Merge branch 'develop' into mlir-nonstandard-shapes

README.md

@@ -67,7 +67,7 @@ The following is a list of prerequisites for building MIGraphX.
 3. Build MIGraphX source code:
 
     ```bash
-    rbuild build -d depend -B build
+    rbuild build -d depend -B build -DGPU_TARGETS=$(/opt/rocm/bin/rocminfo | grep -o -m1 'gfx.*')
     ```
 
 Once completed, all prerequisites are in the `depend` folder and MIGraphX is in the `build` directory.
@@ -106,7 +106,7 @@ the folder to `PATH`, or add the option `--prefix /usr/local` in the pip3 comman
 3. Configure CMake. If the prerequisites are installed at the default location `/usr/local`, use:
 
     ```bash
-    CXX=/opt/rocm/llvm/bin/clang++ cmake ..
+    CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DGPU_TARGETS=$(/opt/rocm/bin/rocminfo | grep -o -m1 'gfx.*')
     ```
 
     Otherwise, you need to set `-DCMAKE_PREFIX_PATH=$your_loc` to configure CMake.

docs/.sphinx/requirements.txt

@@ -75,7 +75,9 @@ pygments==2.15.0
     #   pydata-sphinx-theme
     #   sphinx
 pyjwt[crypto]==2.6.0
-    # via pygithub
+    # via
+    #   pygithub
+    #   pyjwt
 pynacl==1.5.0
     # via pygithub
 pyyaml==6.0
@@ -87,7 +89,7 @@ requests==2.28.2
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==0.26.0
+rocm-docs-core==0.27.0
     # via -r requirements.in
 smmap==5.0.0
     # via gitdb

docs/contributor_guide.rst

 Contributor Guide
-===============
+=================
 
 .. toctree::
    :maxdepth: 2
    :caption: Contents:
 
-   dev_intro
+   dev/dev_intro
    dev/data
    dev/operators
    dev/program

docs/dev_intro.rst → docs/dev/dev_intro.rst

-MIGraphX Fundamentals
+Developer Introduction
 ======================
 
 MIGraphX provides an optimized execution engine for deep learning neural networks.

docs/driver.rst

 MIGraphX Driver
 ===============
 
+The MIGraphX driver is a tool that allows you to utilize many of the core functions of MIGraphX without having to write your own program. It can read, compile, run, and test the performance of a model with randomized data.
+
 read
 ----
 
@@ -17,6 +19,7 @@ compile
 
 Compiles and prints input graph.
 
+.. include:: ./driver/read.rst
 .. include:: ./driver/compile.rst
 
 run
@@ -26,6 +29,7 @@ run
 
 Loads and prints input graph.
 
+.. include:: ./driver/read.rst
 .. include:: ./driver/compile.rst
 
 perf
@@ -35,6 +39,7 @@ perf
 
 Compiles and runs input graph then prints performance report.
 
+.. include:: ./driver/read.rst
 .. include:: ./driver/compile.rst
 
 .. option::  --iterations, -n [unsigned int]
@@ -48,6 +53,7 @@ verify
 
 Runs reference and CPU or GPU implementations and checks outputs for consistency.
 
+.. include:: ./driver/read.rst
 .. include:: ./driver/compile.rst
 
 .. option::  --rms-tol [double]
@@ -71,7 +77,7 @@ Verify each instruction
 Reduce program and verify
 
 roctx
-----
+-----
 
 .. program:: migraphx-driver roctx
 
@@ -86,4 +92,5 @@ An example command line combined with rocprof for tracing purposes is given belo
 After `rocprof` is run, the output directory will contain trace information for HIP, HCC and ROCTX in seperate `.txt` files.
 To understand the interactions between API calls, it is recommended to utilize `roctx.py` helper script as desribed in :ref:`dev/tools:rocTX` section. 
 
+.. include:: ./driver/read.rst
 .. include:: ./driver/compile.rst

docs/driver/compile.rst

-.. include:: ./driver/read.rst
-
 .. option::  --fill0 [std::vector<std::string>]
 
 Fill parameter with 0s

docs/driver/read.rst

@@ -46,11 +46,11 @@ Trim instructions from the end (Default: 0)
 
 Dim of a parameter (format: "@name d1 d2 dn")
 
-.. options:: --dyn-input-dim [std::vector<std::string>]
+.. option:: --dyn-input-dim [std::vector<std::string>]
 
 Set dynamic dimensions of a parameter using JSON formatting (format "@name" "dynamic_dimension_json")
 
-.. options:: --default-dyn-dim
+.. option:: --default-dyn-dim
 
 Set the default dynamic dimension (format {min:x, max:y, optimals:[o1,o2,...]})
 

docs/reference/py.rst

@@ -95,7 +95,7 @@ shape
     :rtype: bool
 
 dynamic_dimension
---------
+-----------------
 
 .. py:class:: dynamic_dimension(min, max, optimals)
 

src/CMakeLists.txt

@@ -175,6 +175,7 @@ register_migraphx_ops(
     mul
     multibroadcast
     multinomial
+    nearbyint
     neg
     nonmaxsuppression
     nonzero
@@ -205,7 +206,6 @@ register_migraphx_ops(
     rnn_last_hs_output
     rnn_var_sl_last_output
     roialign
-    round
     rsqrt
     run_on_target
     scalar

src/driver/main.cpp

@@ -59,6 +59,13 @@ namespace migraphx {
 namespace driver {
 inline namespace MIGRAPHX_INLINE_NS {
 
