public ushort8(ushort2 x01, ushort3 x234, ushort3 x567)
{
if (Sse2.IsSse2Supported)
{
v128 mid = Sse2.bslli_si128(x234, 2 * sizeof(ushort));
v128 hi = Sse2.bslli_si128(x567, 5 * sizeof(ushort));
if (Sse4_1.IsSse41Supported)
{
hi = Sse4_1.blend_epi16(mid, hi, 0b1110_0000);
this = Sse4_1.blend_epi16(x01, hi, 0b1111_1100);
}
else
{
hi = Mask.BlendEpi16_SSE2(mid, hi, 0b1110_0000);
this = Mask.BlendEpi16_SSE2(x01, hi, 0b1111_1100);
}
}
else
{
this.x0 = x01.x;
this.x1 = x01.y;
this.x2 = x234.x;
this.x3 = x234.y;
this.x4 = x234.z;
this.x5 = x567.x;
this.x6 = x567.y;
this.x7 = x567.z;
}
}
public ushort2x3(ushort m00, ushort m01, ushort m02,
ushort m10, ushort m11, ushort m12)
{
this.c0 = new ushort2(m00, m10);
this.c1 = new ushort2(m01, m11);
this.c2 = new ushort2(m02, m12);
}
public static ushort2 lcm(short2 x, short2 y)
{
ushort2 absX = (ushort2)abs(x);
ushort2 absY = (ushort2)abs(y);
return((absX / gcd(absX, absY)) * absY);
}
public ushort8(ushort3 x012, ushort2 x34, ushort3 x567)
{
if (Sse2.IsSse2Supported)
{
v128 hi = Sse2.bslli_si128(x567, 2 * sizeof(short));
if (Sse4_1.IsSse41Supported)
{
hi = Sse4_1.blend_epi16(x34, hi, 0b0001_1100);
hi = Sse2.bslli_si128(hi, 3 * sizeof(short));
this = Sse4_1.blend_epi16(x012, hi, 0b1111_1000);
}
else
{
hi = Mask.BlendEpi16_SSE2(x34, hi, 0b0001_1100);
hi = Sse2.bslli_si128(hi, 3 * sizeof(short));
this = Mask.BlendEpi16_SSE2(x012, hi, 0b1111_1000);
}
}
else
{
this.x0 = x012.x;
this.x1 = x012.y;
this.x2 = x012.z;
this.x3 = x34.x;
this.x4 = x34.y;
this.x5 = x567.x;
this.x6 = x567.y;
this.x7 = x567.z;
}
}
public ushort8(ushort3 x012, ushort3 x345, ushort2 x67)
{
if (Sse2.IsSse2Supported)
{
v128 mid = Sse2.bslli_si128(x345, 3 * sizeof(ushort));
v128 hi = Sse2.bslli_si128(x67, 6 * sizeof(ushort));
if (Sse4_1.IsSse41Supported)
{
mid = Sse4_1.blend_epi16(x012, mid, 0b0011_1000);
this = Sse4_1.blend_epi16(mid, hi, 0b1100_0000);
}
else
{
mid = Mask.BlendEpi16_SSE2(x012, mid, 0b0011_1000);
this = Mask.BlendEpi16_SSE2(mid, hi, 0b1100_0000);
}
}
else
{
this.x0 = x012.x;
this.x1 = x012.y;
this.x2 = x012.z;
this.x3 = x345.x;
this.x4 = x345.y;
this.x5 = x345.z;
this.x6 = x67.x;
this.x7 = x67.y;
}
}
public ushort2x4(ushort v)
{
this.c0 = v;
this.c1 = v;
this.c2 = v;
this.c3 = v;
}
public ushort2x4(ushort2 c0, ushort2 c1, ushort2 c2, ushort2 c3)
{
this.c0 = c0;
this.c1 = c1;
this.c2 = c2;
this.c3 = c3;
}
public static ushort2 tzcnt(ushort2 x)
{
if (Ssse3.IsSsse3Supported)
{
v128 NIBBLE_MASK = new v128(0x0F0F_0F0F);
v128 SHUFFLE_MASK_LO = new v128(16, 0, 1, 0, 2, 0, 1, 0, 3, 0, 1, 0, 2, 0, 1, 0);
v128 SHUFFLE_MASK_HI = new v128(16, 4, 5, 4, 6, 4, 5, 4, 7, 4, 5, 4, 6, 4, 5, 4);
v128 tzcnt_bytes = Sse2.min_epu8(Ssse3.shuffle_epi8(SHUFFLE_MASK_LO, Sse2.and_si128(NIBBLE_MASK, x)),
Ssse3.shuffle_epi8(SHUFFLE_MASK_HI, Sse2.and_si128(NIBBLE_MASK, Sse2.srli_epi16(x, 4))));
return(Sse2.min_epu8(tzcnt_bytes,
Sse2.srli_epi16(Sse2.add_epi8(tzcnt_bytes, Sse2.set1_epi8(8)), 8)));
}
else if (Sse2.IsSse2Supported)
{
v128 compareMask = x & (ushort2)(-((short2)x));
ushort2 first = Mask.BlendV(default(v128), new ushort2(1), Sse2.cmpeq_epi16(compareMask, default(v128)));
ushort2 second = Mask.BlendV(default(v128), new ushort2(8), Sse2.cmpeq_epi16(compareMask & (ushort2)0x00FF, default(v128)));
ushort2 third = Mask.BlendV(default(v128), new ushort2(4), Sse2.cmpeq_epi16(compareMask & (ushort2)0x0F0F, default(v128)));
ushort2 fourth = Mask.BlendV(default(v128), new ushort2(2), Sse2.cmpeq_epi16(compareMask & (ushort2)0x3333, default(v128)));
ushort2 fifth = Mask.BlendV(default(v128), new ushort2(1), Sse2.cmpeq_epi16(compareMask & (ushort2)0x5555, default(v128)));
return((first + second) + ((third + fourth) + fifth));
}
else
{
return(new ushort2(tzcnt(x.x), tzcnt(x.y)));
}
}
public static ushort2 gcd(ushort2 x, ushort2 y)
{
if (Sse2.IsSse2Supported)
{
v128 ZERO = default(v128);
v128 result = ZERO;
v128 result_if_zero_any = ZERO;
v128 x_is_zero = Sse2.cmpeq_epi16(x, ZERO);
v128 y_is_zero = Sse2.cmpeq_epi16(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;
ushort2 shift = tzcnt(x | y);
x = shrl(x, tzcnt(x));
do
{
y = shrl(y, tzcnt(y));
if (Sse4_1.IsSse41Supported)
{
v128 tempX = x;
x = Sse4_1.min_epu16(x, y);
y = Sse4_1.max_epu16(y, tempX);
}
else
{
v128 tempX = x;
v128 x_greater_y = Operator.greater_mask_ushort(x, y);
x = Mask.BlendV(x, y, x_greater_y);
y = Mask.BlendV(y, tempX, x_greater_y);
}
y -= x;
v128 loopCheck = Sse2.andnot_si128(doneMask, Sse2.cmpeq_epi16(y, ZERO));
result = Mask.BlendV(result, x, loopCheck);
doneMask = Sse2.or_si128(doneMask, loopCheck);
} while (-1 != doneMask.SInt0);
result = shl(result, shift);
result = Mask.BlendV(result, result_if_zero_any, any_zero);
return(result);
}
else
{
return(new ushort2((ushort)gcd((uint)x.x, (uint)y.x), (ushort)gcd((uint)x.y, (uint)y.y)));
}
}
public ushort2x4(ushort m00, ushort m01, ushort m02, ushort m03,
ushort m10, ushort m11, ushort m12, ushort m13)
{
this.c0 = new ushort2(m00, m10);
this.c1 = new ushort2(m01, m11);
this.c2 = new ushort2(m02, m12);
this.c3 = new ushort2(m03, m13);
}
public static ushort2 reversebits(ushort2 x)
{
x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
x = ((x >> 4) & 0x0F0F) | ((x & 0x0F0F) << 4);
return((x >> 8) | (x << 8));
}
public static ushort2 floorpow2(ushort2 x)
{
x |= x >> 1;
x |= x >> 2;
x |= x >> 4;
x |= x >> 8;
return(x - (x >> 1));
}
public static ushort2 ceilpow2(ushort2 x)
{
x -= 1;
x |= x >> 1;
x |= x >> 2;
x |= x >> 4;
x |= x >> 8;
return(x + 1);
}
public static ushort avg(ushort2 c)
{
if (Sse2.IsSse2Supported)
{
return(Sse2.avg_epu16(c, Sse2.bsrli_si128(c, 1 * sizeof(ushort))).UShort0);
}
else
{
return((ushort)((1u + csum(c)) / 2u));
}
}
public static ushort2 avg(ushort2 x, ushort2 y)
{
if (Sse2.IsSse2Supported)
{
return(Sse2.avg_epu16(x, y));
}
else
{
return(new ushort2((ushort)((x.x + y.x + 1) >> 1), (ushort)((x.y + y.y + 1) >> 1)));
}
}
public static ushort cmin(ushort2 x)
{
if (Sse2.IsSse2Supported)
{
return(min(x, x.yy).x);
}
else
{
return((ushort)math.min((uint)x.x, (uint)x.y));
}
}
public static ushort2 andnot(ushort2 left, ushort2 right)
{
if (Sse2.IsSse2Supported)
{
return(Sse2.andnot_si128(right, left));
}
else
{
return(left & ~right);
}
}
public static ushort2 subadd(ushort2 a, ushort2 b)
{
if (Ssse3.IsSsse3Supported)
{
return(a + Ssse3.sign_epi16(b, new ushort2(ushort.MaxValue, 1)));
}
else
{
return(a - select(b, (ushort2)(-(short2)b), new bool2(false, true)));
}
}
public static ushort2 divrem(ushort2 dividend, ushort divisor, out ushort2 remainder)
{
if (Constant.IsConstantExpression(divisor))
{
remainder = dividend % divisor;
return(dividend / divisor);
}
else
{
return(divrem(dividend, (ushort2)divisor, out remainder));
}
}
public static bool2 ispow2(ushort2 x)
{
if (Sse2.IsSse2Supported)
{
v128 result = (byte2)(new ushort2(1) & Sse2.and_si128(Operator.greater_mask_ushort(x, default(v128)),
Sse2.cmpeq_epi16(default(v128), x & (x - 1))));
return(*(bool2 *)&result);
}
else
{
return(new bool2(math.ispow2((uint)x.x), math.ispow2((uint)x.y)));
}
}
internal static v128 greater_mask_ushort(ushort2 left, ushort2 right)
{
if (Sse2.IsSse2Supported)
{
ushort2 mask = 1 << 15;
return(Sse2.cmpgt_epi16(Sse2.xor_si128(left, mask),
Sse2.xor_si128(right, mask)));
}
else
{
throw new CPUFeatureCheckException();
}
}
public ushort2 NextUShort2(ushort2 max)
{
Assert.IsPositive(max.x);
Assert.IsPositive(max.y);
if (Sse2.IsSse2Supported)
{
return(Sse2.mulhi_epi16(max, new ushort2((ushort)NextState(), (ushort)NextState())));
}
else
{
return((ushort2)(((uint2)max * new uint2(NextState(), NextState())) >> 16));
}
}
public ushort2 NextUShort2(ushort2 min, ushort2 max)
{
Assert.IsNotSmaller(max.x, min.x);
Assert.IsNotSmaller(max.y, min.y);
if (Sse2.IsSse2Supported)
{
return(min + Sse2.mulhi_epi16(max - min, new ushort2((ushort)NextState(), (ushort)NextState())));
}
else
{
return(min + (ushort2)(((uint2)(max - min) * new uint2(NextState(), NextState())) >> 16));
}
}
public static ushort2 max(ushort2 a, ushort2 b)
{
if (Sse4_1.IsSse41Supported)
{
return(Sse4_1.max_epu16(a, b));
}
else if (Sse2.IsSse2Supported)
{
return(Mask.BlendV(a, b, Operator.greater_mask_ushort(b, a)));
}
else
{
return(new ushort2((ushort)math.max((uint)a.x, (uint)b.x), (ushort)math.max((uint)a.y, (uint)b.y)));
}
}
public static short2 compareto(ushort2 x, ushort2 y)
{
