perf(hex8): prepare guarded private four-column paired tangent packet

src/fesa/elements/hex8_pair_simd_dispatch.cpp

@@ -0,0 +1,53 @@
+#include "hex8_pair_simd.h"
+
+#include "hex8_pair_arithmetic.h"
+
+#if defined(FESA_HEX8_PAIR_AVX2) && defined(_MSC_VER) && defined(_M_X64)
+#include <float.h>
+#include <intrin.h>
+#include <xmmintrin.h>
+#endif
+
+namespace fesa::hex8_internal {
+void EvaluateMaterialPacketScalar(const MaterialPacketInput& input,
+                                  PairPacket* output) {
+  for (std::size_t lane = 0; lane < output->size(); ++lane) {
+    const auto first = MultiplyPair(input.deviatoric_left[lane],
+                                    input.deviatoric_right[lane]);
+    const auto second = MultiplyPair(input.shear_left[lane],
+                                     input.shear_right[lane]);
+    const auto third = MultiplyPair(input.tensor_left[lane],
+                                    input.gradient_dot[lane]);
+    const auto material = AddPair(AddPair(first, second), third);
+    (*output)[lane] = AddPair(input.accumulated[lane],
+                             MultiplyPair(material, input.weight));
+  }
+}
+
+MaterialPacketEvaluator SelectMaterialPacketEvaluator() {
+#if defined(FESA_HEX8_PAIR_AVX2) && defined(_MSC_VER) && defined(_M_X64)
+  static const bool hardware_supported = [] {
+    int info[4]{};
+    __cpuid(info, 0);
+    if (info[0] < 7) return false;
+    __cpuidex(info, 1, 0);
+    constexpr int required = (1 << 12) | (1 << 26) | (1 << 27) | (1 << 28);
+    if ((info[2] & required) != required) return false;
+    if ((_xgetbv(0) & 6U) != 6U) return false;
+    __cpuidex(info, 7, 0);
+    return (info[1] & (1 << 5)) != 0;
+  }();
+  if (!hardware_supported || !math_internal::ArithmeticEnvironmentSupported())
+    return EvaluateMaterialPacketScalar;
+  if ((_mm_getcsr() & 0x1f80U) != 0x1f80U)
+    return EvaluateMaterialPacketScalar;
+  unsigned int control = 0;
+  if (_controlfp_s(&control, 0, 0) != 0 ||
+      (control & _MCW_EM) != _MCW_EM)
+    return EvaluateMaterialPacketScalar;
+  return EvaluateMaterialPacketAvx2;
+#else
+  return EvaluateMaterialPacketScalar;
+#endif
+}
+}  // namespace fesa::hex8_internal