+inline std::string get_version()
+{
+    return "MIGraphX Version: " + std::to_string(MIGRAPHX_VERSION_MAJOR) + "." +
+           std::to_string(MIGRAPHX_VERSION_MINOR) + "." + std::to_string(MIGRAPHX_VERSION_PATCH) +
+           "." MIGRAPHX_VERSION_TWEAK;
+}
+
 struct loader
 {
     std::string model;
@@ -597,16 +604,6 @@ struct verify : command<verify>
     }
 };
 
-struct version : command<version>
-{
-    void parse(const argument_parser&) {}
-    void run() const
-    {
-        std::cout << "MIGraphX Version: " << MIGRAPHX_VERSION_MAJOR << "." << MIGRAPHX_VERSION_MINOR
-                  << "." << MIGRAPHX_VERSION_PATCH << "." MIGRAPHX_VERSION_TWEAK << std::endl;
-    }
-};
-
 struct compile : command<compile>
 {
     compiler c;
@@ -759,16 +756,14 @@ struct main_command
     }
     void parse(argument_parser& ap)
     {
-        std::string version_str = "MIGraphX Version: " + std::to_string(MIGRAPHX_VERSION_MAJOR) +
-                                  "." + std::to_string(MIGRAPHX_VERSION_MINOR) + "." +
-                                  std::to_string(MIGRAPHX_VERSION_PATCH) +
-                                  "." MIGRAPHX_VERSION_TWEAK;
+        std::string version_str = get_version();
         ap(wrong_commands, {}, ap.metavar("<command>"), ap.append());
         ap(nullptr, {"-h", "--help"}, ap.help("Show help"), ap.show_help(get_command_help()));
         ap(nullptr,
            {"-v", "--version"},
            ap.help("Show MIGraphX version"),
            ap.show_help(version_str));
+        ap(nullptr, {"--ort-sha"}, ap.help("Show MIGraphX onnx runtime SHA"));
 
         // Trim command off of exe name
         ap.set_exe_name(ap.get_exe_name().substr(0, ap.get_exe_name().size() - 5));
@@ -811,7 +806,6 @@ using namespace migraphx::driver; // NOLINT
 int main(int argc, const char* argv[])
 {
     std::vector<std::string> args(argv + 1, argv + argc);
-
     // no argument, print the help infomration by default
     if(args.empty())
     {
@@ -821,15 +815,27 @@ int main(int argc, const char* argv[])
     auto&& m = get_commands();
     auto cmd = args.front();
 
-    if(cmd == "ort-sha")
+    if(cmd == "--ort-sha")
     {
         std::cout << MIGRAPHX_ORT_SHA1 << std::endl;
         return 0;
     }
+    if(cmd == "-v" or cmd == "--version")
+    {
+        std::cout << get_version() << std::endl;
+        return 0;
+    }
 
     if(m.count(cmd) > 0)
     {
-        m.at(cmd)(argv[0], {args.begin() + 1, args.end()});
+        std::string driver_invocation =
+            std::string(argv[0]) + " " + migraphx::to_string_range(args, " ");
+        std::cout << "Running [ " << get_version() << " ]: " << driver_invocation << std::endl;
+
+        m.at(cmd)(argv[0],
+                  {args.begin() + 1, args.end()}); // run driver command found in commands map
+
+        std::cout << "[ " << get_version() << " ] Complete: " << driver_invocation << std::endl;
     }
     else
     {

src/include/migraphx/module.hpp

@@ -42,6 +42,7 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
+MIGRAPHX_EXPORT
 const operation& get_operation(instruction_ref ins);
 
 struct module_impl;

src/include/migraphx/op/isinf.hpp

@@ -35,7 +35,7 @@ struct isinf : unary<isinf>
 {
     auto apply() const
     {
-        return [&](auto x) { return std::isinf(x); };
+        return [&](auto x) { return std::isinf(static_cast<double>(x)); };
     }
 
     std::string name() const { return "isinf"; }

src/include/migraphx/op/round.hpp → src/include/migraphx/op/nearbyint.hpp

 /*
  * 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
@@ -21,24 +21,28 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  * THE SOFTWARE.
  */
-#ifndef MIGRAPHX_GUARD_OPERATORS_ROUND_HPP
-#define MIGRAPHX_GUARD_OPERATORS_ROUND_HPP
+#ifndef MIGRAPHX_GUARD_OPERATORS_NEARBYINT_HPP
+#define MIGRAPHX_GUARD_OPERATORS_NEARBYINT_HPP
 
 #include <migraphx/op/unary.hpp>
 #include <migraphx/config.hpp>
+#include <fenv.h>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
-
-struct round : unary<round>
+struct nearbyint : unary<nearbyint>
 {
     auto apply() const
     {
-        return [](auto x) { return std::round(x); };
+        return [](auto x) {
+            auto rounding_mode = fegetround();
+            fesetround(FE_TONEAREST);
+            return std::nearbyint(x);
+            fesetround(rounding_mode);
+        };
     }
 };
-
 } // namespace op
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/op/normalize_attribute.hpp

