提交 5f863682 编写于 作者: M Megvii Engine Team

Revert "feat(dnn): add elemwise modes"

This reverts commit cb713ddb24e09b9426eb12c5a9767c308d65fd60.

GitOrigin-RevId: 02adf025e602415be9ea0367e6fd1cb8feca6492
上级 d2a1905a
......@@ -14,27 +14,23 @@ MODES = {
1: ['RELU', 'ABS', 'NEGATE', 'ACOS', 'ASIN', 'CEIL', 'COS',
'EXP', 'EXPM1', 'FLOOR', 'LOG', 'LOG1P', 'SIGMOID', 'SIN',
'TANH', 'FAST_TANH', 'ROUND', 'ERF', 'ERFINV', 'ERFC',
'ERFCINV', 'H_SWISH', 'SILU', 'GELU', 'SINH', 'COSH',
'ASINH', 'ACOSH', 'ATANH', 'TAN', 'SOFTPLUS', 'RELU6',
'HSIGMOID', 'LOGSIGMOID', 'SQRT', 'SQUARE', 'SIGN'],
'ERFCINV', 'H_SWISH', 'SILU', 'GELU'],
2: ['ABS_GRAD', 'ADD', 'FLOOR_DIV', 'MAX', 'MIN', 'MOD', 'MUL',
'SIGMOID_GRAD', 'SUB', 'SWITCH_GT0', 'TANH_GRAD', 'LT',
'LEQ', 'EQ', 'FUSE_ADD_RELU', 'TRUE_DIV', 'POW',
'LOG_SUM_EXP', 'FUSE_ADD_TANH', 'FAST_TANH_GRAD',
'FUSE_ADD_SIGMOID', 'ATAN2', 'H_SWISH_GRAD',
'FUSE_ADD_H_SWISH', 'SILU_GRAD', 'GELU_GRAD', 'PRELU',
'ASINH_GRAD', 'ACOSH_GRAD', 'ATANH_GRAD', 'SOFTPLUS_GRAD',
'RELU6_GRAD', 'HSIGMOID_GRAD'],
3: ['COND_LEQ_MOV', 'COND_LT_MOV', 'FUSE_MUL_ADD3', 'CLIP', 'PRELU_GRAD'],
'FUSE_ADD_H_SWISH', 'SILU_GRAD', 'GELU_GRAD'],
3: ['COND_LEQ_MOV', 'COND_LT_MOV', 'FUSE_MUL_ADD3'],
}
QINT4_MODES = {
1: ['RELU', 'ABS', 'NEGATE', 'CEIL', 'FLOOR', 'SIGMOID',
'TANH', 'FAST_TANH', 'ROUND', 'H_SWISH'],
2: ['ADD', 'MAX', 'MIN', 'MUL', 'SUB', 'SWITCH_GT0',
'LT', 'LEQ', 'EQ', 'FUSE_ADD_RELU', 'FUSE_ADD_TANH',
'FUSE_ADD_SIGMOID', 'FUSE_ADD_H_SWISH', 'PRELU'],
3: ['COND_LEQ_MOV', 'COND_LT_MOV', 'FUSE_MUL_ADD3', 'CLIP'],
2: ['ADD', 'MAX', 'MIN', 'MUL', 'SUB', 'SWITCH_GT0',
'LT', 'LEQ', 'EQ', 'FUSE_ADD_RELU', 'FUSE_ADD_TANH',
'FUSE_ADD_SIGMOID', 'FUSE_ADD_H_SWISH'],
3: ['COND_LEQ_MOV', 'COND_LT_MOV', 'FUSE_MUL_ADD3'],
}
QINT32_MODES = {
......
