Merge branch 'develop' into optimize

.github/workflows/performance.yaml

@@ -5,14 +5,14 @@ on:
     branches: [develop]
     types: [opened, synchronize, closed]
   schedule:
-    - cron: "0 5 * * 1-6"
+    - cron: "0 6 * * 1-6"
 
   workflow_dispatch:
     inputs:
       rocm_release:
         description: ROCm Version
         required: true
-        default: '5.2'
+        default: '5.3'
       performance_reports_repo:
         description: Result repository
         required: true
@@ -30,9 +30,9 @@ concurrency: "perftest-${{ github.head_ref ||  github.base_ref || 'schedule' }}"
 
 jobs:
   release:
-    uses: rocmsoftwareplatform/migraphx-benchmark/.github/workflows/perf-test.yml@main
+    uses: ROCmSoftwarePlatform/migraphx-benchmark/.github/workflows/perf-test.yml@main
     with:
-      rocm_release: ${{ github.event.inputs.rocm_release || '5.2' }}
+      rocm_release: ${{ github.event.inputs.rocm_release || '5.3' }}
       result_number: ${{ github.event.inputs.result_number || '10' }}
       flags: ${{ github.event.inputs.flags || '-s' }} 
       performance_reports_repo: ${{ github.event.inputs.performance_reports_repo || 'ROCmSoftwarePlatform/migraphx-reports' }} 

Dockerfile

@@ -87,7 +87,7 @@ RUN git clone --single-branch --branch ${ONNXRUNTIME_BRANCH} --recursive ${ONNXR
 
 ADD tools/build_and_test_onnxrt.sh /onnxruntime/build_and_test_onnxrt.sh
 
-RUN cget -p /usr/local install ROCmSoftwarePlatform/llvm-project-mlir@c0723a7e50043d973cb73ae51dc30d36679ee7e5 -DBUILD_MIXR_TARGET=On
+RUN cget -p /usr/local install ROCmSoftwarePlatform/rocMLIR@0f38fb33f518b53b94b541feb9b079668c5518e8 -DBUILD_MIXR_TARGET=On -DLLVM_ENABLE_ZSTD=Off -DLLVM_ENABLE_THREADS=Off
 
 ENV MIOPEN_FIND_DB_PATH=/tmp/miopen/find-db
 ENV MIOPEN_USER_DB_PATH=/tmp/miopen/user-db

examples/migraphx/migraphx_driver/README.md

@@ -29,6 +29,7 @@ See below for a comprehensive list of commands and option arguments, as well as
 | --tf | Load file as a tensorflow graph |
 | --migraphx | Load file as a migraphx graph |
 | --migraphx-json | Load file as a migraphx JSON graph |
+| --batch | Set batch size for the model | 
 | --nhwc | Treat tensorflow format as nhwc | 
 | --nchw | Treat tensorflow format as nchw |
 | --skip-unknown-operators | Skip unknown operators when parsing and continue to parse |

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/common.cpp

@@ -77,7 +77,6 @@ std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, sha
     }
     auto offset = s1.ndim() - s0.ndim();
     std::vector<shape::dynamic_dimension> out_dims(s1.dyn_dims());
-    shape::dynamic_dimension one_dyn_dim{1, 1, 0};
     std::transform(
         s0.dyn_dims().cbegin(),
         s0.dyn_dims().cend(),
@@ -88,7 +87,7 @@ std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, sha
             {
                 return a;
             }
-            else if(a == one_dyn_dim or b == one_dyn_dim)
+            else if(a == 1 or b == 1)
             {
                 // setting opt to 0, may need to be changed
                 return shape::dynamic_dimension{std::max(a.min, b.min), std::max(a.max, b.max), 0};

src/dead_code_elimination.cpp

@@ -51,8 +51,8 @@ void dead_code_elimination::apply(module& m) const
         // Skip instruction with empty shape as output unless its [dynamic, builtin, undefined,
         // identity, allocate]
         if((not i->get_shape().dynamic() and i->get_shape().elements() == 0) and
-           i->name().front() != '@' and
-           not contains({"undefined", "identity", "allocate"}, i->name()))
+           not(i->name().front() == '@') and not contains({"identity", "allocate"}, i->name()) and
+           not i->is_undefined())
             continue;
         assert(std::distance(m.begin(), i) <= std::distance(m.begin(), last));
         std::unordered_set<instruction_ref> visited;

