Fix Round operator inaccuracy (#2244)

The inaccuracy was caused by ONNX round requires nearest integer rounding for halway (0.5) cases.
std::round rounds away from zero, thus giving wrong results with halfway cases.
Replaced std::round with std::nearbyint which uses the correct rounding by default.

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/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/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/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/rewrite_quantization.cpp

@@ -47,7 +47,7 @@ void apply_quantizelinear(module& m, instruction_ref ins)
             ins, make_op("convert", {{"target_type", y_scale->get_shape().type()}}), x);
     }
     auto div            = m.insert_instruction(ins, make_op("div"), x, y_scale);
-    auto add_zero_point = m.insert_instruction(ins, make_op("round"), div);
+    auto add_zero_point = m.insert_instruction(ins, make_op("nearbyint"), div);
 
     if(ins->inputs().size() == 3)
     {

src/targets/gpu/kernels/include/migraphx/kernels/math.hpp

@@ -103,6 +103,7 @@ MIGRAPHX_DEVICE_MATH(floor, ::floor)
 MIGRAPHX_DEVICE_MATH(isnan, ::isnan)
 MIGRAPHX_DEVICE_MATH(isinf, ::isinf)
 MIGRAPHX_DEVICE_MATH(log, ::log)
+MIGRAPHX_DEVICE_MATH(nearbyint, ::nearbyint)
 MIGRAPHX_DEVICE_MATH(pow, ::pow)
 MIGRAPHX_DEVICE_MATH(remainder, ::remainder)
 MIGRAPHX_DEVICE_MATH(round, ::round)
@@ -152,6 +153,7 @@ MIGRAPHX_DEVICE_MATH_HALF(atan, ::atan)
 MIGRAPHX_DEVICE_MATH_HALF(atanh, ::atanh)
 MIGRAPHX_DEVICE_MATH_HALF(cosh, ::cosh)
 MIGRAPHX_DEVICE_MATH_HALF(erf, ::erf)
+MIGRAPHX_DEVICE_MATH_HALF(nearbyint, ::nearbyint)
 MIGRAPHX_DEVICE_MATH_HALF(pow, ::pow)
 MIGRAPHX_DEVICE_MATH_HALF(remainder, ::remainder)
 MIGRAPHX_DEVICE_MATH_HALF(round, ::round)
@@ -236,6 +238,7 @@ MIGRAPHX_DEVICE_MATH_VEC(isnan)
 MIGRAPHX_DEVICE_MATH_VEC(log)
 MIGRAPHX_DEVICE_MATH_VEC(max)
 MIGRAPHX_DEVICE_MATH_VEC(min)
+MIGRAPHX_DEVICE_MATH_VEC(nearbyint)
 MIGRAPHX_DEVICE_MATH_VEC(pow)
 MIGRAPHX_DEVICE_MATH_VEC(remainder)
 MIGRAPHX_DEVICE_MATH_VEC(round)

test/gpu/codegen_literal.cpp

@@ -64,7 +64,7 @@ TEST_CASE(mul_literal_round_test)
     auto l1 = mm->add_literal(1 / 0.00787402f);
 
     auto mul   = mm->add_instruction(migraphx::make_op("mul"), l0, l1);
-    auto round = mm->add_instruction(migraphx::make_op("round"), mul);
+    auto round = mm->add_instruction(migraphx::make_op("nearbyint"), mul);
 
     mm->add_return({round});
 

test/onnx/gen_onnx.py

@@ -7087,6 +7087,16 @@ def roialign_test():
     return ([node], [x, roi, bi], [y])
 
 
+@onnx_test()
+def round_half_test():
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT16, [4, 4])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT16, [4, 4])
+
+    node = onnx.helper.make_node('Round', inputs=['x'], outputs=['y'])
+
+    return ([node], [x], [y])
+
+
 @onnx_test()
 def scatter_add_test():
     x = helper.make_tensor_value_info('data', TensorProto.FLOAT, [3, 4, 5, 6])

