Changeset 1584 for trunk/lib/idisa_cpp


Ignore:
Timestamp:
Oct 24, 2011, 1:20:43 PM (8 years ago)
Author:
huah
Message:

fixed the bitblock::load/store for AVX; added testing modules for bitblock::srl/sll/srli/slli

Location:
trunk/lib/idisa_cpp
Files:
6 edited

Legend:

Unmodified
Added
Removed
  • trunk/lib/idisa_cpp/idisa_avx.cpp

    r1580 r1584  
    8585{
    8686public:
    87         static IDISA_ALWAYS_INLINE bitblock256_t load_unaligned(bitblock256_t* arg1);
     87        static IDISA_ALWAYS_INLINE bitblock256_t load_unaligned(float const* arg1);
    8888        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t srli(bitblock256_t arg1);
    89         static IDISA_ALWAYS_INLINE void store_aligned(bitblock256_t* arg1, bitblock256_t arg2);
     89        static IDISA_ALWAYS_INLINE void store_aligned(float* arg1, bitblock256_t arg2);
    9090        static IDISA_ALWAYS_INLINE bool all(bitblock256_t arg1);
    9191        static IDISA_ALWAYS_INLINE bool any(bitblock256_t arg1);
    9292        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock256_t arg1);
    9393        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock256_t slli(bitblock256_t arg1);
    94         static IDISA_ALWAYS_INLINE bitblock256_t load_aligned(bitblock256_t* arg1);
    95         static IDISA_ALWAYS_INLINE void store_unaligned(bitblock256_t* arg1, bitblock256_t arg2);
     94        static IDISA_ALWAYS_INLINE bitblock256_t load_aligned(float const* arg1);
     95        static IDISA_ALWAYS_INLINE void store_unaligned(float* arg1, bitblock256_t arg2);
    9696};
    9797
     
    33053305
    33063306//The total number of operations is 1
    3307 IDISA_ALWAYS_INLINE bitblock256_t bitblock256::load_unaligned(bitblock256_t* arg1)
    3308 {
    3309         return _mm256_loadu_ps((bitblock256_t*)(arg1));
     3307IDISA_ALWAYS_INLINE bitblock256_t bitblock256::load_unaligned(float const* arg1)
     3308{
     3309        return _mm256_loadu_ps((float const*)(arg1));
    33103310}
    33113311
     
    33173317
    33183318//The total number of operations is 1
    3319 IDISA_ALWAYS_INLINE void bitblock256::store_aligned(bitblock256_t* arg1, bitblock256_t arg2)
    3320 {
    3321         _mm256_store_ps((bitblock256_t*)(arg1), arg2);
     3319IDISA_ALWAYS_INLINE void bitblock256::store_aligned(float* arg1, bitblock256_t arg2)
     3320{
     3321        _mm256_store_ps((float*)(arg1), arg2);
    33223322}
    33233323
     
    33473347
    33483348//The total number of operations is 1
    3349 IDISA_ALWAYS_INLINE bitblock256_t bitblock256::load_aligned(bitblock256_t* arg1)
    3350 {
    3351         return _mm256_load_ps((bitblock256_t*)(arg1));
    3352 }
    3353 
    3354 //The total number of operations is 1
    3355 IDISA_ALWAYS_INLINE void bitblock256::store_unaligned(bitblock256_t* arg1, bitblock256_t arg2)
    3356 {
    3357         _mm256_storeu_ps((bitblock256_t*)(arg1), arg2);
     3349IDISA_ALWAYS_INLINE bitblock256_t bitblock256::load_aligned(float const* arg1)
     3350{
     3351        return _mm256_load_ps((float const*)(arg1));
     3352}
     3353
     3354//The total number of operations is 1
     3355IDISA_ALWAYS_INLINE void bitblock256::store_unaligned(float* arg1, bitblock256_t arg2)
     3356{
     3357        _mm256_storeu_ps((float*)(arg1), arg2);
    33583358}
    33593359
  • trunk/lib/idisa_cpp/idisa_sse2.cpp

    r1580 r1584  
    8989public:
    9090        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
    91         static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(bitblock128_t* arg1);
     91        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(float const* arg1);
    9292        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
    9393        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
    94         static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t* arg1, bitblock128_t arg2);
     94        static IDISA_ALWAYS_INLINE void store_aligned(float* arg1, bitblock128_t arg2);
    9595        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
    9696        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
    9797        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
    9898        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
    99         static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(bitblock128_t* arg1);
    100         static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t* arg1, bitblock128_t arg2);
     99        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(float const* arg1);
     100        static IDISA_ALWAYS_INLINE void store_unaligned(float* arg1, bitblock128_t arg2);
    101101};
    102102
     
    29862986
    29872987//The total number of operations is 1
    2988 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(bitblock128_t* arg1)
     2988IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(float const* arg1)
    29892989{
    29902990        return _mm_loadu_si128((bitblock128_t*)(arg1));
     
    29982998
    29992999//The total number of operations is 1
    3000 IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t* arg1, bitblock128_t arg2)
     3000IDISA_ALWAYS_INLINE void bitblock128::store_aligned(float* arg1, bitblock128_t arg2)
    30013001{
    30023002        _mm_store_si128((bitblock128_t*)(arg1), arg2);
     
    30343034
    30353035//The total number of operations is 1
    3036 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(bitblock128_t* arg1)
     3036IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(float const* arg1)
    30373037{
    30383038        return _mm_load_si128((bitblock128_t*)(arg1));
     
    30403040
    30413041//The total number of operations is 1
    3042 IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t* arg1, bitblock128_t arg2)
     3042IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(float* arg1, bitblock128_t arg2)
    30433043{
    30443044        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
  • trunk/lib/idisa_cpp/idisa_sse3.cpp

    r1580 r1584  
    8989public:
    9090        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
    91         static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(bitblock128_t* arg1);
     91        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(float const* arg1);
    9292        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
    9393        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
    94         static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t* arg1, bitblock128_t arg2);
     94        static IDISA_ALWAYS_INLINE void store_aligned(float* arg1, bitblock128_t arg2);
    9595        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
    9696        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
    9797        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
    9898        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
    99         static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(bitblock128_t* arg1);
    100         static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t* arg1, bitblock128_t arg2);
     99        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(float const* arg1);
     100        static IDISA_ALWAYS_INLINE void store_unaligned(float* arg1, bitblock128_t arg2);
    101101};
    102102
     
    29862986
    29872987//The total number of operations is 1
    2988 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(bitblock128_t* arg1)
     2988IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(float const* arg1)
    29892989{
    29902990        return _mm_loadu_si128((bitblock128_t*)(arg1));
     
    29982998
    29992999//The total number of operations is 1
    3000 IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t* arg1, bitblock128_t arg2)
     3000IDISA_ALWAYS_INLINE void bitblock128::store_aligned(float* arg1, bitblock128_t arg2)
    30013001{
    30023002        _mm_store_si128((bitblock128_t*)(arg1), arg2);
     
    30343034
    30353035//The total number of operations is 1
    3036 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(bitblock128_t* arg1)
     3036IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(float const* arg1)
    30373037{
    30383038        return _mm_load_si128((bitblock128_t*)(arg1));
     
    30403040
    30413041//The total number of operations is 1
    3042 IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t* arg1, bitblock128_t arg2)
     3042IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(float* arg1, bitblock128_t arg2)
    30433043{
    30443044        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
  • trunk/lib/idisa_cpp/idisa_sse4_1.cpp

    r1580 r1584  
    9090public:
    9191        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
    92         static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(bitblock128_t* arg1);
     92        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(float const* arg1);
    9393        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
    9494        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
    95         static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t* arg1, bitblock128_t arg2);
     95        static IDISA_ALWAYS_INLINE void store_aligned(float* arg1, bitblock128_t arg2);
    9696        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
    9797        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
    9898        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
    9999        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
    100         static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(bitblock128_t* arg1);
    101         static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t* arg1, bitblock128_t arg2);
     100        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(float const* arg1);
     101        static IDISA_ALWAYS_INLINE void store_unaligned(float* arg1, bitblock128_t arg2);
    102102};
    103103
     
    30183018
    30193019//The total number of operations is 1
    3020 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(bitblock128_t* arg1)
     3020IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(float const* arg1)
    30213021{
    30223022        return _mm_loadu_si128((bitblock128_t*)(arg1));
     
    30303030
    30313031//The total number of operations is 1
    3032 IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t* arg1, bitblock128_t arg2)
     3032IDISA_ALWAYS_INLINE void bitblock128::store_aligned(float* arg1, bitblock128_t arg2)
    30333033{
    30343034        _mm_store_si128((bitblock128_t*)(arg1), arg2);
     
    30663066
    30673067//The total number of operations is 1
    3068 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(bitblock128_t* arg1)
     3068IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(float const* arg1)
    30693069{
    30703070        return _mm_load_si128((bitblock128_t*)(arg1));
     
