diff --git a/dnn/scripts/gen_elemwise_multi_type_utils.py b/dnn/scripts/gen_elemwise_multi_type_utils.py
index 39aec81819da22113d9180e506cca3daeed5fbe7..6de5a5124476564c33555fe6614c1d85b5f3152a 100755
--- a/dnn/scripts/gen_elemwise_multi_type_utils.py
+++ b/dnn/scripts/gen_elemwise_multi_type_utils.py
@@ -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 = {
diff --git a/dnn/scripts/gen_elemwise_utils.py b/dnn/scripts/gen_elemwise_utils.py
index 52f28bb79f15661e31128d1921ff0156b72a075d..84bc541f315ed0f338858891b00b81ec45642c44 100755
--- a/dnn/scripts/gen_elemwise_utils.py
+++ b/dnn/scripts/gen_elemwise_utils.py
@@ -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'): []
diff --git a/dnn/scripts/opr_param_defs.py b/dnn/scripts/opr_param_defs.py
index 0e901b57c6f6538843354fddb5f91627a3ef7b1e..1c7dd1938f6cdec17e2c655e5456777a390a22d7 100755
--- a/dnn/scripts/opr_param_defs.py
+++ b/dnn/scripts/opr_param_defs.py
@@ -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(
diff --git a/dnn/src/common/elemwise/each_mode.inl b/dnn/src/common/elemwise/each_mode.inl
index 48ca51dc65a2b7466e2f57bacbd1a4e656e337da..67a144ca63fa5256c4eed54791e75855ff184ba0 100644
--- a/dnn/src/common/elemwise/each_mode.inl
+++ b/dnn/src/common/elemwise/each_mode.inl
@@ -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)
diff --git a/dnn/src/common/elemwise/kern_defs.cuh b/dnn/src/common/elemwise/kern_defs.cuh
index 951054708b4901bd5bf26d80f4c2ae2030fa328a..87788c0d8217e7930836144e6138dd8c6994e99f 100644
--- a/dnn/src/common/elemwise/kern_defs.cuh
+++ b/dnn/src/common/elemwise/kern_defs.cuh
@@ -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
 
diff --git a/dnn/src/common/elemwise/opr_impl.cpp b/dnn/src/common/elemwise/opr_impl.cpp
index 2f0a1d5ab2991ab44789dbd3a7de4a5de26f173b..ed994866c653b3baf6241b00ecb7853569c617c0 100644
--- a/dnn/src/common/elemwise/opr_impl.cpp
+++ b/dnn/src/common/elemwise/opr_impl.cpp
@@ -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,
diff --git a/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cu
deleted file mode 100644
index fc3af7a753d1384d5ef5bf97c728b9bcfcc12d99..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_float16.cu
deleted file mode 100644
index 1c4f89c8f04c1117f9d69bf522d4025a33918fa8..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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_float16
-#include "../kern_impl.inl"
-#endif
diff --git a/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_float32.cu
deleted file mode 100644
index 7674459b3698a33b1ec6a21093cd9027af0f4e4c..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ACOSH_GRAD_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_bfloat16.cu
deleted file mode 100644
index 80411b71b2cbf6f5e253705896bc8bffe65ff478..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_float16.cu
deleted file mode 100644
index aa417709bbb389dc4ecb2036fab84f9b7f517243..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_float32.cu
deleted file mode 100644
index cfcf7ad3488aa53bd75ee02994ffa4849a804147..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ACOSH_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cu
deleted file mode 100644
index 393728a4da3ea08b625e5feca888a419e821802b..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_float16.cu
deleted file mode 100644
index 807bedca44031f25d83edf83d1753f9312dcb903..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_float32.cu
deleted file mode 100644
index 7a3d5a3edf854c6c184d3eb98247793c8c9dbefc..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ASINH_GRAD_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/ASINH_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/ASINH_dt_bfloat16.cu
deleted file mode 100644
index 7a9123815dbdd5952974f112a679e5f2b86fe974..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ASINH_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ASINH_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/ASINH_dt_float16.cu
deleted file mode 100644
index 29c9307d2cf90ee005479bae3482adbf8c5c48a7..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ASINH_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ASINH_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/ASINH_dt_float32.cu
deleted file mode 100644
index b6fc8ab6412d8753cb5aa7c0b233cc369fc91199..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ASINH_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cu
deleted file mode 100644
index cf7afe1b76c63a1cf00a903ed1f5d5f0cadbc391..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_float16.cu
deleted file mode 100644
index 333b8f606c6be46342e53fda9eecdebf100ef005..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_float32.cu
deleted file mode 100644
index be794c32d835e9cf4fd8789f559ccd67910942b4..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ATANH_GRAD_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/ATANH_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/ATANH_dt_bfloat16.cu
deleted file mode 100644
index fa6683dfeaa38572803e57dc67edf14dd10592ad..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ATANH_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ATANH_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/ATANH_dt_float16.cu
deleted file mode 100644
index 804e5361338cc49ad03fb3b8630762327cb17d1a..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ATANH_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/ATANH_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/ATANH_dt_float32.cu
deleted file mode 100644
index 5fd222b3a2e5422cb79163f8a9f5f79f2154baa1..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/ATANH_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/CLIP_dt_bfloat16.cu
deleted file mode 100644
index de73cd15d2c246cf5006379238ae4fc6b3f49682..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/CLIP_dt_float16.cu
deleted file mode 100644
index 25e351c123538c392bbc936332893723ba6410a5..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/CLIP_dt_float32.cu
deleted file mode 100644
index cf8dc776ff0ed87a3167dbe07fbce5346b30f357..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int16.cu b/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int16.cu
deleted file mode 100644
index f60b5c4ced3b3541b2734339dd3d4fae7c6e9b5e..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int16.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int32.cu b/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int32.cu
deleted file mode 100644
index c003f595ea958af3884485680a3678b1eb8598cf..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int8.cu b/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int8.cu
deleted file mode 100644
index cb0ec0463dfe7d38043d321085f150fe8bee1826..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_int8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_uint8.cu b/dnn/src/cuda/elemwise/kimpl/CLIP_dt_uint8.cu
deleted file mode 100644
index b0198d93d68ba3cf90d60bbf262e6f44a02e1ffe..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/CLIP_dt_uint8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/COSH_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/COSH_dt_bfloat16.cu
deleted file mode 100644
index 7cb17527beb5abb7f2cdba58f3633c9c5f309be2..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/COSH_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/COSH_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/COSH_dt_float16.cu
deleted file mode 100644
index 5f42f2350232482a8fa6e3b2bdc5e3841baa89da..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/COSH_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/COSH_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/COSH_dt_float32.cu
deleted file mode 100644
index 94ea18705cbb641fb4d8afb3bc23544db120e892..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/COSH_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cu
deleted file mode 100644
index a8115bff8428750645a818544b7fa829248824b4..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cu
deleted file mode 100644
index a1fb7ee3ee5f1c288281a8bb7cb5445c0c596d43..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cu
deleted file mode 100644
index 9c0a4aebd89bd9eb4aca61a27e5b4b4f8e393b23..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_bfloat16.cu
deleted file mode 100644
index 28a83976f4009d40d50e31fdeb6d94cf9bee5ab0..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_float16.cu
deleted file mode 100644
index cdb77455f7aeb8a6207d40d0c38d0acdb94d77ba..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_float32.cu
deleted file mode 100644
index 528f944c2a66eb3dcc36868dfc4f89c8c9bae099..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/HSIGMOID_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cu
deleted file mode 100644
index 06322df63cecfe034ba9b53b9a7105c163e57e89..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_float16.cu
deleted file mode 100644
index d0b6c026851ea01f3f796b8e9ccda787f51fceb3..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_float32.cu
deleted file mode 100644
index ea1bcf1a3ab05774fd28e69152f18f37c80cbdbf..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/LOGSIGMOID_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cu
deleted file mode 100644
index dc33150072bc4d40f3d56a9b45e9d26c2985b559..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_float16.cu
deleted file mode 100644
index 3441181806c090103f7e7711b52ce4c417252204..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_float32.cu
deleted file mode 100644
index 1fedc850076ee187afd30c48b51785fb958dc357..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_GRAD_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_dt_bfloat16.cu
deleted file mode 100644
index 78c18f9d716524a01880336c03e906b9d646cd9c..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_dt_float16.cu
deleted file mode 100644
index 33e6ce7393622702eca9c187eaa1cb964b8f13b6..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_dt_float32.cu
deleted file mode 100644
index 46f2d36705d6ae3c09361b540bd5650750e71ae8..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int16.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int16.cu
deleted file mode 100644
index d1dfa9ace337bbda9b40e88137fa43a59baa9eb0..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int16.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int32.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int32.cu
deleted file mode 100644
index d6d7332f8ab7761a13ea68eacc38005b10e0b62b..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int8.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int8.cu
deleted file mode 100644
index 621a7dd311d38be2b39117270e7398de22086edb..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_int8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_uint8.cu b/dnn/src/cuda/elemwise/kimpl/PRELU_dt_uint8.cu
deleted file mode 100644
index 86ff475d046afda072fd0ba6de02b432d79cad72..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/PRELU_dt_uint8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cu
deleted file mode 100644
index 906996656ee3c6f0cfb24a3a31d00a3e8544f046..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_float16.cu
deleted file mode 100644
index efb61fa6b23dce1fdc68fca7d9e047d3d57cf94d..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_float32.cu
deleted file mode 100644
index 6088f41deae5038ebf6772f76b69cfdb97e81192..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_GRAD_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_dt_bfloat16.cu
deleted file mode 100644
index cf79b7e965a6d0eb7b0a3c83669ac7bc9c808d72..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_dt_float16.cu
deleted file mode 100644
index 0646045d1ec3a0c56117c3c321cf68fd1a656ccf..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_dt_float32.cu
deleted file mode 100644
index 2fe7746f70244b9c2cad4e2ca0a8dee11244cb2a..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int16.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int16.cu
deleted file mode 100644
index 32c2dab37cfed97cb66a8b5352712312dd6817eb..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int16.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int32.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int32.cu
deleted file mode 100644
index e59877c32838c7a7a1b2943c76a4af25af1e56c5..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int8.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int8.cu
deleted file mode 100644
index 6f6f77416c5b1da4255c2eae3ba5ebe2c47ba3ff..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_int8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_uint8.cu b/dnn/src/cuda/elemwise/kimpl/RELU6_dt_uint8.cu
deleted file mode 100644
index 60812b5509ddcc27abe7507bc85461a6b0de0d94..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/RELU6_dt_uint8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/SIGN_dt_bfloat16.cu
deleted file mode 100644
index 34316156e524f88bcfc8e683fed67c69a4bba202..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/SIGN_dt_float16.cu
deleted file mode 100644
index 04ac0b8687bdeac1be29f152c33278ecc17a3257..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/SIGN_dt_float32.cu
deleted file mode 100644
index 0402184f056634b18572acbe4735bee01b9b39d6..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int16.cu b/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int16.cu
deleted file mode 100644
index 0a854c234513c1f50230e8a62d9a1066dbb84ae5..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int16.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int32.cu b/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int32.cu
deleted file mode 100644
index 5f3aa92707bc2a41e90984fee2d88c1a70e123eb..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int8.cu b/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int8.cu
deleted file mode 100644
index c0d44608aa081a9b4cfcd23f6da42800c06aadcd..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_int8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_uint8.cu b/dnn/src/cuda/elemwise/kimpl/SIGN_dt_uint8.cu
deleted file mode 100644
index 37f4b4b24ef2f272c6fb465bf45eff288c9417d9..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SIGN_dt_uint8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SINH_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/SINH_dt_bfloat16.cu
deleted file mode 100644
index 19b3b24d4bccffd1d20b6776fcacc1937219341c..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SINH_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SINH_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/SINH_dt_float16.cu
deleted file mode 100644
index 0298140ee348480cc71f7dc46b9700adfc2815f8..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SINH_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SINH_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/SINH_dt_float32.cu
deleted file mode 100644
index d781a287d648e3a91deded69a105ae383af35f31..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SINH_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cu
deleted file mode 100644
index 9769ef8793294c868338e59ed687b048f4cda809..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cu
deleted file mode 100644
index 694fea1bfbea814472c444b8f4d4203406c89256..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cu
deleted file mode 100644
index 05710880904d2589837ffbc52b3d5c17f98a95f1..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cu
deleted file mode 100644
index 7df279c7fb7c8ddfc8b7374998997c83420f373f..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_float16.cu
deleted file mode 100644
index 98d84daddbd4734908f999caf3fbe8b77e4e3cac..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_float32.cu
deleted file mode 100644
index 898996df7c3d2615134dd72c2f3c1ec9b8ee713f..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SOFTPLUS_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SQRT_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/SQRT_dt_bfloat16.cu
deleted file mode 100644
index b6483dfefcbccb56f73d5bb8ffa46d5416bb5975..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQRT_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SQRT_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/SQRT_dt_float16.cu
deleted file mode 100644
index 262e68d4b517151dd28071fddff8e8ddba3afec3..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQRT_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SQRT_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/SQRT_dt_float32.cu
deleted file mode 100644
index 1c6aa2af998578e8a18b868df8af41d0aece06bf..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQRT_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_bfloat16.cu
deleted file mode 100644
index 59f5383ecc121f378bf65621e9fbd2447597df21..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_float16.cu
deleted file mode 100644
index c53551bb16fcd02dba621a2332b4fbe37a35eba7..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_float32.cu
deleted file mode 100644
index 4282b4793173f75f5e55a7bf8a47171d10dc120a..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int16.cu b/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int16.cu
deleted file mode 100644
index 5d7bec087512364120790473ebb7623635054a0e..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int16.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int32.cu b/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int32.cu
deleted file mode 100644
index c1237c3775c46fce8f4e32eb6272b5cb7677c737..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int8.cu b/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int8.cu
deleted file mode 100644
index 8c01483e5b305d6a6f0804d753ced10d920f5aa5..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_int8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_uint8.cu b/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_uint8.cu
deleted file mode 100644
index 094c419155572eec5310118820a9df803b894b96..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/SQUARE_dt_uint8.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise/kimpl/TAN_dt_bfloat16.cu b/dnn/src/cuda/elemwise/kimpl/TAN_dt_bfloat16.cu
deleted file mode 100644
index 415ed3fc49e9d6247e037e1083409e6cf79a1a8c..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/TAN_dt_bfloat16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/TAN_dt_float16.cu b/dnn/src/cuda/elemwise/kimpl/TAN_dt_float16.cu
deleted file mode 100644
index dd5339a35c4cc5c7ea6b5294713a85bdee765bf2..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/TAN_dt_float16.cu
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/cuda/elemwise/kimpl/TAN_dt_float32.cu b/dnn/src/cuda/elemwise/kimpl/TAN_dt_float32.cu
deleted file mode 100644
index 796c4e65abc8f34328a146fe79af06e04442cdd1..0000000000000000000000000000000000000000
--- a/dnn/src/cuda/elemwise/kimpl/TAN_dt_float32.cu
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/ACOSH_GRAD_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/ACOSH_GRAD_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 74e69bb6b2c6805d293277299c271caa49a7c7d1..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/ACOSH_GRAD_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/ACOSH_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/ACOSH_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 6734fd674d5641e1c985b27b3e942ca3e73c3096..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/ACOSH_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/ASINH_GRAD_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/ASINH_GRAD_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 6af7f4060e1dad5488d42ac80efb69151f2d57bf..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/ASINH_GRAD_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/ASINH_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/ASINH_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 9b080b7369a4178755c151ca72e473096d101ce6..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/ASINH_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/ATANH_GRAD_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/ATANH_GRAD_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 1d477ec4b2c0a891712c808b6eb97a1d829a74be..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/ATANH_GRAD_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/ATANH_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/ATANH_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 934699b3207bfad342e8aafe6dda481762e3fe26..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/ATANH_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_qint4_dt_qint4.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_qint4_dt_qint4.cu
deleted file mode 100644
index accf56fa04b1d33e2e142f45285f23ee512cdff6..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_qint4_dt_qint4.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 78de5e8a6ae6843e9c90f0565ef2d1bb2966e71a..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_quint4_dt_quint4.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_quint4_dt_quint4.cu
deleted file mode 100644
index 754de3e6edf2e9d96cec4518272a677ca51a11ba..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/CLIP_dt_quint4_dt_quint4.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/COSH_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/COSH_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 2bcc45ab6702ce3a242b29db41666ba6226149f7..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/COSH_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/HSIGMOID_GRAD_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/HSIGMOID_GRAD_dt_qint8_dt_qint8.cu
deleted file mode 100644
index a50f44c9f09e013eca2a1374300c2f8b2d4fba1a..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/HSIGMOID_GRAD_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/HSIGMOID_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/HSIGMOID_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 5cac0b94bcd6223e7f57b62b0a8655e18d50bcd9..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/HSIGMOID_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/LOGSIGMOID_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/LOGSIGMOID_dt_qint8_dt_qint8.cu
deleted file mode 100644
index bb2abf6aa57caae6de6d05f3f2008ad2650c2b72..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/LOGSIGMOID_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/PRELU_GRAD_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/PRELU_GRAD_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 1518830ea8e4f5de3281a9b059fbcd74b4c2a7e4..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/PRELU_GRAD_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/PRELU_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/PRELU_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 60fa38fa2d968c7ac1e392266a3f8158c740c119..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/PRELU_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/RELU6_GRAD_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/RELU6_GRAD_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 49ef8a946a5415d193380638e82a71736b045b12..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/RELU6_GRAD_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/RELU6_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/RELU6_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 7cf2b51230ac42bac52d9df41795e0fe15718606..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/RELU6_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/SIGN_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/SIGN_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 279bb2bbb8a8321c316f26bce9903f0ed30369be..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/SIGN_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/SINH_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/SINH_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 68f6f9a7d1613bf060d217037c73401ae68102df..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/SINH_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/SOFTPLUS_GRAD_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/SOFTPLUS_GRAD_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 2c8aa8b61f48dab8cf290db6f4ca395063d8942b..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/SOFTPLUS_GRAD_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/SOFTPLUS_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/SOFTPLUS_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 7c250442daf19731b1f73a094a53f41a36017cac..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/SOFTPLUS_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/SQRT_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/SQRT_dt_qint8_dt_qint8.cu
deleted file mode 100644
index 48290ef97f6853e4abc5cb11e6c5eb690511a48a..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/SQRT_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/SQUARE_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/SQUARE_dt_qint8_dt_qint8.cu
deleted file mode 100644
index a24e045707f53379b0599df1ae3bb426058edd4c..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/SQUARE_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/kimpl/TAN_dt_qint8_dt_qint8.cu b/dnn/src/cuda/elemwise_multi_type/kimpl/TAN_dt_qint8_dt_qint8.cu
deleted file mode 100644
index ecf0bcb60437b0e6449787326530988e82244bd9..0000000000000000000000000000000000000000
Binary files a/dnn/src/cuda/elemwise_multi_type/kimpl/TAN_dt_qint8_dt_qint8.cu and /dev/null differ
diff --git a/dnn/src/cuda/elemwise_multi_type/opr_impl.cpp b/dnn/src/cuda/elemwise_multi_type/opr_impl.cpp
index 2f45e524199928a474d2e73ca5e803c843ab92d0..64a90258a26a60b58bdc3cadae6002990e1eb588 100644
--- a/dnn/src/cuda/elemwise_multi_type/opr_impl.cpp
+++ b/dnn/src/cuda/elemwise_multi_type/opr_impl.cpp
@@ -267,10 +267,7 @@ IMPL_MODE_DISPATCHER(2, dt_qint4, dt_qint4);
 IMPL_MODE_DISPATCHER(2, dt_quint4, dt_quint4);
 #undef FOREACH
 
