Merge from develop

src/include/migraphx/operators.hpp

@@ -35,7 +35,6 @@
 #include <migraphx/op/as_shape.hpp>
 #include <migraphx/op/atan.hpp>
 #include <migraphx/op/atanh.hpp>
-#include <migraphx/op/batch_norm_inference.hpp>
 #include <migraphx/op/binary.hpp>
 #include <migraphx/op/broadcast.hpp>
 #include <migraphx/op/capture.hpp>

src/include/migraphx/pad_calc.hpp

@@ -24,9 +24,10 @@
 #ifndef MIGRAPHX_GUARD_OPERATORS_PAD_CALC_HPP
 #define MIGRAPHX_GUARD_OPERATORS_PAD_CALC_HPP
 
-#include <migraphx/config.hpp>
 #include <cstdint>
 #include <vector>
+#include <migraphx/config.hpp>
+#include <migraphx/shape.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -42,18 +43,21 @@ void calculate_padding(int64_t idx,
 /*!
  * Calculate the padding for auto_padding. Used for dynamic shapes
  * where the padding calculation must be done at evaluation time.
- * \param tensor_lens input tensor image shape
- * \param k_lens weights kernel shape
- * \param strides strides for the kernel
- * \param dilations dilations for the kernel
- * \param use_upper put odd padding on upper or lower side
  * \return padding in the form of {x0_begin, x1_begin, ... x0_end , x1_end, ...}
  */
-std::vector<std::size_t> calc_dyn_auto_pad(std::vector<std::size_t> tensor_lens,
-                                           std::vector<std::size_t> k_lens,
-                                           std::vector<std::size_t> strides,
-                                           std::vector<std::size_t> dilations,
-                                           bool use_upper = true);
+std::vector<std::size_t> calc_dyn_auto_pad(const std::vector<std::size_t>& input_lens,
+                                           const std::vector<std::size_t>& wei_lens,
+                                           const std::vector<std::size_t>& strides,
+                                           const std::vector<std::size_t>& dilations,
+                                           bool use_upper);
+
+// Used for dynamic auto padding of convolution operators since padding needs to be computed at
+// evaulation time.
+shape compute_padded_shape(const shape& input,
+                           const shape& weights,
+                           const std::vector<std::size_t>& padding,
+                           const std::vector<std::size_t>& stride,
+                           const std::vector<std::size_t>& dilation);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/reflect.hpp

@@ -56,11 +56,11 @@ auto reflect_impl(rank<0>, T&, Selector)
 }
 
 template <class T>
-auto reflectable_impl(rank<1>, T&& x)
+auto reflectable_impl(rank<1>, const T& x)
     -> decltype(T::reflect(x, reflect_placeholder{}), std::true_type{});
 
 template <class T>
-auto reflectable_impl(rank<0>, T &&) -> decltype(std::false_type{});
+auto reflectable_impl(rank<0>, const T&) -> decltype(std::false_type{});
 
 template <class T>
 struct remove_rvalue_reference
@@ -111,8 +111,18 @@ auto reflect(T& x, Selector f)
 template <class T>
 auto reflect_tie(T& x)
 {
-    return reflect(x, [](auto&& y, auto&&...) { return detail::wrap<decltype(y)>(y); })(
-        [](auto&&... xs) { return detail::auto_tuple(xs.get()...); });
+    return reflect(x, [](auto&& y, auto&&...) {
+        // cppcheck-suppress UnnecessaryElseStatement
+        if constexpr(is_reflectable<decltype(y)>{})
+        {
+            auto t = reflect_tie(y);
+            return detail::wrap<decltype(t)>(t);
+        }
+        else
+        {
+            return detail::wrap<decltype(y)>(y);
+        }
+    })([](auto&&... xs) { return detail::auto_tuple(xs.get()...); });
 }
 
 template <class T, class F>

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/include/migraphx/streamutils.hpp

@@ -26,7 +26,9 @@
 
 #include <ostream>
 #include <algorithm>
+#include <migraphx/reflect.hpp>
 #include <migraphx/rank.hpp>
+#include <migraphx/requires.hpp>
 #include <migraphx/config.hpp>
 #include <vector>
 
@@ -83,6 +85,20 @@ auto stream_write_value_impl(rank<0>, std::ostream& os, const Range& r)
     os << "}";
 }
 
