From 027d281c9957690bf02f2cdf9b3e74cf47b50dea Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?D=C3=A9nes=20Tarj=C3=A1n?= <Denes.Tarjan@arm.com>
Date: Mon, 18 Mar 2024 18:20:08 +0000
Subject: [PATCH 1/2] Separate morphology sve2/sme2 functions

---
 intrinsiccv/include/intrinsiccv/config.h.in   | 10 +++-
 .../include/intrinsiccv/containers/stack.h    |  4 +-
 .../intrinsiccv/morphology/workspace.h        | 54 ++++++-------------
 intrinsiccv/include/intrinsiccv/operations.h  | 10 ++--
 intrinsiccv/include/intrinsiccv/sve2.h        |  6 ++-
 intrinsiccv/include/intrinsiccv/traits.h      |  7 +--
 intrinsiccv/include/intrinsiccv/types.h       |  4 +-
 intrinsiccv/include/intrinsiccv/utils.h       | 30 ++++++-----
 .../include/intrinsiccv/workspace/borders.h   |  4 +-
 .../include/intrinsiccv/workspace/separable.h |  4 +-
 intrinsiccv/src/arithmetics/absdiff_sme2.cpp  |  4 ++
 intrinsiccv/src/arithmetics/add_sme2.cpp      |  4 ++
 intrinsiccv/src/arithmetics/sub_sme2.cpp      |  4 ++
 .../src/morphology/morphology_neon.cpp        |  3 ++
 intrinsiccv/src/morphology/morphology_sc.h    | 53 +++++++++++++-----
 .../src/morphology/morphology_sme2.cpp        |  8 +--
 .../src/morphology/morphology_sve2.cpp        |  6 ++-
 scripts/ci-opencv.sh                          |  2 +-
 18 files changed, 124 insertions(+), 93 deletions(-)

diff --git a/intrinsiccv/include/intrinsiccv/config.h.in b/intrinsiccv/include/intrinsiccv/config.h.in
index 710049712..ee791b73c 100644
--- a/intrinsiccv/include/intrinsiccv/config.h.in
+++ b/intrinsiccv/include/intrinsiccv/config.h.in
@@ -17,6 +17,8 @@
 
 #cmakedefine01 INTRINSICCV_CANNY_ALGORITHM_CONFORM_OPENCV
 
+#cmakedefine01 INTRINSICCV_COMPILER_SUPPORTS_SME2
+
 // Set to '1' if compiling NEON code paths, otherwise it is set to '0'.
 #ifndef INTRINSICCV_TARGET_NEON
 #define INTRINSICCV_TARGET_NEON 0
@@ -47,10 +49,16 @@
 #define INTRINSICCV_ASSUME_128BIT_SVE2 0
 #define INTRINSICCV_TARGET_FN_ATTRS INTRINSICCV_ATTR_SECTION(".text.sme2")
 #define INTRINSICCV_LOCALLY_STREAMING __arm_locally_streaming
-#define INTRINSICCV_STREAMING_COMPATIBLE __arm_streaming_compatible
 #else
 #define INTRINSICCV_LOCALLY_STREAMING
+#endif
+
+#if ((INTRINSICCV_TARGET_SVE2 || INTRINSICCV_TARGET_SME2) && INTRINSICCV_COMPILER_SUPPORTS_SME2)
+#define INTRINSICCV_STREAMING_COMPATIBLE __arm_streaming_compatible
+#define INTRINSICCV_SC_NAMESPACE intrinsiccv::sve2
+#else
 #define INTRINSICCV_STREAMING_COMPATIBLE
+#define INTRINSICCV_SC_NAMESPACE intrinsiccv
 #endif
 
 #ifdef __linux__
diff --git a/intrinsiccv/include/intrinsiccv/containers/stack.h b/intrinsiccv/include/intrinsiccv/containers/stack.h
index 52c346434..f1aac1429 100644
--- a/intrinsiccv/include/intrinsiccv/containers/stack.h
+++ b/intrinsiccv/include/intrinsiccv/containers/stack.h
@@ -13,7 +13,7 @@
 #include "intrinsiccv/traits.h"
 #include "intrinsiccv/utils.h"
 
