manual merge

.github/workflows/ci.yaml

@@ -7,7 +7,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Cancel Previous Runs
-        uses: styfle/cancel-workflow-action@0.6.0
+        uses: styfle/cancel-workflow-action@0.11.0
         with:
           access_token: ${{ github.token }}
   tidy:
@@ -15,9 +15,19 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-
-    - uses: actions/checkout@v2
+      run: |
+        sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku 
+        du . --max-depth=1 -h
+        ls -la
+        cd /usr/local
+        du . --max-depth=1 -h
+        ls -la
+        cd /usr/local/lib
+        echo $(pwd)
+        du . --max-depth=1 -h
+        ls -la
+
+    - uses: actions/checkout@v3
 
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
@@ -34,7 +44,7 @@ jobs:
         message("::set-output name=timestamp::${current_date}")
 
     - name: Cache files for tidy
-      uses: pat-s/always-upload-cache@v2.1.3
+      uses: pat-s/always-upload-cache@v3.0.11
       with:
         path: tidy-cache
         key: tidy-cache-${{ steps.cache_timestamp.outputs.timestamp }}
@@ -65,8 +75,8 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
 
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
@@ -110,8 +120,8 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku 
+    - uses: actions/checkout@v3
 
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
@@ -146,10 +156,10 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
         python-version: 3.8
     - name: Install pyflakes
@@ -167,10 +177,10 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
         python-version: 3.8
     - name: run License Check
@@ -198,16 +208,16 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws*  /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
         python-version: 3.7
     - name: Cache dependencies
       # Ignore the failure of a step and avoid terminating the job.
       continue-on-error: true
-      uses: actions/cache@v2
+      uses: actions/cache@v3
       with:
         # This path is specific to Ubuntu
         path: ${{ github.workspace }}/cget
@@ -294,16 +304,16 @@ jobs:
 
     steps:
     - name: Free space
-      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android
-    - uses: actions/checkout@v2
+      run: sudo rm -rf /usr/local/android /usr/share/dotnet /usr/local/share/boost /opt/ghc /usr/local/share/chrom* /usr/share/swift /usr/local/julia* /usr/local/lib/android /usr/local/graalvm /usr/local/aws* /usr/local/lib/heroku
+    - uses: actions/checkout@v3
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v4
       with:
         python-version: 3.7
     - name: Cache dependencies
       # Ignore the failure of a step and avoid terminating the job.
       continue-on-error: true
-      uses: actions/cache@v2
+      uses: actions/cache@v3
       with:
         # This path is specific to Ubuntu
         path: ${{ github.workspace }}/cget

Dockerfile

@@ -74,7 +74,8 @@ RUN cget -p $PREFIX install facebook/zstd@v1.4.5 -X subdir -DCMAKE_DIR=build/cma
 RUN cget -p $PREFIX install ccache@v4.1 -DENABLE_TESTING=OFF
 
 # Install newer cmake for onnx runtime
-RUN cget -p /opt/cmake install kitware/cmake@v3.13.4
+ARG CMAKE_VERSION=3.24.2
+RUN cget -p /opt/cmake install -X binary https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-Linux-x86_64.tar.gz
 
 ARG ONNXRUNTIME_REPO=https://github.com/Microsoft/onnxruntime
 ARG ONNXRUNTIME_BRANCH=main

src/api/include/migraphx/migraphx.hpp

@@ -32,6 +32,7 @@
 #include <memory>
 #include <numeric>
 #include <exception>
+#include <array>
 #include <vector>
 #include <cassert>
 #include <iostream>

src/common.cpp

@@ -27,6 +27,7 @@
 #include <migraphx/algorithm.hpp>
 #include <migraphx/stringutils.hpp>
 #include <migraphx/instruction.hpp>
+#include <migraphx/ranges.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -43,6 +44,7 @@ inline namespace MIGRAPHX_INLINE_NS {
 // In this case we need to broadcast the (:,:,1:,:) axis
 // of s0 plus the 1st dimension of s1 giving
 // output_lens = (3,2,7,5)
+//
 std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
                                                   std::vector<std::size_t> s1)
 {
@@ -50,25 +52,63 @@ std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
         return s0;
     if(s0.size() > s1.size())
         s0.swap(s1);
-
     std::vector<std::size_t> out_lens(s1);
     auto offset = s1.size() - s0.size();
     std::transform(
         s0.begin(), s0.end(), s1.begin() + offset, out_lens.begin() + offset, [&](auto a, auto b) {
             if(a != b and a != 1 and b != 1)
             {
-                MIGRAPHX_THROW("COMPUTE_BROADCASTLEN: shape {" + to_string_range(s0) + "} and {" +
-                               to_string_range(s1) + "} mismatch!");
+                MIGRAPHX_THROW("COMPUTE_BROADCASTLEN: shape {" + migraphx::to_string_range(s0) +
+                               "} and {" + migraphx::to_string_range(s1) + "} mismatch!");
             }
             return std::max(a, b);
         });
-
     return out_lens;
 }
 