+template <class T, MIGRAPHX_REQUIRES(is_reflectable<T>{})>
+void stream_write_value_impl(rank<0>, std::ostream& os, const T& x)
+{
+    char delim = '{';
+    reflect_each(x, [&](auto&& y, auto name) {
+        os << delim;
+        os << name << "=";
+        stream_write_value_impl(rank<2>{}, os, y);
+        delim = ',';
+    });
+    if(delim == ',')
+        os << "}";
+}
+
 } // namespace detail
 
 template <class T>

src/rewrite_batchnorm.cpp → src/layout_nhwc.cpp

@@ -21,63 +21,98 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-#include <migraphx/rewrite_batchnorm.hpp>
-#include <migraphx/program.hpp>
+#include <migraphx/layout_nhwc.hpp>
+#include <migraphx/module.hpp>
 #include <migraphx/instruction.hpp>
-#include <migraphx/op/batch_norm_inference.hpp>
-#include <migraphx/op/broadcast.hpp>
-#include <migraphx/op/add.hpp>
-#include <migraphx/op/mul.hpp>
 #include <migraphx/iterator_for.hpp>
+#include <migraphx/permutation.hpp>
+#include <migraphx/functional.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/make_op.hpp>
-
-#include <migraphx/dfor.hpp>
+#include <migraphx/eliminate_contiguous.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/pass_manager.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-void rewrite_batchnorm::apply(module& m) const
+template <class Predicate>
+std::vector<instruction_ref> find_lasts(const module& m, Predicate pred)
+{
+    std::vector<instruction_ref> result;
+    fix([&](auto self, auto ins) {
+        if(pred(ins))
+        {
+            result.push_back(ins);
+            return;
+        }
+        for(auto input : ins->inputs())
+            self(input);
+    })(std::prev(m.end()));
+    return result;
+}
+
+std::unordered_set<instruction_ref> preserve_output_layout(module& m)
+{
+    std::unordered_set<instruction_ref> result;
+    std::vector<instruction_ref> outputs = find_lasts(m, [](auto ins) {
+        return ins->name() == "convolution" and ins->get_shape().lens().size() == 4;
+    });
+    for(auto output : outputs)
+    {
+        auto permutation = find_permutation(output->get_shape());
+        auto layout      = m.insert_instruction(
+            std::next(output), make_op("layout", {{"permutation", permutation}}), output);
+        result.insert(m.replace_instruction(output, layout));
+    }
+    return result;
+}
+
+void transform_convolutions(module& m)
 {
     for(auto ins : iterator_for(m))
     {
-        if(ins->name() != "batch_norm_inference")
+        if(ins->name() != "convolution")
             continue;
-        // Get scale, bias, mean, variance from inputs
-        auto gamma    = ins->inputs()[1]->eval();
-        auto bias     = ins->inputs()[2]->eval();
-        auto mean     = ins->inputs()[3]->eval();
-        auto variance = ins->inputs()[4]->eval();
-        if(any_of({gamma, bias, mean, variance}, [](auto arg) { return arg.empty(); }))
+        if(ins->get_shape().lens().size() != 4)
             continue;
+        auto v = ins->get_operator().to_value();
+        if(v.at("group").to<int>() > 1)
+            continue;
+        auto args = ins->inputs();
+        std::transform(args.begin(), args.end(), args.begin(), [&](const auto& i) {
+            return m.insert_instruction(ins, make_op("layout", {{"permutation", {0, 2, 3, 1}}}), i);
+        });
+        auto conv = m.insert_instruction(ins, ins->get_operator(), args);
+        auto c    = m.insert_instruction(ins, make_op("contiguous"), conv);
+        m.replace_instruction(ins, c);
+    }
+}
 
-        std::vector<std::size_t> lens = ins->inputs()[1]->get_shape().lens();
-        shape s{ins->get_shape().type(), lens};
-        // Get epsilon
-        auto bn_op   = any_cast<op::batch_norm_inference>(ins->get_operator());
-        auto epsilon = bn_op.epsilon;
-
-        argument a{s};
-        argument b{s};
-        visit_all(gamma, bias, mean, variance, a, b)(
-            [&](auto gamma2, auto bias2, auto mean2, auto variance2, auto a2, auto b2) {
-                dfor(a.get_shape().elements())(
-                    [&](std::size_t c) { a2[c] = gamma2[c] / std::sqrt(variance2[c] + epsilon); });
-                dfor(b.get_shape().elements())([&](std::size_t c) {
-                    b2[c] = bias2[c] - (gamma2[c] * mean2[c] / std::sqrt(variance2[c] + epsilon));
-                });
-            });
-
-        auto broadcast   = op::broadcast{1, ins->get_shape().lens()};
-        auto a_ins       = m.add_literal({a.get_shape(), a.data()});
-        auto a_broadcast = m.insert_instruction(ins, broadcast, a_ins);
-        auto mul   = m.insert_instruction(ins, make_op("mul"), ins->inputs().front(), a_broadcast);
-        auto b_ins = m.add_literal({b.get_shape(), b.data()});
-        auto b_broadcast = m.insert_instruction(ins, broadcast, b_ins);
-        auto add         = m.insert_instruction(ins, make_op("add"), mul, b_broadcast);
-        m.replace_instruction(ins, add);
+void remove_layout(module& m, const std::unordered_set<instruction_ref>& output_layouts)
+{
+    for(auto ins : iterator_for(m))
+    {
+        if(ins->name() != "layout")
+            continue;
+        if(ins->get_shape() != ins->inputs().front()->get_shape())
+            continue;
+        if(contains(output_layouts, ins))
+            continue;
+        m.replace_instruction(ins, ins->inputs().front());
     }
 }
 
+void layout_nhwc::apply(module_pass_manager& mpm) const
+{
+    std::unordered_set<instruction_ref> output_layouts = preserve_output_layout(mpm.get_module());
+    transform_convolutions(mpm.get_module());
+    mpm.run_pass(dead_code_elimination{});
+    mpm.run_pass(eliminate_contiguous{"contiguous"});
+    mpm.run_pass(dead_code_elimination{});
+    remove_layout(mpm.get_module(), output_layouts);
+    mpm.run_pass(dead_code_elimination{});
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/load_save.cpp

@@ -25,7 +25,6 @@
 #include <migraphx/file_buffer.hpp>
 #include <migraphx/json.hpp>
 #include <migraphx/msgpack.hpp>
-#include <migraphx/file_buffer.hpp>
 #include <fstream>
 
 namespace migraphx {

src/module.cpp

@@ -34,7 +34,6 @@
 #include <migraphx/pass_manager.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/register_target.hpp>
-#include <migraphx/make_op.hpp>
 #include <migraphx/json.hpp>
 #include <iostream>
 #include <sstream>

src/onnx/conv.cpp

@@ -30,7 +30,7 @@ namespace onnx {
 
 void recalc_conv_attributes(value& v, size_t kdims)
 {
-    if(not(v["padding"].size() == kdims or v["padding"].size() == kdims * 2))
+    if(v["padding"].size() != kdims and v["padding"].size() != kdims * 2)
     {
         v["padding"].resize(kdims);
         std::fill_n(v["padding"].begin(), kdims, 0);

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) {
@@ -54,18 +54,19 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
             MIGRAPHX_THROW("PARSE_BATCHNORM: argument scale, bias, mean, or var rank != 1");
         }
 
-        if(x_lens.size() == 1)
+        auto x_rank = x_lens.size();
+        if(x_rank == 1 or x_rank == 2)
         {
-            auto rt   = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
-            auto eps  = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
-            auto n0   = info.add_broadcastable_binary_op("sub", args[0], args[3]);
-            auto d0   = info.add_broadcastable_binary_op("add", args[4], eps);
-            auto d1   = info.add_broadcastable_binary_op("pow", d0, rt);
-            auto div0 = info.add_broadcastable_binary_op("div", n0, d1);
-            auto r0   = info.add_broadcastable_binary_op("mul", div0, args[1]);
+            auto rt      = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
+            auto eps     = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
+            auto numer   = info.add_broadcastable_binary_op("sub", args[0], args[3]);
+            auto var_eps = info.add_broadcastable_binary_op("add", args[4], eps);
+            auto denom   = info.add_broadcastable_binary_op("pow", var_eps, rt);
+            auto div0    = info.add_broadcastable_binary_op("div", numer, denom);
+            auto r0      = info.add_broadcastable_binary_op("mul", div0, args[1]);
             return info.add_broadcastable_binary_op("add", r0, args[2]);
         }
-        else if(x_lens.size() > 2)
+        else if(x_rank > 2)
         {
             // unsqueeze tensors of shape (C) to broadcast correctly
             std::vector<int64_t> unsqueeze_axes(x_lens.size() - 2);
@@ -89,7 +90,7 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
         }
         else
         {
-            // num dims either 0 or 2
+            // rank == 0
             MIGRAPHX_THROW("PARSE_BATCHNORM: rank " + std::to_string(x_lens.size()) +
                            " input tensor, unhandled data format");
         }

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_convolution.cpp

@@ -125,11 +125,9 @@ struct parse_convolution : op_parser<parse_convolution>
                     values["padding_mode"] = is_same_upper
                                                  ? to_value(op::padding_mode_t::same_upper)
                                                  : to_value(op::padding_mode_t::same_lower);
-                    values["use_dynamic_same_auto_pad"] = true;
                 }
                 else
                 {
-                    values["padding_mode"] = to_value(op::padding_mode_t::same);
                     // kernel shape will be fixed, so max_lens() == min_len() for kernel lengths
                     auto weight_lens = weights->get_shape().max_lens();
                     std::vector<std::size_t> k_lens(weight_lens.begin() + 2, weight_lens.end());

src/onnx/parse_deconvolution.cpp

@@ -95,6 +95,8 @@ struct parse_deconvolution : op_parser<parse_deconvolution>
             check_attr_sizes(
                 kdims, values["dilation"].size(), "PARSE_CONV_TRANSPOSE: inconsistent dilations");
         }
+
+        // TODO: auto padding needs to be implemented for this parser and operator
         if(contains(info.attributes, "auto_pad"))
         {
             auto s = info.attributes["auto_pad"].s();
@@ -106,7 +108,9 @@ struct parse_deconvolution : op_parser<parse_deconvolution>
 
             if(s.find("SAME") != std::string::npos)
             {
-                values["padding_mode"] = to_value(op::padding_mode_t::same);
+                bool is_same_upper     = (s.find("SAME_UPPER") != std::string::npos);
+                values["padding_mode"] = is_same_upper ? to_value(op::padding_mode_t::same_upper)
+                                                       : to_value(op::padding_mode_t::same_lower);
             }
         }
 

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/pad_calc.cpp

@@ -52,19 +52,21 @@ void calculate_padding(int64_t idx,
     }
 }
 
-std::vector<std::size_t> calc_dyn_auto_pad(std::vector<std::size_t> tensor_lens,
-                                           std::vector<std::size_t> k_lens,
-                                           std::vector<std::size_t> strides,
-                                           std::vector<std::size_t> dilations,
+std::vector<std::size_t> calc_dyn_auto_pad(const std::vector<std::size_t>& input_lens,
+                                           const std::vector<std::size_t>& wei_lens,
+                                           const std::vector<std::size_t>& strides,
+                                           const std::vector<std::size_t>& dilations,
                                            bool use_upper)
 {
     std::vector<std::size_t> padding;
-    padding.resize(2 * k_lens.size());
-    for(std::size_t i = 0; i < padding.size() / 2; i++)
+    assert(input_lens.size() >= 3);
+    std::size_t num_spatial_dims = input_lens.size() - 2;
+    padding.resize(2 * num_spatial_dims);
+    for(std::size_t i = 0; i < num_spatial_dims; i++)
     {
-        std::ptrdiff_t input_dim      = tensor_lens[i];
+        std::ptrdiff_t input_dim      = input_lens[i + 2];
         std::ptrdiff_t stride         = strides[i];
-        std::ptrdiff_t weight_dim     = k_lens[i];
+        std::ptrdiff_t weight_dim     = wei_lens[i + 2];
         std::ptrdiff_t dilation       = dilations[i];
         std::ptrdiff_t output_dim     = (input_dim + stride - 1) / stride; // round up result
         std::ptrdiff_t new_weight_dim = weight_dim + (weight_dim - 1) * (dilation - 1);
@@ -86,5 +88,28 @@ std::vector<std::size_t> calc_dyn_auto_pad(std::vector<std::size_t> tensor_lens,
     return padding;
 }
 
