$F3
d dl Z d dlZd dlZd dlZd dlZd dlmZmZmZ d dl m
Z
d*dZ G d d Z G d d e Z
G d
d e Z G d d
e Z G d de Z G d de Z G d de ZdZdZdZeez Zee
eeeededeeeiZ ej D ]<\ ZZerej ndZ e d Z e e dk rd d e d Z ne d Z dZ! e" Z#esd e d Z!n.ej sd e d!Z!n ej$ s d e d"e#d#< ej% s d e d$e#d%< e D ]\ Z&Z'e&D ]Z(e#) e(e! Z*e'fZ+ e"ee e(e& Z, e-d'e'j. d(e d(e d(e( e+e, Z/e*r$ e j0 ! e*) e/ e/ e1 e/j. < >dS )+ N)targetsclear_floatstatusget_floatstatus)__cpu_baseline__Fc t }|s| o|dz dk }||s|o|dz dk z }||s|o|dz dk z }||s|o|dz dk z }|S )N r )r ) divbyzerooverflow underflowinvalidallerrrets n/builddir/build/BUILD/cloudlinux-venv-1.0.10/venv/lib64/python3.11/site-packages/numpy/core/tests/test_simd.pycheck_floatstatusr s}
C )
/#'a CCO8 /#'a /CC 9 037q.0CCN7 .qQ.C J c | e Zd ZdZdZdZd Zd ZddZd Z d Z
d Z d Z d
Z
d Zd Zd
Zd Zd Zd Zd Zd ZdS )
_Test_UtilityNc B t | j |dz | j z S )z
To call NPV intrinsics without the attribute 'npyv' and
auto suffixing intrinsics according to class attribute 'sfx'
_getattrnpyvsfx)selfattrs r __getattr__z_Test_Utility.__getattr__ s!
ty$*tx"7888r c B t | j | d| j d S )Nr x2r )r intrin_names r _x2z_Test_Utility._x2& s' ty["?"?48"?"?"?@@@r Fc |d}|| j }t |||z }|rt | }| r d |D S t | S )z[
Create list of consecutive numbers according to number of vector's lanes.
Nr c g | ]}|d z S ) ? .0xs r
z'_Test_Utility._data..5 s )))AG)))r )nlanesrangereversed_is_fplist)r startcountreverserngs r _dataz_Test_Utility._data) sp =E =KEE55=)) 3--C ;;== *))S))) )Cyyr c $ | j d dk S )Nr ur r s r _is_unsignedz_Test_Utility._is_unsigned8 x{c!!r c $ | j d dk S )Nr sr9 r: s r
_is_signedz_Test_Utility._is_signed; r< r c $ | j d dk S )Nr fr9 r: s r r0 z_Test_Utility._is_fp> r< r c : t | j dd S )Nr )intr r: s r _scalar_sizez_Test_Utility._scalar_sizeA s 48ABB< r c | r|S | | fd|D S )Nc L g | ] }t t | !S r( )minmax)r* vmax_intmin_ints r r, z+_Test_Utility._int_clip..I s+ ;;;!C7OOW--;;;r )r0 _int_max_int_min)r seqrJ rK s @@r _int_clipz_Test_Utility._int_clipD sP ;;== J--//--//;;;;;s;;;;r c | rd S | | d d }| r|dz S |S )Nr r )r0 _to_unsignedsetallr? )r max_us r rL z_Test_Utility._int_maxK sZ ;;== 4!!$++b//2215 ?? A: r c | rd S | rdS | dz S )Nr r )r0 r; rL r: s r rM z_Test_Utility._int_minS sF ;;== 4 1 1$%%r c n t | j d| j dd z d }|d S )Nsetall_ur rQ r r )r max_unsigs r
_true_maskz_Test_Utility._true_maskZ s6 AGDIzDHQRRL'@AA"EE |r c b t |t t f r. t | j d| j dd z | S |j dd }|d dk rd}nd} t | j | |dd | | S ) Nload_ur npyv_ r bz
cvt_u{0}_b{0}zreinterpret_u{0}_{1})
isinstancer1 tupler r r __name__replaceformat)r vectorr
cvt_intrins r rR z_Test_Utility._to_unsigned^ s ftUm , , O>749h!"" &=>>vFF F/))'266C1v}} ,
3
F749j&7&7ABB&E&EFFvNN Nr c t d S Ninffloatr: s r
_pinfinityz_Test_Utility._pinfinityi U||r c " t d S rg ri r: s r
_ninfinityz_Test_Utility._ninfinityl s e } r c t d S )Nnanri r: s r _nanz_Test_Utility._nano rl r c | j }|dk rt }n| d }d | S )Nbaseline__ ) target_namer splitjoin)r targets r
_cpu_featuresz_Test_Utility._cpu_featuresr s? ! Z %FF\\$''Fxx r NNF)ra
__module__ __qualname__r r rv r r$ r6 r; r? r0 rD rO rL rM rY rR rk rn rq rz r( r r r r s D CK9 9 9A A A
" " "" " "" " "! ! !< < < & & & O O O r r c e Zd ZdZd ZddZd Zd Zd Zd Z e
j d
d d g e
j d
ddgddgdgdgf d Z
dS )
_SIMD_BOOLz2
To test all boolean vector types at once
c L t | j d| j dd z S )Nnlanes_ur r r: s r _nlanesz_SIMD_BOOL._nlanes~ s# ty*tx|";<<. s& 777aQU) 777r )rY r. r r/ )r r2 r3 r4 r5 r s @r r6 z_SIMD_BOOL._data sT OO%% DLLNN## 3--C777737777r c | j dd }t | j d|z }t | j d| d| } | || S )Nr r[ cvt_b_u)r r r )r datalen_strloadcvts r _load_bz_SIMD_BOOL._load_b s_ (122,ty(W"455di!=!=!=G!=!=>>s44:: r c | }| d }| | }| | }d t || D } t | d || }||k sJ d t || D } t | d || }||k sJ d t || D } t | d || }
|
| k sJ t | d | } | |k sJ | j d
vrd S d t || D } t | d
|| }
| |
k sJ d t || D } t | d || }||k sJ d t || D } t | d || }||k sJ d S )z
Logical operations for boolean types.