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 // A simple last-in first-out container.
 template <typename T>
@@ -171,6 +171,6 @@ class Stack final : public NonCopyable {
   Block *first_block_;
 };  // end of class Stack<T>
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_CONTAINERS_STACK_H
diff --git a/intrinsiccv/include/intrinsiccv/morphology/workspace.h b/intrinsiccv/include/intrinsiccv/morphology/workspace.h
index 022e1ab72..e6c89977a 100644
--- a/intrinsiccv/include/intrinsiccv/morphology/workspace.h
+++ b/intrinsiccv/include/intrinsiccv/morphology/workspace.h
@@ -13,11 +13,7 @@
 #include "intrinsiccv/intrinsiccv.h"
 #include "intrinsiccv/types.h"
 
-#if INTRINSICCV_TARGET_SME2
-#include "intrinsiccv/sve2.h"
-#endif
-
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 // Forward declarations.
 class MorphologyWorkspace;
@@ -55,6 +51,18 @@ class MorphologyWorkspace final {
     }
   }
 
+  template <typename T>
+  class CopyDataMemcpy {
+   public:
+    constexpr void operator()(Rows<const T> src_rows, Rows<T> dst_rows,
+                              size_t length) const
+        INTRINSICCV_STREAMING_COMPATIBLE {
+      std::memcpy(static_cast<void *>(&dst_rows[0]),
+                  static_cast<const void *>(&src_rows[0]),
+                  length * sizeof(T) * dst_rows.channels());
+    }
+  };
+
   // MorphologyWorkspace is only constructible with create().
   MorphologyWorkspace() = delete;
 
@@ -160,6 +168,7 @@ class MorphologyWorkspace final {
                O operation) INTRINSICCV_STREAMING_COMPATIBLE {
     using S = typename O::SourceType;
     using B = typename O::BufferType;
+    typename O::CopyData copy_data{};
 
     if (INTRINSICCV_UNLIKELY(rect.width() == 0 || rect.height() == 0)) {
       return;
@@ -326,39 +335,6 @@ class MorphologyWorkspace final {
     return height;
   }
 
-#if INTRINSICCV_TARGET_SME2
-  template <typename ScalarType>
-  class CopyOperation final : public UnrollTwice {
-   public:
-    using ContextType = sve2::Context;
-    using VecTraits = sve2::VecTraits<ScalarType>;
-    using VectorType = typename VecTraits::VectorType;
-
-    VectorType vector_path(ContextType,
-                           VectorType src) INTRINSICCV_STREAMING_COMPATIBLE {
-      return src;
-    }
-  };  // end of class CopyOperation<ScalarType>
-#endif
-
-  template <typename T>
-  void copy_data(Rows<const T> src_rows, Rows<T> dst_rows,
-                 size_t length) INTRINSICCV_STREAMING_COMPATIBLE {
-#if INTRINSICCV_TARGET_SME2
-    // 'sve2::apply_operation_by_rows' can only handle one channel well
-    // so width must be multiplied in order to copy all the data
-    Rectangle rect{length * dst_rows.channels(), std::size_t{1}};
-    Rows<const T> src_1ch{&src_rows[0], src_rows.stride(), 1};
-    Rows<T> dst_1ch{&dst_rows[0], dst_rows.stride(), 1};
-    CopyOperation<T> operation;
-    sve2::apply_operation_by_rows(operation, rect, src_1ch, dst_1ch);
-#else
-    std::memcpy(static_cast<void *>(&dst_rows[0]),
-                static_cast<const void *>(&src_rows[0]),
-                length * sizeof(T) * dst_rows.channels());
-#endif
-  }
-
   template <typename T, typename BorderType>
   void make_constant_border(Rows<T> dst_rows, size_t dst_index, size_t count,
                             BorderType value) INTRINSICCV_STREAMING_COMPATIBLE {
@@ -417,6 +393,6 @@ class MorphologyWorkspace final {
   uint8_t data_[0] INTRINSICCV_ATTR_ALIGNED(sizeof(void *));
 };  // end of class MorphologyWorkspace
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_MORPHOLOGY_WORKSPACE_H
diff --git a/intrinsiccv/include/intrinsiccv/operations.h b/intrinsiccv/include/intrinsiccv/operations.h
index 7dc888a47..1a8943ab2 100644
--- a/intrinsiccv/include/intrinsiccv/operations.h
+++ b/intrinsiccv/include/intrinsiccv/operations.h
@@ -11,7 +11,7 @@
 #include "intrinsiccv/traits.h"
 #include "intrinsiccv/types.h"
 
