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/compile.rst
\ No newline at end of file
+.. 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/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())
         {
+            // 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();
+               }))
+            {
+                return vec;
+            }
             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())
-                {
-                    MIGRAPHX_THROW(
-                        "NORMALIZE_ATTR: 'use_lens' on a non-fixed dynamic dimension, axis=" +
-                        std::to_string(i));
-                }
-                return dd.max;
+                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

@@ -181,6 +181,76 @@ static std::string get_nearest_mode(const onnx_parser::attribute_map& attr)
     return nearest_mode;
 }
 
+static std::vector<double> get_scales(const onnx_parser::attribute_map& attr)
+{
+    std::vector<double> scales;
+    if(contains(attr, "scales"))
+    {
+        copy(attr.at("scales").floats(), std::back_inserter(scales));
+    }
+
+    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())
+        {
+            continue;
+        }
+
+        // skipped empty input
+        auto lens = arg->get_shape().lens();
+        if(lens.empty())
+        {
+            continue;
+        }
+
+        auto type = arg->get_shape().type();
+        // output size
+        if(type == shape::int64_type)
+        {
+            auto arg_out_s = arg->eval();
+            check_arg_empty(arg_out_s,
+                            "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_" + op_name +
+                               ": specified output size does not match input size");
+            }
+
+            // compute the scale
+            vec_scale.resize(in_lens.size());
+            std::transform(in_lens.begin(),
+                           in_lens.end(),
+                           out_lens.begin(),
+                           vec_scale.begin(),
+                           [](auto iss, auto oss) { return 1.0 * oss / iss; });
+        }
+        else
+        {
+
+            // scale input
+            if(lens[0] == in_lens.size())
+            {
+                auto arg_scale = arg->eval();
+                check_arg_empty(arg_scale,
+                                "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"}}; }
@@ -214,72 +284,30 @@ struct parse_resize : op_parser<parse_resize>
         std::vector<std::size_t> out_lens(in_lens.size());
 
         // scale
-        std::vector<double> vec_scale;
+        std::vector<double> vec_scale = get_scales(info.attributes);
 
-        for(const auto& arg : args)
+        // If `scales` was not an attribute, it must be an input
+        if(vec_scale.empty())
         {
-            if(arg->name() == "undefined" or arg == args.front())
-            {
-                continue;
-            }
-
-            // skipped empty input
-            auto lens = arg->get_shape().lens();
-            if(lens.empty())
-            {
-                continue;
-            }
-
-            auto type = arg->get_shape().type();
-            // output size
-            if(type == shape::int64_type)
-            {
-                auto arg_out_s = arg->eval();
-                check_arg_empty(arg_out_s,
-                                "PARSE_" + opd.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 +
-                                   ": specified output size does not match input size");
-                }
+            // 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);
+        }
 
-                // compute the scale
-                vec_scale.resize(in_lens.size());
-                std::transform(in_lens.begin(),
-                               in_lens.end(),
-                               out_lens.begin(),
-                               vec_scale.begin(),
-                               [](auto iss, auto oss) { return 1.0 * oss / iss; });
-            }
-            else
-            {
+        if(in_lens.size() != vec_scale.size())
+        {
+            MIGRAPHX_THROW("PARSE_" + opd.op_name + ": ranks of input and scale are different!");
+        }
 
-                // scale input
-                if(lens[0] == in_lens.size())
-                {
-                    auto arg_scale = arg->eval();
-                    check_arg_empty(arg_scale,
-                                    "PARSE_" + opd.op_name +
-                                        ": dynamic input scale is not supported!");
-
-                    arg_scale.visit([&](const auto& v) { vec_scale.assign(v.begin(), v.end()); });
-                    if(in_lens.size() != vec_scale.size())
-                    {
-                        MIGRAPHX_THROW("PARSE_" + opd.op_name +
-                                       ": ranks of input and scale are different!");
-                    }
-
-                    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);
-                                   });
-                }
-            }
+        // 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); });
         }
 
         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")