Merge branch 'develop' into navi-reduce

src/onnx/parse_unique.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include <migraphx/onnx/op_parser.hpp>
+#include <migraphx/ranges.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/tune_axis.hpp>
+#include <optional>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+// generate unique output stream y, given input stream x;
+//
+// case unsorted:
+// input x: [2, 1, 1, 3, 4, 3], attr_sorted = 0;
+// output(s):
+// y: [2, 1, 3, 4]   --- the unique output
+// y_indices: [0, 1, 3, 4]  --- first incidence, in terms of indices of x
+// x_rev_indices: [0, 1, 1, 2, 3, 2] --- x seen in terms of indices of y
+// y_count: [1, 2, 2, 1] -- count at each y_index. sum = len(x)
+//
+// case sorted:
+// input x: [2, 1, 1, 3, 4, 3], attr_sorted = 1;
+// output(s):
+// y: [1, 2, 3, 4]   --- the unique output
+// y_indices: [1, 0, 3, 4]  --- first incidence, in terms of indices of x
+// x_rev_indices: [1, 0, 0, 2, 3, 2] --- x seen in terms of indices of y
+// y_count: [2, 1, 2, 1] -- count at each y_index. sum = len(x)
+
+struct parse_unique : op_parser<parse_unique>
+{
+
+    std::vector<op_desc> operators() const { return {{"Unique"}}; }
+
+    std::vector<instruction_ref> parse(const op_desc& opd,
+                                       const onnx_parser& parser,
+                                       const onnx_parser::node_info& info,
+                                       std::vector<instruction_ref> args) const
+    {
+        int64_t sorted = 1; // default = sorted.
+
+        if(contains(info.attributes, "sorted"))
+            sorted = parser.parse_value(info.attributes.at("sorted")).at<int>();
+
+        std::optional<int64_t> axis;
+        if(contains(info.attributes, "axis"))
+        {
+            auto n_dim = args[0]->get_shape().ndim();
+            axis       = parser.parse_value(info.attributes.at("axis")).at<int>();
+            axis       = tune_axis(n_dim, *axis, opd.op_name);
+        }
+        migraphx::argument data_arg = args.back()->eval();
+
+        auto opr     = axis ? migraphx::make_op("unique", {{"axis", *axis}, {"sorted", sorted}})
+                            : migraphx::make_op("unique", {{"sorted", sorted}});
+        auto u_opr   = info.add_instruction(opr, args.at(0));
+        auto i_y     = info.add_instruction(make_op("get_tuple_elem", {{"index", 0}}), u_opr);
+        auto i_y_idx = info.add_instruction(make_op("get_tuple_elem", {{"index", 1}}), u_opr);
+        auto i_x_idx = info.add_instruction(make_op("get_tuple_elem", {{"index", 2}}), u_opr);
+        auto i_count = info.add_instruction(make_op("get_tuple_elem", {{"index", 3}}), u_opr);
+
+        return {i_y, i_y_idx, i_x_idx, i_count};
+    }
+};
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/onnx/pooling.cpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+
+#include <migraphx/onnx/pooling.hpp>
+#include <migraphx/onnx/checks.hpp>
+#include <migraphx/onnx/padding.hpp>
+#include <migraphx/stringutils.hpp>
+#include <migraphx/make_op.hpp>
+#include <migraphx/op/pooling.hpp>
+#include <migraphx/op/pad.hpp>
+#include <migraphx/ranges.hpp>
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+namespace onnx {
+
+value handle_pooling_values(const op_desc& opd,
+                            onnx_parser::node_info info,
+                            const shape& in_shape,
+                            value values)
+{
+    auto kdims = in_shape.ndim() - 2;
+    if(starts_with(opd.onnx_name, "Global") or starts_with(opd.onnx_name, "QLinearGlobal"))
+    {
+        // if spatial dimensions are dynamic use dyn_global flag
+        if(in_shape.dynamic() and std::any_of(in_shape.dyn_dims().cbegin() + 2,
+                                              in_shape.dyn_dims().cend(),
+                                              [](auto dd) { return not dd.is_fixed(); }))
+        {
+            values["dyn_global"] = true;
+            values["lengths"]    = std::vector<size_t>();
+        }
+        else
+        {
+            // works with static and fixed dynamic shape
+            auto m_lens       = in_shape.max_lens();
+            values["lengths"] = std::vector<size_t>(m_lens.begin() + 2, m_lens.end());
+        }
+    }
+
+    if(contains(info.attributes, "ceil_mode"))
+    {
+        values["ceil_mode"] = static_cast<bool>(info.attributes.at("ceil_mode").i());
+    }
+
+    if(contains(info.attributes, "strides"))
+    {
+        values["stride"].clear();
+        copy(info.attributes["strides"].ints(), std::back_inserter(values["stride"]));
+        check_attr_sizes(kdims, values["stride"].size(), "PARSE_POOLING: inconsistent strides");
+    }
+
+    if(contains(info.attributes, "kernel_shape"))
+    {
+        values["lengths"].clear();
+        copy(info.attributes["kernel_shape"].ints(), std::back_inserter(values["lengths"]));
+        check_attr_sizes(kdims, values["lengths"].size(), "PARSE_POOLING: inconsistent lengths");
+    }
+
+    if(contains(info.attributes, "dilations"))
+    {
+        values["dilations"].clear();
+        copy(info.attributes["dilations"].ints(), std::back_inserter(values["dilations"]));
+        check_attr_sizes(
+            kdims, values["dilations"].size(), "PARSE_POOLING: inconsistent dilations");
+    }
+
+    // lp_order attribute
+    if(contains(info.attributes, "p"))
+    {
+        values["lp_order"] = info.attributes.at("p").i();
+    }
+
+    // ensure pads available only when auto_pad is "NOT_SET"
+    check_padding_mode(info, "POOLING");
+
+    return values;
+}
+
+instruction_ref add_pooling_op(const op_desc& opd, onnx_parser::node_info info, instruction_ref l0)
+{
+    std::string mode                                                 = opd.op_name;
+    const std::unordered_map<std::string, op::pooling_mode> mode_map = {
+        {"max", op::pooling_mode::max},
+        {"average", op::pooling_mode::average},
+        {"lpnorm", op::pooling_mode::lpnorm}};
+    if(not contains(mode_map, mode))
+    {
+        MIGRAPHX_THROW(
+            "PARSE_POOLING: onnx pooling mode must be [\"max\", \"average\", \"lpnorm\"]");
+    }
+    operation op  = make_op("pooling", {{"mode", mode_map.at(mode)}});
+    value values  = op.to_value();
+    auto in_shape = l0->get_shape();
+    assert(in_shape.ndim() > 2);
+    auto kdims = in_shape.ndim() - 2;
+
+    values = handle_pooling_values(opd, info, in_shape, values);
+
+    // count include padding, if count include pad is 1, we always use
+    // explicit pad
+    int count_include_pad = 0;
+    if(contains(info.attributes, "count_include_pad"))
+    {
+        if(in_shape.dynamic())
+        {
+            MIGRAPHX_THROW("PARSE_POOLING: count_include_pad attribute is not supported for "
+                           "dynamic input shape");
+        }
+        count_include_pad = info.attributes.at("count_include_pad").i();
+    }
+
+    std::vector<int64_t> paddings;
+    float pad_val = ((mode == "max") ? std::numeric_limits<float>::lowest() : 0.0f);
+
+    if(contains(info.attributes, "pads"))
+    {
+        values["padding"].clear();
+        copy(info.attributes["pads"].ints(), std::back_inserter(paddings));
+        check_attr_sizes(
+            kdims, paddings.size() / 2, "PARSE_POOLING: inconsistent explicit paddings");
+    }
+
+    if(paddings.size() != 2 * kdims)
+    {
+        paddings.resize(kdims * 2);
+        std::fill_n(paddings.begin(), 2 * kdims, 0);
+    }
+
+    if(values["padding"].size() != kdims)
+    {
+        values["padding"].resize(kdims);
+        std::fill_n(values["padding"].begin(), kdims, 0);
+    }
+
+    if(values["stride"].size() != kdims)
+    {
+        values["stride"].resize(kdims);
+        std::fill_n(values["stride"].begin(), kdims, 1);
+    }
+
+    if(values["dilations"].size() != kdims)
+    {
+        values["dilations"].resize(kdims);
+        std::fill_n(values["dilations"].begin(), kdims, 1);
+    }
+
+    // used to calculate the supposed output shape
+    std::vector<int64_t> orig_padding = paddings;
+
+    // TODO:  add parsing for dilations
+    if(contains(info.attributes, "auto_pad") and
+       to_upper(info.attributes["auto_pad"].s()) != "NOTSET")
+    {
+        auto auto_pad = to_upper(info.attributes["auto_pad"].s());
+        // don't use the given padding sizes, if any
+        // values["padding"].clear();
+        if(in_shape.dynamic())
+        {
+            // set padding_mode to trigger auto padding at runtime
+            bool is_same_upper     = (auto_pad.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);
+        }
+        else
+        {
+            // Calculate auto padding
+            // dilations (argument 4) not supported; default to all 1's
+            cal_auto_padding_size(info,
+                                  values,
+                                  values["lengths"].to_vector<std::size_t>(),
+                                  values["dilations"].to_vector<std::size_t>(),
+                                  in_shape.lens(),
+                                  paddings);
+            values["padding"] = paddings;
+            // default padding_mode indicates that padding sizes are not calculated dynamically
+            values["padding_mode"] = migraphx::op::padding_mode_t::default_;
+        }
+    }
+
+    std::vector<int64_t> slice_start;
+    std::vector<int64_t> slice_end;
+    tune_padding_size(values, paddings, count_include_pad, slice_start);
+
+    if(not slice_start.empty())
+    {
+        if(in_shape.dynamic())
+        {
+            MIGRAPHX_THROW(
+                "PARSE_POOLING: asymmetric padding not supported for dynamic input shape");
+        }
+        // calculate expected output shape
+        orig_padding.insert(orig_padding.begin() + kdims, 2, 0);
+        orig_padding.insert(orig_padding.begin(), 2, 0);
+        op::pad pad{orig_padding, 0.0f};
+        shape padded_shape = pad.compute_shape({l0->get_shape()});
+
+        // make an op just to get its output shape
+        auto out_lens = make_op("pooling", values).compute_shape({padded_shape}).lens();
+        // compute slice_end information
+        slice_end.resize(slice_start.size());
+        std::transform(out_lens.begin() + 2,
+                       out_lens.end(),
+                       slice_start.begin(),
+                       slice_end.begin(),
+                       [](auto i, auto j) { return i + j; });
+    }
+    values["padding"] = std::vector<size_t>(paddings.begin(), paddings.end());
+
+    check_asym_padding(info, l0, paddings, values, count_include_pad, pad_val);
+    op.from_value(values);
+
+    auto l1 = info.add_instruction(op, l0);
+    if(not slice_start.empty())
+    {
+        std::vector<int64_t> axes(kdims);
+        std::iota(axes.begin(), axes.end(), 2);
+        l1 = info.add_instruction(
+            make_op("slice", {{"axes", axes}, {"starts", slice_start}, {"ends", slice_end}}), l1);
+    }
+
+    return l1;
+}
+
+} // namespace onnx
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/py/migraphx_py.cpp