test/onnx/onnx_test.cpp

@@ -5788,9 +5788,9 @@ TEST_CASE(quantizelinear_test)
     auto l1_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {5}}}), l1);
     auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
-    auto round = mm->add_instruction(migraphx::make_op("round"), div);
-    auto s     = round->get_shape();
-    auto clip  = insert_quantizelinear_clip(*mm, div, round, s, 0, 255);
+    auto nearbyint = mm->add_instruction(migraphx::make_op("nearbyint"), div);
+    auto s         = nearbyint->get_shape();
+    auto clip      = insert_quantizelinear_clip(*mm, div, nearbyint, s, 0, 255);
     mm->add_instruction(
         migraphx::make_op("convert",
                           {{"target_type", migraphx::to_value(migraphx::shape::uint8_type)}}),
@@ -5813,9 +5813,9 @@ TEST_CASE(quantizelinear_int32_test)
                           {{"target_type", migraphx::to_value(migraphx::shape::float_type)}}),
         l0);
     auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
-    auto round = mm->add_instruction(migraphx::make_op("round"), div);
-    auto s     = round->get_shape();
-    auto clip  = insert_quantizelinear_clip(*mm, div, round, s, 0, 255);
+    auto nearbyint = mm->add_instruction(migraphx::make_op("nearbyint"), div);
+    auto s         = nearbyint->get_shape();
+    auto clip      = insert_quantizelinear_clip(*mm, div, nearbyint, s, 0, 255);
     mm->add_instruction(
         migraphx::make_op("convert",
                           {{"target_type", migraphx::to_value(migraphx::shape::uint8_type)}}),
@@ -5835,7 +5835,7 @@ TEST_CASE(quantizelinear_zero_point_test)
     auto l1_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {5}}}), l1);
     auto div   = mm->add_instruction(migraphx::make_op("div"), l0, l1_mbcast);
-    auto round = mm->add_instruction(migraphx::make_op("round"), div);
+    auto round = mm->add_instruction(migraphx::make_op("nearbyint"), div);
     auto l2_mbcast =
         mm->add_instruction(migraphx::make_op("multibroadcast", {{"out_lens", {5}}}), l2);
     l2_mbcast = mm->add_instruction(
@@ -5868,7 +5868,7 @@ migraphx::program make_quantizelinear_axis_prog()
         migraphx::make_op("broadcast", {{"axis", axis}, {"out_lens", input_lens}}), l1);
 
     auto div      = mm->add_instruction(migraphx::make_op("div"), l0, l1_bcast);
-    auto round    = mm->add_instruction(migraphx::make_op("round"), div);
+    auto round    = mm->add_instruction(migraphx::make_op("nearbyint"), div);
     auto l2_bcast = mm->add_instruction(
         migraphx::make_op("broadcast", {{"axis", axis}, {"out_lens", input_lens}}), l2);
     l2_bcast = mm->add_instruction(
@@ -6997,7 +6997,7 @@ TEST_CASE(round_test)
     migraphx::program p;
     auto* mm   = p.get_main_module();
     auto input = mm->add_parameter("x", migraphx::shape{migraphx::shape::double_type, {10, 5}});
-    mm->add_instruction(migraphx::make_op("round"), input);
+    mm->add_instruction(migraphx::make_op("nearbyint"), input);
 
     auto prog = optimize_onnx("round_test.onnx");
     EXPECT(p == prog);

test/onnx/round_half_test.onnx

+	round_half_test:J
+
+
x
y"Roundround_half_testZ
+
x
+
+
+
+b
+
y
+
+
+
+B
\ No newline at end of file

test/onnx/verify_onnx.cpp

@@ -2056,6 +2056,43 @@ TEST_CASE(reversesequence_time_verify_test)
     EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
 }
 
+TEST_CASE(round_half_test)
+{
+    migraphx::program p = migraphx::parse_onnx("round_half_test.onnx");
+    p.compile(migraphx::make_target("ref"));
+
+    migraphx::shape xs{migraphx::shape::half_type, {4, 4}};
+    std::vector<float> tmp = {-3.51,
+                              -3.5,
+                              -3.49,
+                              -2.51,
+                              -2.50,
+                              -2.49,
+                              -1.6,
+                              -1.5,
+                              -0.51,
+                              -0.5,
+                              0.5,
+                              0.6,
+                              2.4,
+                              2.5,
+                              3.5,
+                              4.5};
+    std::vector<migraphx::half> data{tmp.cbegin(), tmp.cend()};
+    migraphx::parameter_map param_map;
+    param_map["x"] = migraphx::argument(xs, data.data());
+
+    auto result = p.eval(param_map).back();
+
+    std::vector<migraphx::half> result_vector;
+    result.visit([&](auto output) { result_vector.assign(output.begin(), output.end()); });
+
+    tmp = {-4.0, -4.0, -3.0, -3.0, -2.0, -2.0, -2.0, -2.0, -1.0, 0.0, 0.0, 1.0, 2.0, 2.0, 4.0, 4.0};
+    std::vector<migraphx::half> gold{tmp.cbegin(), tmp.cend()};
+
+    EXPECT(migraphx::verify::verify_rms_range(result_vector, gold));
+}
+
 TEST_CASE(selu_test)
 {
     migraphx::program p = migraphx::parse_onnx("selu_test.onnx");

