Merge branch 'develop' into ci_cdna_to_mi100

docs/.sphinx/requirements.txt

@@ -21,7 +21,7 @@ charset-normalizer==3.1.0
     # via requests
 click==8.1.3
     # via sphinx-external-toc
-cryptography==41.0.3
+cryptography==41.0.4
     # via pyjwt
 deprecated==1.2.13
     # via pygithub
@@ -87,7 +87,7 @@ requests==2.28.2
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core>=0.20.0
+rocm-docs-core==0.24.2
     # via -r requirements.in
 smmap==5.0.0
     # via gitdb

docs/dev_intro.rst

@@ -131,7 +131,7 @@ In this case, we can create `argument <migraphx::argument>` objects directly fro
     std::vector<float> results_vector(64);
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
 
-    EXPECT(migraphx::verify::verify_range(results_vector, sol));
+    EXPECT(migraphx::verify::verify_rms_range(results_vector, sol));
 
 An `argument <migraphx::argument>` can handle memory buffers from either the GPU or the CPU.
 By default when running the `program <migraphx::program>`, buffers are allocated on the corresponding target.

docs/driver.rst

@@ -50,9 +50,17 @@ Runs reference and CPU or GPU implementations and checks outputs for consistency
 
 .. include:: ./driver/compile.rst
 
-.. option::  --tolerance [double]
+.. option::  --rms-tol [double]
 
-Tolerance for errors (Default: 80)
+Tolerance for RMS error (Default: 0.001)
+
+.. option::  --atol [double]
+
+Tolerance for elementwise absolute difference (Default: 0.001)
+
+.. option::  --rtol [double]
+
+Tolerance for elementwise relative difference (Default: 0.001)
 
 .. option::  -i, --per-instruction
 

examples/migraphx/migraphx_driver/README.md

@@ -55,7 +55,9 @@ See below for a comprehensive list of commands and option arguments, as well as
 | --exhaustive-tune                        | Enable exhaustive search to find fastest kernel           |
 | --fp16                                   | Quantize for fp16                                         |
 | --int8                                   | Quantize for int8                                         |
-| --tolerance                              | Tolerance for errors                                      |
+| --rms-tol                                | Tolerance for the RMS error (Default: 0.001)              |
+| --atol                                   | Tolerance for elementwise absolute difference (Default: 0.001) |
+| --rtol                                   | Tolerance for elementwise relative difference (Default: 0.001) |
 | --per-instruction \| -i                  | Verify each instruction                                   |
 | --reduce \| -r                           | Reduce program and verify                                 |
 | --iterations \| -n                       | Number of iterations to run for perf report               |

src/driver/main.cpp

@@ -536,13 +536,19 @@ struct params : command<params>
 struct verify : command<verify>
 {
     compiler c;
-    double tolerance     = 80;
+    migraphx::verify::tolerance tols;
     bool per_instruction = false;
     bool reduce          = false;
     void parse(argument_parser& ap)
     {
         c.parse(ap);
-        ap(tolerance, {"--tolerance"}, ap.help("Tolerance for errors"));
+        ap(tols.rms_tol, {"--rms-tol"}, ap.help("Tolerance for the RMS error (Default: 0.001)"));
+        ap(tols.atol,
+           {"--atol"},
+           ap.help("Tolerance for the elementwise absolute difference (Default: 0.001)"));
+        ap(tols.rtol,
+           {"--rtol"},
+           ap.help("Tolerance for the elementwise relative difference (Default: 0.001)"));
         ap(per_instruction,
            {"-i", "--per-instruction"},
            ap.help("Verify each instruction"),
@@ -567,15 +573,15 @@ struct verify : command<verify>
 
         if(per_instruction)
         {
-            verify_instructions(p, t, c.co, quantize, tolerance);
+            verify_instructions(p, t, c.co, quantize, tols);
         }
         else if(reduce)
         {
-            verify_reduced_program(p, t, c.co, quantize, m, tolerance);
+            verify_reduced_program(p, t, c.co, quantize, m, tols);
         }
         else
         {
-            verify_program(c.l.file, p, t, c.co, quantize, m, tolerance);
+            verify_program(c.l.file, p, t, c.co, quantize, m, tols);
         }
     }
 };

src/driver/verify.cpp

@@ -77,24 +77,24 @@ void verify_program(const std::string& name,
                     compile_options options,
                     precision quantize,
                     const parameter_map& inputs,
-                    double tolerance)
+                    verify::tolerance tols)
 {
-    auto x = run_ref(p, inputs);
-    auto y = run_target(p, t, options, quantize, inputs);
+    auto ref_outs    = run_ref(p, inputs);
+    auto target_outs = run_target(p, t, options, quantize, inputs);
 