-#define FOREACH(cb)                                \
-    MEGDNN_ELEMWISE_MODE_ENABLE(COND_LEQ_MOV, cb)  \
-    MEGDNN_ELEMWISE_MODE_ENABLE(FUSE_MUL_ADD3, cb) \
-    MEGDNN_ELEMWISE_MODE_ENABLE(CLIP, cb)
+#define FOREACH MEGDNN_FOREACH_ELEMWISE_MODE_TERNARY_FLOAT
 IMPL_MODE_DISPATCHER(3, dt_qint4, dt_qint4);
 IMPL_MODE_DISPATCHER(3, dt_quint4, dt_quint4);
 #undef FOREACH
diff --git a/dnn/src/fallback/elemwise/fallback_impl/opr_binary_impl.cpp b/dnn/src/fallback/elemwise/fallback_impl/opr_binary_impl.cpp
index f0cf6c13ef97260c56ca7212b4eb217a549003f5..4197a2da522fb14fe95f315f008fcc82119b04ef 100644
--- a/dnn/src/fallback/elemwise/fallback_impl/opr_binary_impl.cpp
+++ b/dnn/src/fallback/elemwise/fallback_impl/opr_binary_impl.cpp
@@ -228,7 +228,6 @@ INST(Mode::SHL);
 INST(Mode::SHR);
 INST(Mode::FUSE_ADD_RELU);
 INST(Mode::RMULH);
-INST(Mode::PRELU);
 #undef INST
 
 #define INST(mode) \
@@ -259,13 +258,6 @@ INST(Mode::H_SWISH_GRAD);
 INST(Mode::FUSE_ADD_H_SWISH);
 INST(Mode::SILU_GRAD);
 INST(Mode::GELU_GRAD);
-INST(Mode::PRELU);
-INST(Mode::ASINH_GRAD);
-INST(Mode::ACOSH_GRAD);
-INST(Mode::ATANH_GRAD);
-INST(Mode::SOFTPLUS_GRAD);
-INST(Mode::RELU6_GRAD);
-INST(Mode::HSIGMOID_GRAD);
 #undef INST
 }  // namespace fallback
 }  // namespace megdnn
diff --git a/dnn/src/fallback/elemwise/fallback_impl/opr_unary_impl.cpp b/dnn/src/fallback/elemwise/fallback_impl/opr_unary_impl.cpp
index bf7acf095dfc157ca74c97a2c90071bf9478fb57..b36e7dae23f65bdd6468f29586e1dcbd2f45428d 100644
--- a/dnn/src/fallback/elemwise/fallback_impl/opr_unary_impl.cpp
+++ b/dnn/src/fallback/elemwise/fallback_impl/opr_unary_impl.cpp
@@ -77,9 +77,6 @@ using Mode = param_enumv::Elemwise::Mode;
 INST(Mode::RELU);
 INST(Mode::ABS);
 INST(Mode::NEGATE);
-INST(Mode::RELU6);
-INST(Mode::SQUARE);
-INST(Mode::SIGN);
 #undef INST
 
