diff options
Diffstat (limited to 'candle-kernels/src/unary.cu')
| -rw-r--r-- | candle-kernels/src/unary.cu | 4 |
1 files changed, 4 insertions, 0 deletions
diff --git a/candle-kernels/src/unary.cu b/candle-kernels/src/unary.cu index 0aab40cb..85d74b82 100644 --- a/candle-kernels/src/unary.cu +++ b/candle-kernels/src/unary.cu @@ -80,6 +80,7 @@ extern "C" __global__ void FN_NAME( \ #if __CUDA_ARCH__ >= 800 UNARY_OP(__nv_bfloat16, ucopy_bf16, x) UNARY_OP(__nv_bfloat16, uneg_bf16, -x) +UNARY_OP(__nv_bfloat16, urecip_bf16, recipg(x)) UNARY_OP(__nv_bfloat16, uexp_bf16, expg(x)) UNARY_OP(__nv_bfloat16, ulog_bf16, logg(x)) UNARY_OP(__nv_bfloat16, usin_bf16, sing(x)) @@ -95,6 +96,7 @@ UNARY_OP1(__nv_bfloat16, uelu_bf16, elu_fwd(x, param)) #if __CUDA_ARCH__ >= 530 UNARY_OP(__half, ucopy_f16, x) UNARY_OP(__half, uneg_f16, -x) +UNARY_OP(__half, urecip_f16, recipg(x)) UNARY_OP(__half, uexp_f16, expg(x)) UNARY_OP(__half, ulog_f16, logg(x)) UNARY_OP(__half, usin_f16, sing(x)) @@ -113,6 +115,8 @@ 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, urecip_f32, recipg(x)) +UNARY_OP(double, urecip_f64, recipg(x)) UNARY_OP(float, uexp_f32, expg(x)) UNARY_OP(double, uexp_f64, expg(x)) UNARY_OP(float, ulog_f32, logg(x)) |