-    std::size_t output_num = x.size();
+    std::size_t output_num = ref_outs.size();
     for(std::size_t i = 0; i < output_num; ++i)
     {
-        if(x[i].get_shape().type() != y[i].get_shape().type() or
-           x[i].get_shape().lens() != y[i].get_shape().lens())
+        if(ref_outs[i].get_shape().type() != target_outs[i].get_shape().type() or
+           ref_outs[i].get_shape().lens() != target_outs[i].get_shape().lens())
         {
             std::cout << "FAILED: " << name << std::endl;
-            std::cout << "Shape mismatch {" << x[i].get_shape() << "} != {" << y[i].get_shape()
-                      << "}" << std::endl;
+            std::cout << "Shape mismatch {" << ref_outs[i].get_shape() << "} != {"
+                      << target_outs[i].get_shape() << "}" << std::endl;
         }
         else
         {
-            verify_args(name, x[i], y[i], tolerance);
+            verify_args(name, target_outs[i], verify::expected{ref_outs[i]}, tols);
         }
     }
 }
@@ -103,7 +103,7 @@ void verify_instructions(const program& prog,
                          const target& t,
                          compile_options options,
                          precision quantize,
-                         double tolerance)
+                         verify::tolerance tols)
 {
     const auto* mm_prog = prog.get_main_module();
     for(auto&& ins : (*mm_prog))
@@ -134,8 +134,7 @@ 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, quantize, create_param_map(p, false), tolerance);
+            verify_program(ins.name(), p, t, options, quantize, create_param_map(p, false), tols);
         }
         catch(...)
         {
@@ -151,7 +150,7 @@ void verify_reduced(program p,
                     compile_options options,
                     precision quantize,
                     const parameter_map& inputs,
-                    double tolerance)
+                    verify::tolerance tols)
 {
     auto* mm  = p.get_main_module();
     auto last = std::prev(mm->end(), n);
@@ -160,7 +159,7 @@ void verify_reduced(program p,
     std::cout << p << std::endl;
     try
     {
-        verify_program(std::to_string(n), p, t, options, quantize, inputs, tolerance);
+        verify_program(std::to_string(n), p, t, options, quantize, inputs, tols);
     }
     catch(const std::exception& e)
     {
@@ -174,7 +173,7 @@ void verify_reduced_program(const program& p,
                             compile_options options,
                             precision quantize,
                             const parameter_map& inputs,
-                            double tolerance)
+                            verify::tolerance tols)
 {
     const auto* mm = p.get_main_module();
     auto n         = std::distance(mm->begin(), mm->end());
@@ -187,7 +186,7 @@ void verify_reduced_program(const program& p,
             std::cout << "Skip: " << i << std::endl;
             continue;
         }
-        verify_reduced(p, i, t, options, quantize, inputs, tolerance);
+        verify_reduced(p, i, t, options, quantize, inputs, tols);
     }
 }
 

src/driver/verify.hpp

@@ -26,6 +26,7 @@
 
 #include "precision.hpp"
 #include <migraphx/program.hpp>
+#include <migraphx/verify.hpp>
 
 namespace migraphx {
 namespace driver {
@@ -37,18 +38,18 @@ void verify_program(const std::string& name,
                     compile_options options     = compile_options{},
                     precision quantize          = precision::fp32,
                     const parameter_map& inputs = {},
-                    double tolerance            = 100);
+                    verify::tolerance tols      = verify::tolerance{});
 void verify_instructions(const program& prog,
                          const target& t,
                          compile_options options = compile_options{},
                          precision quantize      = precision::fp32,
-                         double tolerance        = 80);
+                         verify::tolerance tols  = verify::tolerance{});
 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);
+                            verify::tolerance tols      = verify::tolerance{});
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace driver

src/include/migraphx/verify.hpp

@@ -29,10 +29,13 @@
 #include <functional>
 #include <iostream>
 #include <numeric>
+#include <assert.h>
 
 #include <migraphx/float_equal.hpp>
 #include <migraphx/config.hpp>
+#include <migraphx/env.hpp>
 
