arno/hashcat
1
0
Fork 0
mirror of https://github.com/hashcat/hashcat.git synced 2024-11-26 09:58:16 +00:00
Code Issues Releases Wiki Activity

Switch HIP vector datatypes to OpenCL like ext_vector_type()

Browse Source
This commit is contained in:
Jens Steube 2021-07-19 20:24:30 +02:00
parent 9421e1f89d
commit bdb7999f07
12 changed files with 142 additions and 799 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
OpenCL/inc_hash_md4.h
View File

@ -31,7 +31,7 @@
#define MD4_STEP(f,a,b,c,d,x,K,s) \
{ \
a += K; \
a += make_u32x (K); \
a = hc_add3 (a, x, f (b, c, d)); \
a = hc_rotl32 (a, s); \
}

2
OpenCL/inc_hash_md5.h
View File

@ -36,7 +36,7 @@
#define MD5_STEP(f,a,b,c,d,x,K,s) \
{ \
a += K; \
a += make_u32x (K); \
a = hc_add3 (a, x, f (b, c, d)); \
a = hc_rotl32 (a, s); \
a += b; \

4
OpenCL/inc_hash_ripemd160.h
View File

@ -32,7 +32,7 @@
#define RIPEMD160_STEP(f,a,b,c,d,e,x,K,s) \
{ \
a += K; \
a += make_u32x (K); \
a += x; \
a += f (b, c, d); \
a = hc_rotl32 (a, s); \
@ -54,7 +54,7 @@
#define RIPEMD160_STEP_WORKAROUND_BUG(f,a,b,c,d,e,x,K,s) \
{ \
a += K; \
a += make_u32x (K); \
a += x; \
a += f (b, c, d); \
a = ROTATE_LEFT_WORKAROUND_BUG (a, s); \

2
OpenCL/inc_hash_sha1.h
View File

@ -28,7 +28,7 @@
#define SHA1_STEP(f,a,b,c,d,e,x) \
{ \
e += K; \
e += make_u32x (K); \
e = hc_add3 (e, x, f (b, c, d)); \
e += hc_rotl32 (a, 5u); \
b = hc_rotl32 (b, 30u); \

2
OpenCL/inc_hash_sha384.h
View File

@ -44,7 +44,7 @@
#define SHA384_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K) \
{ \
h += K; \
h += make_u32x (K); \
h += x; \
h += SHA384_S1 (e); \
h += F0 (e, f, g); \

2
OpenCL/inc_hash_sha512.h
View File

@ -44,7 +44,7 @@
#define SHA512_STEP(F0,F1,a,b,c,d,e,f,g,h,x,K) \
{ \
h += K; \
h += make_u32x (K); \
h += x; \
h += SHA512_S1 (e); \
h += F0 (e, f, g); \

17
OpenCL/inc_platform.cl
View File

@ -220,20 +220,19 @@ extern "C" __device__ __attribute__((const)) size_t __ockl_get_group_id(uint);
extern "C" __device__ __attribute__((const)) size_t __ockl_get_local_size(uint);
extern "C" __device__ __attribute__((const)) size_t __ockl_get_num_groups(uint);
DECLSPEC size_t get_global_id (const u32 dimindx __attribute__((unused)))
DECLSPEC size_t get_global_id (const u32 dimindx)
{
return (__ockl_get_group_id (0) * __ockl_get_local_size (0)) + __ockl_get_local_id (0);
return (__ockl_get_group_id (dimindx) * __ockl_get_local_size (dimindx)) + __ockl_get_local_id (dimindx);
}
DECLSPEC size_t get_local_id (const u32 dimindx __attribute__((unused)))
DECLSPEC size_t get_local_id (const u32 dimindx)
{
return __ockl_get_local_id (0);
return __ockl_get_local_id (dimindx);
}
DECLSPEC size_t get_local_size (const u32 dimindx __attribute__((unused)))
DECLSPEC size_t get_local_size (const u32 dimindx)
{
// verify
return __ockl_get_local_size (0);
return __ockl_get_local_size (dimindx);
}
DECLSPEC u32x rotl32 (const u32x a, const int n)
@ -309,8 +308,10 @@ DECLSPEC u64 rotr64_S (const u64 a, const int n)
return out.v64;
}
#define FIXED_THREAD_COUNT(n) __attribute__((amdgpu_flat_work_group_size (1, (n))))
extern "C" __device__ int printf(const char *fmt, ...);
//int printf(__constant const char* st, ...) __attribute__((format(printf, 1, 2)));
#define FIXED_THREAD_COUNT(n) __attribute__((amdgpu_flat_work_group_size (1, (n))))
#define SYNC_THREADS() __builtin_amdgcn_s_barrier ()
#endif

6
OpenCL/inc_platform.h
View File

@ -48,9 +48,9 @@ DECLSPEC u32 hc_atomic_dec (volatile GLOBAL_AS u32 *p);
DECLSPEC u32 hc_atomic_inc (volatile GLOBAL_AS u32 *p);
DECLSPEC u32 hc_atomic_or (volatile GLOBAL_AS u32 *p, volatile const u32 val);
DECLSPEC size_t get_global_id (const u32 dimindx __attribute__((unused)));
DECLSPEC size_t get_local_id (const u32 dimindx __attribute__((unused)));
DECLSPEC size_t get_local_size (const u32 dimindx __attribute__((unused)));
DECLSPEC size_t get_global_id (const u32 dimindx);
DECLSPEC size_t get_local_id (const u32 dimindx);
DECLSPEC size_t get_local_size (const u32 dimindx);
DECLSPEC u32x rotl32 (const u32x a, const int n);
DECLSPEC u32x rotr32 (const u32x a, const int n);