@@ -40,6 +40,8 @@ namespace op {
  * 2. use_rank (default) vs use_len:
  *  `use_rank` sets the max value/index of the attribute as the rank of lens.
  *  `use_lens` sets the max value/index as the corresponding value in lens at the axes index.
+ *      Uses the dynamic_dimension.max value for dynamic shapes. Returns the original vector
+ *      (no normalization) if any of dynamic_dimension[axes] are not fixed.
  * 3. `clip_min` vs. `not_clip_min` (default):
  *  Clip values less than the minimum to the minimum or not.
  * 4. `include_min` vs. `exclude_min` (default):

src/include/migraphx/op/quantizelinear.hpp

@@ -30,11 +30,11 @@
 #include <migraphx/par_for.hpp>
 #include <migraphx/value.hpp>
 #include <cmath>
+#include <fenv.h>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace op {
-
 struct quantizelinear
 {
     std::string name() const { return "quantizelinear"; }
@@ -71,26 +71,26 @@ struct quantizelinear
         {
             y_zero_point = args.at(2);
         }
-
         argument result{output_shape};
+        auto rounding_mode = fegetround();
+        fesetround(FE_TONEAREST);
         visit_all(result, y_zero_point)([&](auto output, auto zero_pts) {
             visit_all(x, y_scale)([&](auto input, auto scales) {
                 using quant_type = typename decltype(output)::value_type;
                 auto min_value   = std::numeric_limits<quant_type>::min();
                 auto max_value   = std::numeric_limits<quant_type>::max();
                 par_for(output_shape.elements(), [&](auto i) {
-                    int64_t quantized = static_cast<int64_t>(std::round(input[i] / scales[i])) +
+                    int64_t quantized = static_cast<int64_t>(std::nearbyint(input[i] / scales[i])) +
                                         static_cast<int64_t>(zero_pts[i]);
                     output[i] = std::max(static_cast<int64_t>(min_value),
                                          std::min(static_cast<int64_t>(max_value), quantized));
                 });
             });
         });
-
+        fesetround(rounding_mode);
         return result;
     }
 };
-
 } // namespace op
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/op/slice.hpp

@@ -38,6 +38,18 @@ namespace op {
 
 /**
  * Slice operator that accepts variable axes, starts and ends.
+ * All of `starts`, `ends`, and `axes` must be supplied by either
+ * their attribute or an input (but not both).
+ *
+ * Valid calls:
+ * slice(input); axes, starts, ends set
+ * slice(input, starts); axes, ends set
+ * slice(input, ends); starts, axes set
+ * slice(input, axes); starts, ends set
+ * slice(input, starts, ends); axes set
+ * slice(input, starts, axes); ends set
+ * slice(input, ends, axes); starts set
+ * slice(input, start, ends, axes); none set
  *
  * Attributes:
  * axes: constant axes to slice over (optional)
@@ -46,8 +58,8 @@ namespace op {
  *
  * Parameters:
  * data: the input tensor to slice (dynamic or static shape)
- * input_starts: starting indicies of slice (optional, static shape)
- * input_ends: ending indicies of slice (optional, static shape)
+ * input_starts: starting indices of slice (optional, static shape)
+ * input_ends: ending indices of slice (optional, static shape)
  * input_axes: axes to slice over (optional, static shape)
  */
 struct slice
@@ -56,6 +68,18 @@ struct slice
     std::vector<int64_t> starts{};
     std::vector<int64_t> ends{};
 
+    /**
+     * Named arrays for the set attribute possibilities.
+     */
+    static constexpr std::array<bool, 3> all_set     = {true, true, true};
+    static constexpr std::array<bool, 3> ends_axes   = {false, true, true};
+    static constexpr std::array<bool, 3> starts_axes = {true, false, true};
+    static constexpr std::array<bool, 3> starts_ends = {true, true, false};
+    static constexpr std::array<bool, 3> axes_only   = {false, false, true};
+    static constexpr std::array<bool, 3> ends_only   = {false, true, false};
+    static constexpr std::array<bool, 3> starts_only = {true, false, false};
+    static constexpr std::array<bool, 3> none_set    = {false, false, false};
+
     template <class Self, class F>
     static auto reflect(Self& self, F f)
     {
@@ -63,24 +87,26 @@ struct slice
     }
 
     /**
-     * Ensure that attribute vectors axes, starts, and ends are all the same size and values are
-     * within limits.
+     * Ensure that attribute axes is within limits.
+     * Will attempt to normalize starts and ends; but will use the dynamic_dimension.max
+     * values for dynamic shapes. This makes it so you have to renormalize for
+     * non-fixed dynamic_dimensions.
      */
     value attributes() const
     {
-        value normalize     = value::object{};
-        normalize["axes"]   = value::array{normalize_attribute::include_min};
-        normalize["starts"] = value::array{normalize_attribute::clip_max,
+        value normalize_axes     = value::object{};
+        normalize_axes["axes"]   = value::array{normalize_attribute::include_min};
+        normalize_axes["starts"] = value::array{normalize_attribute::clip_max,
                                                 normalize_attribute::clip_min,
                                                 normalize_attribute::include_max,
                                                 normalize_attribute::use_len,
                                                 normalize_attribute::include_min};
-        normalize["ends"]   = value::array{normalize_attribute::clip_max,
+        normalize_axes["ends"]   = value::array{normalize_attribute::clip_max,
                                               normalize_attribute::clip_min,
                                               normalize_attribute::include_max,
                                               normalize_attribute::use_len,
                                               normalize_attribute::include_min};
-        return {{"normalize_axes", normalize}};
+        return {{"normalize_axes", normalize_axes}};
     }
 