 #define INST(mode) \
@@ -108,19 +105,6 @@ INST(Mode::ERFCINV);
 INST(Mode::H_SWISH);
 INST(Mode::SILU);
 INST(Mode::GELU);
-INST(Mode::SINH);
-INST(Mode::COSH);
-INST(Mode::ASINH);
-INST(Mode::ACOSH);
-INST(Mode::ATANH);
-INST(Mode::TAN);
-INST(Mode::SOFTPLUS);
-INST(Mode::RELU6);
-INST(Mode::HSIGMOID);
-INST(Mode::LOGSIGMOID);
-INST(Mode::SQRT);
-INST(Mode::SQUARE);
-INST(Mode::SIGN);
 #undef INST
 }  // namespace fallback
 }  // namespace megdnn
diff --git a/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cpp
deleted file mode 100644
index fc3af7a753d1384d5ef5bf97c728b9bcfcc12d99..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_float16.cpp
deleted file mode 100644
index 1c4f89c8f04c1117f9d69bf522d4025a33918fa8..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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_float16
-#include "../kern_impl.inl"
-#endif
diff --git a/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_float32.cpp
deleted file mode 100644
index 7674459b3698a33b1ec6a21093cd9027af0f4e4c..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ACOSH_GRAD_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/ACOSH_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/ACOSH_dt_bfloat16.cpp
deleted file mode 100644
index 80411b71b2cbf6f5e253705896bc8bffe65ff478..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ACOSH_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ACOSH_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/ACOSH_dt_float16.cpp
deleted file mode 100644
index aa417709bbb389dc4ecb2036fab84f9b7f517243..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ACOSH_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ACOSH_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/ACOSH_dt_float32.cpp
deleted file mode 100644
index cfcf7ad3488aa53bd75ee02994ffa4849a804147..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ACOSH_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cpp
deleted file mode 100644
index 393728a4da3ea08b625e5feca888a419e821802b..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_float16.cpp
deleted file mode 100644
index 807bedca44031f25d83edf83d1753f9312dcb903..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_float32.cpp
deleted file mode 100644
index 7a3d5a3edf854c6c184d3eb98247793c8c9dbefc..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ASINH_GRAD_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/ASINH_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/ASINH_dt_bfloat16.cpp
deleted file mode 100644
index 7a9123815dbdd5952974f112a679e5f2b86fe974..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ASINH_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ASINH_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/ASINH_dt_float16.cpp
deleted file mode 100644
index 29c9307d2cf90ee005479bae3482adbf8c5c48a7..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ASINH_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ASINH_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/ASINH_dt_float32.cpp
deleted file mode 100644
index b6fc8ab6412d8753cb5aa7c0b233cc369fc91199..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ASINH_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cpp
deleted file mode 100644
index cf7afe1b76c63a1cf00a903ed1f5d5f0cadbc391..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_float16.cpp
deleted file mode 100644
index 333b8f606c6be46342e53fda9eecdebf100ef005..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_float32.cpp
deleted file mode 100644
index be794c32d835e9cf4fd8789f559ccd67910942b4..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ATANH_GRAD_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/ATANH_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/ATANH_dt_bfloat16.cpp
deleted file mode 100644
index fa6683dfeaa38572803e57dc67edf14dd10592ad..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ATANH_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ATANH_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/ATANH_dt_float16.cpp
deleted file mode 100644
index 804e5361338cc49ad03fb3b8630762327cb17d1a..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ATANH_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/ATANH_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/ATANH_dt_float32.cpp
deleted file mode 100644
index 5fd222b3a2e5422cb79163f8a9f5f79f2154baa1..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/ATANH_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/CLIP_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/CLIP_dt_bfloat16.cpp
deleted file mode 100644
index de73cd15d2c246cf5006379238ae4fc6b3f49682..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/CLIP_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/CLIP_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/CLIP_dt_float16.cpp
deleted file mode 100644
index 25e351c123538c392bbc936332893723ba6410a5..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/CLIP_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/CLIP_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/CLIP_dt_float32.cpp
deleted file mode 100644
index cf8dc776ff0ed87a3167dbe07fbce5346b30f357..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/CLIP_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/CLIP_dt_int16.cpp b/dnn/src/naive/elemwise/kimpl/CLIP_dt_int16.cpp
deleted file mode 100644
index f60b5c4ced3b3541b2734339dd3d4fae7c6e9b5e..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/CLIP_dt_int16.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/CLIP_dt_int32.cpp b/dnn/src/naive/elemwise/kimpl/CLIP_dt_int32.cpp
deleted file mode 100644
index c003f595ea958af3884485680a3678b1eb8598cf..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/CLIP_dt_int32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/CLIP_dt_int8.cpp b/dnn/src/naive/elemwise/kimpl/CLIP_dt_int8.cpp
deleted file mode 100644
index cb0ec0463dfe7d38043d321085f150fe8bee1826..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/CLIP_dt_int8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/CLIP_dt_uint8.cpp b/dnn/src/naive/elemwise/kimpl/CLIP_dt_uint8.cpp
deleted file mode 100644
index b0198d93d68ba3cf90d60bbf262e6f44a02e1ffe..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/CLIP_dt_uint8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/COSH_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/COSH_dt_bfloat16.cpp
deleted file mode 100644
index 7cb17527beb5abb7f2cdba58f3633c9c5f309be2..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/COSH_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/COSH_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/COSH_dt_float16.cpp
deleted file mode 100644
index 5f42f2350232482a8fa6e3b2bdc5e3841baa89da..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/COSH_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/COSH_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/COSH_dt_float32.cpp
deleted file mode 100644
index 94ea18705cbb641fb4d8afb3bc23544db120e892..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/COSH_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cpp
deleted file mode 100644
index a8115bff8428750645a818544b7fa829248824b4..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cpp
deleted file mode 100644
index a1fb7ee3ee5f1c288281a8bb7cb5445c0c596d43..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cpp
deleted file mode 100644
index 9c0a4aebd89bd9eb4aca61a27e5b4b4f8e393b23..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_bfloat16.cpp
deleted file mode 100644
index 28a83976f4009d40d50e31fdeb6d94cf9bee5ab0..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_float16.cpp
deleted file mode 100644
index cdb77455f7aeb8a6207d40d0c38d0acdb94d77ba..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_float32.cpp
deleted file mode 100644
index 528f944c2a66eb3dcc36868dfc4f89c8c9bae099..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/HSIGMOID_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cpp
deleted file mode 100644
index 06322df63cecfe034ba9b53b9a7105c163e57e89..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_float16.cpp
deleted file mode 100644
index d0b6c026851ea01f3f796b8e9ccda787f51fceb3..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_float32.cpp
deleted file mode 100644
index ea1bcf1a3ab05774fd28e69152f18f37c80cbdbf..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/LOGSIGMOID_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cpp
deleted file mode 100644
index dc33150072bc4d40f3d56a9b45e9d26c2985b559..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_float16.cpp
deleted file mode 100644
index 3441181806c090103f7e7711b52ce4c417252204..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_float32.cpp
deleted file mode 100644
index 1fedc850076ee187afd30c48b51785fb958dc357..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_GRAD_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_dt_bfloat16.cpp
deleted file mode 100644
index 78c18f9d716524a01880336c03e906b9d646cd9c..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_dt_float16.cpp
deleted file mode 100644
index 33e6ce7393622702eca9c187eaa1cb964b8f13b6..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_dt_float32.cpp
deleted file mode 100644
index 46f2d36705d6ae3c09361b540bd5650750e71ae8..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_dt_int16.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_dt_int16.cpp
deleted file mode 100644
index d1dfa9ace337bbda9b40e88137fa43a59baa9eb0..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_dt_int16.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_dt_int32.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_dt_int32.cpp
deleted file mode 100644
index d6d7332f8ab7761a13ea68eacc38005b10e0b62b..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_dt_int32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_dt_int8.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_dt_int8.cpp
deleted file mode 100644
index 621a7dd311d38be2b39117270e7398de22086edb..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_dt_int8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/PRELU_dt_uint8.cpp b/dnn/src/naive/elemwise/kimpl/PRELU_dt_uint8.cpp
deleted file mode 100644
index 86ff475d046afda072fd0ba6de02b432d79cad72..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/PRELU_dt_uint8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cpp
deleted file mode 100644
index 906996656ee3c6f0cfb24a3a31d00a3e8544f046..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_float16.cpp
deleted file mode 100644
index efb61fa6b23dce1fdc68fca7d9e047d3d57cf94d..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_float32.cpp
deleted file mode 100644
index 6088f41deae5038ebf6772f76b69cfdb97e81192..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_GRAD_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_dt_bfloat16.cpp
deleted file mode 100644
index cf79b7e965a6d0eb7b0a3c83669ac7bc9c808d72..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_dt_float16.cpp
deleted file mode 100644
index 0646045d1ec3a0c56117c3c321cf68fd1a656ccf..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_dt_float32.cpp
deleted file mode 100644
index 2fe7746f70244b9c2cad4e2ca0a8dee11244cb2a..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_dt_int16.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_dt_int16.cpp
deleted file mode 100644
index 32c2dab37cfed97cb66a8b5352712312dd6817eb..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_dt_int16.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_dt_int32.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_dt_int32.cpp
deleted file mode 100644
index e59877c32838c7a7a1b2943c76a4af25af1e56c5..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_dt_int32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_dt_int8.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_dt_int8.cpp
deleted file mode 100644
index 6f6f77416c5b1da4255c2eae3ba5ebe2c47ba3ff..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_dt_int8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/RELU6_dt_uint8.cpp b/dnn/src/naive/elemwise/kimpl/RELU6_dt_uint8.cpp
deleted file mode 100644
index 60812b5509ddcc27abe7507bc85461a6b0de0d94..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/RELU6_dt_uint8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SIGN_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/SIGN_dt_bfloat16.cpp
deleted file mode 100644
index 34316156e524f88bcfc8e683fed67c69a4bba202..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SIGN_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SIGN_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/SIGN_dt_float16.cpp
deleted file mode 100644
index 04ac0b8687bdeac1be29f152c33278ecc17a3257..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SIGN_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SIGN_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/SIGN_dt_float32.cpp
deleted file mode 100644
index 0402184f056634b18572acbe4735bee01b9b39d6..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SIGN_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SIGN_dt_int16.cpp b/dnn/src/naive/elemwise/kimpl/SIGN_dt_int16.cpp
deleted file mode 100644
index 0a854c234513c1f50230e8a62d9a1066dbb84ae5..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SIGN_dt_int16.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SIGN_dt_int32.cpp b/dnn/src/naive/elemwise/kimpl/SIGN_dt_int32.cpp
deleted file mode 100644
index 5f3aa92707bc2a41e90984fee2d88c1a70e123eb..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SIGN_dt_int32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SIGN_dt_int8.cpp b/dnn/src/naive/elemwise/kimpl/SIGN_dt_int8.cpp
deleted file mode 100644
index c0d44608aa081a9b4cfcd23f6da42800c06aadcd..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SIGN_dt_int8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SIGN_dt_uint8.cpp b/dnn/src/naive/elemwise/kimpl/SIGN_dt_uint8.cpp
deleted file mode 100644
index 37f4b4b24ef2f272c6fb465bf45eff288c9417d9..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SIGN_dt_uint8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SINH_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/SINH_dt_bfloat16.cpp
deleted file mode 100644
index 19b3b24d4bccffd1d20b6776fcacc1937219341c..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SINH_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SINH_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/SINH_dt_float16.cpp
deleted file mode 100644
index 0298140ee348480cc71f7dc46b9700adfc2815f8..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SINH_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SINH_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/SINH_dt_float32.cpp
deleted file mode 100644
index d781a287d648e3a91deded69a105ae383af35f31..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SINH_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cpp
deleted file mode 100644
index 9769ef8793294c868338e59ed687b048f4cda809..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cpp
deleted file mode 100644
index 694fea1bfbea814472c444b8f4d4203406c89256..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cpp
deleted file mode 100644
index 05710880904d2589837ffbc52b3d5c17f98a95f1..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cpp
deleted file mode 100644
index 7df279c7fb7c8ddfc8b7374998997c83420f373f..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_float16.cpp
deleted file mode 100644
index 98d84daddbd4734908f999caf3fbe8b77e4e3cac..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_float32.cpp
deleted file mode 100644
index 898996df7c3d2615134dd72c2f3c1ec9b8ee713f..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SOFTPLUS_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SQRT_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/SQRT_dt_bfloat16.cpp
deleted file mode 100644
index b6483dfefcbccb56f73d5bb8ffa46d5416bb5975..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQRT_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SQRT_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/SQRT_dt_float16.cpp
deleted file mode 100644
index 262e68d4b517151dd28071fddff8e8ddba3afec3..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQRT_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SQRT_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/SQRT_dt_float32.cpp
deleted file mode 100644
index 1c6aa2af998578e8a18b868df8af41d0aece06bf..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQRT_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/SQUARE_dt_bfloat16.cpp
deleted file mode 100644
index 59f5383ecc121f378bf65621e9fbd2447597df21..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/SQUARE_dt_float16.cpp
deleted file mode 100644
index c53551bb16fcd02dba621a2332b4fbe37a35eba7..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/SQUARE_dt_float32.cpp
deleted file mode 100644
index 4282b4793173f75f5e55a7bf8a47171d10dc120a..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int16.cpp b/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int16.cpp
deleted file mode 100644
index 5d7bec087512364120790473ebb7623635054a0e..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int16.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int32.cpp b/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int32.cpp
deleted file mode 100644
index c1237c3775c46fce8f4e32eb6272b5cb7677c737..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int8.cpp b/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int8.cpp
deleted file mode 100644
index 8c01483e5b305d6a6f0804d753ced10d920f5aa5..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_int8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_uint8.cpp b/dnn/src/naive/elemwise/kimpl/SQUARE_dt_uint8.cpp
deleted file mode 100644
index 094c419155572eec5310118820a9df803b894b96..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/SQUARE_dt_uint8.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/naive/elemwise/kimpl/TAN_dt_bfloat16.cpp b/dnn/src/naive/elemwise/kimpl/TAN_dt_bfloat16.cpp
deleted file mode 100644
index 415ed3fc49e9d6247e037e1083409e6cf79a1a8c..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/TAN_dt_bfloat16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/TAN_dt_float16.cpp b/dnn/src/naive/elemwise/kimpl/TAN_dt_float16.cpp
deleted file mode 100644
index dd5339a35c4cc5c7ea6b5294713a85bdee765bf2..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/TAN_dt_float16.cpp
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/naive/elemwise/kimpl/TAN_dt_float32.cpp b/dnn/src/naive/elemwise/kimpl/TAN_dt_float32.cpp
deleted file mode 100644
index 796c4e65abc8f34328a146fe79af06e04442cdd1..0000000000000000000000000000000000000000
--- a/dnn/src/naive/elemwise/kimpl/TAN_dt_float32.cpp
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cpp.hip
deleted file mode 100644
index a231d0a33fec72bff4cfbda682224c8a3f215b39..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_float16.cpp.hip
deleted file mode 100644
index ca103ee3b573124f63edf218416c4b7f2838af94..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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_float16