847
OpenCL/inc_types.h
View File

@ -17,22 +17,93 @@
#endif
#ifdef IS_CUDA
//https://docs.nvidia.com/cuda/nvrtc/index.html#integer-size
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef unsigned int uint;
typedef unsigned long long ulong;
// https://docs.nvidia.com/cuda/nvrtc/index.html#integer-size
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef unsigned int uint;
typedef unsigned long long ulong;
#endif
#ifdef IS_HIP
typedef unsigned char uchar;
typedef unsigned short ushort;
typedef unsigned int uint;
typedef unsigned long long ulong;
// https://github.com/llvm-mirror/clang/blob/master/lib/Headers/opencl-c-base.h
// built-in scalar data types:
/**
* An unsigned 8-bit integer.
*/
typedef unsigned char uchar;
/**
* An unsigned 16-bit integer.
*/
typedef unsigned short ushort;
/**
* An unsigned 32-bit integer.
*/
typedef unsigned int uint;
/**
* An unsigned 64-bit integer.
*/
typedef unsigned long ulong;
/**
* The unsigned integer type of the result of the sizeof operator. This
* is a 32-bit unsigned integer if CL_DEVICE_ADDRESS_BITS
* defined in table 4.3 is 32-bits and is a 64-bit unsigned integer if
* CL_DEVICE_ADDRESS_BITS is 64-bits.
*/
typedef __SIZE_TYPE__ size_t;
// built-in vector data types:
typedef char char2 __attribute__((ext_vector_type(2)));
typedef char char3 __attribute__((ext_vector_type(3)));
typedef char char4 __attribute__((ext_vector_type(4)));
typedef char char8 __attribute__((ext_vector_type(8)));
typedef char char16 __attribute__((ext_vector_type(16)));
typedef uchar uchar2 __attribute__((ext_vector_type(2)));
typedef uchar uchar3 __attribute__((ext_vector_type(3)));
typedef uchar uchar4 __attribute__((ext_vector_type(4)));
typedef uchar uchar8 __attribute__((ext_vector_type(8)));
typedef uchar uchar16 __attribute__((ext_vector_type(16)));
typedef short short2 __attribute__((ext_vector_type(2)));
typedef short short3 __attribute__((ext_vector_type(3)));
typedef short short4 __attribute__((ext_vector_type(4)));
typedef short short8 __attribute__((ext_vector_type(8)));
typedef short short16 __attribute__((ext_vector_type(16)));
typedef ushort ushort2 __attribute__((ext_vector_type(2)));
typedef ushort ushort3 __attribute__((ext_vector_type(3)));
typedef ushort ushort4 __attribute__((ext_vector_type(4)));
typedef ushort ushort8 __attribute__((ext_vector_type(8)));
typedef ushort ushort16 __attribute__((ext_vector_type(16)));
typedef int int2 __attribute__((ext_vector_type(2)));
typedef int int3 __attribute__((ext_vector_type(3)));
typedef int int4 __attribute__((ext_vector_type(4)));
typedef int int8 __attribute__((ext_vector_type(8)));
typedef int int16 __attribute__((ext_vector_type(16)));
typedef uint uint2 __attribute__((ext_vector_type(2)));
typedef uint uint3 __attribute__((ext_vector_type(3)));
typedef uint uint4 __attribute__((ext_vector_type(4)));
typedef uint uint8 __attribute__((ext_vector_type(8)));
typedef uint uint16 __attribute__((ext_vector_type(16)));
typedef long long2 __attribute__((ext_vector_type(2)));
typedef long long3 __attribute__((ext_vector_type(3)));
typedef long long4 __attribute__((ext_vector_type(4)));
typedef long long8 __attribute__((ext_vector_type(8)));
typedef long long16 __attribute__((ext_vector_type(16)));
typedef ulong ulong2 __attribute__((ext_vector_type(2)));
typedef ulong ulong3 __attribute__((ext_vector_type(3)));
typedef ulong ulong4 __attribute__((ext_vector_type(4)));
typedef ulong ulong8 __attribute__((ext_vector_type(8)));
typedef ulong ulong16 __attribute__((ext_vector_type(16)));
typedef float float2 __attribute__((ext_vector_type(2)));
typedef float float3 __attribute__((ext_vector_type(3)));
typedef float float4 __attribute__((ext_vector_type(4)));
typedef float float8 __attribute__((ext_vector_type(8)));
typedef float float16 __attribute__((ext_vector_type(16)));
#endif
#ifdef KERNEL_STATIC
@ -835,764 +906,8 @@ typedef __device_builtin__ struct u64x u64x;
#define make_u32x u32x
#define make_u64x u64x
#elif defined IS_HIP
#if VECT_SIZE == 2
struct u8x
{
u8 s0;
u8 s1;
inline __device__ u8x (const u8 a, const u8 b) : s0(a), s1(b) { }
inline __device__ u8x (const u8 a) : s0(a), s1(a) { }
inline __device__ u8x (void) : s0(0), s1(0) { }
inline __device__ ~u8x (void) { }
};
struct u16x
{
u16 s0;
u16 s1;
inline __device__ u16x (const u16 a, const u16 b) : s0(a), s1(b) { }
inline __device__ u16x (const u16 a) : s0(a), s1(a) { }
inline __device__ u16x (void) : s0(0), s1(0) { }
inline __device__ ~u16x (void) { }
};
struct u32x
{
u32 s0;
u32 s1;
inline __device__ u32x (const u32 a, const u32 b) : s0(a), s1(b) { }
inline __device__ u32x (const u32 a) : s0(a), s1(a) { }
inline __device__ u32x (void) : s0(0), s1(0) { }
inline __device__ ~u32x (void) { }
};
struct u64x
{
u64 s0;
u64 s1;
inline __device__ u64x (const u64 a, const u64 b) : s0(a), s1(b) { }
inline __device__ u64x (const u64 a) : s0(a), s1(a) { }
inline __device__ u64x (void) : s0(0), s1(0) { }
inline __device__ ~u64x (void) { }
};
inline __device__ bool operator != (const u32x a, const u32 b) { return ((a.s0 != b) && (a.s1 != b)); }
inline __device__ bool operator != (const u32x a, const u32x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1)); }
inline __device__ void operator ^= (u32x &a, const u32 b) { a.s0 ^= b; a.s1 ^= b; }
inline __device__ void operator ^= (u32x &a, const u32x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; }
inline __device__ void operator |= (u32x &a, const u32 b) { a.s0 |= b; a.s1 |= b; }
inline __device__ void operator |= (u32x &a, const u32x b) { a.s0 |= b.s0; a.s1 |= b.s1; }
inline __device__ void operator &= (u32x &a, const u32 b) { a.s0 &= b; a.s1 &= b; }
inline __device__ void operator &= (u32x &a, const u32x b) { a.s0 &= b.s0; a.s1 &= b.s1; }
inline __device__ void operator += (u32x &a, const u32 b) { a.s0 += b; a.s1 += b; }
inline __device__ void operator += (u32x &a, const u32x b) { a.s0 += b.s0; a.s1 += b.s1; }
inline __device__ void operator -= (u32x &a, const u32 b) { a.s0 -= b; a.s1 -= b; }
inline __device__ void operator -= (u32x &a, const u32x b) { a.s0 -= b.s0; a.s1 -= b.s1; }
inline __device__ void operator *= (u32x &a, const u32 b) { a.s0 *= b; a.s1 *= b; }
inline __device__ void operator *= (u32x &a, const u32x b) { a.s0 *= b.s0; a.s1 *= b.s1; }
inline __device__ void operator >>= (u32x &a, const u32 b) { a.s0 >>= b; a.s1 >>= b; }
inline __device__ void operator >>= (u32x &a, const u32x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; }
inline __device__ void operator <<= (u32x &a, const u32 b) { a.s0 <<= b; a.s1 <<= b; }
inline __device__ void operator <<= (u32x &a, const u32x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; }
inline __device__ u32x operator << (const u32x a, const u32 b) { return u32x ((a.s0 << b), (a.s1 << b) ); }
inline __device__ u32x operator << (const u32x a, const u32x b) { return u32x ((a.s0 << b.s0), (a.s1 << b.s1)); }
inline __device__ u32x operator >> (const u32x a, const u32 b) { return u32x ((a.s0 >> b), (a.s1 >> b) ); }
inline __device__ u32x operator >> (const u32x a, const u32x b) { return u32x ((a.s0 >> b.s0), (a.s1 >> b.s1)); }
inline __device__ u32x operator ^ (const u32x a, const u32 b) { return u32x ((a.s0 ^ b), (a.s1 ^ b) ); }
inline __device__ u32x operator ^ (const u32x a, const u32x b) { return u32x ((a.s0 ^ b.s0), (a.s1 ^ b.s1)); }
inline __device__ u32x operator | (const u32x a, const u32 b) { return u32x ((a.s0 | b), (a.s1 | b) ); }
inline __device__ u32x operator | (const u32x a, const u32x b) { return u32x ((a.s0 | b.s0), (a.s1 | b.s1)); }
inline __device__ u32x operator & (const u32x a, const u32 b) { return u32x ((a.s0 & b), (a.s1 & b) ); }
inline __device__ u32x operator & (const u32x a, const u32x b) { return u32x ((a.s0 & b.s0), (a.s1 & b.s1)); }
inline __device__ u32x operator + (const u32x a, const u32 b) { return u32x ((a.s0 + b), (a.s1 + b) ); }
inline __device__ u32x operator + (const u32x a, const u32x b) { return u32x ((a.s0 + b.s0), (a.s1 + b.s1)); }
inline __device__ u32x operator - (const u32x a, const u32 b) { return u32x ((a.s0 - b), (a.s1 - b) ); }
inline __device__ u32x operator - (const u32x a, const u32x b) { return u32x ((a.s0 - b.s0), (a.s1 - b.s1)); }
inline __device__ u32x operator * (const u32x a, const u32 b) { return u32x ((a.s0 * b), (a.s1 * b) ); }
inline __device__ u32x operator * (const u32x a, const u32x b) { return u32x ((a.s0 * b.s0), (a.s1 * b.s1)); }
inline __device__ u32x operator % (const u32x a, const u32 b) { return u32x ((a.s0 % b), (a.s1 % b) ); }
inline __device__ u32x operator % (const u32x a, const u32x b) { return u32x ((a.s0 % b.s0), (a.s1 % b.s1)); }
inline __device__ u32x operator ~ (const u32x a) { return u32x (~a.s0, ~a.s1); }
inline __device__ bool operator != (const u64x a, const u64 b) { return ((a.s0 != b) && (a.s1 != b)); }
inline __device__ bool operator != (const u64x a, const u64x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1)); }
inline __device__ void operator ^= (u64x &a, const u64 b) { a.s0 ^= b; a.s1 ^= b; }
inline __device__ void operator ^= (u64x &a, const u64x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; }
inline __device__ void operator |= (u64x &a, const u64 b) { a.s0 |= b; a.s1 |= b; }
inline __device__ void operator |= (u64x &a, const u64x b) { a.s0 |= b.s0; a.s1 |= b.s1; }
inline __device__ void operator &= (u64x &a, const u64 b) { a.s0 &= b; a.s1 &= b; }
inline __device__ void operator &= (u64x &a, const u64x b) { a.s0 &= b.s0; a.s1 &= b.s1; }
inline __device__ void operator += (u64x &a, const u64 b) { a.s0 += b; a.s1 += b; }
inline __device__ void operator += (u64x &a, const u64x b) { a.s0 += b.s0; a.s1 += b.s1; }
inline __device__ void operator -= (u64x &a, const u64 b) { a.s0 -= b; a.s1 -= b; }
inline __device__ void operator -= (u64x &a, const u64x b) { a.s0 -= b.s0; a.s1 -= b.s1; }
inline __device__ void operator *= (u64x &a, const u64 b) { a.s0 *= b; a.s1 *= b; }
inline __device__ void operator *= (u64x &a, const u64x b) { a.s0 *= b.s0; a.s1 *= b.s1; }
inline __device__ void operator >>= (u64x &a, const u64 b) { a.s0 >>= b; a.s1 >>= b; }
inline __device__ void operator >>= (u64x &a, const u64x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; }
inline __device__ void operator <<= (u64x &a, const u64 b) { a.s0 <<= b; a.s1 <<= b; }
inline __device__ void operator <<= (u64x &a, const u64x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; }
inline __device__ u64x operator << (const u64x a, const u64 b) { return u64x ((a.s0 << b), (a.s1 << b) ); }
inline __device__ u64x operator << (const u64x a, const u64x b) { return u64x ((a.s0 << b.s0), (a.s1 << b.s1)); }