......@@ -12,27 +12,23 @@ DTYPES = {'dt_int32': ('Int32', 'INT'),
}
MODES = {
(1, 'INT'): ['RELU', 'ABS', 'NEGATE', 'RELU6', 'SQUARE', 'SIGN'],
(1, 'INT'): ['RELU', 'ABS', 'NEGATE'],
(2, 'INT'): ['ABS_GRAD', 'ADD', 'FLOOR_DIV', 'MAX', 'MIN', 'MOD', 'MUL',
'SIGMOID_GRAD', 'SUB', 'SWITCH_GT0', 'TANH_GRAD', 'LT', 'LEQ',
'EQ', 'FUSE_ADD_RELU', 'SHL', 'SHR', 'RMULH', 'PRELU'],
(3, 'INT'): ['COND_LEQ_MOV', 'COND_LT_MOV', 'CLIP'],
'EQ', 'FUSE_ADD_RELU', 'SHL', 'SHR', 'RMULH'],
(3, 'INT'): ['COND_LEQ_MOV', 'COND_LT_MOV'],
(1, 'FLOAT'): ['RELU', 'ABS', 'NEGATE', 'ACOS', 'ASIN', 'CEIL', 'COS',
'EXP', 'EXPM1', 'FLOOR', 'LOG', 'LOG1P', 'SIGMOID', 'SIN',
'TANH', 'FAST_TANH', 'ROUND', 'ERF', 'ERFINV', 'ERFC',
'ERFCINV', 'H_SWISH', 'SILU', 'GELU', 'SINH', 'COSH',
'ASINH', 'ACOSH', 'ATANH', 'TAN', 'SOFTPLUS', 'RELU6',
'HSIGMOID', 'LOGSIGMOID', 'SQRT', 'SQUARE', 'SIGN'],
'ERFCINV', 'H_SWISH', 'SILU', 'GELU'],
(2, 'FLOAT'): ['ABS_GRAD', 'ADD', 'FLOOR_DIV', 'MAX', 'MIN', 'MOD', 'MUL',
'SIGMOID_GRAD', 'SUB', 'SWITCH_GT0', 'TANH_GRAD', 'LT',
'LEQ', 'EQ', 'FUSE_ADD_RELU', 'TRUE_DIV', 'POW',
'LOG_SUM_EXP', 'FUSE_ADD_TANH', 'FAST_TANH_GRAD',
'FUSE_ADD_SIGMOID', 'ATAN2', 'H_SWISH_GRAD',
'FUSE_ADD_H_SWISH', 'SILU_GRAD', 'GELU_GRAD', 'PRELU',
'ASINH_GRAD', 'ACOSH_GRAD', 'ATANH_GRAD', 'SOFTPLUS_GRAD',
'RELU6_GRAD', 'HSIGMOID_GRAD'],
(3, 'FLOAT'): ['COND_LEQ_MOV', 'COND_LT_MOV', 'FUSE_MUL_ADD3', 'CLIP', 'PRELU_GRAD'],
'FUSE_ADD_H_SWISH', 'SILU_GRAD', 'GELU_GRAD'],
(3, 'FLOAT'): ['COND_LEQ_MOV', 'COND_LT_MOV', 'FUSE_MUL_ADD3'],
(1, 'BOOL'): ['NOT'],
(2, 'BOOL'): ['AND', 'OR', 'XOR', 'LT', 'LEQ', 'EQ'],
(3, 'BOOL'): []
......