     std::string name() const { return "slice"; }
@@ -88,7 +114,7 @@ struct slice
     /**
      * Computes the slice output shape dimensions for given starts, ends,and axes.
      * Templated to also handle tensor views.
-     * Possibily different type between [in_starts, in_ends] and [in_axes] if in_axes is this
+     * Possibly different type between [in_starts, in_ends] and [in_axes] if in_axes is this
      * object's axes attribute. Assumes in_starts and in_ends are normalized; in_axes are valid.
      */
     template <class A, class B>
@@ -104,63 +130,161 @@ struct slice
         return new_lens;
     }
 
-    shape normalize_compute_shape(std::vector<shape> inputs) const
+    /// Get the attributes that are non-empty
+    std::array<bool, 3> get_set_attributes() const
+    {
+        std::array<std::vector<int64_t>, 3> attrs = {this->starts, this->ends, this->axes};
+        std::array<bool, 3> bool_vec;
+        std::transform(
+            attrs.cbegin(), attrs.cend(), bool_vec.begin(), [](auto a) { return not a.empty(); });
+        return bool_vec;
+    }
+
+    /// Helper function for normalize_compute_shape()
+    shape compute_two_or_more(std::vector<shape> inputs) const
     {
-        check_shapes{inputs, *this, true}.has(1, 3, 4);
         auto input_shape    = inputs[0];
-        if(inputs.size() == 1)
+        auto set_attributes = get_set_attributes();
+        // check that inputs [1, end) are all 1D, have the same
+        // dimension, and are static
+        check_shapes{inputs.begin() + 1,
+                     inputs.end(),
+                     std::string("SLICE: inputs (starts, ends, and input_axes)"),
+                     false}
+            .only_dims(1)
+            .same_dims();
+        auto dds = input_shape.to_dynamic().dyn_dims();
+        if(inputs.size() == 2)
         {
-            auto t = input_shape.type();
-            if(input_shape.dynamic() and std::any_of(axes.begin(), axes.end(), [&](auto axis) {
-                   return not input_shape.dyn_dims()[axis].is_fixed();
-               }))
+            if(set_attributes == ends_axes)
             {
-                MIGRAPHX_THROW("SLICE: slicing is not allowed on non-fixed dynamic input axis ");
+                // attr ends and axes set; inputs are (data, input_starts)
+                if(inputs[1].lens().at(0) != axes.size())
+                {
+                    MIGRAPHX_THROW("SLICE: 2 input and attributes mismatch");
                 }
-            if(input_shape.dynamic())
+                std::for_each(axes.cbegin(), axes.cend(), [&](const auto& axis) {
+                    dds.at(axis) = {0, dds.at(axis).max};
+                });
+            }
+            else if(set_attributes == starts_axes)
             {
-                return shape{t,
-                             lens_calc(input_shape.min_lens(), starts, ends, axes),
-                             lens_calc(input_shape.max_lens(), starts, ends, axes),
-                             {}};
+                // attr starts and axes set; inputs are (data, input_ends)
+                if(inputs[1].lens().at(0) != axes.size())
+                {
+                    MIGRAPHX_THROW("SLICE: 2 input and attributes mismatch");
                 }
-            else
+                std::for_each(axes.cbegin(), axes.cend(), [&](const auto& axis) {
+                    dds.at(axis) = {0, dds.at(axis).max};
+                });
+            }
+            else if(set_attributes == starts_ends)
             {
-                return shape{
-                    t, lens_calc(input_shape.lens(), starts, ends, axes), input_shape.strides()};
+                // attr starts and ends set; inputs are (data, input_axes)
+                if(inputs[1].lens().at(0) != starts.size())
+                {
+                    MIGRAPHX_THROW("SLICE: 2 input and attributes mismatch");
                 }
+                std::transform(dds.begin(), dds.end(), dds.begin(), [](auto dd) {
+                    return shape::dynamic_dimension{0, dd.max};
+                });
             }
             else
             {
-            // check that starts, ends, and optionally input_axes are all 1D, have the same
-            // dimension, and are static
-            check_shapes{inputs.begin() + 1,
-                         inputs.end(),
-                         std::string("SLICE: inputs (starts, ends, and input_axes)"),
-                         false}
-                .only_dims(1)
-                .same_dims();
-            auto dds = input_shape.to_dynamic().dyn_dims();
-            if(inputs.size() == 3)
+                MIGRAPHX_THROW("SLICE: Invalid 2 input and attributes configuration");
+            }
+        }
+        else if(inputs.size() == 3)
+        {
+            if(set_attributes == axes_only)
             {
+                // attr axes set; inputs are (data, input_starts, input_ends)
                 if(inputs[1].lens().at(0) != axes.size())
                 {
-                    MIGRAPHX_THROW("SLICE: inputs starts and ends do not have the same dimension "
-                                   "as the axes attribute");
+                    MIGRAPHX_THROW("SLICE: 3 input and attributes mismatch");
                 }
                 std::for_each(axes.cbegin(), axes.cend(), [&](const auto& axis) {
                     dds.at(axis) = {0, dds.at(axis).max};
                 });
             }
+            else if(set_attributes == ends_only)
+            {
+                // attr ends set; inputs are (data, input_starts, input_axes)
+                if(inputs[1].lens().at(0) != ends.size())
+                {
+                    MIGRAPHX_THROW("SLICE: 3 input and attributes mismatch");
+                }
+                std::transform(dds.begin(), dds.end(), dds.begin(), [](auto dd) {
+                    return shape::dynamic_dimension{0, dd.max};
+                });
+            }
+            else if(set_attributes == starts_only)
+
+            {
+                // attr starts set; inputs are (data, input_ends, input_axes)
+                if(inputs[1].lens().at(0) != starts.size())
+                {
+                    MIGRAPHX_THROW("SLICE: 3 input and attributes mismatch");
+                }
+                std::transform(dds.begin(), dds.end(), dds.begin(), [](auto dd) {
+                    return shape::dynamic_dimension{0, dd.max};
+                });
+            }
             else
             {
-                // if axes is an input, then all the output dimensions could be 0 to the max value
+                MIGRAPHX_THROW("Invalid 3 input and attributes configuration");
+            }
+        }
+        else
+        {
+            // all 4 inputs (data, inputs_starts, input_ends, input_axes)
             std::transform(dds.begin(), dds.end(), dds.begin(), [](auto dd) {
                 return shape::dynamic_dimension{0, dd.max};
             });
         }
         return shape{input_shape.type(), dds};
     }
+
+    // uses the normalize_axes flag to normalize axes, starts, and ends
+    shape normalize_compute_shape(std::vector<shape> inputs) const
+    {
+        check_shapes{inputs, *this, true}.has(1, 2, 3, 4);
+        if(inputs.size() == 1)
+        {
+            auto input_shape    = inputs[0];
+            auto set_attributes = get_set_attributes();
+            if(set_attributes != all_set)
+            {
+                MIGRAPHX_THROW("SLICE 1_arg: Invalid 1 input and attributes configuration");
+            }
+            // NOTE: make sure to update how normalization works here if this type of slicing is
+            // changed to be allowed
+            if(input_shape.dynamic() and std::any_of(axes.begin(), axes.end(), [&](auto axis) {
+                   return not input_shape.dyn_dims()[axis].is_fixed();
+               }))
+            {
+                MIGRAPHX_THROW(
+                    "SLICE 1_arg: slicing is not allowed on non-fixed dynamic input axis ");
+            }
+            if(input_shape.dynamic())
+            {
+                return shape{
+                    input_shape.type(),
+                    lens_calc(input_shape.min_lens(), this->starts, this->ends, this->axes),
+                    lens_calc(input_shape.max_lens(), this->starts, this->ends, this->axes),
+                    {}};
+            }
+            else
+            {
+                return shape{input_shape.type(),
+                             lens_calc(input_shape.lens(), this->starts, this->ends, this->axes),
+                             input_shape.strides()};
+            }
+        }
+        else
+        {
+            return compute_two_or_more(inputs);
+        }
     }
 