inline __device__ u64x operator >> (const u64x a, const u64 b) { return u64x ((a.s0 >> b), (a.s1 >> b) ); }
inline __device__ u64x operator >> (const u64x a, const u64x b) { return u64x ((a.s0 >> b.s0), (a.s1 >> b.s1)); }
inline __device__ u64x operator ^ (const u64x a, const u64 b) { return u64x ((a.s0 ^ b), (a.s1 ^ b) ); }
inline __device__ u64x operator ^ (const u64x a, const u64x b) { return u64x ((a.s0 ^ b.s0), (a.s1 ^ b.s1)); }
inline __device__ u64x operator | (const u64x a, const u64 b) { return u64x ((a.s0 | b), (a.s1 | b) ); }
inline __device__ u64x operator | (const u64x a, const u64x b) { return u64x ((a.s0 | b.s0), (a.s1 | b.s1)); }
inline __device__ u64x operator & (const u64x a, const u64 b) { return u64x ((a.s0 & b), (a.s1 & b) ); }
inline __device__ u64x operator & (const u64x a, const u64x b) { return u64x ((a.s0 & b.s0), (a.s1 & b.s1)); }
inline __device__ u64x operator + (const u64x a, const u64 b) { return u64x ((a.s0 + b), (a.s1 + b) ); }
inline __device__ u64x operator + (const u64x a, const u64x b) { return u64x ((a.s0 + b.s0), (a.s1 + b.s1)); }
inline __device__ u64x operator - (const u64x a, const u64 b) { return u64x ((a.s0 - b), (a.s1 - b) ); }
inline __device__ u64x operator - (const u64x a, const u64x b) { return u64x ((a.s0 - b.s0), (a.s1 - b.s1)); }
inline __device__ u64x operator * (const u64x a, const u64 b) { return u64x ((a.s0 * b), (a.s1 * b) ); }
inline __device__ u64x operator * (const u64x a, const u64x b) { return u64x ((a.s0 * b.s0), (a.s1 * b.s1)); }
inline __device__ u64x operator % (const u64x a, const u64 b) { return u64x ((a.s0 % b), (a.s1 % b) ); }
inline __device__ u64x operator % (const u64x a, const u64x b) { return u64x ((a.s0 % b.s0), (a.s1 % b.s1)); }
inline __device__ u64x operator ~ (const u64x a) { return u64x (~a.s0, ~a.s1); }
#endif
#if VECT_SIZE == 4
struct u8x
{
u8 s0;
u8 s1;
u8 s2;
u8 s3;
inline __device__ u8x (const u8 a, const u8 b, const u8 c, const u8 d) : s0(a), s1(b), s2(c), s3(d) { }
inline __device__ u8x (const u8 a) : s0(a), s1(a), s2(a), s3(a) { }
inline __device__ u8x (void) : s0(0), s1(0), s2(0), s3(0) { }
inline __device__ ~u8x (void) { }
};
struct u16x
{
u16 s0;
u16 s1;
u16 s2;
u16 s3;
inline __device__ u16x (const u16 a, const u16 b, const u16 c, const u16 d) : s0(a), s1(b), s2(c), s3(d) { }
inline __device__ u16x (const u16 a) : s0(a), s1(a), s2(a), s3(a) { }
inline __device__ u16x (void) : s0(0), s1(0), s2(0), s3(0) { }
inline __device__ ~u16x (void) { }
};
struct u32x
{
u32 s0;
u32 s1;
u32 s2;
u32 s3;
inline __device__ u32x (const u32 a, const u32 b, const u32 c, const u32 d) : s0(a), s1(b), s2(c), s3(d) { }
inline __device__ u32x (const u32 a) : s0(a), s1(a), s2(a), s3(a) { }
inline __device__ u32x (void) : s0(0), s1(0), s2(0), s3(0) { }
inline __device__ ~u32x (void) { }
};
struct u64x
{
u64 s0;
u64 s1;
u64 s2;
u64 s3;
inline __device__ u64x (const u64 a, const u64 b, const u64 c, const u64 d) : s0(a), s1(b), s2(c), s3(d) { }
inline __device__ u64x (const u64 a) : s0(a), s1(a), s2(a), s3(a) { }
inline __device__ u64x (void) : s0(0), s1(0), s2(0), s3(0) { }
inline __device__ ~u64x (void) { }
};
inline __device__ bool operator != (const u32x a, const u32 b) { return ((a.s0 != b) && (a.s1 != b) && (a.s2 != b) && (a.s3 != b) ); }
inline __device__ bool operator != (const u32x a, const u32x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1) && (a.s2 != b.s2) && (a.s3 != b.s3)); }
inline __device__ void operator ^= (u32x &a, const u32 b) { a.s0 ^= b; a.s1 ^= b; a.s2 ^= b; a.s3 ^= b; }
inline __device__ void operator ^= (u32x &a, const u32x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; a.s2 ^= b.s2; a.s3 ^= b.s3; }
inline __device__ void operator |= (u32x &a, const u32 b) { a.s0 |= b; a.s1 |= b; a.s2 |= b; a.s3 |= b; }
inline __device__ void operator |= (u32x &a, const u32x b) { a.s0 |= b.s0; a.s1 |= b.s1; a.s2 |= b.s2; a.s3 |= b.s3; }
inline __device__ void operator &= (u32x &a, const u32 b) { a.s0 &= b; a.s1 &= b; a.s2 &= b; a.s3 &= b; }
inline __device__ void operator &= (u32x &a, const u32x b) { a.s0 &= b.s0; a.s1 &= b.s1; a.s2 &= b.s2; a.s3 &= b.s3; }
inline __device__ void operator += (u32x &a, const u32 b) { a.s0 += b; a.s1 += b; a.s2 += b; a.s3 += b; }
inline __device__ void operator += (u32x &a, const u32x b) { a.s0 += b.s0; a.s1 += b.s1; a.s2 += b.s2; a.s3 += b.s3; }
inline __device__ void operator -= (u32x &a, const u32 b) { a.s0 -= b; a.s1 -= b; a.s2 -= b; a.s3 -= b; }
inline __device__ void operator -= (u32x &a, const u32x b) { a.s0 -= b.s0; a.s1 -= b.s1; a.s2 -= b.s2; a.s3 -= b.s3; }
inline __device__ void operator *= (u32x &a, const u32 b) { a.s0 *= b; a.s1 *= b; a.s2 *= b; a.s3 *= b; }
inline __device__ void operator *= (u32x &a, const u32x b) { a.s0 *= b.s0; a.s1 *= b.s1; a.s2 *= b.s2; a.s3 *= b.s3; }
inline __device__ void operator >>= (u32x &a, const u32 b) { a.s0 >>= b; a.s1 >>= b; a.s2 >>= b; a.s3 >>= b; }
inline __device__ void operator >>= (u32x &a, const u32x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; a.s2 >>= b.s2; a.s3 >>= b.s3; }
inline __device__ void operator <<= (u32x &a, const u32 b) { a.s0 <<= b; a.s1 <<= b; a.s2 <<= b; a.s3 <<= b; }
inline __device__ void operator <<= (u32x &a, const u32x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; a.s2 <<= b.s2; a.s3 <<= b.s3; }
inline __device__ u32x operator << (const u32x a, const u32 b) { return u32x ((a.s0 << b), (a.s1 << b) , (a.s2 << b), (a.s3 << b) ); }
inline __device__ u32x operator << (const u32x a, const u32x b) { return u32x ((a.s0 << b.s0), (a.s1 << b.s1), (a.s2 << b.s2), (a.s3 << b.s3)); }
inline __device__ u32x operator >> (const u32x a, const u32 b) { return u32x ((a.s0 >> b), (a.s1 >> b) , (a.s2 >> b), (a.s3 >> b) ); }
inline __device__ u32x operator >> (const u32x a, const u32x b) { return u32x ((a.s0 >> b.s0), (a.s1 >> b.s1), (a.s2 >> b.s2), (a.s3 >> b.s3)); }
inline __device__ u32x operator ^ (const u32x a, const u32 b) { return u32x ((a.s0 ^ b), (a.s1 ^ b) , (a.s2 ^ b), (a.s3 ^ b) ); }
inline __device__ u32x operator ^ (const u32x a, const u32x b) { return u32x ((a.s0 ^ b.s0), (a.s1 ^ b.s1), (a.s2 ^ b.s2), (a.s3 ^ b.s3)); }
inline __device__ u32x operator | (const u32x a, const u32 b) { return u32x ((a.s0 | b), (a.s1 | b) , (a.s2 | b), (a.s3 | b) ); }
inline __device__ u32x operator | (const u32x a, const u32x b) { return u32x ((a.s0 | b.s0), (a.s1 | b.s1), (a.s2 | b.s2), (a.s3 | b.s3)); }
inline __device__ u32x operator & (const u32x a, const u32 b) { return u32x ((a.s0 & b), (a.s1 & b) , (a.s2 & b), (a.s3 & b) ); }
inline __device__ u32x operator & (const u32x a, const u32x b) { return u32x ((a.s0 & b.s0), (a.s1 & b.s1), (a.s2 & b.s2), (a.s3 & b.s3)); }
inline __device__ u32x operator + (const u32x a, const u32 b) { return u32x ((a.s0 + b), (a.s1 + b) , (a.s2 + b), (a.s3 + b) ); }
inline __device__ u32x operator + (const u32x a, const u32x b) { return u32x ((a.s0 + b.s0), (a.s1 + b.s1), (a.s2 + b.s2), (a.s3 + b.s3)); }
inline __device__ u32x operator - (const u32x a, const u32 b) { return u32x ((a.s0 - b), (a.s1 - b) , (a.s2 - b), (a.s3 - b) ); }
inline __device__ u32x operator - (const u32x a, const u32x b) { return u32x ((a.s0 - b.s0), (a.s1 - b.s1), (a.s2 - b.s2), (a.s3 - b.s3)); }
inline __device__ u32x operator * (const u32x a, const u32 b) { return u32x ((a.s0 * b), (a.s1 * b) , (a.s2 * b), (a.s3 * b) ); }
inline __device__ u32x operator * (const u32x a, const u32x b) { return u32x ((a.s0 * b.s0), (a.s1 * b.s1), (a.s2 * b.s2), (a.s3 * b.s3)); }
inline __device__ u32x operator % (const u32x a, const u32 b) { return u32x ((a.s0 % b), (a.s1 % b) , (a.s2 % b), (a.s3 % b) ); }
inline __device__ u32x operator % (const u32x a, const u32x b) { return u32x ((a.s0 % b.s0), (a.s1 % b.s1), (a.s2 % b.s2), (a.s3 % b.s3)); }
inline __device__ u32x operator ~ (const u32x a) { return u32x (~a.s0, ~a.s1, ~a.s2, ~a.s3); }
inline __device__ bool operator != (const u64x a, const u64 b) { return ((a.s0 != b) && (a.s1 != b) && (a.s2 != b) && (a.s3 != b) ); }
inline __device__ bool operator != (const u64x a, const u64x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1) && (a.s2 != b.s2) && (a.s3 != b.s3)); }
inline __device__ void operator ^= (u64x &a, const u64 b) { a.s0 ^= b; a.s1 ^= b; a.s2 ^= b; a.s3 ^= b; }
inline __device__ void operator ^= (u64x &a, const u64x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; a.s2 ^= b.s2; a.s3 ^= b.s3; }
inline __device__ void operator |= (u64x &a, const u64 b) { a.s0 |= b; a.s1 |= b; a.s2 |= b; a.s3 |= b; }
inline __device__ void operator |= (u64x &a, const u64x b) { a.s0 |= b.s0; a.s1 |= b.s1; a.s2 |= b.s2; a.s3 |= b.s3; }
inline __device__ void operator &= (u64x &a, const u64 b) { a.s0 &= b; a.s1 &= b; a.s2 &= b; a.s3 &= b; }
inline __device__ void operator &= (u64x &a, const u64x b) { a.s0 &= b.s0; a.s1 &= b.s1; a.s2 &= b.s2; a.s3 &= b.s3; }
inline __device__ void operator += (u64x &a, const u64 b) { a.s0 += b; a.s1 += b; a.s2 += b; a.s3 += b; }
inline __device__ void operator += (u64x &a, const u64x b) { a.s0 += b.s0; a.s1 += b.s1; a.s2 += b.s2; a.s3 += b.s3; }
inline __device__ void operator -= (u64x &a, const u64 b) { a.s0 -= b; a.s1 -= b; a.s2 -= b; a.s3 -= b; }
inline __device__ void operator -= (u64x &a, const u64x b) { a.s0 -= b.s0; a.s1 -= b.s1; a.s2 -= b.s2; a.s3 -= b.s3; }
inline __device__ void operator *= (u64x &a, const u64 b) { a.s0 *= b; a.s1 *= b; a.s2 *= b; a.s3 *= b; }
inline __device__ void operator *= (u64x &a, const u64x b) { a.s0 *= b.s0; a.s1 *= b.s1; a.s2 *= b.s2; a.s3 *= b.s3; }
inline __device__ void operator >>= (u64x &a, const u64 b) { a.s0 >>= b; a.s1 >>= b; a.s2 >>= b; a.s3 >>= b; }
inline __device__ void operator >>= (u64x &a, const u64x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; a.s2 >>= b.s2; a.s3 >>= b.s3; }
inline __device__ void operator <<= (u64x &a, const u64 b) { a.s0 <<= b; a.s1 <<= b; a.s2 <<= b; a.s3 <<= b; }
inline __device__ void operator <<= (u64x &a, const u64x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; a.s2 <<= b.s2; a.s3 <<= b.s3; }
inline __device__ u64x operator << (const u64x a, const u64 b) { return u64x ((a.s0 << b), (a.s1 << b) , (a.s2 << b), (a.s3 << b) ); }
inline __device__ u64x operator << (const u64x a, const u64x b) { return u64x ((a.s0 << b.s0), (a.s1 << b.s1), (a.s2 << b.s2), (a.s3 << b.s3)); }
inline __device__ u64x operator >> (const u64x a, const u64 b) { return u64x ((a.s0 >> b), (a.s1 >> b) , (a.s2 >> b), (a.s3 >> b) ); }
inline __device__ u64x operator >> (const u64x a, const u64x b) { return u64x ((a.s0 >> b.s0), (a.s1 >> b.s1), (a.s2 >> b.s2), (a.s3 >> b.s3)); }