test/py/onnx_backend_test.py

@@ -83,7 +83,6 @@ def disabled_tests_onnx_1_7_0(backend_test):
     backend_test.exclude(r'test_nonmaxsuppression_two_batches_cpu')
     backend_test.exclude(r'test_nonmaxsuppression_two_classes_cpu')
     backend_test.exclude(r'test_nonzero_example_cpu')
-    backend_test.exclude(r'test_round_cpu')
     backend_test.exclude(r'test_softmax_axis_0_cpu')
     backend_test.exclude(r'test_softmax_axis_1_cpu')
     backend_test.exclude(r'test_softmax_default_axis_cpu')

test/ref/round.cpp → test/ref/nearbyint.cpp

@@ -30,39 +30,71 @@
 
 #include <test.hpp>
 
-TEST_CASE(round_test)
+TEST_CASE(nearbyint_test)
 {
     migraphx::program p;
     auto* mm = p.get_main_module();
-    migraphx::shape s{migraphx::shape::float_type, {9}};
-    auto l =
-        mm->add_literal(migraphx::literal{s, {1.1, 1.5, 1.6, -1.1, -1.5, -1.6, 0.0, 2.0, -2.0}});
-    mm->add_instruction(migraphx::make_op("round"), l);
+    migraphx::shape s{migraphx::shape::float_type, {4, 4}};
+    auto l = mm->add_literal(migraphx::literal{s,
+                                               {-3.51,
+                                                -3.5,
+                                                -3.49,
+                                                -2.51,
+                                                -2.50,
+                                                -2.49,
+                                                -1.6,
+                                                -1.5,
+                                                -0.51,
+                                                -0.5,
+                                                0.5,
+                                                0.6,
+                                                2.4,
+                                                2.5,
+                                                3.5,
+                                                4.5}});
+    mm->add_instruction(migraphx::make_op("nearbyint"), l);
     p.compile(migraphx::make_target("ref"));
     auto result = p.eval({}).back();
     std::vector<float> results_vector;
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
-    std::vector<float> gold = {1.0, 2.0, 2.0, -1.0, -2.0, -2.0, 0.0, 2.0, -2.0};
+    std::vector<float> gold = {
+        -4.0, -4.0, -3.0, -3.0, -2.0, -2.0, -2.0, -2.0, -1.0, 0.0, 0.0, 1.0, 2.0, 2.0, 4.0, 4.0};
     EXPECT(migraphx::verify::verify_rms_range(results_vector, gold));
 }
 
-TEST_CASE(round_dyn_test)
+TEST_CASE(nearbyint_dyn_test)
 {
     migraphx::program p;
     auto* mm = p.get_main_module();
     migraphx::shape::dynamic_dimension dd{4, 10};
     migraphx::shape s{migraphx::shape::float_type, {dd}};
     auto input = mm->add_parameter("X", s);
-    mm->add_instruction(migraphx::make_op("round"), input);
+    mm->add_instruction(migraphx::make_op("nearbyint"), input);
     p.compile(migraphx::make_target("ref"));
 
-    std::vector<float> input_data{1.1, 1.5, 1.6, -1.1, -1.5, -1.6, 0.0, 2.0, -2.0};
+    std::vector<float> input_data{-3.51,
+                                  -3.5,
+                                  -3.49,
+                                  -2.51,
+                                  -2.50,
+                                  -2.49,
+                                  -1.6,
+                                  -1.5,
+                                  -0.51,
+                                  -0.5,
+                                  0.5,
+                                  0.6,
+                                  2.4,
+                                  2.5,
+                                  3.5,
+                                  4.5};
     migraphx::parameter_map params0;
-    migraphx::shape input_fixed_shape0{migraphx::shape::float_type, {9}};
+    migraphx::shape input_fixed_shape0{migraphx::shape::float_type, {16}};
     params0["X"] = migraphx::argument(input_fixed_shape0, input_data.data());
     auto result  = p.eval(params0).back();
     std::vector<float> results_vector;
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
-    std::vector<float> gold = {1.0, 2.0, 2.0, -1.0, -2.0, -2.0, 0.0, 2.0, -2.0};
+    std::vector<float> gold = {
+        -4.0, -4.0, -3.0, -3.0, -2.0, -2.0, -2.0, -2.0, -1.0, 0.0, 0.0, 1.0, 2.0, 2.0, 4.0, 4.0};
     EXPECT(migraphx::verify::verify_rms_range(results_vector, gold));
 }