src/file_buffer.cpp

@@ -30,23 +30,31 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
 template <class T>
-T generic_read_file(const std::string& filename)
+T generic_read_file(const std::string& filename, size_t offset = 0, size_t nbytes = 0)
 {
     std::ifstream is(filename, std::ios::binary | std::ios::ate);
-    std::streamsize size = is.tellg();
-    if(size < 1)
+    if(nbytes == 0)
+    {
+        // if there is a non-zero offset and nbytes is not set,
+        // calculate size of remaining bytes to read
+        nbytes = is.tellg();
+        if(offset > nbytes)
+            MIGRAPHX_THROW("offset is larger than file size");
+        nbytes -= offset;
+    }
+    if(nbytes < 1)
         MIGRAPHX_THROW("Invalid size for: " + filename);
-    is.seekg(0, std::ios::beg);
+    is.seekg(offset, std::ios::beg);
 
-    T buffer(size, 0);
-    if(not is.read(&buffer[0], size))
+    T buffer(nbytes, 0);
+    if(not is.read(&buffer[0], nbytes))
         MIGRAPHX_THROW("Error reading file: " + filename);
     return buffer;
 }
 
-std::vector<char> read_buffer(const std::string& filename)
+std::vector<char> read_buffer(const std::string& filename, size_t offset, size_t nbytes)
 {
-    return generic_read_file<std::vector<char>>(filename);
+    return generic_read_file<std::vector<char>>(filename, offset, nbytes);
 }
 
 std::string read_string(const std::string& filename)

src/include/migraphx/check_shapes.hpp

@@ -198,7 +198,7 @@ struct check_shapes
      */
     const check_shapes& same_ndims() const
     {
-        if(not this->same([](const shape& s) { return s.max_lens().size(); }))
+        if(not this->same([](const shape& s) { return s.ndim(); }))
             MIGRAPHX_THROW(prefix() + "Number of dimensions do not match");
         return *this;
     }

src/include/migraphx/file_buffer.hpp

@@ -31,7 +31,7 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-std::vector<char> read_buffer(const std::string& filename);
+std::vector<char> read_buffer(const std::string& filename, size_t offset = 0, size_t nbytes = 0);
 std::string read_string(const std::string& filename);
 
 void write_buffer(const std::string& filename, const char* buffer, std::size_t size);

src/include/migraphx/instruction.hpp

@@ -121,6 +121,8 @@ struct instruction
 
     bool can_eval() const;
 
+    bool is_undefined() const;
+
     argument eval(bool check_eval = true) const;
 
     void finalize(context& ctx);

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/argmax.hpp

@@ -30,6 +30,7 @@
 #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 {
@@ -56,12 +57,20 @@ struct argmax
 
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
-        auto lens = inputs[0].lens();
-
-        lens[axis] = 1;
-
-        return {shape::int64_type, lens};
+        check_shapes{inputs, *this, true}.has(1);
+        const auto& s0 = inputs[0];
+        if(s0.dynamic())
+        {
+            auto dyn_dims  = s0.dyn_dims();
+            dyn_dims[axis] = {1, 1, 0};
+            return {shape::int64_type, dyn_dims};
+        }
+        else
+        {
+            auto lens  = s0.lens();
+            lens[axis] = 1;
+            return {shape::int64_type, lens};
+        }
     }
 
     template <class T>
@@ -79,19 +88,18 @@ struct argmax
                 max_index = i;
             }
         }
-
         return max_index;
     }
 
-    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 batch_item_num = args.front().get_shape().lens()[axis];
 
         result.visit([&](auto output) {
             args[0].visit([&](auto input) {
-                par_for(output_shape.elements(), [&](auto i) {
-                    auto data_idx = output_shape.multi(i);
+                par_for(dyn_out.computed_shape.elements(), [&](auto i) {
+                    auto data_idx = dyn_out.computed_shape.multi(i);
                     output[i]     = this->calc_argmax(input, data_idx, batch_item_num);
                 });
             });

src/include/migraphx/op/dot.hpp

@@ -28,6 +28,7 @@
 #include <migraphx/argument.hpp>
 #include <migraphx/config.hpp>
 #include <migraphx/gemm.hpp>
+#include <migraphx/dyn_output.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -38,41 +39,69 @@ struct dot
     std::string name() const { return "dot"; }
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.same_type().has(2);
+        check_shapes{inputs, *this, true}.same_type().same_ndims().has(2);
         const shape& a = inputs.at(0);
         const shape& b = inputs.at(1);
         auto t         = a.type();
 