inline __device__ u64x operator ^ (const u64x a, const u64 b) { return u64x ((a.s0 ^ b), (a.s1 ^ b) , (a.s2 ^ b), (a.s3 ^ b) ); }
inline __device__ u64x operator ^ (const u64x a, const u64x b) { return u64x ((a.s0 ^ b.s0), (a.s1 ^ b.s1), (a.s2 ^ b.s2), (a.s3 ^ b.s3)); }
inline __device__ u64x operator | (const u64x a, const u64 b) { return u64x ((a.s0 | b), (a.s1 | b) , (a.s2 | b), (a.s3 | b) ); }
inline __device__ u64x operator | (const u64x a, const u64x b) { return u64x ((a.s0 | b.s0), (a.s1 | b.s1), (a.s2 | b.s2), (a.s3 | b.s3)); }
inline __device__ u64x operator & (const u64x a, const u64 b) { return u64x ((a.s0 & b), (a.s1 & b) , (a.s2 & b), (a.s3 & b) ); }
inline __device__ u64x operator & (const u64x a, const u64x b) { return u64x ((a.s0 & b.s0), (a.s1 & b.s1), (a.s2 & b.s2), (a.s3 & b.s3)); }
inline __device__ u64x operator + (const u64x a, const u64 b) { return u64x ((a.s0 + b), (a.s1 + b) , (a.s2 + b), (a.s3 + b) ); }
inline __device__ u64x operator + (const u64x a, const u64x b) { return u64x ((a.s0 + b.s0), (a.s1 + b.s1), (a.s2 + b.s2), (a.s3 + b.s3)); }
inline __device__ u64x operator - (const u64x a, const u64 b) { return u64x ((a.s0 - b), (a.s1 - b) , (a.s2 - b), (a.s3 - b) ); }
inline __device__ u64x operator - (const u64x a, const u64x b) { return u64x ((a.s0 - b.s0), (a.s1 - b.s1), (a.s2 - b.s2), (a.s3 - b.s3)); }
inline __device__ u64x operator * (const u64x a, const u64 b) { return u64x ((a.s0 * b), (a.s1 * b) , (a.s2 * b), (a.s3 * b) ); }
inline __device__ u64x operator * (const u64x a, const u64x b) { return u64x ((a.s0 * b.s0), (a.s1 * b.s1), (a.s2 * b.s2), (a.s3 * b.s3)); }
inline __device__ u64x operator % (const u64x a, const u32 b) { return u64x ((a.s0 % b), (a.s1 % b) , (a.s2 % b), (a.s3 % b) ); }
inline __device__ u64x operator % (const u64x a, const u64x b) { return u64x ((a.s0 % b.s0), (a.s1 % b.s1), (a.s2 % b.s2), (a.s3 % b.s3)); }
inline __device__ u64x operator ~ (const u64x a) { return u64x (~a.s0, ~a.s1, ~a.s2, ~a.s3); }
#endif
#if VECT_SIZE == 8
struct u8x
{
u8 s0;
u8 s1;
u8 s2;
u8 s3;
u8 s4;
u8 s5;
u8 s6;
u8 s7;
inline __device__ u8x (const u8 a, const u8 b, const u8 c, const u8 d, const u8 e, const u8 f, const u8 g, const u8 h) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h) { }
inline __device__ u8x (const u8 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a) { }
inline __device__ u8x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0) { }
inline __device__ ~u8x (void) { }
};
struct u16x
{
u16 s0;
u16 s1;
u16 s2;
u16 s3;
u16 s4;
u16 s5;
u16 s6;
u16 s7;
inline __device__ u16x (const u16 a, const u16 b, const u16 c, const u16 d, const u16 e, const u16 f, const u16 g, const u16 h) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h) { }
inline __device__ u16x (const u16 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a) { }
inline __device__ u16x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0) { }
inline __device__ ~u16x (void) { }
};
struct u32x
{
u32 s0;
u32 s1;
u32 s2;
u32 s3;
u32 s4;
u32 s5;
u32 s6;
u32 s7;
inline __device__ u32x (const u32 a, const u32 b, const u32 c, const u32 d, const u32 e, const u32 f, const u32 g, const u32 h) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h) { }
inline __device__ u32x (const u32 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a) { }
inline __device__ u32x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0) { }
inline __device__ ~u32x (void) { }
};
struct u64x
{
u64 s0;
u64 s1;
u64 s2;
u64 s3;
u64 s4;
u64 s5;
u64 s6;
u64 s7;
inline __device__ u64x (const u64 a, const u64 b, const u64 c, const u64 d, const u64 e, const u64 f, const u64 g, const u64 h) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h) { }
inline __device__ u64x (const u64 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a) { }
inline __device__ u64x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0) { }
inline __device__ ~u64x (void) { }
};
inline __device__ bool operator != (const u32x a, const u32 b) { return ((a.s0 != b) && (a.s1 != b) && (a.s2 != b) && (a.s3 != b) && (a.s4 != b) && (a.s5 != b) && (a.s6 != b) && (a.s7 != b) ); }
inline __device__ bool operator != (const u32x a, const u32x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1) && (a.s2 != b.s2) && (a.s3 != b.s3) && (a.s4 != b.s4) && (a.s5 != b.s5) && (a.s6 != b.s6) && (a.s7 != b.s7)); }
inline __device__ void operator ^= (u32x &a, const u32 b) { a.s0 ^= b; a.s1 ^= b; a.s2 ^= b; a.s3 ^= b; a.s4 ^= b; a.s5 ^= b; a.s6 ^= b; a.s7 ^= b; }
inline __device__ void operator ^= (u32x &a, const u32x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; a.s2 ^= b.s2; a.s3 ^= b.s3; a.s4 ^= b.s4; a.s5 ^= b.s5; a.s6 ^= b.s6; a.s7 ^= b.s7; }
inline __device__ void operator |= (u32x &a, const u32 b) { a.s0 |= b; a.s1 |= b; a.s2 |= b; a.s3 |= b; a.s4 |= b; a.s5 |= b; a.s6 |= b; a.s7 |= b; }
inline __device__ void operator |= (u32x &a, const u32x b) { a.s0 |= b.s0; a.s1 |= b.s1; a.s2 |= b.s2; a.s3 |= b.s3; a.s4 |= b.s4; a.s5 |= b.s5; a.s6 |= b.s6; a.s7 |= b.s7; }
inline __device__ void operator &= (u32x &a, const u32 b) { a.s0 &= b; a.s1 &= b; a.s2 &= b; a.s3 &= b; a.s4 &= b; a.s5 &= b; a.s6 &= b; a.s7 &= b; }
inline __device__ void operator &= (u32x &a, const u32x b) { a.s0 &= b.s0; a.s1 &= b.s1; a.s2 &= b.s2; a.s3 &= b.s3; a.s4 &= b.s4; a.s5 &= b.s5; a.s6 &= b.s6; a.s7 &= b.s7; }
inline __device__ void operator += (u32x &a, const u32 b) { a.s0 += b; a.s1 += b; a.s2 += b; a.s3 += b; a.s4 += b; a.s5 += b; a.s6 += b; a.s7 += b; }
inline __device__ void operator += (u32x &a, const u32x b) { a.s0 += b.s0; a.s1 += b.s1; a.s2 += b.s2; a.s3 += b.s3; a.s4 += b.s4; a.s5 += b.s5; a.s6 += b.s6; a.s7 += b.s7; }
inline __device__ void operator -= (u32x &a, const u32 b) { a.s0 -= b; a.s1 -= b; a.s2 -= b; a.s3 -= b; a.s4 -= b; a.s5 -= b; a.s6 -= b; a.s7 -= b; }
inline __device__ void operator -= (u32x &a, const u32x b) { a.s0 -= b.s0; a.s1 -= b.s1; a.s2 -= b.s2; a.s3 -= b.s3; a.s4 -= b.s4; a.s5 -= b.s5; a.s6 -= b.s6; a.s7 -= b.s7; }
inline __device__ void operator *= (u32x &a, const u32 b) { a.s0 *= b; a.s1 *= b; a.s2 *= b; a.s3 *= b; a.s4 *= b; a.s5 *= b; a.s6 *= b; a.s7 *= b; }
inline __device__ void operator *= (u32x &a, const u32x b) { a.s0 *= b.s0; a.s1 *= b.s1; a.s2 *= b.s2; a.s3 *= b.s3; a.s4 *= b.s4; a.s5 *= b.s5; a.s6 *= b.s6; a.s7 *= b.s7; }
inline __device__ void operator >>= (u32x &a, const u32 b) { a.s0 >>= b; a.s1 >>= b; a.s2 >>= b; a.s3 >>= b; a.s4 >>= b; a.s5 >>= b; a.s6 >>= b; a.s7 >>= b; }
inline __device__ void operator >>= (u32x &a, const u32x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; a.s2 >>= b.s2; a.s3 >>= b.s3; a.s4 >>= b.s4; a.s5 >>= b.s5; a.s6 >>= b.s6; a.s7 >>= b.s7; }
inline __device__ void operator <<= (u32x &a, const u32 b) { a.s0 <<= b; a.s1 <<= b; a.s2 <<= b; a.s3 <<= b; a.s4 <<= b; a.s5 <<= b; a.s6 <<= b; a.s7 <<= b; }
inline __device__ void operator <<= (u32x &a, const u32x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; a.s2 <<= b.s2; a.s3 <<= b.s3; a.s4 <<= b.s4; a.s5 <<= b.s5; a.s6 <<= b.s6; a.s7 <<= b.s7; }
inline __device__ u32x operator << (const u32x a, const u32 b) { return u32x ((a.s0 << b), (a.s1 << b) , (a.s2 << b), (a.s3 << b) , (a.s4 << b), (a.s5 << b) , (a.s6 << b), (a.s7 << b) ); }
inline __device__ u32x operator << (const u32x a, const u32x b) { return u32x ((a.s0 << b.s0), (a.s1 << b.s1), (a.s2 << b.s2), (a.s3 << b.s3), (a.s4 << b.s4), (a.s5 << b.s5), (a.s6 << b.s6), (a.s7 << b.s7)); }
inline __device__ u32x operator >> (const u32x a, const u32 b) { return u32x ((a.s0 >> b), (a.s1 >> b) , (a.s2 >> b), (a.s3 >> b) , (a.s4 >> b), (a.s5 >> b) , (a.s6 >> b), (a.s7 >> b) ); }
inline __device__ u32x operator >> (const u32x a, const u32x b) { return u32x ((a.s0 >> b.s0), (a.s1 >> b.s1), (a.s2 >> b.s2), (a.s3 >> b.s3), (a.s4 >> b.s4), (a.s5 >> b.s5), (a.s6 >> b.s6), (a.s7 >> b.s7)); }
inline __device__ u32x operator ^ (const u32x a, const u32 b) { return u32x ((a.s0 ^ b), (a.s1 ^ b) , (a.s2 ^ b), (a.s3 ^ b) , (a.s4 ^ b), (a.s5 ^ b) , (a.s6 ^ b), (a.s7 ^ b) ); }
inline __device__ u32x operator ^ (const u32x a, const u32x b) { return u32x ((a.s0 ^ b.s0), (a.s1 ^ b.s1), (a.s2 ^ b.s2), (a.s3 ^ b.s3), (a.s4 ^ b.s4), (a.s5 ^ b.s5), (a.s6 ^ b.s6), (a.s7 ^ b.s7)); }
inline __device__ u32x operator | (const u32x a, const u32 b) { return u32x ((a.s0 | b), (a.s1 | b) , (a.s2 | b), (a.s3 | b) , (a.s4 | b), (a.s5 | b) , (a.s6 | b), (a.s7 | b) ); }
inline __device__ u32x operator | (const u32x a, const u32x b) { return u32x ((a.s0 | b.s0), (a.s1 | b.s1), (a.s2 | b.s2), (a.s3 | b.s3), (a.s4 | b.s4), (a.s5 | b.s5), (a.s6 | b.s6), (a.s7 | b.s7)); }
inline __device__ u32x operator & (const u32x a, const u32 b) { return u32x ((a.s0 & b), (a.s1 & b) , (a.s2 & b), (a.s3 & b) , (a.s4 & b), (a.s5 & b) , (a.s6 & b), (a.s7 & b) ); }
inline __device__ u32x operator & (const u32x a, const u32x b) { return u32x ((a.s0 & b.s0), (a.s1 & b.s1), (a.s2 & b.s2), (a.s3 & b.s3), (a.s4 & b.s4), (a.s5 & b.s5), (a.s6 & b.s6), (a.s7 & b.s7)); }
inline __device__ u32x operator + (const u32x a, const u32 b) { return u32x ((a.s0 + b), (a.s1 + b) , (a.s2 + b), (a.s3 + b) , (a.s4 + b), (a.s5 + b) , (a.s6 + b), (a.s7 + b) ); }
inline __device__ u32x operator + (const u32x a, const u32x b) { return u32x ((a.s0 + b.s0), (a.s1 + b.s1), (a.s2 + b.s2), (a.s3 + b.s3), (a.s4 + b.s4), (a.s5 + b.s5), (a.s6 + b.s6), (a.s7 + b.s7)); }
inline __device__ u32x operator - (const u32x a, const u32 b) { return u32x ((a.s0 - b), (a.s1 - b) , (a.s2 - b), (a.s3 - b) , (a.s4 - b), (a.s5 - b) , (a.s6 - b), (a.s7 - b) ); }
inline __device__ u32x operator - (const u32x a, const u32x b) { return u32x ((a.s0 - b.s0), (a.s1 - b.s1), (a.s2 - b.s2), (a.s3 - b.s3), (a.s4 - b.s4), (a.s5 - b.s5), (a.s6 - b.s6), (a.s7 - b.s7)); }
inline __device__ u32x operator * (const u32x a, const u32 b) { return u32x ((a.s0 * b), (a.s1 * b) , (a.s2 * b), (a.s3 * b) , (a.s4 * b), (a.s5 * b) , (a.s6 * b), (a.s7 * b) ); }
inline __device__ u32x operator * (const u32x a, const u32x b) { return u32x ((a.s0 * b.s0), (a.s1 * b.s1), (a.s2 * b.s2), (a.s3 * b.s3), (a.s4 * b.s4), (a.s5 * b.s5), (a.s6 * b.s6), (a.s7 * b.s7)); }
inline __device__ u32x operator % (const u32x a, const u32 b) { return u32x ((a.s0 % b), (a.s1 % b) , (a.s2 % b), (a.s3 % b) , (a.s4 % b), (a.s5 % b) , (a.s6 % b), (a.s7 % b) ); }
inline __device__ u32x operator % (const u32x a, const u32x b) { return u32x ((a.s0 % b.s0), (a.s1 % b.s1), (a.s2 % b.s2), (a.s3 % b.s3), (a.s4 % b.s4), (a.s5 % b.s5), (a.s6 % b.s6), (a.s7 % b.s7)); }
