observe the results of multiply-add across implementations

- add specific test showing diverging results when using FMA
- re-write eval test to not fail due to FMA
- re-write hitcollector test to not fail due to FMA

5 files changed, 63 insertions, 9 deletions

diff --git a/eval/CMakeLists.txt b/eval/CMakeLists.txt
index 3a9aabc83ba..3e81521550a 100644
--- a/eval/CMakeLists.txt
+++ b/eval/CMakeLists.txt
@@ -19,6 +19,7 @@ vespa_define_module(
     src/tests/eval/gbdt
     src/tests/eval/inline_operation
     src/tests/eval/interpreted_function
+    src/tests/eval/multiply_add
     src/tests/eval/node_tools
     src/tests/eval/node_types
     src/tests/eval/param_usage
diff --git a/eval/src/tests/eval/multiply_add/CMakeLists.txt b/eval/src/tests/eval/multiply_add/CMakeLists.txt
new file mode 100644
index 00000000000..c50aa4f50a2
--- /dev/null
+++ b/eval/src/tests/eval/multiply_add/CMakeLists.txt
@@ -0,0 +1,9 @@
+# Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+vespa_add_executable(eval_multiply_add_test_app TEST
+    SOURCES
+    multiply_add_test.cpp
+    DEPENDS
+    vespaeval
+    gtest
+)
+vespa_add_test(NAME eval_multiply_add_test_app COMMAND eval_multiply_add_test_app)
diff --git a/eval/src/tests/eval/multiply_add/multiply_add_test.cpp b/eval/src/tests/eval/multiply_add/multiply_add_test.cpp
new file mode 100644
index 00000000000..35cab0a6030
--- /dev/null
+++ b/eval/src/tests/eval/multiply_add/multiply_add_test.cpp
@@ -0,0 +1,44 @@
+// Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+#include <vespa/eval/eval/function.h>
+#include <vespa/eval/eval/llvm/compiled_function.h>
+#include <vespa/eval/eval/interpreted_function.h>
+#include <vespa/eval/tensor/default_tensor_engine.h>
+#include <vespa/vespalib/gtest/gtest.h>
+
+using namespace vespalib::eval;
+
+using Engine = vespalib::tensor::DefaultTensorEngine;
+
+double gcc_fun(double a, double b) {
+    return (a * 3) + b;
+}
+
+TEST(MultiplyAddTest, multiply_add_gives_same_result) {
+    auto fun = Function::parse("a*3+b");
+    CompiledFunction cfun(*fun, PassParams::ARRAY);
+    NodeTypes node_types = NodeTypes(*fun, {ValueType::double_type(), ValueType::double_type()});
+    InterpretedFunction ifun(Engine::ref(), *fun, node_types);
+    auto llvm_fun = cfun.get_function();
+    //-------------------------------------------------------------------------
+    double a = -1.0/3.0;
+    double b = 1.0;
+    std::vector<double> ab({a, b});
+    SimpleParams params(ab);
+    InterpretedFunction::Context ictx(ifun);
+    //-------------------------------------------------------------------------
+    const Value &result_value = ifun.eval(ictx, params);
+    double ifun_res = result_value.as_double();
+    double llvm_res = llvm_fun(&ab[0]);
+    double gcc_res = gcc_fun(a, b);
+    fprintf(stderr, "ifun_res: %a\n", ifun_res);
+    fprintf(stderr, "llvm_res: %a\n", llvm_res);
+    fprintf(stderr, "gcc_res:  %a\n", gcc_res);
+    EXPECT_EQ(ifun_res, llvm_res);
+    EXPECT_DOUBLE_EQ(llvm_res + 1.0, gcc_res + 1.0);
+    if (llvm_res != gcc_res) {
+        fprintf(stderr, "WARNING: diverging results caused by fused multiply add\n");
+    }
+}
+
+GTEST_MAIN_RUN_ALL_TESTS()
diff --git a/eval/src/vespa/eval/eval/test/eval_spec.cpp b/eval/src/vespa/eval/eval/test/eval_spec.cpp
index 709234a1a2c..dbc20dcf606 100644
--- a/eval/src/vespa/eval/eval/test/eval_spec.cpp
+++ b/eval/src/vespa/eval/eval/test/eval_spec.cpp
@@ -162,11 +162,11 @@ EvalSpec::add_function_call_cases() {
 void
 EvalSpec::add_tensor_operation_cases() {
     add_rule({"a", -1.0, 1.0}, "map(a,f(x)(sin(x)))", [](double x){ return std::sin(x); });
-    add_rule({"a", -1.0, 1.0}, "map(a,f(x)(x+x*3))", [](double x){ return (x + (x * 3)); });
+    add_rule({"a", -1.0, 1.0}, "map(a,f(x)(x*x*3))", [](double x){ return ((x * x) * 3); });
     add_rule({"a", -1.0, 1.0}, {"b", -1.0, 1.0}, "join(a,b,f(x,y)(x+y))", [](double x, double y){ return (x + y); });
-    add_rule({"a", -1.0, 1.0}, {"b", -1.0, 1.0}, "join(a,b,f(x,y)(x+y*3))", [](double x, double y){ return (x + (y * 3)); });
+    add_rule({"a", -1.0, 1.0}, {"b", -1.0, 1.0}, "join(a,b,f(x,y)(x*y*3))", [](double x, double y){ return ((x * y) * 3); });
     add_rule({"a", -1.0, 1.0}, {"b", -1.0, 1.0}, "merge(a,b,f(x,y)(x+y))", [](double x, double y){ return (x + y); });
-    add_rule({"a", -1.0, 1.0}, {"b", -1.0, 1.0}, "merge(a,b,f(x,y)(x+y*3))", [](double x, double y){ return (x + (y * 3)); });
+    add_rule({"a", -1.0, 1.0}, {"b", -1.0, 1.0}, "merge(a,b,f(x,y)(x*y*3))", [](double x, double y){ return ((x * y) * 3); });
     add_rule({"a", -1.0, 1.0}, "reduce(a,avg)", [](double a){ return a; });
     add_rule({"a", -1.0, 1.0}, "reduce(a,count)", [](double){ return 1.0; });
     add_rule({"a", -1.0, 1.0}, "reduce(a,prod)", [](double a){ return a; });
diff --git a/searchlib/src/tests/hitcollector/hitcollector_test.cpp b/searchlib/src/tests/hitcollector/hitcollector_test.cpp
index 31a24d2a8f1..2274314c7da 100644
--- a/searchlib/src/tests/hitcollector/hitcollector_test.cpp
+++ b/searchlib/src/tests/hitcollector/hitcollector_test.cpp
@@ -55,7 +55,7 @@ void checkResult(const ResultSet & rs, const std::vector<RankedHit> & exp)
 