+std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, shape s1)
+{
+    // change both shapes to dynamic_dimension representation
+    s0 = s0.to_dynamic();
+    s1 = s1.to_dynamic();
+    if(s0.ndim() > s1.ndim())
+    {
+        std::swap(s0, s1);
+    }
+    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(),
+        s1.dyn_dims().cbegin() + offset,
+        out_dims.begin() + offset,
+        [&](auto a, auto b) {
+            if(a == b)
+            {
+                return a;
+            }
+            else if(a == one_dyn_dim or b == one_dyn_dim)
+            {
+                // 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};
+            }
+            else
+            {
+                MIGRAPHX_THROW("COMPUTE_BROADCASTED_DYN_DIMS: dynamic shapes {" +
+                               migraphx::to_string_range(s0.dyn_dims()) + "} and {" +
+                               migraphx::to_string_range(s1.dyn_dims()) + "} mismatch!");
+            }
+        });
+    return out_dims;
+}
+
+// Compute the common (broadcasted) dimensions of a list of fixed shapes
 std::vector<std::size_t> compute_common_lens(const std::vector<shape>& shapes)
 {
     assert(not shapes.empty());
+    assert(
+        std::none_of(shapes.cbegin(), shapes.cend(), [](auto shape) { return shape.dynamic(); }));
     return transform_accumulate(shapes.begin() + 1,
                                 shapes.end(),
                                 shapes.front().lens(),
@@ -114,6 +154,48 @@ instruction_ref insert_common_op(module& m,
                                  const operation& op,
                                  std::vector<instruction_ref> inputs)
 {
+    if(std::any_of(
+           inputs.cbegin(), inputs.cend(), [](auto input) { return input->get_shape().dynamic(); }))
+    {
+        // currently only handles the binary case
+        if(inputs.size() != 2)
+        {
+            MIGRAPHX_THROW("INSERT_COMMON_OP: not handled; " + migraphx::to_string(inputs.size()) +
+                           "inputs, only handle two inputs if any are dynamic shape");
+        }
+
+        auto c_type = compute_common_types(to_shapes(inputs));
+        auto c_dyn_dims =
+            compute_broadcasted_dyn_dims(inputs[0]->get_shape(), inputs[1]->get_shape());
+
+        // following should work for a static or dynamic shape
+        if(inputs[0]->get_shape().dyn_dims() != c_dyn_dims)
+        {
+            inputs[0] = m.insert_instruction(
+                ins,
+                make_op("multibroadcast", {{"out_dyn_dims", to_value(c_dyn_dims)}}),
+                inputs[0],
+                inputs[1]);
+        }
+        if(inputs[1]->get_shape().dyn_dims() != c_dyn_dims)
+        {
+            inputs[1] = m.insert_instruction(
+                ins,
+                make_op("multibroadcast", {{"out_dyn_dims", to_value(c_dyn_dims)}}),
+                inputs[1],
+                inputs[0]);
+        }
+        std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](auto input) {
+            if(input->get_shape().type() != c_type)
+            {
+                input =
+                    m.insert_instruction(ins, make_op("convert", {{"target_type", c_type}}), input);
+            }
+            return input;
+        });
+    }
+    else
+    {
         auto common = common_shape(to_shapes(inputs));
         std::transform(inputs.begin(), inputs.end(), inputs.begin(), [&](auto input) {
             if(input->get_shape().lens() != common.lens())
@@ -128,6 +210,7 @@ instruction_ref insert_common_op(module& m,
             }
             return input;
         });
+    }
     return m.insert_instruction(ins, op, inputs);
 }
 

src/include/migraphx/common.hpp

@@ -36,6 +36,9 @@ struct operation;
 
 std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
                                                   std::vector<std::size_t> s1);
+
+std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, shape s1);
+
 shape common_shape(const std::vector<shape>& shapes);
 
