[CPU][ACL] LPT transformations are enabled + FQ decomposition #28981

alvoron · 2025-02-13T15:37:40Z

Details:

It depends on [CPU][ACL] int8 MatMul support #28870 so this PR should be merged after
LPT is enabled on ARM
FQ is decomposed to avoid reference code

Tickets:

CVS-162444

dmitry-gorokhov · 2025-02-19T05:03:24Z

src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp

-                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::Floor>(n) ||


Lets return FQ here. We are disabling Snippets for Quantized models in general.

Reverted this change

dmitry-gorokhov · 2025-02-19T05:04:40Z

src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp

@@ -1325,7 +1326,14 @@ void Transformations::Snippets(void) {
    }

    CPU_DEBUG_CAP_TRANSFORMATION_SCOPE(this, Snippets);
+// Disable MainSnippets for int8 models on arm platforms
+#if defined(OPENVINO_ARCH_ARM64)
+    if (!ov::pass::low_precision::LowPrecision::isFunctionQuantized(model, {ov::pass::low_precision::levels::int8})) {


Should be same list of levels as defined in Transformations::UpToLpt

dmitry-gorokhov · 2025-02-19T05:05:31Z

src/plugins/intel_cpu/tests/functional/custom/subgraph_tests/src/common/convs_and_sums.cpp

@@ -119,8 +119,11 @@ class ConvsAndSums : virtual public SubgraphBaseStaticTest {

        auto result = std::make_shared<ov::op::v0::Result>(relu3);
        function = std::make_shared<ov::Model>(result, params, "SimpleNet");
-
+#if defined(OPENVINO_ARCH_ARM64) || defined(OPENVINO_ARCH_ARM)
+        abs_threshold = 5e-2;


Should depend on FP16 inference precision, not the arch. Please fix here and below.

dmitry-gorokhov · 2025-02-19T05:06:33Z

src/plugins/intel_cpu/src/graph_optimizer.cpp

+        }
+
+        auto fusedOps = parentNode->getFusedWith();
+        fusedOps[fusedOps.size() - 1]->setOriginalOutputPrecisionAtPort(0, childNode->getOriginalOutputPrecisionAtPort(0));


What if fusedOps is empty?

Added the check

dmitry-gorokhov · 2025-02-19T05:09:26Z

src/plugins/intel_cpu/src/graph_optimizer.cpp

+        auto parentEdges = childNode->parentEdges;
+        for (auto& parentEdge : parentEdges) {
+            auto p_edge = parentEdge.lock();
+            if (p_edge->getParent()->getType() == Type::Eltwise) {
+                continue;
+            }
+            graph.RemoveEdge(p_edge);
+        }


Seems like this code is already part of graph.DropNode(childNode);

It looks so, removed this code

dmitry-gorokhov · 2025-02-19T05:10:17Z

src/plugins/intel_cpu/src/nodes/kernels/aarch64/jit_uni_eltwise_generic.cpp

+        } else if (ops_list_[i] == ov::intel_cpu::Type::Convert) {
+            // do nothing


not required anymore

dmitry-gorokhov · 2025-02-19T05:10:26Z

src/plugins/intel_cpu/src/nodes/eltwise.cpp

+            } else if (node->getType() == Type::Convert) {
+                // do nothing


not required anymore

…FakeQuantizeDecomposition

alvoron added 14 commits January 24, 2025 15:58

added lowp_gemm_acl to MM impl priority list

4b05b5c

added LPT MatMulTransformation test for arm

5ae4c4e

add u8 test case and skip it

3091bff

Merge branch 'master' into alvoron_matmut_int8

2765a27

disable fq fusing

c8ae024

address commints

5b2f1df

remove lowp_gemm_acl

696c3db

enable lpt on arm and fq decomposition

5d2fc38

disable MatMulTransformation on ARM

93b2582

remove 3rd party changes

4531f55

Enable MatMulTransformation against FC nodes only

abe2bcb

disable lpt

64956de

skip all MatMulTransformation tests

18d780c

fq decomposition

159653c

alvoron added the platform: arm OpenVINO on ARM / ARM64 label Feb 13, 2025

alvoron requested review from a team as code owners February 13, 2025 15:37

github-actions bot added category: IE Tests OpenVINO Test: plugins and common category: CPU OpenVINO CPU plugin labels Feb 13, 2025

alvoron added 2 commits February 14, 2025 08:07

Merge branch 'master' into alvoron_fq_decomposition_arm

82c9814

Update transformation_pipeline.cpp

0eaeda4

alvoron force-pushed the alvoron_fq_decomposition_arm branch 3 times, most recently from afeccad to d96d1bc Compare February 15, 2025 13:34

adjust tests threshold on arm

c9809be

alvoron force-pushed the alvoron_fq_decomposition_arm branch from d96d1bc to c9809be Compare February 15, 2025 13:44