-        if(not std::all_of(
-               inputs.begin(), inputs.end(), [](auto s) { return s.lens().size() >= 2; }))
+        if(not std::all_of(inputs.begin(), inputs.end(), [](auto s) { return s.ndim() >= 2; }))
         {
-            MIGRAPHX_THROW("DOT: dot only accept 2 or more dims operands");
+            MIGRAPHX_THROW("DOT: dot only accepts operands with 2 or more dimensions ");
         }
-
-        // only handle the case that the batch size of a and b are the same
-        if(not std::equal(
-               a.lens().rbegin() + 2, a.lens().rend(), b.lens().rbegin() + 2, b.lens().rend()))
+        if(a.dynamic() or b.dynamic())
         {
-            MIGRAPHX_THROW("DOT: batch size of A and B mismatch: {" + to_string_range(a.lens()) +
-                           "} x {" + to_string_range(b.lens()) + "}");
+            auto s0 = a.to_dynamic();
+            auto s1 = b.to_dynamic();
+            if(not std::equal(s0.dyn_dims().rbegin() + 2,
+                              s0.dyn_dims().rend(),
+                              s1.dyn_dims().rbegin() + 2,
+                              s1.dyn_dims().rend()))
+            {
+                MIGRAPHX_THROW("DOT: dynamic outer dimensions of A and B mismatch: {" +
+                               to_string_range(s0.dyn_dims()) + "} x {" +
+                               to_string_range(s1.dyn_dims()) + "}");
+            }
+            std::size_t dim_0 = s0.ndim() - 2;
+            std::size_t dim_1 = s0.ndim() - 1;
+            if(s0.dyn_dims()[dim_1] != s1.dyn_dims()[dim_0])
+            {
+                MIGRAPHX_THROW("DOT: dynamic inner dimensions do not match: {" +
+                               to_string_range(s0.dyn_dims()) + "} x {" +
+                               to_string_range(s1.dyn_dims()) + "}");
+            }
+            auto out_dyn_dims   = s0.dyn_dims();
+            out_dyn_dims[dim_1] = s1.dyn_dims()[dim_1];
+            return {t, out_dyn_dims};
         }
-
-        std::size_t dim_0 = a.lens().size() - 2;
-        std::size_t dim_1 = a.lens().size() - 1;
-        if(a.lens()[dim_1] != b.lens()[dim_0])
+        else
         {
-            MIGRAPHX_THROW("DOT: inner dimensions do not match: {" + to_string_range(a.lens()) +
-                           "} x {" + to_string_range(b.lens()) + "}");
-        }
+            // only handle the case that all the dimensions except the last two are the same
+            if(not std::equal(
+                   a.lens().rbegin() + 2, a.lens().rend(), b.lens().rbegin() + 2, b.lens().rend()))
+            {
+                MIGRAPHX_THROW("DOT: static outer dimensions of A and B mismatch: {" +
+                               to_string_range(a.lens()) + "} x {" + to_string_range(b.lens()) +
+                               "}");
+            }
 
-        auto out_lens   = a.lens();
-        out_lens[dim_1] = b.lens()[dim_1];
-        return {t, out_lens};
+            std::size_t dim_0 = a.ndim() - 2;
+            std::size_t dim_1 = a.ndim() - 1;
+            if(a.lens()[dim_1] != b.lens()[dim_0])
+            {
+                MIGRAPHX_THROW("DOT: static inner dimensions do not match: {" +
+                               to_string_range(a.lens()) + "} x {" + to_string_range(b.lens()) +
+                               "}");
+            }
+
+            auto out_lens   = a.lens();
+            out_lens[dim_1] = b.lens()[dim_1];
+            return {t, out_lens};
+        }
     }
 
