SIMT接口精度标准汇总

本节按照数据类型和接口类型对SIMT浮点计算接口进行精度标准的分类汇总。对于非软仿接口,给出对应硬件指令的精度标准;对于软仿接口,由于软仿程度较低且使用到的硬件指令较少,给出其所用硬件指令的综合性精度标准。

bf16数据类型计算API精度标准

数学库函数

接口 硬件指令精度标准 软仿精度标准
hexp - 1-ulp,Not Support denormalized numbers
htanh - 1-ulp,Not Support denormalized numbers
hexp2 - 1-ulp,Not Support denormalized numbers
hexp10 - 1-ulp,Not Support denormalized numbers
hlog - 1-ulp,Not Support denormalized numbers
hlog2 - 1-ulp,Not Support denormalized numbers
hlog10 - 1-ulp,Not Support denormalized numbers
hsqrt - 1-ulp,Not Support denormalized numbers
hrsqrt - 1-ulp,Not Support denormalized numbers
h2tanh - 1-ulp,Not Support denormalized numbers
h2exp - 1-ulp,Not Support denormalized numbers
h2exp2 - 1-ulp,Not Support denormalized numbers
h2exp10 - 1-ulp,Not Support denormalized numbers
h2log - 1-ulp,Not Support denormalized numbers
h2log2 - 1-ulp,Not Support denormalized numbers
h2log10 - 1-ulp,Not Support denormalized numbers
h2sqrt - 1-ulp,Not Support denormalized numbers
h2rsqrt - 1-ulp,Not Support denormalized numbers

算术函数

接口 硬件指令精度标准 软仿精度标准
__habs - 0-ulp, Support denormalized numbers
__hadd 0-ulp, Support denormalized numbers -
__hmul 0-ulp, Support denormalized numbers -
__hdiv 1-ulp,Not Support denormalized numbers -
__haddx2 0-ulp, Support denormalized numbers -
__hmulx2 0-ulp, Support denormalized numbers -
__hdivx2 1-ulp,Not Support denormalized numbers -

fp16数据类型计算API精度标准

数学库函数

接口 硬件指令精度标准 软仿精度标准
hexp 1-ulp -
hexp2 - 1-ulp,Not Support denormalized numbers
hexp10 - 1-ulp,Not Support denormalized numbers
hlog 1-ulp,Not Support denormalized numbers -
hlog2 - 1-ulp,Not Support denormalized numbers
hlog10 - 1-ulp,Not Support denormalized numbers
hsqrt 1-ulp,Not Support denormalized numbers -
hrsqrt - 1-ulp,Not Support denormalized numbers
h2tanh - 1-ulp,Not Support denormalized numbers
h2exp - 1-ulp,Not Support denormalized numbers
h2exp2 - 1-ulp,Not Support denormalized numbers
h2exp10 - 1-ulp,Not Support denormalized numbers
h2log - 1-ulp,Not Support denormalized numbers
h2log2 - 1-ulp,Not Support denormalized numbers
h2log10 - 1-ulp,Not Support denormalized numbers
h2sqrt 1-ulp,Not Support denormalized numbers -
h2rsqrt - 1-ulp,Not Support denormalized numbers

算术函数

接口 硬件指令精度标准 软仿精度标准
__hadd 0-ulp, Support denormalized numbers -
__hmul 0-ulp, Support denormalized numbers -
__hdiv 1-ulp,Not Support denormalized numbers -
__haddx2 0-ulp, Support denormalized numbers -
__hmulx2 0-ulp, Support denormalized numbers -
__hdivx2 1-ulp,Not Support denormalized numbers -

fp32数据类型计算API精度标准

接口 硬件指令精度标准 软仿精度标准
expf 1-ulp,Not Support denormalized numbers -
tanhf - 1-ulp,Not Support denormalized numbers
tanpif - 1-ulp,Not Support denormalized numbers
atanf - 1-ulp,Not Support denormalized numbers
atan2f - 1-ulp,Not Support denormalized numbers
atanhf - 1-ulp,Not Support denormalized numbers
expm1f - 1-ulp,Not Support denormalized numbers
logf - 1-ulp,Not Support denormalized numbers
log2f - 1-ulp,Not Support denormalized numbers
log10f - 1-ulp,Not Support denormalized numbers
log1pf - 1-ulp,Not Support denormalized numbers
cosf - 1-ulp,Not Support denormalized numbers
acoshf - 1-ulp,Not Support denormalized numbers
sqrtf 1-ulp,Not Support denormalized numbers -
rsqrtf - 1-ulp,Not Support denormalized numbers
cbrtf - 1-ulp,Not Support denormalized numbers
rcbrtf - 1-ulp,Not Support denormalized numbers
remquof - 1-ulp,Not Support denormalized numbers
fdividef 1-ulp,Not Support denormalized numbers -