From e55a7017897dd3993f40e324a14aa867a98d353a Mon Sep 17 00:00:00 2001
From: Ahmet Inan <inan@aicodix.de>
Date: Tue, 1 Jul 2025 22:39:07 +0200
Subject: [PATCH] added support for int16_t

---
 polar_helper.hh | 101 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 101 insertions(+)

diff --git a/polar_helper.hh b/polar_helper.hh
index e05c599..a83504a 100644
--- a/polar_helper.hh
+++ b/polar_helper.hh
@@ -153,6 +153,58 @@ struct PolarHelper<SIMD<int8_t, WIDTH>>
 	}
 };
 
+template <int WIDTH>
+struct PolarHelper<SIMD<int16_t, WIDTH>>
+{
+	typedef SIMD<int16_t, WIDTH> TYPE;
+	typedef int PATH;
+	typedef SIMD<uint16_t, WIDTH> MAP;
+	static TYPE one()
+	{
+		return vdup<TYPE>(1);
+	}
+	static TYPE zero()
+	{
+		return vzero<TYPE>();
+	}
+	static TYPE signum(TYPE a)
+	{
+		return vsignum(a);
+	}
+	static TYPE qabs(TYPE a)
+	{
+		return vqabs(a);
+	}
+	static TYPE qadd(TYPE a, TYPE b)
+	{
+		return vqadd(a, b);
+	}
+	static TYPE qmul(TYPE a, TYPE b)
+	{
+#ifdef __ARM_NEON
+		return vmul(a, b);
+#else
+		return vsign(a, b);
+#endif
+	}
+	static TYPE prod(TYPE a, TYPE b)
+	{
+#ifdef __ARM_NEON
+		return vmul(vmul(vsignum(a), vsignum(b)), vmin(vqabs(a), vqabs(b)));
+#else
+		return vsign(vmin(vqabs(a), vqabs(b)), vsign(vsignum(a), b));
+#endif
+	}
+	static TYPE madd(TYPE a, TYPE b, TYPE c)
+	{
+#ifdef __ARM_NEON
+		return vmax(vqadd(vmul(a, vmax(b, vdup<TYPE>(-32767))), c), vdup<TYPE>(-32767));
+#else
+		return vmax(vqadd(vsign(vmax(b, vdup<TYPE>(-32767)), a), c), vdup<TYPE>(-32767));
+#endif
+	}
+};
+
 template <>
 struct PolarHelper<int8_t>
 {
@@ -202,5 +254,54 @@ struct PolarHelper<int8_t>
 	}
 };
 
+template <>
+struct PolarHelper<int16_t>
+{
+	typedef int PATH;
+	static int16_t one()
+	{
+		return 1;
+	}
+	static int16_t zero()
+	{
+		return 0;
+	}
+	static int16_t signum(int16_t v)
+	{
+		return (v > 0) - (v < 0);
+	}
+	template <typename IN>
+	static int16_t quant(IN in)
+	{
+		return std::min<IN>(std::max<IN>(std::nearbyint(in), -32767), 32767);
+	}
+	static int16_t qabs(int16_t a)
+	{
+		return std::abs(std::max<int16_t>(a, -32767));
+	}
+	static int16_t qmin(int16_t a, int16_t b)
+	{
+		return std::min(a, b);
+	}
+	static int16_t qadd(int16_t a, int16_t b)
+	{
+		return std::min<int32_t>(std::max<int32_t>(int32_t(a) + int32_t(b), -32767), 32767);
+	}
+	static int16_t qmul(int16_t a, int16_t b)
+	{
+		// return std::min<int32_t>(std::max<int32_t>(int32_t(a) * int32_t(b), -32767), 32767);
+		// only used for hard decision values anyway
+		return a * b;
+	}
+	static int16_t prod(int16_t a, int16_t b)
+	{
+		return signum(a) * signum(b) * qmin(qabs(a), qabs(b));
+	}
+	static int16_t madd(int16_t a, int16_t b, int16_t c)
+	{
+		return std::min<int32_t>(std::max<int32_t>(int32_t(a) * int32_t(b) + int32_t(c), -32767), 32767);
+	}
+};
+
 }