use threshold instead of tolerance

docs/driver.rst

@@ -50,9 +50,9 @@ Runs reference and CPU or GPU implementations and checks outputs for consistency
 
 .. include:: ./driver/compile.rst
 
-.. option::  --tolerance [double]
+.. option::  --threshold [double]
 
-Tolerance for errors (Default: 80)
+Threshold for RMS error (Default: 0.001)
 
 .. option::  -i, --per-instruction
 

examples/migraphx/migraphx_driver/README.md

@@ -55,7 +55,7 @@ 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                                      |
+| --threshold                              | threshold for errors                                      |
 | --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 +536,13 @@ struct params : command<params>
 struct verify : command<verify>
 {
     compiler c;
-    double tolerance     = 80;
+    double threshold     = 0.001;
     bool per_instruction = false;
     bool reduce          = false;
     void parse(argument_parser& ap)
     {
         c.parse(ap);
-        ap(tolerance, {"--tolerance"}, ap.help("Tolerance for errors"));
+        ap(threshold, {"--threshold"}, ap.help("threshold for the RMS error"));
         ap(per_instruction,
            {"-i", "--per-instruction"},
            ap.help("Verify each instruction"),
@@ -567,15 +567,15 @@ struct verify : command<verify>
 
         if(per_instruction)
         {
-            verify_instructions(p, t, c.co, quantize, tolerance);
+            verify_instructions(p, t, c.co, quantize, threshold);
         }
         else if(reduce)
         {
-            verify_reduced_program(p, t, c.co, quantize, m, tolerance);
+            verify_reduced_program(p, t, c.co, quantize, m, threshold);
         }
         else
         {
-            verify_program(c.l.file, p, t, c.co, quantize, m, tolerance);
+            verify_program(c.l.file, p, t, c.co, quantize, m, threshold);
         }
     }
 };

src/driver/verify.cpp

@@ -76,7 +76,7 @@ void verify_program(const std::string& name,
                     compile_options options,
                     precision quantize,
                     const parameter_map& inputs,
-                    double tolerance)
+                    double threshold)
 {
     auto x = run_ref(p, inputs);
     auto y = run_target(p, t, options, quantize, inputs);
@@ -84,7 +84,7 @@ void verify_program(const std::string& name,
     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);
+        verify_args(name, x[i], y[i], threshold);
     }
 }
 
@@ -92,7 +92,7 @@ void verify_instructions(const program& prog,
                          const target& t,
                          compile_options options,
                          precision quantize,
-                         double tolerance)
+                         double threshold)
 {
     const auto* mm_prog = prog.get_main_module();
     for(auto&& ins : (*mm_prog))
@@ -124,7 +124,7 @@ 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);
+                ins.name(), p, t, options, quantize, create_param_map(p, false), threshold);
         }
         catch(...)
         {
@@ -140,14 +140,14 @@ void verify_reduced(program p,
                     compile_options options,
                     precision quantize,
                     const parameter_map& inputs,
-                    double tolerance)
+                    double threshold)
 {
     auto* mm  = p.get_main_module();
     auto last = std::prev(mm->end(), n + 1);
     mm->remove_instructions(last, mm->end());
     std::cout << "Verify: " << n << std::endl;
     std::cout << p << std::endl;
-    verify_program(std::to_string(n), p, t, options, quantize, inputs, tolerance);
+    verify_program(std::to_string(n), p, t, options, quantize, inputs, threshold);
 }
 
 void verify_reduced_program(const program& p,
@@ -155,14 +155,14 @@ void verify_reduced_program(const program& p,
                             compile_options options,
                             precision quantize,
                             const parameter_map& inputs,
-                            double tolerance)
+                            double threshold)
 {
     const auto* mm = p.get_main_module();
     auto n         = std::distance(mm->begin(), mm->end());
     std::cout << "Verify steps: " << n << std::endl;
     for(std::size_t i = 0; i < n; i++)
     {
-        verify_reduced(p, i, t, options, quantize, inputs, tolerance);
+        verify_reduced(p, i, t, options, quantize, inputs, threshold);
     }
 }
 

src/driver/verify.hpp

@@ -37,18 +37,18 @@ void verify_program(const std::string& name,
                     compile_options options     = compile_options{},
                     precision quantize          = precision::fp32,
                     const parameter_map& inputs = {},
-                    double tolerance            = 100);
+                    double threshold            = 0.001);
 void verify_instructions(const program& prog,
                          const target& t,
                          compile_options options = compile_options{},
                          precision quantize      = precision::fp32,
-                         double tolerance        = 80);
+                         double threshold        = 0.001);
 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);
+                            double threshold            = 0.001);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace driver

src/include/migraphx/verify.hpp

@@ -187,10 +187,31 @@ double rms_range(const R1& r1, const R2& r2)
         return std::numeric_limits<range_value<R1>>::max();
 }
 
