manual merge

.github/workflows/ci.yaml

@@ -19,7 +19,7 @@ jobs:
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
     # It also restores the cache if it exists.
-    - uses: satackey/action-docker-layer-caching@v0.0.8
+    - uses: satackey/action-docker-layer-caching@v0.0.11
       # Ignore the failure of a step and avoid terminating the job.
       continue-on-error: true
 
@@ -65,7 +65,7 @@ jobs:
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
     # It also restores the cache if it exists.
-    - uses: satackey/action-docker-layer-caching@v0.0.8
+    - uses: satackey/action-docker-layer-caching@v0.0.11
       # Ignore the failure of a step and avoid terminating the job.
       continue-on-error: true
 
@@ -108,7 +108,7 @@ jobs:
     # In this step, this action saves a list of existing images,
     # the cache is created without them in the post run.
     # It also restores the cache if it exists.
-    - uses: satackey/action-docker-layer-caching@v0.0.8
+    - uses: satackey/action-docker-layer-caching@v0.0.11
       # Ignore the failure of a step and avoid terminating the job.
       continue-on-error: true
 

CMakeLists.txt

@@ -200,6 +200,7 @@ rocm_enable_cppcheck(
     RULE_FILE
         ${CMAKE_CURRENT_SOURCE_DIR}/cppcheck.rules
     SOURCES
+        examples/
         src/
         test/
     INCLUDE

cmake/TargetFlags.cmake

 
+function(eval_and_strip_genex OUTPUT_VAR INPUT)
+    string(REPLACE "$<LINK_LANGUAGE:CXX>" "1" INPUT "${INPUT}")
+    string(REPLACE "$<COMPILE_LANGUAGE:CXX>" "1" INPUT "${INPUT}")
+    string(REPLACE "SHELL:" "" INPUT "${INPUT}")
+    string(REPLACE "$<BOOL:>" "0" INPUT "${INPUT}")
+    string(REGEX REPLACE "\\$<BOOL:(0|FALSE|false|OFF|off|N|n|IGNORE|ignore|NOTFOUND|notfound)>" "0" INPUT "${INPUT}")
+    string(REGEX REPLACE "\\$<BOOL:[^<>]*-NOTFOUND>" "0" INPUT "${INPUT}")
+    string(REGEX REPLACE "\\$<BOOL:[^$<>]*>" "1" INPUT "${INPUT}")
+    string(REPLACE "$<NOT:0>" "1" INPUT "${INPUT}")
+    string(REPLACE "$<NOT:1>" "0" INPUT "${INPUT}")
+    string(REGEX REPLACE "\\$<0:[^<>]*>" "" INPUT "${INPUT}")
+    string(REGEX REPLACE "\\$<1:([^<>]*)>" "\\1" INPUT "${INPUT}")
+    string(GENEX_STRIP "${INPUT}" INPUT)
+    set(${OUTPUT_VAR} "${INPUT}" PARENT_SCOPE)
+endfunction()
+
 function(get_target_property2 VAR TARGET PROPERTY)
     get_target_property(_pflags ${TARGET} ${PROPERTY})
     if(_pflags)
+        eval_and_strip_genex(_pflags "${_pflags}")
         set(${VAR} ${_pflags} PARENT_SCOPE)
     else()
         set(${VAR} "" PARENT_SCOPE)
     endif()
 endfunction()
 
+function(flags_requires_arg OUTPUT_VAR FLAG)
+    set(_args -x -isystem)
+    if(FLAG IN_LIST _args)
+        set(${OUTPUT_VAR} 1 PARENT_SCOPE)
+    else()
+        set(${OUTPUT_VAR} 0 PARENT_SCOPE)
+    endif()
+endfunction()
 
 macro(append_flags FLAGS TARGET PROPERTY PREFIX)
     get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    set(_requires_arg 0)
     foreach(FLAG ${_pflags})
-        if(TARGET ${FLAG})
-            target_flags(_pflags2 ${FLAG})
-            string(APPEND ${FLAGS} " ${_pflags2}")
-        else()
-            string(APPEND ${FLAGS} " ${PREFIX}${FLAG}")
+        string(STRIP "${FLAG}" FLAG)
+        if(FLAG)
+            if(TARGET ${FLAG} AND NOT _requires_arg)
+                target_flags(_pflags2 ${FLAG})
+                string(APPEND ${FLAGS} " ${_pflags2}")
+            else()
+                string(APPEND ${FLAGS} " ${PREFIX}${FLAG}")
+            endif()
+            flags_requires_arg(_requires_arg "${FLAG}")
         endif()
     endforeach()
 endmacro()
 