     /**
@@ -194,14 +318,14 @@ struct slice
 
     /**
      * Calculates the starting offset for the sliced tensor (for aliasing).
-     * Used when the starts and/or the axes are inputs.
+     * Used for 2-4 inputs to `slice.
      *
      * \param s static input shape
      * \param input_starts starting indices of slice
      * \param ax_vec axes to slice on
      */
-    template <class IndView, class Axes>
-    auto compute_offset(const shape& s, const IndView& input_starts, const Axes& ax_vec) const
+    template <class T>
+    auto compute_offset(const shape& s, const T& input_starts, const T& ax_vec) const
     {
         auto ret = 0;
         for(std::size_t i = 0; i < ax_vec.size(); ++i)
@@ -212,106 +336,168 @@ struct slice
         return ret * s.type_size();
     }
 
-    std::unordered_map<std::string, std::vector<int64_t>>
-    normalize_inputs(const shape& input_shape,
-                     const std::vector<int64_t>& input_starts,
-                     const std::vector<int64_t>& input_ends) const
-    {
-        auto attrs = this->attributes().at("normalize_axes");
-        return {{"input_starts",
-                 normalize_indices(input_starts,
-                                   this->axes,
-                                   input_shape,
-                                   attrs.at("starts"),
-                                   "Slice variable input_starts")},
-                {"input_ends",
-                 normalize_indices(input_ends,
-                                   this->axes,
-                                   input_shape,
-                                   attrs.at("ends"),
-                                   "Slice variable input_ends")}};
-    }
-
     /**
-     * Three input version of the normalize_inputs.
-     * This one also checks that the input_axes are valid.
+     * If given, normalize the inputs. Otherwise get from operator attributes.
+     * Return the values in a map.
+     *
+     * Parameters
+     * input_shape: static shape of the input
+     * input_starts: optional
+     * input_ends: optional
+     * input_ends: optional
      */
     std::unordered_map<std::string, std::vector<int64_t>>
-    normalize_inputs(shape input_shape,
-                     const std::vector<int64_t>& input_starts,
-                     const std::vector<int64_t>& input_ends,
-                     const std::vector<int64_t>& input_axes) const
-    {
-        auto attrs = this->attributes().at("normalize_axes");
-        auto norm_axes =
-            normalize_axes(input_axes, input_shape, attrs.at("axes"), "Slice variable input_axes");
-        return {{"input_starts",
-                 normalize_indices(input_starts,
+    normalize_starts_ends_axes(shape input_shape,
+                               const optional<std::vector<int64_t>>& input_starts,
+                               const optional<std::vector<int64_t>>& input_ends,
+                               const optional<std::vector<int64_t>>& input_axes) const
+    {
+        auto axes_attrs = this->attributes().at("normalize_axes");
+        std::vector<int64_t> norm_starts;
+        std::vector<int64_t> norm_ends;
+        std::vector<int64_t> norm_axes;
+        if(input_axes)
+        {
+            norm_axes = normalize_axes(input_axes.value(),
+                                       input_shape,
+                                       axes_attrs.at("axes"),
+                                       "Slice variable input_axes");
+        }
+        else
+        {
+            norm_axes = this->axes;
+        }
+        if(input_starts)
+        {
+            norm_starts = normalize_indices(input_starts.value(),
                                             norm_axes,
                                             input_shape,
-                                   attrs.at("starts"),
-                                   "Slice variable input_starts")},
-                {"input_ends",
-                 normalize_indices(input_ends,
+                                            axes_attrs.at("starts"),
+                                            "Slice variable input_starts");
+        }
+        else
+        {
+            norm_starts = this->starts;
+        }
+        if(input_ends)
+        {
+            norm_ends = normalize_indices(input_ends.value(),
                                           norm_axes,
                                           input_shape,
-                                   attrs.at("ends"),
-                                   "Slice variable input ends")},
-                {"input_axes", norm_axes}};
+                                          axes_attrs.at("ends"),
+                                          "Slice variable input ends");
+        }
+        else
+        {
+            norm_ends = this->ends;
+        }
+        return {{"norm_starts", norm_starts}, {"norm_ends", norm_ends}, {"norm_axes", norm_axes}};
     }
 
     argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
     {
         auto input       = args[0];
         auto input_shape = input.get_shape();
-        switch(args.size())
+        if(args.size() == 1)
         {
-        case 1: {
             std::size_t offset = compute_offset(input_shape);
             return {dyn_out.computed_shape, [=] { return input.data() + offset; }};
         }
-        case 3: {
-            shape calc_shape;
-            std::size_t offset = 0;
+        else
+        {
+            // Note that we re-normalize both the attributes and inputs because of the non-fixed
+            // dynamic input shape case. It's possible to only re-normalize if slicing over
+            // non-fixed dynamic_dimensions.
+            auto set_attributes = get_set_attributes();
+            std::unordered_map<std::string, std::vector<int64_t>> norm_inputs;
+            if(set_attributes == ends_axes)
+            {
+                // attr ends and axes set; inputs are (data, input_starts)
+                args[1].visit([&](auto input_starts) {
+                    norm_inputs =
+                        normalize_starts_ends_axes(input_shape,