inline __device__ u32x operator ~ (const u32x a) { return u32x (~a.s0, ~a.s1, ~a.s2, ~a.s3, ~a.s4, ~a.s5, ~a.s6, ~a.s7); }
inline __device__ bool operator != (const u64x a, const u64 b) { return ((a.s0 != b) && (a.s1 != b) && (a.s2 != b) && (a.s3 != b) && (a.s4 != b) && (a.s5 != b) && (a.s6 != b) && (a.s7 != b) ); }
inline __device__ bool operator != (const u64x a, const u64x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1) && (a.s2 != b.s2) && (a.s3 != b.s3) && (a.s4 != b.s4) && (a.s5 != b.s5) && (a.s6 != b.s6) && (a.s7 != b.s7)); }
inline __device__ void operator ^= (u64x &a, const u64 b) { a.s0 ^= b; a.s1 ^= b; a.s2 ^= b; a.s3 ^= b; a.s4 ^= b; a.s5 ^= b; a.s6 ^= b; a.s7 ^= b; }
inline __device__ void operator ^= (u64x &a, const u64x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; a.s2 ^= b.s2; a.s3 ^= b.s3; a.s4 ^= b.s4; a.s5 ^= b.s5; a.s6 ^= b.s6; a.s7 ^= b.s7; }
inline __device__ void operator |= (u64x &a, const u64 b) { a.s0 |= b; a.s1 |= b; a.s2 |= b; a.s3 |= b; a.s4 |= b; a.s5 |= b; a.s6 |= b; a.s7 |= b; }
inline __device__ void operator |= (u64x &a, const u64x b) { a.s0 |= b.s0; a.s1 |= b.s1; a.s2 |= b.s2; a.s3 |= b.s3; a.s4 |= b.s4; a.s5 |= b.s5; a.s6 |= b.s6; a.s7 |= b.s7; }
inline __device__ void operator &= (u64x &a, const u64 b) { a.s0 &= b; a.s1 &= b; a.s2 &= b; a.s3 &= b; a.s4 &= b; a.s5 &= b; a.s6 &= b; a.s7 &= b; }
inline __device__ void operator &= (u64x &a, const u64x b) { a.s0 &= b.s0; a.s1 &= b.s1; a.s2 &= b.s2; a.s3 &= b.s3; a.s4 &= b.s4; a.s5 &= b.s5; a.s6 &= b.s6; a.s7 &= b.s7; }
inline __device__ void operator += (u64x &a, const u64 b) { a.s0 += b; a.s1 += b; a.s2 += b; a.s3 += b; a.s4 += b; a.s5 += b; a.s6 += b; a.s7 += b; }
inline __device__ void operator += (u64x &a, const u64x b) { a.s0 += b.s0; a.s1 += b.s1; a.s2 += b.s2; a.s3 += b.s3; a.s4 += b.s4; a.s5 += b.s5; a.s6 += b.s6; a.s7 += b.s7; }
inline __device__ void operator -= (u64x &a, const u64 b) { a.s0 -= b; a.s1 -= b; a.s2 -= b; a.s3 -= b; a.s4 -= b; a.s5 -= b; a.s6 -= b; a.s7 -= b; }
inline __device__ void operator -= (u64x &a, const u64x b) { a.s0 -= b.s0; a.s1 -= b.s1; a.s2 -= b.s2; a.s3 -= b.s3; a.s4 -= b.s4; a.s5 -= b.s5; a.s6 -= b.s6; a.s7 -= b.s7; }
inline __device__ void operator *= (u64x &a, const u64 b) { a.s0 *= b; a.s1 *= b; a.s2 *= b; a.s3 *= b; a.s4 *= b; a.s5 *= b; a.s6 *= b; a.s7 *= b; }
inline __device__ void operator *= (u64x &a, const u64x b) { a.s0 *= b.s0; a.s1 *= b.s1; a.s2 *= b.s2; a.s3 *= b.s3; a.s4 *= b.s4; a.s5 *= b.s5; a.s6 *= b.s6; a.s7 *= b.s7; }
inline __device__ void operator >>= (u64x &a, const u64 b) { a.s0 >>= b; a.s1 >>= b; a.s2 >>= b; a.s3 >>= b; a.s4 >>= b; a.s5 >>= b; a.s6 >>= b; a.s7 >>= b; }
inline __device__ void operator >>= (u64x &a, const u64x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; a.s2 >>= b.s2; a.s3 >>= b.s3; a.s4 >>= b.s4; a.s5 >>= b.s5; a.s6 >>= b.s6; a.s7 >>= b.s7; }
inline __device__ void operator <<= (u64x &a, const u64 b) { a.s0 <<= b; a.s1 <<= b; a.s2 <<= b; a.s3 <<= b; a.s4 <<= b; a.s5 <<= b; a.s6 <<= b; a.s7 <<= b; }
inline __device__ void operator <<= (u64x &a, const u64x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; a.s2 <<= b.s2; a.s3 <<= b.s3; a.s4 <<= b.s4; a.s5 <<= b.s5; a.s6 <<= b.s6; a.s7 <<= b.s7; }
inline __device__ u64x operator << (const u64x a, const u64 b) { return u64x ((a.s0 << b), (a.s1 << b) , (a.s2 << b), (a.s3 << b) , (a.s4 << b), (a.s5 << b) , (a.s6 << b), (a.s7 << b) ); }
inline __device__ u64x operator << (const u64x a, const u64x b) { return u64x ((a.s0 << b.s0), (a.s1 << b.s1), (a.s2 << b.s2), (a.s3 << b.s3), (a.s4 << b.s4), (a.s5 << b.s5), (a.s6 << b.s6), (a.s7 << b.s7)); }
inline __device__ u64x operator >> (const u64x a, const u64 b) { return u64x ((a.s0 >> b), (a.s1 >> b) , (a.s2 >> b), (a.s3 >> b) , (a.s4 >> b), (a.s5 >> b) , (a.s6 >> b), (a.s7 >> b) ); }
inline __device__ u64x operator >> (const u64x a, const u64x b) { return u64x ((a.s0 >> b.s0), (a.s1 >> b.s1), (a.s2 >> b.s2), (a.s3 >> b.s3), (a.s4 >> b.s4), (a.s5 >> b.s5), (a.s6 >> b.s6), (a.s7 >> b.s7)); }
inline __device__ u64x operator ^ (const u64x a, const u64 b) { return u64x ((a.s0 ^ b), (a.s1 ^ b) , (a.s2 ^ b), (a.s3 ^ b) , (a.s4 ^ b), (a.s5 ^ b) , (a.s6 ^ b), (a.s7 ^ b) ); }
inline __device__ u64x operator ^ (const u64x a, const u64x b) { return u64x ((a.s0 ^ b.s0), (a.s1 ^ b.s1), (a.s2 ^ b.s2), (a.s3 ^ b.s3), (a.s4 ^ b.s4), (a.s5 ^ b.s5), (a.s6 ^ b.s6), (a.s7 ^ b.s7)); }
inline __device__ u64x operator | (const u64x a, const u64 b) { return u64x ((a.s0 | b), (a.s1 | b) , (a.s2 | b), (a.s3 | b) , (a.s4 | b), (a.s5 | b) , (a.s6 | b), (a.s7 | b) ); }
inline __device__ u64x operator | (const u64x a, const u64x b) { return u64x ((a.s0 | b.s0), (a.s1 | b.s1), (a.s2 | b.s2), (a.s3 | b.s3), (a.s4 | b.s4), (a.s5 | b.s5), (a.s6 | b.s6), (a.s7 | b.s7)); }
inline __device__ u64x operator & (const u64x a, const u64 b) { return u64x ((a.s0 & b), (a.s1 & b) , (a.s2 & b), (a.s3 & b) , (a.s4 & b), (a.s5 & b) , (a.s6 & b), (a.s7 & b) ); }
inline __device__ u64x operator & (const u64x a, const u64x b) { return u64x ((a.s0 & b.s0), (a.s1 & b.s1), (a.s2 & b.s2), (a.s3 & b.s3), (a.s4 & b.s4), (a.s5 & b.s5), (a.s6 & b.s6), (a.s7 & b.s7)); }
inline __device__ u64x operator + (const u64x a, const u64 b) { return u64x ((a.s0 + b), (a.s1 + b) , (a.s2 + b), (a.s3 + b) , (a.s4 + b), (a.s5 + b) , (a.s6 + b), (a.s7 + b) ); }
inline __device__ u64x operator + (const u64x a, const u64x b) { return u64x ((a.s0 + b.s0), (a.s1 + b.s1), (a.s2 + b.s2), (a.s3 + b.s3), (a.s4 + b.s4), (a.s5 + b.s5), (a.s6 + b.s6), (a.s7 + b.s7)); }
inline __device__ u64x operator - (const u64x a, const u64 b) { return u64x ((a.s0 - b), (a.s1 - b) , (a.s2 - b), (a.s3 - b) , (a.s4 - b), (a.s5 - b) , (a.s6 - b), (a.s7 - b) ); }
inline __device__ u64x operator - (const u64x a, const u64x b) { return u64x ((a.s0 - b.s0), (a.s1 - b.s1), (a.s2 - b.s2), (a.s3 - b.s3), (a.s4 - b.s4), (a.s5 - b.s5), (a.s6 - b.s6), (a.s7 - b.s7)); }
inline __device__ u64x operator * (const u64x a, const u64 b) { return u64x ((a.s0 * b), (a.s1 * b) , (a.s2 * b), (a.s3 * b) , (a.s4 * b), (a.s5 * b) , (a.s6 * b), (a.s7 * b) ); }
inline __device__ u64x operator * (const u64x a, const u64x b) { return u64x ((a.s0 * b.s0), (a.s1 * b.s1), (a.s2 * b.s2), (a.s3 * b.s3), (a.s4 * b.s4), (a.s5 * b.s5), (a.s6 * b.s6), (a.s7 * b.s7)); }
inline __device__ u64x operator % (const u64x a, const u64 b) { return u64x ((a.s0 % b), (a.s1 % b) , (a.s2 % b), (a.s3 % b) , (a.s4 % b), (a.s5 % b) , (a.s6 % b), (a.s7 % b) ); }
inline __device__ u64x operator % (const u64x a, const u64x b) { return u64x ((a.s0 % b.s0), (a.s1 % b.s1), (a.s2 % b.s2), (a.s3 % b.s3), (a.s4 % b.s4), (a.s5 % b.s5), (a.s6 % b.s6), (a.s7 % b.s7)); }
inline __device__ u64x operator ~ (const u64x a) { return u64x (~a.s0, ~a.s1, ~a.s2, ~a.s3, ~a.s4, ~a.s5, ~a.s6, ~a.s7); }
#endif
#if VECT_SIZE == 16
struct u8x
{
u8 s0;
u8 s1;
u8 s2;
u8 s3;
u8 s4;
u8 s5;
u8 s6;
u8 s7;
u8 s8;
u8 s9;
u8 sa;
u8 sb;
u8 sc;
u8 sd;
u8 se;
u8 sf;
inline __device__ u8x (const u8 a, const u8 b, const u8 c, const u8 d, const u8 e, const u8 f, const u8 g, const u8 h, const u8 i, const u8 j, const u8 k, const u8 l, const u8 m, const u8 n, const u8 o, const u8 p) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h), s8(i), s9(j), sa(k), sb(l), sc(m), sd(n), se(o), sf(p) { }
inline __device__ u8x (const u8 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a), s8(a), s9(a), sa(a), sb(a), sc(a), sd(a), se(a), sf(a) { }
inline __device__ u8x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0), s8(0), s9(0), sa(0), sb(0), sc(0), sd(0), se(0), sf(0) { }
inline __device__ ~u8x (void) { }
};
struct u16x
{
u16 s0;
u16 s1;
u16 s2;
u16 s3;
u16 s4;
u16 s5;
u16 s6;
u16 s7;
u16 s8;
u16 s9;
u16 sa;
u16 sb;
u16 sc;
u16 sd;
u16 se;
u16 sf;
inline __device__ u16x (const u16 a, const u16 b, const u16 c, const u16 d, const u16 e, const u16 f, const u16 g, const u16 h, const u16 i, const u16 j, const u16 k, const u16 l, const u16 m, const u16 n, const u16 o, const u16 p) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h), s8(i), s9(j), sa(k), sb(l), sc(m), sd(n), se(o), sf(p) { }
inline __device__ u16x (const u16 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a), s8(a), s9(a), sa(a), sb(a), sc(a), sd(a), se(a), sf(a) { }
inline __device__ u16x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0), s8(0), s9(0), sa(0), sb(0), sc(0), sd(0), se(0), sf(0){ }
inline __device__ ~u16x (void) { }
};
struct u32x
{
u32 s0;
u32 s1;
u32 s2;
u32 s3;
u32 s4;
u32 s5;
u32 s6;
u32 s7;
u32 s8;
u32 s9;
u32 sa;
u32 sb;
u32 sc;
u32 sd;
u32 se;
u32 sf;
inline __device__ u32x (const u32 a, const u32 b, const u32 c, const u32 d, const u32 e, const u32 f, const u32 g, const u32 h, const u32 i, const u32 j, const u32 k, const u32 l, const u32 m, const u32 n, const u32 o, const u32 p) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h), s8(i), s9(j), sa(k), sb(l), sc(m), sd(n), se(o), sf(p) { }
inline __device__ u32x (const u32 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a), s8(a), s9(a), sa(a), sb(a), sc(a), sd(a), se(a), sf(a) { }
inline __device__ u32x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0), s8(0), s9(0), sa(0), sb(0), sc(0), sd(0), se(0), sf(0){ }
inline __device__ ~u32x (void) { }
};
struct u64x
{
u64 s0;
u64 s1;
u64 s2;
u64 s3;
u64 s4;
u64 s5;
u64 s6;
u64 s7;
u64 s8;
u64 s9;
u64 sa;
u64 sb;
u64 sc;
u64 sd;
u64 se;
u64 sf;
inline __device__ u64x (const u64 a, const u64 b, const u64 c, const u64 d, const u64 e, const u64 f, const u64 g, const u64 h, const u64 i, const u64 j, const u64 k, const u64 l, const u64 m, const u64 n, const u64 o, const u64 p) : s0(a), s1(b), s2(c), s3(d), s4(e), s5(f), s6(g), s7(h), s8(i), s9(j), sa(k), sb(l), sc(m), sd(n), se(o), sf(p) { }
inline __device__ u64x (const u64 a) : s0(a), s1(a), s2(a), s3(a), s4(a), s5(a), s6(a), s7(a), s8(a), s9(a), sa(a), sb(a), sc(a), sd(a), se(a), sf(a) { }
inline __device__ u64x (void) : s0(0), s1(0), s2(0), s3(0), s4(0), s5(0), s6(0), s7(0), s8(0), s9(0), sa(0), sb(0), sc(0), sd(0), se(0), sf(0) { }
inline __device__ ~u64x (void) { }
};
inline __device__ bool operator != (const u32x a, const u32 b) { return ((a.s0 != b) && (a.s1 != b) && (a.s2 != b) && (a.s3 != b) && (a.s4 != b) && (a.s5 != b) && (a.s6 != b) && (a.s7 != b) && (a.s8 != b) && (a.s9 != b) && (a.sa != b) && (a.sb != b) && (a.sc != b) && (a.sd != b) && (a.se != b) && (a.sf != b) ); }
inline __device__ bool operator != (const u32x a, const u32x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1) && (a.s2 != b.s2) && (a.s3 != b.s3) && (a.s4 != b.s4) && (a.s5 != b.s5) && (a.s6 != b.s6) && (a.s7 != b.s7) && (a.s8 != b.s8) && (a.s9 != b.s9) && (a.sa != b.sa) && (a.sb != b.sb) && (a.sc != b.sc) && (a.sd != b.sd) && (a.se != b.se) && (a.sf != b.sf)); }
inline __device__ void operator ^= (u32x &a, const u32 b) { a.s0 ^= b; a.s1 ^= b; a.s2 ^= b; a.s3 ^= b; a.s4 ^= b; a.s5 ^= b; a.s6 ^= b; a.s7 ^= b; a.s8 ^= b; a.s9 ^= b; a.sa ^= b; a.sb ^= b; a.sc ^= b; a.sd ^= b; a.se ^= b; a.sf ^= b; }
inline __device__ void operator ^= (u32x &a, const u32x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; a.s2 ^= b.s2; a.s3 ^= b.s3; a.s4 ^= b.s4; a.s5 ^= b.s5; a.s6 ^= b.s6; a.s7 ^= b.s7; a.s8 ^= b.s8; a.s9 ^= b.s9; a.sa ^= b.sa; a.sb ^= b.sb; a.sc ^= b.sc; a.sd ^= b.sd; a.se ^= b.se; a.sf ^= b.sf; }
inline __device__ void operator |= (u32x &a, const u32 b) { a.s0 |= b; a.s1 |= b; a.s2 |= b; a.s3 |= b; a.s4 |= b; a.s5 |= b; a.s6 |= b; a.s7 |= b; a.s8 |= b; a.s9 |= b; a.sa |= b; a.sb |= b; a.sc |= b; a.sd |= b; a.se |= b; a.sf |= b; }
inline __device__ void operator |= (u32x &a, const u32x b) { a.s0 |= b.s0; a.s1 |= b.s1; a.s2 |= b.s2; a.s3 |= b.s3; a.s4 |= b.s4; a.s5 |= b.s5; a.s6 |= b.s6; a.s7 |= b.s7; a.s8 |= b.s8; a.s9 |= b.s9; a.sa |= b.sa; a.sb |= b.sb; a.sc |= b.sc; a.sd |= b.sd; a.se |= b.se; a.sf |= b.sf; }
inline __device__ void operator &= (u32x &a, const u32 b) { a.s0 &= b; a.s1 &= b; a.s2 &= b; a.s3 &= b; a.s4 &= b; a.s5 &= b; a.s6 &= b; a.s7 &= b; a.s8 &= b; a.s9 &= b; a.sa &= b; a.sb &= b; a.sc &= b; a.sd &= b; a.se &= b; a.sf &= b; }
inline __device__ void operator &= (u32x &a, const u32x b) { a.s0 &= b.s0; a.s1 &= b.s1; a.s2 &= b.s2; a.s3 &= b.s3; a.s4 &= b.s4; a.s5 &= b.s5; a.s6 &= b.s6; a.s7 &= b.s7; a.s8 &= b.s8; a.s9 &= b.s9; a.sa &= b.sa; a.sb &= b.sb; a.sc &= b.sc; a.sd &= b.sd; a.se &= b.se; a.sf &= b.sf; }
inline __device__ void operator += (u32x &a, const u32 b) { a.s0 += b; a.s1 += b; a.s2 += b; a.s3 += b; a.s4 += b; a.s5 += b; a.s6 += b; a.s7 += b; a.s8 += b; a.s9 += b; a.sa += b; a.sb += b; a.sc += b; a.sd += b; a.se += b; a.sf += b; }
inline __device__ void operator += (u32x &a, const u32x b) { a.s0 += b.s0; a.s1 += b.s1; a.s2 += b.s2; a.s3 += b.s3; a.s4 += b.s4; a.s5 += b.s5; a.s6 += b.s6; a.s7 += b.s7; a.s8 += b.s8; a.s9 += b.s9; a.sa += b.sa; a.sb += b.sb; a.sc += b.sc; a.sd += b.sd; a.se += b.se; a.sf += b.sf; }
inline __device__ void operator -= (u32x &a, const u32 b) { a.s0 -= b; a.s1 -= b; a.s2 -= b; a.s3 -= b; a.s4 -= b; a.s5 -= b; a.s6 -= b; a.s7 -= b; a.s8 -= b; a.s9 -= b; a.sa -= b; a.sb -= b; a.sc -= b; a.sd -= b; a.se -= b; a.sf -= b; }
inline __device__ void operator -= (u32x &a, const u32x b) { a.s0 -= b.s0; a.s1 -= b.s1; a.s2 -= b.s2; a.s3 -= b.s3; a.s4 -= b.s4; a.s5 -= b.s5; a.s6 -= b.s6; a.s7 -= b.s7; a.s8 -= b.s8; a.s9 -= b.s9; a.sa -= b.sa; a.sb -= b.sb; a.sc -= b.sc; a.sd -= b.sd; a.se -= b.se; a.sf -= b.sf; }
inline __device__ void operator *= (u32x &a, const u32 b) { a.s0 *= b; a.s1 *= b; a.s2 *= b; a.s3 *= b; a.s4 *= b; a.s5 *= b; a.s6 *= b; a.s7 *= b; a.s8 *= b; a.s9 *= b; a.sa *= b; a.sb *= b; a.sc *= b; a.sd *= b; a.se *= b; a.sf *= b; }
inline __device__ void operator *= (u32x &a, const u32x b) { a.s0 *= b.s0; a.s1 *= b.s1; a.s2 *= b.s2; a.s3 *= b.s3; a.s4 *= b.s4; a.s5 *= b.s5; a.s6 *= b.s6; a.s7 *= b.s7; a.s8 *= b.s8; a.s9 *= b.s9; a.sa *= b.sa; a.sb *= b.sb; a.sc *= b.sc; a.sd *= b.sd; a.se *= b.se; a.sf *= b.sf; }
inline __device__ void operator >>= (u32x &a, const u32 b) { a.s0 >>= b; a.s1 >>= b; a.s2 >>= b; a.s3 >>= b; a.s4 >>= b; a.s5 >>= b; a.s6 >>= b; a.s7 >>= b; a.s8 >>= b; a.s9 >>= b; a.sa >>= b; a.sb >>= b; a.sc >>= b; a.sd >>= b; a.se >>= b; a.sf >>= b; }
inline __device__ void operator >>= (u32x &a, const u32x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; a.s2 >>= b.s2; a.s3 >>= b.s3; a.s4 >>= b.s4; a.s5 >>= b.s5; a.s6 >>= b.s6; a.s7 >>= b.s7; a.s8 >>= b.s8; a.s9 >>= b.s9; a.sa >>= b.sa; a.sb >>= b.sb; a.sc >>= b.sc; a.sd >>= b.sd; a.se >>= b.se; a.sf >>= b.sf; }
inline __device__ void operator <<= (u32x &a, const u32 b) { a.s0 <<= b; a.s1 <<= b; a.s2 <<= b; a.s3 <<= b; a.s4 <<= b; a.s5 <<= b; a.s6 <<= b; a.s7 <<= b; a.s8 <<= b; a.s9 <<= b; a.sa <<= b; a.sb <<= b; a.sc <<= b; a.sd <<= b; a.se <<= b; a.sf <<= b; }
inline __device__ void operator <<= (u32x &a, const u32x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; a.s2 <<= b.s2; a.s3 <<= b.s3; a.s4 <<= b.s4; a.s5 <<= b.s5; a.s6 <<= b.s6; a.s7 <<= b.s7; a.s8 <<= b.s8; a.s9 <<= b.s9; a.sa <<= b.sa; a.sb <<= b.sb; a.sc <<= b.sc; a.sd <<= b.sd; a.se <<= b.se; a.sf <<= b.sf; }
inline __device__ u32x operator << (const u32x a, const u32 b) { return u32x ((a.s0 << b), (a.s1 << b) , (a.s2 << b), (a.s3 << b) , (a.s4 << b), (a.s5 << b) , (a.s6 << b), (a.s7 << b), (a.s8 << b), (a.s9 << b) , (a.sa << b), (a.sb << b) , (a.sc << b), (a.sd << b) , (a.se << b), (a.sf << b) ); }
inline __device__ u32x operator << (const u32x a, const u32x b) { return u32x ((a.s0 << b.s0), (a.s1 << b.s1), (a.s2 << b.s2), (a.s3 << b.s3), (a.s4 << b.s4), (a.s5 << b.s5), (a.s6 << b.s6), (a.s7 << b.s7), (a.s8 << b.s8), (a.s9 << b.s9), (a.sa << b.sa), (a.sb << b.sb), (a.sc << b.sc), (a.sd << b.sd), (a.se << b.se), (a.sf << b.sf)); }
inline __device__ u32x operator >> (const u32x a, const u32 b) { return u32x ((a.s0 >> b), (a.s1 >> b) , (a.s2 >> b), (a.s3 >> b) , (a.s4 >> b), (a.s5 >> b) , (a.s6 >> b), (a.s7 >> b), (a.s8 >> b), (a.s9 >> b) , (a.sa >> b), (a.sb >> b) , (a.sc >> b), (a.sd >> b) , (a.se >> b), (a.sf >> b) ); }
inline __device__ u32x operator >> (const u32x a, const u32x b) { return u32x ((a.s0 >> b.s0), (a.s1 >> b.s1), (a.s2 >> b.s2), (a.s3 >> b.s3), (a.s4 >> b.s4), (a.s5 >> b.s5), (a.s6 >> b.s6), (a.s7 >> b.s7), (a.s8 >> b.s8), (a.s9 >> b.s9), (a.sa >> b.sa), (a.sb >> b.sb), (a.sc >> b.sc), (a.sd >> b.sd), (a.se >> b.se), (a.sf >> b.sf)); }
inline __device__ u32x operator ^ (const u32x a, const u32 b) { return u32x ((a.s0 ^ b), (a.s1 ^ b) , (a.s2 ^ b), (a.s3 ^ b) , (a.s4 ^ b), (a.s5 ^ b) , (a.s6 ^ b), (a.s7 ^ b), (a.s8 ^ b), (a.s9 ^ b) , (a.sa ^ b), (a.sb ^ b) , (a.sc ^ b), (a.sd ^ b) , (a.se ^ b), (a.sf ^ b) ); }
inline __device__ u32x operator ^ (const u32x a, const u32x b) { return u32x ((a.s0 ^ b.s0), (a.s1 ^ b.s1), (a.s2 ^ b.s2), (a.s3 ^ b.s3), (a.s4 ^ b.s4), (a.s5 ^ b.s5), (a.s6 ^ b.s6), (a.s7 ^ b.s7), (a.s8 ^ b.s8), (a.s9 ^ b.s9), (a.sa ^ b.sa), (a.sb ^ b.sb), (a.sc ^ b.sc), (a.sd ^ b.sd), (a.se ^ b.se), (a.sf ^ b.sf)); }
inline __device__ u32x operator | (const u32x a, const u32 b) { return u32x ((a.s0 | b), (a.s1 | b) , (a.s2 | b), (a.s3 | b) , (a.s4 | b), (a.s5 | b) , (a.s6 | b), (a.s7 | b), (a.s8 | b), (a.s9 | b) , (a.sa | b), (a.sb | b) , (a.sc | b), (a.sd | b) , (a.se | b), (a.sf | b) ); }
inline __device__ u32x operator | (const u32x a, const u32x b) { return u32x ((a.s0 | b.s0), (a.s1 | b.s1), (a.s2 | b.s2), (a.s3 | b.s3), (a.s4 | b.s4), (a.s5 | b.s5), (a.s6 | b.s6), (a.s7 | b.s7), (a.s8 | b.s8), (a.s9 | b.s9), (a.sa | b.sa), (a.sb | b.sb), (a.sc | b.sc), (a.sd | b.sd), (a.se | b.se), (a.sf | b.sf)); }
inline __device__ u32x operator & (const u32x a, const u32 b) { return u32x ((a.s0 & b), (a.s1 & b) , (a.s2 & b), (a.s3 & b) , (a.s4 & b), (a.s5 & b) , (a.s6 & b), (a.s7 & b), (a.s8 & b), (a.s9 & b) , (a.sa & b), (a.sb & b) , (a.sc & b), (a.sd & b) , (a.se & b), (a.sf & b) ); }
inline __device__ u32x operator & (const u32x a, const u32x b) { return u32x ((a.s0 & b.s0), (a.s1 & b.s1), (a.s2 & b.s2), (a.s3 & b.s3), (a.s4 & b.s4), (a.s5 & b.s5), (a.s6 & b.s6), (a.s7 & b.s7), (a.s8 & b.s8), (a.s9 & b.s9), (a.sa & b.sa), (a.sb & b.sb), (a.sc & b.sc), (a.sd & b.sd), (a.se & b.se), (a.sf & b.sf)); }
inline __device__ u32x operator + (const u32x a, const u32 b) { return u32x ((a.s0 + b), (a.s1 + b) , (a.s2 + b), (a.s3 + b) , (a.s4 + b), (a.s5 + b) , (a.s6 + b), (a.s7 + b), (a.s8 + b), (a.s9 + b) , (a.sa + b), (a.sb + b) , (a.sc + b), (a.sd + b) , (a.se + b), (a.sf + b) ); }
inline __device__ u32x operator + (const u32x a, const u32x b) { return u32x ((a.s0 + b.s0), (a.s1 + b.s1), (a.s2 + b.s2), (a.s3 + b.s3), (a.s4 + b.s4), (a.s5 + b.s5), (a.s6 + b.s6), (a.s7 + b.s7), (a.s8 + b.s8), (a.s9 + b.s9), (a.sa + b.sa), (a.sb + b.sb), (a.sc + b.sc), (a.sd + b.sd), (a.se + b.se), (a.sf + b.sf)); }
inline __device__ u32x operator - (const u32x a, const u32 b) { return u32x ((a.s0 - b), (a.s1 - b) , (a.s2 - b), (a.s3 - b) , (a.s4 - b), (a.s5 - b) , (a.s6 - b), (a.s7 - b), (a.s8 - b), (a.s9 - b) , (a.sa - b), (a.sb - b) , (a.sc - b), (a.sd - b) , (a.se - b), (a.sf - b) ); }
inline __device__ u32x operator - (const u32x a, const u32x b) { return u32x ((a.s0 - b.s0), (a.s1 - b.s1), (a.s2 - b.s2), (a.s3 - b.s3), (a.s4 - b.s4), (a.s5 - b.s5), (a.s6 - b.s6), (a.s7 - b.s7), (a.s8 - b.s8), (a.s9 - b.s9), (a.sa - b.sa), (a.sb - b.sb), (a.sc - b.sc), (a.sd - b.sd), (a.se - b.se), (a.sf - b.sf)); }
inline __device__ u32x operator * (const u32x a, const u32 b) { return u32x ((a.s0 * b), (a.s1 * b) , (a.s2 * b), (a.s3 * b) , (a.s4 * b), (a.s5 * b) , (a.s6 * b), (a.s7 * b), (a.s8 * b), (a.s9 * b) , (a.sa * b), (a.sb * b) , (a.sc * b), (a.sd * b) , (a.se * b), (a.sf * b) ); }
inline __device__ u32x operator * (const u32x a, const u32x b) { return u32x ((a.s0 * b.s0), (a.s1 * b.s1), (a.s2 * b.s2), (a.s3 * b.s3), (a.s4 * b.s4), (a.s5 * b.s5), (a.s6 * b.s6), (a.s7 * b.s7), (a.s8 * b.s8), (a.s9 * b.s9), (a.sa * b.sa), (a.sb * b.sb), (a.sc * b.sc), (a.sd * b.sd), (a.se * b.se), (a.sf * b.sf)); }
inline __device__ u32x operator % (const u32x a, const u32 b) { return u32x ((a.s0 % b), (a.s1 % b) , (a.s2 % b), (a.s3 % b) , (a.s4 % b), (a.s5 % b) , (a.s6 % b), (a.s7 % b), (a.s8 % b), (a.s9 % b) , (a.sa % b), (a.sb % b) , (a.sc % b), (a.sd % b) , (a.se % b), (a.sf % b) ); }
inline __device__ u32x operator % (const u32x a, const u32x b) { return u32x ((a.s0 % b.s0), (a.s1 % b.s1), (a.s2 % b.s2), (a.s3 % b.s3), (a.s4 % b.s4), (a.s5 % b.s5), (a.s6 % b.s6), (a.s7 % b.s7), (a.s8 % b.s8), (a.s9 % b.s9), (a.sa % b.sa), (a.sb % b.sb), (a.sc % b.sc), (a.sd % b.sd), (a.se % b.se), (a.sf % b.sf)); }
inline __device__ u32x operator ~ (const u32x a) { return u32x (~a.s0, ~a.s1, ~a.s2, ~a.s3, ~a.s4, ~a.s5, ~a.s6, ~a.s7, ~a.s8, ~a.s9, ~a.sa, ~a.sb, ~a.sc, ~a.sd, ~a.se, ~a.sf); }
inline __device__ bool operator != (const u64x a, const u64 b) { return ((a.s0 != b) && (a.s1 != b) && (a.s2 != b) && (a.s3 != b) && (a.s4 != b) && (a.s5 != b) && (a.s6 != b) && (a.s7 != b) && (a.s8 != b) && (a.s9 != b) && (a.sa != b) && (a.sb != b) && (a.sc != b) && (a.sd != b) && (a.se != b) && (a.sf != b) ); }
inline __device__ bool operator != (const u64x a, const u64x b) { return ((a.s0 != b.s0) && (a.s1 != b.s1) && (a.s2 != b.s2) && (a.s3 != b.s3) && (a.s4 != b.s4) && (a.s5 != b.s5) && (a.s6 != b.s6) && (a.s7 != b.s7) && (a.s8 != b.s8) && (a.s9 != b.s9) && (a.sa != b.sa) && (a.sb != b.sb) && (a.sc != b.sc) && (a.sd != b.sd) && (a.se != b.se) && (a.sf != b.sf)); }
inline __device__ void operator ^= (u64x &a, const u64 b) { a.s0 ^= b; a.s1 ^= b; a.s2 ^= b; a.s3 ^= b; a.s4 ^= b; a.s5 ^= b; a.s6 ^= b; a.s7 ^= b; a.s8 ^= b; a.s9 ^= b; a.sa ^= b; a.sb ^= b; a.sc ^= b; a.sd ^= b; a.se ^= b; a.sf ^= b; }
inline __device__ void operator ^= (u64x &a, const u64x b) { a.s0 ^= b.s0; a.s1 ^= b.s1; a.s2 ^= b.s2; a.s3 ^= b.s3; a.s4 ^= b.s4; a.s5 ^= b.s5; a.s6 ^= b.s6; a.s7 ^= b.s7; a.s8 ^= b.s8; a.s9 ^= b.s9; a.sa ^= b.sa; a.sb ^= b.sb; a.sc ^= b.sc; a.sd ^= b.sd; a.se ^= b.se; a.sf ^= b.sf; }
inline __device__ void operator |= (u64x &a, const u64 b) { a.s0 |= b; a.s1 |= b; a.s2 |= b; a.s3 |= b; a.s4 |= b; a.s5 |= b; a.s6 |= b; a.s7 |= b; a.s8 |= b; a.s9 |= b; a.sa |= b; a.sb |= b; a.sc |= b; a.sd |= b; a.se |= b; a.sf |= b; }
inline __device__ void operator |= (u64x &a, const u64x b) { a.s0 |= b.s0; a.s1 |= b.s1; a.s2 |= b.s2; a.s3 |= b.s3; a.s4 |= b.s4; a.s5 |= b.s5; a.s6 |= b.s6; a.s7 |= b.s7; a.s8 |= b.s8; a.s9 |= b.s9; a.sa |= b.sa; a.sb |= b.sb; a.sc |= b.sc; a.sd |= b.sd; a.se |= b.se; a.sf |= b.sf; }
inline __device__ void operator &= (u64x &a, const u64 b) { a.s0 &= b; a.s1 &= b; a.s2 &= b; a.s3 &= b; a.s4 &= b; a.s5 &= b; a.s6 &= b; a.s7 &= b; a.s8 &= b; a.s9 &= b; a.sa &= b; a.sb &= b; a.sc &= b; a.sd &= b; a.se &= b; a.sf &= b; }
inline __device__ void operator &= (u64x &a, const u64x b) { a.s0 &= b.s0; a.s1 &= b.s1; a.s2 &= b.s2; a.s3 &= b.s3; a.s4 &= b.s4; a.s5 &= b.s5; a.s6 &= b.s6; a.s7 &= b.s7; a.s8 &= b.s8; a.s9 &= b.s9; a.sa &= b.sa; a.sb &= b.sb; a.sc &= b.sc; a.sd &= b.sd; a.se &= b.se; a.sf &= b.sf; }
inline __device__ void operator += (u64x &a, const u64 b) { a.s0 += b; a.s1 += b; a.s2 += b; a.s3 += b; a.s4 += b; a.s5 += b; a.s6 += b; a.s7 += b; a.s8 += b; a.s9 += b; a.sa += b; a.sb += b; a.sc += b; a.sd += b; a.se += b; a.sf += b; }
inline __device__ void operator += (u64x &a, const u64x b) { a.s0 += b.s0; a.s1 += b.s1; a.s2 += b.s2; a.s3 += b.s3; a.s4 += b.s4; a.s5 += b.s5; a.s6 += b.s6; a.s7 += b.s7; a.s8 += b.s8; a.s9 += b.s9; a.sa += b.sa; a.sb += b.sb; a.sc += b.sc; a.sd += b.sd; a.se += b.se; a.sf += b.sf; }
inline __device__ void operator -= (u64x &a, const u64 b) { a.s0 -= b; a.s1 -= b; a.s2 -= b; a.s3 -= b; a.s4 -= b; a.s5 -= b; a.s6 -= b; a.s7 -= b; a.s8 -= b; a.s9 -= b; a.sa -= b; a.sb -= b; a.sc -= b; a.sd -= b; a.se -= b; a.sf -= b; }
inline __device__ void operator -= (u64x &a, const u64x b) { a.s0 -= b.s0; a.s1 -= b.s1; a.s2 -= b.s2; a.s3 -= b.s3; a.s4 -= b.s4; a.s5 -= b.s5; a.s6 -= b.s6; a.s7 -= b.s7; a.s8 -= b.s8; a.s9 -= b.s9; a.sa -= b.sa; a.sb -= b.sb; a.sc -= b.sc; a.sd -= b.sd; a.se -= b.se; a.sf -= b.sf; }
inline __device__ void operator *= (u64x &a, const u64 b) { a.s0 *= b; a.s1 *= b; a.s2 *= b; a.s3 *= b; a.s4 *= b; a.s5 *= b; a.s6 *= b; a.s7 *= b; a.s8 *= b; a.s9 *= b; a.sa *= b; a.sb *= b; a.sc *= b; a.sd *= b; a.se *= b; a.sf *= b; }
inline __device__ void operator *= (u64x &a, const u64x b) { a.s0 *= b.s0; a.s1 *= b.s1; a.s2 *= b.s2; a.s3 *= b.s3; a.s4 *= b.s4; a.s5 *= b.s5; a.s6 *= b.s6; a.s7 *= b.s7; a.s8 *= b.s8; a.s9 *= b.s9; a.sa *= b.sa; a.sb *= b.sb; a.sc *= b.sc; a.sd *= b.sd; a.se *= b.se; a.sf *= b.sf; }
inline __device__ void operator >>= (u64x &a, const u64 b) { a.s0 >>= b; a.s1 >>= b; a.s2 >>= b; a.s3 >>= b; a.s4 >>= b; a.s5 >>= b; a.s6 >>= b; a.s7 >>= b; a.s8 >>= b; a.s9 >>= b; a.sa >>= b; a.sb >>= b; a.sc >>= b; a.sd >>= b; a.se >>= b; a.sf >>= b; }
inline __device__ void operator >>= (u64x &a, const u64x b) { a.s0 >>= b.s0; a.s1 >>= b.s1; a.s2 >>= b.s2; a.s3 >>= b.s3; a.s4 >>= b.s4; a.s5 >>= b.s5; a.s6 >>= b.s6; a.s7 >>= b.s7; a.s8 >>= b.s8; a.s9 >>= b.s9; a.sa >>= b.sa; a.sb >>= b.sb; a.sc >>= b.sc; a.sd >>= b.sd; a.se >>= b.se; a.sf >>= b.sf; }
inline __device__ void operator <<= (u64x &a, const u64 b) { a.s0 <<= b; a.s1 <<= b; a.s2 <<= b; a.s3 <<= b; a.s4 <<= b; a.s5 <<= b; a.s6 <<= b; a.s7 <<= b; a.s8 <<= b; a.s9 <<= b; a.sa <<= b; a.sb <<= b; a.sc <<= b; a.sd <<= b; a.se <<= b; a.sf <<= b; }
inline __device__ void operator <<= (u64x &a, const u64x b) { a.s0 <<= b.s0; a.s1 <<= b.s1; a.s2 <<= b.s2; a.s3 <<= b.s3; a.s4 <<= b.s4; a.s5 <<= b.s5; a.s6 <<= b.s6; a.s7 <<= b.s7; a.s8 <<= b.s8; a.s9 <<= b.s9; a.sa <<= b.sa; a.sb <<= b.sb; a.sc <<= b.sc; a.sd <<= b.sd; a.se <<= b.se; a.sf <<= b.sf; }
inline __device__ u64x operator << (const u64x a, const u64 b) { return u64x ((a.s0 << b), (a.s1 << b) , (a.s2 << b), (a.s3 << b) , (a.s4 << b), (a.s5 << b) , (a.s6 << b), (a.s7 << b), (a.s8 << b), (a.s9 << b) , (a.sa << b), (a.sb << b) , (a.sc << b), (a.sd << b) , (a.se << b), (a.sf << b) ); }
inline __device__ u64x operator << (const u64x a, const u64x b) { return u64x ((a.s0 << b.s0), (a.s1 << b.s1), (a.s2 << b.s2), (a.s3 << b.s3), (a.s4 << b.s4), (a.s5 << b.s5), (a.s6 << b.s6), (a.s7 << b.s7), (a.s8 << b.s8), (a.s9 << b.s9), (a.sa << b.sa), (a.sb << b.sb), (a.sc << b.sc), (a.sd << b.sd), (a.se << b.se), (a.sf << b.sf)); }
inline __device__ u64x operator >> (const u64x a, const u64 b) { return u64x ((a.s0 >> b), (a.s1 >> b) , (a.s2 >> b), (a.s3 >> b) , (a.s4 >> b), (a.s5 >> b) , (a.s6 >> b), (a.s7 >> b), (a.s8 >> b), (a.s9 >> b) , (a.sa >> b), (a.sb >> b) , (a.sc >> b), (a.sd >> b) , (a.se >> b), (a.sf >> b) ); }
inline __device__ u64x operator >> (const u64x a, const u64x b) { return u64x ((a.s0 >> b.s0), (a.s1 >> b.s1), (a.s2 >> b.s2), (a.s3 >> b.s3), (a.s4 >> b.s4), (a.s5 >> b.s5), (a.s6 >> b.s6), (a.s7 >> b.s7), (a.s8 >> b.s8), (a.s9 >> b.s9), (a.sa >> b.sa), (a.sb >> b.sb), (a.sc >> b.sc), (a.sd >> b.sd), (a.se >> b.se), (a.sf >> b.sf)); }
inline __device__ u64x operator ^ (const u64x a, const u64 b) { return u64x ((a.s0 ^ b), (a.s1 ^ b) , (a.s2 ^ b), (a.s3 ^ b) , (a.s4 ^ b), (a.s5 ^ b) , (a.s6 ^ b), (a.s7 ^ b), (a.s8 ^ b), (a.s9 ^ b) , (a.sa ^ b), (a.sb ^ b) , (a.sc ^ b), (a.sd ^ b) , (a.se ^ b), (a.sf ^ b) ); }
inline __device__ u64x operator ^ (const u64x a, const u64x b) { return u64x ((a.s0 ^ b.s0), (a.s1 ^ b.s1), (a.s2 ^ b.s2), (a.s3 ^ b.s3), (a.s4 ^ b.s4), (a.s5 ^ b.s5), (a.s6 ^ b.s6), (a.s7 ^ b.s7), (a.s8 ^ b.s8), (a.s9 ^ b.s9), (a.sa ^ b.sa), (a.sb ^ b.sb), (a.sc ^ b.sc), (a.sd ^ b.sd), (a.se ^ b.se), (a.sf ^ b.sf)); }
inline __device__ u64x operator | (const u64x a, const u64 b) { return u64x ((a.s0 | b), (a.s1 | b) , (a.s2 | b), (a.s3 | b) , (a.s4 | b), (a.s5 | b) , (a.s6 | b), (a.s7 | b), (a.s8 | b), (a.s9 | b) , (a.sa | b), (a.sb | b) , (a.sc | b), (a.sd | b) , (a.se | b), (a.sf | b) ); }
inline __device__ u64x operator | (const u64x a, const u64x b) { return u64x ((a.s0 | b.s0), (a.s1 | b.s1), (a.s2 | b.s2), (a.s3 | b.s3), (a.s4 | b.s4), (a.s5 | b.s5), (a.s6 | b.s6), (a.s7 | b.s7), (a.s8 | b.s8), (a.s9 | b.s9), (a.sa | b.sa), (a.sb | b.sb), (a.sc | b.sc), (a.sd | b.sd), (a.se | b.se), (a.sf | b.sf)); }
inline __device__ u64x operator & (const u64x a, const u64 b) { return u64x ((a.s0 & b), (a.s1 & b) , (a.s2 & b), (a.s3 & b) , (a.s4 & b), (a.s5 & b) , (a.s6 & b), (a.s7 & b), (a.s8 & b), (a.s9 & b) , (a.sa & b), (a.sb & b) , (a.sc & b), (a.sd & b) , (a.se & b), (a.sf & b) ); }
inline __device__ u64x operator & (const u64x a, const u64x b) { return u64x ((a.s0 & b.s0), (a.s1 & b.s1), (a.s2 & b.s2), (a.s3 & b.s3), (a.s4 & b.s4), (a.s5 & b.s5), (a.s6 & b.s6), (a.s7 & b.s7), (a.s8 & b.s8), (a.s9 & b.s9), (a.sa & b.sa), (a.sb & b.sb), (a.sc & b.sc), (a.sd & b.sd), (a.se & b.se), (a.sf & b.sf)); }
inline __device__ u64x operator + (const u64x a, const u64 b) { return u64x ((a.s0 + b), (a.s1 + b) , (a.s2 + b), (a.s3 + b) , (a.s4 + b), (a.s5 + b) , (a.s6 + b), (a.s7 + b), (a.s8 + b), (a.s9 + b) , (a.sa + b), (a.sb + b) , (a.sc + b), (a.sd + b) , (a.se + b), (a.sf + b) ); }
inline __device__ u64x operator + (const u64x a, const u64x b) { return u64x ((a.s0 + b.s0), (a.s1 + b.s1), (a.s2 + b.s2), (a.s3 + b.s3), (a.s4 + b.s4), (a.s5 + b.s5), (a.s6 + b.s6), (a.s7 + b.s7), (a.s8 + b.s8), (a.s9 + b.s9), (a.sa + b.sa), (a.sb + b.sb), (a.sc + b.sc), (a.sd + b.sd), (a.se + b.se), (a.sf + b.sf)); }
inline __device__ u64x operator - (const u64x a, const u64 b) { return u64x ((a.s0 - b), (a.s1 - b) , (a.s2 - b), (a.s3 - b) , (a.s4 - b), (a.s5 - b) , (a.s6 - b), (a.s7 - b), (a.s8 - b), (a.s9 - b) , (a.sa - b), (a.sb - b) , (a.sc - b), (a.sd - b) , (a.se - b), (a.sf - b) ); }
inline __device__ u64x operator - (const u64x a, const u64x b) { return u64x ((a.s0 - b.s0), (a.s1 - b.s1), (a.s2 - b.s2), (a.s3 - b.s3), (a.s4 - b.s4), (a.s5 - b.s5), (a.s6 - b.s6), (a.s7 - b.s7), (a.s8 - b.s8), (a.s9 - b.s9), (a.sa - b.sa), (a.sb - b.sb), (a.sc - b.sc), (a.sd - b.sd), (a.se - b.se), (a.sf - b.sf)); }
inline __device__ u64x operator * (const u64x a, const u64 b) { return u64x ((a.s0 * b), (a.s1 * b) , (a.s2 * b), (a.s3 * b) , (a.s4 * b), (a.s5 * b) , (a.s6 * b), (a.s7 * b), (a.s8 * b), (a.s9 * b) , (a.sa * b), (a.sb * b) , (a.sc * b), (a.sd * b) , (a.se * b), (a.sf * b) ); }
inline __device__ u64x operator * (const u64x a, const u64x b) { return u64x ((a.s0 * b.s0), (a.s1 * b.s1), (a.s2 * b.s2), (a.s3 * b.s3), (a.s4 * b.s4), (a.s5 * b.s5), (a.s6 * b.s6), (a.s7 * b.s7), (a.s8 * b.s8), (a.s9 * b.s9), (a.sa * b.sa), (a.sb * b.sb), (a.sc * b.sc), (a.sd * b.sd), (a.se * b.se), (a.sf * b.sf)); }
inline __device__ u64x operator % (const u64x a, const u64 b) { return u64x ((a.s0 % b), (a.s1 % b) , (a.s2 % b), (a.s3 % b) , (a.s4 % b), (a.s5 % b) , (a.s6 % b), (a.s7 % b), (a.s8 % b), (a.s9 % b) , (a.sa % b), (a.sb % b) , (a.sc % b), (a.sd % b) , (a.se % b), (a.sf % b) ); }
inline __device__ u64x operator % (const u64x a, const u64x b) { return u64x ((a.s0 % b.s0), (a.s1 % b.s1), (a.s2 % b.s2), (a.s3 % b.s3), (a.s4 % b.s4), (a.s5 % b.s5), (a.s6 % b.s6), (a.s7 % b.s7), (a.s8 % b.s8), (a.s9 % b.s9), (a.sa % b.sa), (a.sb % b.sb), (a.sc % b.sc), (a.sd % b.sd), (a.se % b.se), (a.sf % b.sf)); }
inline __device__ u64x operator ~ (const u64x a) { return u64x (~a.s0, ~a.s1, ~a.s2, ~a.s3, ~a.s4, ~a.s5, ~a.s6, ~a.s7, ~a.s8, ~a.s9, ~a.sa, ~a.sb, ~a.sc, ~a.sd, ~a.se, ~a.sf); }
#endif
typedef struct u8x u8x;
typedef struct u16x u16x;
typedef struct u32x u32x;
typedef struct u64x u64x;
#define make_u8x u8x
#define make_u16x u16x
#define make_u32x u32x
#define make_u64x u64x
#else
typedef VTYPE(uchar, VECT_SIZE) u8x;
typedef VTYPE(uchar, VECT_SIZE) u8x;
typedef VTYPE(ushort, VECT_SIZE) u16x;
typedef VTYPE(uint, VECT_SIZE) u32x;
typedef VTYPE(ulong, VECT_SIZE) u64x;

19
OpenCL/m08900-pure.cl
View File

@ -24,7 +24,7 @@ typedef struct
} scrypt_tmp_t;
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
inline __device__ uint4 operator & (const uint4 a, const u32 b) { return make_uint4 ((a.x & b ), (a.y & b ), (a.z & b ), (a.w & b )); }
inline __device__ uint4 operator << (const uint4 a, const u32 b) { return make_uint4 ((a.x << b ), (a.y << b ), (a.z << b ), (a.w << b )); }
@ -41,6 +41,15 @@ inline __device__ uint4 rotate (const uint4 a, const int n)
#endif
#if defined IS_HIP
inline __device__ uint4 rotate (const uint4 a, const int n)
{
return ((a << n) | ((a >> (32 - n))));
}
#endif
DECLSPEC uint4 hc_swap32_4 (uint4 v)
{
return (rotate ((v & 0x00FF00FF), 24u) | rotate ((v & 0xFF00FF00), 8u));
@ -57,7 +66,7 @@ DECLSPEC uint4 hc_swap32_4 (uint4 v)
#define ADD_ROTATE_XOR(r,i1,i2,s) (r) ^= rotate ((i1) + (i2), (s));
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
#define SALSA20_2R() \
{ \
@ -303,7 +312,7 @@ KERNEL_FQ void m08900_init (KERN_ATTR_TMPS (scrypt_tmp_t))
digest[6] = sha256_hmac_ctx2.opad.h[6];
digest[7] = sha256_hmac_ctx2.opad.h[7];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
const uint4 tmp0 = make_uint4 (digest[0], digest[1], digest[2], digest[3]);
const uint4 tmp1 = make_uint4 (digest[4], digest[5], digest[6], digest[7]);
#else
@ -331,7 +340,7 @@ KERNEL_FQ void m08900_init (KERN_ATTR_TMPS (scrypt_tmp_t))
uint4 X[4];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
X[0] = make_uint4 (T[0].x, T[1].y, T[2].z, T[3].w);
X[1] = make_uint4 (T[1].x, T[2].y, T[3].z, T[0].w);
X[2] = make_uint4 (T[2].x, T[3].y, T[0].z, T[1].w);
@ -441,7 +450,7 @@ KERNEL_FQ void m08900_comp (KERN_ATTR_TMPS (scrypt_tmp_t))
uint4 T[4];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
T[0] = make_uint4 (X[0].x, X[3].y, X[2].z, X[1].w);
T[1] = make_uint4 (X[1].x, X[0].y, X[3].z, X[2].w);
T[2] = make_uint4 (X[2].x, X[1].y, X[0].z, X[3].w);

19
OpenCL/m15700-pure.cl
View File

@ -31,7 +31,7 @@ typedef struct ethereum_scrypt
} ethereum_scrypt_t;
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
inline __device__ uint4 operator & (const uint4 a, const u32 b) { return make_uint4 ((a.x & b ), (a.y & b ), (a.z & b ), (a.w & b )); }
inline __device__ uint4 operator << (const uint4 a, const u32 b) { return make_uint4 ((a.x << b ), (a.y << b ), (a.z << b ), (a.w << b )); }
@ -48,6 +48,15 @@ inline __device__ uint4 rotate (const uint4 a, const int n)
#endif
#if defined IS_HIP
inline __device__ uint4 rotate (const uint4 a, const int n)
{
return ((a << n) | ((a >> (32 - n))));
}
#endif
DECLSPEC uint4 hc_swap32_4 (uint4 v)
{
return (rotate ((v & 0x00FF00FF), 24u) | rotate ((v & 0xFF00FF00), 8u));
@ -64,7 +73,7 @@ DECLSPEC uint4 hc_swap32_4 (uint4 v)
#define ADD_ROTATE_XOR(r,i1,i2,s) (r) ^= rotate ((i1) + (i2), (s));
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
#define SALSA20_2R() \
{ \
@ -439,7 +448,7 @@ KERNEL_FQ void m15700_init (KERN_ATTR_TMPS_ESALT (scrypt_tmp_t, ethereum_scrypt_
digest[6] = sha256_hmac_ctx2.opad.h[6];
digest[7] = sha256_hmac_ctx2.opad.h[7];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
const uint4 tmp0 = make_uint4 (digest[0], digest[1], digest[2], digest[3]);
const uint4 tmp1 = make_uint4 (digest[4], digest[5], digest[6], digest[7]);
#else
@ -467,7 +476,7 @@ KERNEL_FQ void m15700_init (KERN_ATTR_TMPS_ESALT (scrypt_tmp_t, ethereum_scrypt_
uint4 X[4];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
X[0] = make_uint4 (T[0].x, T[1].y, T[2].z, T[3].w);
X[1] = make_uint4 (T[1].x, T[2].y, T[3].z, T[0].w);
X[2] = make_uint4 (T[2].x, T[3].y, T[0].z, T[1].w);
@ -577,7 +586,7 @@ KERNEL_FQ void m15700_comp (KERN_ATTR_TMPS_ESALT (scrypt_tmp_t, ethereum_scrypt_
uint4 T[4];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
T[0] = make_uint4 (X[0].x, X[3].y, X[2].z, X[1].w);
T[1] = make_uint4 (X[1].x, X[0].y, X[3].z, X[2].w);
T[2] = make_uint4 (X[2].x, X[1].y, X[0].z, X[3].w);

19
OpenCL/m22700-pure.cl
View File

@ -72,7 +72,7 @@ DECLSPEC int is_valid_bitcoinj (const u32 *w)
return 1;
}
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
inline __device__ uint4 operator & (const uint4 a, const u32 b) { return make_uint4 ((a.x & b ), (a.y & b ), (a.z & b ), (a.w & b )); }
inline __device__ uint4 operator << (const uint4 a, const u32 b) { return make_uint4 ((a.x << b ), (a.y << b ), (a.z << b ), (a.w << b )); }
@ -89,6 +89,15 @@ inline __device__ uint4 rotate (const uint4 a, const int n)
#endif
#if defined IS_HIP
inline __device__ uint4 rotate (const uint4 a, const int n)
{
return ((a << n) | ((a >> (32 - n))));
}
#endif
DECLSPEC uint4 hc_swap32_4 (uint4 v)
{
return (rotate ((v & 0x00FF00FF), 24u) | rotate ((v & 0xFF00FF00), 8u));
@ -105,7 +114,7 @@ DECLSPEC uint4 hc_swap32_4 (uint4 v)
#define ADD_ROTATE_XOR(r,i1,i2,s) (r) ^= rotate ((i1) + (i2), (s));
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
#define SALSA20_2R() \
{ \
@ -374,7 +383,7 @@ KERNEL_FQ void m22700_init (KERN_ATTR_TMPS (scrypt_tmp_t))
digest[6] = sha256_hmac_ctx2.opad.h[6];
digest[7] = sha256_hmac_ctx2.opad.h[7];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
const uint4 tmp0 = make_uint4 (digest[0], digest[1], digest[2], digest[3]);
const uint4 tmp1 = make_uint4 (digest[4], digest[5], digest[6], digest[7]);
#else
@ -402,7 +411,7 @@ KERNEL_FQ void m22700_init (KERN_ATTR_TMPS (scrypt_tmp_t))
uint4 X[4];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
X[0] = make_uint4 (T[0].x, T[1].y, T[2].z, T[3].w);
X[1] = make_uint4 (T[1].x, T[2].y, T[3].z, T[0].w);
X[2] = make_uint4 (T[2].x, T[3].y, T[0].z, T[1].w);
@ -575,7 +584,7 @@ KERNEL_FQ void m22700_comp (KERN_ATTR_TMPS (scrypt_tmp_t))
uint4 T[4];
#if defined IS_CUDA || defined IS_HIP
#if defined IS_CUDA
T[0] = make_uint4 (X[0].x, X[3].y, X[2].z, X[1].w);
T[1] = make_uint4 (X[1].x, X[0].y, X[3].z, X[2].w);
T[2] = make_uint4 (X[2].x, X[1].y, X[0].z, X[3].w);