@@ -40,7 +40,7 @@
 #include <migraphx/json.hpp>
 #include <migraphx/make_op.hpp>
 #include <migraphx/op/common.hpp>
-
+#include <migraphx/float8.hpp>
 #ifdef HAVE_GPU
 #include <migraphx/gpu/hip.hpp>
 #endif
@@ -144,6 +144,18 @@ struct npy_format_descriptor<half>
     static constexpr auto name() { return _("half"); }
 };
 
+template <>
+struct npy_format_descriptor<migraphx::fp8::fp8e4m3fnuz>
+{
+    static std::string format()
+    {
+        // following: https://docs.python.org/3/library/struct.html#format-characters
+        // TODO: need to figure out correct encoding
+        return "z";
+    }
+    static constexpr auto name() { return _("fp8e4m3fnuz"); }
+};
+
 } // namespace detail
 } // namespace pybind11
 

src/register_target.cpp

@@ -56,7 +56,11 @@ target make_target(const std::string& name)
 {
     if(not contains(target_map(), name))
     {
+#ifdef _WIN32
+        std::string target_name = "migraphx_" + name + ".dll";
+#else
         std::string target_name = "libmigraphx_" + name + ".so";
+#endif
         store_target_lib(dynamic_loader(target_name));
     }
     const auto it = target_map().find(name);

src/rewrite_pooling.cpp

@@ -35,6 +35,110 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
+static void replace_with_reduce(module& m, instruction_ref ins)
+{
+    auto&& s  = ins->inputs().front()->get_shape();
+    auto&& op = any_cast<op::pooling>(ins->get_operator());
+    auto lens = s.lens();
+    std::vector<std::int64_t> axes(lens.size() - 2);
+    std::iota(axes.begin(), axes.end(), 2);
+
+    // average pooling
+    if(op.mode == op::pooling_mode::average)
+    {
+        m.replace_instruction(ins, make_op("reduce_mean", {{"axes", axes}}), ins->inputs());
+    }
+    // max pooling
+    else
+    {
+        m.replace_instruction(ins, make_op("reduce_max", {{"axes", axes}}), ins->inputs());
+    }
+}
+
+static void replace_dilations_with_gather_pooling(module& m, instruction_ref ins)
+{
+    // TODO remove this when MIOpen supports dilated pooling
+    auto&& s  = ins->inputs().front()->get_shape();
+    auto&& op = any_cast<op::pooling>(ins->get_operator());
+    // Ignore N, C axes
+    std::vector<size_t> dims = {s.lens().cbegin() + 2, s.lens().cend()};
+
+    bool default_padding =
+        std::all_of(op.padding.cbegin(), op.padding.cend(), [](auto i) { return i == 0; });
+
+    if(not default_padding)
+    {
+        for(size_t idx{0}; idx < op.padding.size(); ++idx)
+        {
+            // We need to pad both ends
+            dims[idx] += op.padding.at(idx) * 2;
+        }
+    }
+    std::vector<size_t> kernels   = op.lengths;
+    std::vector<size_t> strides   = op.stride;
+    std::vector<size_t> dilations = op.dilations;
+
+    std::vector<std::vector<int>> axis_indices;
+    axis_indices.resize(dims.size());
+
+    for(auto idx{0}; idx < dims.size(); ++idx)
+    {
+        // Only consider if iw fits into the window
+        for(size_t stride{0}; stride < dims.at(idx) - dilations.at(idx) * (kernels.at(idx) - 1);
+            stride += strides.at(idx))
+        {
+            for(size_t step{0}; step < kernels.at(idx); ++step)
+            {
+                axis_indices.at(idx).push_back(stride + dilations.at(idx) * step);
+            }
+        }
+    }
+
+    auto elements = ins->inputs().front();
+    if(not default_padding)
+    {
+        // Pad supports asym, we need to provide both ends
+        std::vector<size_t> padding(2 * s.lens().size(), 0);
+        // Format will be e.g {N, C, P1, P2, N, C, P1, P2}
+        for(size_t idx{0}; idx < op.padding.size(); ++idx)
+        {
+            // Ignore N, C axes
+            padding.at(2 + idx)                   = op.padding.at(idx);
+            padding.at(2 + idx + s.lens().size()) = op.padding.at(idx);
+        }
+
+        // Default value needed for Max pooling
+        elements = m.insert_instruction(
+            ins,
+            make_op("pad", {{"pads", padding}, {"value", std::numeric_limits<float>::lowest()}}),
+            elements);
+    }
+
+    for(auto idx{0}; idx < axis_indices.size(); ++idx)
+    {
+        migraphx::shape s_indices{migraphx::shape::int32_type, {axis_indices.at(idx).size()}};
+        auto indices = m.add_literal(migraphx::literal{s_indices, axis_indices.at(idx)});
+        elements     = m.insert_instruction(
+            ins, make_op("gather", {{"axis", idx + 2 /*ignore N,C*/}}), elements, indices);
+    }
+
+    // Ignore padding
+    std::vector<size_t> new_padding(kernels.size(), 0);
+    // The kernel window elements are places next to each other. E.g. {x1, y1, x2, y2, ...}
+    // We need to skip them to not overlap
+    std::vector<size_t> new_strides(kernels);
+    // Ignore dilations
+    std::vector<size_t> new_dilations(kernels.size(), 1);
+    m.replace_instruction(ins,
+                          make_op("pooling",
+                                  {{"mode", op.mode},
+                                   {"padding", new_padding},
+                                   {"stride", new_strides},
+                                   {"lengths", kernels},
+                                   {"dilations", new_dilations}}),
+                          elements);
+}
+
 void rewrite_pooling::apply(module& m) const
 {
     for(auto ins : iterator_for(m))
@@ -43,26 +147,36 @@ void rewrite_pooling::apply(module& m) const
             continue;
         if(ins->inputs().empty())
             continue;
-        auto&& s  = ins->inputs().front()->get_shape();
-        auto&& op = any_cast<op::pooling>(ins->get_operator());
-        if(not std::all_of(op.padding.begin(), op.padding.end(), [](auto i) { return i == 0; }))
-            continue;
-        if(not std::all_of(op.stride.begin(), op.stride.end(), [](auto i) { return i == 1; }))
-            continue;
-        auto lens = s.lens();
-        if(not std::equal(lens.begin() + 2, lens.end(), op.lengths.begin(), op.lengths.end()))
-            continue;
-        std::vector<std::int64_t> axes(lens.size() - 2);
-        std::iota(axes.begin(), axes.end(), 2);
-        // average pooling
-        if(op.mode == op::pooling_mode::average)
+        auto&& s                  = ins->inputs().front()->get_shape();
+        auto&& op                 = any_cast<op::pooling>(ins->get_operator());
+        bool same_kernel_as_shape = std::equal(
+            s.lens().cbegin() + 2, s.lens().cend(), op.lengths.cbegin(), op.lengths.cend());
+        bool default_strides =
+            std::all_of(op.stride.cbegin(), op.stride.cend(), [](auto i) { return i == 1; });
+        bool default_padding =
+            std::all_of(op.padding.cbegin(), op.padding.cend(), [](auto i) { return i == 0; });
+        bool default_dilations =
+            std::all_of(op.dilations.cbegin(), op.dilations.cend(), [](auto i) { return i == 1; });
+        if(same_kernel_as_shape and default_strides and default_padding and default_dilations)
         {
-            m.replace_instruction(ins, make_op("reduce_mean", {{"axes", axes}}), ins->inputs());
+            replace_with_reduce(m, ins);
         }
-        // max pooling
-        else
+        else if(not default_dilations)
         {
-            m.replace_instruction(ins, make_op("reduce_max", {{"axes", axes}}), ins->inputs());
+            // Dilated AvgPool with padding is not supported
+            if(not default_padding and op.mode == op::pooling_mode::average)
+            {
+                continue;
+            }
+            auto size =
+                std::accumulate(s.lens().cbegin(), s.lens().cend(), 1, std::multiplies<size_t>());
+            // Can't handle too much size because of literal size
+            if(size > 100000)
+            {
+                continue;
+            }
+
+            replace_dilations_with_gather_pooling(m, ins);
         }
     }
 }

