public uint4x3(uint4 c0, uint4 c1, uint4 c2)
{
this.c0 = c0;
this.c1 = c1;
this.c2 = c2;
}
public uint4x2(sfloat v)
{
this.c0 = new uint4(v);
this.c1 = new uint4(v);
}
public uint4x2(float4x2 v)
{
this.c0 = (uint4)v.c0;
this.c1 = (uint4)v.c1;
}
public uint4x3(sfloat v)
{
this.c0 = (uint4)v;
this.c1 = (uint4)v;
this.c2 = (uint4)v;
}
public Hash128(uint4 value) => Value = value;
public uint4x3(bool v)
{
this.c0 = math.select(new uint4(0u), new uint4(1u), v);
this.c1 = math.select(new uint4(0u), new uint4(1u), v);
this.c2 = math.select(new uint4(0u), new uint4(1u), v);
}
public uint4x3(int v)
{
this.c0 = (uint4)v;
this.c1 = (uint4)v;
this.c2 = (uint4)v;
}
public uint8(uint4 x0123, uint2 x45, uint2 x67)
{
this = new uint8(x0123, new uint4(x45, x67));
}
public uint8(uint2 x01, uint2 x23, uint4 x4567)
{
this = new uint8(new uint4(x01, x23), x4567);
}
public override int GetHashCode()
{
uint4 primes = new uint4(863, 5471887, 13143149, 15485291);
return((int)math.csum(Value * primes));
}
public static int ConvertToInt(uint4 result)
{
return((int)(result.x + result.y * 10 + result.z * 100 + result.w * 1000));
}
public uint4x2(float v)
{
this.c0 = (uint4)v;
this.c1 = (uint4)v;
}
public uint4x2(double4x2 v)
{
this.c0 = (uint4)v.c0;
this.c1 = (uint4)v.c1;
}
public uint4x2(double v)
{
this.c0 = (uint4)v;
this.c1 = (uint4)v;
}
public uint4x3(uint v)
{
this.c0 = v;
this.c1 = v;
this.c2 = v;
}
public uint4x2(uint4 c0, uint4 c1)
{
this.c0 = c0;
this.c1 = c1;
}
public static uint4x3 uint4x3(uint4 c0, uint4 c1, uint4 c2)
{
return(new uint4x3(c0, c1, c2));
}
public uint4x2(uint v)
{
this.c0 = v;
this.c1 = v;
}
public uint4x3(bool4x3 v)
{
this.c0 = math.select(new uint4(0u), new uint4(1u), v.c0);
this.c1 = math.select(new uint4(0u), new uint4(1u), v.c1);
this.c2 = math.select(new uint4(0u), new uint4(1u), v.c2);
}
public static uint4x2 uint4x2(uint4 c0, uint4 c1)
{
return(new uint4x2(c0, c1));
}
public uint4x3(int4x3 v)
{
this.c0 = (uint4)v.c0;
this.c1 = (uint4)v.c1;
this.c2 = (uint4)v.c2;
}
public uint4x2(int v)
{
this.c0 = (uint4)v;
this.c1 = (uint4)v;
}
public static extern CUResult cuMemcpyDtoH_v2(ref uint4 dstHost, CUdeviceptr srcDevice, SizeT ByteCount);
public static uint4 gcd(uint4 x, uint4 y)
{
if (Sse2.IsSse2Supported)
{
v128 ZERO = default(v128);
v128 _x = *(v128 *)&x;
v128 _y = *(v128 *)&y;
v128 result = ZERO;
v128 result_if_zero_any = ZERO;
v128 x_is_zero = Sse2.cmpeq_epi32(_x, ZERO);
v128 y_is_zero = Sse2.cmpeq_epi32(_y, ZERO);
v128 any_zero = Sse2.or_si128(x_is_zero, y_is_zero);
result_if_zero_any = Mask.BlendV(result_if_zero_any, _y, x_is_zero);
result_if_zero_any = Mask.BlendV(result_if_zero_any, _x, y_is_zero);
v128 doneMask = any_zero;
int4 shift = math.tzcnt(x | y);
x = shrl(x, math.tzcnt(x));
_x = *(v128 *)&x;
do
{
uint4 temp_y = shrl(*(uint4 *)&_y, math.tzcnt(*(uint4 *)&_y));
_y = *(v128 *)&temp_y;
if (Sse4_1.IsSse41Supported)
{
v128 tempX = _x;
_x = Sse4_1.min_epu32(_x, _y);
_y = Sse4_1.max_epu32(_y, tempX);
}
else
{
v128 tempX = _x;
v128 x_greater_y = Operator.greater_mask_uint(_x, _y);
_x = Mask.BlendV(_x, _y, x_greater_y);
_y = Mask.BlendV(_y, tempX, x_greater_y);
}
_y = Sse2.sub_epi32(_y, _x);
v128 loopCheck = Sse2.andnot_si128(doneMask, Sse2.cmpeq_epi32(_y, ZERO));
result = Mask.BlendV(result, _x, loopCheck);
doneMask = Sse2.or_si128(doneMask, loopCheck);
} while (bitmask32(4 * sizeof(uint)) != Sse2.movemask_epi8(doneMask));
uint4 result_temp = shl(*(uint4 *)&result, shift);
result = *(v128 *)&result_temp;
result = Mask.BlendV(result, result_if_zero_any, any_zero);
return(*(uint4 *)&result);
}
else
{
return(new uint4(gcd(x.x, y.x), gcd(x.y, y.y), gcd(x.z, y.z), gcd(x.w, y.w)));
}
}
