NPUW: Use dynamic dispatch for AVX2 code

openvinotoolkit · Oct 14, 2024 · 1e6748b · 1e6748b
1 parent 81e141c
commit 1e6748b
Show file tree

Hide file tree

Showing 7 changed files with 1,671 additions and 1,584 deletions.
diff --git a/src/plugins/intel_npu/src/plugin/CMakeLists.txt b/src/plugins/intel_npu/src/plugin/CMakeLists.txt
@@ -66,9 +66,12 @@ target_include_directories(${TARGET_NAME}
         $<TARGET_PROPERTY:LevelZero::NPUExt,INTERFACE_INCLUDE_DIRECTORIES>
 )
 
-if(ENABLE_AVX2)
-    ov_avx2_optimization_flags(avx2_flags)
-    target_compile_options(${TARGET_NAME} PRIVATE "${avx2_flags}")
-endif()
+cross_compiled_file(${TARGET_NAME}
+        ARCH AVX2 ANY
+                    npuw/util_xarch.cpp
+        API         npuw/util_xarch.hpp
+        NAME        unpack unpack_scale unpack_scale_zp to_f16
+        NAMESPACE   ov::npuw::util::XARCH
+)
 
 ov_add_api_validator_post_build_step(TARGET ${NPU_PLUGIN_TARGET})
diff --git a/src/plugins/intel_npu/src/plugin/npuw/just_sync_infer_request.cpp b/src/plugins/intel_npu/src/plugin/npuw/just_sync_infer_request.cpp
@@ -19,6 +19,7 @@
 #include "openvino/runtime/make_tensor.hpp"
 #include "plugin.hpp"
 #include "util.hpp"
+#include "util_xarch.hpp"
 #include "weights_bank.hpp"
 
 ov::npuw::JustInferRequest::JustInferRequest(const std::shared_ptr<ov::npuw::CompiledModel>& compiled_model)
@@ -617,18 +618,18 @@ void ov::npuw::JustInferRequest::unpack_closure(std::size_t idx, RqPtr request)
 
         if (!comp_model_desc.scales.empty() && comp_model_desc.scales[cidx] && comp_model_desc.zerops[cidx]) {
             // Unpacking this weight requires scaling with zero points...
-            ov::npuw::util::unpack(ov::get_tensor_impl(closure),
-                                   ov::get_tensor_impl(comp_model_desc.zerops[cidx]),
-                                   ov::get_tensor_impl(comp_model_desc.scales[cidx]),
-                                   clparam);
+            ov::npuw::util::XARCH::unpack_scale_zp(ov::get_tensor_impl(closure),
+                                    ov::get_tensor_impl(comp_model_desc.zerops[cidx]),
+                                    ov::get_tensor_impl(comp_model_desc.scales[cidx]),
+                                    clparam);
         } else if (!comp_model_desc.scales.empty() && comp_model_desc.scales[cidx]) {
             // Unpacking this weight requires scaling
-            ov::npuw::util::unpack(ov::get_tensor_impl(closure),
+            ov::npuw::util::XARCH::unpack_scale(ov::get_tensor_impl(closure),
                                    ov::get_tensor_impl(comp_model_desc.scales[cidx]),
                                    clparam);
         } else {
             // Unpacking this weight doesn't require scaling
-            ov::npuw::util::unpack(ov::get_tensor_impl(closure), clparam);
+            ov::npuw::util::XARCH::unpack(ov::get_tensor_impl(closure), clparam);
         }
     }
 }