Test intrinsics:
npyv_xor_##SFX, npyv_and_##SFX, npyv_or_##SFX, npyv_not_##SFX,
npyv_andc_b8, npvy_orc_b8, nvpy_xnor_b8
Tr4 c g | ]
\ }}||z S r( r( r* ar^ s r r, z5_SIMD_BOOL.test_operators_logical.. :::daAE:::r andc g | ]
\ }}||z S r( r( r s r r, z5_SIMD_BOOL.test_operators_logical.. s 999TQ1q5999r orc g | ]
\ }}||z S r( r( r s r r, z5_SIMD_BOOL.test_operators_logical.. r r xornotb8Nc & g | ]\ }}|| z d z S r( r s r r, z5_SIMD_BOOL.test_operators_logical.. s& EEEAa1"f_EEEr andcc & g | ]\ }}|| z d z S r r( r s r r, z5_SIMD_BOOL.test_operators_logical.. s& DDD1Q!VtODDDr orcc & g | ]\ }}||z d z S r r( r s r r, z5_SIMD_BOOL.test_operators_logical.. s& EEEAq1uX_EEEr xnor)r6 r zipr r )r data_adata_bvdata_avdata_bdata_andvanddata_orvordata_xorvxorvnot data_andcvandcdata_orcvorc data_xnorvxnors r test_operators_logicalz!_SIMD_BOOL.test_operators_logical s8 D)),,v&&,,v&&::c&&&9&9:::#wtU##GW55x 99S%8%8999!gdD!!'733g~~~~::c&&&9&9:::#wtU##GW55x #wtU##G,,v~~~~ 8D ! ! FEEVV1D1DEEE %f%%gw77E!!!!DDFF0C0CDDD#wtU##GW554 EEVV1D1DEEE %f%%gw77E!!!!!!r c d }| | d fD ][}| | } || }| | }t | }|t | k sJ \d S )Nc P t d t | d D S )Nc @ g | ]\ }}t |d k |z S )r rC )r* ir+ s r r, z<_SIMD_BOOL.test_tobits.... s+ %V%V%V41ac!q&kkQ&6%V%V%Vr r )sum enumerate)r s r z(_SIMD_BOOL.test_tobits.. s( %V%V9TSTCUCU%V%V%V!W!W r Tr )r6 r tobitsbin)r data2bitsr vdata data_bitsr
bin_tobitss r test_tobitsz_SIMD_BOOL.test_tobits s WW ZZ\\4::d:#;#;< 0 0DLL&&E! $I[[''F VJ Y ///// 0 0r c
| j dvrdS | }| d }| | }| | }t | j d| j }| j dk r6d t | t | z D } ||| }n| j dk r>d d
t | z d
t | z z D } ||||| }nL| j d k rAd d
t | z d
t | z z D } ||||||||| }||k sJ dS )z
Pack multiple vectors into one
Test intrinsics:
npyv_pack_b8_b16
npyv_pack_b8_b32
npyv_pack_b8_b64
)b16b32b64NTr pack_b8_r c g | ]}|d z S r r( r* r s r r, z(_SIMD_BOOL.test_pack.. s DDDAa$hDDDr r c g | ]}|d z S r r( r s r r, z(_SIMD_BOOL.test_pack.. HHHAa$hHHHr r r c g | ]}|d z S r r( r s r r, z(_SIMD_BOOL.test_pack.. r r r
)r r6 r r r r1 )r r rdatar vrdata pack_simdspackvpacks r test_packz_SIMD_BOOL.test_pack sw 80 0 0 Fzz||
4
(( T"" e$$DI'<$('<'<== 8u DD$u++T
*BDDDE Ife,,EE
X
HH!DKK-!DJJ,*FHHHE IffeU;;EE
X
HH!DKK-!DJJ,*FHHHE Iffff $uuu> >E~~~~~~r intrinanyr r rQ r c | || z }t | }t | | } || } || }||k sJ dS zX
Test intrinsics:
npyv_any_##SFX
npyv_all_##SFX
N)r r evalr r r r r funcdesiredsimds r test_operators_crosstestz#_SIMD_BOOL.test_operators_crosstest sh dT\\^^ 344F||v&&$v,,vf~~w&r r{ )ra r| r} __doc__r r6 r r r r pytestmark parametrizer r( r r r r z s = = =8 8 8 8 )" )" )"V0 0 0 > [ Xu~66 [ V Q
B
& ' '
76 ' ' 'r r c p e Zd ZdZd Zd Zd Zej dg d d Z
dS ) _SIMD_INTz2
To test all integer vector types at once
c | j dv rd S | | | j z
}| | d }| | | | }}t
| D ] | fd|D }| | }||k sJ | fd|D }| | }||k sJ t
d| D ] | fd|D }|
| } | |k sJ | fd|D }| | }
|
|k sJ d S ) N)u8s8Tr c g | ]}|z S r( r( r* r r3 s r r, z2_SIMD_INT.test_operators_shift.. #?#?#?1AJ#?#?#?r c g | ]}|z S r( r( r s r r, z2_SIMD_INT.test_operators_shift.. r r r c g | ]}|z S r( r( r s r r, z2_SIMD_INT.test_operators_shift.. r r c g | ]}|z S r( r( r s r r, z2_SIMD_INT.test_operators_shift.. r r ) r r6 rL r- rM r r. rD shlshrshlishri) r r r r r
data_shl_ar
data_shr_ar r r r3 s @r test_operators_shiftz _SIMD_INT.test_operators_shift s 8| # # FDMMOOdk 9::DMMOOT:: 99V,,dii.?.?4,,..//
%
%E #?#?#?#?#?#?#?@@J((7E**C*$$$$ #?#?#?#?#?#?#?@@J((7E**C*$$$$$ 1d / / 1 122
&
&E #?#?#?#?#?#?#?@@J99We,,D:%%%% #?#?#?#?#?#?#?@@J99We,,D:%%%%%
&
&r c L | j dv rd S | | | j z
}| | d }| | | | }}| d t || D }| || }||k sJ | d t || D }| || }||k sJ d S )N)u32s32u64s64Tr c g | ]
\ }}||z S r( r( r s r r, z>_SIMD_INT.test_arithmetic_subadd_saturated.." #J#J#JdaAE#J#J#Jr c g | ]
\ }}||z
S r( r( r s r r, z>_SIMD_INT.test_arithmetic_subadd_saturated..& r r )
r r6 rL r- rM r rO r addssubs) r r r r r data_addsr data_subsr s r test_arithmetic_subadd_saturatedz*_SIMD_INT.test_arithmetic_subadd_saturated s 83 3 3 FDMMOOdk 9::DMMOOT:: 99V,,dii.?.?NN#J#Jc&&6I6I#J#J#JKK yy'**y NN#J#Jc&&6I6I#J#J#JKK yy'**y r c | }| | j }| | | | }}d t || D }| || }||k sJ d t || D }| || }||k sJ d S )Nc 4 g | ]\ }}t || S r( )rH r s r r, z/_SIMD_INT.test_math_max_min../ $ >>>$!QC1II>>>r c 4 g | ]\ }}t || S r( )rG r s r r, z/_SIMD_INT.test_math_max_min..3 r r )r6 r- r r rH rG ) r r r r r data_maxsimd_maxdata_minsimd_mins r test_math_max_minz_SIMD_INT.test_math_max_min* s DK(( 99V,,dii.?.?>>#ff*=*=>>>88GW--8####>>#ff*=*=>>>88GW--8######r r2 )iir d i' c | | | }| | t | k sJ | | t | k sJ dS )zf
Test intrinsics:
npyv_reduce_max_##sfx
npyv_reduce_min_##sfx
N)r r6
reduce_maxrH
reduce_minrG )r r2 r s r test_reduce_max_minz _SIMD_INT.test_reduce_max_min7 sm ))DJJu - -..w''3w<<7777w''3w<<777777r N) ra r| r} r r r r
r r r r r( r r r r s & & &B! ! ! $ $ $ [ W&C&C&CDD8 8 ED8 8 8r r c e Zd ZdZd ZdS )
_SIMD_FP32z'
To only test single precision
c ~ | }| j j st j d| rd nt
| | }| || d }fd|D }|
| }||k sJ dS )
Round to nearest even integer, assume CPU control register is set to rounding.
Test intrinsics:
npyv_round_s32_##SFX
z.*(NEON|ASIMD)c 6 t | | dk rdndz S )Nr ? r )rI s r r z-_SIMD_FP32.test_conversions..Q s s1qAvv4#@ A A r r c & g | ]
} | S r( r( )r* r+ _rounds r r, z/_SIMD_FP32.test_conversions..V s! 111AffQii111r N) rz r simd_f64rematchroundr r6 subrS round_s32)r featuresr
data_roundvroundr s @r test_conversionsz_SIMD_FP32.test_conversionsG s %%''y! bh/@(&K&K BAFFF))DJJLL))((7DKK$4$4551111111
((######r Nra r| r} r r" r( r r r r C s- $ $ $ $ $r r c e Zd ZdZd ZdS )
_SIMD_FP64z'
To only test double precision
c | | }| || d }| || d }d t | t | z D }| || }||k sJ dS )r r g c , g | ]}t | S r( )r r) s r r, z/_SIMD_FP64.test_conversions..g s FFF1eAhhFFFr N)r r6 r rS mulr1 r )r r r r r! s r r" z_SIMD_FP64.test_conversions^ s ))DJJLL))((7DKK$4$455((7DKK$5$566FFW
W
(EFFF
11######r Nr# r( r r r% r% Z s- $ $ $ $ $r r% c
e Zd ZdZd Zd Zd Zd Zej
dde j fde j
fd e j fd
efg d Zej
d g d
d Zd Zd Zej
dg d d Zej
dej dfej dfej dfej dfej dfej dfg d Z ej
d d d g ej
d e d d gd e d g e d d!gd! e d g e d e d gd"d#gd#d"gd$d#gf d% Z d&S )'_SIMD_FPz0
To test all float vector types at once
c r | | gdz \ }}}| || }| d t ||| D }| ||| }||k sJ | ||| }| || }||k sJ | ||| } | || }
| |
k sJ | ||| } | ||
d } | | k sJ t | ||| }
|
dd d t | dd d k sJ |
dd d t | dd d k sJ d S )N c & g | ]\ }}}||z |z S r( r( )r* r r^ cs r r, z2_SIMD_FP.test_arithmetic_fused..s s& V V VGAq!a!eai V V Vr rQ r r r )
r r6 addr muladdmulsubr nmuladdnmulsubr( rS r1 muladdsub)r r r vdata_c vdata_cx2data_fmafmafmsdata_fmsnfma data_nfmanfms data_nfmsfmass r test_arithmetic_fusedz _SIMD_FP.test_arithmetic_fusedo s %)YYtzz||%<%<$=a$?!'HHWg.. 99 V Vs7GW7U7U V V VWWkk'7G44hkk'7G4488Hi00h||GWg66HHY11 y ||GWg66HHXt{{277 y DNN7GW==>>ADqDzT(^^ADqD 11111ADqDzT(^^ADqD 1111111r c B | | | }}}| }| | }d||f||f||ff}|D ]T\ }}|g| j z } | | | }
|
t j | d k sJ U| |
|| d }
|
|k sJ d S )N)r r Tnan_okrQ ) rk rn rq r6 r r- absrS r approxr( ) r pinfninfrp r r abs_casescaser data_absvabss r test_absz_SIMD_FP.test_abs s //++T__->-> Cdzz|| $**,,'' tTlT4L3*E & @ @MD' y,H88DKK - -..D6=$????????xx B8899t||||||r c 4 | | | }}}| }| | }ddd|f||f||ff}|D ]T\ }}|g| j z } | | | }
|
t j | d k sJ U| d |D } | | }
|
| k sJ d S )N) rN ) rO g TrB c 6 g | ]}t j | S r( )mathsqrtr) s r r, z&_SIMD_FP.test_sqrt.. s : : :ty|| : : :r )
rk rn rq r6 r r- rR rS r rE ) r rF rG rp r r
sqrt_casesrI r data_sqrtrR s r test_sqrtz_SIMD_FP.test_sqrt s //++T__->-> Cdzz|| $**,,''"Js dC[4QU,W
' A AMD' $+-IIIdkk$ / /00D6=4@@@@@@@@II : :T : : :;; yy y r c
| | | }}}| }| | }||f||f||ff}|D ]T\ }}|g| j z } | | | }
|
t j | d k sJ Ud |D } | | }
|
| k sJ d S )NTrB c g | ]}||z S r( r( r) s r r, z(_SIMD_FP.test_square.. s )))qqs)))r )
rk rn rq r6 r r- squarerS r rE ) r rF rG rp r r square_casesrI r data_squarerX s r test_squarez_SIMD_FP.test_square s //++T__->-> Cdzz|| $**,,'' c
T4L4,? ) E EMD'")DK/Kkk$++d"3"344FV];t D D DDDDDD))D))) U##$$$$$$r z intrin, funcceiltruncfloorrintc |}t | | }| | | }}}||f||f||ff}|D ]J\ }} | g| j z }
|| | } | t
j |
d k sJ Kt ddd D ]\dD ]W| fdt | j D } fd| D }
|| } | |
k sJ X]d D ]8}
| |
| }fd
D }
||
k sJ 9|d k rd }nd
}|D ]c|
|| } |
| d }
| |
k sJ ddS )z
Test intrinsics:
npyv_rint_##SFX
npyv_ceil_##SFX
npyv_trunc_##SFX
npyv_floor##SFX
TrB r i )gggffffffg?g?gffffff?c g | ]
}|z z S r( r( )r* r wr+ s r r, z*_SIMD_FP.test_rounding.. s! !F!F!Fa1Q3'!F!F!Fr c & g | ]
} | S r( r( )r* r+ r s r r, z*_SIMD_FP.test_rounding.. s! 4 4 4!dd1gg 4 4 4r )g/ Cg/ Cg ?Dg _Dc & g | ]
} | S r( r( )r* nr s r r, z*_SIMD_FP.test_rounding.. s! ---a$$q''---r r^ )rN )rN g пg333333ӿgܿr rN N) r rk rn rq r- rS r rE r. r rR )r r r r# rF rG rp round_casesrI r r r r r y
data_szerorc r+ s ` @@r
test_roundingz_SIMD_FP.test_rounding s< v&& //++T__->-> CdSzD$<$ > ( D DMD'!4;.JVDKK - -..FV]:d C C CCCCCCq%(( , ,A<
,
,yy!F!F!F!F!F53E3E!F!F!FGG 4 4 4 4t 4 4 4
+++++
,
# #A AAq A----1---J
????? ' ! ! JJ:J ( (A&&vvdkk!nn'='=>>F **4;;t+<+<==JZ''''' ( (r r )rH maxpmaxnrG minpminnc @ | | | c}}ddddd |dd d }t |dd }t | d|z }t | | }| j dz }d d
gd
d gfd d gd d gf|d gd |gfd |g|d gfd d gd d gfd d gd d gff}|D ]s\ } }
| | |z } | |
|z } || | }
|| | }||
k sJ || }
|| }||
k sJ t|sdS |dk rd
}nfd}d fd f|f|fff}|D ]\ } }
| | |
g|z } || |
}
|| }|t j |
d k sJ |
| } |
|
} |
g| j z }
|| | }|t j |
d k sJ dS )a
Test intrinsics:
npyv_max_##sfx
npyv_maxp_##sfx
npyv_maxn_##sfx
npyv_min_##sfx
npyv_minp_##sfx
npyv_minn_##sfx
npyv_reduce_max_##sfx
npyv_reduce_maxp_##sfx
npyv_reduce_maxn_##sfx
npyv_reduce_min_##sfx
npyv_reduce_minp_##sfx
npyv_reduce_minn_##sfx
r r )xpnpnnxnNr r, reduce_rO rN
ic ^ t j | r|nt j | r| n|S NrQ isnan)r r^ s r r z'_SIMD_FP.test_max_min.. s, Z]]ATZ]](A r c \ t j | st j | rn|S rx ry )r r^ rp s r r z'_SIMD_FP.test_max_min.. s) z!}}<
1
<1 r TrB ) rk rn rq getr r r- r r rE rS )r r rF rG chk_nanr
reduce_intrin hf_nlanescasesop1op2r r r r test_nanvdata_abrp s @r test_max_minz_SIMD_FP.test_max_min s & //++T__->-> dC!1A66::6"##;JJF2A2J i&&899
v&&KN 4[4+ &#Y"c #BZ"d $$Z$ $#YS "2Yb "
HCiiI
..GiiI
..G4))D6'7++D4<<<<4==D =))D4<<<<< F a<< HH H "I
I
$K
3K
#J
< .7 s 2 2 2! 2 2 2r )
rk rn rq r6 r r- reciprS r rE ) r rF rG rp r r recip_casesrI r
data_recipr s r test_reciprocalz_SIMD_FP.test_reciprocal, s! //++T__->-> Cdzz|| $**,,''SzD#;t sDkDRV<X ( C CMD'!4;.JJJt{{4 0 011EFM*T B B BBBBBBYY 2 2T 2 2 233
5!!
""""""r c | | | }|dg| j z k sJ dS )zQ
Compare Not NaN. Test intrinsics:
npyv_notnan_##SFX
r N)notnanrS rq r- )r nnans r test_special_casesz_SIMD_FP.test_special_cases; sH
{{4;;tyy{{3344s4;&&&&&&r r# )r_ r] r\ r^ c t | | }t d t d t d fD ]F}| | }t || t d dk sJ Gd S )Nrp rh T)r F)r rj rS r r )r r# r drI s r test_unary_invalid_fpexceptionz'_SIMD_FP.test_unary_invalid_fpexceptionC s {++,,e uU||m< < .to_bool..[ 999$DI%999r r( rd r s r to_boolz2_SIMD_FP.test_comparison_with_nan..to_boolZ 9999&999 9r r )rN rO c . g | ]\ }} || S r( r( )r* r r^ py_comps r r, z5_SIMD_FP.test_comparison_with_nan..f s' FFF$!Q1
FFFr )rk rn rq rY r r- rS r )r r np_comprF rG rp r r cmp_cases
case_operand1
case_operand2r r r r vcmpdata_cmpr s ` @r test_comparison_with_nanz!_SIMD_FP.test_comparison_with_nanN s? //++T__->-> CdOO%% : : : : : w''XQx#sdC[C[,0 ,5 $ $ (M=#_T[0F#_T[0Fkk-00Gkk-00G766'73344DFFFF#ff2E2EFFFH8##### $ $r r r r rp r r rO rN r' c | || j z }t | }t | | } || } || }||k sJ dS r r r- r r r s r r z!_SIMD_FP.test_operators_crosstesti b " 4$+-..F||v&&$v,,vf~~w&r N) ra r| r} r r@ rL rU r[ r r r rQ r\ r] r^ r rj r r r r operatorltlegtgeeqner rj r r( r r r* r* k s 2 2 22
!
!
!
%
%
% [ ^vty.A djGTZ065/.C D D,( ,(D D,(\ [ X ( ( ( C< C< C Z"d? Z#d@ Z$dA Z%dBS )C _SIMD_ALLz*
To test all vector types at once
c | }| | }||k sJ | | }||k sJ | | }||k sJ | | }t | d | j dz }|d | j dz }||k sJ ||k sJ d S )Nr )r6 r loadaloadsloadlr1 r- )r r load_data
loada_data
loads_datar
loadl_half data_halfs r test_memory_loadz_SIMD_ALL.test_memory_load s zz||IIdOO D ZZ%%
T!!!!ZZ%%
T!!!!
4 %[[ $+q. 1
$+q. ) Y&&&&}}}}}}r c | }| | }dg| j z }| || ||k sJ dg| j z }| || ||k sJ dg| j z }| || ||k sJ dg| j z }| || |d | j dz |d | j dz k sJ ||k sJ dg| j z }| || |d | j dz || j dz d k sJ ||k sJ d S )Nr r )r6 r r- storestoreastoresstorelstoreh)r r r r store_astore_sstore_lstore_hs r test_memory_storez_SIMD_ALL.test_memory_store su zz|| $dk!
5% }}}}# # GU###$# # GU###$# # GU### Q'4a +@@@@@% # # GU### Q'4 Q+@@@@@% r z intrin, elsizes, scale, fillz self.load_tillz, self.load_till @ r i z!self.load2_tillz, self.load2_tillr c | |vrd S t | \ }}| }t t d| j dz }|| j dz | j dz gz
}|D ]a} ||| g|R }
||| } | |z } |d | || j | z
|z z z } |
| k sJ |d | dg| j | z
z z }
| |
k sJ bd S )Nr r r
r )rD r r6 r1 r. r- )r r elsizesscalefillnpyv_load_tillznpyv_load_tillr lanesrf load_till
load_tillz data_till
data_tillzs r test_memory_partial_loadz"_SIMD_ALL.test_memory_partial_load s'
g - - F*.v,,' zz||U1dkAo..//
$+q.$+q.11 , ,A&tQ6666I( q11J
JA RaR4DKMe+C#DDI )))) bqbQC4;q=$99J +++++ , ,r zintrin, elsizes, scale)zself.store_tillr r )zself.store2_tillr r c | |vrd S t | }| }| d }| | }t t d| j dz }|| j dz | j dz gz
}|D ]V} | }
|d | |z |
d | |z < | d } || | | | |
k sJ Wd S )NTr r r r
rD r r6 r r1 r. r- copy) r r r r npyv_store_tillr data_revr r rf r
store_tills r test_memory_partial_storez#_SIMD_ALL.test_memory_partial_store s
g - - F v,,zz||::d:++ $U1dkAo..//
$+q.$+q.11 + +A
I"&x%x.Ihqwh D11J OJ5 1 1 1 ***** + +r )z
self.loadnr r )z self.loadn2r r c
| |vrd S t | }t dd D ]dk r^| | j z t t
j t fdt |dd D }ndk r*| d| | j |z z }nZ| | j z t t
j t fdt | D }| | } | }||k sJ d S )Nr r c ( g | ]}| d S rx r( r* r r strides r r, z6_SIMD_ALL.test_memory_noncont_load.. & III4
F
+IIIr rQ r3 c & g | ]
}|d S rx r( r s r r, z6_SIMD_ALL.test_memory_noncont_load.. # AAAq4 6 ?AAAr )
rD r r. r6 r- r1 itertoolschainr r ) r r r r
npyv_loadn data_strideloadnr r s @@r test_memory_noncont_loadz"_SIMD_ALL.test_memory_noncont_load s}
g - - F&\\
Cnn ( (Fzzzz&6'$+*=>> "9?IIIIIU5!R5H5HIIIJ$ 1zz|| "1U7mt{E/A B zzt{(:z;; "9?AAAAAE%LLAAAB$D ))K00K JtV,,EK'''''! ( (r z!self.loadn_tillz, self.loadn_tillz#self.loadn2_tillz, self.loadn2_tillc
| |vrd S t | \ }}t t d| j dz }|| j dz | j dz gz
}t dd D ]sdk r^| | j z t t
j t fdt |dd D }ndk r*| d| | j |z z }nZ| | j z t t
j t fd
t | D }t | | }|D ]]} | |z }
| j |
z
} |d |
|| |z z z } || g|R }
|
| k sJ |d |
dg| z z } || }||k sJ ^ud S ) Nr r r
r r r c ( g | ]}| d S rx r( r s r r, z>_SIMD_ALL.test_memory_noncont_partial_load.. r r rQ r c & g | ]
}|d S rx r( r s r r, z>_SIMD_ALL.test_memory_noncont_partial_load.. r r )
rD r r1 r. r- r6 r r r r )r r r r r npyv_loadn_tillznpyv_loadn_tillr r rf nscalellanesdata_stride_till
loadn_tilldata_stride_tillz loadn_tillzr r s @@r test_memory_noncont_partial_loadz*_SIMD_ALL.test_memory_noncont_partial_load sW
g - - F,0LL)/U1dkAo..//
$+q.$+q.11Cnn 8 8Fzzzz&6'$+*=>> "9?IIIIIU5!R5H5HIIIJ$ 1zz|| "1U7mt{E/A B zzt{(:z;; "9?AAAAAE%LLAAAB$ tyy 5 566K
8
8U v- (465=+AA ! -_T61 Dt D D D
!%55555$/$8A3<$G! . .tVQ ? ? "&777777
8 8 8r )z self.storenr r )z self.storen2r r c | |vrd S t | }| }| | }| j |z }t dd D ]}dg|z | j z } t d||z | D ] }
|
|z |z } || | |z | |
|
|z < dg|z | j z } | dgdz z
} || || | d d | k sJ | dd dgdz k sJ t dd D ]}dg| z | j z } t d||z | D ]!}
|
|z |z } || | |z | |
|z
|
pd < "dgdz } | dg| z | j z z
} || || | dd | k sJ | d d dgdz k sJ dg| j z } | } || d | d|< || d| | | k sJ d S )Nr r r r r )rD r r6 r r- r. r )
r r r r npyv_storenr r hlanesr data_storenr> r storens
r test_memory_noncont_storez#_SIMD_ALL.test_memory_noncont_store s
g - - F6ll zz|| $%Arll - -F &6/DK7K1fVmV44
9
9Y%)-a%i AagI&&Vf_t{2F tfRi F K . . .$3$<;....#$$ r storen_tills r !test_memory_noncont_partial_storez+_SIMD_ALL.test_memory_noncont_partial_store? s
g - - F <<zz|| $U1dkAo..//
$+q.$+q.11%Arll 6 6F
6
6 !FVOdk 9 XagX$4;qw3F)GG q&-88!< < . 9
!
56A
r c g | ] }|D ]}|
S r( r( r/ s r r, z*_SIMD_ALL.test_reorder.. r1 r r r r )
r6 r r- combinelcombinehcombiner r$ r1 unzip)r r r r r data_a_lo data_b_lo data_a_hi data_b_hir3 r4 r5 data_zipl data_ziphvzipr6 s r test_reorderz_SIMD_ALL.test_reorder s **,,
4
(@(@ 99V,,dii.?.??DKN?+ ?DKN?+ 4;>??+ 4;>??+ =='229y00000=='229y00000,,w009y0)i2GHHHHH II
9i00
II
9i00
xx)) 9-----s4;q $' 2333tIi 88888
9i00((((( i 4 ??!BCC(((((((r c | dk rd S fdt d| j dz D }| | t | j }||k sJ d S )Nr c
` g | ]*}t t ||d z z D ]}|+S )r )r/ r. )r* r+ rh ssizes r r, z0_SIMD_ALL.test_reorder_rev64.. sX
aRY 7 788
r r )rD r. r- rev64r )r
data_rev64rB rA s @r test_reorder_rev64z _SIMD_ALL.test_reorder_rev64 s !!## B;; F
Q RY77
499U4;%7%78899
""""""r c
| }|dk rdS | | d|z
dz
| j
z }|dk rddgng d}t
D ]F fd|D } | j g|R }
fd t
||z D }||k sJ GdS )
z\
Test permuting elements for each 128-bit lane.
npyv_permi128_##sfx
r N r r r )r r r
c g | ]
}|z z S r( r( )r* shfr permds r r, z3_SIMD_ALL.test_reorder_permi128.. s" :::cSE):::r c 4 g | ]\ }}|| z z S r( r( )r* ejr permns r r, z3_SIMD_ALL.test_reorder_permi128.. s= Aq Q!uf*%& r )rD r r6 r- r. permi128r ) r rA nlane128shflindicesvperm
data_vpermr r rJ rN s @@@@r test_reorder_permi128z _SIMD_ALL.test_reorder_permi128 s
!!## 2:: Fyy &&U
a;% "1vv,,,u ' 'A:::::T:::G!DM$1111E %gh&6 7 7 J J&&&&& ' 'r z func, intrinr r r r r c $
| r| }n/| | | j z
}| | d }| | | | }}t
| | }|
fd}fdt || D } | ||| } | |k sJ d S )NTr c fd| D S )Nc g | ]}|k S r( r( r s r r, zH_SIMD_ALL.test_operators_comparison..to_bool..
r r r( r s r r z4_SIMD_ALL.test_operators_comparison..to_bool r r c . g | ]\ }} || S r( r( )r* r r^ r s r r, z7_SIMD_ALL.test_operators_comparison.. s' ???41aDDAJJ???r ) r0 r6 rL r- rM r r rY r ) r r r r r r r r r cmpr s ` @r test_operators_comparisonz#_SIMD_ALL.test_operators_comparison s ;;== ?ZZ\\FFZZ
$+ =>>FDMMOOT:: 99V,,dii.?.?v&&OO%% : : : : : @???3vv+>+>???gffWg..//hr c | r| }n/| | | j z
}| | d }| | | | }}| r9| | }| | }| j | j }}n||}}d | j }} |d t || D } || || }
|
| k sJ |d t || D } | t | d || } | | k sJ |d t || D }
| t | d || }||
k sJ |d |D } | t | d
| }||k sJ | j
d vrd S d t || D } | t | d
|| }||k sJ d S )NTr c | S rx r( )r s r r z2_SIMD_ALL.test_operators_logical.. s r c g | ]
\ }}||z S r( r( r s r r, z4_SIMD_ALL.test_operators_logical.. N N N1a!e N N Nr c g | ]
\ }}||z S r( r( r s r r, z4_SIMD_ALL.test_operators_logical..$ r_ r r c g | ]
\ }}||z S r( r( r s r r, z4_SIMD_ALL.test_operators_logical..( r_ r r c g | ]}| S r( r( )r* r s r r, z4_SIMD_ALL.test_operators_logical.., s 6 6 6Qqb 6 6 6r r r c g | ] \ }}|| z S r( r( r s r r, z4_SIMD_ALL.test_operators_logical..2 s" FFF1Q!VFFFr r ) r0 r6 rL r- rM r rR r r r r )r r r r r data_cast_a data_cast_bcast cast_datar r r r r r data_notr r r s r r z _SIMD_ALL.test_operators_logical s ;;== ?ZZ\\FFZZ
$+ =>>FDMMOOT:: 99V,,dii.?.? ;;== 5 ++G44K ++G44K " /1B)DD'-vK )k49)D 9 N NK0M0M N N NOOtDHHWg..//x 9 N NK0M0M N N NOOt'GD$''99::g~~~~ 9 N NK0M0M N N NOOt(GD%(('::;;x 9 6 6+ 6 6 677t(GD%((1122x 8D ! ! FFFK(E(EFFF *WT6**7G<<== !!!!!!r r r r r )r r r, r
)rQ rt )r r r r, r
)r r il )r rQ rt ri r
r rQ c | || j z }t | }t | | } || } || }||k sJ dS r r r s r r z"_SIMD_ALL.test_operators_crosstest6 r r c d| j dd z }t | j d|d| j }t | j d| j d| } || d }| | d | d }||k sJ || } || }||k sJ d S )Nr^ r cvt_r r )r r r rS r )r bsfx
to_boolean from_booleanfalse_vbtrue_vb
false_vsfx true_vsfxs r test_conversion_booleanz!_SIMD_ALL.test_conversion_booleanN s TXabb\! TYYtttTXX(FGG
tyy44*HII :dkk!nn--::dkk!nndkk!nn==7""""!\(++
L)) Y&&&&&&r c | j dvrdS | j d t t | j dd dz z }t | j d| d| j }| | | j z
}| | } || }|d| j dz }|| j dz d }|||fk sJ dS )zh
Test expand intrinsics:
npyv_expand_u16_u8
npyv_expand_u32_u16
r r Nr r r expand_r ) r strrC r r r6 rL r- r )r totypeexpandr r edatadata_lodata_his r test_conversion_expandz _SIMD_ALL.test_conversion_expand[ s 8= ( ( F!STXabb\!2!21!4 5 55$Af$A$Atx$A$ABB
4==??T[899 $u
Q't{A~''*******r c | r| }n/| | | j z
}| | d }| | | | }}| d t
|| D }| || }||k sJ | d t
|| D }| || }||k sJ d S )NTr c g | ]
\ }}||z S r( r( r s r r, z4_SIMD_ALL.test_arithmetic_subadd..x D D D1a!e D D Dr c g | ]
\ }}||z
S r( r( r s r r, z4_SIMD_ALL.test_arithmetic_subadd..{ s E E EAq1u E E Er ) r0 r6 rL r- rM r r r/ r ) r r r r r data_addr/ data_subr s r test_arithmetic_subaddz _SIMD_ALL.test_arithmetic_subaddo s ;;== ?ZZ\\FFZZ
$+ =>>FDMMOOT:: 99V,,dii.?.? 99 D DFF0C0C D D DEExx))hII E EVV1D1D E E EFFxx))hr c | j dv rd S | r| }n/| | | j z
}| | d }| | | | }}| d t || D }| || }||k sJ d S )N)r r Tr c g | ]
\ }}||z S r( r( r s r r, z1_SIMD_ALL.test_arithmetic_mul.. r r ) r r0 r6 rL r- rM r r r( )r r r r r data_mulr( s r test_arithmetic_mulz _SIMD_ALL.test_arithmetic_mul s 8~ % % F ;;== ?ZZ\\FFZZ
$+ =>>FDMMOOT:: 99V,,dii.?.?99 D DFF0C0C D D DEEhhw((hr c p | sd S | | d }}| | | | }}| d t || D }| || }||k sJ d S )NTr c g | ]
\ }}||z S r( r( r s r r, z1_SIMD_ALL.test_arithmetic_div.. r r )r0 r6 r r div)r r r r r data_divr s r test_arithmetic_divz _SIMD_ALL.test_arithmetic_div s {{}} F tzz$z'?'? 99V,,dii.?.? 99 D DFF0C0C D D DEEhhw((hr c l | rdS | fd d g}|t ddd z
}|t ddd z
}| }|dk r(|t dd d
z
}|t dd d z
}|d k r(|t d d
d z
}|t d d
d z
}|dk r(|t d
dd z
}|t d
dd z
}|d |D z
}t j || D ]\ }| d dk r'| | | } fd|D }| }|
|| }||k sJ dS )zn
Test integer division intrinsics:
npyv_divisor_##sfx
npyv_divc_##sfx
Nc v |dk r| k r| S | dk |dk }}| dk s||k r| |z S | |z |z
|z dz S )z
Divide towards zero works with large integers > 2^53,
and wrap around overflow similar to what C does.