+                                                   input_starts.template to_vector<int64_t>(),
+                                                   this->ends,
+                                                   this->axes);
+                });
+            }
+            else if(set_attributes == starts_axes)
+            {
+                // attr starts and axes set; inputs are (data, input_ends)
+                args[1].visit([&](auto input_ends) {
+                    norm_inputs =
+                        normalize_starts_ends_axes(input_shape,
+                                                   this->starts,
+                                                   input_ends.template to_vector<int64_t>(),
+                                                   this->axes);
+                });
+            }
+            else if(set_attributes == starts_ends)
+            {
+                // attr starts and ends set; inputs are (data, input_axes)
+                args[1].visit([&](auto input_axes) {
+                    norm_inputs =
+                        normalize_starts_ends_axes(input_shape,
+                                                   this->starts,
+                                                   this->ends,
+                                                   input_axes.template to_vector<int64_t>());
+                });
+            }
+            else if(set_attributes == axes_only)
+            {
+                // attr axes set; inputs are (data, input_starts, input_ends)
                 visit_all(args[1], args[2])([&](auto input_starts, auto input_ends) {
-                auto norm_inputs = normalize_inputs(input_shape,
+                    norm_inputs =
+                        normalize_starts_ends_axes(input_shape,
                                                    input_starts.template to_vector<int64_t>(),
-                                                    input_ends.template to_vector<int64_t>());
-                offset = compute_offset(input_shape, norm_inputs.at("input_starts"), this->axes);
-                calc_shape = {input_shape.type(),
-                              lens_calc(input_shape.lens(),
-                                        norm_inputs.at("input_starts"),
-                                        norm_inputs.at("input_ends"),
-                                        this->axes),
-                              input_shape.strides()};
+                                                   input_ends.template to_vector<int64_t>(),
+                                                   this->axes);
+                });
+            }
+            else if(set_attributes == ends_only)
+            {
+                // attr ends set; inputs are (data, input_starts, input_axes)
+                visit_all(args[1], args[2])([&](auto input_starts, auto input_axes) {
+                    norm_inputs =
+                        normalize_starts_ends_axes(input_shape,
+                                                   input_starts.template to_vector<int64_t>(),
+                                                   this->ends,
+                                                   input_axes.template to_vector<int64_t>());
                 });
-            return {calc_shape, [=] { return input.data() + offset; }};
             }
-        case 4: {
-            shape calc_shape;
-            std::size_t offset = 0;
+            else if(set_attributes == starts_only)
+            {
+                // attr starts set; inputs are (data, input_ends, input_axes)
+                visit_all(args[1], args[2])([&](auto input_ends, auto input_axes) {
+                    norm_inputs =
+                        normalize_starts_ends_axes(input_shape,
+                                                   this->starts,
+                                                   input_ends.template to_vector<int64_t>(),
+                                                   input_axes.template to_vector<int64_t>());
+                });
+            }
+            else
+            {
+                // no attr set, all inputs
                 visit_all(args[1], args[2], args[3])(
                     [&](auto input_starts, auto input_ends, auto input_axes) {
-                    auto norm_inputs = normalize_inputs(input_shape,
+                        norm_inputs =
+                            normalize_starts_ends_axes(input_shape,
                                                        input_starts.template to_vector<int64_t>(),
                                                        input_ends.template to_vector<int64_t>(),
                                                        input_axes.template to_vector<int64_t>());
-                    offset           = compute_offset(
-                        input_shape, norm_inputs.at("input_starts"), norm_inputs.at("input_axes"));
-                    calc_shape = shape{input_shape.type(),
+                    });
+            }
+            auto offset = compute_offset(
+                input_shape, norm_inputs.at("norm_starts"), norm_inputs.at("norm_axes"));
+            shape calc_shape = shape{input_shape.type(),
                                      lens_calc(input_shape.lens(),
-                                                 norm_inputs.at("input_starts"),
-                                                 norm_inputs.at("input_ends"),
-                                                 norm_inputs.at("input_axes")),
+                                               norm_inputs.at("norm_starts"),
+                                               norm_inputs.at("norm_ends"),
+                                               norm_inputs.at("norm_axes")),
                                      input_shape.strides()};
-                });
             return {calc_shape, [=] { return input.data() + offset; }};
         }
-        default: {
-            // Should never get here; covering in case some code change occurs
-            MIGRAPHX_THROW("SLICE: invalid number of inputs");
-        }
-        }
     }
 