 instruction_ref insert_common_op(module& m,

src/include/migraphx/op/binary.hpp

@@ -28,6 +28,7 @@
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/argument.hpp>
 #include <migraphx/value.hpp>
+#include <migraphx/dyn_output.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -60,10 +61,19 @@ struct binary : op_name<Derived>
     value attributes() const { return base_attributes(); }
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, static_cast<const Derived&>(*this)}.has(2).same_type().same_dims();
+        check_shapes{inputs, static_cast<const Derived&>(*this), true}
+            .has(2)
+            .same_type()
+            .same_dims();
         auto s0 = inputs.at(0);
         auto s1 = inputs.at(1);
-        if(s0 == s1 and s0.packed())
+        if(s0.dynamic() or s1.dynamic())
+        {
+            if(s0 == s1)
+                return s0;
+            MIGRAPHX_THROW("BINARY: " + point_function() + ": fixed-dyn shape for inputs");
+        }
+        else if(s0 == s1 and s0.packed())
         {
             return s0;
         }
@@ -81,9 +91,9 @@ struct binary : 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
     {
-        argument result{output_shape};
+        argument result{dyn_out.computed_shape};
         visit_all(result, args[0], args[1])([&](auto output, auto input1, auto input2) {
             std::transform(input1.begin(),
                            input1.end(),

src/include/migraphx/op/broadcast.hpp

@@ -27,23 +27,30 @@
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/argument.hpp>
 #include <migraphx/config.hpp>
+#include <migraphx/dyn_output.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
-/// The broadcast operator performs the numpy-style broadcasting of an axis of a given tensor. This
-/// is achieved primarily by setting the stride of the broadcasted axis to zero. Linear indicies are
-/// computed from multi-indicies by computing the inner product on the multi-index with the strides.
-/// For example, if we have a tensor A(2,3) it has lengths of (2,3) and strides of (3,1). If we want
-/// to compute the linear offset that corresponds to the element on the 2nd row (i = 1) and 3rd
-/// column (j = 2), we compute the following inner product (1,2) dot (3, 1) = 1*3 + 2*1 = 5. It is
-/// obvious from there that we can negate the effects of a given axis by setting the stride of that
-/// axis to zero.
+/**
+ * 1 input version:
+ * Broadcasts a tensor from the original shape to the broadcast_lens by setting the stride of
+ * broadcasted dimensions to zero. `axis` attribute for a 1D input shape is the output dimension
+ * that stays the same. ex: broadcasting shape [1024] -> [4, 1024, 3] has axis = 1 For higher rank
+ * input shapes, axis is an offset parameter for the broadcasting. Such that this operator would
+ * work in the opposite direction of NumPy broadcasting. ex: broadcasting shape [2, 2] -> [2, 2, 3]
+ * with axis = 0
+ *
+ * 2 input version:
+ * Broadcast the first input 1D shape into the second input shape based on the axis parameter.
+ * Handles broadcasting a 1D static shape into a higher rank dynamic shape.
+ * broadcast_lens is not used
+ */
 struct broadcast
 {
     uint64_t axis                           = 0;
-    std::vector<std::size_t> broadcast_lens;
+    std::vector<std::size_t> broadcast_lens = {};
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
@@ -54,36 +61,86 @@ struct broadcast
     std::string name() const { return "broadcast"; }
     shape compute_shape(std::vector<shape> inputs) const
     {
-        auto input = inputs.at(0);
-        auto t     = input.type();
-
-        std::vector<size_t> bcast_strides(broadcast_lens.size(), 0);
-        // the broacast op is deprecated now, so not handling the negative
+        check_shapes{inputs, *this, true}.has(1, 2);
+        auto s0 = inputs.at(0);
+        auto t  = s0.type();
+        if(inputs.size() == 1)
+        {
+            // the ONNX broadcast op is deprecated now, so not handling the negative
             // value of axis anymore
             if(axis >= broadcast_lens.size())
             {
-            MIGRAPHX_THROW("BROADCAST : axis is out of range");
+                MIGRAPHX_THROW("BROADCAST : axis " + migraphx::to_string(axis) +
+                               " is out of range");
             }
-
-        if(broadcast_lens.size() - axis < input.lens().size())
+            if(broadcast_lens.size() - axis < s0.lens().size())
             {
-            MIGRAPHX_THROW("BROADCAST: (broadcast ndims - axis) is less than input ndims");
+                MIGRAPHX_THROW("BROADCAST: (broadcast ndims - axis) is less than s0 ndims");
             }
-
-        if(not std::equal(input.lens().begin(), input.lens().end(), broadcast_lens.begin() + axis))
+            if(not std::equal(s0.lens().begin(), s0.lens().end(), broadcast_lens.begin() + axis))
             {
                 MIGRAPHX_THROW("BROADCAST: when broadcasting, succeeding sizes must match");
             }
-        std::copy(input.strides().begin(), input.strides().end(), bcast_strides.begin() + axis);
 