 macro(append_link_flags FLAGS TARGET PROPERTY)
     get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    set(_requires_arg 0)
     foreach(FLAG ${_pflags})
-        if(TARGET ${FLAG})
-            target_flags(_pflags2 ${FLAG})
-            string(APPEND ${FLAGS} " ${_pflags2}")
-        elseif(FLAG MATCHES "^-.*")
-            string(APPEND ${FLAGS} " ${FLAG}")
-        elseif(EXISTS ${FLAG})
-            string(APPEND ${FLAGS} " ${FLAG}")
-        else()
-            string(APPEND ${FLAGS} " -l${FLAG}")
+        string(STRIP "${FLAG}" FLAG)
+        if(FLAG)
+            if(TARGET ${FLAG} AND NOT _requires_arg)
+                target_flags(_pflags2 ${FLAG})
+                string(APPEND ${FLAGS} " ${_pflags2}")
+            elseif(FLAG MATCHES "^-.*")
+                string(APPEND ${FLAGS} " ${FLAG}")
+            elseif(EXISTS ${FLAG})
+                string(APPEND ${FLAGS} " ${FLAG}")
+            else()
+                string(APPEND ${FLAGS} " -l${FLAG}")
+            endif()
+            flags_requires_arg(_requires_arg "${FLAG}")
         endif()
     endforeach()
 endmacro()

doc/src/conf.py

@@ -32,8 +32,10 @@ import re
 # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
 # ones.
 extensions = [
-    'breathe', 'sphinx.ext.mathjax', 'sphinx.ext.viewcode', 'sphinx_rtd_theme'
+    'breathe', 'sphinx.ext.mathjax', 'sphinx.ext.viewcode', 'sphinx_rtd_theme',
+    'sphinx.ext.autosectionlabel'
 ]
+autosectionlabel_prefix_document = True
 
 # Add any paths that contain templates here, relative to this directory.
 templates_path = ['_templates']

doc/src/dev/tools.rst

+Tools
+=====
+
+roctx.py
+--------
+MIGraphX driver provides `roctx` command which can be used with `rocprof` binary to get marker timing information for each MIGraphX operator.  
+In order to help user to process timing information, rocTX helper script is provided at `tools/roctx.py`.
+The `roctx.py` helper script provides two main functionality: `run` and `parse`. Available knobs and usage are given below:
+
+::
+
+    Usage: roctx.py [-h] [--json-path json_path] [--out out]
+    [--study-name study-name] [--repeat repeat] [--parse]
+    [--run run] [--debug]
+
+.. option::  --run
+
+Runs `migraphx-driver roctx` command and given `migraphx-driver` knobs, and then parses the results, providing GPU kernel timing information.
+MIGraphX knobs can be given via a string to `--run` knob. Please see the examples below.
+
+.. option::  --parse
+
+Given `--json-path`, parses JSON file and provides GPU kernel timing information.
+
+.. option::  --out
+
+Output folder
+
+.. option::  --study-name
+
+Optional. Allows user to name a study for easier interpretation. Defaults to timestamp.
+
+.. option::  --repeat
+
+Number of iterations. Set to **2** by default.
+
+.. option::  --debug
+
+Provides additional debug information related to data. Only use for debugging purposes.
+
+**Examples:**
+
+**Running inference with rocTX for a given ONNX file:**
+::
+    python roctx.py --run '--onnx --gpu fcn-resnet50-11.onnx' --out output_folder --repeat 5
+
+After a run, similar to output given below is expected at terminal. The output will provide `SUM`, `MIN`, `MAX` and `COUNT` information for each kernel executed for a given model.
+Average total time is also provided. There are three files provided for reference:
+
+1. `OUTPUT CSV FILE` provides a summary of the run, providing utilized MIGraphX knobs and related kernel timing information
+2. `KERNEL TIMING DETAILS` provides the hotspot kernel timing information
+3. This will provide all output data related to all iterations executed during a run.
+
+An example output:
+
+.. image:: ./roctx1.jpg
+
+Hotspot kerel timing information:
+
+.. image:: ./roctx2.jpg
+
+**Parsing an already existing JSON file:**
+::
+
+    python roctx.py --parse --json-path ../trace.json
\ No newline at end of file

