Mirza Halilcevic
d425db338b
Change-Id: Idf52308186a73ae1c4d815eb59f20ea7da99c964
[ROCm/hip-tests commit: dc78fafa49]
739 строки
9.1 KiB
Plaintext
739 строки
9.1 KiB
Plaintext
Atomics [
|
|
atomicCAS
|
|
atomicCAS_system
|
|
atomicAdd
|
|
atomicAdd_system
|
|
unsafeAtomicAdd
|
|
safeAtomicAdd
|
|
atomicSub
|
|
atomicSub_system
|
|
atomicExch
|
|
atomicExch_system
|
|
atomicMin
|
|
atomicMin_system
|
|
unsafeAtomicMin
|
|
safeAtomicMin
|
|
atomicMax
|
|
atomicMax_system
|
|
unsafeAtomicMax
|
|
safeAtomicMax
|
|
atomicInc
|
|
atomicDec
|
|
atomicAnd
|
|
atomicAnd_system
|
|
atomicOr
|
|
atomicOr_system
|
|
atomicXor
|
|
atomicXor_system
|
|
]
|
|
|
|
Builtin atomics [
|
|
__hip_atomic_load
|
|
__hip_atomic_store
|
|
__hip_atomic_compare_exchange_weak
|
|
__hip_atomic_compare_exchange_strong
|
|
__hip_atomic_exchange
|
|
__hip_atomic_fetch_add
|
|
__hip_atomic_fetch_and
|
|
__hip_atomic_fetch_or
|
|
__hip_atomic_fetch_xor
|
|
__hip_atomic_fetch_min
|
|
__hip_atomic_fetch_max
|
|
]
|
|
|
|
Cooperative groups [
|
|
File restriction: cooperative_groups
|
|
Device groups: (
|
|
coalesced_group_tile
|
|
coalesced_group
|
|
grid_group
|
|
multi_grid_group
|
|
thread_block_tile
|
|
thread_block
|
|
)
|
|
size
|
|
thread_rank
|
|
group_index
|
|
thread_index
|
|
shfl
|
|
shfl_up
|
|
shfl_down
|
|
shfl_xor
|
|
sync
|
|
]
|
|
|
|
Warp [
|
|
File restriction: cooperative_groups
|
|
Device groups: (
|
|
ballot
|
|
any
|
|
all
|
|
shfl_down
|
|
shfl_up
|
|
shfl
|
|
shfl_xor
|
|
)
|
|
__ballot
|
|
__any
|
|
__all
|
|
__shfl_down
|
|
__shfl_up
|
|
__shfl
|
|
__shfl_xor
|
|
]
|
|
|
|
Launch bounds [
|
|
Device groups: (
|
|
launch_bounds
|
|
)
|
|
__launch_bounds__
|
|
]
|
|
|
|
Channel descriptor [
|
|
hipCreateChannelDesc
|
|
hipCreateChannelDescHalf
|
|
hipCreateChannelDescHalf1
|
|
hipCreateChannelDescHalf2
|
|
]
|
|
|
|
Device assert [
|
|
assert
|
|
static_assert
|
|
]
|
|
|
|
Device clock [
|
|
Device groups: (
|
|
clock
|
|
)
|
|
clock
|
|
clock64
|
|
wall_clock64
|
|
]
|
|
|
|
Device printf [
|
|
printf
|
|
]
|
|
|
|
HIP specific [
|
|
Device groups: (
|
|
hip_hc_8pk
|
|
)
|
|
__hip_hc_add8pk
|
|
__hip_hc_sub8pk
|
|
__hip_hc_mul8pk
|
|
]
|
|
|
|
Device synchronization [
|
|
__syncthreads
|
|
__syncthreads_and
|
|
__syncthreads_count
|
|
__syncthreads_or
|
|
]
|
|
|
|
Memory fence [
|
|
__threadfence
|
|
__threadfence_block
|
|
__threadfence_system
|
|
]
|
|
|
|
Device memory [
|
|
memcpy
|
|
memset
|
|
]
|
|
|
|
Device math [
|
|
rsqrt
|
|
sqrt
|
|
cbrt
|
|
rcbrt
|
|
hypot
|
|
rhypot
|
|
norm3d
|
|
rnorm3d
|
|
norm4d
|
|
rnorm4d
|
|
norm
|
|
rnorm
|
|
rsqrtf
|
|
sqrtf
|
|
cbrtf
|
|
rcbrtf
|
|
hypotf
|
|
rhypotf
|
|
norm3df
|
|
rnorm3df
|
|
norm4df
|
|
rnorm4df
|
|
normf
|
|
rnormf
|
|
rsqrt
|
|
sqrt
|
|
cbrt
|
|
rcbrt
|
|
hypot
|
|
rhypot
|
|
norm3d
|
|
rnorm3d
|
|
norm4d
|
|
rnorm4d
|
|
norm
|
|
rnorm
|
|
expf
|
|
exp2f
|
|
exp10f
|
|
expm1f
|
|
frexpf
|
|
ldexpf
|
|
powf
|
|
powif
|
|
scalbnf
|
|
scalblnf
|
|
log
|
|
log2
|
|
log10
|
|
log1p
|
|
logb
|
|
ilogb
|
|
logf
|
|
log2f
|
|
log10f
|
|
log1pf
|
|
logbf
|
|
ilogbf
|
|
sin
|
|
cos
|
|
tan
|
|
sincos
|
|
sinpi
|
|
cospi
|
|
sincospi
|
|
asin
|
|
acos
|
|
atan
|
|
atan2
|
|
sinh
|
|
cosh
|
|
tanh
|
|
asinh
|
|
acosh
|
|
atanh
|
|
sinf
|
|
cosf
|
|
tanf
|
|
sincosf
|
|
sinpif
|
|
cospif
|
|
sincospif
|
|
asinf
|
|
acosf
|
|
atanf
|
|
atan2f
|
|
sinhf
|
|
coshf
|
|
tanhf
|
|
asinhf
|
|
acoshf
|
|
atanhf
|
|
erf
|
|
erfc
|
|
erfinv
|
|
erfcinv
|
|
erfcx
|
|
normcdf
|
|
normcdfinv
|
|
lgamma
|
|
tgamma
|
|
j0
|
|
j1
|
|
jn
|
|
y0
|
|
y1
|
|
yn
|
|
cyl_bessel_i0
|
|
cyl_bessel_i1
|
|
erff
|
|
erfcf
|
|
erfinvf
|
|
erfcinvf
|
|
erfcxf
|
|
normcdff
|
|
normcdfinvf
|
|
lgammaf
|
|
tgammaf
|
|
j0f
|
|
j1f
|
|
jnf
|
|
y0f
|
|
y1f
|
|
ynf
|
|
cyl_bessel_i0f
|
|
cyl_bessel_i1f
|
|
fmod
|
|
remainder
|
|
remquo
|
|
modf
|
|
fdim
|
|
trunc
|
|
round
|
|
rint
|
|
nearbyint
|
|
ceil
|
|
floor
|
|
lrint
|
|
lround
|
|
llrint
|
|
llround
|
|
fmodf
|
|
remainderf
|
|
remquof
|
|
modff
|
|
fdimf
|
|
truncf
|
|
roundf
|
|
rintf
|
|
nearbyintf
|
|
ceilf
|
|
floorf
|
|
lrintf
|
|
lroundf
|
|
llrintf
|
|
llroundf
|
|
fma
|
|
abs
|
|
fabs
|
|
copysign
|
|
fmax
|
|
fmin
|
|
nan
|
|
nextafter
|
|
signbit
|
|
fmaf
|
|
abs
|
|
fabsf
|
|
copysignf
|
|
fdividef
|
|
fmaxf
|
|
fminf
|
|
isfinite
|
|
isinf
|
|
isnan
|
|
nanf
|
|
nextafterf
|
|
signbit
|
|
__brev
|
|
__brevll
|
|
__clz
|
|
__clzll
|
|
__ffs
|
|
__ffsll
|
|
__popc
|
|
__popcll
|
|
__mul24
|
|
__umul24
|
|
__dadd_rn
|
|
__dsub_rn
|
|
__dmul_rn
|
|
__fma_rn
|
|
__drcp_rn
|
|
__dsqrt_rn
|
|
__ddiv_rn
|
|
__fadd_rn
|
|
__fsub_rn
|
|
__fmul_rn
|
|
__fmaf_rn
|
|
__frcp_rn
|
|
__fsqrt_rn
|
|
__frsqrt_rn
|
|
__fdiv_rn
|
|
__fdividef
|
|
__expf
|
|
__exp10f
|
|
__logf
|
|
__log2f
|
|
__log10f
|
|
__sinf
|
|
__cosf
|
|
__sincosf
|
|
__tanf
|
|
__powf
|
|
]
|
|
|
|
Device conversion functions [
|
|
__double2float_rd
|
|
__double2float_rn
|
|
__double2float_ru
|
|
__double2float_rz
|
|
__double2hiint
|
|
__double2loint
|
|
__double2int_rd
|
|
__double2int_rn
|
|
__double2int_ru
|
|
__double2int_rz
|
|
__double2ll_rd
|
|
__double2ll_rn
|
|
__double2ll_ru
|
|
__double2ll_rz
|
|
__double2uint_rd
|
|
__double2uint_rn
|
|
__double2uint_ru
|
|
__double2uint_rz
|
|
__double2ull_rd
|
|
__double2ull_rn
|
|
__double2ull_ru
|
|
__double2ull_rz
|
|
__double_as_longlong
|
|
__float2int_rd
|
|
__float2int_rn
|
|
__float2int_ru
|
|
__float2int_rz
|
|
__float2ll_rd
|
|
__float2ll_rn
|
|
__float2ll_ru
|
|
__float2ll_rz
|
|
__float2uint_rd
|
|
__float2uint_rn
|
|
__float2uint_ru
|
|
__float2uint_rz
|
|
__float2ull_rd
|
|
__float2ull_rn
|
|
__float2ull_ru
|
|
__float2ull_rz
|
|
__float_as_int
|
|
__float_as_uint
|
|
__hiloint2double
|
|
__int2double_rn
|
|
__int2float_rd
|
|
__int2float_rn
|
|
__int2float_ru
|
|
__int2float_rz
|
|
__int_as_float
|
|
__ll2double_rd
|
|
__ll2double_rn
|
|
__ll2double_ru
|
|
__ll2double_rz
|
|
__ll2float_rd
|
|
__ll2float_rn
|
|
__ll2float_ru
|
|
__ll2float_rz
|
|
__longlong_as_double
|
|
__uint2double_rn
|
|
__uint2float_rd
|
|
__uint2float_rn
|
|
__uint2float_ru
|
|
__uint2float_rz
|
|
__uint_as_float
|
|
__ull2double_rd
|
|
__ull2double_rn
|
|
__ull2double_ru
|
|
__ull2double_rz
|
|
__ull2float_rd
|
|
__ull2float_rn
|
|
__ull2float_ru
|
|
__ull2float_rz
|
|
]
|
|
|
|
Device complex type functions [
|
|
hipCrealf
|
|
hipCimagf
|
|
make_hipFloatComplex
|
|
hipConjf
|
|
hipCsqabsf
|
|
hipCaddf
|
|
hipCsubf
|
|
hipCmulf
|
|
hipCdivf
|
|
hipCabsf
|
|
hipCreal
|
|
hipCimag
|
|
make_hipDoubleComplex
|
|
hipConj
|
|
hipCsqabs
|
|
hipCadd
|
|
hipCsub
|
|
hipCmul
|
|
hipCdiv
|
|
hipCabs
|
|
make_hipComplex
|
|
hipComplexDoubleToFloat
|
|
hipComplexFloatToDouble
|
|
hipCfmaf
|
|
hipCfma
|
|
]
|
|
|
|
Device vector types [
|
|
make_uchar1
|
|
make_uchar2
|
|
make_uchar3
|
|
make_uchar4
|
|
make_char1
|
|
make_char2
|
|
make_char3
|
|
make_char4
|
|
make_ushort1
|
|
make_ushort2
|
|
make_ushort3
|
|
make_ushort4
|
|
make_short1
|
|
make_short2
|
|
make_short3
|
|
make_short4
|
|
make_uint1
|
|
make_uint2
|
|
make_uint3
|
|
make_uint4
|
|
make_int1
|
|
make_int2
|
|
make_int3
|
|
make_int4
|
|
make_float1
|
|
make_float2
|
|
make_float3
|
|
make_float4
|
|
make_double1
|
|
make_double2
|
|
make_double3
|
|
make_double4
|
|
make_ulong1
|
|
make_ulong2
|
|
make_ulong3
|
|
make_ulong4
|
|
make_long1
|
|
make_long2
|
|
make_long3
|
|
make_long4
|
|
make_ulonglong1
|
|
make_ulonglong2
|
|
make_ulonglong3
|
|
make_ulonglong4
|
|
make_longlong1
|
|
make_longlong2
|
|
make_longlong3
|
|
make_longlong4
|
|
]
|
|
|
|
Device surface functions [
|
|
__hipGetPixelAddr
|
|
__hipMapToNativeFloat4
|
|
__hipMapFromNativeFloat4
|
|
surf1Dread
|
|
surf1Dwrite
|
|
surf2Dread
|
|
surf2Dwrite
|
|
surf3Dread
|
|
surf3Dwrite
|
|
surf1DLayeredread
|
|
surf1DLayeredwrite
|
|
surf2DLayeredread
|
|
surf2DLayeredwrite
|
|
surfCubemapread
|
|
surfCubemapwrite
|
|
surfCubemapLayeredread
|
|
surfCubemapLayeredwrite
|
|
]
|
|
|
|
Device texture functions [
|
|
tex1Dfetch
|
|
tex1D
|
|
tex2D
|
|
tex1DLayered
|
|
tex2DLayered
|
|
tex3D
|
|
texCubemap
|
|
tex1DLod
|
|
tex2DLod
|
|
tex1DLayeredLod
|
|
tex2DLayeredLod
|
|
tex3DLod
|
|
texCubemapLod
|
|
texCubemapLayered
|
|
texCubemapLayeredLod
|
|
texCubemapGrad
|
|
texCubemapLayeredGrad
|
|
tex1DGrad
|
|
tex2DGrad
|
|
tex1DLayeredGrad
|
|
tex2DLayeredGrad
|
|
tex3DGrad
|
|
tex2Dgather
|
|
]
|
|
|
|
Device float16 functions [
|
|
make_half2
|
|
__low2half
|
|
__high2half
|
|
__half2half2
|
|
__halves2half2
|
|
__low2half2
|
|
__high2half2
|
|
__lows2half2
|
|
__highs2half2
|
|
__lowhigh2highlow
|
|
__half_as_short
|
|
__half_as_ushort
|
|
__short_as_half
|
|
__ushort_as_half
|
|
__float2half
|
|
__float2half_rn
|
|
__float2half_rz
|
|
__float2half_rd
|
|
__float2half_ru
|
|
__float2half2_rn
|
|
__floats2half2_rn
|
|
__float22half2_rn
|
|
__half2float
|
|
__low2float
|
|
__high2float
|
|
__half22float2
|
|
__half2int_rn
|
|
__half2int_rz
|
|
__half2int_rd
|
|
__half2int_ru
|
|
__int2half_rn
|
|
__int2half_rz
|
|
__int2half_rd
|
|
__int2half_ru
|
|
__half2short_rn
|
|
__half2short_rz
|
|
__half2short_rd
|
|
__half2short_ru
|
|
__short2half_rn
|
|
__short2half_rz
|
|
__short2half_rd
|
|
__short2half_ru
|
|
__half2ll_rn
|
|
__half2ll_rz
|
|
__half2ll_rd
|
|
__half2ll_ru
|
|
__ll2half_rn
|
|
__ll2half_rz
|
|
__ll2half_rd
|
|
__ll2half_ru
|
|
__half2uint_rn
|
|
__half2uint_rz
|
|
__half2uint_rd
|
|
__half2uint_ru
|
|
__uint2half_rn
|
|
__uint2half_rz
|
|
__uint2half_rd
|
|
__uint2half_ru
|
|
__half2ushort_rn
|
|
__half2ushort_rz
|
|
__half2ushort_rd
|
|
__half2ushort_ru
|
|
__ushort2half_rn
|
|
__ushort2half_rz
|
|
__ushort2half_rd
|
|
__ushort2half_ru
|
|
__half2ull_rn
|
|
__half2ull_rz
|
|
__half2ull_rd
|
|
__half2ull_ru
|
|
__ull2half_rn
|
|
__ull2half_rz
|
|
__ull2half_rd
|
|
__ull2half_ru
|
|
__ldg
|
|
__ldcg
|
|
__ldca
|
|
__ldcs
|
|
__heq
|
|
__hne
|
|
__hle
|
|
__hge
|
|
__hlt
|
|
__hgt
|
|
__hequ
|
|
__hneu
|
|
__hleu
|
|
__hgeu
|
|
__hltu
|
|
__hgtu
|
|
__heq2
|
|
__hne2
|
|
__hle2
|
|
__hge2
|
|
__hlt2
|
|
__hgt2
|
|
__hequ2
|
|
__hneu2
|
|
__hleu2
|
|
__hgeu2
|
|
__hltu2
|
|
__hgtu2
|
|
__hbeq2
|
|
__hbne2
|
|
__hble2
|
|
__hbge2
|
|
__hblt2
|
|
__hbgt2
|
|
__hbequ2
|
|
__hbneu2
|
|
__hbleu2
|
|
__hbgeu2
|
|
__hbltu2
|
|
__hbgtu2
|
|
__clamp_01
|
|
__hadd
|
|
__habs
|
|
__hsub
|
|
__hmul
|
|
__hadd_sat
|
|
__hsub_sat
|
|
__hmul_sat
|
|
__hfma
|
|
__hfma_sat
|
|
__hdiv
|
|
__hadd2
|
|
__habs2
|
|
__hsub2
|
|
__hmul2
|
|
__hadd2_sat
|
|
__hsub2_sat
|
|
__hmul2_sat
|
|
__hfma2
|
|
__hfma2_sat
|
|
__h2div
|
|
amd_mixed_dot
|
|
htrunc
|
|
hceil
|
|
hfloor
|
|
hrint
|
|
hsin
|
|
hcos
|
|
hexp
|
|
hexp2
|
|
hexp10
|
|
hlog2
|
|
hlog
|
|
hlog10
|
|
hrcp
|
|
hrsqrt
|
|
hsqrt
|
|
__hisinf
|
|
__hisnan
|
|
__hneg
|
|
h2trunc
|
|
h2ceil
|
|
h2floor
|
|
h2rint
|
|
h2sin
|
|
h2cos
|
|
h2exp
|
|
h2exp2
|
|
h2exp10
|
|
h2log2
|
|
h2log
|
|
h2log10
|
|
h2rcp
|
|
h2rsqrt
|
|
h2sqrt
|
|
__hisinf2
|
|
__hisnan2
|
|
__hneg2
|
|
]
|
|
|
|
OpenGL Interop [
|
|
hipGLGetDevices
|
|
hipGraphicsGLRegisterBuffer
|
|
hipGraphicsGLRegisterImage
|
|
] |