+            std::vector<size_t> bcast_strides(broadcast_lens.size(), 0);
+            std::copy(s0.strides().begin(), s0.strides().end(), bcast_strides.begin() + axis);
             shape output{t, broadcast_lens, std::move(bcast_strides)};
-        if(output.elements() < input.elements())
-            MIGRAPHX_THROW("BROADCAST: output size must be greater than or equal to input size");
+            if(output.elements() < s0.elements())
+            {
+                // don't think this can occur?
+                MIGRAPHX_THROW("BROADCAST: output size must be greater than or equal to s0 size");
+            }
+            return output;
+        }
+        else
+        {
+            // two inputs
+            auto s1 = inputs.at(1);
+            if(s0.dynamic())
+            {
+                MIGRAPHX_THROW("BROADCAST_2in: s0 is a dynamic shape, does not handle broadcasting "
+                               "a dynamic shape");
+            }
+            if(s0.ndim() != 1)
+            {
+                MIGRAPHX_THROW("BROADCAST_2in: s0 has ndim " + migraphx::to_string(s0.ndim()) +
+                               ", only handle ndim = 1");
+            }
+            if(axis >= s1.ndim())
+            {
+                MIGRAPHX_THROW("BROADCAST_2in: axis " + migraphx::to_string(axis) +
+                               " is out of range");
+            }
+            if(s1.dynamic())
+            {
+                s0 = s0.to_dynamic();
+                if(s0.dyn_dims()[0] != s1.dyn_dims()[axis])
+                {
+                    MIGRAPHX_THROW("BROADCAST_2in: s0 length doesn't match with dynamic s1 axis "
+                                   "dimension length (" +
+                                   migraphx::to_string(s0.dyn_dims()[0]) +
+                                   " != " + migraphx::to_string(s1.dyn_dims()[axis]) + ")");
+                }
+                return s1;
+            }
+
+            if(s0.lens()[0] != s1.lens()[axis])
+            {
+                MIGRAPHX_THROW("BROADCAST_2in: s0 length doesn't match with static s1 axis "
+                               "dimension length (" +
+                               migraphx::to_string(s0.lens()[0]) +
+                               " != " + migraphx::to_string(s1.lens()[axis]) + ")");
+            }
+            std::vector<size_t> bcast_strides(s1.ndim(), 0);
+            std::copy(s0.strides().begin(), s0.strides().end(), bcast_strides.begin() + axis);
+            shape output{t, s1.lens(), std::move(bcast_strides)};
             return output;
         }
-    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/multibroadcast.hpp

@@ -26,64 +26,105 @@
 
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/argument.hpp>
+#include <migraphx/dyn_output.hpp>
+#include <migraphx/common.hpp>
 #include <migraphx/config.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
 
+/**
+ * Broadcast multiple dimensions between two tensors.
+ * Two versions of this operator: one input and two inputs.
+ * One input version uses output_lens attribute and broadcasts to it.
+ * Two inputs version broadcasts both inputs to the common shape at evaluation time.
+ */
 struct multibroadcast
 {
-    std::vector<std::size_t> output_lens;
+    std::vector<std::size_t> output_lens = {};
+
+    // optional attribute
+    std::vector<shape::dynamic_dimension> output_dyn_dims = {};
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.output_lens, "out_lens"));
+        return pack(f(self.output_lens, "out_lens"), f(self.output_dyn_dims, "out_dyn_dims"));
     }
 
     std::string name() const { return "multibroadcast"; }
 
     shape compute_shape(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this}.has(1);
+        check_shapes{inputs, *this, true}.has(1, 2);
+
         auto t  = inputs.at(0).type();
-        auto input = inputs.at(0);
+        auto s0 = inputs.at(0);
 
-        if(input.lens().empty())
+        if(s0.max_lens().empty())
         {
-            MIGRAPHX_THROW("MULTIBROADCAST: inputs dimensions should be > 0");
+            MIGRAPHX_THROW("MULTIBROADCAST: input dimensions should be > 0");
         }
 
-        if(input.lens().size() > output_lens.size())
+        auto make_bcast_strides = [&](std::vector<std::size_t> bcast_lens, std::size_t offset) {
+            std::vector<size_t> bcast_strides(bcast_lens.size(), 0);
+            for(std::ptrdiff_t i = s0.lens().size() - 1; i >= 0; i--)
+            {
+                if(bcast_lens[i + offset] == s0.lens()[i])
                 {
-            MIGRAPHX_THROW("MULTIBROADCAST: inputs dimensions should <= output size");
+                    bcast_strides[i + offset] = s0.strides()[i];
                 }
+            }
+            return bcast_strides;
+        };
 