-#include "../kern_impl.inl"
-#endif
diff --git a/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_float32.cpp.hip
deleted file mode 100644
index 11f768c9c52e1503c2bd5899c850ccc63857ab2b..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ACOSH_GRAD_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_bfloat16.cpp.hip
deleted file mode 100644
index 28c3f17367487d9a035c3d634b61380ff5f92fce..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_float16.cpp.hip
deleted file mode 100644
index 06d2c12e34e64b82460bcd9f227cff2cebb99653..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_float32.cpp.hip
deleted file mode 100644
index 9e96939608a8158628eea58c29fb1904abd25e99..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ACOSH_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cpp.hip
deleted file mode 100644
index d0f61ca23ce4e96f1634c82e561c08e7fc75a376..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_float16.cpp.hip
deleted file mode 100644
index 0840a54c1f583c8f1460c8c8b448c37033beff07..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_float32.cpp.hip
deleted file mode 100644
index c239ddb2d9035e1e903437fc19bef6877592ce3e..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ASINH_GRAD_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/ASINH_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ASINH_dt_bfloat16.cpp.hip
deleted file mode 100644
index ea0f11ecf5242865af1286a77a89e173326f07dd..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ASINH_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ASINH_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ASINH_dt_float16.cpp.hip
deleted file mode 100644
index cc5d73023c2f321aa6d1c35469783936d266538e..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ASINH_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ASINH_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ASINH_dt_float32.cpp.hip
deleted file mode 100644
index 9607630222a5105655a4b63e5a0a365a34d2b6c3..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ASINH_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cpp.hip
deleted file mode 100644
index 3b0496f40b7c662e2177be83d64938eb242d57a2..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_float16.cpp.hip
deleted file mode 100644
index 1a0c841a628be653ac2d89d63063717ab5efdaa2..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_float32.cpp.hip
deleted file mode 100644
index ce400c030129f96d30165f471910e3abe1418548..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ATANH_GRAD_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/ATANH_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ATANH_dt_bfloat16.cpp.hip
deleted file mode 100644
index 9ac3b63acf34f05af2fb0429c44ab516101a17e4..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ATANH_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ATANH_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ATANH_dt_float16.cpp.hip
deleted file mode 100644
index 56902e43605aac21546d8e020e461ede7d156a28..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ATANH_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/ATANH_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/ATANH_dt_float32.cpp.hip
deleted file mode 100644
index 048cab3a76bc4e25f096d4d6ae1ec20cbd242cd8..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/ATANH_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/CLIP_dt_bfloat16.cpp.hip
deleted file mode 100644
index 91c018ae1e4ad41e1d9f1e21376fdf8d82697f90..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/CLIP_dt_float16.cpp.hip
deleted file mode 100644
index 1d06d8d1b7e24602b50298994f6657d5dee1128d..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/CLIP_dt_float32.cpp.hip
deleted file mode 100644
index 346efcbd2b6713709af8eacdf8f012e24e8aa234..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int16.cpp.hip
deleted file mode 100644
index 4394848f2d91b795f6939ccf8c71f99ff54337c5..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int16.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int32.cpp.hip
deleted file mode 100644
index ed51ebd8ea93ad5cd36a12f38141e7e4e967a8cd..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int8.cpp.hip
deleted file mode 100644
index dea82a24f65ec5da7bcdc817f05ec268df720ecf..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_int8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_uint8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/CLIP_dt_uint8.cpp.hip
deleted file mode 100644
index 0d798b15d923ad378a9ed67a4e2f5c68fce38dc8..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/CLIP_dt_uint8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/COSH_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/COSH_dt_bfloat16.cpp.hip
deleted file mode 100644
index 25c504a30ce6958e2dec00b1503891e7b7b7703c..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/COSH_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/COSH_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/COSH_dt_float16.cpp.hip
deleted file mode 100644
index a5e92a13ffccdf5f22d970ea1cbafa0efbf1b484..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/COSH_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/COSH_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/COSH_dt_float32.cpp.hip
deleted file mode 100644
index 2d4c2784a4d0a51abe2d3692724440d52a553a99..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/COSH_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cpp.hip
deleted file mode 100644
index 54b03d90afc2bdede7486a0f8725cb6f3a1c91a2..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cpp.hip
deleted file mode 100644
index eaa54cca410540dabc4a7b6e5e0135a3ec2fdaae..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cpp.hip
deleted file mode 100644
index a000bb105eafbee679bca1bf9404f9b498c7aba2..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_GRAD_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_bfloat16.cpp.hip
deleted file mode 100644
index fdb642b662e6f834b7389ec84277ea7d602fcff4..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_float16.cpp.hip
deleted file mode 100644
index 94e88cd6427c5b3a7b4a30bb3df9bc55f87906dc..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_float32.cpp.hip
deleted file mode 100644
index 8e13dd5352d54d06badfdff5a441efdc821a40a1..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/HSIGMOID_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cpp.hip
deleted file mode 100644
index fa85681cc7d363efc4dd575ea0eadf13c7598907..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_float16.cpp.hip
deleted file mode 100644
index 4992d81da43177d15475c3362e41b12aa1887b27..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_float32.cpp.hip
deleted file mode 100644
index 72e6e3a958a16b30705625c866f87695e0a0baac..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/LOGSIGMOID_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cpp.hip
deleted file mode 100644
index d24ceb4ad6f2fd07b6f9bdbfb63185ff4ab9eb38..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_float16.cpp.hip
deleted file mode 100644
index 4665a27755bb279048dd6cd509b2189230ce9330..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_float32.cpp.hip
deleted file mode 100644
index 023f6fe1af2e3a52c4027332be09037f5032bca8..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_GRAD_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_dt_bfloat16.cpp.hip
deleted file mode 100644
index 2ae7e683c775c2aacab44de03588ced34a075bb6..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_dt_float16.cpp.hip
deleted file mode 100644
index 1e1253caa4d24906ed02537625f7c742406ec9f0..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_dt_float32.cpp.hip
deleted file mode 100644
index 9d3b9676e38029ae699ec55303140f2a5f117f28..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int16.cpp.hip
deleted file mode 100644
index 03846baf8a3d2da8abe530d4c3fe225e23c6139c..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int16.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int32.cpp.hip
deleted file mode 100644
index 41f41670806ad556eab8fb4a2f984d747bdfb692..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int8.cpp.hip
deleted file mode 100644
index 88c2bbb698ec02bd8cd4fa9ed82d2e703c5b208e..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_int8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_uint8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/PRELU_dt_uint8.cpp.hip
deleted file mode 100644
index a9febeadddd443eb523209408545a97be637af0b..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/PRELU_dt_uint8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cpp.hip
deleted file mode 100644
index f2099c6bb68cff5c4633b5649ff075cacf5d1cda..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_float16.cpp.hip
deleted file mode 100644
index b46f1c52df62d2a07fef279206084935d23a25c5..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_float32.cpp.hip
deleted file mode 100644
index 37b7d1baa0e613dc3d01b02fdeeb23e3c9bc99db..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_GRAD_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_dt_bfloat16.cpp.hip
deleted file mode 100644
index bb39a1a8f83758894115bf5f4d2110dc4fc6b9f8..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_dt_float16.cpp.hip
deleted file mode 100644
index f84eac713c0aa82c3d341ac6c65ef7c0ff8e0656..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_dt_float32.cpp.hip
deleted file mode 100644
index fd3fe60f2f66b89082e0594e986a8b0d4a430859..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int16.cpp.hip
deleted file mode 100644
index 66a41225b1810515d4b0bc5fa47787a0adf5d670..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int16.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int32.cpp.hip
deleted file mode 100644
index 1a5eed8257d459878fc8a49096470df6d95ddd60..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int8.cpp.hip
deleted file mode 100644
index f2ecc40acf7105edc7d341f6a781db28baa643ab..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_int8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_uint8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/RELU6_dt_uint8.cpp.hip
deleted file mode 100644
index b13aad0d21d6a3f111954c44dcd1b56bdf014fac..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/RELU6_dt_uint8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SIGN_dt_bfloat16.cpp.hip
deleted file mode 100644
index ee5373bfd14eb062305b570a0574f586f979541c..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SIGN_dt_float16.cpp.hip
deleted file mode 100644
index 69277cd56003e5dc284bfb06fa589e61d96efef9..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SIGN_dt_float32.cpp.hip
deleted file mode 100644
index 718709a16157ae4fdff1f0cf1ce1510d0ee7bb10..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int16.cpp.hip
deleted file mode 100644
index d1cdfbd5e604089f20afedddda535bb414b00dbe..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int16.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int32.cpp.hip
deleted file mode 100644
index 2955aef0b0c79f4fbdd5c1996cdc84dabda2a73d..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int8.cpp.hip
deleted file mode 100644
index e6cc08494c5354658654d3403d1ecd02dd5d6e89..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_int8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_uint8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SIGN_dt_uint8.cpp.hip
deleted file mode 100644
index b39727a4f782cdd847a35533a11b35e15d987821..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SIGN_dt_uint8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SINH_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SINH_dt_bfloat16.cpp.hip
deleted file mode 100644
index 7b0d47402fae6599505e936a8b9fbebfcb4c7de8..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SINH_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SINH_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SINH_dt_float16.cpp.hip
deleted file mode 100644
index 070496fb23061a428a3b897b68998804cfddd9c4..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SINH_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SINH_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SINH_dt_float32.cpp.hip
deleted file mode 100644
index d295a78f4b8551e050acfd5c60cfa09447f78607..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SINH_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cpp.hip
deleted file mode 100644
index 2de73538a4b8cb79e21a508f5d2df571f646305a..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cpp.hip
deleted file mode 100644
index da7dabfd444afd92c39034900b3afe521ef29e9f..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cpp.hip
deleted file mode 100644
index f123cd7e99eac6351fca70fafb068cc7fc881974..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_GRAD_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cpp.hip
deleted file mode 100644
index f86b264d9c03efb8e6d497aa43edbbf94bbfc866..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_float16.cpp.hip
deleted file mode 100644
index 991ba23079dd46974f7f87df14fb9b0bc7dee4cd..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_float32.cpp.hip
deleted file mode 100644
index 55716f22b90b8fd22cc3834aeb28b3ed46a14130..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SOFTPLUS_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SQRT_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQRT_dt_bfloat16.cpp.hip
deleted file mode 100644
index 9c4b92115a0969f676f7c50bdd9990f25eb1bc71..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQRT_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SQRT_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQRT_dt_float16.cpp.hip
deleted file mode 100644
index ad05f946340f8f78d57642ba620b64730059d46e..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQRT_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SQRT_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQRT_dt_float32.cpp.hip
deleted file mode 100644
index 448f9a0caa72c1540db2c7b594963341fbba4340..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQRT_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_bfloat16.cpp.hip
deleted file mode 100644
index aa6d5912f8837696302bcd27ea97185e54d34059..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_float16.cpp.hip
deleted file mode 100644
index 1db3690dbe89f8fea70c13248bf87aa06ea6fdf8..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_float32.cpp.hip
deleted file mode 100644
index da4072aee13f210ba74f53b577fec03eb9062bcf..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int16.cpp.hip
deleted file mode 100644
index 487b9824921f6425c7aaa6c41ecd53963eaaa913..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int16.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int32.cpp.hip
deleted file mode 100644
index ae617e29fefe53c5a5a2f0ba0c2802b9cc8743a9..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int8.cpp.hip
deleted file mode 100644
index 9ebe2f858068f43828bc0b1b58391b03f5b6b8a3..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_int8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_uint8.cpp.hip b/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_uint8.cpp.hip
deleted file mode 100644
index b01411f101f44c8225ab1fbb80abec10455daed1..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/SQUARE_dt_uint8.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/src/rocm/elemwise/kimpl/TAN_dt_bfloat16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/TAN_dt_bfloat16.cpp.hip
deleted file mode 100644
index 9a274f50ad898d5bc8fd0048c10682acfeaa83ec..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/TAN_dt_bfloat16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/TAN_dt_float16.cpp.hip b/dnn/src/rocm/elemwise/kimpl/TAN_dt_float16.cpp.hip
deleted file mode 100644
index 522e02f835a7b11547191ddf7d719603b73a1174..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/TAN_dt_float16.cpp.hip
+++ /dev/null
@@ -1,7 +0,0 @@
-// 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
diff --git a/dnn/src/rocm/elemwise/kimpl/TAN_dt_float32.cpp.hip b/dnn/src/rocm/elemwise/kimpl/TAN_dt_float32.cpp.hip
deleted file mode 100644
index 42a0c3b1971843605554f9c712fb4f91d4251337..0000000000000000000000000000000000000000
--- a/dnn/src/rocm/elemwise/kimpl/TAN_dt_float32.cpp.hip
+++ /dev/null
@@ -1,5 +0,0 @@
-// 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"
diff --git a/dnn/test/common/elemwise.cpp b/dnn/test/common/elemwise.cpp
index 2ef46962e6b757987c6489c6dd1693a9a9c416ab..31f9bcf02f63af3fd8279dbfdae143c409dd91cf 100644
--- a/dnn/test/common/elemwise.cpp
+++ b/dnn/test/common/elemwise.cpp
@@ -744,8 +744,8 @@ DEF_TEST(all_modes) {
     TensorShapeArray shapes;
     UniformFloatRNG default_rng_f32{-100.f, 100.f}, pos_rng_f32{.1f, 1000.f},
             small_pos_rng_f32{.1f, .10f}, small_rng_f32{-3.f, 3.f},
-            abslt1_rng_f32{-0.95f, 0.95f}, uniform_0_2_rng{0.f, 2.f},
-            tanh_rng_f32{-5.f, 5.f}, lt1_rng_f32{1.f, 10.f};
+            abslt1_rng_f32{-1.f, 1.f}, uniform_0_2_rng{0.f, 2.f},
+            tanh_rng_f32{-5.f, 5.f};
     UniformFloatNonZeroRNG nonzero_rng_f32{.1f, 1000.f},
             big_nonzero_rng_f32{100.f, 1000.f};
     UniformIntRNG default_rng_i32{-100, 100}, small_rng_i32{-2, 2},
@@ -786,14 +786,12 @@ DEF_TEST(all_modes) {
         shapes[shapes.size() - 1] = {};
         auto do_run = [&](DType dtype, float eps = 1e-3) {
             // limit value ranges for some modes
-            if (mode == Mode::LOG || mode == Mode::LOG1P || mode == Mode::SQRT) {
+            if (mode == Mode::LOG || mode == Mode::LOG1P) {
                 checker.set_rng(0, &pos_rng_f32);
-            } else if (mode == Mode::POW || mode == Mode::SOFTPLUS_GRAD) {
+            } else if (mode == Mode::POW) {
                 checker.set_rng(0, &small_pos_rng_f32);
                 checker.set_rng(1, &small_rng_f32);
-            } else if (
-                    mode == Mode::EXP || mode == Mode::EXPM1 || mode == Mode::SINH ||
-                    mode == Mode::COSH) {
+            } else if (mode == Mode::EXP || mode == Mode::EXPM1) {
                 checker.set_rng(0, &small_rng_f32);
             } else if (mode == Mode::FAST_TANH) {
                 checker.set_rng(0, &tanh_rng_f32);
@@ -809,10 +807,6 @@ DEF_TEST(all_modes) {
                 checker.set_rng(1, &default_rng_f32);
             } else if (mode == Mode::ERFCINV) {
                 checker.set_rng(0, &uniform_0_2_rng);
-            } else if (mode == Mode::ACOSH_GRAD || mode == Mode::ACOSH) {
-                checker.set_rng(0, &lt1_rng_f32);
-            } else if (mode == Mode::ATANH_GRAD || mode == Mode::ATANH) {
-                checker.set_rng(0, &abslt1_rng_f32);
             } else if (
                     mode == Mode::MOD || mode == Mode::TRUE_DIV ||
                     mode == Mode::FLOOR_DIV) {
diff --git a/imperative/python/megengine/functional/elemwise.py b/imperative/python/megengine/functional/elemwise.py
index b4568c6e744fdafebe0a39ed7bf94ef76df56694..e92cb2d9861893f0e236806da1f8e1ed8c838ad7 100644
--- a/imperative/python/megengine/functional/elemwise.py
+++ b/imperative/python/megengine/functional/elemwise.py
@@ -467,12 +467,12 @@ def log1p(x):
 
 def sqrt(x: Tensor) -> Tensor:
     r"""Element-wise `sqrt`."""
-    return _elwise(x, mode=Elemwise.Mode.SQRT)
+    return x ** 0.5
 
 
 def square(x: Tensor) -> Tensor:
     r"""Element-wise `square`."""
-    return _elwise(x, mode=Elemwise.Mode.SQUARE)
+    return x ** 2
 
 
 def round(x):
@@ -515,7 +515,7 @@ def sin(x):
 
 def tan(x):
     r"""Element-wise `tangent`."""
-    return _elwise(x, mode=Elemwise.Mode.TAN)
+    return sin(x) / cos(x)
 
 
 def acos(x):
@@ -544,12 +544,13 @@ def atan2(y, x):
 
 def cosh(x):
     r"""Element-wise `hyperbolic cosine`."""
-    return _elwise(x, mode=Elemwise.Mode.COSH)
+    return 0.5 * (exp(x) + exp(-x))
 
 
 def sinh(x):
     r"""Element-wise `hyperbolic sine`."""
-    return _elwise(x, mode=Elemwise.Mode.SINH)
+    u = expm1(x)
+    return 0.5 * u / (u + 1) * (u + 2)
 
 
 def tanh(x):
@@ -559,17 +560,17 @@ def tanh(x):
 
 def asinh(x):
     r"""Element-wise `inverse hyperbolic sine`."""
-    return _elwise(x, mode=Elemwise.Mode.ASINH)
+    return log(x + (x ** 2 + 1) ** 0.5)
 
 
 def acosh(x):
     r"""Element-wise `inverse hyperbolic cosine`."""
-    return _elwise(x, mode=Elemwise.Mode.ACOSH)
+    return log(x + (x ** 2 - 1) ** 0.5)
 
 
 def atanh(x):
     r"""Element-wise `inverse hyperbolic tangent`."""
-    return _elwise(x, mode=Elemwise.Mode.ATANH)
+    return log1p(2 * x / (1 - x)) / 2
 
 
 # bit-twiddling functions
@@ -679,7 +680,7 @@ def clip(x: Tensor, lower=None, upper=None) -> Tensor:
     ), "At least one of 'lower' or 'upper' must not be None"
     if lower is not None:
         if upper is not None:
-            return _elwise(x, lower, upper, mode=Elemwise.Mode.CLIP)
+            return minimum(maximum(x, lower), upper)
         else:
             return maximum(x, lower)
     else:
diff --git a/imperative/python/megengine/functional/math.py b/imperative/python/megengine/functional/math.py
index 42fa014e2213ef157f4267ccc05881ab1627974c..978221ac48c180736cfcf87ea5c362e1aea7b5d2 100644
--- a/imperative/python/megengine/functional/math.py
+++ b/imperative/python/megengine/functional/math.py
@@ -6,7 +6,7 @@ from typing import Iterable, Optional, Sequence, Tuple, Union
 from ..core._imperative_rt.core2 import Const, apply
 from ..core._imperative_rt.ops import SubgraphBuilder as _SubgraphBuilder
 from ..core.ops import builtin
-from ..core.tensor.array_method import _elwise, _matmul
+from ..core.tensor.array_method import _matmul
 from ..core.tensor.utils import _normalize_axis
 from ..tensor import Tensor
 from ..utils.deprecation import deprecated_kwargs_default
@@ -86,7 +86,7 @@ def sign(inp: Tensor):
         >>> F.sign(x)
         Tensor([ 1 -1  0], dtype=int32, device=xpux:0)
     """
-    return _elwise(inp, mode=builtin.Elemwise.Mode.SIGN)
+    return (inp > 0).astype(inp.dtype) - (inp < 0).astype(inp.dtype)
 
 
 def sum(
diff --git a/imperative/python/megengine/functional/nn.py b/imperative/python/megengine/functional/nn.py
index 8a42e9065593ba557c18862d8a5a468daf30d875..58a60b4d543c68243419d8e964a24c9c909373b2 100644
--- a/imperative/python/megengine/functional/nn.py
+++ b/imperative/python/megengine/functional/nn.py
@@ -753,9 +753,37 @@ def sigmoid(x):
     return _elwise(x, mode=Elemwise.Mode.SIGMOID)
 
 
+@lru_cache(maxsize=None)
+def _get_hsigmoid_op(dtype=None, device=None):
+    @subgraph_fn(
+        "Hsigmoid",
+        dtype=dtype,
+        device=device,
+        nr_inputs=1,
+        jit_fusion=True,
+        custom_grad=True,
+    )
+    def hsigmoid(inputs, f, c):
+        (inp,) = inputs[0:1]
+        inp = f("+", inp, c(3))
+        max_0 = f("max", inp, c(0))
+        min_6 = f("min", max_0, c(6))
+        oup = f("/", min_6, c(6))
+        (oup_grad,) = yield (oup,)
+        inp_grad = f("/", oup_grad, c(6))
+        inp_grad = f("cond_leq_mov", max_0, c(6), inp_grad)
+        inp_grad = f("cond_leq_mov", c(0), inp, inp_grad)
+        yield (inp_grad,)
+
+    return hsigmoid
+
+
 def hsigmoid(x):
     r"""Element-wise `relu6(x + 3) / 6`."""
-    return _elwise(x, mode=Elemwise.Mode.HSIGMOID)
+    hsigmoid = _get_hsigmoid_op(x.dtype, x.device)
+    (x,) = hsigmoid(x)
+    return x
+    # return relu6(x + 3) / 6
 
 
 def relu(x):
@@ -763,14 +791,95 @@ def relu(x):
     return _elwise(x, mode=Elemwise.Mode.RELU)
 