    30723072
    30733073//The total number of operations is 1
    3074 IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t* arg1, bitblock128_t arg2)
     3074IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(float* arg1, bitblock128_t arg2)
    30753075{
    30763076        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
  • trunk/lib/idisa_cpp/idisa_sse4_2.cpp

    r1580 r1584  
    9090public:
    9191        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
    92         static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(bitblock128_t* arg1);
     92        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(float const* arg1);
    9393        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
    9494        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
    95         static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t* arg1, bitblock128_t arg2);
     95        static IDISA_ALWAYS_INLINE void store_aligned(float* arg1, bitblock128_t arg2);
    9696        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
    9797        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
    9898        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
    9999        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
    100         static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(bitblock128_t* arg1);
    101         static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t* arg1, bitblock128_t arg2);
     100        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(float const* arg1);
     101        static IDISA_ALWAYS_INLINE void store_unaligned(float* arg1, bitblock128_t arg2);
    102102};
    103103
     
    30003000
    30013001//The total number of operations is 1
    3002 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(bitblock128_t* arg1)
     3002IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(float const* arg1)
    30033003{
    30043004        return _mm_loadu_si128((bitblock128_t*)(arg1));
     
    30123012
    30133013//The total number of operations is 1
    3014 IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t* arg1, bitblock128_t arg2)
     3014IDISA_ALWAYS_INLINE void bitblock128::store_aligned(float* arg1, bitblock128_t arg2)
    30153015{
    30163016        _mm_store_si128((bitblock128_t*)(arg1), arg2);
     
    30483048
    30493049//The total number of operations is 1
    3050 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(bitblock128_t* arg1)
     3050IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(float const* arg1)
    30513051{
    30523052        return _mm_load_si128((bitblock128_t*)(arg1));
     
    30543054
    30553055//The total number of operations is 1
    3056 IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t* arg1, bitblock128_t arg2)
     3056IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(float* arg1, bitblock128_t arg2)
    30573057{
    30583058        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
  • trunk/lib/idisa_cpp/idisa_ssse3.cpp

    r1580 r1584  
    9090public:
    9191        static IDISA_ALWAYS_INLINE bitblock128_t sll(bitblock128_t arg1, bitblock128_t arg2);
    92         static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(bitblock128_t* arg1);
     92        static IDISA_ALWAYS_INLINE bitblock128_t load_unaligned(float const* arg1);
    9393        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t srli(bitblock128_t arg1);
    9494        static IDISA_ALWAYS_INLINE bitblock128_t srl(bitblock128_t arg1, bitblock128_t arg2);
    95         static IDISA_ALWAYS_INLINE void store_aligned(bitblock128_t* arg1, bitblock128_t arg2);
     95        static IDISA_ALWAYS_INLINE void store_aligned(float* arg1, bitblock128_t arg2);
    9696        static IDISA_ALWAYS_INLINE bool all(bitblock128_t arg1);
    9797        static IDISA_ALWAYS_INLINE bool any(bitblock128_t arg1);
    9898        static IDISA_ALWAYS_INLINE uint64_t popcount(bitblock128_t arg1);
    9999        template <uint64_t sh> static IDISA_ALWAYS_INLINE bitblock128_t slli(bitblock128_t arg1);
    100         static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(bitblock128_t* arg1);
    101         static IDISA_ALWAYS_INLINE void store_unaligned(bitblock128_t* arg1, bitblock128_t arg2);
     100        static IDISA_ALWAYS_INLINE bitblock128_t load_aligned(float const* arg1);
     101        static IDISA_ALWAYS_INLINE void store_unaligned(float* arg1, bitblock128_t arg2);
    102102};
    103103
     
    30213021
    30223022//The total number of operations is 1
    3023 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(bitblock128_t* arg1)
     3023IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_unaligned(float const* arg1)
    30243024{
    30253025        return _mm_loadu_si128((bitblock128_t*)(arg1));
     
    30333033
    30343034//The total number of operations is 1
    3035 IDISA_ALWAYS_INLINE void bitblock128::store_aligned(bitblock128_t* arg1, bitblock128_t arg2)
     3035IDISA_ALWAYS_INLINE void bitblock128::store_aligned(float* arg1, bitblock128_t arg2)
    30363036{
    30373037        _mm_store_si128((bitblock128_t*)(arg1), arg2);
     
    30693069
    30703070//The total number of operations is 1
    3071 IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(bitblock128_t* arg1)
     3071IDISA_ALWAYS_INLINE bitblock128_t bitblock128::load_aligned(float const* arg1)
    30723072{
    30733073        return _mm_load_si128((bitblock128_t*)(arg1));
     
    30753075
    30763076//The total number of operations is 1
    3077 IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(bitblock128_t* arg1, bitblock128_t arg2)
     3077IDISA_ALWAYS_INLINE void bitblock128::store_unaligned(float* arg1, bitblock128_t arg2)
    30783078{
    30793079        _mm_storeu_si128((bitblock128_t*)(arg1), arg2);
Note: See TracChangeset for help on using the changeset viewer.