NPUW: Fix for dynamic dispatch for AVX2 code

openvinotoolkit · Oct 16, 2024 · 7060669 · 7060669
1 parent 6eca67c
commit 7060669
Show file tree

Hide file tree

Showing 9 changed files with 1,646 additions and 1,614 deletions.
diff --git a/src/plugins/intel_npu/src/plugin/CMakeLists.txt b/src/plugins/intel_npu/src/plugin/CMakeLists.txt
@@ -70,7 +70,7 @@ 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
+        NAME        unpack_i4i8 unpack_u4i8 unpack_i4f16 unpack_i4f16_scale unpack_i4f16_z unpack_u4f16 unpack_u4f16_scale_zp unpack_u4f16_asymm_zp unpack_u4f16_z unpack_u4f32 unpack_i8f16 unpack_i8f16_scale unpack_u8f16 to_f16
         NAMESPACE   ov::npuw::util::XARCH
 )
 

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
@@ -18,7 +18,6 @@
 #include "openvino/runtime/iasync_infer_request.hpp"
 #include "plugin.hpp"
 #include "util.hpp"
-#include "util_xarch.hpp"
 #include "weights_bank.hpp"
 
 ov::npuw::MemAccessSim::MemAccessSim(const std::shared_ptr<ov::npuw::CompiledModel>& compiled_model) {
@@ -777,18 +776,18 @@ 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::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);
+            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);
         } else if (!comp_model_desc.scales.empty() && comp_model_desc.scales[cidx]) {
             // Unpacking this weight requires scaling
-            ov::npuw::util::XARCH::unpack_scale(ov::get_tensor_impl(closure),
-                                                ov::get_tensor_impl(comp_model_desc.scales[cidx]),
-                                                clparam);
+            ov::npuw::util::unpack(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::XARCH::unpack(ov::get_tensor_impl(closure), clparam);
+            ov::npuw::util::unpack(ov::get_tensor_impl(closure), clparam);
         }
     }
 }

diff --git a/src/plugins/intel_npu/src/plugin/npuw/lazy_tensor.cpp b/src/plugins/intel_npu/src/plugin/npuw/lazy_tensor.cpp
@@ -207,9 +207,9 @@ ov::Tensor LazyTensorImpl::eval() const {
             const auto& ts = cs.get_orig_tensor();
             ov::Tensor dst(type, shape);
             if (tw && tz && ts) {
-                ov::npuw::util::XARCH::unpack_scale_zp(gti(tw), gti(tz), gti(ts), gti(dst));
+                ov::npuw::util::unpack(gti(tw), gti(tz), gti(ts), gti(dst));
             } else if (tw && ts) {
-                ov::npuw::util::XARCH::unpack_scale(gti(tw), gti(ts), gti(dst));
+                ov::npuw::util::unpack(gti(tw), gti(ts), gti(dst));
             } else {
                 NPUW_ASSERT(false && "Unsupported combination");
             }
@@ -224,7 +224,7 @@ ov::Tensor LazyTensorImpl::eval() const {
     case TransformType::PERMUTE:
         return ov::npuw::util::permute(m_parent->eval(), std::get<std::vector<std::size_t>>(m_transform.second));
     case TransformType::CONVERT:
-        return ov::npuw::util::XARCH::to_f16(m_parent->eval());
+        return ov::npuw::util::to_f16(m_parent->eval());
     default:
         NPUW_ASSERT(false);
     }

diff --git a/src/plugins/intel_npu/src/plugin/npuw/lazy_tensor.hpp b/src/plugins/intel_npu/src/plugin/npuw/lazy_tensor.hpp
@@ -14,7 +14,6 @@
 #include "openvino/runtime/make_tensor.hpp"
 #include "openvino/runtime/tensor.hpp"
 #include "util.hpp"
-#include "util_xarch.hpp"
 
 namespace ov {
 namespace npuw {