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 // Base for classes that change the behaviour or extend the interface of an
 // operation.
@@ -110,19 +110,19 @@ class OperationBase {
 
   // Returns true if the innermost operation is unrolled twice, otherwise false.
   static constexpr bool is_unrolled_twice() INTRINSICCV_STREAMING_COMPATIBLE {
-    return intrinsiccv::is_unrolled_twice<
+    return ::INTRINSICCV_SC_NAMESPACE::is_unrolled_twice<
         concrete_operation_type_t<OperationType>>;
   }
 
   // Returns true if the innermost operation is unrolled once, otherwise false.
   static constexpr bool is_unrolled_once() INTRINSICCV_STREAMING_COMPATIBLE {
-    return intrinsiccv::is_unrolled_once<
+    return ::INTRINSICCV_SC_NAMESPACE::is_unrolled_once<
         concrete_operation_type_t<OperationType>>;
   }
 
   // Returns true if the innermost operation uses tail path, otherwise false.
   static constexpr bool uses_tail_path() INTRINSICCV_STREAMING_COMPATIBLE {
-    return intrinsiccv::uses_tail_path<
+    return ::INTRINSICCV_SC_NAMESPACE::uses_tail_path<
         concrete_operation_type_t<OperationType>>;
   }
 
@@ -1270,6 +1270,6 @@ class RemoveContextAdapter : public OperationBase<OperationType> {
   }
 };  // end of class RemoveContextAdapter<OperationType>
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_OPERATIONS_H
diff --git a/intrinsiccv/include/intrinsiccv/sve2.h b/intrinsiccv/include/intrinsiccv/sve2.h
index 8193c9b38..4d43d86a2 100644
--- a/intrinsiccv/include/intrinsiccv/sve2.h
+++ b/intrinsiccv/include/intrinsiccv/sve2.h
@@ -517,7 +517,8 @@ class SeparableFilter<FilterType, 3UL> {
   using SourceVectorType = typename SourceVecTraits::VectorType;
   using BufferVecTraits = typename sve2::VecTraits<BufferType>;
   using BufferVectorType = typename BufferVecTraits::VectorType;
-  using BorderInfoType = typename intrinsiccv::FixedBorderInfo3x3<SourceType>;
+  using BorderInfoType =
+      typename ::INTRINSICCV_SC_NAMESPACE::FixedBorderInfo3x3<SourceType>;
   using BorderType = FixedBorderType;
   using BorderOffsets = typename BorderInfoType::Offsets;
 
@@ -653,7 +654,8 @@ class SeparableFilter<FilterType, 5UL> {
   using SourceVectorType = typename SourceVecTraits::VectorType;
   using BufferVecTraits = typename sve2::VecTraits<BufferType>;
   using BufferVectorType = typename BufferVecTraits::VectorType;
-  using BorderInfoType = typename intrinsiccv::FixedBorderInfo5x5<SourceType>;
+  using BorderInfoType =
+      typename ::INTRINSICCV_SC_NAMESPACE::FixedBorderInfo5x5<SourceType>;
   using BorderType = FixedBorderType;
   using BorderOffsets = typename BorderInfoType::Offsets;
 
diff --git a/intrinsiccv/include/intrinsiccv/traits.h b/intrinsiccv/include/intrinsiccv/traits.h
index c5d92a77e..cf77cdf0b 100644
--- a/intrinsiccv/include/intrinsiccv/traits.h
+++ b/intrinsiccv/include/intrinsiccv/traits.h
@@ -9,7 +9,7 @@
 
 #include "intrinsiccv/config.h"
 
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 // An empty class.
 class Monostate {};
@@ -17,7 +17,8 @@ class Monostate {};
 template <typename FnType>
 class remove_streaming_compatible;
 