-    argument compute(shape output_shape, std::vector<argument> args) const
+    argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
-        argument result = argument{output_shape};
+        argument result = argument{dyn_out.computed_shape};
         visit_all(result, args[0], args[1])(
             [&](auto cmat, auto amat, auto bmat) { gemm(cmat, amat, bmat, 1.0f, 0.0f); });
         return result;

src/include/migraphx/op/flatten.hpp

@@ -55,17 +55,47 @@ struct flatten
     std::string name() const { return "flatten"; }
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1).standard();
-        auto&& lens = inputs.front().lens();
-        auto x =
-            std::accumulate(lens.begin(), lens.begin() + axis, std::size_t{1}, std::multiplies<>{});
-        auto y =
-            std::accumulate(lens.begin() + axis, lens.end(), std::size_t{1}, std::multiplies<>{});
-        return {inputs.at(0).type(), {x, y}};
+        check_shapes{inputs, *this, true}.has(1);
+        auto s = inputs[0];
+        if(s.dynamic())
+        {
+            auto min_lens = s.min_lens();
+            auto max_lens = s.max_lens();
+            auto opt_lens = s.opt_lens();
+            // If any of the opt values is 0, output opt will be 0
+            shape::dynamic_dimension x = {
+                std::accumulate(
+                    min_lens.begin(), min_lens.begin() + axis, std::size_t{1}, std::multiplies<>{}),
+                std::accumulate(
+                    max_lens.begin(), max_lens.begin() + axis, std::size_t{1}, std::multiplies<>{}),
+                std::accumulate(opt_lens.begin(),
+                                opt_lens.begin() + axis,
+                                std::size_t{1},
+                                std::multiplies<>{})};
+            shape::dynamic_dimension y = {
+                std::accumulate(
+                    min_lens.begin() + axis, min_lens.end(), std::size_t{1}, std::multiplies<>{}),
+                std::accumulate(
+                    max_lens.begin() + axis, max_lens.end(), std::size_t{1}, std::multiplies<>{}),
+                std::accumulate(
+                    opt_lens.begin() + axis, opt_lens.end(), std::size_t{1}, std::multiplies<>{}),
+            };
+            return {s.type(), {x, y}};
+        }
+        else
+        {
+            check_shapes{inputs, *this}.standard();
+            auto&& lens = s.lens();
+            auto x      = std::accumulate(
+                lens.begin(), lens.begin() + axis, std::size_t{1}, std::multiplies<>{});
+            auto y = std::accumulate(
+                lens.begin() + axis, lens.end(), std::size_t{1}, std::multiplies<>{});
+            return {s.type(), {x, y}};
+        }
     }
-    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/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/softmax.hpp

@@ -53,15 +53,15 @@ struct softmax
     std::string name() const { return "softmax"; }
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
-        if(inputs.at(0).packed())
+        check_shapes{inputs, *this, true}.has(1);
+        auto s0 = inputs[0];
+        if(s0.dynamic() or s0.packed())
         {
-            return inputs.at(0);
+            return s0;
         }
         else
         {
-            auto lens = inputs.at(0).lens();
-            return {inputs.at(0).type(), lens};
+            return {s0.type(), s0.lens()};
         }
     }
 