 
+@lru_cache(maxsize=None)
+def _get_relu6_op(dtype=None, device=None):
+    @subgraph_fn(
+        "ReLU6",
+        dtype=dtype,
+        device=device,
+        nr_inputs=1,
+        jit_fusion=True,
+        custom_grad=True,
+    )
+    def relu6(inputs, f, c):
+        (inp,) = inputs[0:1]
+        max_0 = f("max", inp, c(0))
+        min_6 = f("min", max_0, c(6))
+        oup = min_6
+        (oup_grad,) = yield (oup,)
+        inp_grad = f("cond_leq_mov", max_0, c(6), oup_grad)
+        inp_grad = f("cond_leq_mov", c(0), inp, inp_grad)
+        yield (inp_grad,)
+
+    return relu6
+
+
 def relu6(x):
     r"""Element-wise `min(max(x, 0), 6)`."""
-    return _elwise(x, mode=Elemwise.Mode.RELU6)
+    relu6 = _get_relu6_op(x.dtype, x.device)
+    (x,) = relu6(x)
+    return x
+
+
+@lru_cache(maxsize=None)
+def _get_prelu_op(dtype=None, device=None):
+    @subgraph_fn(
+        "PReLU",
+        dtype=dtype,
+        device=device,
+        nr_inputs=2,
+        jit_fusion=True,
+        custom_grad=True,
+    )
+    def prelu(inputs, f, c):
+        (inp, weight) = inputs[0:2]
+        max_0 = f("max", inp, c(0))
+        min_0 = f("min", inp, c(0))
+        oup = f("fma3", min_0, weight, max_0)
+        (oup_grad,) = yield (oup,)
+        inp_grad_0 = f("cond_leq_mov", c(0), inp, oup_grad)
+        inp_grad_1 = f("*", oup_grad, weight)
+        inp_grad_1 = f("cond_leq_mov", inp, c(0), inp_grad_1)
+        inp_grad = f("+", inp_grad_0, inp_grad_1)
+        weight_grad = f("*", oup_grad, min_0)
+        yield (inp_grad, weight_grad)
+
+    return prelu
+
+
+def prelu(inp: Tensor, weight: Tensor) -> Tensor:
+    r"""Element-wise PReLU function.
+
+    Refer to :class:`~.PReLU` for more information.
+    """
+    prelu = _get_prelu_op(dtype=inp.dtype, device=inp.device)
+    (oup,) = prelu(inp, broadcast_to(weight, inp.shape))
+    return oup
 
 
-def prelu(x, y):
-    r"""Element-wise `max(x, 0) + y * min(x, 0)`."""
-    return _elwise(x, y, mode=Elemwise.Mode.PRELU)
+@lru_cache(maxsize=None)
+def _get_leaky_relu_op(negative_slope, *, dtype=None, device=None):
+    @subgraph_fn(
+        "LeakyReLU",
+        dtype=dtype,
+        device=device,
+        nr_inputs=1,
+        jit_fusion=True,
+        custom_grad=True,
+    )
+    def leakyReLU(inputs, f, c):
+        (inp,) = inputs[0:1]
+        max_0 = f("max", inp, c(0))
+        min_0 = f("min", inp, c(0))
+        oup = f("+", max_0, f("*", min_0, c(negative_slope)))
+        (oup_grad,) = yield (oup,)
+        inp_grad_0 = f("cond_leq_mov", c(0), inp, oup_grad)
+        inp_grad_1 = f("*", oup_grad, c(negative_slope))
+        inp_grad_1 = f("cond_leq_mov", inp, c(0), inp_grad_1)
+        inp_grad = f("+", inp_grad_0, inp_grad_1)
+        yield (inp_grad,)
+
+    return leakyReLU
 
 
 def leaky_relu(inp: Tensor, negative_slope: float = 0.01) -> Tensor:
@@ -778,7 +887,9 @@ def leaky_relu(inp: Tensor, negative_slope: float = 0.01) -> Tensor:
 
     Refer to :class:`~.LeakyReLU` for more information.
     """
-    return _elwise(inp, negative_slope, mode=Elemwise.Mode.PRELU)
+    leakyReLU = _get_leaky_relu_op(negative_slope, dtype=inp.dtype, device=inp.device)
+    (oup,) = leakyReLU(inp)
+    return oup
 
 
 def silu(x):
@@ -797,6 +908,36 @@ def gelu(x):
     return _elwise(x, mode=Elemwise.Mode.GELU)
 
 
+@lru_cache(maxsize=None)
+def _get_softplus_op(dtype=None, device=None):
+    @subgraph_fn(
+        "Softplus",
+        dtype=dtype,
+        device=device,
+        nr_inputs=1,
+        jit_fusion=True,
+        custom_grad=True,
+    )
+    def softplus(inputs, f, c):
+        (inp,) = inputs[0:1]
+        neg_abs = f("-", f("abs", inp))
+        exp = f("exp", neg_abs)
+        oup0 = f("log1p", exp)
+        oup1 = f("relu", inp)
+        oup = f("+", oup0, oup1)
+        (oup_grad,) = yield (oup,)
+        inp_grad_0 = f("switch_gt0", oup1, oup_grad)
+        inp_grad_1 = oup_grad
+        inp_grad_1 = f("/", oup_grad, f("+", exp, c(1)))
+        inp_grad_1 = f("*", inp_grad_1, exp)
+        inp_grad_1 = f("-", inp_grad_1)
+        inp_grad_1 = f("abs_grad", inp, inp_grad_1)
+        inp_grad = f("+", inp_grad_0, inp_grad_1)
+        yield (inp_grad,)
+
+    return softplus
+
+
 def softplus(inp: Tensor) -> Tensor:
     r"""Applies the element-wise function:
 
@@ -819,7 +960,9 @@ def softplus(inp: Tensor) -> Tensor:
         >>> y.numpy().round(decimals=4)
         array([0.0486, 0.1269, 0.3133, 0.6931, 1.3133, 2.1269], dtype=float32)
     """
-    return _elwise(inp, mode=Elemwise.Mode.SOFTPLUS)
+    softplus = _get_softplus_op(inp.dtype, inp.device)
+    (oup,) = softplus(inp)
+    return oup
 
 
 def logsoftmax(inp: Tensor, axis: Union[int, Sequence[int]]) -> Tensor:
@@ -848,6 +991,39 @@ def logsoftmax(inp: Tensor, axis: Union[int, Sequence[int]]) -> Tensor:
     return inp - logsumexp(inp, axis, keepdims=True)
 
 
+@lru_cache(maxsize=None)
+def _get_logsigmoid_op(dtype=None, device=None):
+    @subgraph_fn(
+        "LogSigmoid",
+        dtype=dtype,
+        device=device,
+        nr_inputs=1,
+        jit_fusion=True,
+        custom_grad=True,
+    )
+    def logsigmoid(inputs, f, c):
+        (inp,) = inputs[0:1]
+        neg_abs = f("-", f("abs", inp))
+        exp = f("exp", neg_abs)
+        oup0 = f("log1p", exp)
+        oup1 = f("relu", f("-", inp))
+        oup = f("+", oup0, oup1)
+        oup = f("-", oup)
+        (oup_grad,) = yield (oup,)
+        oup_grad = f("-", oup_grad)
+        inp_grad_0 = f("switch_gt0", oup1, oup_grad)
+        inp_grad_0 = f("-", inp_grad_0)
+        inp_grad_1 = oup_grad
+        inp_grad_1 = f("/", inp_grad_1, f("+", exp, c(1)))
+        inp_grad_1 = f("*", inp_grad_1, exp)
+        inp_grad_1 = f("-", inp_grad_1)
+        inp_grad_1 = f("abs_grad", inp, inp_grad_1)
+        inp_grad = f("+", inp_grad_0, inp_grad_1)
+        yield (inp_grad,)
+
+    return logsigmoid
+
+
 def logsigmoid(inp: Tensor) -> Tensor:
     r"""Applies the element-wise function:
 
@@ -865,7 +1041,9 @@ def logsigmoid(inp: Tensor) -> Tensor:
         array([-5.0067, -4.0182, -3.0486, -2.1269, -1.3133, -0.6931, -0.3133,
                -0.1269, -0.0486, -0.0181], dtype=float32)
     """
-    return _elwise(inp, mode=Elemwise.Mode.LOGSIGMOID)
+    logsigmoid = _get_logsigmoid_op(inp.dtype, inp.device)
+    (oup,) = logsigmoid(inp)
+    return oup
 
 
 def logsumexp(
diff --git a/imperative/src/impl/transformations/dtype_promote.cpp b/imperative/src/impl/transformations/dtype_promote.cpp
index 5e7743d77c097dc42ef41cd65bcb1a7258f4e0cd..58de880a7364ff5785c6c6b87f3e026d2fe9f6d0 100644
--- a/imperative/src/impl/transformations/dtype_promote.cpp
+++ b/imperative/src/impl/transformations/dtype_promote.cpp
@@ -116,17 +116,12 @@ ValueRefList elemwise_rule(const OpDef& op, Span<ValueRef> inputs) {
     }
 