-#if INTRINSICCV_TARGET_SME2
+#if ((INTRINSICCV_TARGET_SVE2 || INTRINSICCV_TARGET_SME2) && \
+     INTRINSICCV_COMPILER_SUPPORTS_SME2)
 template <typename Ret, typename Impl, typename... Args>
 class remove_streaming_compatible<Ret (Impl::*)(Args...)
                                       INTRINSICCV_STREAMING_COMPATIBLE> {
@@ -191,6 +192,6 @@ class NonCopyable {
   NonCopyable &operator=(const NonCopyable &) = delete;
 };  // end of class NonCopyable
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_TYPE_TRAITS_H
diff --git a/intrinsiccv/include/intrinsiccv/types.h b/intrinsiccv/include/intrinsiccv/types.h
index 41510e109..5c30a4575 100644
--- a/intrinsiccv/include/intrinsiccv/types.h
+++ b/intrinsiccv/include/intrinsiccv/types.h
@@ -13,7 +13,7 @@
 #include "intrinsiccv/ctypes.h"
 #include "intrinsiccv/utils.h"
 
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 // Represents a point on a 2D plane.
 class Point final {
@@ -668,6 +668,6 @@ class RowsOverUniquePtr {
   std::unique_ptr<T[]> data_;
 };
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_TYPES_H
diff --git a/intrinsiccv/include/intrinsiccv/utils.h b/intrinsiccv/include/intrinsiccv/utils.h
index 54765279a..0a15fd6ad 100644
--- a/intrinsiccv/include/intrinsiccv/utils.h
+++ b/intrinsiccv/include/intrinsiccv/utils.h
@@ -13,7 +13,7 @@
 #include "intrinsiccv/config.h"
 #include "intrinsiccv/ctypes.h"
 
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 // Saturating cast from signed to unsigned type.
 template <typename S, typename U,
@@ -402,20 +402,22 @@ std::enable_if_t<alignof(T) == 1, intrinsiccv_error_t> check_pointer_and_stride(
   return INTRINSICCV_OK;
 }
 
-#define CHECK_POINTER_AND_STRIDE(pointer, stride)        \
-  do {                                                   \
-    if (intrinsiccv_error_t ptr_stride_err =             \
-            check_pointer_and_stride(pointer, stride)) { \
-      return ptr_stride_err;                             \
-    }                                                    \
+#define CHECK_POINTER_AND_STRIDE(pointer, stride)                           \
+  do {                                                                      \
+    if (intrinsiccv_error_t ptr_stride_err =                                \
+            ::INTRINSICCV_SC_NAMESPACE::check_pointer_and_stride(pointer,   \
+                                                                 stride)) { \
+      return ptr_stride_err;                                                \
+    }                                                                       \
   } while (false)
 
-#define MAKE_POINTER_CHECK_ALIGNMENT(ElementType, name, from)            \
-  if constexpr (alignof(ElementType) > 1) {                              \
-    if (is_misaligned<ElementType>(reinterpret_cast<uintptr_t>(from))) { \
-      return INTRINSICCV_ERROR_ALIGNMENT;                                \
-    }                                                                    \
-  }                                                                      \
+#define MAKE_POINTER_CHECK_ALIGNMENT(ElementType, name, from)   \
+  if constexpr (alignof(ElementType) > 1) {                     \
+    if (::INTRINSICCV_SC_NAMESPACE::is_misaligned<ElementType>( \
+            reinterpret_cast<uintptr_t>(from))) {               \
+      return INTRINSICCV_ERROR_ALIGNMENT;                       \
+    }                                                           \
+  }                                                             \
   ElementType *name = reinterpret_cast<ElementType *>(from)
 
 // Check whether the image size is acceptable by limiting it.
@@ -434,6 +436,6 @@ std::enable_if_t<alignof(T) == 1, intrinsiccv_error_t> check_pointer_and_stride(
 // Check whether the rectangle size is acceptable by limiting it.
 #define CHECK_RECTANGLE_SIZE(rect) CHECK_IMAGE_SIZE(rect.width, rect.height)
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_UTILS_H
diff --git a/intrinsiccv/include/intrinsiccv/workspace/borders.h b/intrinsiccv/include/intrinsiccv/workspace/borders.h
index 9f72b9c55..c225ec8ab 100644
--- a/intrinsiccv/include/intrinsiccv/workspace/borders.h
+++ b/intrinsiccv/include/intrinsiccv/workspace/borders.h
@@ -9,7 +9,7 @@
 
 #include "intrinsiccv/intrinsiccv.h"
 
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 enum class FixedBorderType {
   REPLICATE,
@@ -278,6 +278,6 @@ using FixedBorderInfo3x3 = FixedBorderInfo<T, 3UL>;
 template <typename T>
 using FixedBorderInfo5x5 = FixedBorderInfo<T, 5UL>;
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_WORKSPACE_BORDERS_H
diff --git a/intrinsiccv/include/intrinsiccv/workspace/separable.h b/intrinsiccv/include/intrinsiccv/workspace/separable.h
index 599e1186b..9a32c2f7d 100644
--- a/intrinsiccv/include/intrinsiccv/workspace/separable.h
+++ b/intrinsiccv/include/intrinsiccv/workspace/separable.h
@@ -12,7 +12,7 @@
 #include "intrinsiccv/intrinsiccv.h"
 #include "intrinsiccv/types.h"
 
-namespace intrinsiccv {
+namespace INTRINSICCV_SC_NAMESPACE {
 
 // Forward declarations.
 class SeparableFilterWorkspace;
@@ -215,6 +215,6 @@ class SeparableFilterWorkspace final {
   uint8_t data_[0] INTRINSICCV_ATTR_ALIGNED(kAlignment);
 };  // end of class SeparableFilterWorkspace
 
-}  // namespace intrinsiccv
+}  // namespace INTRINSICCV_SC_NAMESPACE
 
 #endif  // INTRINSICCV_WORKSPACE_SEPARABLE_H
diff --git a/intrinsiccv/src/arithmetics/absdiff_sme2.cpp b/intrinsiccv/src/arithmetics/absdiff_sme2.cpp
index 29895f1a0..85f3e00ec 100644
--- a/intrinsiccv/src/arithmetics/absdiff_sme2.cpp
+++ b/intrinsiccv/src/arithmetics/absdiff_sme2.cpp
@@ -7,6 +7,10 @@
 
 namespace intrinsiccv::sme2 {
 
+using sve2::Rectangle;
+using sve2::Rows;
+using sve2::UnrollTwice;
+
 template <typename ScalarType, typename VectorType,
           std::enable_if_t<std::is_signed<ScalarType>::value, bool> = true>
 VectorType vector_path_impl(svbool_t pg, VectorType src_a,
diff --git a/intrinsiccv/src/arithmetics/add_sme2.cpp b/intrinsiccv/src/arithmetics/add_sme2.cpp
index 11cc62908..2063d5366 100644
--- a/intrinsiccv/src/arithmetics/add_sme2.cpp
+++ b/intrinsiccv/src/arithmetics/add_sme2.cpp
@@ -9,6 +9,10 @@
 
 namespace intrinsiccv::sme2 {
 
+using sve2::Rectangle;
+using sve2::Rows;
+using sve2::UnrollTwice;
+
 template <typename ScalarType>
 class SaturatingAdd final : public UnrollTwice {
  public:
diff --git a/intrinsiccv/src/arithmetics/sub_sme2.cpp b/intrinsiccv/src/arithmetics/sub_sme2.cpp
index 312b512f6..1b756d482 100644
--- a/intrinsiccv/src/arithmetics/sub_sme2.cpp
+++ b/intrinsiccv/src/arithmetics/sub_sme2.cpp
@@ -9,6 +9,10 @@
 
 namespace intrinsiccv::sme2 {
 
+using sve2::Rectangle;
+using sve2::Rows;
+using sve2::UnrollTwice;
+
 template <typename ScalarType>
 class SaturatingSub final : public UnrollTwice {
  public:
diff --git a/intrinsiccv/src/morphology/morphology_neon.cpp b/intrinsiccv/src/morphology/morphology_neon.cpp
index 98aac273b..9bc91c368 100644
--- a/intrinsiccv/src/morphology/morphology_neon.cpp
+++ b/intrinsiccv/src/morphology/morphology_neon.cpp
@@ -467,6 +467,7 @@ class DilateOperation final {
   using SourceType = ScalarType;
   using BufferType = ScalarType;
   using DestinationType = ScalarType;
+  using CopyData = MorphologyWorkspace::CopyDataMemcpy<ScalarType>;
 