......@@ -421,31 +421,9 @@ pdef('Elemwise').add_enum(
Doc('GELU = 58', 'unary: x Phi(x)'),
Doc('GELU_GRAD = 59', 'binary: grad(x Phi(x))'),
Doc('COND_LT_MOV = 60', 'ternary: x < y ? z : 0'),
Doc('SINH = 61', 'unary: sinh(x)'),
Doc('COSH = 62', 'unary: cosh(x)'),
Doc('ASINH = 63', 'unary: asinh(x)'),
Doc('ACOSH = 64', 'unary: acosh(x)'),
Doc('ATANH = 65', 'unary: atanh(x)'),
Doc('TAN = 66', 'unary: tan(x)'),
Doc('ASINH_GRAD = 67', 'binary: y / sqrt(x^2 + 1)'),
Doc('ACOSH_GRAD = 68', 'binary: y / sqrt(x^2 - 1) (x > 1)'),
Doc('ATANH_GRAD = 69', 'binary: y / (1 - x^2) (|x| < 1)'),
Doc('PRELU = 70', 'binary: x > 0 ? x : x * y'),
Doc('CLIP = 71', 'ternary: x <= y ? y : (x <= z ? x : z)'),
Doc('PRELU_GRAD = 72', 'ternary: x > 0 ? y : y * z'),
Doc('SOFTPLUS = 73', 'unary: log(1 + e^x)'),
Doc('SOFTPLUS_GRAD = 74', 'binary: y * e^x / (1 + e^x)'),
Doc('RELU6 = 75', 'unary: min(max(0, x), 6)'),
Doc('RELU6_GRAD = 76', 'binary: x < 0 ? 0 : (x > 6 ? 0 : y)'),
Doc('HSIGMOID = 77', 'unary: relu6(x + 3) / 6'),
Doc('HSIGMOID_GRAD = 78', 'binary: x < -3 ? 0 : (x > 3 ? 0 : y / 6)'),
Doc('LOGSIGMOID = 79', 'unary: -log(1 + e^(-x))'),
Doc('SQRT = 80', 'unary: x^(1/2)'),
Doc('SQUARE = 81', 'unary: x^2'),
Doc('SIGN = 82', 'unary: sgn(x)'),
Doc('NEQ = 83', 'binary: x != y'),
Doc('ISNAN = 84', 'unary: isnan(x)'),
Doc('ISINF = 85', 'unary: isinf(x)'),
Doc('NEQ = 61', 'binary: x != y'),
Doc('ISNAN = 62', 'unary: isnan(x)'),
Doc('ISINF = 63', 'unary: isinf(x)'),
)
pdef('ElemwiseMultiType').add_enum(
......
......@@ -25,28 +25,12 @@
MEGDNN_ELEMWISE_MODE_ENABLE(ERFCINV, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(H_SWISH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SILU, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(GELU, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SINH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(COSH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ASINH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ATANH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(TAN, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(LOGSIGMOID, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SQRT, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
MEGDNN_ELEMWISE_MODE_ENABLE(GELU, cb)
#define MEGDNN_FOREACH_ELEMWISE_MODE_UNARY_INT(cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(RELU, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ABS, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(NEGATE, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
MEGDNN_ELEMWISE_MODE_ENABLE(NEGATE, cb)
#define MEGDNN_FOREACH_ELEMWISE_MODE_BINARY_BOOL(cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(AND, cb) \
......@@ -82,14 +66,7 @@
MEGDNN_ELEMWISE_MODE_ENABLE(H_SWISH_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(FUSE_ADD_H_SWISH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SILU_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(GELU_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ASINH_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ATANH_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(RELU6_GRAD, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID_GRAD, cb)
MEGDNN_ELEMWISE_MODE_ENABLE(GELU_GRAD, cb)
#define MEGDNN_FOREACH_ELEMWISE_MODE_BINARY_INT(cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(ABS_GRAD, cb) \
......@@ -109,19 +86,15 @@
MEGDNN_ELEMWISE_MODE_ENABLE(FUSE_ADD_RELU, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SHL, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(SHR, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(RMULH, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
MEGDNN_ELEMWISE_MODE_ENABLE(RMULH, cb)
#define MEGDNN_FOREACH_ELEMWISE_MODE_TERNARY_BOOL(cb)
#define MEGDNN_FOREACH_ELEMWISE_MODE_TERNARY_FLOAT(cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(COND_LEQ_MOV, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(COND_LT_MOV, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(FUSE_MUL_ADD3, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(PRELU_GRAD, cb)
