Add fp16 verify to driver (#988)

Allows --fp16 to be used in the driver to compare the target fp16 result and the ref fp32 result.

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);
 

src/include/migraphx/stringutils.hpp

@@ -168,7 +168,8 @@ inline std::string to_string_range(const std::initializer_list<T>& r)
 }
 
 template <class T>
-inline std::string to_string(const T& x)
+inline auto to_string(const T& x)
+    -> decltype((std::declval<std::stringstream>() << x), std::string{})
 {
     std::stringstream ss;
     ss << x;