   explicit DilateOperation(Rectangle kernel) : kernel_{kernel} {}
 
@@ -537,6 +538,7 @@ class ErodeOperation final {
   using SourceType = ScalarType;
   using BufferType = ScalarType;
   using DestinationType = ScalarType;
+  using CopyData = MorphologyWorkspace::CopyDataMemcpy<ScalarType>;
 
   explicit ErodeOperation(Rectangle kernel) : kernel_{kernel} {}
 
@@ -566,6 +568,7 @@ intrinsiccv_error_t erode(const T *src, size_t src_stride, T *dst,
   CHECK_IMAGE_SIZE(width, height);
 
   auto *workspace = reinterpret_cast<MorphologyWorkspace *>(context);
+
   if (workspace->type_size() != sizeof(T)) {
     return INTRINSICCV_ERROR_CONTEXT_MISMATCH;
   }
diff --git a/intrinsiccv/src/morphology/morphology_sc.h b/intrinsiccv/src/morphology/morphology_sc.h
index 437970404..9294f218f 100644
--- a/intrinsiccv/src/morphology/morphology_sc.h
+++ b/intrinsiccv/src/morphology/morphology_sc.h
@@ -15,6 +15,33 @@
 
 namespace intrinsiccv::sve2 {
 
+template <typename T>
+class CopyDataSVE2 {
+  class CopyOperation final : public UnrollTwice {
+   public:
+    using ContextType = sve2::Context;
+    using VecTraits = sve2::VecTraits<T>;
+    using VectorType = typename VecTraits::VectorType;
+
+    VectorType vector_path(ContextType,
+                           VectorType src) INTRINSICCV_STREAMING_COMPATIBLE {
+      return src;
+    }
+  };  // end of class CopyOperation
+
+ public:
+  void operator()(Rows<const T> src_rows, Rows<T> dst_rows,
+                  size_t length) const INTRINSICCV_STREAMING_COMPATIBLE {
+    // 'apply_operation_by_rows' can only handle one channel well
+    // so width must be multiplied in order to copy all the data
+    Rectangle rect{length * dst_rows.channels(), std::size_t{1}};
+    Rows<const T> src_1ch{&src_rows[0], src_rows.stride(), 1};
+    Rows<T> dst_1ch{&dst_rows[0], dst_rows.stride(), 1};
+    CopyOperation op{};
+    apply_operation_by_rows(op, rect, src_1ch, dst_1ch);
+  }
+};
+
 template <typename ScalarType, typename O>
 class VerticalOp final {
  public:
@@ -419,12 +446,13 @@ using HorizontalMin = HorizontalOp<T, Min<T>>;
 template <typename T>
 using HorizontalMax = HorizontalOp<T, Max<T>>;
 
-template <typename ScalarType>
+template <typename ScalarType, typename CopyDataOperation>
 class DilateOperation final {
  public:
   using SourceType = ScalarType;
   using BufferType = ScalarType;
   using DestinationType = ScalarType;
+  using CopyData = CopyDataOperation;
 
   explicit DilateOperation(Rectangle kernel) INTRINSICCV_STREAMING_COMPATIBLE
       : kernel_{kernel} {}
@@ -432,22 +460,20 @@ class DilateOperation final {
   void process_horizontal(Rectangle rect, Rows<const SourceType> src_rows,
                           Rows<BufferType> dst_rows)
       INTRINSICCV_STREAMING_COMPATIBLE {
-    sve2::HorizontalMax<ScalarType>{rect, kernel_}.process_rows(src_rows,
-                                                                dst_rows);
+    HorizontalMax<ScalarType>{rect, kernel_}.process_rows(src_rows, dst_rows);
   }
 