-        auto offset = output_lens.size() - input.lens().size();
-        for(std::ptrdiff_t i = input.lens().size() - 1; i >= 0; i--)
+        if(inputs.size() == 1)
+        {
+            if(s0.lens().size() > output_lens.size())
             {
-            if(output_lens[i + offset] != input.lens()[i] and input.lens()[i] != 1)
+                MIGRAPHX_THROW("MULTIBROADCAST: input dimensions should <= output size");
+            }
+
+            auto offset = output_lens.size() - s0.lens().size();
+            for(std::ptrdiff_t i = s0.lens().size() - 1; i >= 0; i--)
             {
-                MIGRAPHX_THROW("MULTIBROADCAST: input shape {" + to_string_range(input.lens()) +
+                if(output_lens[i + offset] != s0.lens()[i] and s0.lens()[i] != 1)
+                {
+                    MIGRAPHX_THROW("MULTIBROADCAST: input shape {" + to_string_range(s0.lens()) +
                                    "} cannot be broadcasted to {" + to_string_range(output_lens) +
                                    "}!");
                 }
             }
 
-        std::vector<size_t> bcast_strides(output_lens.size(), 0);
-        for(std::ptrdiff_t i = input.lens().size() - 1; i >= 0; i--)
+            auto bcast_strides = make_bcast_strides(output_lens, offset);
+            return {t, output_lens, std::move(bcast_strides)};
+        }
+        else
         {
-            if(output_lens[i + offset] == input.lens()[i])
+            // two inputs
+            auto s1 = inputs.at(1);
+            if(s0.dynamic() or s1.dynamic())
             {
-                bcast_strides[i + offset] = input.strides()[i];
+                if(not output_dyn_dims.empty())
+                {
+                    return {t, output_dyn_dims};
                 }
+                return {t, compute_broadcasted_dyn_dims(s0, s1)};
             }
-        return {t, output_lens, bcast_strides};
+            else
+            {
+                auto bcast_lens    = compute_broadcasted_lens(s0.lens(), s1.lens());
+                auto offset        = bcast_lens.size() - s0.lens().size();
+                auto bcast_strides = make_bcast_strides(bcast_lens, offset);
+                return {t, std::move(bcast_lens), std::move(bcast_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

@@ -30,6 +30,7 @@
 #include <numeric>
 #include <memory>
 
+#include <migraphx/functional.hpp>
 #include <migraphx/errors.hpp>
 #include <migraphx/half.hpp>
 #include <migraphx/config.hpp>
@@ -89,7 +90,10 @@ struct shape
         std::size_t opt = 0;
 
         template <class Self, class F>
-        static auto reflect(Self& self, F f);
+        static auto reflect(Self& self, F f)
+        {
+            return pack(f(self.min, "min"), f(self.max, "max"), f(self.opt, "opt"));
+        }
 
         bool is_fixed() const;
         bool has_optimal() const;
@@ -115,6 +119,12 @@ struct shape
 
     shape(type_t t, std::vector<dynamic_dimension> dims);
 
+    // Construct a dynamic shape from three sets of lengths (of the same rank)
+    shape(type_t t,
+          std::vector<std::size_t> mins,
+          std::vector<std::size_t> maxes,
+          std::vector<std::size_t> opts);
+
     template <class Range>
     shape(type_t t, const Range& l) : shape(t, std::vector<std::size_t>(l.begin(), l.end()))
     {
@@ -136,6 +146,12 @@ struct shape
     const std::vector<std::size_t>& lens() const;
     const std::vector<std::size_t>& strides() const;
 
+    /*!
+     * The number of dimensions in the shape.
+     * Same as the number of indices required to get a data value.
+     */
+    std::size_t ndim() const;
+
     /*!
      * Return the number of elements in the tensor.
      */
@@ -221,6 +237,9 @@ struct shape
 
     shape with_type(type_t t) const;
 
+    // convert the shape to an equivalent dynamic shape
+    shape to_dynamic() const;
+
     friend bool operator==(const shape& x, const shape& y);
     friend bool operator!=(const shape& x, const shape& y);
     friend std::ostream& operator<<(std::ostream& os, const shape& x);

src/onnx/parse_batchnorm.cpp

@@ -44,7 +44,7 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
         {
             epsilon = parser.parse_value(info.attributes.at("epsilon")).at<float>();
         }
-        auto x_lens = args[0]->get_shape().lens();
+        auto x_lens = args[0]->get_shape().max_lens();
         auto x_type = args[0]->get_shape().type();
 