doc/src/developer_guide.rst

@@ -13,3 +13,4 @@ Developer Guide
    dev/quantization
    dev/pass
    dev/matchers
+   dev/tools

doc/src/driver.rst

@@ -61,3 +61,21 @@ Verify each instruction
 .. option::  -r, --reduce
 
 Reduce program and verify
+
+roctx
+----
+
+.. program:: migraphx-driver roctx
+
+Provides marker information for each operation, allowing MIGraphX to be used with `rocprof <https://rocmdocs.amd.com/en/latest/ROCm_Tools/ROCm-Tools.html>`_ for performance analysis.
+This allows user to get GPU-level kernel timing information.
+An example command line combined with rocprof for tracing purposes is given below:
+
+.. code-block:: bash
+
+    /opt/rocm/bin/rocprof --hip-trace --roctx-trace --flush-rate 1ms --timestamp on -d <OUTPUT_PATH> --obj-tracking on /opt/rocm/bin/migraphx-driver roctx <ONNX_FILE> <MIGRAPHX_OPTIONS>
+
+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

examples/migraphx/cpp_parse_load_save/parse_load_save.cpp

@@ -24,7 +24,6 @@ int main(int argc, char** argv)
         return 0;
     }
 
-    char* parse_arg        = getCmdOption(argv + 2, argv + argc, "--parse");
     char* load_arg         = getCmdOption(argv + 2, argv + argc, "--load");
     char* save_arg         = getCmdOption(argv + 2, argv + argc, "--save");
     const char* input_file = argv[1];

examples/nlp/python_bert_squad/requirements_bertsquad.txt

-tensorflow==2.5.1
+tensorflow==2.5.2
 onnxruntime
 tokenizers
\ No newline at end of file

examples/vision/python_3dunet/3dunet_inference.ipynb

@@ -10,6 +10,16 @@
     "https://github.com/naomifridman/Unet_Brain_tumor_segmentation"
    ]
   },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "09ceec31",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "!pip install SimpleITK matplotlib scikit-image"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,

examples/vision/python_nfnet/nfnet_inference.ipynb

@@ -17,7 +17,9 @@
     "- How to optimize NFNet ONNX model with AMD MIGraphX.\n",
     "- How to run inference on AMD GPU with the optimized ONNX model.\n",
     "\n",
-    "The NFNet utilized in this example is the smallest NFNet version, F0: 71.5M parameters (83.6% top-1 accuracy on ImageNet)"
+    "The NFNet utilized in this example is the smallest NFNet version, F0: 71.5M parameters (83.6% top-1 accuracy on ImageNet)\n",
+    "\n",
+    "Please make sure MIGraphX Python API is installed following the instructions at Github page."
    ]
   },
   {
@@ -107,7 +109,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "with open('../python_api_inference/imagenet_simple_labels.json') as json_data:\n",
+    "with open('../python_resnet50/imagenet_simple_labels.json') as json_data:\n",
     "    labels = json.load(json_data)"
    ]
   },

examples/vision/python_nfnet/requirements_nfnet.txt

 opencv-python
-onnxruntime
+onnxruntime
\ No newline at end of file
+image
\ No newline at end of file

src/cpp_generator.cpp

 #include <migraphx/cpp_generator.hpp>
 #include <migraphx/module.hpp>
 #include <migraphx/operation.hpp>
+#include <migraphx/ranges.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/builtin.hpp>
 #include <migraphx/stringutils.hpp>
@@ -51,6 +52,7 @@ cpp_generator::function& cpp_generator::function::set_types(const module& m)
 cpp_generator::function&
 cpp_generator::function::set_types(const module& m, const std::function<std::string(shape)>& parse)
 {
+    this->params.clear();
     auto pmap = m.get_parameter_shapes();
     std::map<std::string, shape> input_map(pmap.begin(), pmap.end());
     std::transform(
@@ -63,11 +65,30 @@ cpp_generator::function::set_types(const module& m, const std::function<std::str
     return *this;
 }
 
