Function CPPSPMD_DECL

encoder/cppspmd_math.h:474–523

{
	const float fPi = 3.141592653589793f, fOneOverPi = 0.3183098861837907f;
	CPPSPMD_DECL(const uint8_t, s_table0[16]) =	{ 128 + 0, 128 + 2, 128 + -2, 128 + 4,    128 + 0, 128 + 2, 128 + -2, 128 + 4,	  128 + 0, 128 + 2, 128 + -2, 128 + 4,   128 + 0, 128 + 2, 128 + -2, 128 + 4 };

	vint table = init_lookup4(s_table0); // a load
	vint sgn = cast_vfloat_to_vint(x) & 0x80000000;

	store_all(x, abs(x));
	vfloat orig_x = x;

	vfloat q = x * fOneOverPi;
	store_all(x, q - floor(q));

	vfloat x4 = x * 4.0f;
	vint octant = (vint)(x4);

	vfloat x0 = spmd_ternaryf((octant & 1) != 0, -x4, x4);

	vint k = table_lookup4_8(octant, table) & 0xFF; // a shuffle

	vfloat bias = (vfloat)k + -128.0f;
	vfloat y = x0 + bias;

	vfloat z = tan82(y);

	vfloat r;

	vbool octant_one_or_two = (octant == 1) || (octant == 2);

	// SPMD optimization - skip costly divide if we can
	if (spmd_any(octant_one_or_two))
	{
		const float fDivThresh = .4371e-7f;
		vfloat one_over_z = 1.0f / spmd_ternaryf(abs(z) > fDivThresh, z, spmd_ternaryf(z < 0.0f, -fDivThresh, fDivThresh));

		vfloat b = spmd_ternaryf(octant_one_or_two, one_over_z, z);
		store_all(r, spmd_ternaryf((octant & 2) != 0, -b, b));
	}
	else
	{
		store_all(r, spmd_ternaryf(octant == 0, z, -z));
	}

	// Small angle approximation, to decrease the max rel error near Pi.
	SPMD_SIF(x >= (1.0f - .0003125f*4.0f))
	{
		store(r, vfnma(floor(q) + 1.0f, fPi, orig_x));
	}
	SPMD_SENDIF

	return cast_vint_to_vfloat(cast_vfloat_to_vint(r) ^ sgn);
}

inline void spmd_kernel::seed_rand(rand_context& x, vint seed)
{