         if(std::any_of(args.cbegin() + 1, args.cend(), [](auto a) {

src/onnx/parse_binary_op.cpp

@@ -57,6 +57,12 @@ struct parse_binary_op : op_parser<parse_binary_op>
                 parser.parse_value(info.attributes.at("broadcast")).at<uint64_t>();
             if(broadcasted != 0)
             {
+                if(std::any_of(
+                       args.cbegin(), args.cend(), [](auto a) { return a->get_shape().dynamic(); }))
+                {
+                    MIGRAPHX_THROW(
+                        "Binary op broadcast attribute not supported for dynamic input shapes");
+                }
                 uint64_t axis = parser.parse_value(info.attributes.at("axis")).at<uint64_t>();
                 auto l        = info.add_instruction(
                     make_op("broadcast",

src/onnx/parse_split.cpp

@@ -26,6 +26,9 @@
 #include <migraphx/ranges.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/tune_axis.hpp>
+#include <migraphx/onnx/checks.hpp>
+
+#include <migraphx/stringutils.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -55,12 +58,12 @@ struct parse_split : op_parser<parse_split>
         {
             literal s = parser.parse_value(info.attributes.at("split"));
             s.visit([&](auto v) { vec_splits.assign(v.begin(), v.end()); });
-
-            if(std::accumulate(vec_splits.begin(), vec_splits.end(), int64_t(0)) !=
-               static_cast<int64_t>(lens[tuned_axis]))
-            {
-                MIGRAPHX_THROW("PARSE_SPLIT: sum of split attribute unequal to dim size of axis!");
         }
+        else if(args.size() == 2)
+        {
+            auto s = args[1]->eval();
+            check_arg_empty(s, "Split: dynamic shape is not supported");
+            s.visit([&](auto v) { vec_splits.assign(v.begin(), v.end()); });
         }
         // no split attribute, input is equally divided
         else
@@ -74,6 +77,15 @@ struct parse_split : op_parser<parse_split>
             vec_splits.resize(info.num_outputs, dl);
         }
 
+        if(std::accumulate(vec_splits.begin(), vec_splits.end(), int64_t(0)) !=
+           static_cast<int64_t>(lens[tuned_axis]))
+        {
+            MIGRAPHX_THROW(
+                "PARSE_SPLIT: sum of split attribute unequal to dim size of axis! tuned axis:" +
+                std::to_string(lens[tuned_axis]) + " Output " + to_string_range(vec_splits) +
+                " Rank " + std::to_string(n_rank) + " Len outs " + to_string_range(lens));
+        }
+
         std::vector<instruction_ref> ret_ins;
         int64_t start = 0;
         for(auto sl : vec_splits)

src/pass_manager.cpp

@@ -94,11 +94,19 @@ struct module_pm : module_pass_manager
     virtual void run_pass(const pass& p) override
     {
         assert(mod);
+
+        timer ts{};
+        using seconds = std::chrono::duration<double>;
+
         trace("Module: ", mod->name(), ", Pass: ", p.name());
+        const double t1 = ts.record<seconds>();
         assert(mod->validate() == mod->end());
         p.apply(*this);
         trace(*mod);
         validate_pass(*mod, p, *t);
+
+        const double t2 = ts.record<seconds>();
+        trace("Pass: ", p.name(), " completed in (s): ", (t2 - t1));
     }
 };
 

src/shape.cpp

@@ -71,6 +71,19 @@ struct shape_impl
     {
     }
 
+    shape_impl(shape::type_t t,
+               std::vector<std::size_t> mins,
+               std::vector<std::size_t> maxes,
+               std::vector<std::size_t> opts)
+        : m_type(t)
+    {
+        assert(mins.size() == maxes.size() and maxes.size() == opts.size());
+        for(size_t i = 0; i < mins.size(); ++i)
+        {
+            m_dyn_dims.push_back(shape::dynamic_dimension{mins[i], maxes[i], opts[i]});
+        }
+    }
+
     shape_impl(const std::vector<shape>& subs) : m_type(shape::tuple_type), m_shapes(subs) {}
 
     shape::type_t m_type;
@@ -224,6 +237,14 @@ shape::shape(type_t t, std::vector<shape::dynamic_dimension> dims)
 {
 }
 
+shape::shape(type_t t,
+             std::vector<std::size_t> mins,
+             std::vector<std::size_t> maxes,
+             std::vector<std::size_t> opts)
+    : impl(std::make_shared<shape_impl>(t, std::move(mins), std::move(maxes), std::move(opts)))
+{
+}
+
 shape::shape(const std::vector<shape>& subs) : impl(std::make_shared<shape_impl>(subs)) {}
 
 shape::shape(std::shared_ptr<shape_impl> pimpl) : impl(std::move(pimpl)) {}
@@ -244,6 +265,15 @@ const std::vector<std::size_t>& shape::lens() const { return impl->m_lens; }
 
 const std::vector<std::size_t>& shape::strides() const { return impl->m_strides; }
 