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_VERIFY_ENABLE_ALLCLOSE)
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 namespace verify {
@@ -187,16 +190,103 @@ double rms_range(const R1& r1, const R2& r2)
         return std::numeric_limits<range_value<R1>>::max();
 }
 
+template <class R>
+double get_rms_tol(const R&, std::size_t tolerance = 80)
+{
+    double threshold = std::numeric_limits<range_value<R>>::epsilon() * tolerance;
+    return threshold;
+}
+
+/*
+C++ doesn't support named arguments, this is just wrapper that helps distinguish between actual
+results v/s expected results arguments.
+*/
+template <class T>
+struct expected
+{
+    expected() = default;
+    explicit expected(const T& input) : x(&input) {}
+    const T& data() const
+    {
+        assert(x != nullptr);
+        return *x;
+    }
+
+    private:
+    const T* x = nullptr;
+};
+
+// deduction guide for templated expected class
+template <class T>
+expected(const T&) -> expected<T>;
+
+struct tolerance
+{
+    double rms_tol = 0.001;
+    double atol    = 0.001;
+    double rtol    = 0.001;
+};
+
+/*
+MIGraphX implementation of numpy's np.allclose() which checks if elementwise absolute diff is within
+tolerance using this formula:  abs(a - b) < atol + rtol(abs(b))
+*/
+template <class R1, class R2>
+bool allclose(const R1& r1, const R2& r2, tolerance tols)
+{
+    std::size_t n = range_distance(r1);
+    if(n == range_distance(r2))
+    {
+        auto idx = mismatch_idx(r1, r2, [&](auto x, auto y) {
+            return abs_diff(double(x), double(y)) < tols.atol + tols.rtol * std::abs(double(y));
+        });
+        return idx >= range_distance(r1);
+    }
+    return false;
+}
+
 template <class R1, class R2>
-bool verify_range(const R1& r1, const R2& r2, double tolerance = 80, double* out_error = nullptr)
+bool verify_rms_range(const R1& r1,
+                      const R2& r2,
+                      std::size_t tolerance = 80,
+                      double* out_rms_error = nullptr)
 {
-    double threshold = std::numeric_limits<range_value<R1>>::epsilon() * tolerance;
+    double threshold = get_rms_tol(r1, tolerance);
     auto error       = rms_range(r1, r2);
-    if(out_error != nullptr)
-        *out_error = error;
+    if(out_rms_error != nullptr)
+        *out_rms_error = error;
     return error <= threshold;
 }
 
+template <class R1, class R2>
+bool verify_range_with_tolerance(const R1& r1,
+                                 const expected<R2>& r2,
+                                 tolerance tols        = tolerance{},
+                                 double* out_rms_error = nullptr)
+{
+    auto rms_error = rms_range(r1, r2.data());
+    // disable ewise_verify by default for now, it requires lot of tests to be fixed
+    bool ewise_verify = true;
+    if(enabled(MIGRAPHX_VERIFY_ENABLE_ALLCLOSE{}))
+    {
+        ewise_verify = allclose(r1, r2.data(), tols);
+    }
+    if(out_rms_error != nullptr)
+        *out_rms_error = rms_error;
+    return rms_error <= tols.rms_tol and ewise_verify;
+}
+
+// expected argument should be passed as second, but if it is passed as the first by mistake then
+// flip the order
+template <class R1, class R2>
+bool verify_range_with_tolerance(const expected<R1>& r1,
+                                 const R2& r2,
+                                 tolerance tols        = tolerance{},
+                                 double* out_rms_error = nullptr)
+{
+    return verify_rms_range(r2, r1, tols, out_rms_error);
+}
+
 } // namespace verify
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/include/migraphx/verify_args.hpp

@@ -31,11 +31,15 @@
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
 
-MIGRAPHX_EXPORT
-bool verify_args(const std::string& name,
-                 const argument& ref_arg,
-                 const argument& target_arg,
-                 double tolerance = 80);
+MIGRAPHX_EXPORT bool verify_args(const std::string& name,
+                                 const argument& target_arg,
+                                 const verify::expected<argument>& ref_arg,
+                                 verify::tolerance);
+
+MIGRAPHX_EXPORT bool verify_args_with_tolerance(const std::string& name,
+                                                const argument& target_arg,
+                                                const verify::expected<argument>& ref_arg,
+                                                std::size_t tolerance = 80);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx