Add missing get_unpack_fn

Author: AntoinePrv

cpp/src/arrow/util/bpacking.cc

@@ -68,4 +68,22 @@ template int unpack<uint16_t>(const uint8_t*, uint16_t*, int, int);
 template int unpack<uint32_t>(const uint8_t*, uint32_t*, int, int);
 template int unpack<uint64_t>(const uint8_t*, uint64_t*, int, int);
 
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn(int num_bits) {
+  if constexpr (std::is_same_v<Uint, uint16_t>) {
+    // Current SIMD unpack function do not out beat scalar implementation for uin16_t
+    return get_unpack_fn_scalar<uint16_t>(num_bits);
+  } else {
+#if defined(ARROW_HAVE_NEON)
+    return get_unpack_fn_neon<Uint>(num_bits);
+#else
+    // TODO
+#endif
+  }
+}
+
+template UnpackFn<uint16_t> get_unpack_fn<uint16_t>(int);
+template UnpackFn<uint32_t> get_unpack_fn<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn<uint64_t>(int);
+
 }  // namespace arrow::internal

cpp/src/arrow/util/bpacking_internal.h

@@ -35,4 +35,16 @@ extern template ARROW_TEMPLATE_EXPORT int unpack<uint32_t>(const uint8_t*, uint3
 extern template ARROW_TEMPLATE_EXPORT int unpack<uint64_t>(const uint8_t*, uint64_t*, int,
                                                            int);
 
+template <typename Uint>
+using UnpackFn = int (*)(const uint8_t* in, Uint* out, int batch_size);
+
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint16_t> get_unpack_fn<uint16_t>(int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn<uint32_t>(int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn<uint64_t>(int);
+
 }  // namespace arrow::internal

cpp/src/arrow/util/bpacking_scalar.cc

@@ -30,4 +30,13 @@ template int unpack_scalar<uint16_t>(const uint8_t*, uint16_t*, int, int);
 template int unpack_scalar<uint32_t>(const uint8_t*, uint32_t*, int, int);
 template int unpack_scalar<uint64_t>(const uint8_t*, uint64_t*, int, int);
 
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn_scalar(int num_bits) {
+  return get_unpack_fn<ScalarUnpackerForWidth, Uint>(num_bits);
+}
+
+template UnpackFn<uint16_t> get_unpack_fn_scalar<uint16_t>(int);
+template UnpackFn<uint32_t> get_unpack_fn_scalar<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn_scalar<uint64_t>(int);
+
 }  // namespace arrow::internal

cpp/src/arrow/util/bpacking_scalar_internal.h

@@ -17,6 +17,7 @@
 
 #pragma once
 
+#include "arrow/util/bpacking_internal.h"
 #include "arrow/util/visibility.h"
 
 #include <cstdint>
@@ -36,4 +37,16 @@ extern template ARROW_TEMPLATE_EXPORT int unpack_scalar<uint32_t>(const uint8_t*
 extern template ARROW_TEMPLATE_EXPORT int unpack_scalar<uint64_t>(const uint8_t*,
                                                                   uint64_t*, int, int);
 
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn_scalar(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint16_t> get_unpack_fn_scalar<uint16_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn_scalar<uint32_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn_scalar<uint64_t>(
+    int);
+
 }  // namespace arrow::internal

cpp/src/arrow/util/bpacking_simd_internal.h

@@ -17,6 +17,7 @@
 
 #pragma once
 
+#include "arrow/util/bpacking_internal.h"
 #include "arrow/util/visibility.h"
 
 #include <cstdint>
@@ -37,6 +38,18 @@ extern template ARROW_TEMPLATE_EXPORT int unpack_neon<uint32_t>(const uint8_t*,
 extern template ARROW_TEMPLATE_EXPORT int unpack_neon<uint64_t>(const uint8_t*, uint64_t*,
                                                                 int, int);
 
+template <typename Uint>
+ARROW_EXPORT UnpackFn<Uint> get_unpack_fn_neon(int num_bits);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint16_t> get_unpack_fn_neon<uint16_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint32_t> get_unpack_fn_neon<uint32_t>(
+    int);
+
+extern template ARROW_TEMPLATE_EXPORT UnpackFn<uint64_t> get_unpack_fn_neon<uint64_t>(
+    int);
+
 #endif
 
 #if defined(ARROW_HAVE_SSE4_2)

cpp/src/arrow/util/bpacking_simd_min.cc

@@ -32,6 +32,15 @@ template int unpack_neon<uint16_t>(const uint8_t*, uint16_t*, int, int);
 template int unpack_neon<uint32_t>(const uint8_t*, uint32_t*, int, int);
 template int unpack_neon<uint64_t>(const uint8_t*, uint64_t*, int, int);
 
+template <typename Uint>
+UnpackFn<Uint> get_unpack_fn_neon(int num_bits) {
+  return get_unpack_fn<Simd128UnpackerForWidth, Uint>(num_bits);
+}
+
+template UnpackFn<uint16_t> get_unpack_fn_neon<uint16_t>(int);
+template UnpackFn<uint32_t> get_unpack_fn_neon<uint32_t>(int);
+template UnpackFn<uint64_t> get_unpack_fn_neon<uint64_t>(int);
+
 #endif
 
 #if defined(ARROW_HAVE_SSE4_2)