+template <class R>
+double get_threshold(const R&, std::size_t tolerance = 80)
+{
+    double threshold = std::numeric_limits<range_value<R>>::epsilon() * tolerance;
+    return threshold;
+}
+
 template <class R1, class R2>
-bool verify_range(const R1& r1, const R2& r2, double tolerance = 80, double* out_error = nullptr)
+bool verify_range(const R1& r1,
+                  const R2& r2,
+                  std::size_t tolerance = 80,
+                  double* out_error     = nullptr)
 {
+    // double threshold = get_threshold(r1, tolerance);
     double threshold = std::numeric_limits<range_value<R1>>::epsilon() * tolerance;
+
+    auto error = rms_range(r1, r2);
+    if(out_error != nullptr)
+        *out_error = error;
+    return error <= threshold;
+}
+
+template <class R1, class R2>
+bool verify_range(const R1& r1, const R2& r2, double threshold, double* out_error = nullptr)
+{
     auto error = rms_range(r1, r2);
     if(out_error != nullptr)
         *out_error = error;

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,
+MIGRAPHX_EXPORT bool verify_args(const std::string& name,
                                  const argument& ref_arg,
                                  const argument& target_arg,
-                 double tolerance = 80);
+                                 double threshold);
+
+MIGRAPHX_EXPORT bool verify_args(const std::string& name,
+                                 const argument& ref_arg,
+                                 const argument& target_arg,
+                                 std::size_t tolerance = 80);
 
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

src/verify_args.cpp

@@ -30,12 +30,12 @@ inline namespace MIGRAPHX_INLINE_NS {
 bool verify_args(const std::string& name,
                  const argument& ref_arg,
                  const argument& target_arg,
-                 double tolerance)
+                 double threshold)
 {
     bool passed = true;
     visit_all(ref_arg, target_arg)([&](auto ref, auto target) {
         double error;
-        passed = verify::verify_range(ref, target, tolerance, &error);
+        passed = verify::verify_range(ref, target, threshold, &error);
         if(not passed)
         {
             // TODO: Check for nans
@@ -93,5 +93,15 @@ bool verify_args(const std::string& name,
     return passed;
 }
 
+bool verify_args(const std::string& name,
+                 const argument& ref_arg,
+                 const argument& target_arg,
+                 std::size_t tolerance)
+{
+    double threshold = 0.001;
+    target_arg.visit([&](auto ta) { threshold = verify::get_threshold(ta, tolerance); });
+    return verify_args(name, ref_arg, target_arg, threshold);
+}
+
 } // namespace MIGRAPHX_INLINE_NS
 } // namespace migraphx

test/gpu/quantization.cpp

@@ -118,7 +118,7 @@ TEST_CASE(int8_quantization)
         // the regular pipeline uses the rewrite_quantization in the much
         // earlier stage.
         if(migraphx::gpu::mlir_enabled())
-            EXPECT(migraphx::verify::verify_range(ref_result, gpu_result, 1e5));
+            EXPECT(migraphx::verify::verify_range(ref_result, gpu_result, 0.0119209));
         else
             EXPECT(migraphx::verify::verify_range(ref_result, gpu_result));
     }

test/quantization.cpp

@@ -1077,7 +1077,7 @@ TEST_CASE(int8_quantization_dot)
         std::vector<float> no_quant_result;
         run_prog(p, ref_t, m, no_quant_result);
 
-        EXPECT(migraphx::verify::verify_range(quant_result, no_quant_result, 30000));
+        EXPECT(migraphx::verify::verify_range(quant_result, no_quant_result, 0.003576));
     }
 }
 

test/ref/multinomial.cpp

@@ -78,5 +78,5 @@ TEST_CASE(multinomial_test)
     std::transform(res_dist.begin(), res_dist.end(), res_norm.begin(), [&](auto n) {
         return static_cast<double>(n) / res_dist_sum;
     });
-    EXPECT(migraphx::verify::verify_range(norm, res_norm, 100000));
+    EXPECT(migraphx::verify::verify_range(norm, res_norm, 0.01));
 }

test/ref/random_uniform.cpp

@@ -68,7 +68,7 @@ TEST_CASE(random_uniform_test)
     std::uniform_real_distribution<> dis(0.0, 1.0);
     std::vector<float> rand_samples(sample_size);
     std::generate(rand_samples.begin(), rand_samples.end(), [&]() { return dis(gen); });
-    EXPECT(migraphx::verify::verify_range(result_vec, rand_samples, 100));
+    EXPECT(migraphx::verify::verify_range(result_vec, rand_samples, 0.00001));
 }
 
 TEST_CASE(random_uniform_int_test)

test/ref/rnn_ops.cpp

@@ -1008,7 +1008,7 @@ TEST_CASE(rnn_fp16)
 
     std::vector<float> last_output_data_gold{
         0.2935145, -0.23719997, -0.31123261, -0.18357255, 0., 0., 0., 0.};
-    EXPECT(migraphx::verify::verify_range(last_output_data, last_output_data_gold, 5e4));
+    EXPECT(migraphx::verify::verify_range(last_output_data, last_output_data_gold, 0.005));
 }
 
 TEST_CASE(gru_forward)
@@ -2983,7 +2983,7 @@ TEST_CASE(gru_fp16)
                                     -0.3969709,  0.43360898, 0.35775262,  0.23280787, -0.52179873,
                                     -0.21944991, 0.4535257,  -0.13735442, 0.51757574, 0.50380427};
 
-    EXPECT(migraphx::verify::verify_range(hs_data, hs_data_gold, 5e4));
+    EXPECT(migraphx::verify::verify_range(hs_data, hs_data_gold, 0.00596));
 }
 
 TEST_CASE(lstm_forward)