[LPT] ReduceSum: zero point support

openvinotoolkit · Jun 12, 2024 · 48526d8 · 48526d8
1 parent 107869e
commit 48526d8
Show file tree

Hide file tree

Showing 3 changed files with 70 additions and 2 deletions.
diff --git a/src/common/low_precision_transformations/src/reduce_sum.cpp b/src/common/low_precision_transformations/src/reduce_sum.cpp
@@ -68,7 +68,10 @@ 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 reductionSizeConstant = ov::opset1::Constant::create(
+            dequantization.subtractConstant->get_element_type(),
+            Shape{},
+            { static_cast<float>(reductionSize) });
         const auto result = fold<ov::opset1::Multiply>(dequantization.subtractConstant, reductionSizeConstant);
 
         replace_node(dequantization.subtractConstant, result);

diff --git a/src/common/low_precision_transformations/tests/reduce_sum_transformation.cpp b/src/common/low_precision_transformations/tests/reduce_sum_transformation.cpp
@@ -30,6 +30,7 @@ using namespace ov::builder::subgraph;
 class ReduceSumTransformation : public ReduceTransformation<ov::op::v1::ReduceSum> {
     void SetUp() override {
         ReduceTransformation::SetUp();
+
         const auto transformationParams = std::get<1>(GetParam()).params;
 
         SimpleLowPrecisionTransformer transform;
@@ -336,4 +337,68 @@ 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},
+                    {{2.f, 4.f, 6.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,
+                {
+                    {},
+                    {{8.f, 16.f, 24.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}}
+                }
+            }
+        },
+        {
+            LayerTransformation::createParamsU8I8(),
+            {0},
+            true,
+            {
+                ov::element::i8,
+                {
+                    {ov::element::f32},
+                    {{2.f, 4.f, 6.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,
+                {
+                    {},
+                    {{8.f, 16.f, 24.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}}
+                }
+            }
+        }
+    };
+
+    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],