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

2026-10-14 20:12:20 +09:00
parent b474d127d8
commit 9b8e53dcfd
6 changed files with 498 additions and 5 deletions
@@ -0,0 +1,28 @@
+#ifndef FESA_SRC_FESA_ELEMENTS_HEX8_PAIR_SIMD_H_
+#define FESA_SRC_FESA_ELEMENTS_HEX8_PAIR_SIMD_H_
+
+#include <array>
+
+#include "../math/high_low_arithmetic.h"
+
+namespace fesa::hex8_internal {
+using PairPacket = std::array<math_internal::HighLow, 4>;
+struct MaterialPacketInput {
+  PairPacket deviatoric_left, deviatoric_right;
+  PairPacket shear_left, shear_right;
+  PairPacket tensor_left, gradient_dot;
+  PairPacket accumulated;
+  math_internal::HighLow weight;
+};
+using MaterialPacketEvaluator = void (*)(const MaterialPacketInput&, PairPacket*);
+
+void EvaluateMaterialPacketScalar(const MaterialPacketInput& input,
+                                  PairPacket* output);
+void EvaluateMaterialPacketAvx2(const MaterialPacketInput& input,
+                                PairPacket* output);
+/// @brief Selects a private backend from CPU/OS and this thread's FP state.
+/// @note Never mutates FP state or caches a thread's environment decision.
+MaterialPacketEvaluator SelectMaterialPacketEvaluator();
+}  // namespace fesa::hex8_internal
+
+#endif  // FESA_SRC_FESA_ELEMENTS_HEX8_PAIR_SIMD_H_