MEGDNN_ELEMWISE_MODE_ENABLE(FUSE_MUL_ADD3, cb)
#define MEGDNN_FOREACH_ELEMWISE_MODE_TERNARY_INT(cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(COND_LEQ_MOV, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(COND_LT_MOV, cb) \
MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
MEGDNN_ELEMWISE_MODE_ENABLE(COND_LT_MOV, cb)
......@@ -154,18 +154,11 @@ struct ElemwiseKern;
// int and float
DEF_KERN_ALL(NEGATE, -x);
DEF_KERN_ALL(SQUARE, x* x);
#if defined(__HIP_PLATFORM_HCC__) && !defined(__HIP_PLATFORM_NVCC__)
DEF_KERN_INT(RELU, x <= ctype(0) ? ctype(0) : x);
DEF_KERN_INT(RELU6, x <= ctype(0) ? ctype(0) : (x <= ctype(6) ? x : ctype(6)));
DEF_KERN_INT(SIGN, x < ctype(0) ? ctype(-1) : (x > ctype(0) ? ctype(1) : ctype(0)));
DEF_KERN_FLOAT(RELU, x <= 0.f ? ctype(0) : x);
DEF_KERN_FLOAT(RELU6, x <= 6.f ? ctype(0) : (x <= 6.f ? x : ctype(6)));
DEF_KERN_FLOAT(SIGN, x < 0.f ? -1.f : (x > 0.f ? 1.f : 0.f));
#else
DEF_KERN_ALL(RELU, x <= ctype(0) ? ctype(0) : x);
DEF_KERN_ALL(RELU6, x <= ctype(0) ? ctype(0) : (x <= ctype(6) ? x : ctype(6)));
DEF_KERN_ALL(SIGN, x < ctype(0) ? ctype(-1) : (x > ctype(0) ? ctype(1) : ctype(0)));
#endif
DEF_KERN_INT(ABS, abs(int(x)));
// DEF_KERN_INT(ABS, x > ctype(0) ? x : -x);
......@@ -193,18 +186,6 @@ DEF_KERN_FLOAT(ERFCINV, erfcinvf(x));
DEF_KERN_FLOAT(H_SWISH, x* min(max(x + 3, 0.f), 6.f) * (1.f / 6.f));
DEF_KERN_FLOAT(SILU, x / (expf(-x) + 1.f));
DEF_KERN_FLOAT(GELU, x* normcdf(x));
DEF_KERN_FLOAT(SINH, sinhf(x));
DEF_KERN_FLOAT(COSH, coshf(x));
DEF_KERN_FLOAT(ASINH, asinhf(x));
DEF_KERN_FLOAT(ACOSH, acoshf(x));
DEF_KERN_FLOAT(ATANH, atanhf(x));
DEF_KERN_FLOAT(TAN, tanf(x));
DEF_KERN_FLOAT(SOFTPLUS, log1pf(expf(-fabsf(x))) + (x <= ctype(0) ? ctype(0) : x));
DEF_KERN_FLOAT(
HSIGMOID,
x <= ctype(-3) ? ctype(0) : (x >= ctype(3) ? ctype(1) : ((x + 3.f) / 6.f)));
DEF_KERN_FLOAT(SQRT, sqrtf(x));
DEF_KERN_FLOAT(LOGSIGMOID, -log1pf(expf(-fabsf(x))) + (x >= ctype(0) ? ctype(0) : x));
// int only
DEF_KERN(dt_bool, NOT, x ^ 1);
......@@ -259,12 +240,6 @@ DEF_KERN_FLOAT(FUSE_ADD_RELU, (x + y) <= 0.f ? ctype(0) : (x + y));
#else
DEF_KERN_ALL(FUSE_ADD_RELU, (x + y) <= ctype(0) ? ctype(0) : (x + y));
#endif
#if defined(__HIP_PLATFORM_HCC__) && !defined(__HIP_PLATFORM_NVCC__)
DEF_KERN_INT(PRELU, x > ctype(0) ? x : (x * y));
DEF_KERN_FLOAT(PRELU, x > 0.f ? x : (x * y));
#else
DEF_KERN_ALL(PRELU, x > ctype(0) ? x : (x * y));
#endif
// float only
DEF_KERN_FLOAT(TRUE_DIV, x / y);
......@@ -284,14 +259,6 @@ DEF_KERN_FLOAT(
DEF_KERN_FLOAT(FUSE_ADD_H_SWISH, fuse_add_hswish(x, y));
DEF_KERN_FLOAT(SILU_GRAD, silu_grad(x, y));
DEF_KERN_FLOAT(GELU_GRAD, gelu_grad(x, y));
DEF_KERN_FLOAT(ASINH_GRAD, y / sqrt(x * x + 1.f));
DEF_KERN_FLOAT(ACOSH_GRAD, y / sqrt(x * x - 1.f));
DEF_KERN_FLOAT(ATANH_GRAD, y / (1.f - x * x));
DEF_KERN_FLOAT(SOFTPLUS_GRAD, y* expf(x) / (1.f + expf(x)));
DEF_KERN_FLOAT(RELU6_GRAD, x <= ctype(0) ? ctype(0) : (x >= ctype(6) ? ctype(0) : y));
DEF_KERN_FLOAT(
HSIGMOID_GRAD,
x <= ctype(-3) ? ctype(0) : (x >= ctype(3) ? ctype(0) : (y / 6.f)));
#undef KERN_SIG
/* ================== ternary kernels ================== */
......@@ -301,8 +268,6 @@ DEF_KERN_FLOAT(
DEF_KERN_ALL(COND_LEQ_MOV, x <= y ? z : ctype(0));
DEF_KERN_ALL(COND_LT_MOV, x < y ? z : ctype(0));
DEF_KERN_ALL(FUSE_MUL_ADD3, x* y + z);
DEF_KERN_ALL(CLIP, x <= y ? y : (x <= z ? x : z));
DEF_KERN_FLOAT(PRELU_GRAD, x >= 0.f ? y : (y * z));
#undef KERN_SIG
......
......@@ -220,28 +220,6 @@ const ModeTrait& ModeTrait::from_mode(Mode mode) {
CB_MODE(Mode::GELU);
CB_MODE(Mode::GELU_GRAD);
CB_MODE(Mode::COND_LT_MOV);
CB_MODE(Mode::SINH);
CB_MODE(Mode::COSH);
CB_MODE(Mode::ASINH);
CB_MODE(Mode::ACOSH);
CB_MODE(Mode::ATANH);
CB_MODE(Mode::TAN);
CB_MODE(Mode::ASINH_GRAD);
CB_MODE(Mode::ACOSH_GRAD);
CB_MODE(Mode::ATANH_GRAD);
CB_MODE(Mode::PRELU);
CB_MODE(Mode::PRELU_GRAD);
CB_MODE(Mode::CLIP);
CB_MODE(Mode::SOFTPLUS);