         for (uint32_t i = 0; i < exp.size(); ++i) {
             EXPECT_EQUAL(rh[i]._docId, exp[i]._docId);
-            EXPECT_EQUAL(rh[i]._rankValue, exp[i]._rankValue);
+            EXPECT_EQUAL(rh[i]._rankValue + 1.0, exp[i]._rankValue + 1.0);
         }
     } else {
         ASSERT_TRUE(rs.getArray() == nullptr);
@@ -328,7 +328,7 @@ TEST("testScaling") {
         finalScores[3] = 300;
         finalScores[4] = 400;
 
-        testScaling(initScores, std::move(finalScores), exp);
+        TEST_DO(testScaling(initScores, std::move(finalScores), exp));
     }
     { // scale down and adjust up
         exp[0]._rankValue = 200; // scaled
@@ -342,7 +342,7 @@ TEST("testScaling") {
         finalScores[3] = 500;
         finalScores[4] = 600;
 
-        testScaling(initScores, std::move(finalScores), exp);
+        TEST_DO(testScaling(initScores, std::move(finalScores), exp));
     }
     { // scale up and adjust down
 
@@ -357,7 +357,7 @@ TEST("testScaling") {
         finalScores[3] = 3250;
         finalScores[4] = 4500;
 
-        testScaling(initScores, std::move(finalScores), exp);
+        TEST_DO(testScaling(initScores, std::move(finalScores), exp));
     }
     { // minimal scale (second phase range = 0 (4 - 4) -> 1)
         exp[0]._rankValue = 1; // scaled
@@ -371,7 +371,7 @@ TEST("testScaling") {
         finalScores[3] = 4;
         finalScores[4] = 4;
 
-        testScaling(initScores, std::move(finalScores), exp);
+        TEST_DO(testScaling(initScores, std::move(finalScores), exp));
     }
     { // minimal scale (first phase range = 0 (4000 - 4000) -> 1)
         std::vector<feature_t> is(initScores);
@@ -387,7 +387,7 @@ TEST("testScaling") {
         finalScores[3] = 400;
         finalScores[4] = 500;
 
-        testScaling(is, std::move(finalScores), exp);
+        TEST_DO(testScaling(is, std::move(finalScores), exp));
     }
 }