     std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 0; }

src/include/migraphx/operators.hpp

@@ -84,6 +84,7 @@
 #include <migraphx/op/mod.hpp>
 #include <migraphx/op/mul.hpp>
 #include <migraphx/op/multibroadcast.hpp>
+#include <migraphx/op/nearbyint.hpp>
 #include <migraphx/op/neg.hpp>
 #include <migraphx/op/nonmaxsuppression.hpp>
 #include <migraphx/op/nonzero.hpp>
@@ -110,7 +111,6 @@
 #include <migraphx/op/rnn_variable_seq_lens.hpp>
 #include <migraphx/op/rnn_var_sl_last_output.hpp>
 #include <migraphx/op/roialign.hpp>
-#include <migraphx/op/round.hpp>
 #include <migraphx/op/rsqrt.hpp>
 #include <migraphx/op/scalar.hpp>
 #include <migraphx/op/scatter_add.hpp>

src/normalize_attributes.cpp

@@ -66,15 +66,15 @@ auto tune_attribute(const std::vector<int64_t>& vec,
     {
         if(input_shape.dynamic())
         {
-            std::transform(axes.begin(), axes.end(), max_vals.begin(), [&](auto i) {
-                const auto& dd = input_shape.dyn_dims().at(i);
-                if(not dd.is_fixed())
+            // return the unchanged `vec` if the dynamic_dimensions at `axes` are not fixed
+            if(std::any_of(axes.begin(), axes.end(), [&](auto ax) {
+                   return not input_shape.dyn_dims().at(ax).is_fixed();
+               }))
             {
-                    MIGRAPHX_THROW(
-                        "NORMALIZE_ATTR: 'use_lens' on a non-fixed dynamic dimension, axis=" +
-                        std::to_string(i));
+                return vec;
             }
-                return dd.max;
+            std::transform(axes.begin(), axes.end(), max_vals.begin(), [&](auto i) {
+                return input_shape.dyn_dims().at(i).max;
             });
         }
         else

src/onnx/parse_generic_op.cpp

@@ -60,7 +60,7 @@ struct parse_generic_op : op_parser<parse_generic_op>
                 {"Neg", "neg"},
                 {"Reciprocal", "recip"},
                 {"Relu", "relu"},
-                {"Round", "round"},
+                {"Round", "nearbyint"},
                 {"Sigmoid", "sigmoid"},
                 {"Sign", "sign"},
                 {"Sin", "sin"},

src/onnx/parse_resize.cpp

@@ -181,41 +181,23 @@ static std::string get_nearest_mode(const onnx_parser::attribute_map& attr)
     return nearest_mode;
 }
 
-struct parse_resize : op_parser<parse_resize>
+static std::vector<double> get_scales(const onnx_parser::attribute_map& attr)
 {
-    std::vector<op_desc> operators() const { return {{"Resize"}, {"Upsample"}}; }
-
-    instruction_ref parse(const op_desc& opd,
-                          const onnx_parser& /*parser*/,
-                          onnx_parser::node_info info,
-                          std::vector<instruction_ref> args) const
-    {
-        // coord transform mode
-        std::string coord_trans_mode = get_coord_trans_mode(info.attributes);
-
-        // mode: only nearest and linear modes are supported for now
-        std::string mode = get_mode(info.attributes);
-
-        // nearest mode
-        std::string nearest_mode = get_nearest_mode(info.attributes);
-
-        // check exclude_outside, only support 0
-        if(contains(info.attributes, "exclude_outside") and
-           info.attributes.at("exclude_outside").i() == 1)
+    std::vector<double> scales;
+    if(contains(attr, "scales"))
     {
-            MIGRAPHX_THROW("PARSE_" + opd.op_name + ": exclude_outside 1 is not supported!");
+        copy(attr.at("scales").floats(), std::back_inserter(scales));
     }
 