src/rewrite_quantization.cpp

@@ -47,7 +47,7 @@ void apply_quantizelinear(module& m, instruction_ref ins)
             ins, make_op("convert", {{"target_type", y_scale->get_shape().type()}}), x);
     }
     auto div            = m.insert_instruction(ins, make_op("div"), x, y_scale);
-    auto add_zero_point = m.insert_instruction(ins, make_op("round"), div);
+    auto add_zero_point = m.insert_instruction(ins, make_op("nearbyint"), div);
 
     if(ins->inputs().size() == 3)
     {

src/schedule.cpp

@@ -27,7 +27,7 @@
 #include <migraphx/iterator_for.hpp>
 #include <migraphx/iterator.hpp>
 #include <migraphx/dfor.hpp>
-#include <migraphx/par_for.hpp>
+#include <migraphx/simple_par_for.hpp>
 #include <migraphx/functional.hpp>
 #include <migraphx/ranges.hpp>
 #include <migraphx/dom_info.hpp>
@@ -461,7 +461,7 @@ struct stream_info
                        std::back_inserter(index_to_ins),
                        [](auto&& it) { return it.first; });
 
-        par_for(concur_ins.size(), [&](auto ins_index, auto tid) {
+        simple_par_for(concur_ins.size(), [&](auto ins_index, auto tid) {
             auto merge_first = index_to_ins[ins_index];
             assert(concur_ins.count(merge_first) > 0);
             auto& merge_second = concur_ins.at(merge_first);

src/simplify_algebra.cpp

@@ -941,15 +941,6 @@ struct find_splits
                 {
                     auto split = i->inputs()[split_idx];
                     assert(split->name() == "slice");
-                    // Insert contiguous for reshapes
-                    auto outputs = i->outputs();
-                    for(auto output : outputs)
-                    {
-                        if(output->name() != "reshape")
-                            continue;
-                        auto x = m.insert_instruction(output, make_op("contiguous"), i);
-                        m.replace_instruction(output, output->get_operator(), x);
-                    }
 
                     m.replace_instruction(i, split->get_operator(), c);
                 }
@@ -1181,13 +1172,6 @@ struct find_conv_dot_horiz_fusion
             for(auto arg : range(start, last))
             {
                 auto outputs = arg->outputs();
-                for(auto output : outputs)
-                {
-                    if(output->name() != "reshape")
-                        continue;
-                    auto x = m.insert_instruction(output, make_op("contiguous"), arg);
-                    m.replace_instruction(output, output->get_operator(), x);
-                }
 
                 int64_t len = arg->get_shape().lens()[axis];
                 m.replace_instruction(
@@ -1487,11 +1471,6 @@ struct find_split_reshape
                    slc_axis_len;
         });
 
-        // insert the reshape instruction and add contiguous if needed
-        if(not input->get_shape().standard())
-        {
-            input = m.insert_instruction(std::next(input), make_op("contiguous"), input);
-        }
         auto rsp_ins = m.insert_instruction(
             std::next(input), make_op("reshape", {{"dims", rsp_out_lens}}), input);
 

src/simplify_dyn_ops.cpp

@@ -22,8 +22,10 @@
  * THE SOFTWARE.
  */
 #include <migraphx/simplify_dyn_ops.hpp>
+#include <migraphx/op/slice.hpp>
 #include <migraphx/matcher.hpp>
 #include <migraphx/make_op.hpp>
+#include <migraphx/literal.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -32,6 +34,10 @@ inline namespace MIGRAPHX_INLINE_NS {
  * Convert 2 input static shape broadcast/multibroadcast into 1 input version.
  * Some compiler passes (ex. simplify_algebra) only support the 1 input versions
  * of the broadcasting operators.
+ * From:
+ * broadcast_op(argument_with_static_shape, argument_with_static_shape)
+ * To:
+ * broadcast_op(argument_with_static_shape); broadcast_op.out_lens = constant_output_dims
  */
 struct find_static_2in_broadcasts
 {
@@ -60,8 +66,65 @@ struct find_static_2in_broadcasts
 };
 