   void process_vertical(Rectangle rect, IndirectRows<BufferType> src_rows,
                         Rows<DestinationType> dst_rows)
       INTRINSICCV_STREAMING_COMPATIBLE {
-    sve2::VerticalMax<ScalarType>{rect, kernel_}.process_rows(src_rows,
-                                                              dst_rows);
+    VerticalMax<ScalarType>{rect, kernel_}.process_rows(src_rows, dst_rows);
   }
 
  private:
   Rectangle kernel_;
 };  // end of class DilateOperation<ScalarType>
 
-template <typename T>
+template <typename T, typename CopyOperation>
 static intrinsiccv_error_t dilate_sc(const T *src, size_t src_stride, T *dst,
                                      size_t dst_stride, size_t width,
                                      size_t height,
@@ -484,7 +510,7 @@ static intrinsiccv_error_t dilate_sc(const T *src, size_t src_stride, T *dst,
   Rows<const T> current_src_rows = src_rows;
   Rows<T> current_dst_rows = dst_rows;
   for (size_t iteration = 0; iteration < workspace->iterations(); ++iteration) {
-    DilateOperation<T> operation{kernel};
+    DilateOperation<T, CopyOperation> operation{kernel};
     workspace->process(rect, current_src_rows, current_dst_rows, margin, border,
                        workspace->border_type(), operation);
     // Update source for the next iteration.
@@ -494,12 +520,13 @@ static intrinsiccv_error_t dilate_sc(const T *src, size_t src_stride, T *dst,
 }
 
 // Helper structure for erode.
-template <typename ScalarType>
+template <typename ScalarType, typename CopyDataOperation>
 class ErodeOperation final {
  public:
   using SourceType = ScalarType;
   using BufferType = ScalarType;
   using DestinationType = ScalarType;
+  using CopyData = CopyDataOperation;
 
   explicit ErodeOperation(Rectangle kernel) INTRINSICCV_STREAMING_COMPATIBLE
       : kernel_{kernel} {}
@@ -507,22 +534,20 @@ class ErodeOperation final {
   void process_horizontal(Rectangle rect, Rows<const SourceType> src_rows,
                           Rows<BufferType> dst_rows)
       INTRINSICCV_STREAMING_COMPATIBLE {
-    sve2::HorizontalMin<ScalarType>{rect, kernel_}.process_rows(src_rows,
-                                                                dst_rows);
+    HorizontalMin<ScalarType>{rect, kernel_}.process_rows(src_rows, dst_rows);
   }
 
   void process_vertical(Rectangle rect, IndirectRows<BufferType> src_rows,
                         Rows<DestinationType> dst_rows)
       INTRINSICCV_STREAMING_COMPATIBLE {
-    sve2::VerticalMin<ScalarType>{rect, kernel_}.process_rows(src_rows,
-                                                              dst_rows);
+    VerticalMin<ScalarType>{rect, kernel_}.process_rows(src_rows, dst_rows);
   }
 
  private:
   Rectangle kernel_;
 };  // end of class ErodeOperation<ScalarType>
 
-template <typename T>
+template <typename T, typename CopyOperation>
 static intrinsiccv_error_t erode_sc(const T *src, size_t src_stride, T *dst,
                                     size_t dst_stride, size_t width,
                                     size_t height,
@@ -559,7 +584,7 @@ static intrinsiccv_error_t erode_sc(const T *src, size_t src_stride, T *dst,
   Rows<const T> current_src_rows = src_rows;
   Rows<T> current_dst_rows = dst_rows;
   for (size_t iteration = 0; iteration < workspace->iterations(); ++iteration) {
-    ErodeOperation<T> operation{kernel};
+    ErodeOperation<T, CopyOperation> operation{kernel};
     workspace->process(rect, current_src_rows, current_dst_rows, margin, border,
                        workspace->border_type(), operation);
     // Update source for the next iteration.
diff --git a/intrinsiccv/src/morphology/morphology_sme2.cpp b/intrinsiccv/src/morphology/morphology_sme2.cpp
index 32f9e2d7f..b1864a3af 100644
--- a/intrinsiccv/src/morphology/morphology_sme2.cpp
+++ b/intrinsiccv/src/morphology/morphology_sme2.cpp
@@ -10,16 +10,16 @@ template <typename T>
 INTRINSICCV_LOCALLY_STREAMING INTRINSICCV_TARGET_FN_ATTRS intrinsiccv_error_t
 dilate(const T *src, size_t src_stride, T *dst, size_t dst_stride, size_t width,
        size_t height, intrinsiccv_morphology_context_t *context) {
-  return intrinsiccv::sve2::dilate_sc<T>(src, src_stride, dst, dst_stride,
-                                         width, height, context);
+  return intrinsiccv::sve2::dilate_sc<T, sve2::CopyDataSVE2<T> >(
+      src, src_stride, dst, dst_stride, width, height, context);
 }
 