+cpp_generator::function& cpp_generator::function::set_generic_types(const module& m)
+{
+    this->params.clear();
+    auto pmap = m.get_parameter_shapes();
+    std::map<std::string, shape> input_map(pmap.begin(), pmap.end());
+    std::transform(
+        input_map.begin(), input_map.end(), std::back_inserter(this->params), [&](auto&& p) {
+            return param{p.first, "T" + p.first};
+        });
+
+    std::transform(input_map.begin(),
+                   input_map.end(),
+                   std::back_inserter(this->tparams),
+                   [&](auto&& p) { return "class T" + p.first; });
+    this->return_type = "auto";
+    return *this;
+}
+
 struct cpp_generator_impl
 {
     std::stringstream fs{};
-    std::size_t function_count                   = 0;
-    std::function<std::string(std::string)> fmap = nullptr;
+    std::size_t function_count                                = 0;
+    std::function<std::string(std::string)> fmap              = nullptr;
+    std::unordered_map<std::string, std::string> point_op_map = {};
 };
 cpp_generator::cpp_generator() : impl(std::make_unique<cpp_generator_impl>()) {}
 
@@ -83,41 +104,54 @@ cpp_generator::~cpp_generator() noexcept = default;
 
 void cpp_generator::fmap(const std::function<std::string(std::string)>& f) { impl->fmap = f; }
 
+void cpp_generator::add_point_op(const std::string& op_name, const std::string& code)
+{
+    impl->point_op_map[op_name] = code;
+}
+
 std::string cpp_generator::generate_point_op(const operation& op,
                                              const std::vector<std::string>& args)
 {
-    auto v          = op.to_value();
-    auto attributes = op.attributes();
-    if(not attributes.contains("point_op"))
-        MIGRAPHX_THROW("op is missing point_op attribute: " + op.name());
-    return interpolate_string(attributes["point_op"].to<std::string>(),
-                              [&](auto start, auto last) -> std::string {
-                                  auto key = trim({start, last});
-                                  if(key.empty())
-                                      MIGRAPHX_THROW("Empty parameter");
-                                  std::string fselector = "function:";
-                                  if(starts_with(key, fselector))
-                                  {
-                                      auto fname = key.substr(fselector.size());
-                                      if(impl->fmap == nullptr)
-                                          return fname;
-                                      else
-                                          return impl->fmap(fname);
-                                  }
-                                  else if(with_char(::isdigit)(key[0]))
-                                  {
-                                      auto i = std::stoul(key);
-                                      return args.at(i);
-                                  }
-                                  else if(v.contains(key))
-                                  {
-                                      return v[key].template to<std::string>();
-                                  }
-                                  else
-                                  {
-                                      return key;
-                                  }
-                              });
+    auto v = op.to_value();
+    std::string code;
+    if(contains(impl->point_op_map, op.name()))
+    {
+        code = impl->point_op_map.at(op.name());
+    }
+    else
+    {
+        auto attributes = op.attributes();
+        if(not attributes.contains("point_op"))
+            MIGRAPHX_THROW("op is missing point_op attribute: " + op.name());
+        code = attributes["point_op"].to<std::string>();
+    }
+    return interpolate_string(code, [&](auto start, auto last) -> std::string {
+        auto key = trim({start, last});
+        if(key.empty())
+            MIGRAPHX_THROW("Empty parameter");
+        std::string fselector = "function:";
+        if(starts_with(key, fselector))
+        {
+            auto fname = key.substr(fselector.size());
+            if(impl->fmap == nullptr)
+                return fname;
+            else
+                return impl->fmap(fname);
+        }
+        else if(with_char(::isdigit)(key[0]))
+        {
+            auto i = std::stoul(key);
+            return args.at(i);
+        }
+        else if(v.contains(key))
+        {
+            return v[key].template to<std::string>();
+        }
+        else
+        {
+            return key;
+        }
+    });
 }
 
 std::string cpp_generator::str() const { return impl->fs.str(); }
@@ -148,6 +182,8 @@ cpp_generator::function cpp_generator::generate_module(const module& m)
 std::string cpp_generator::create_function(const cpp_generator::function& f)
 {
     impl->function_count++;
+    if(not f.tparams.empty())
+        impl->fs << "template<" << join_strings(f.tparams, ", ") << ">\n";
     std::string name = f.name.empty() ? "f" + std::to_string(impl->function_count) : f.name;
     impl->fs << join_strings(f.attributes, " ") << " " << f.return_type << " " << name;
     char delim = '(';

src/driver/argument_parser.hpp

@@ -17,6 +17,7 @@
 #include <migraphx/type_name.hpp>
 #include <migraphx/functional.hpp>
 #include <migraphx/stringutils.hpp>
+#include <migraphx/rank.hpp>
 
 namespace migraphx {
 namespace driver {
@@ -106,10 +107,22 @@ struct argument_parser
         return to_string_range(x);
     }
 