 /**
- * Simplify slice with variable `starts` and `ends` to the constant version if
- * the `input_starts` and `input_ends` inputs are constant.
+ * Simplify slice with 2 inputs to the 1 input version if inputs[1] is constant.
+ * From:
+ * slice(data, constant_input); two attributes set
+ * To:
+ * slice(data); slice.starts, slice.ends. slice.axes set
+ */
+struct find_const_2in_slice
+{
+    auto matcher() const
+    {
+        return match::name("slice")(match::nargs(2), match::arg(1)(match::is_constant()));
+    }
+
+    void apply(module& m, const match::matcher_result& mr) const
+    {
+        auto ins       = mr.result;
+        auto inputs    = ins->inputs();
+        auto slice_op  = any_cast<op::slice>(ins->get_operator());
+        auto set_attrs = slice_op.get_set_attributes();
+        std::vector<int64_t> starts_vec;
+        std::vector<int64_t> ends_vec;
+        std::vector<int64_t> axes_vec;
+        if(set_attrs == op::slice::ends_axes)
+        {
+            // slice(data, starts)
+            inputs.at(1)->eval().visit(
+                [&](auto output) { starts_vec.assign(output.begin(), output.end()); });
+            ends_vec = slice_op.ends;
+            axes_vec = slice_op.axes;
+        }
+        else if(set_attrs == op::slice::starts_axes)
+        {
+            // slice(data, ends)
+            inputs.at(1)->eval().visit(
+                [&](auto output) { ends_vec.assign(output.begin(), output.end()); });
+            starts_vec = slice_op.starts;
+            axes_vec   = slice_op.axes;
+        }
+        else
+        {
+            // slice(data, axes)
+            inputs.at(1)->eval().visit(
+                [&](auto output) { axes_vec.assign(output.begin(), output.end()); });
+            starts_vec = slice_op.starts;
+            ends_vec   = slice_op.ends;
+        }
+        m.replace_instruction(
+            ins,
+            make_op("slice", {{"starts", starts_vec}, {"ends", ends_vec}, {"axes", axes_vec}}),
+            inputs.at(0));
+    }
+};
+
+/**
+ * Simplify slice with 3 inputs to the 1 input version if inputs[1:2] are constant.
+ * From:
+ * slice(data, constant_input1, constant_input2); one attribute set
+ * To:
+ * slice(data); slice.starts, slice.ends. slice.axes set
  */
 struct find_const_3in_slice
 {
@@ -76,27 +139,51 @@ struct find_const_3in_slice
     {
         auto ins            = mr.result;
         auto inputs         = ins->inputs();
-        argument starts_arg = inputs.at(1)->eval();
-        argument ends_arg   = inputs.at(2)->eval();
-        if(not starts_arg.empty() and not ends_arg.empty())
+        auto slice_op       = any_cast<op::slice>(ins->get_operator());
+        auto set_attrs      = slice_op.get_set_attributes();
+        std::vector<int64_t> starts_vec;
+        std::vector<int64_t> ends_vec;
+        std::vector<int64_t> axes_vec;
+        if(set_attrs == op::slice::axes_only)
         {
-            std::vector<int64_t> starts_vec;
-            std::vector<int64_t> ends_vec;
-            starts_arg.visit([&](auto output) { starts_vec.assign(output.begin(), output.end()); });
-            ends_arg.visit([&](auto output) { ends_vec.assign(output.begin(), output.end()); });
-            auto slice_val = ins->get_operator().to_value();
-            auto axes_vec  = slice_val.at("axes").to_vector<int64_t>();
-            m.replace_instruction(
-                ins,
-                make_op("slice", {{"starts", starts_vec}, {"ends", ends_vec}, {"axes", axes_vec}}),
-                inputs.at(0));
+            // slice(data, starts, ends)
+            inputs.at(1)->eval().visit(
+                [&](auto output) { starts_vec.assign(output.begin(), output.end()); });
+            inputs.at(2)->eval().visit(
+                [&](auto output) { ends_vec.assign(output.begin(), output.end()); });
+            axes_vec = slice_op.axes;
+        }
+        else if(set_attrs == op::slice::ends_only)
+        {
+            // slice(data, starts, axes)
+            inputs.at(1)->eval().visit(
+                [&](auto output) { starts_vec.assign(output.begin(), output.end()); });
+            inputs.at(2)->eval().visit(
+                [&](auto output) { axes_vec.assign(output.begin(), output.end()); });
+            ends_vec = slice_op.ends;
+        }
+        else
+        {
+            // slice(data, ends, axes)
+            inputs.at(1)->eval().visit(
+                [&](auto output) { ends_vec.assign(output.begin(), output.end()); });
+            inputs.at(2)->eval().visit(
+                [&](auto output) { axes_vec.assign(output.begin(), output.end()); });
+            starts_vec = slice_op.starts;
         }
+        m.replace_instruction(
+            ins,
+            make_op("slice", {{"starts", starts_vec}, {"ends", ends_vec}, {"axes", axes_vec}}),
+            inputs.at(0));
     }
 };
 
 /**
- * Simplify slice with variable `starts`, `ends`, and `input_axes` to the constant version if
- * the `input_starts`, `input_ends`, and `input_axes` inputs are constant.
+ * Simplify slice with 4 inputs to the 1 input version if inputs[1:3] are constant.
+ * From:
+ * slice(data, constant_starts, constant_ends, constant_axes)
+ * To:
+ * slice(data); slice.starts, slice.ends. slice.axes set
  */
 struct find_const_4in_slice
 {
@@ -112,9 +199,9 @@ struct find_const_4in_slice
     {
         auto ins            = mr.result;
         auto inputs         = ins->inputs();
-        argument starts_arg = inputs.at(1)->eval();
-        argument ends_arg   = inputs.at(2)->eval();
-        argument axes_arg   = inputs.at(3)->eval();
+        argument starts_arg = inputs.at(1)->eval(false);
+        argument ends_arg   = inputs.at(2)->eval(false);
+        argument axes_arg   = inputs.at(3)->eval(false);
         if(not starts_arg.empty() and not ends_arg.empty() and not axes_arg.empty())
         {
             std::vector<int64_t> starts_vec;
@@ -131,10 +218,116 @@ struct find_const_4in_slice
     }
 };
 
+/**
+ * Simplify dimensions_of to a literal when the input arugment has a static shape
+ * or the dynamic dimensions from `start` to `end` are fixed.
+ */
+struct find_static_dimensions_of
+{
+    auto matcher() const { return match::name("dimensions_of")(); }
+
+    void apply(module& m, const match::matcher_result& mr) const
+    {
+        auto ins                 = mr.result;
+        auto input               = ins->inputs().at(0);
+        auto dimensions_of_value = ins->get_operator().to_value();
+        auto start               = dimensions_of_value.at("start").to<std::size_t>();
+        auto end                 = dimensions_of_value.at("end").to<std::size_t>();
+        if(input->get_shape().dynamic())
+        {
+            // check if dynamic dimensions from start to end are fixed
+            auto dds = input->get_shape().dyn_dims();
+            if(std::any_of(dds.begin() + start, dds.begin() + end, [](auto dd) {
+                   return not dd.is_fixed();
+               }))
+            {
+                return;
+            }
+        }
+        std::size_t output_ndim = end - start;
+        std::vector<int64_t> vec_shape(output_ndim);
+        migraphx::shape s(migraphx::shape::int64_type, {output_ndim});
+        std::vector<std::size_t> input_lens = input->get_shape().to_static(1).lens();
+        std::transform(input_lens.begin() + start,
+                       input_lens.begin() + end,
+                       vec_shape.begin(),
+                       [](auto i) { return int64_t(i); });
+        migraphx::shape output_shape{migraphx::shape::int64_type, {end - start}};
+        auto lit_ins = m.add_literal(migraphx::literal{output_shape, vec_shape});
+        m.replace_instruction(ins, lit_ins);
+    }
+};
+
+/**
+ * Simplify allocate into 2 argument reshape that has constant output dimensions into a static 1
+ * argument reshape. Intended to simplify what ONNX parse_reshape creates for dynamic reshapes.
+ * This matcher can be generalized to matching reshape(data, static_shape_output_tensor).
+ * From:
+ * x = allocate(constant_output_dims) -> reshape(data, x)
+ * To:
+ * reshape(data); reshape.dims = constant_output_dims
+ */
+struct find_const_alloc_reshapes
+{
+    auto matcher() const
+    {
+        return match::name("reshape")(match::nargs(2),
+                                      match::arg(1)(match::name("allocate")(match::is_constant())));
+    }
+
+    void apply(module& m, const match::matcher_result& mr) const
+    {
+        auto reshape_ins         = mr.result;
+        auto reshape_inputs      = reshape_ins->inputs();
+        auto alloc_ins           = reshape_inputs.at(1);
+        argument output_dims_arg = alloc_ins->inputs().at(0)->eval(false);
+        std::vector<int64_t> output_dims_vec;
+        output_dims_arg.visit(
+            [&](auto output) { output_dims_vec.assign(output.begin(), output.end()); });
+        m.replace_instruction(
+            reshape_ins, make_op("reshape", {{"dims", output_dims_vec}}), reshape_inputs.at(0));
+        // have dead_code_elimination remove the previous allocate
+    }
+};
+
+/**
+ * Simplify allocate into fill operator that has constant output dimensions and constant value.
+ * The allocate into fill instructions is what is produced when parsing the ONNX
+ * ConstantOfShape operator. This replacement could be handled with propagate_constant, but
+ * would rather have the simplification happen earlier during compiling.
+ * This matcher can be generalized to matching fill(constant_value, static_shape_output_tensor).
+ * From:
+ * x = allocate(constant_ouptut_dims) -> fill(constant_value, x)
+ * To:
+ * literal
+ */
+struct find_const_alloc_fill
+{
+    auto matcher() const
+    {
+        return match::name("fill")(match::arg(0)(match::is_constant()),
+                                   match::arg(1)(match::name("allocate")(match::is_constant())));
+    }
+
+    void apply(module& m, const match::matcher_result& mr) const
+    {
+        auto fill_ins = mr.result;
+        auto fill_arg = fill_ins->eval(false);
+        auto l        = m.add_literal(fill_arg.get_shape(), fill_arg.data());
+        m.replace_instruction(fill_ins, l);
+    }
+};
+
 void simplify_dyn_ops::apply(module& m) const
 {
-    match::find_matches(
-        m, find_static_2in_broadcasts{}, find_const_3in_slice{}, find_const_4in_slice{});
+    match::find_matches(m,
+                        find_static_dimensions_of{},
+                        find_const_alloc_reshapes{},
+                        find_static_2in_broadcasts{},
+                        find_const_2in_slice{},
+                        find_const_3in_slice{},
+                        find_const_4in_slice{},
+                        find_const_alloc_fill{});
 }
 