 template <typename T>
 INTRINSICCV_LOCALLY_STREAMING INTRINSICCV_TARGET_FN_ATTRS intrinsiccv_error_t
 erode(const T *src, size_t src_stride, T *dst, size_t dst_stride, size_t width,
       size_t height, intrinsiccv_morphology_context_t *context) {
-  return intrinsiccv::sve2::erode_sc<T>(src, src_stride, dst, dst_stride, width,
-                                        height, context);
+  return intrinsiccv::sve2::erode_sc<T, sve2::CopyDataSVE2<T> >(
+      src, src_stride, dst, dst_stride, width, height, context);
 }
 
 #define INTRINSICCV_INSTANTIATE_TEMPLATE(name, type)                    \
diff --git a/intrinsiccv/src/morphology/morphology_sve2.cpp b/intrinsiccv/src/morphology/morphology_sve2.cpp
index b66abd956..6be0cf693 100644
--- a/intrinsiccv/src/morphology/morphology_sve2.cpp
+++ b/intrinsiccv/src/morphology/morphology_sve2.cpp
@@ -10,14 +10,16 @@ template <typename T>
 INTRINSICCV_TARGET_FN_ATTRS intrinsiccv_error_t
 dilate(const T *src, size_t src_stride, T *dst, size_t dst_stride, size_t width,
        size_t height, intrinsiccv_morphology_context_t *context) {
-  return dilate_sc(src, src_stride, dst, dst_stride, width, height, context);
+  return dilate_sc<T, MorphologyWorkspace::CopyDataMemcpy<T> >(
+      src, src_stride, dst, dst_stride, width, height, context);
 }
 
 template <typename T>
 INTRINSICCV_TARGET_FN_ATTRS intrinsiccv_error_t
 erode(const T *src, size_t src_stride, T *dst, size_t dst_stride, size_t width,
       size_t height, intrinsiccv_morphology_context_t *context) {
-  return erode_sc(src, src_stride, dst, dst_stride, width, height, context);
+  return erode_sc<T, MorphologyWorkspace::CopyDataMemcpy<T> >(
+      src, src_stride, dst, dst_stride, width, height, context);
 }
 
 #define INTRINSICCV_INSTANTIATE_TEMPLATE(name, type)                    \
diff --git a/scripts/ci-opencv.sh b/scripts/ci-opencv.sh
index 3e6e9923b..993827248 100755
--- a/scripts/ci-opencv.sh
+++ b/scripts/ci-opencv.sh
@@ -22,7 +22,7 @@ tar xf build/opencv.tar.gz -C build
 rm build/opencv.tar.gz
 mv build/opencv-${OPENCV_VER} build/opencv
 patch -d build/opencv -p1<adapters/opencv/opencv-${OPENCV_PATCH_VER}.patch
-cmake \
+LDFLAGS="--rtlib=compiler-rt -fuse-ld=lld" cmake \
   -S build/opencv \
   -B build/build-opencv \
   -G Ninja \
-- 
GitLab


From 128086d262b89f3aa6a5fd65d7ac7377d52aeec5 Mon Sep 17 00:00:00 2001
From: Michael Platings <michael.platings@arm.com>
Date: Tue, 19 Mar 2024 11:44:37 +0000
Subject: [PATCH 2/2] Update to Docker image with latest LLVM

---
 .gitlab-ci.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index c5b101b59..8da6ec095 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
-image: registry.gitlab.arm.com/intrinsiccv/intrinsiccv:6
+image: registry.gitlab.arm.com/intrinsiccv/intrinsiccv:7
 
 # Only run CI for main branch & merge requests
 workflow:
-- 
GitLab