[CPU][ARM] Enable fast math in ACL deconvolution executor (#26615)

### Details: - ACL deconvolution `fast_math` option is enabled on `PERFORMANCE` mode. - This option enables fast math computation in ACL. In case this flag were set, ACL could dispatch the fastest implementation available which may introduce a drop of accuracy as well. - Accuracy testing on dataset subset highlights some deviations from reference values. Results are attached to the ticket. ### Tickets: - CVS-152534
openvinotoolkit · Oct 21, 2024 · d0056bd · d0056bd
1 parent 2be7e5f
commit d0056bd
Show file tree

Hide file tree

Showing 5 changed files with 18 additions and 2 deletions.
diff --git a/src/plugins/intel_cpu/src/config.cpp b/src/plugins/intel_cpu/src/config.cpp
@@ -400,6 +400,12 @@ void Config::readProperties(const ov::AnyMap& prop, const ModelType modelType) {
             inferencePrecision = ov::element::undefined;
         }
     }
+    // enable ACL fast math in PERFORMANCE mode
+#if defined(OV_CPU_WITH_ACL)
+    if (executionMode == ov::hint::ExecutionMode::PERFORMANCE) {
+        aclFastMath = true;
+    }
+#endif
     // disable dynamic quantization and kv quantization for best accuracy
     if (executionMode == ov::hint::ExecutionMode::ACCURACY) {
         if (!fcDynamicQuantizationGroupSizeSetExplicitly) {

diff --git a/src/plugins/intel_cpu/src/config.h b/src/plugins/intel_cpu/src/config.h
@@ -53,6 +53,9 @@ struct Config {
     uint64_t fcDynamicQuantizationGroupSize = 32;
     ov::element::Type kvCachePrecision = ov::element::f16;
     bool fcDynamicQuantizationGroupSizeSetExplicitly = false;
+#if defined(OV_CPU_WITH_ACL)
+    bool aclFastMath = false;
+#endif
 #if defined(OPENVINO_ARCH_X86_64)
     size_t rtCacheCapacity = 5000ul;
 #else

diff --git a/src/plugins/intel_cpu/src/nodes/deconv.cpp b/src/plugins/intel_cpu/src/nodes/deconv.cpp
@@ -219,6 +219,9 @@ Deconvolution::Deconvolution(const std::shared_ptr<ov::Node>& op,
     for (size_t i = 0; i < deconvAttrs.dilation.size(); i++) {
         deconvAttrs.kernel.push_back(weightDims[withGroups + 2 + i]);
     }
+#if defined(OV_CPU_WITH_ACL)
+    deconvAttrs.aclFastMath = context->getConfig().aclFastMath;
+#endif
 
     externOutShape = inputShapes.size() == 3;
     biasPort = externOutShape ? 3 : 2;

diff --git a/src/plugins/intel_cpu/src/nodes/executors/acl/acl_deconv.cpp b/src/plugins/intel_cpu/src/nodes/executors/acl/acl_deconv.cpp
@@ -99,7 +99,7 @@ bool AclDeconvExecutor::init(const DeconvAttrs& deconvAttrs,
 
     deconv = std::make_unique<arm_compute::NEDeconvolutionLayer>();
     configureThreadSafe([&] {
-        deconv->configure(&srcTensor, &weiTensor, deconvAttrs.withBiasesParam ? &biasTensor : nullptr, &dstTensor, deconv_info);
+        deconv->configure(&srcTensor, &weiTensor, deconvAttrs.withBiasesParam ? &biasTensor : nullptr, &dstTensor, deconv_info, deconvAttrs.aclFastMath);
     });
     return true;
 }
@@ -271,7 +271,8 @@ bool AclDeconvExecutorBuilder::customIsSupported(const DeconvAttrs &deconvAttrs,
                                                                                  &weiTensorInfo,
                                                                                  deconvAttrs.withBiasesParam ? &biasTensorInfo : nullptr,
                                                                                  &dstTensorInfo,
-                                                                                 deconv_info);
+                                                                                 deconv_info,
+                                                                                 deconvAttrs.aclFastMath);
         if (!status) {
             DEBUG_LOG("NEDeconvolutionLayer validation failed: ", status.error_description());
             return false;

diff --git a/src/plugins/intel_cpu/src/nodes/executors/deconv.hpp b/src/plugins/intel_cpu/src/nodes/executors/deconv.hpp
@@ -22,6 +22,9 @@ struct DeconvAttrs {
     std::vector<ptrdiff_t> paddingR;
     ov::CoordinateDiff outputPadding;
     bool withBiasesParam = false;
+#if defined(OV_CPU_WITH_ACL)
+    bool aclFastMath = false;
+#endif
 };
 
 class DeconvExecutor {