 } // namespace MIGRAPHX_INLINE_NS

src/simplify_qdq.cpp

@@ -45,77 +45,145 @@ std::unordered_set<std::string> get_quantizable_op_names()
     return s;
 }
 
-MIGRAPHX_PRED_MATCHER(has_same_value, instruction_ref ins)
+struct match_find_quantizable_ops
 {
-    if(ins->name() != "@literal")
-        return false;
-    bool all_same = false;
-    ins->get_literal().visit([&](auto s) {
-        all_same = std::all_of(s.begin() + 1, s.end(), [&](const auto& scale) {
-            return float_equal(scale, s.front());
+    static bool
+    is_valid_scale(instruction_ref scale, std::vector<std::size_t> lens, std::size_t axis)
+    {
+        return scale->get_shape().scalar() or scale->get_shape().elements() == lens.at(axis);
+    }
+
+    static bool is_valid_zero_point(instruction_ref zp)
+    {
+        if(not zp->can_eval())
+            return false;
+
+        bool all_zeros = false;
+        zp->eval().visit([&](auto z) {
+            all_zeros =
+                std::all_of(z.begin(), z.end(), [&](auto val) { return float_equal(val, 0); });
         });
-    });
-    return all_same;
-}
+        return all_zeros;
+    }
 
-struct match_find_quantizable_ops
-{
+    static auto
+    scale_broadcast_op(instruction_ref scale, std::vector<std::size_t> lens, std::size_t axis)
+    {
+        if(scale->get_shape().scalar())
+        {
+            return migraphx::make_op("multibroadcast", {{"out_lens", lens}});
+        }
+        else
+        {
+            return migraphx::make_op("broadcast", {{"out_lens", lens}, {"axis", axis}});
+        }
+    }
 
-    static auto dequantizelinear_op(const std::string& name, const std::string& scale)
+    // Helper function to insert quantized versions of any broadcasts and transpose ops that
+    // occur between dequantizelinear and the quantized op
+    static auto
+    propagate_quantized_ins(module& m, const instruction_ref dqins, const instruction_ref qop)
+    {
+        auto qinp     = dqins->inputs().front();
+        auto next_ins = dqins;
+
+        while(next_ins != qop)
+        {
+            if(next_ins->name() != "dequantizelinear")
+            {
+                qinp = m.insert_instruction(qop, next_ins->get_operator(), qinp);
+            }
+            next_ins = next_ins->outputs().front();
+        }
+        return qinp;
+    }
+
+    static auto dequantizelinear_op(const std::string& scale, const std::string& zp)
     {
         return match::name("dequantizelinear")(
-            match::arg(0)(match::skip(match::name("quantizelinear"))(match::any().bind(name))),
-            match::arg(1)(match::skip_broadcasts(has_same_value().bind(scale))),
-            match::arg(2)(match::skip_broadcasts(match::all_of(match::has_value(0)))));
+            match::arg(0)(match::skip(match::name("quantizelinear"))(match::any())),
+            match::arg(1)(match::skip_broadcasts(match::is_constant().bind(scale))),
+            match::arg(2)(match::skip_broadcasts(match::is_constant().bind(zp))));
     }
 
     auto matcher() const
     {
         return match::name(get_quantizable_op_names())(
-            match::arg(0)(dequantizelinear_op("x1", "scale1")),
-            match::arg(1)(dequantizelinear_op("x2", "scale2")));
+            match::arg(0)(match::skip_broadcasts_transposes_contiguous(
+                dequantizelinear_op("scale1", "zp1").bind("dq1"))),
+            match::arg(1)(match::skip_broadcasts_transposes_contiguous(
+                dequantizelinear_op("scale2", "zp2").bind("dq2"))));
     }
 
     void apply(module& m, const match::matcher_result& r) const
     {
         auto qop    = r.result;
-        auto q1     = r.instructions["x1"];
-        auto q2     = r.instructions["x2"];
+        auto dq1    = r.instructions["dq1"];
+        auto dq2    = r.instructions["dq2"];
         auto scale1 = r.instructions["scale1"];
         auto scale2 = r.instructions["scale2"];
+        auto zp1    = r.instructions["zp1"];
+        auto zp2    = r.instructions["zp2"];
 
         // Only INT8 type currently supported
-        if(q1->get_shape().type() != migraphx::shape::int8_type or
-           q2->get_shape().type() != migraphx::shape::int8_type)
+        if(dq1->inputs().front()->get_shape().type() != migraphx::shape::int8_type or
+           dq2->inputs().front()->get_shape().type() != migraphx::shape::int8_type)
             return;
 
-        double scale;
-        visit_all(scale1->get_literal(), scale2->get_literal())(
-            [&](const auto s1, const auto s2) { scale = s1.front() * s2.front(); });
+        // Only symmetric quantization supported (ie. non-zero zero_points not allowed)
+        if(not(is_valid_zero_point(zp1) and is_valid_zero_point(zp2)))
+            return;
 
+        // Only support scalar and 1D scales
+        if(scale1->get_shape().lens().size() != 1 or scale2->get_shape().lens().size() != 1)
+            return;
+
+        // Propagate q1 and q2 through any broadcasts and transposes before qop
         auto qop_args  = qop->inputs();
-        qop_args.at(0) = q1;
-        qop_args.at(1) = q2;
+        qop_args.at(0) = propagate_quantized_ins(m, dq1, qop);
+        qop_args.at(1) = propagate_quantized_ins(m, dq2, qop);
         instruction_ref dq;
-        instruction_ref dq_scale;
+        instruction_ref out_scale;
         instruction_ref zero_point;
         if(qop->name() == "convolution")
         {
             auto conv_val = qop->get_operator().to_value();
             dq            = m.insert_instruction(
                 qop, migraphx::make_op("quant_convolution", conv_val), qop_args);
+            auto out_lens = dq->get_shape().lens();
+
+            // Input scale should always be scalar and weight scale can be scalar or 1D of the
+            // same lens as the output channel dim (dim 1 in the output)
+            if(not(is_valid_scale(scale1, out_lens, 1) and is_valid_scale(scale2, out_lens, 1)))
+                return;
+
+            auto s1_bcast =
+                m.insert_instruction(qop, scale_broadcast_op(scale1, out_lens, 1), scale1);
+            auto s2_bcast =
+                m.insert_instruction(qop, scale_broadcast_op(scale2, out_lens, 1), scale2);
+
+            out_scale = m.insert_instruction(qop, migraphx::make_op("mul"), s1_bcast, s2_bcast);
         }
         else if(qop->name() == "dot")
         {
-            dq = m.insert_instruction(qop, migraphx::make_op("quant_dot"), qop_args);
+            dq            = m.insert_instruction(qop, migraphx::make_op("quant_dot"), qop_args);
+            auto out_lens = dq->get_shape().lens();
+
+            // For (..., M, N) x (..., N, K) dot, only support cases where quantization axis is M
+            // for input1 and K for input 2
+            if(not(is_valid_scale(scale1, out_lens, out_lens.size() - 2) and
+                   is_valid_scale(scale2, out_lens, out_lens.size() - 1)))
+                return;
+
+            auto s1_bcast = m.insert_instruction(
+                qop, scale_broadcast_op(scale1, out_lens, out_lens.size() - 2), scale1);
+            auto s2_bcast = m.insert_instruction(
+                qop, scale_broadcast_op(scale2, out_lens, out_lens.size() - 1), scale2);
+
+            out_scale = m.insert_instruction(qop, migraphx::make_op("mul"), s1_bcast, s2_bcast);
         }
-        auto ins_type = qop->get_shape().type();
-        dq_scale      = m.add_literal(literal({ins_type}, {scale}));
 
-        auto lens = dq->get_shape().lens();
-        auto scale_mb =
-            m.insert_instruction(qop, make_op("multibroadcast", {{"out_lens", lens}}), dq_scale);
-        dq = m.insert_instruction(qop, make_op("dequantizelinear"), dq, scale_mb);
+        dq = m.insert_instruction(qop, make_op("dequantizelinear"), dq, out_scale);
         m.replace_instruction(qop, dq);
     }
 };

src/simplify_reshapes.cpp

@@ -103,8 +103,6 @@ struct find_reshaper
         auto input = mr.instructions["x"];
         auto dims  = ins->get_shape().lens();
 
