refactor(numpycpp): update init, linalg, numpy headers and svml_bridge

Author: peng.li24

numpycpp/detail/svml_bridge.h

@@ -148,29 +148,39 @@ NUMPY_SVML_F32(cbrt,  "__svml_cbrtf16", "npy_cbrtf")
 NUMPY_SVML_F32(expm1, "__svml_expm1f16","npy_expm1f")
 NUMPY_SVML_F32(log1p, "__svml_log1pf16","npy_log1pf")
 
-// pow / atan2 — SVML 2-arg
+// pow / atan2 — SVML 2-arg: 使用 __svml_pow8 / __svml_atan28 向量符号，
+// 广播标量到 __m512，调用 SVML 向量函数，提取结果。确保与 numpy 内部使用的
+// SVML 实现位级一致（npy_pow / npy_atan2 是标量 libm 回退，会差 1 ULP）。
 __attribute__((target("avx512f")))
 inline double pow_svml_f64(double x, double e) {
-    static auto fn = (double (*)(double, double))resolve_svml("npy_pow");
-    if (fn) return fn(x, e);
+    static auto fn = (__m512d (*)(__m512d, __m512d))resolve_svml("__svml_pow8");
+    if (fn) return _mm512_cvtsd_f64(fn(_mm512_set1_pd(x), _mm512_set1_pd(e)));
+    static auto scalar_fn = (double (*)(double, double))resolve_svml("npy_pow");
+    if (scalar_fn) return scalar_fn(x, e);
     return std::pow(x, e);
 }
 __attribute__((target("avx512f")))
 inline float pow_svml_f32(float x, float e) {
-    static auto fn = (float (*)(float, float))resolve_svml("npy_powf");
-    if (fn) return fn(x, e);
+    static auto fn = (__m512 (*)(__m512, __m512))resolve_svml("__svml_powf16");
+    if (fn) return _mm512_cvtss_f32(fn(_mm512_set1_ps(x), _mm512_set1_ps(e)));
+    static auto scalar_fn = (float (*)(float, float))resolve_svml("npy_powf");
+    if (scalar_fn) return scalar_fn(x, e);
     return std::pow(x, e);
 }
 __attribute__((target("avx512f")))
 inline double atan2_svml_f64(double y, double x) {
-    static auto fn = (double (*)(double, double))resolve_svml("npy_atan2");
-    if (fn) return fn(y, x);
+    static auto fn = (__m512d (*)(__m512d, __m512d))resolve_svml("__svml_atan28");
+    if (fn) return _mm512_cvtsd_f64(fn(_mm512_set1_pd(y), _mm512_set1_pd(x)));
+    static auto scalar_fn = (double (*)(double, double))resolve_svml("npy_atan2");
+    if (scalar_fn) return scalar_fn(y, x);
     return std::atan2(y, x);
 }
 __attribute__((target("avx512f")))
 inline float atan2_svml_f32(float y, float x) {
-    static auto fn = (float (*)(float, float))resolve_svml("npy_atan2f");
-    if (fn) return fn(y, x);
+    static auto fn = (__m512 (*)(__m512, __m512))resolve_svml("__svml_atan2f16");
+    if (fn) return _mm512_cvtss_f32(fn(_mm512_set1_ps(y), _mm512_set1_ps(x)));
+    static auto scalar_fn = (float (*)(float, float))resolve_svml("npy_atan2f");
+    if (scalar_fn) return scalar_fn(y, x);
     return std::atan2(y, x);
 }
 

numpycpp/init.h

@@ -97,7 +97,7 @@ inline void logspace(T* dst, T start, T stop, size_t num,
                      bool endpoint = true, T base = T(10)) {
     linspace(dst, start, stop, num, endpoint);
     for (size_t i = 0; i < num; ++i)
-        dst[i] = std::pow(base, dst[i]);
+        dst[i] = detail::pow(base, dst[i]);
 }
 
 // ============================================================================

numpycpp/linalg.h

@@ -20,6 +20,7 @@
 #include <map>
 #include <set>
 #include <algorithm>
+#include <numeric>
 #include <stdexcept>
 #include <type_traits>
 #include <immintrin.h>   // SSE/AVX intrinsics (SSE2 is baseline on x86_64)

numpycpp/numpy.h

@@ -17,6 +17,18 @@
 // ════════════════════════════════════════════════════════════════════════════
 #pragma once
 
+// Internal math backend — loaded first so init.h et al. can use detail::pow etc.
+#ifndef NUMPYCPP_INTERNAL_INCLUDE
+#  define NUMPYCPP_INTERNAL_INCLUDE
+#  ifdef NUMPYCPP_STD_ONLY
+#    include "detail/std_math_backend.h"
+#  else
+#    include "detail/npy_math_float.h"
+#    include "detail/svml_bridge.h"
+#  endif
+#  undef NUMPYCPP_INTERNAL_INCLUDE
+#endif
+
 #include "init.h"
 #include "elementwise.h"
 #include "reduce.h"