+std::size_t shape::ndim() const
+{
+    if(this->dynamic())
+    {
+        return dyn_dims().size();
+    }
+    return lens().size();
+}
+
 std::size_t shape::elements() const { return impl->elements(); }
 
 std::size_t shape::bytes() const
@@ -437,6 +467,16 @@ shape shape::with_type(type_t t) const
     return {c};
 }
 
+shape shape::to_dynamic() const
+{
+    if(this->dynamic())
+    {
+        return *this;
+    }
+    std::vector<std::size_t> zeroes(this->ndim(), 0);
+    return {type(), lens(), lens(), zeroes};
+}
+
 std::size_t shape::element_space() const { return impl->element_space(); }
 
 std::string shape::type_string() const { return name(this->type()); }
@@ -464,15 +504,11 @@ bool shape::dynamic_dimension::is_fixed() const { return this->min == this->max;
 
 bool shape::dynamic_dimension::has_optimal() const { return opt != 0; }
 
-template <class Self, class F>
-auto shape::dynamic_dimension::reflect(Self& self, F f)
-{
-    return pack(f(self.min, "min"), f(self.max, "max"), f(self.opt, "opt"));
-}
-
 bool operator==(const shape::dynamic_dimension& x, const shape::dynamic_dimension& y)
 {
-    return (x.min == y.min and x.max == y.max and x.opt == y.opt);
+    // don't check opt if both are fixed
+    return (x.min == y.min and x.max == y.max and
+            ((x.is_fixed() and y.is_fixed()) or (x.opt == y.opt)));
 }
 
 bool operator!=(const shape::dynamic_dimension& x, const shape::dynamic_dimension& y)

src/targets/gpu/compile_miopen.cpp

@@ -64,7 +64,7 @@ std::size_t compile_miopen::compile(operation& op, instruction_ref ins, bool for
 {
     op.from_value({{"int8_x4_format", format}});
     auto v = op.compile(*ctx, ins->get_shape(), to_shapes(ins->inputs()));
-    return v.get("workspace", 0);
+    return v.get<std::size_t>("workspace", 0);
 }
 
 void compile_miopen::apply(module& m) const

src/targets/gpu/compile_ops.cpp

@@ -40,18 +40,25 @@ MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_GPU_COMPILE_PARALLEL);
 struct precompile_op
 {
     operation op                = op::identity{};
+    std::size_t additional_args = 1;
+    bool ignore_modules         = false;
 
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
-        return pack(f(self.op, "op"));
+        return pack(f(self.op, "op"),
+                    f(self.additional_args, "additional_args"),
+                    f(self.ignore_modules, "ignore_modules"));
     }
 
     std::string name() const { return "gpu::precompile_op"; }
 
     shape compute_shape(std::vector<shape> inputs, const std::vector<module_ref>& mods) const
     {
-        inputs.pop_back();
+        // Pop off additional args
+        inputs.resize(inputs.size() - additional_args);
+        if(ignore_modules)
+            return op.compute_shape(inputs);
         return op.compute_shape(inputs, mods);
     }
 

src/targets/gpu/fuse_ops.cpp

@@ -772,11 +772,9 @@ struct find_layernorm_pointwise
     {
         auto ins       = r.result;
         auto layernorm = r.instructions["layernorm"];
-        auto* pm       = ins->module_inputs().front();
-
         if(not layernorm->module_inputs().empty())
             return;
-
+        auto* pm    = ins->module_inputs().front();
         auto inputs = layernorm->inputs();
         inputs.pop_back();
         inputs.insert(inputs.end(), ins->inputs().begin() + 1, ins->inputs().end());
@@ -785,6 +783,37 @@ struct find_layernorm_pointwise
     }
 };
 
