[CPU][ARM64] Enable FakeQuantize tokenization and decomposition (open…

…vinotoolkit#28700) ### Details: - **Enabled tests for FakeQuantizeDecomposition** in `src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp` - **Enabled FakeQuantize op tokenization** by adding it in `is_supported_op` in `transformation_pipeline.cpp` - Updated `FakeQuantizeDecompositionTest::SetUp()`, replaced MaxPool ops (`makePrerequisitesOriginal()`) with empty vector due to outdated limitations in Snippets, to support the operation in AArch64 platforms. - Adjusted expected node counts in tests since `reorder` ops are missed in AArch64 platforms and disabled `legacyFuse` tests for non-x64 platforms. ### Tests <img width="873" alt="Screenshot 2025-01-27 at 4 53 59 PM" src="https://github.com/user-attachments/assets/8184923e-85bb-45c0-bbca-59db7e6a5381" /> Passed local tests ### Tickets: - Closes openvinotoolkit#28508
mryzhov · Feb 3, 2025 · 91a8476 · 91a8476
1 parent 26a3ab1
commit 91a8476
Show file tree

Hide file tree

Showing 4 changed files with 21 additions and 23 deletions.
diff --git a/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp b/src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp
@@ -1120,17 +1120,18 @@ void Transformations::MainSnippets(void) {
                 ov::is_type<ov::op::v0::Clamp>(n) || ov::is_type<ov::op::v0::Ceiling>(n) ||
                 ov::is_type<ov::op::v0::Convert>(n) || ov::is_type<ov::op::v1::Divide>(n) ||
                 ov::is_type<ov::op::v0::Elu>(n) || ov::is_type<ov::op::v0::Exp>(n) ||
-                ov::is_type<ov::op::v1::Equal>(n) || ov::is_type<ov::op::v0::Floor>(n) ||
-                ov::is_type<ov::op::v1::FloorMod>(n) || ov::is_type<ov::op::v0::Gelu>(n) ||
-                ov::is_type<ov::op::v7::Gelu>(n) || ov::is_type<ov::op::v1::Greater>(n) ||
-                ov::is_type<ov::op::v1::GreaterEqual>(n) || ov::is_type<ov::op::v4::HSwish>(n) ||
-                ov::is_type<ov::op::v1::LessEqual>(n) || ov::is_type<ov::op::v1::Maximum>(n) ||
-                ov::is_type<ov::op::v1::Minimum>(n) || ov::is_type<ov::op::v4::Mish>(n) ||
-                ov::is_type<ov::op::v1::Mod>(n) || ov::is_type<ov::op::v1::Multiply>(n) ||
-                ov::is_type<ov::op::v0::PRelu>(n) || ov::is_type<ov::op::v0::Relu>(n) ||
-                ov::is_type<ov::op::v5::Round>(n) || ov::is_type<ov::op::v0::Sigmoid>(n) ||
-                ov::is_type<ov::op::v0::Sqrt>(n) || ov::is_type<ov::op::v1::Subtract>(n) ||
-                ov::is_type<ov::op::v4::Swish>(n) || ov::is_type<ov::op::v0::Tanh>(n));
+                ov::is_type<ov::op::v1::Equal>(n) || ov::is_type<ov::op::v0::FakeQuantize>(n) ||
+                ov::is_type<ov::op::v0::Floor>(n) || ov::is_type<ov::op::v1::FloorMod>(n) || 
+                ov::is_type<ov::op::v0::Gelu>(n) || ov::is_type<ov::op::v7::Gelu>(n) || 
+                ov::is_type<ov::op::v1::Greater>(n) || ov::is_type<ov::op::v1::GreaterEqual>(n) || 
+                ov::is_type<ov::op::v4::HSwish>(n) || ov::is_type<ov::op::v1::LessEqual>(n) || 
+                ov::is_type<ov::op::v1::Maximum>(n) || ov::is_type<ov::op::v1::Minimum>(n) || 
+                ov::is_type<ov::op::v4::Mish>(n) || ov::is_type<ov::op::v1::Mod>(n) || 
+                ov::is_type<ov::op::v1::Multiply>(n) || ov::is_type<ov::op::v0::PRelu>(n) || 
+                ov::is_type<ov::op::v0::Relu>(n) || ov::is_type<ov::op::v5::Round>(n) || 
+                ov::is_type<ov::op::v0::Sigmoid>(n) || ov::is_type<ov::op::v0::Sqrt>(n) || 
+                ov::is_type<ov::op::v1::Subtract>(n) || ov::is_type<ov::op::v4::Swish>(n) || 
+                ov::is_type<ov::op::v0::Tanh>(n));
 #else
         // CPU Plugin support Swish in Subgraph via conversion to SwichCPU which assumes second input to be constant,
         // and CPU Plugin does not support Mish for x64

