[CPU][ARM] ACL int8 Convolution executor shouldn't support non-i32 bias

src/plugins/intel_cpu/src/graph_optimizer.cpp

@@ -919,6 +919,10 @@ void GraphOptimizer::FuseFCAndTransposeOnWeights(Graph& graph) {
 
 void GraphOptimizer::FuseConvolutionAndZeroPoints(Graph& graph) {
     const auto& graphNodes = graph.GetNodes();
+// zero points fusing is skipped on ARM platforms because oneDNN is not involved into int8 convolution inference
+#if defined(OPENVINO_ARCH_ARM) || defined(OPENVINO_ARCH_ARM64)
+    return;
+#endif
 
     auto isSuitableConvNode = [](const NodePtr& node) {
         bool retVal = false;

src/plugins/intel_cpu/src/nodes/executors/acl/acl_conv.cpp

@@ -94,6 +94,9 @@ bool ACLConvolutionExecutor::supports(const ConvConfig& config) {
                        config.descs.at(ARG_WEI)->getPrecision() == ov::element::i8;
 
     VERIFY(isQuantized, UNSUPPORTED_SRC_PRECISIONS);
+    if (config.attrs.withBias) {
+        VERIFY(config.descs.at(ARG_BIAS)->getPrecision() == ov::element::i32, UNSUPPORTED_BIAS_PRECISIONS);
+    }
     VERIFY(config.attrs.postOps.size() <= 1U, UNSUPPORTED_BY_EXECUTOR);
 
     return true;

src/plugins/intel_cpu/src/nodes/executors/convolution_implementations.cpp

@@ -59,20 +59,23 @@ static const TypeMapping dnnlConvTypeMapping {
     {{_f32, _half_float | _i8, _any, _any},                   {bypass(), bypass(), use<0>(), use<0>()}},
     {{_bf16, _f16, _any, _any},                               {bypass(), bypass(), use<0>(), use<0>()}},
     {{_f16, _bf16, _any, _any},                               {bypass(), bypass(), use<0>(), use<0>()}},
-    // quantization configuration
+    // quantization configuration is not applicable for ARM
+    // because there is the dedicated low-precision implementation for ARM
+#if !defined(OPENVINO_ARCH_ARM64) && !defined(OPENVINO_ARCH_ARM)
     // int8 conv does not support f16 output and bias
     {{_u8 | _i8, _i8,  _quant |_bf16 | _f32 | _i32 | _dynamic,  _quant | _bf16 | _f32 | _i32 | _dynamic}, {bypass(), bypass(), bypass(),  bypass()}},
     {{_u8 | _i8, _i8, _f16, _u8 | _i8 | _i32 | _bf16 | _f32}, {bypass(), bypass(), just<f32>(), bypass()}},
     {{_u8 | _i8, _i8, _any, _any}, {bypass(), bypass(), just<f32>(), just<f32>()}},
+#endif
     // @todo should we fallback to FPXX instead of _f32?
     {{_any, _any, _any, _any},                                {just<f32>(), just<f32>(), just<f32>(), just<f32>()}},
     // @todo explicitly cover configuration limitations for oneDNN on ARM
 };
 
 static const TypeMapping aclLowpConvTypeMapping {
     // {src, wei, bia, dst}                            pt<src, wei, bias, dst>
-    {{_u8, _u8 | _i8, _any, _u8},                      {bypass(), bypass(), just<i32>(), bypass()}},
-    {{_i8, _i8, _any, _i8},                            {bypass(), bypass(), just<i32>(), bypass()}},
+    {{_u8, _u8 | _i8, _i32 | _dynamic, _u8},                      {bypass(), bypass(), bypass(), bypass()}},
+    {{_i8, _i8, _i32 | _dynamic, _i8},                            {bypass(), bypass(), bypass(), bypass()}},
 };
 // clang-format on
 struct CreateOptimalConfigDefault {
@@ -245,13 +248,23 @@ const std::vector<ExecutorImplementation<ConvAttrs>>& getImplementations() {
             AcceptsAnyShape<ConvAttrs>,
             CreateDnnlDefault<DnnlConvolutionPrimitive, ConvAttrs>{}
             )
+        OV_CPU_INSTANCE_ACL(
+            "convolution_acl_lowp", ExecutorType::Acl, OperationType::Convolution,
+            // supports
+            [](const ConvConfig& config, [[maybe_unused]] const MemoryFormatFilter& memoryFormatFilter) -> bool {
+                VERIFY(ACLConvolutionExecutor::supports(config), UNSUPPORTED_BY_EXECUTOR);
+                return true;
+            },
+            CreateOptimalConfigAclLowp{{LayoutType::ncsp, LayoutType::ncsp, LayoutType::ncsp, LayoutType::ncsp}},
+            AcceptsAnyShape<ConvAttrs>,
+            CreateDefault<ACLConvolutionExecutor, ConvAttrs>{}
+            )
         OV_CPU_INSTANCE_ACL(
             "convolution_dnnl_nspc_nspc_unconditional_acl", ExecutorType::Dnnl, OperationType::Convolution,
             // supports
             [](const ConvConfig& config, const MemoryFormatFilter& memoryFormatFilter) -> bool {
                 VERIFY(MatchesMemoryFormatFilter(config.descs, LayoutConfig{LayoutType::nspc, LayoutType::ncsp, LayoutType::nspc, LayoutType::nspc},
                                                  memoryFormatFilter, dnnlConvolutionMappingNotation), MEMORY_FORMAT_MISMATCH);
-                VERIFY(!isQuantized(config), UNSUPPORTED_SRC_PRECISIONS);
                 return true;
             },
             CreateOptimalConfigDefault{{LayoutType::nspc, LayoutType::ncsp, LayoutType::nspc, LayoutType::nspc}},
@@ -274,17 +287,6 @@ const std::vector<ExecutorImplementation<ConvAttrs>>& getImplementations() {
             AcceptsAnyShape<ConvAttrs>,
             CreateDnnlDefault<DnnlConvolutionPrimitive, ConvAttrs>{}
             )
-        OV_CPU_INSTANCE_ACL(
-            "convolution_acl_lowp", ExecutorType::Acl, OperationType::Convolution,
-            // supports
-            [](const ConvConfig& config, [[maybe_unused]] const MemoryFormatFilter& memoryFormatFilter) -> bool {
-                VERIFY(ACLConvolutionExecutor::supports(config), UNSUPPORTED_BY_EXECUTOR);
-                return true;
-            },
-            CreateOptimalConfigAclLowp{{LayoutType::ncsp, LayoutType::ncsp, LayoutType::ncsp, LayoutType::ncsp}},
-            AcceptsAnyShape<ConvAttrs>,
-            CreateDefault<ACLConvolutionExecutor, ConvAttrs>{}
-            )
     };
 
     return convolutionImplementations;

src/plugins/intel_cpu/src/nodes/executors/debug_messages.hpp

@@ -11,6 +11,7 @@
 #define UNSUPPORTED_TYPE_OF_POSTOPS          " the type of post ops is not supported"
 #define UNSUPPORTED_SRC_PRECISIONS           " unsupported src precisions"
 #define UNSUPPORTED_WEI_PRECISIONS           " unsupported wei precisions"
+#define UNSUPPORTED_BIAS_PRECISIONS          " unsupported bias precisions"
 #define UNSUPPORTED_DST_PRECISIONS           " unsupported dst precisions"
 #define UNSUPPORTED_ISA                      " unsupported isa"
 #define UNSUPPORTED_SRC_RANK                 " unsupported src rank"