if (Sse2.IsSse2Supported)
{
short2 xGreatery = Operator.greater_mask_ushort(x, y);
short2 yGreaterx = Operator.greater_mask_ushort(y, x);
return((0 - xGreatery) + yGreaterx);
}
else
{
return(new short2((short)compareto(x.x, y.x),
(short)compareto(x.y, y.y)));
}
}
public static bool2 isdivisible(ushort2 dividend, ushort2 divisor)
{
Assert.AreNotEqual(0, divisor.x);
Assert.AreNotEqual(0, divisor.y);
if (Constant.IsConstantExpression(divisor))
{
uint2 compile = (new uint2(uint.MaxValue) / divisor) + 1;
return(dividend * compile <= compile - 1);
}
else
{
return(dividend % divisor == 0);
}
}
public sbyte2 NextSByte2(sbyte2 min, sbyte2 max)
{
Assert.IsNotSmaller(max.x, min.x);
Assert.IsNotSmaller(max.y, min.y);
if (Ssse3.IsSsse3Supported)
{
ushort2 temp = (ushort2)(max - min) * new ushort2(NextState(), NextState());
return(min + Ssse3.shuffle_epi8(temp, new byte4(1, 3, 0, 0)));
}
else
{
return(min + (sbyte2)(((ushort2)(max - min) * new ushort2(NextState(), NextState())) >> 8));
}
}
public static ushort2 divrem(ushort2 dividend, ushort2 divisor, out ushort2 remainder)
{
if (Sse2.IsSse2Supported)
{
ushort2 quotient = dividend / divisor;
remainder = dividend - (quotient * divisor);
return(quotient);
}
else
{
remainder = dividend % divisor;
return(dividend / divisor);
}
}
public static ushort2 lzcnt(ushort2 x)
{
if (Ssse3.IsSsse3Supported)
{
v128 NIBBLE_MASK = new v128(0x0F0F_0F0F);
v128 SHUFFLE_MASK_LO = new v128(16, 7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4);
v128 SHUFFLE_MASK_HI = new v128(16, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0);
v128 lzcnt_bytes = Sse2.min_epu8(Ssse3.shuffle_epi8(SHUFFLE_MASK_LO, Sse2.and_si128(NIBBLE_MASK, x)),
Ssse3.shuffle_epi8(SHUFFLE_MASK_HI, Sse2.and_si128(NIBBLE_MASK, Sse2.srli_epi16(x, 4))));
return(Sse2.min_epu8(Sse2.add_epi8(lzcnt_bytes, Sse2.set1_epi16(8)),
Sse2.srli_epi16(lzcnt_bytes, 8)));
}
else if (Sse2.IsSse2Supported)
{
ushort2 y;
ushort2 n = 16;
ushort2 mask;
y = x >> 8;
mask = Sse2.cmpeq_epi16(y, default(v128));
n = Mask.BlendV(n - 8, n, mask);
x = Mask.BlendV(y, x, mask);
y = x >> 4;
mask = Sse2.cmpeq_epi16(y, default(v128));
n = Mask.BlendV(n - 4, n, mask);
x = Mask.BlendV(y, x, mask);
y = x >> 2;
mask = Sse2.cmpeq_epi16(y, default(v128));
n = Mask.BlendV(n - 2, n, mask);
x = Mask.BlendV(y, x, mask);
y = x >> 1;
mask = Sse2.cmpeq_epi16(y, default(v128));
return(Mask.BlendV(n - 2, n - x, mask));
}
else
{
return(new ushort2(lzcnt(x.x), lzcnt(x.y)));
}
}
public ushort8(ushort2 x01, ushort2 x23, ushort2 x45, ushort2 x67)
{
if (Sse2.IsSse2Supported)
{
this = new ushort8(new ushort4(x01, x23), new ushort4(x45, x67));
}
else
{
this.x0 = x01.x;
this.x1 = x01.y;
this.x2 = x23.x;
this.x3 = x23.y;
this.x4 = x45.x;
this.x5 = x45.y;
this.x6 = x67.x;
this.x7 = x67.y;
}
}