-        if(not input->get_shape().standard())
-            input = m.insert_instruction(ins, make_op("contiguous"), input);
         m.replace_instruction(ins, make_op("reshape", {{"dims", dims}}), input);
     }
 };
@@ -475,9 +473,8 @@ struct find_resize
             ins_rsp, migraphx::make_op("reshape", {{"dims", in_dims}}), in_rsp);
         auto mb_rsp = m.insert_instruction(
             ins_rsp, migraphx::make_op("multibroadcast", {{"out_lens", out_dims}}), rsp_data);
-        auto std_mb = m.insert_instruction(ins, migraphx::make_op("contiguous"), mb_rsp);
         std::vector<int64_t> rsp_dims(out_lens.begin(), out_lens.end());
-        m.replace_instruction(ins, migraphx::make_op("reshape", {{"dims", rsp_dims}}), std_mb);
+        m.replace_instruction(ins, migraphx::make_op("reshape", {{"dims", rsp_dims}}), mb_rsp);
     }
 };
 
@@ -626,9 +623,8 @@ struct find_transpose_contiguous_reshaper_unary
         auto cont_ins      = r.instructions["cont_ins"];
         auto unary_op_name = ins->get_operator().name();
         auto unary_ins     = m.insert_instruction(cont_ins, make_op(unary_op_name), trans_ins);
-        auto new_cont_ins  = m.insert_instruction(cont_ins, make_op("contiguous"), unary_ins);
         // older cont and reshape are removed by deadcode elimination
-        m.replace_instruction(ins, reshaper_ins->get_operator(), new_cont_ins);
+        m.replace_instruction(ins, reshaper_ins->get_operator(), unary_ins);
     }
 };
 
@@ -647,8 +643,8 @@ struct find_broadcast_transpose
     {
         auto transpose      = r.result;
         auto transpose_lens = transpose->get_shape().lens();
-        auto bcast_ins = r.instructions["bcast_ins"];
-        auto input     = bcast_ins->inputs().front();
+        auto bcast_ins      = r.instructions["bcast_ins"];
+        auto input          = bcast_ins->inputs().front();
         // scalar transformation does not need extra transpose
         if(not input->get_shape().scalar())
         {

src/targets/cpu/CMakeLists.txt

@@ -74,21 +74,27 @@ if(MIGRAPHX_ENABLE_ZENDNN)
     target_link_libraries(migraphx_cpu PRIVATE ${BLIS_LIB})
     target_link_libraries(migraphx_cpu PRIVATE ${ZENDNN_LIB})
 else()
-    target_link_libraries(migraphx_cpu PRIVATE DNNL::dnnl)
+    target_link_libraries(migraphx_cpu PUBLIC DNNL::dnnl)
 endif()
 target_link_libraries(migraphx_cpu PRIVATE migraphx)
 
 migraphx_generate_export_header(migraphx_cpu)
 
 find_package(OpenMP)
-target_link_libraries(migraphx_cpu PUBLIC OpenMP::OpenMP_CXX)
-# Add library path to rpath to workaround issues with our broken packages
-foreach(LIBRARY ${OpenMP_CXX_LIBRARIES})
-    if(LIBRARY MATCHES "libomp")
-        get_filename_component(LIBRARY_PATH "${LIBRARY}" PATH)
-        target_link_libraries(migraphx_cpu PUBLIC -Wl,-rpath=${LIBRARY_PATH} -Wl,-rpath-link=${LIBRARY_PATH})
-    endif()
-endforeach()
+if(WIN32)
+    target_link_libraries(migraphx_cpu PUBLIC libomp)
+    target_include_directories(migraphx_cpu PUBLIC ${OpenMP_CXX_INCLUDE_DIRS})
+    target_compile_options(migraphx_cpu PUBLIC ${OpenMP_CXX_FLAGS})
+else()
+    target_link_libraries(migraphx_cpu PUBLIC OpenMP::OpenMP_CXX)
+    # Add library path to rpath to workaround issues with our broken packages
+    foreach(LIBRARY ${OpenMP_CXX_LIBRARIES})
+        if(LIBRARY MATCHES "libomp")
+            get_filename_component(LIBRARY_PATH "${LIBRARY}" PATH)
+            target_link_libraries(migraphx_cpu PUBLIC -Wl,-rpath=${LIBRARY_PATH} -Wl,-rpath-link=${LIBRARY_PATH})
+        endif()
+    endforeach()
+endif()
 
 rocm_install_targets(
   TARGETS migraphx_cpu

src/targets/cpu/dnnl.cpp

@@ -68,6 +68,7 @@ dnnl::memory::data_type to_dnnl_memory_data_type(shape::type_t t)
     case st::int32_type: return dt::s32;
     case st::int8_type: return dt::s8;
     case st::uint8_type: return dt::u8;
+    case st::fp8e4m3fnuz_type: MIGRAPHX_THROW("fp8e4m3fnuz unsupported in DNNL");
     default: MIGRAPHX_THROW("Unsupported data type");
     }
 }

src/targets/cpu/lowering.cpp

@@ -340,7 +340,6 @@ struct cpu_apply
                               {"reduce_min", "reduction_min"},
                               {"reduce_sum", "reduction_sum"},
                           });
-
         extend_op("concat", "dnnl::concat");
         extend_op("contiguous", "dnnl::reorder");
         extend_op("convolution", "dnnl::convolution");
@@ -376,6 +375,12 @@ struct cpu_apply
         // Apply these operators first so the inputs can be const folded
         for(auto it : iterator_for(*modl))
         {
+            // skip lowering if input has fp8 as one of the inputs since oneDNN doesn't have fp8
+            // supported yet.
+            if(std::any_of(it->inputs().begin(), it->inputs().end(), [](const auto& i) {
+                   return i->get_shape().type() == migraphx::shape::fp8e4m3fnuz_type;
+               }))
+                continue;
             if(it->name() == "pow")
             {
                 apply_pow(it);
@@ -383,6 +388,12 @@ struct cpu_apply
         }
         for(auto it : iterator_for(*modl))
         {
+            // skip lowering if input has fp8 as one of the inputs since oneDNN doesn't have fp8
+            // supported yet.
+            if(std::any_of(it->inputs().begin(), it->inputs().end(), [](const auto& i) {
+                   return i->get_shape().type() == migraphx::shape::fp8e4m3fnuz_type;
+               }))
+                continue;
             if(it->name() == "pooling")
             {
                 apply_pooling(it);

src/targets/cpu/pooling.cpp

@@ -34,23 +34,32 @@ namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace cpu {
 
-struct dnnl_pooling : dnnl_extend_op<dnnl_pooling, dnnl::pooling_forward, op::pooling>
+struct dnnl_pooling : dnnl_extend_op<dnnl_pooling, dnnl::pooling_v2_forward, op::pooling>
 {
     std::vector<int> arg_map(int) const { return {MIGRAPHX_DNNL_PREFIX(ARG_SRC)}; }
 
-    dnnl::pooling_forward::desc get_desc(const std::unordered_map<int, dnnl::memory::desc>& m) const
+    dnnl::pooling_v2_forward::desc
+    get_desc(const std::unordered_map<int, dnnl::memory::desc>& m) const
     {
-        auto algo = op.mode == op::pooling_mode::max ? dnnl::algorithm::pooling_max
-                                                     : dnnl::algorithm::pooling_avg;
+        auto algo  = op.mode == op::pooling_mode::max ? dnnl::algorithm::pooling_max
+                                                      : dnnl::algorithm::pooling_avg;
         auto kdims = op.kdims();
         std::vector<size_t> padding_l(op.padding.begin(), op.padding.begin() + kdims);
         std::vector<size_t> padding_r(op.padding.begin() + kdims, op.padding.end());
+        // Note: It is not documented, but the default dilation seems to be 0 instead of 1.
+        //       We need to offset dilations with -1.
+        std::vector<size_t> dilations;
+        std::transform(op.dilations.cbegin(),
+                       op.dilations.cend(),
+                       std::back_inserter(dilations),
+                       [](size_t d) { return d - 1; });
         return {dnnl::prop_kind::forward_inference,
                 algo,
                 m.at(MIGRAPHX_DNNL_PREFIX(ARG_SRC)),
                 m.at(MIGRAPHX_DNNL_PREFIX(ARG_DST)),
                 to_dnnl_dims(op.stride),
                 to_dnnl_dims(op.lengths),
+                to_dnnl_dims(dilations),
                 to_dnnl_dims(padding_l),
                 to_dnnl_dims(padding_r)};
     }

src/targets/gpu/CMakeLists.txt

 # ####################################################################################
 # The MIT License (MIT)
 #
-# Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
+# Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
 #
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -22,20 +22,20 @@
 # THE SOFTWARE.
 # ####################################################################################
 