diff --git a/src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp b/src/plugins/intel_cpu/tests/functional/shared_tests_instances/skip_tests_config.cpp
@@ -460,7 +460,7 @@ std::vector<std::string> disabledTestPatterns() {
     retVector.emplace_back(R"(smoke_Snippets.*\[.*\?.*\].*)");
     retVector.emplace_back(R"(smoke_Snippets_Eltwise.*\[1.1..10.1..8.1..4\].*)");
     // smoke_Snippets test cases are not supported on arm64 platforms, except for smoke_Snippets_Eltwise
-    retVector.emplace_back(R"(smoke_Snippets(?!_Eltwise|_Convert).*)");
+    retVector.emplace_back(R"(smoke_Snippets(?!_Eltwise|_Convert|_FQDecomposition_).*)");
     // arm snippets doesn't support sve_128 that required by dnnl injector jit_uni_eltwise_injector_f32 yet
     retVector.emplace_back(R"(smoke_Snippets_Eltwise_TwoResults.*)");
     retVector.emplace_back(R"(smoke_Snippets_Eltwise/TwoInputsAndOutputs.*)");

diff --git a/...cpu/tests/functional/shared_tests_instances/snippets/fake_quantize_decomposition_test.cpp b/...cpu/tests/functional/shared_tests_instances/snippets/fake_quantize_decomposition_test.cpp
@@ -73,8 +73,7 @@ INSTANTIATE_TEST_SUITE_P(
     ::testing::Combine(
         ::testing::ValuesIn(testValuesDecompositionScalars),
         ::testing::ValuesIn(operations),
-        // reorder (nChw[16|8]c) + MaxPool + Subgraph + reorder(nchw)
-        ::testing::Values(std::pair<size_t, size_t>{4, 1}),
+        ::testing::Values(std::pair<size_t, size_t>{1, 1}),
         ::testing::Values(ov::test::utils::DEVICE_CPU)),
     FakeQuantizeDecompositionTest::getTestCaseName);
 
@@ -84,8 +83,7 @@ INSTANTIATE_TEST_SUITE_P(
     ::testing::Combine(
         ::testing::ValuesIn(testValuesDecompositionPerChannel),
         ::testing::ValuesIn(operations),
-        // reorder (nChw[16|8]c) + MaxPool + reorder(nChw[16|8]c) x6 + Subgraph + reorder(nchw)
-        ::testing::Values(std::pair<size_t, size_t>{10, 1}),
+        ::testing::Values(std::pair<size_t, size_t>{1, 1}),
         ::testing::Values(ov::test::utils::DEVICE_CPU)),
     FakeQuantizeDecompositionTest::getTestCaseName);
 
@@ -95,13 +93,12 @@ INSTANTIATE_TEST_SUITE_P(
     ::testing::Combine(
         ::testing::ValuesIn(testValuesDecompositionPerChannelInput),
         ::testing::ValuesIn(operations),
-        // reorder (nChw[16|8]c) + MaxPool + reorder(nChw[16|8]c) x4 + Subgraph + reorder(nchw)
-        ::testing::Values(std::pair<size_t, size_t>{8, 1}),
+        ::testing::Values(std::pair<size_t, size_t>{1, 1}),
         ::testing::Values(ov::test::utils::DEVICE_CPU)),
     FakeQuantizeDecompositionTest::getTestCaseName);
 }  // namespace decompositionInSubgraph
 
-
+#ifdef OPENVINO_ARCH_X86_64
 namespace legacyFuse {
 const std::vector<TestValues> testValuesLegacyFuse = {
     {
@@ -144,11 +141,11 @@ INSTANTIATE_TEST_SUITE_P(
     ::testing::Combine(
         ::testing::ValuesIn(testValuesLegacyFuse),
         ::testing::ValuesIn(operations),
-        // reorder (nChw[16|8]c) + MaxPool + reorder(nhwc) + Convolution(with internal weight reordering) + reorder(nchw)
-        ::testing::Values(std::pair<size_t, size_t>{5, 0}),
+        // reorder (nChw[16|8]c) + Convolution(with internal weight reordering) + reorder(nchw)
+        ::testing::Values(std::pair<size_t, size_t>{3, 0}),
         ::testing::Values(ov::test::utils::DEVICE_CPU)),
     FakeQuantizeDecompositionTest::getTestCaseName);
-
 }  // namespace legacyFuse
+#endif
 
 }  // namespace
diff --git a/src/tests/functional/plugin/shared/src/snippets/fake_quantize_decomposition_test.cpp b/src/tests/functional/plugin/shared/src/snippets/fake_quantize_decomposition_test.cpp
@@ -61,7 +61,7 @@ void FakeQuantizeDecompositionTest::SetUp() {
         values.inputType,
         values.fakeQuantizeShapes,
         values.zeroPoint,
-        ov::test::snippets::FunctionHelper::makePrerequisitesOriginal(),
+        {},
         op);
 }