+    template <class T>
+    auto as_string_value(rank<1>, const T& x) -> decltype(to_string(x))
+    {
+        return to_string(x);
+    }
+
+    template <class T>
+    std::string as_string_value(rank<0>, const T&)
+    {
+        throw std::runtime_error("Can't convert to string");
+    }
+
     template <class T, MIGRAPHX_REQUIRES(not is_multi_value<T>{})>
     std::string as_string_value(const T& x)
     {
-        return to_string(x);
+        return as_string_value(rank<1>{}, x);
     }
 
     template <class T, class... Fs>
@@ -122,10 +135,11 @@ struct argument_parser
                                  return false;
                              }});
 
-        argument& arg     = arguments.back();
-        arg.type          = migraphx::get_type_name<T>();
-        arg.default_value = as_string_value(x);
+        argument& arg = arguments.back();
+        arg.type      = migraphx::get_type_name<T>();
         migraphx::each_args([&](auto f) { f(x, arg); }, fs...);
+        if(not arg.default_value.empty() and arg.nargs > 0)
+            arg.default_value = as_string_value(x);
     }
 
     template <class... Fs>

src/driver/main.cpp

+#include "verify.hpp"
 #include "argument_parser.hpp"
 #include "command.hpp"
-#include "verify.hpp"
+#include "precision.hpp"
 #include "perf.hpp"
 #include "models.hpp"
 #include "marker_roctx.hpp"
@@ -288,14 +289,12 @@ struct compiler_target
 
 struct compiler
 {
-    static const int q_fp16 = 1;
-    static const int q_int8 = 2;
     loader l;
     program_params parameters;
     compiler_target ct;
-    bool offload_copy = false;
-    bool fast_math    = true;
-    int quantize      = 0;
+    bool offload_copy  = false;
+    bool fast_math     = true;
+    precision quantize = precision::fp32;
 
     std::vector<std::string> fill0;
     std::vector<std::string> fill1;
@@ -312,8 +311,8 @@ struct compiler
            {"--disable-fast-math"},
            ap.help("Disable fast math optimization"),
            ap.set_value(false));
-        ap(quantize, {"--fp16"}, ap.help("Quantize for fp16"), ap.set_value(q_fp16));
-        ap(quantize, {"--int8"}, ap.help("Quantize for int8"), ap.set_value(q_int8));
+        ap(quantize, {"--fp16"}, ap.help("Quantize for fp16"), ap.set_value(precision::fp16));
+        ap(quantize, {"--int8"}, ap.help("Quantize for int8"), ap.set_value(precision::int8));
     }
 
     auto params(const program& p) { return parameters.generate(p, ct.get_target(), offload_copy); }
@@ -325,11 +324,11 @@ struct compiler
         if(p.is_compiled())
             return p;
         auto t = ct.get_target();
-        if(quantize == q_fp16)
+        if(quantize == precision::fp16)
         {
             quantize_fp16(p);
         }
-        else if(quantize == q_int8)
+        else if(quantize == precision::int8)
         {
             quantize_int8(p, t, {params(p)});
         }
@@ -377,6 +376,7 @@ struct verify : command<verify>
     bool reduce          = false;
     bool offload_copy    = false;
     bool fast_math       = true;
+    precision quantize   = precision::fp32;
     void parse(argument_parser& ap)
     {
         l.parse(ap);
@@ -396,6 +396,7 @@ struct verify : command<verify>
            ap.help("Verify each instruction"),
            ap.set_value(true));
         ap(reduce, {"-r", "--reduce"}, ap.help("Reduce program and verify"), ap.set_value(true));
+        ap(quantize, {"--fp16"}, ap.help("Quantize for fp16"), ap.set_value(precision::fp16));
     }
 
     void run()
@@ -412,15 +413,15 @@ struct verify : command<verify>
 
         if(per_instruction)
         {
-            verify_instructions(p, t, options, tolerance);
+            verify_instructions(p, t, options, quantize, tolerance);
         }
         else if(reduce)
         {
-            verify_reduced_program(p, t, options, m, tolerance);
+            verify_reduced_program(p, t, options, quantize, m, tolerance);
         }
         else
         {
-            verify_program(l.file, p, t, options, m, tolerance);
+            verify_program(l.file, p, t, options, quantize, m, tolerance);
         }
     }
 };

src/driver/precision.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_PRECISION_HPP
+#define MIGRAPHX_GUARD_RTGLIB_PRECISION_HPP
+
+namespace migraphx {
+namespace driver {
+inline namespace MIGRAPHX_INLINE_NS {
+
+enum class precision
+{
+    fp32,
+    fp16,
+    int8
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace driver
+} // namespace migraphx
+
+#endif

src/driver/verify.cpp

@@ -6,6 +6,7 @@
 #include <migraphx/verify_args.hpp>
 #include <migraphx/instruction.hpp>
 #include <migraphx/compile_options.hpp>
+#include <migraphx/quantization.hpp>
 
 namespace migraphx {
 namespace driver {
@@ -19,9 +20,16 @@ std::vector<argument> run_ref(program p, const parameter_map& inputs)
     return out;
 }
 