rQ r r r( )r r sign_asign_dint_mins r trunc_divz3_SIMD_ALL.test_arithmetic_intdiv..trunc_div sa
Bww1<< UAEFFAvv6))Av
J'A-1 1r r r r r ra i i r l i i l l l c g | ]}| S r( r( r) s r r, z4_SIMD_ALL.test_arithmetic_intdiv.. s """!"""r c ( g | ]} | S r( r( )r* r divisorr s r r, z4_SIMD_ALL.test_arithmetic_intdiv.. s% AAA11g..AAAr ) r0 rM r. rD r productrS r r6 r divc)
r r bsizedividend data_divc
divisor_parmsr r r r s
@@@r test_arithmetic_intdivz _SIMD_ALL.test_arithmetic_intdiv s ;;== F--//
2
2
2
2
2 G8} at$$$ av&&&!!## 199 E$u-- -D E$w// /D 2:: E%.. .D E%00 0D 2:: E%.. .D E%00 0D ""T""""!*!24!>!> % % Hgkk'**1-G!||yyH!5!566HAAAAAAAAI LL 1 1M99X}55D9$$$$$ % %r c | j dvrdS | }| | }t | }| | }||k sJ dS )zH
Test reduce sum intrinsics:
npyv_sum_##sfx
)r r r r N)r r6 r r )r r r data_sumvsums r test_arithmetic_reduce_sumz$_SIMD_ALL.test_arithmetic_reduce_sum sd
87 7 7 Fzz|| $t99xxx r c @ | j dvrdS d| j | | | j z
f}|D ]X}| | }| | }t
| }| | }||k sJ YdS )zQ
Test extend reduce sum intrinsics:
npyv_sumup_##sfx
rw Nr )r r- rM rL r6 r r sumup)r r rr r r r s r test_arithmetic_reduce_sumupz&_SIMD_ALL.test_arithmetic_reduce_sumup s
8= ( ( FDK $--//$+2MN $ $A::a==DIIdOOE4yyH::e$$D8##### $ $r c ^ | | }| | d }| | | }| | | }| || }| |||| }||k sJ | |||| }||k sJ | || }| |||| }||k sJ | |||| }||k sJ | sdS |
|| } | |||| }
|
| k sJ | ||| } | | k sJ | |||| }
|
|k sJ | ||| } | | k sJ dS )z
Conditional addition and subtraction for all supported data types.
Test intrinsics:
npyv_ifadd_##SFX, npyv_ifsub_##SFX
Tr N)
r r6 r r r r ifsubr/ ifaddr0 r ifdivifdivz) r r r r
false_maskr r r r r r r s r test_mask_conditionalz _SIMD_ALL.test_mask_conditional s ))DJJLL))))DJJtJ 4 455ZZ TYY[[99 [[diikk::
88GW--
9gw@@
:wAA 88GW--
9gw@@
:wAA {{}} F88GW--
9gw@@ Y99!!!!
:wAA Z'::$$$$$$r N)&ra r| r} r r r r r r r r r r r r
r r, r> rD rU r r r r r r r[ r r ru r r r r r r r r r( r r r r s $ 4 [ ; *Ha&B ,hFF;KL> , , , [ 5()8 + + + [ 5#$8 ( ( (, [ ; ,hF8D .!ff=MN> 8 8 8B [ 5$%8 !% !% !%F [ 5)*8 ), ), ),V [ 9 < ! ! !"+ + + Z$) $) $)L
#
#
#' ' ', [ ^ g g g g g . "$" $" $"L [ Xu~66 [ V 666
& ' ' 76 ' ' ' '+ + +(
-% -% -%^
$
$
$ !% !% !% !% !%r r )r r r r r )r r )r )r r] rt r (ru )ztarget 'z$' isn't supported by current machinez' isn't supported by NPYVz"' doesn't support single-precisionr z!' doesn'tsupport double-precisionr )r r rv Testr )reason)FFFFF)2r rQ r r r numpy.core._simdr r r numpy.core._multiarray_umathr r r r r r r% r* r bool_sfxint_sfxfp_sfxall_sfxtests_registryitemsrv r r
simd_widthrw pretty_namelenrx skipdictskip_sfxr r r( clsr r| skip_minhrr typera tclsr globalsr( r r r s# H H H H H H H H H H 9 9 9 9 9 905/4 a a a a a a a a Fy' y' y' y' y' y' y' y'vK8 K8 K8 K8 K8
K8 K8 K8\$ $ $ $ $ $ $ $.$ $ $ $ $ $ $ $"T' T' T' T' T'} T' T' T'lG
% G
% G
% G
% G
%
G
% G
% G
%R '
@
F
j i
h j j i
' , ,K "*J ##D))K
s;!2#((;//222 !!n
DtvvH
:K+KKK
Y :@+@@@} B B B B BHUO} : : : : :HUO %**,, , ,
s , ,C \\#t,,F6D4W[ 1s TTTD4Ns|NNjNN;NNNNPTVZ[[D
6/ //555'+GGIIdm $ $ ,,/ , ,r