+shape compute_padded_shape(const shape& input,
+                           const shape& weights,
+                           const std::vector<std::size_t>& padding,
+                           const std::vector<std::size_t>& stride,
+                           const std::vector<std::size_t>& dilation)
+{
+    const size_t num_spatial_dims = input.lens().size() - 2;
+
+    std::vector<size_t> output_lens{input.lens()[0], weights.lens()[0]};
+    // calculate the output shape of the convolution: ((W - K + 2P) / S) + 1
+    for(size_t i = 0; i < num_spatial_dims; ++i)
+    {
+        auto padding_factor = padding[i] + padding[i + num_spatial_dims];
+        output_lens.push_back(std::size_t(std::max<std::ptrdiff_t>(
+            1,
+            (input.lens()[i + 2] - (1 + dilation[i] * (weights.lens()[i + 2] - 1)) +
+             padding_factor) /
+                    stride[i] +
+                1)));
+    }
+    return input.with_lens(output_lens);
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

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/rewrite_rnn.cpp

@@ -46,9 +46,6 @@
 #include <migraphx/iterator_for.hpp>
 #include <migraphx/dfor.hpp>
 #include <migraphx/ranges.hpp>
-#include <migraphx/op/common.hpp>
-#include <migraphx/op/rnn_var_sl_last_output.hpp>
-#include <migraphx/op/rnn_variable_seq_lens.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

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/simplify_algebra.cpp

@@ -827,7 +827,7 @@ MIGRAPHX_PRED_MATCHER(horiz_conv_dot, instruction_ref ins)
     };
     auto dots  = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("dot"));
     auto convs = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("convolution"));
-    return not(dots < 2 and convs < 2);
+    return (dots >= 2 or convs >= 2);
 }
 
 struct find_conv_dot_horiz_fusion
@@ -933,6 +933,73 @@ struct find_div_const
     }
 };
 
+struct find_unit_ops
+{
+    auto matcher() const
+    {
+        auto mul_1 = match::name("mul")(
+            match::either_arg(0, 1)(match::has_value(1.0f), match::any().bind("x")));
+        auto div_1 =
+            match::name("div")(match::args(match::any().bind("x"), match::has_value(1.0f)));
+        auto add_0 = match::name("add")(
+            match::either_arg(0, 1)(match::has_value(0.0f, 1e-12), match::any().bind("x")));
+        auto sub_0 =
+            match::name("sub")(match::args(match::any().bind("x"), match::has_value(0.0f)));
+        return match::any_of(mul_1, div_1, add_0, sub_0);
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins  = r.result;
+        auto c_in = r.instructions["x"];
+
+        m.replace_instruction(ins, c_in);
+    }
+};
+
+struct find_neg_unit_ops
+{
+    auto matcher() const
+    {
+        auto mul_neg_1 = match::name("mul")(
+            match::either_arg(0, 1)(match::has_value(-1.0f), match::any().bind("x")));
+        auto div_neg_1 =
+            match::name("div")(match::args(match::any().bind("x"), match::has_value(-1.0f)));
+        auto sub_0 =
+            match::name("sub")(match::args(match::has_value(0.0f), match::any().bind("x")));
+        return match::any_of(mul_neg_1, div_neg_1, sub_0);
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins  = r.result;
+        auto c_in = r.instructions["x"];
+
+        auto neg = m.add_instruction(make_op("neg"), c_in);
+        m.replace_instruction(ins, neg);
+    }
+};
+
+struct find_zero_ops
+{
+    auto matcher() const
+    {
+        auto mul_zero = match::name("mul")(
+            match::either_arg(0, 1)(match::has_value(0.0f).bind("x"), match::any()));
+        auto div_zero =
+            match::name("div")(match::args(match::has_value(0.0f).bind("x"), match::any()));
+        return match::any_of(mul_zero, div_zero);
+    }
+
+    void apply(module& m, const match::matcher_result& r) const
+    {
+        auto ins      = r.result;
+        auto zero_ins = r.instructions["x"];
+
+        m.replace_instruction(ins, zero_ins);
+    }
+};
+
 struct find_sub_const
 {
     auto matcher() const
@@ -1149,6 +1216,9 @@ void simplify_algebra::apply(module& m) const
                             find_mul_conv{},
                             find_mul_slice_conv{},
                             find_mul_add{},
+                            find_unit_ops{},
+                            find_neg_unit_ops{},
+                            find_zero_ops{},
                             find_dot_add{},
                             find_div_const{},
                             find_sub_const{},