CB_MODE(Mode::SOFTPLUS_GRAD);
CB_MODE(Mode::RELU6);
CB_MODE(Mode::RELU6_GRAD);
CB_MODE(Mode::HSIGMOID);
CB_MODE(Mode::HSIGMOID_GRAD);
CB_MODE(Mode::LOGSIGMOID);
CB_MODE(Mode::SQRT);
CB_MODE(Mode::SQUARE);
CB_MODE(Mode::SIGN);
default:
megdnn_assert(
0,
......
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ACOSH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ASINH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ASINH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ASINH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ASINH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ASINH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ASINH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ATANH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ATANH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ATANH_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ATANH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ATANH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(ATANH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_int16
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_int32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_int8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_uint8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(COSH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(COSH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(COSH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(HSIGMOID, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(LOGSIGMOID, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(LOGSIGMOID, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(LOGSIGMOID, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU_GRAD, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU_GRAD, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU_GRAD, cb)
#define KERN_IMPL_ARITY 3
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_int16
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_int32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_int8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(PRELU, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_uint8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int16
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(RELU6, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_uint8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int16
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SIGN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_uint8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SINH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SINH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SINH, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS_GRAD, cb)
#define KERN_IMPL_ARITY 2
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SOFTPLUS, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQRT, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQRT, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQRT, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int16
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int32
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_int8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(SQUARE, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_uint8
#include "../kern_impl.inl"
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(TAN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_bfloat16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#if !MEGDNN_DISABLE_FLOAT16
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(TAN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float16
#include "../kern_impl.inl"
#endif
// generated by gen_elemwise_kern_impls.py
#define KERN_IMPL_MODE(cb) MEGDNN_ELEMWISE_MODE_ENABLE(TAN, cb)
#define KERN_IMPL_ARITY 1
#define KERN_IMPL_CTYPE dt_float32
#include "../kern_impl.inl"
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册