-std::vector<argument>
-run_target(program p, const target& t, const compile_options& options, const parameter_map& inputs)
+std::vector<argument> run_target(program p,
+                                 const target& t,
+                                 const compile_options& options,
+                                 precision quantize,
+                                 const parameter_map& inputs)
 {
+    if(quantize == precision::fp16)
+    {
+        quantize_fp16(p);
+    }
     p.compile(t, options);
 
     parameter_map m;
@@ -43,24 +51,24 @@ void verify_program(const std::string& name,
                     const program& p,
                     const target& t,
                     compile_options options,
+                    precision quantize,
                     const parameter_map& inputs,
                     double tolerance)
 {
     auto x = run_ref(p, inputs);
-    auto y = run_target(p, t, options, inputs);
+    auto y = run_target(p, t, options, quantize, inputs);
 
     std::size_t output_num = x.size();
     for(std::size_t i = 0; i < output_num; ++i)
     {
         verify_args(name, x[i], y[i], tolerance);
     }
-    // std::cout << "cpu: " << x << std::endl;
-    // std::cout << "gpu: " << y << std::endl;
 }
 
 void verify_instructions(const program& prog,
                          const target& t,
                          compile_options options,
+                         precision quantize,
                          double tolerance)
 {
     const auto* mm_prog = prog.get_main_module();
@@ -92,7 +100,8 @@ void verify_instructions(const program& prog,
         {
             std::cout << "Verify: " << ins.name() << std::endl;
             std::cout << p << std::endl;
-            verify_program(ins.name(), p, t, options, create_param_map(p, false), tolerance);
+            verify_program(
+                ins.name(), p, t, options, quantize, create_param_map(p, false), tolerance);
         }
         catch(...)
         {
@@ -106,6 +115,7 @@ void verify_reduced(program p,
                     int n,
                     const target& t,
                     compile_options options,
+                    precision quantize,
                     const parameter_map& inputs,
                     double tolerance)
 {
@@ -114,12 +124,13 @@ void verify_reduced(program p,
     mm->remove_instructions(last, mm->end());
     std::cout << "Verify: " << std::endl;
     std::cout << p << std::endl;
-    verify_program(std::to_string(n), p, t, options, inputs, tolerance);
+    verify_program(std::to_string(n), p, t, options, quantize, inputs, tolerance);
 }
 
 void verify_reduced_program(const program& p,
                             const target& t,
                             compile_options options,
+                            precision quantize,
                             const parameter_map& inputs,
                             double tolerance)
 {
@@ -127,7 +138,7 @@ void verify_reduced_program(const program& p,
     auto n         = std::distance(mm->begin(), mm->end());
     for(std::size_t i = 0; i < n; i++)
     {
-        verify_reduced(p, i, t, options, inputs, tolerance);
+        verify_reduced(p, i, t, options, quantize, inputs, tolerance);
     }
 }
 

src/driver/verify.hpp

 #ifndef MIGRAPHX_GUARD_RTGLIB_DRIVER_VERIFY_HPP
 #define MIGRAPHX_GUARD_RTGLIB_DRIVER_VERIFY_HPP
 
+#include "precision.hpp"
 #include <migraphx/program.hpp>
 
 namespace migraphx {
@@ -11,15 +12,18 @@ void verify_program(const std::string& name,
                     const program& p,
                     const target& t,
                     compile_options options     = compile_options{},
+                    precision quantize          = precision::fp32,
                     const parameter_map& inputs = {},
                     double tolerance            = 100);
 void verify_instructions(const program& prog,
                          const target& t,
                          compile_options options = compile_options{},
+                         precision quantize      = precision::fp32,
                          double tolerance        = 80);
 void verify_reduced_program(const program& p,
                             const target& t,
                             compile_options options     = compile_options{},
+                            precision quantize          = precision::fp32,
                             const parameter_map& inputs = {},
                             double tolerance            = 80);