Merge branch 'master' into alvoron_fq_decomposition_arm

39ce30d

alvoron force-pushed the alvoron_fq_decomposition_arm branch from 6cd6a6e to 3fb8e85 Compare February 15, 2025 18:59

alvoron force-pushed the alvoron_fq_decomposition_arm branch 3 times, most recently from b42fbb5 to 444fe5e Compare February 17, 2025 10:05

adjust threshold and fix clang

6c8f3c8

alvoron force-pushed the alvoron_fq_decomposition_arm branch from 444fe5e to 6c8f3c8 Compare February 17, 2025 12:31

alvoron added 2 commits February 18, 2025 09:04

Merge branch 'master' into alvoron_fq_decomposition_arm

979ce24

implemented MergeEltwiseAndConvert and disanled snippets on int8

fbe7849

dmitry-gorokhov added this to the 2025.1 milestone Feb 18, 2025

fix MergeEltwiseAndConvert

5963734

dmitry-gorokhov reviewed Feb 19, 2025

View reviewed changes

alvoron added 2 commits February 19, 2025 09:55

address comments

3075684

fix clang and InitGraphStatefulDiffPrimitiveModel test

3222eb5

alvoron force-pushed the alvoron_fq_decomposition_arm branch from 7b27f5c to 3222eb5 Compare February 19, 2025 12:15

Merge branch 'master' into alvoron_fq_decomposition_arm

36698da

alvoron force-pushed the alvoron_fq_decomposition_arm branch 3 times, most recently from 605f5c3 to ef61161 Compare February 19, 2025 15:22

fix arm32 test

e1293b1

alvoron force-pushed the alvoron_fq_decomposition_arm branch from ef61161 to e1293b1 Compare February 20, 2025 09:20

alvoron added 2 commits February 20, 2025 14:19

add arm 32 bit defines

9d18773

replace ov::snippets::pass::FakeQuantizeDecomposition with ov::pass::…

b82c950

…FakeQuantizeDecomposition

alvoron force-pushed the alvoron_fq_decomposition_arm branch from 527acf8 to b82c950 Compare February 26, 2025 13:05

github-actions bot removed the category: IE Tests OpenVINO Test: plugins and common label Feb 26, 2025

remove excessive thresholds

2385139

alvoron force-pushed the alvoron_fq_decomposition_arm branch from acc7c09 to 2385139 Compare February 26, 2025 16:59

alvoron added 2 commits February 26, 2025 20:33

revert ConvertSaturation support

d1b76b8

split lpt manager into common and arm

0c9aedb

alvoron force-pushed the alvoron_fq_decomposition_arm branch from f2f1b1b to 0c9aedb Compare February 27, 2025 10:18

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[CPU][ACL] LPT transformations are enabled + FQ decomposition #28981

[CPU][ACL] LPT transformations are enabled + FQ decomposition #28981

alvoron commented Feb 13, 2025

dmitry-gorokhov Feb 19, 2025

alvoron Feb 19, 2025

dmitry-gorokhov Feb 19, 2025

alvoron Feb 19, 2025

dmitry-gorokhov Feb 19, 2025

alvoron Feb 19, 2025

dmitry-gorokhov Feb 19, 2025

alvoron Feb 19, 2025

dmitry-gorokhov Feb 19, 2025

alvoron Feb 19, 2025

dmitry-gorokhov Feb 19, 2025

alvoron Feb 19, 2025

dmitry-gorokhov Feb 19, 2025

alvoron Feb 19, 2025

		} else if (ops_list_[i] == ov::intel_cpu::Type::Convert) {
		// do nothing

[CPU][ACL] LPT transformations are enabled + FQ decomposition #28981

Are you sure you want to change the base?

[CPU][ACL] LPT transformations are enabled + FQ decomposition #28981

Conversation

alvoron commented Feb 13, 2025

Details:

Tickets:

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment