24.02.1
|
Go to the documentation of this file.
27 #if GPU_ARCH == GPU_ARCH_BIFROST
28 #define MLA(a, b, c) (fma(c, b, a))
29 #else // GPU_ARCH == GPU_ARCH_BIFROST
30 #define MLA(a, b, c) ((b) * (c) + (a))
31 #endif // GPU_ARCH == GPU_ARCH_BIFROST
34 #define hard_swish_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) \
35 (x * ((min(max((x + (DATA_TYPE)3.0), (DATA_TYPE)0.0), (DATA_TYPE)6.0)) * (DATA_TYPE)0.166666667))
38 #define logistic_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) ((DATA_TYPE)1.0 / ((DATA_TYPE)1.0 + exp(-x)))
41 #define tanh_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) ((DATA_TYPE)A_VAL * tanh((DATA_TYPE)B_VAL * x))
44 #define relu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (max((DATA_TYPE)0.0, x))
47 #define brelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)0.0, x)))
50 #define lu_brelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL))
53 #define lrelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) \
54 ((min(x, (DATA_TYPE)0.0) * (DATA_TYPE)A_VAL) + max(x, (DATA_TYPE)0.0))
57 #define srelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (log((DATA_TYPE)1.0 + exp(x)))
60 #define elu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) \
61 (select(((DATA_TYPE)A_VAL * (exp(x) - (DATA_TYPE)1.0)), x, \
62 (SELECT_VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))isgreaterequal(x, (DATA_TYPE)0.0)))
65 #define abs_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (fabs(x))
68 #define square_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (x * x)
71 #define sqrt_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (sqrt(x))
74 #define linear_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (MLA((DATA_TYPE)B_VAL, (DATA_TYPE)A_VAL, x))
77 #define gelu_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) \
78 (x * (DATA_TYPE)0.5 * ((DATA_TYPE)1.0 + erf(x / (DATA_TYPE)1.41421356237)))
81 #define identity_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) (x)
83 #define ACT_OP(op, DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) op##_op(DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL)
85 #define ACTIVATION(op, DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL) ACT_OP(op, DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL)