globallppool_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(globallppool_dyn_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto input =
+        mm->add_parameter("0",
+                          migraphx::shape{migraphx::shape::float_type,
+                                          {{1, 1, 0}, {3, 3, 0}, {16, 32, 0}, {16, 32, 0}}});
+    auto ret = mm->add_instruction(migraphx::make_op("pooling",
+                                                     {{"mode", migraphx::op::pooling_mode::lpnorm},
+                                                      {"dyn_global", true},
+                                                      {"padding", {0, 0, 0, 0}},
+                                                      {"lengths", {}}}),
+                                   input);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {16, 32, 0};
+    auto prog                     = migraphx::parse_onnx("globallppool_dyn_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(globalmaxpool_test)
 {
     migraphx::program p;
@@ -2178,6 +2268,28 @@ TEST_CASE(globalmaxpool_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(globalmaxpool_dyn_test)
+{
+    migraphx::program p;
+    auto* mm = p.get_main_module();
+    auto input =
+        mm->add_parameter("0",
+                          migraphx::shape{migraphx::shape::float_type,
+                                          {{1, 4, 0}, {3, 3, 0}, {32, 32, 0}, {32, 32, 0}}});
+    auto ret = mm->add_instruction(migraphx::make_op("pooling",
+                                                     {{"mode", migraphx::op::pooling_mode::max},
+                                                      {"lengths", {32, 32}},
+                                                      {"padding", {0, 0, 0, 0}}}),
+                                   input);
+    mm->add_return({ret});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    auto prog                     = parse_onnx("globalmaxpool_dyn_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(greater_test)
 {
     migraphx::program p;
@@ -5973,6 +6085,24 @@ TEST_CASE(transpose_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(transpose_dyn_test)
+{
+    migraphx::program p;
+    auto* mm   = p.get_main_module();
+    auto input = mm->add_parameter(
+        "0",
+        migraphx::shape{migraphx::shape::float_type, {{1, 4, 0}, {2, 2, 0}, {2, 2, 0}, {3, 3, 0}}});
+    std::vector<int64_t> perm{0, 3, 1, 2};
+    auto t0 = mm->add_instruction(migraphx::make_op("transpose", {{"permutation", perm}}), input);
+    mm->add_return({t0});
+
+    migraphx::onnx_options options;
+    options.default_dyn_dim_value = {1, 4, 0};
+    auto prog                     = migraphx::parse_onnx("transpose_dyn_test.onnx", options);
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(topk_attrk_test)
 {
     migraphx::program p;