[LPT] ReduceSum: zero point support (openvinotoolkit#24977)

### Details: - *[LPT] ReduceSum: zero point support* ### Tickets: - *CVS-142256*
evkotov · Jun 13, 2024 · 5769fab · 5769fab
1 parent d2019eb
commit 5769fab
Show file tree

Hide file tree

Showing 3 changed files with 78 additions and 3 deletions.
diff --git a/src/common/low_precision_transformations/src/reduce_sum.cpp b/src/common/low_precision_transformations/src/reduce_sum.cpp
@@ -69,10 +69,21 @@ void ReduceSumTransformation::changeDequantizationValues(
 
         // (a1 - s) + (a2 - s) + ... + (an - s) = (a1 + a2 + ... + an) - n * s
         const auto reductionSizeConstant = ov::opset1::Constant::create(deqPrecision, Shape{}, { static_cast<float>(reductionSize) });
-        const auto result = fold<ov::opset1::Multiply>(dequantization.subtractConstant, reductionSizeConstant);
+        OPENVINO_ASSERT(deqPrecision == dequantization.subtract->get_input_element_type(0),
+                        "dequantization precision ", deqPrecision,
+                        " differs from zero point 0 input ", dequantization.subtract->get_input_element_type(0));
+        const auto result = fold<ov::opset1::Multiply>(
+            foldConvert(dequantization.subtractConstant, deqPrecision),
+            reductionSizeConstant);
+
+        replace_node(
+            dequantization.subtractConvert != nullptr ?
+                std::dynamic_pointer_cast<ov::Node>(dequantization.subtractConvert) :
+                dequantization.subtractConstant,
+            result);
 
-        replace_node(dequantization.subtractConstant, result);
         dequantization.subtractConstant = ov::as_type_ptr<ov::opset1::Constant>(result);
+        dequantization.subtractConvert = nullptr;
     }
 }
 

diff --git a/src/common/low_precision_transformations/tests/reduce_sum_transformation.cpp b/src/common/low_precision_transformations/tests/reduce_sum_transformation.cpp
@@ -336,4 +336,68 @@ INSTANTIATE_TEST_SUITE_P(
         ::testing::ValuesIn(reduceSumTransformationTestValues)),
     ReduceSumTransformation::getTestCaseName);
 } // namespace testValues2
+
+namespace testValues3 {
+    const std::vector<ov::PartialShape> inputShapes = {
+        {4, 3, 16, 16}
+    };
+
+    const std::vector<ReduceTransformationTestValues> reduceSumTransformationTestValues = {
+        {
+            LayerTransformation::createParamsU8I8(),
+            {0},
+            true,
+            {
+                ov::element::u8,
+                {
+                    {ov::element::f32},
+                    {{40.f, 80.f, 120.f}, ov::element::f32, {1, 3, 1, 1}, false, 1ul, ov::element::u8, true},
+                    {{0.1f, 1.f, 10.f}, ov::element::f32, {1, 3, 1, 1}}
+                }
+            },
+            {
+                ov::element::u8,
+                {},
+                ov::element::f32,
+                {
+                    {},
+                    {{160.f, 320.f, 480.f}, ov::element::f32, {1, 3, 1, 1}},
+                    {{0.1f, 1.f, 10.f}, ov::element::f32, {1, 3, 1, 1}}
+                }
+            }
+        },
+        {
+            LayerTransformation::createParamsU8I8(),
+            {0},
+            true,
+            {
+                ov::element::i8,
+                {
+                    {ov::element::f32},
+                    {{40.f, 80.f, 120.f}, ov::element::f32, {1, 3, 1, 1}, false, 1ul, ov::element::i8, true},
+                    {{0.1f, 1.f, 10.f}, ov::element::f32, {1, 3, 1, 1}}
+                }
+            },
+            {
+                ov::element::i8,
+                {},
+                ov::element::f32,
+                {
+                    {},
+                    {{160.f, 320.f, 480.f}, ov::element::f32, {1, 3, 1, 1}},
+                    {{0.1f, 1.f, 10.f}, ov::element::f32, {1, 3, 1, 1}}
+                }
+            }
+        }
+    };
+
+    INSTANTIATE_TEST_SUITE_P(
+        smoke_LPT,
+        ReduceSumTransformation,
+        ::testing::Combine(
+            ::testing::ValuesIn(inputShapes),
+            ::testing::ValuesIn(reduceSumTransformationTestValues)),
+        ReduceSumTransformation::getTestCaseName);
+} // namespace testValues3
+
 } // namespace
diff --git a/src/core/src/op/multiply.cpp b/src/core/src/op/multiply.cpp
@@ -51,7 +51,7 @@ bool Multiply::evaluate(TensorVector& outputs, const TensorVector& inputs) const
                                       this,
                                       outputs,
                                       inputs,
-                                      OV_PP_ET_LIST(f32, f64, i32, i64, u32, u64),
+                                      OV_PP_ET_LIST(f32, f64, i8, i32, i64, u8, u32, u64),
                                       multiply::Evaluate,
                                       inputs[0].get_element_type(),
                                       inputs[0],