-list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
-find_package(hip)
+find_package(hip REQUIRED)
 if(NOT GPU_TARGETS)
-    message(FATAL_ERROR "HIP package is broken and has no GPU_TARGETS, please pass -DGPU_TARGETS=$(/opt/rocm/bin/rocminfo | grep -o -m1 'gfx.*') to cmake to build for your gpu.")
+    set(fatal_msg "HIP package is broken and has no GPU_TARGETS. Please pass GPU_TARGETS to cmake.")
+    if(NOT WIN32)
+        set(fatal_msg "${fatal_msg}\nUse -DGPU_TARGETS=$(/opt/rocm/bin/rocminfo | grep -o -m1 'gfx.*') to build for your GPU.")
+    endif()
+    message(FATAL_ERROR ${fatal_msg})
 endif()
-find_package(miopen)
+find_package(miopen REQUIRED)
+message(STATUS "MIGraphX is using MIOpen")
 
 # rocblas
-find_package(rocblas REQUIRED PATHS /opt/rocm)
-message(STATUS "Build with rocblas")
-
-if(NOT TARGET MIOpen)
-    message(SEND_ERROR "Cant find miopen")
-endif()
+find_package(rocblas REQUIRED)
+message(STATUS "MIGraphX build with rocBLAS")
 
 if(MIGRAPHX_USE_COMPOSABLEKERNEL)
     find_package(composable_kernel 1.0.0 REQUIRED COMPONENTS jit_library)
@@ -49,7 +49,6 @@ endif()
 
 file(GLOB KERNEL_FILES CONFIGURE_DEPENDS
     ${CMAKE_CURRENT_SOURCE_DIR}/kernels/include/migraphx/kernels/*.hpp)
-message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
 
 if(NOT MIGRAPHX_USE_COMPOSABLEKERNEL)
     list(REMOVE_ITEM KERNEL_FILES
@@ -66,8 +65,10 @@ file(GLOB DEVICE_GPU_SRCS CONFIGURE_DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/device/*
 add_library(migraphx_device ${DEVICE_GPU_SRCS})
 
 add_library(compile_for_gpu INTERFACE)
-target_compile_options(compile_for_gpu INTERFACE -std=c++17 -fno-gpu-rdc -Wno-cuda-compat -Wno-unused-command-line-argument -Xclang -fallow-half-arguments-and-returns)
-target_link_libraries(compile_for_gpu INTERFACE hip::device -fno-gpu-rdc -Wno-invalid-command-line-argument -Wno-unused-command-line-argument -Wno-option-ignored)
+target_compile_features(compile_for_gpu INTERFACE cxx_std_17)
+target_compile_options(compile_for_gpu INTERFACE -fno-gpu-rdc -Wno-cuda-compat -Wno-unused-command-line-argument -Xclang -fallow-half-arguments-and-returns)
+target_link_options(compile_for_gpu INTERFACE  -fno-gpu-rdc -Wno-invalid-command-line-argument -Wno-unused-command-line-argument -Wno-option-ignored)
+target_link_libraries(compile_for_gpu INTERFACE hip::device)
 check_cxx_compiler_flag("--cuda-host-only -fhip-lambda-host-device -x hip" HAS_HIP_LAMBDA_HOST_DEVICE)
 
 if(HAS_HIP_LAMBDA_HOST_DEVICE)
@@ -211,8 +212,10 @@ if(MIGRAPHX_ENABLE_MLIR)
 endif()
 
 if(MIGRAPHX_USE_HIPRTC)
+    find_package(hiprtc REQUIRED)
     message(STATUS "MIGraphX is using hipRTC")
     target_compile_definitions(migraphx_gpu PRIVATE -DMIGRAPHX_USE_HIPRTC=1)
+    target_link_libraries(migraphx_gpu PUBLIC hiprtc::hiprtc)
 else()
     message(STATUS "MIGraphX is using HIP Clang")
 
@@ -221,12 +224,15 @@ else()
     target_flags(HIP_COMPILER_FLAGS hip::device)
 
     # Remove cuda arch flags
-    string(REGEX REPLACE --cuda-gpu-arch=[a-z0-9]+ "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
-    string(REGEX REPLACE --offload-arch=[a-z0-9:+-]+ "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
+    string(REGEX REPLACE "--cuda-gpu-arch=[a-z0-9]+ ?" "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
+    string(REGEX REPLACE "--offload-arch=[a-z0-9:+-]+ ?" "" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
 
     # Skip library paths since hip will incorrectly treat it as a source file
     string(APPEND HIP_COMPILER_FLAGS " ")
 
+    if(WIN32)
+        string(REPLACE "\\" "/" HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
+    endif()
     foreach(_unused RANGE 2)
         string(REGEX REPLACE " /[^ ]+\\.(a|so) " " " HIP_COMPILER_FLAGS "${HIP_COMPILER_FLAGS}")
     endforeach()
@@ -238,17 +244,25 @@ else()
     )
 
     if(DEFINED CMAKE_CXX_COMPILER_LAUNCHER)
-        execute_process(COMMAND which ${CMAKE_CXX_COMPILER_LAUNCHER} OUTPUT_VARIABLE MIGRAPHX_HIP_COMPILER_LAUNCHER)
+        if(WIN32)
+            execute_process(COMMAND where ${CMAKE_CXX_COMPILER_LAUNCHER} OUTPUT_VARIABLE MIGRAPHX_HIP_COMPILER_LAUNCHER)
+        else()
+            execute_process(COMMAND which ${CMAKE_CXX_COMPILER_LAUNCHER} OUTPUT_VARIABLE MIGRAPHX_HIP_COMPILER_LAUNCHER)
+        endif()
         string(STRIP "${MIGRAPHX_HIP_COMPILER_LAUNCHER}" MIGRAPHX_HIP_COMPILER_LAUNCHER)
         target_compile_definitions(migraphx_gpu PRIVATE -DMIGRAPHX_HIP_COMPILER_LAUNCHER="${MIGRAPHX_HIP_COMPILER_LAUNCHER}")
     endif()
 endif()
 
 # Check miopen find mode api
+
 include(CheckLibraryExists)
 get_target_property(MIOPEN_LOCATION MIOpen LOCATION)
+get_target_property(ROCBLAS_LOCATION roc::rocblas LOCATION)
 check_library_exists(MIOpen "miopenHiddenSetConvolutionFindMode" "${MIOPEN_LOCATION}" HAS_FIND_MODE_API)
 check_library_exists(MIOpen "miopenFindSolutions" "${MIOPEN_LOCATION}" HAS_FIND_2_API)
+# Beta API for automated GEMM tuning
+check_library_exists(roc::rocblas "rocblas_gemm_ex_get_solutions" "${ROCBLAS_LOCATION}" HAS_ROCBLAS_TUNING_BETA_FEATURE_API)
 
 set(MIGRAPHX_USE_FIND_2_API "${HAS_FIND_2_API}" CACHE BOOL "")
 
@@ -271,6 +285,13 @@ else()
     message(STATUS "MIOpen does not have find mode api")
 endif()
 