test/ref/quantizelinear.cpp

@@ -55,7 +55,7 @@ TEST_CASE(quantizelinear_1)
     std::vector<float> results_vector(18);
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
     std::vector<float> gold{
-        -128, 127, 65, -128, 1, 1, -1, 100, 92, -128, 127, 65, -128, 1, 1, -1, 100, 92};
+        -128, 127, 64, -128, 1, 1, -1, 100, 92, -128, 127, 64, -128, 1, 1, -1, 100, 92};
     EXPECT(results_vector == gold);
 }
 
@@ -80,6 +80,6 @@ TEST_CASE(quantizelinear_2)
     auto result = p1.eval({}).back();
     std::vector<float> results_vector(18);
     result.visit([&](auto output) { results_vector.assign(output.begin(), output.end()); });
-    std::vector<float> gold{0, 255, 65, 0, 2, 2, 0, 255, 255, 0, 255, 65, 0, 2, 2, 0, 255, 255};
+    std::vector<float> gold{0, 255, 64, 0, 2, 2, 0, 255, 255, 0, 255, 64, 0, 2, 2, 0, 255, 255};
     EXPECT(results_vector == gold);
 }

test/simplify_reshapes_test.cpp

@@ -1345,7 +1345,7 @@ TEST_CASE(transpose_contiguous_unsqueeze_unary)
         auto cont_ins = m1.add_instruction(migraphx::make_op("contiguous"), transpose_ins);
         auto unsq_ins =
             m1.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}}), cont_ins);
-        auto round = m1.add_instruction(migraphx::make_op("round"), unsq_ins);
+        auto round = m1.add_instruction(migraphx::make_op("nearbyint"), unsq_ins);
         m1.add_instruction(pass_op{}, round);
     }
     run_pass(m1);
@@ -1354,7 +1354,7 @@ TEST_CASE(transpose_contiguous_unsqueeze_unary)
         auto x = m2.add_parameter("x", {migraphx::shape::float_type, {2, 8, 5, 5}});
         auto transpose_ins =
             m2.add_instruction(migraphx::make_op("transpose", {{"permutation", {0, 2, 3, 1}}}), x);
-        auto round    = m2.add_instruction(migraphx::make_op("round"), transpose_ins);
+        auto round    = m2.add_instruction(migraphx::make_op("nearbyint"), transpose_ins);
         auto cont_ins = m2.add_instruction(migraphx::make_op("contiguous"), round);
         auto unsq_ins =
             m2.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {2}}}), cont_ins);

test/verify/test_round.cpp → test/verify/test_nearbyint.cpp

 /*
  * 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
@@ -27,16 +27,21 @@
 #include <migraphx/generate.hpp>
 #include <migraphx/make_op.hpp>
 
-struct test_round : verify_program<test_round>
+template <class T>
+struct test_nearbyint : verify_program<test_nearbyint<T>>
 {
     migraphx::program create_program() const
     {
         migraphx::program p;
+        std::vector<float> tmp{-4.5, -3.5, 0.5, 2.5, 3.5};
+        std::vector<T> data{tmp.cbegin(), tmp.cend()};
+        migraphx::shape s1{migraphx::shape::get_type<T>(), {5}};
         auto* mm = p.get_main_module();
-
-        migraphx::shape s{migraphx::shape::float_type, {2, 3, 4, 6}};
-        auto param = mm->add_parameter("x", s);
-        mm->add_instruction(migraphx::make_op("round"), param);
+        auto l0  = mm->add_literal(migraphx::literal{s1, data});
+        mm->add_instruction(migraphx::make_op("isinf"), l0);
         return p;
     };
 };
+
+template struct test_nearbyint<migraphx::half>;
+template struct test_nearbyint<float>;