Add a couple unitary ops.

This commit is contained in:
laurent
2023-06-23 20:19:20 +01:00
parent fe75a01188
commit 8ed350dc94
3 changed files with 127 additions and 0 deletions

View File

@ -37,6 +37,11 @@ __device__ T gelu_fwd(T x) {
#if __CUDA_ARCH__ >= 530
UNARY_OP(__half, ucopy_f16, x)
UNARY_OP(__half, uneg_f16, -x)
UNARY_OP(__half, uexp_f16, expg(x))
UNARY_OP(__half, ulog_f16, logg(x))
UNARY_OP(__half, usin_f16, sing(x))
UNARY_OP(__half, ucos_f16, cosg(x))
UNARY_OP(__half, uabs_f16, absg(x))
UNARY_OP(__half, usqr_f16, x*x)
UNARY_OP(__half, usqrt_f16, sqrtg(x))
UNARY_OP(__half, gelu_f16, gelu_fwd(x))
@ -46,6 +51,16 @@ UNARY_OP(float, ucopy_f32, x)
UNARY_OP(double, ucopy_f64, x)
UNARY_OP(float, uneg_f32, -x)
UNARY_OP(double, uneg_f64, -x)
UNARY_OP(float, uexp_f32, expg(x))
UNARY_OP(double, uexp_f64, expg(x))
UNARY_OP(float, ulog_f32, logg(x))
UNARY_OP(double, ulog_f64, logg(x))
UNARY_OP(float, usin_f32, sing(x))
UNARY_OP(double, usin_f64, sing(x))
UNARY_OP(float, ucos_f32, cosg(x))
UNARY_OP(double, ucos_f64, cosg(x))
UNARY_OP(float, uabsg_f32, absg(x))
UNARY_OP(double, uabsg_f64, absg(x))
UNARY_OP(float, usqr_f32, x*x)
UNARY_OP(double, usqr_f64, x*x)
UNARY_OP(float, usqrt_f32, sqrtg(x))