+if(HAS_ROCBLAS_TUNING_BETA_FEATURE_API)
+    target_compile_definitions(migraphx_gpu PUBLIC -DMIGRAPHX_USE_ROCBLAS_TUNING_API -DROCBLAS_BETA_FEATURES_API -DROCBLAS_NO_DEPRECATED_WARNINGS)
+    message(STATUS "MIGraphx is using Beta API of rocBLAS")
+else()
+    message(STATUS "rocBLAS does not have User Tuning Beta API")
+endif()
+
 target_link_libraries(migraphx_gpu PUBLIC migraphx MIOpen roc::rocblas)
 target_link_libraries(migraphx_gpu PRIVATE migraphx_device migraphx_kernels)
 if(MIGRAPHX_USE_COMPOSABLEKERNEL)

src/targets/gpu/compile_gen.cpp

@@ -54,6 +54,11 @@ vectorize vectorize::elements(std::size_t axis,
                               const std::vector<shape>& inputs,
                               const std::vector<std::size_t>& sizes)
 {
+    // disable vectorization for fp8 types
+    if(std::any_of(inputs.begin(), inputs.end(), [&](auto ishape) {
+           return ishape.type() == migraphx::shape::fp8e4m3fnuz_type;
+       }))
+        return {1, axis};
     if(std::all_of(
            inputs.begin(), inputs.end(), [&](const auto& s) { return s.lens()[axis] == 1; }))
         return {1, axis};
@@ -86,6 +91,11 @@ vectorize vectorize::elements(std::size_t axis,
 
 vectorize vectorize::elements(context& ctx, std::size_t axis, const std::vector<shape>& inputs)
 {
+    // disable vectorization for fp8 types
+    if(std::any_of(inputs.begin(), inputs.end(), [&](auto ishape) {
+           return ishape.type() == migraphx::shape::fp8e4m3fnuz_type;
+       }))
+        return {1, axis};
     if(inputs.empty())
         return {1, axis};
     std::size_t n = std::max_element(inputs.begin(),

src/targets/gpu/compile_hip.cpp

@@ -194,7 +194,7 @@ struct hiprtc_program
 };
 
 std::vector<std::vector<char>> compile_hip_src_with_hiprtc(std::vector<hiprtc_src_file> srcs,
-                                                           std::string params,
+                                                           const std::string& params,
                                                            const std::string& arch)
 {
     hiprtc_program prog(std::move(srcs));
@@ -238,8 +238,9 @@ bool hip_has_flags(const std::vector<std::string>& flags)
     }
 }
 
-std::vector<std::vector<char>>
-compile_hip_src(const std::vector<src_file>& srcs, std::string params, const std::string& arch)
+std::vector<std::vector<char>> compile_hip_src(const std::vector<src_file>& srcs,
+                                               const std::string& params,
+                                               const std::string& arch)
 {
     std::vector<hiprtc_src_file> hsrcs{srcs.begin(), srcs.end()};
     if(enabled(MIGRAPHX_GPU_DUMP_SRC{}))
@@ -251,10 +252,21 @@ compile_hip_src(const std::vector<src_file>& srcs, std::string params, const std
             std::cout << std::string(src.content) << std::endl;
         }
     }
+    auto fname = fs::path{"migraphx-hiprtc-driver"};
+#ifdef _WIN32
+    fname.replace_extension(".exe");
+#endif
     auto p      = dynamic_loader::path(&compile_hip_src_with_hiprtc);
-    auto driver = p.parent_path().parent_path() / "bin" / "migraphx-hiprtc-driver";
+    auto driver = p.parent_path() / fname;
 
-    if(fs::exists(driver))
+    bool found = fs::exists(driver);
+    if(not found)
+    {
+        driver = p.parent_path().parent_path() / "bin" / fname;
+        found  = fs::exists(driver);
+    }
+
+    if(found)
     {
         value v;
         v["srcs"]   = to_value(hsrcs);
@@ -270,13 +282,13 @@ compile_hip_src(const std::vector<src_file>& srcs, std::string params, const std
         if(fs::exists(out))
             return {read_buffer(out.string())};
     }
-    return compile_hip_src_with_hiprtc(std::move(hsrcs), std::move(params), arch);
+    return compile_hip_src_with_hiprtc(std::move(hsrcs), params, arch);
 }
 
 #else // MIGRAPHX_USE_HIPRTC
 
 std::vector<std::vector<char>> compile_hip_src_with_hiprtc(std::vector<hiprtc_src_file>, // NOLINT
-                                                           std::string,                  // NOLINT
+                                                           const std::string&,           // NOLINT
                                                            const std::string&)
 {
     MIGRAPHX_THROW("Not using hiprtc");
@@ -305,29 +317,15 @@ src_compiler assemble(src_compiler compiler)
     return compiler;
 }
 
-std::vector<std::vector<char>>
-compile_hip_src(const std::vector<src_file>& srcs, std::string params, const std::string& arch)
+std::vector<std::vector<char>> compile_hip_src(const std::vector<src_file>& srcs,
+                                               const std::string& params,
+                                               const std::string& arch)
 {
     assert(not srcs.empty());
+
     if(not is_hip_clang_compiler())
         MIGRAPHX_THROW("Unknown hip compiler: " MIGRAPHX_HIP_COMPILER);
 
-    if(params.find("-std=") == std::string::npos)
-        params += " --std=c++17";
-    params += " -fno-gpu-rdc";
-    if(enabled(MIGRAPHX_GPU_DEBUG_SYM{}))
-        params += " -g";
-    params += " -c";
-    params += " --offload-arch=" + arch;
-    params += " --cuda-device-only";
-    params += " -O" + string_value_of(MIGRAPHX_GPU_OPTIMIZE{}, "3") + " ";
-
-    if(enabled(MIGRAPHX_GPU_DEBUG{}))
-        params += " -DMIGRAPHX_DEBUG";
-
-    params += " -Wno-unused-command-line-argument -Wno-cuda-compat ";
-    params += MIGRAPHX_HIP_COMPILER_FLAGS;
-
     src_compiler compiler;
     compiler.flags    = params;
     compiler.compiler = MIGRAPHX_HIP_COMPILER;
@@ -335,6 +333,23 @@ compile_hip_src(const std::vector<src_file>& srcs, std::string params, const std
     if(has_compiler_launcher())
         compiler.launcher = MIGRAPHX_HIP_COMPILER_LAUNCHER;
 #endif
+
+    if(params.find("-std=") == std::string::npos)
+        compiler.flags += " --std=c++17";
+    compiler.flags += " -fno-gpu-rdc";
+    if(enabled(MIGRAPHX_GPU_DEBUG_SYM{}))
+        compiler.flags += " -g";
+    compiler.flags += " -c";
+    compiler.flags += " --offload-arch=" + arch;
+    compiler.flags += " --cuda-device-only";
+    compiler.flags += " -O" + string_value_of(MIGRAPHX_GPU_OPTIMIZE{}, "3") + " ";
+
+    if(enabled(MIGRAPHX_GPU_DEBUG{}))
+        compiler.flags += " -DMIGRAPHX_DEBUG";
+
+    compiler.flags += " -Wno-unused-command-line-argument -Wno-cuda-compat ";
+    compiler.flags += MIGRAPHX_HIP_COMPILER_FLAGS;
+
     if(enabled(MIGRAPHX_GPU_DUMP_SRC{}))
     {
         for(const auto& src : srcs)

src/targets/gpu/compile_hip_code_object.cpp

@@ -200,7 +200,7 @@ operation compile_hip_code_object(const std::string& content, hip_compile_option
     options.params += " " + join_strings(compiler_warnings(), " ");
     options.params += " -ftemplate-backtrace-limit=0";
     options.params += " -Werror";
-    auto cos = compile_hip_src(srcs, std::move(options.params), get_device_name());
+    auto cos = compile_hip_src(srcs, options.params, get_device_name());
     if(cos.size() != 1)
         MIGRAPHX_THROW("No code object");
     return code_object_op{value::binary{cos.front()},

src/targets/gpu/device/include/migraphx/gpu/device/scan.hpp

@@ -43,24 +43,32 @@ template <index_int N,
 __device__ void block_scan(index idx, Op op, T init, ForStride fs, Input input, Output output)
 {
     using type = decltype(input(deduce_for_stride(fs)));
-    MIGRAPHX_DEVICE_SHARED type buffer[N];
+    MIGRAPHX_DEVICE_SHARED type buffer[2][N];
     type x = init;
     fs([&](auto i) {
+        index_int iout = 0;
+        index_int iin  = 1;
         if(idx.local == 0)
-            buffer[idx.local] = op(input(i), x);
+            buffer[iout][idx.local] = op(input(i), x);
         else
-            buffer[idx.local] = input(i);
+            buffer[iout][idx.local] = input(i);
         __syncthreads();
         for(index_int s = 1; s < idx.nlocal(); s *= 2)
         {
-            if(idx.local + s < idx.nlocal())
+            iout = 1 - iout;
+            iin  = 1 - iin;
+            if(idx.local >= s)
             {
-                buffer[idx.local + s] = op(buffer[idx.local], buffer[idx.local + s]);
+                buffer[iout][idx.local] = op(buffer[iin][idx.local], buffer[iin][idx.local - s]);
+            }
+            else
+            {
+                buffer[iout][idx.local] = buffer[iin][idx.local];
             }
             __syncthreads();
         }
-        x = buffer[idx.nlocal() - 1];
-        output(i, buffer[idx.local]);
+        x = buffer[iout][idx.nlocal() - 1];
+        output(i, buffer[iout][idx.local]);
     });
 }