src/include/migraphx/op/squeeze.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 {
@@ -54,52 +55,85 @@ struct squeeze
     std::string name() const { return "squeeze"; }
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this, true}.has(1);
         auto input_shape = inputs[0];
-        auto type        = input_shape.type();
-        auto old_lens    = input_shape.lens();
-        auto old_strides = input_shape.strides();
-        if(std::any_of(axes.begin(), axes.end(), [&](auto axis) { return old_lens[axis] != 1; }))
+        if(input_shape.dynamic())
         {
-            MIGRAPHX_THROW("squeeze axis dimension should be equal to 1");
-        }
-        std::vector<std::size_t> new_lens;
-        std::vector<std::size_t> new_strides;
-        if(axes.empty())
-        {
-            for(auto i : range(old_lens.size()))
+            if(std::any_of(axes.begin(), axes.end(), [&](auto axis) {
+                   return input_shape.dyn_dims()[axis] != 1;
+               }))
+            {
+                MIGRAPHX_THROW(
+                    "SQUEEZE: dynamic axis dimension should be equal to {1, 1, 0} or {1, 1, 1}");
+            }
+            std::vector<shape::dynamic_dimension> dyn_dims = {};
+            if(axes.empty())
+            {
+                std::copy_if(input_shape.dyn_dims().cbegin(),
+                             input_shape.dyn_dims().cend(),
+                             std::back_inserter(dyn_dims),
+                             [&](auto dd) { return dd != 1; });
+            }
+            else
             {
-                if(old_lens[i] != 1)
+                for(auto i : range(input_shape.ndim()))
                 {
-                    new_lens.push_back(old_lens[i]);
-                    new_strides.push_back(old_strides[i]);
+                    if(std::find(axes.begin(), axes.end(), i) == axes.end())
+                    {
+                        dyn_dims.push_back(input_shape.dyn_dims()[i]);
+                    }
                 }
             }
+            return {input_shape.type(), dyn_dims};
         }
         else
         {
-            for(auto i : range(old_lens.size()))
+            auto type        = input_shape.type();
+            auto old_lens    = input_shape.lens();
+            auto old_strides = input_shape.strides();
+            if(std::any_of(
+                   axes.begin(), axes.end(), [&](auto axis) { return old_lens[axis] != 1; }))
             {
-                if(std::find(axes.begin(), axes.end(), i) == axes.end())
+                MIGRAPHX_THROW("SQUEEZE: static axis dimension should be equal to 1");
+            }
+            std::vector<std::size_t> new_lens;
+            std::vector<std::size_t> new_strides;
+            if(axes.empty())
+            {
+                for(auto i : range(old_lens.size()))
                 {
-                    new_lens.push_back(old_lens[i]);
-                    new_strides.push_back(old_strides[i]);
+                    if(old_lens[i] != 1)
+                    {
+                        new_lens.push_back(old_lens[i]);
+                        new_strides.push_back(old_strides[i]);
+                    }
                 }
             }
-        }
-        if(new_lens.empty())
-        {
-            return shape{type};
-        }
-        else
-        {
-            return shape{type, new_lens, new_strides};
+            else
+            {
+                for(auto i : range(old_lens.size()))
+                {
+                    if(std::find(axes.begin(), axes.end(), i) == axes.end())
+                    {
+                        new_lens.push_back(old_lens[i]);
+                        new_strides.push_back(old_strides[i]);
+                    }
+                }
+            }
+            if(new_lens.empty())
+            {
+                return shape{type};
+            }
+            else
+            {
+                return shape{type, new_lens, new_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/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/op/unsqueeze.hpp

@@ -29,11 +29,20 @@
 #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 {
 namespace op {
 
+/**
+ * Adds dimensions to a tensor based on the axes attribute.
+ * `axes` are based on the number of output shape dimensions and should not contain duplicates.
+ * `steps` are for modifying dimensions added to the middle of the original shape.
+ * Each step must be a factor of the original dimension.
+ * ex: unsqueeze(shape = [3, 4, 10], axes = [2, 4, 5], steps = [2]) -> shape = [3, 4, 2, 5, 1, 1]
+ * Dynamic shape version does not handle `steps`.
+ */
 struct unsqueeze
 {
     std::vector<int64_t> axes;
@@ -56,63 +65,89 @@ struct unsqueeze
     std::string name() const { return "unsqueeze"; }
     shape normalize_compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this, true}.has(1);
         auto input_shape = inputs[0];
-        auto type        = input_shape.type();
-        auto old_lens    = input_shape.lens();
-        auto old_strides = input_shape.strides();
-        if(input_shape.scalar())
+
+        if(input_shape.dynamic())
         {
-            if(old_lens.size() == 1 and old_lens.front() == 1)
-                return shape{type, old_lens};
-            else
-                MIGRAPHX_THROW("UNSQUEEZE: Input must be a scalar");
+            if(not steps.empty())
+            {
+                MIGRAPHX_THROW("UNSQUEEZE_dyn: nonempty steps attribute");
+            }
+            std::vector<shape::dynamic_dimension> dyn_dims = {};
+            auto new_ndim                                  = input_shape.ndim() + axes.size();
+            std::size_t k                                  = 0;
+            for(auto i : range(new_ndim))
+            {
+                if(std::find(axes.begin(), axes.end(), i) != axes.end())
+                {
+                    dyn_dims.push_back({1, 1, 0});
+                }
+                else
+                {
+                    dyn_dims.push_back(input_shape.dyn_dims().at(k++));
+                }
+            }
+            return {input_shape.type(), dyn_dims};
         }
+        else
+        {
+            auto type        = input_shape.type();
+            auto old_lens    = input_shape.lens();
+            auto old_strides = input_shape.strides();
+            if(input_shape.scalar())
+            {
+                if(old_lens.size() == 1 and old_lens.front() == 1)
+                    return shape{type, old_lens};
+                else
+                    MIGRAPHX_THROW("UNSQUEEZE: Input must be a scalar");
+            }
 
