refactor: namespace svml → detail — prevent direct use by applications

Rename numpy::svml → numpy::detail to follow the standard C++ convention
for internal implementation namespaces. Applications should only use the
public numpy::* API from core.h — never numpy::detail::exp() etc.

Also update svml_bridge.h header comment to explicitly warn against
direct use of the detail namespace.

Author: peng.li24

numpy/core.h

@@ -86,37 +86,37 @@ inline void abs(const T* src, T* dst, size_t n) {
 /// numpy.exp(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void exp(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::exp(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::exp(src[i]));
 }
 
 /// numpy.log(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void log(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::log(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::log(src[i]));
 }
 
 /// numpy.sin(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void sin(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::sin(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::sin(src[i]));
 }
 
 /// numpy.cos(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void cos(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::cos(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::cos(src[i]));
 }
 
 /// numpy.tan(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void tan(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::tan(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::tan(src[i]));
 }
 
 /// numpy.power(x1, x2, /, out=None, *, where=True, ...)
 template<typename T>
 inline void power(const T* src, T* dst, size_t n, T exponent) {
-    NUMPY_UNROLL4(i, dst[i] = svml::pow(src[i], exponent));
+    NUMPY_UNROLL4(i, dst[i] = detail::pow(src[i], exponent));
 }
 
 /// numpy.clip(a, a_min, a_max, out=None, **kwargs)
@@ -128,31 +128,31 @@ inline void clip(const T* src, T* dst, size_t n, T min_val, T max_val) {
 /// numpy.log10(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void log10(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::log10(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::log10(src[i]));
 }
 
 /// numpy.log2(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void log2(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::log2(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::log2(src[i]));
 }
 
 /// numpy.arcsin(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void arcsin(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::asin(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::asin(src[i]));
 }
 
 /// numpy.arccos(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void arccos(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::acos(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::acos(src[i]));
 }
 
 /// numpy.arctan(x, /, out=None, *, where=True, ...)
 template<typename T>
 inline void arctan(const T* src, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::atan(src[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::atan(src[i]));
 }
 
 /// numpy.round(a, decimals=0, out=None)
@@ -414,12 +414,12 @@ inline void isfinite(const T* src, bool* dst, size_t n) {
 /// numpy.arctan2(x1, x2, /, out=None, *, where=True, ...) — array-array
 template<typename T>
 inline void arctan2_array(const T* a, const T* b, T* dst, size_t n) {
-    NUMPY_UNROLL4(i, dst[i] = svml::atan2(a[i], b[i]));
+    NUMPY_UNROLL4(i, dst[i] = detail::atan2(a[i], b[i]));
 }
 /// numpy.arctan2(x1, x2, /, out=None, *, where=True, ...) — array-scalar
 template<typename T>
 inline void arctan2_scalar(const T* src, T* dst, size_t n, T b) {
-    NUMPY_UNROLL4(i, dst[i] = svml::atan2(src[i], b));
+    NUMPY_UNROLL4(i, dst[i] = detail::atan2(src[i], b));
 }
 /// numpy.maximum(x1, x2, /, out=None, *, where=True, ...) — array-array
 template<typename T>

numpy/svml_bridge.h

@@ -1,6 +1,9 @@
 // INTERNAL HEADER — DO NOT INCLUDE DIRECTLY.
 // Use #include "numpy/core.h" which pulls this in automatically.
 //
+// All functions live in numpy::detail — do not call numpy::detail::exp()
+// directly. Use numpy::exp() from core.h.
+//
 // SVML/npy bridge — bit-exact math on every x86_64 architecture.
 //
 // numpy uses different math implementations depending on CPU features:
@@ -25,7 +28,9 @@
 #endif
 
 namespace numpy {
-namespace svml {
+namespace detail {
+// Internal dispatch namespace — use numpy::exp() etc., not numpy::detail::exp().
+// All math functions are resolved at runtime from numpy's _multiarray_umath.so.
 
 inline void* g_svml_handle = nullptr;
 
@@ -318,5 +323,5 @@ NUMPY_SVML_D1(sqrt)
 template<typename T> inline T pow(T x, T e)    { return svml_impl<T>::pow(x, e); }
 template<typename T> inline T atan2(T y, T x)  { return svml_impl<T>::atan2(y, x); }
 
-} // namespace svml
+} // namespace detail
 } // namespace numpy

tests/module.cpp

@@ -47,7 +47,7 @@ PYBIND11_MODULE(numpycpp, m) {
     try {
         py::module_ np_core = py::module_::import("numpy.core._multiarray_umath");
         std::string umath_path = np_core.attr("__file__").cast<std::string>();
-        numpy::svml::bridge_init(umath_path.c_str());
+        numpy::detail::bridge_init(umath_path.c_str());
     } catch (...) {
         // Fall back: SVML → libm
     }