-        // input data shape info
-        auto in_s    = args[0]->get_shape();
-        auto in_lens = in_s.lens();
-
-        // output shape is explicitly specified
-        std::vector<std::size_t> out_lens(in_lens.size());
-
-        // scale
-        std::vector<double> vec_scale;
+    return scales;
+}
 
+static void parse_args(const std::vector<instruction_ref>& args,
+                       const std::vector<size_t>& in_lens,
+                       const std::string& op_name,
+                       std::vector<double>& vec_scale,
+                       std::vector<std::size_t>& out_lens)
+{
     for(const auto& arg : args)
     {
         if(arg->name() == "undefined" or arg == args.front())
@@ -236,12 +218,12 @@ struct parse_resize : op_parser<parse_resize>
         {
             auto arg_out_s = arg->eval();
             check_arg_empty(arg_out_s,
-                                "PARSE_" + opd.op_name + ": dynamic output size is not supported!");
+                            "PARSE_" + op_name + ": dynamic output size is not supported!");
             arg_out_s.visit([&](const auto& ol) { out_lens.assign(ol.begin(), ol.end()); });
 
             if(out_lens.size() != in_lens.size())
             {
-                    MIGRAPHX_THROW("PARSE_" + opd.op_name +
+                MIGRAPHX_THROW("PARSE_" + op_name +
                                ": specified output size does not match input size");
             }
 
@@ -261,25 +243,71 @@ struct parse_resize : op_parser<parse_resize>
             {
                 auto arg_scale = arg->eval();
                 check_arg_empty(arg_scale,
-                                    "PARSE_" + opd.op_name +
-                                        ": dynamic input scale is not supported!");
+                                "PARSE_" + op_name + ": dynamic input scale is not supported!");
 
                 arg_scale.visit([&](const auto& v) { vec_scale.assign(v.begin(), v.end()); });
+            }
+        }
+    }
+}
+
+struct parse_resize : op_parser<parse_resize>
+{
+    std::vector<op_desc> operators() const { return {{"Resize"}, {"Upsample"}}; }
+
+    instruction_ref parse(const op_desc& opd,
+                          const onnx_parser& /*parser*/,
+                          onnx_parser::node_info info,
+                          std::vector<instruction_ref> args) const
+    {
+        // coord transform mode
+        std::string coord_trans_mode = get_coord_trans_mode(info.attributes);
+
+        // mode: only nearest and linear modes are supported for now
+        std::string mode = get_mode(info.attributes);
+
+        // nearest mode
+        std::string nearest_mode = get_nearest_mode(info.attributes);
+
+        // check exclude_outside, only support 0
+        if(contains(info.attributes, "exclude_outside") and
+           info.attributes.at("exclude_outside").i() == 1)
+        {
+            MIGRAPHX_THROW("PARSE_" + opd.op_name + ": exclude_outside 1 is not supported!");
+        }
+
+        // input data shape info
+        auto in_s    = args[0]->get_shape();
+        auto in_lens = in_s.lens();
+
+        // output shape is explicitly specified
+        std::vector<std::size_t> out_lens(in_lens.size());
+
+        // scale
+        std::vector<double> vec_scale = get_scales(info.attributes);
+
+        // If `scales` was not an attribute, it must be an input
+        if(vec_scale.empty())
+        {
+            // Depending on the args, it *must* populate the `vec_scale`, and might populate
+            // `out_lens`
+            parse_args(args, in_lens, opd.op_name, vec_scale, out_lens);
+        }
+
         if(in_lens.size() != vec_scale.size())
         {
-                        MIGRAPHX_THROW("PARSE_" + opd.op_name +
-                                       ": ranks of input and scale are different!");
+            MIGRAPHX_THROW("PARSE_" + opd.op_name + ": ranks of input and scale are different!");
         }
 
-                    std::transform(in_lens.begin(),
+        // if the output was not calculated yet, we update it based on the scales
+        if(all_of(out_lens.cbegin(), out_lens.cend(), [](auto o) { return o == 0; }))
+        {
+            std::transform(
+                in_lens.begin(),
                 in_lens.end(),
                 vec_scale.begin(),
                 out_lens.begin(),
-                                   [&](auto idx, auto scale) {
-                                       return static_cast<std::size_t>(idx * scale);
-                                   });
-                }
-            }
+                [&](auto idx, auto scale) { return static_cast<std::size_t>(idx * scale); });
         }
 
         shape out_s{in_s.type(), out_lens};
@@ -288,7 +316,6 @@ struct parse_resize : op_parser<parse_resize>
 
         // reshape input to one-dimension
         std::vector<int64_t> rsp_lens = {static_cast<int64_t>(in_s.elements())};
-        args[0]                       = info.make_contiguous(args[0]);
         auto rsp = info.add_instruction(make_op("reshape", {{"dims", rsp_lens}}), args[0]);
 
         if(mode == "nearest")