+struct find_concat_pointwise
+{
+    auto matcher() const
+    {
+        return precompile_name("pointwise")(
+            match::arg(0)(precompile_name("concat").bind("concat")));
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins    = r.result;
+        auto concat = r.instructions["concat"];
+        if(not concat->module_inputs().empty())
+            return;
+
+        // TODO: Handle type conversions
+        if(ins->get_shape().type() != concat->get_shape().type())
+            return;
+
+        auto* pm    = ins->module_inputs().front();
+        auto inputs = concat->inputs();
+        inputs.pop_back();
+        inputs.insert(inputs.end(), ins->inputs().begin() + 1, ins->inputs().end());
+
+        auto op = concat->get_operator();
+        op.from_value({{"additional_args", ins->inputs().size() - 1}, {"ignore_modules", true}});
+
+        m.replace_instruction(ins, op, inputs, {pm});
+    }
+};
+
 void fuse_ops::apply(module& m) const
 {
     match::find_matches(m, find_contiguous_pointwise{});
@@ -793,6 +822,7 @@ void fuse_ops::apply(module& m) const
     run_passes(m, {dead_code_elimination{}});
     match::find_matches(m,
                         find_layernorm_pointwise{},
+                        find_concat_pointwise{},
                         find_gemm_pointwise{},
                         find_contiguous_tranpose_gemm{},
                         find_commutative_broadcast{});

src/targets/gpu/gemm_impl.cpp

@@ -21,7 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-#include <rocblas.h>
+#include <rocblas/rocblas.h>
 #include <migraphx/gpu/gemm_impl.hpp>
 #include <migraphx/reduce_dims.hpp>
 #include <migraphx/permutation.hpp>

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

@@ -25,7 +25,7 @@
 #define MIGRAPHX_GUARD_MIGRAPHLIB_ROCBLAS_HPP
 #include <migraphx/manage_ptr.hpp>
 #include <migraphx/config.hpp>
-#include <rocblas.h>
+#include <rocblas/rocblas.h>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

src/targets/gpu/jit/concat.cpp

@@ -38,16 +38,19 @@ using namespace migraphx::gpu::gen; // NOLINT
 static const char* const concat_kernel = R"__migraphx__(
 #include <migraphx/kernels/concat.hpp>
 #include <migraphx/kernels/vectorize.hpp>
+#include <migraphx/kernels/ops.hpp>
 #include <args.hpp>
 
 namespace migraphx {
 
+${preamble}
+
 extern "C" {
 
 __global__ void ${kernel}(${params}) 
 {
-    transform_args(make_tensors(), rotate_last(), ${transformers})(${args})([](auto y, auto... xs) {
-        concat<${axis}>(y, xs...);
+    transform_args(make_tensors(), rotate_last(), ${transformers})(${args})([](auto y, ${concat_params}, auto... xs) {
+        concat<${axis}>(${concat_args})(${post}, y, xs...);
     });
 }
 
@@ -68,28 +71,42 @@ struct concat_compiler : compiler<concat_compiler>
 
     operation compile_op(context& ctx, const std::vector<shape>& inputs, const value& v) const
     {
-        // TODO: Use reduce_dims
+        auto num_of_concat_inputs = v.get("concat_inputs", inputs.size() - 1);
         hip_compile_options options;
         options.inputs      = inputs;
         options.output      = inputs.back();
         options.params      = "-Wno-float-equal";
+        options.kernel_name = v.get("kernel", "concat_kernel");
         auto axis           = find_fast_axis(options.inputs);
         auto vec            = vectorize::elements(ctx, axis, options.inputs);
-        options.kernel_name = v.get("kernel", "concat_kernel");
         options.set_launch_params(
             v, compute_global_for(ctx, get_concat_elements(options.inputs) / vec.size, 256));
-        auto src = interpolate_string(concat_kernel,
+        auto src = interpolate_string(
+            concat_kernel,
             {{"kernel", options.kernel_name},
              {"params", enum_params(inputs.size(), "void * private_p")},
              {"args", enum_params(inputs.size(), "private_p")},
+             {"concat_params", enum_params(num_of_concat_inputs, "auto concat_x")},
+             {"concat_args", enum_params(num_of_concat_inputs, "concat_x")},
+             {"post", v.get("post", std::string{"op::id{}"})},
              {"transformers", make_transformer_args(vec)},
+             {"preamble", v.get("preamble", std::string{})},
              {"axis", v.at("axis").to<std::string>()}});
         return compile_hip_code_object(src, options);
     }
 
     compiler_replace compile(context& ctx, instruction_ref ins, const operation& op) const
     {
-        return replace(compile_op(ctx, to_shapes(ins->inputs()), op.to_value()));
+        auto v = op.to_value();
+        if(not ins->module_inputs().empty())
+        {
+            auto* pm           = ins->module_inputs().front();
+            v["concat_inputs"] = ins->inputs().size() - pm->get_parameter_names().size();
+            v["preamble"]      = generate_pointwise(*pm, "post_concat");
+            v["post"]          = "MIGRAPHX_LIFT(post_concat)";
+            v["kernel"]        = "concat_" + generate_name_from_ops(*pm) + "_kernel";
+        }
+        return replace(compile_op(ctx, to_shapes(ins->inputs()), v));
     }
 };