-        if(steps.size() > axes.size())
-            MIGRAPHX_THROW("UNSQUEEZE: Steps provided with no axis");
+            if(steps.size() > axes.size())
+                MIGRAPHX_THROW("UNSQUEEZE: Steps provided with no axis");
 
-        std::size_t new_size = old_lens.size() + axes.size();
+            std::size_t new_size = old_lens.size() + axes.size();
 
-        std::vector<std::size_t> new_lens(new_size);
-        std::vector<std::size_t> new_strides(new_size);
-        std::size_t p = 0;
-        for(auto i : range(new_size))
-        {
-            auto axis_idx = std::find(axes.begin(), axes.end(), i) - axes.begin();
-            if(axis_idx < axes.size())
+            std::vector<std::size_t> new_lens(new_size);
+            std::vector<std::size_t> new_strides(new_size);
+            std::size_t p = 0;
+            for(auto i : range(new_size))
             {
-                std::int64_t step = 1;
-                if(axis_idx < steps.size())
-                    step = steps[axis_idx];
-                if(step == 0)
-                    MIGRAPHX_THROW("UNSQUEEZE: step must be non-zero");
-                new_lens[i] = step;
-                if(p < old_strides.size())
+                auto axis_idx = std::find(axes.begin(), axes.end(), i) - axes.begin();
+                if(axis_idx < axes.size())
                 {
-                    if((old_lens[p] % step) != 0)
-                        MIGRAPHX_THROW("UNSQUEEZE: Axis dimenstion is not divisible by step");
-                    old_lens[p] /= step;
-                    new_strides[i] = old_strides[p] * old_lens[p];
+                    std::int64_t step = 1;
+                    if(axis_idx < steps.size())
+                        step = steps[axis_idx];
+                    if(step == 0)
+                        MIGRAPHX_THROW("UNSQUEEZE: step must be non-zero");
+                    new_lens[i] = step;
+                    if(p < old_strides.size())
+                    {
+                        if((old_lens[p] % step) != 0)
+                            MIGRAPHX_THROW("UNSQUEEZE: Axis dimenstion is not divisible by step");
+                        old_lens[p] /= step;
+                        new_strides[i] = old_strides[p] * old_lens[p];
+                    }
+                    else
+                    {
+                        if(step != 1)
+                            MIGRAPHX_THROW("UNSQUEEZE: Step must be 1 for extra axes");
+                        new_strides[i] = 1;
+                    }
                 }
                 else
                 {
-                    if(step != 1)
-                        MIGRAPHX_THROW("UNSQUEEZE: Step must be 1 for extra axes");
-                    new_strides[i] = 1;
+                    new_lens[i]    = old_lens[p];
+                    new_strides[i] = old_strides[p++];
                 }
             }
-            else
-            {
-                new_lens[i]    = old_lens[p];
-                new_strides[i] = old_strides[p++];
-            }
+            return shape{type, new_lens, new_strides};
         }
-        return shape{type, new_lens, new_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.hpp

@@ -101,6 +101,12 @@ struct shape
         friend bool operator==(const dynamic_dimension& x, const dynamic_dimension& y);
         friend bool operator!=(const dynamic_dimension& x, const dynamic_dimension& y);
         friend std::ostream& operator<<(std::ostream& os, const dynamic_dimension& x);
+
+        // compare to fixed std::size_t dimension
+        friend bool operator==(const dynamic_dimension& x, const std::size_t& y);
+        friend bool operator==(const std::size_t& x, const dynamic_dimension& y);
+        friend bool operator!=(const dynamic_dimension& x, const std::size_t& y);
+        friend bool operator!=(const std::size_t& x, const dynamic_dimension& y);
     };
 
     static const std::vector<type_t>& types();