     static std::unordered_set<Elemwise::Mode> cast_case1 = {
-            Elemwise::Mode::TRUE_DIV,   Elemwise::Mode::EXP,
-            Elemwise::Mode::POW,        Elemwise::Mode::LOG,
-            Elemwise::Mode::EXPM1,      Elemwise::Mode::LOG1P,
-            Elemwise::Mode::ACOS,       Elemwise::Mode::ASIN,
-            Elemwise::Mode::ATAN2,      Elemwise::Mode::COS,
-            Elemwise::Mode::SIN,        Elemwise::Mode::LOG_SUM_EXP,
-            Elemwise::Mode::TAN,        Elemwise::Mode::ASINH,
-            Elemwise::Mode::ACOSH,      Elemwise::Mode::ATANH,
-            Elemwise::Mode::SINH,       Elemwise::Mode::COSH,
-            Elemwise::Mode::SOFTPLUS,   Elemwise::Mode::HSIGMOID,
-            Elemwise::Mode::LOGSIGMOID, Elemwise::Mode::SQRT,
+            Elemwise::Mode::TRUE_DIV, Elemwise::Mode::EXP,
+            Elemwise::Mode::POW,      Elemwise::Mode::LOG,
+            Elemwise::Mode::EXPM1,    Elemwise::Mode::LOG1P,
+            Elemwise::Mode::ACOS,     Elemwise::Mode::ASIN,
+            Elemwise::Mode::ATAN2,    Elemwise::Mode::COS,
+            Elemwise::Mode::SIN,      Elemwise::Mode::LOG_SUM_EXP,
     };
 
     static std::unordered_set<Elemwise::Mode> cast_case2 = {
diff --git a/src/jit/impl/ast_c.cpp b/src/jit/impl/ast_c.cpp
index eb1dc794bacb97e6a70820e8d96e951ef963f89f..edce00074fd3a6a61e0b44709cb285d2cafb0286 100644
--- a/src/jit/impl/ast_c.cpp
+++ b/src/jit/impl/ast_c.cpp
@@ -133,7 +133,7 @@ const ElemGeneratorMap& ast_c::elem_opr_generator() {
                                    0.f}) /
                           6.f),
     };
-    mgb_assert(map.size() + 41 == opr::Elemwise::Param::MODE_NR_MEMBER);
+    mgb_assert(map.size() + 19 == opr::Elemwise::Param::MODE_NR_MEMBER);
     // unimplemented modes: SHL, SHR, FAST_TANH, FAST_TANH_GRAD, ROUND, RMULH,
     // ERFINV, ERFCINV, NOT, AND, OR, XOR, NEQ, ISNAN, ISINF
     return map;
diff --git a/src/opr/impl/basic_arith.cpp b/src/opr/impl/basic_arith.cpp
index 9924f74718f6aaceeb9afafda081acbb3b4131c1..8111b4e0585a049ce5e3f46822709e4f052c4e67 100644
--- a/src/opr/impl/basic_arith.cpp
+++ b/src/opr/impl/basic_arith.cpp
@@ -543,34 +543,6 @@ MGB_IMPL_OPR_GRAD(Elemwise) {
             RET(EL2(SILU_GRAD, i0, og));
         case Mode::GELU:
             RET(EL2(GELU_GRAD, i0, og));
-        case Mode::SINH:
-            RET(EL1(COSH, i0) * og);
-        case Mode::COSH:
-            RET(EL1(SINH, i0) * og);
-        case Mode::ASINH:
-            RET(EL2(ASINH_GRAD, i0, og));
-        case Mode::ACOSH:
-            RET(EL2(ACOSH_GRAD, i0, og));
-        case Mode::ATANH:
-            RET(EL2(ATANH_GRAD, i0, og));
-        case Mode::TAN: {
-            auto two = i0.make_scalar_dt(2);
-            RET(og / (EL2(POW, EL1(COS, i0), two)));
-        }
-        case Mode::RELU6:
-            RET(EL2(RELU6_GRAD, i0, og));
-        case Mode::SOFTPLUS:
-            RET(EL2(SOFTPLUS_GRAD, i0, og));
-        case Mode::HSIGMOID:
-            RET(EL2(HSIGMOID_GRAD, i0, og));
-        case Mode::LOGSIGMOID:
-            RET(EL2(SOFTPLUS_GRAD, EL1(NEGATE, i0), og));
-        case Mode::SQRT:
-            RET(og / EL1(SQRT, i0) / 2);
-        case Mode::SQUARE:
-            RET(og * 2 * i0);
-        case Mode::SIGN:
-            RET(i0.make_scalar_dt(0).broadcast(i0.symshape()));
 
         // binary
         case Mode::ABS_GRAD:
@@ -645,11 +617,6 @@ MGB_IMPL_OPR_GRAD(Elemwise) {
         case Mode::XOR:
         case Mode::AND:
             return nullptr;
-        case Mode::PRELU:
-            if (wrt_idx == 0) {
-                RET(EL3(PRELU_GRAD, i0, og, i1));
-            }
-            RET(EL2(SWITCH_GT0, -i0, og * i0));
 
         // ternary
         case Mode::COND_LEQ_MOV:
@@ -660,15 +627,6 @@ MGB_IMPL_OPR_GRAD(Elemwise) {
             if (wrt_idx <= 1)
                 return nullptr;
             RET(EL3(COND_LT_MOV, i0, i1, og));
-        case Mode::CLIP:
-            if (wrt_idx == 0) {
-                RET(EL3(COND_LEQ_MOV, i1, i0, EL3(COND_LEQ_MOV, i0, i2, og)));
-            }
-            if (wrt_idx == 1) {
-                RET(EL3(COND_LEQ_MOV, i0, i1, og));
-            }
-            RET(EL3(COND_LEQ_MOV, i2, i0, og));
-
         // fuse oprs
         case Mode::FUSE_MUL_ADD3:
             if (wrt_idx < 2) {
diff --git a/src/opr/test/basic_arith/elemwise.cpp b/src/opr/test/basic_arith/elemwise.cpp
index 89ae9ca3b60e59877cd8c04d497745c5c7257093..b4bab90d59305c16e0d5aaff1e80ed335b6692ed 100644
--- a/src/opr/test/basic_arith/elemwise.cpp
+++ b/src/opr/test/basic_arith/elemwise.cpp
@@ -349,99 +349,6 @@ struct CheckerConfig<H_SWISH> : public CheckerConfig<void> {};
 template <>
 struct CheckerConfig<H_SWISH_GRAD> : public NoGradCheckerConfig {};
 
-template <>
-struct CheckerConfig<TAN> : public NoGradCheckerConfig {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(-1.2, 1.2);
-    }
-};
-template <>
-struct CheckerConfig<SINH> : public CheckerConfig<void> {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(-5, 5);
-    }
-    template <class Opt>
-    static void update_opt(Opt& opt) {
-        opt.numdiff_eps = 1e-2;
-        opt.numdiff_max_err = 0.1;
-    }
-};
-template <>
-struct CheckerConfig<COSH> : public CheckerConfig<SINH> {};
-template <>
-struct CheckerConfig<ASINH> : public CheckerConfig<void> {
-    template <class Opt>
-    static void update_opt(Opt& opt) {
-        opt.numdiff_eps = 1e-2;
-        opt.numdiff_max_err = 0.1;
-    }
-};
-template <>
-struct CheckerConfig<ACOSH> : public CheckerConfig<ASINH> {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(1.05, 5);
-    }
-};
-template <>
-struct CheckerConfig<ATANH> : public CheckerConfig<ASINH> {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(-0.95, 0.95);
-    }
-};
-template <>
-struct CheckerConfig<SOFTPLUS> : public CheckerConfig<void> {};
-template <>
-struct CheckerConfig<LOGSIGMOID> : public CheckerConfig<void> {};
-template <>
-struct CheckerConfig<SQUARE> : public CheckerConfig<void> {};
-template <>
-struct CheckerConfig<SQRT> : public CheckerConfig<void> {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(0.05, 5);
-    }
-    template <class Opt>
-    static void update_opt(Opt& opt) {
-        opt.numdiff_eps = 1e-2;
-        opt.numdiff_max_err = 0.1;
-    }
-};
-template <>
-struct CheckerConfig<RELU6> : public CheckerConfig<void> {
-    template <typename ctype, class Checker>
-    static void do_update_checker(Checker& checker) {
-        auto icoord = [](const typename Checker::NumInpArray& inp) {
-            auto p0 = inp[0]->template ptr<ctype>();
-            for (size_t i = 0, it = inp[0]->shape().total_nr_elems(); i < it; ++i) {
-                if (std::abs(p0[i]) < 1) {
-                    p0[i] += 2;
-                } else if (std::abs(p0[i] - 6) < 1) {
-                    p0[i] += 2;
-                }
-            }
-        };
-        checker.set_input_coordinator(icoord);
-    }
-    template <class Checker>
-    static void update_checker(Checker& checker) {
-        using ctype = typename Checker::ctype;
-        return do_update_checker<ctype>(checker);
-    }
-};
-template <>
-struct CheckerConfig<HSIGMOID> : public CheckerConfig<void> {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(-2.95, 2.95);
-    }
-};
-template <>
-struct CheckerConfig<SIGN> : public NoZeroCheckerConfig<0> {};
-
 /* ======================= binary config ======================= */
 template <bool for_mod>
 struct BinaryInputMinGap : public CheckerConfig<void> {
@@ -660,85 +567,13 @@ template <>
 struct CheckerConfig<SILU_GRAD> : public NoGradCheckerConfig {};
 template <>
 struct CheckerConfig<GELU_GRAD> : public NoGradCheckerConfig {};
-template <>
-struct CheckerConfig<PRELU> : public NoZeroCheckerConfig<0> {};
 
-template <>
-struct CheckerConfig<ASINH_GRAD> : public NoGradCheckerConfig {};
-template <>
-struct CheckerConfig<ACOSH_GRAD> : public NoGradCheckerConfig {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(1.05, 5);
-    }
-};
-template <>
-struct CheckerConfig<ATANH_GRAD> : public NoGradCheckerConfig {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(-0.95, 0.95);
-    }
-};
-template <>
-struct CheckerConfig<RELU6_GRAD> : public NoGradCheckerConfig {};
-template <>
-struct CheckerConfig<SOFTPLUS_GRAD> : public NoGradCheckerConfig {};
-template <>
-struct CheckerConfig<HSIGMOID_GRAD> : public NoGradCheckerConfig {
-    template <typename ctype>
-    static InputGenerator get_inp_gen(size_t) {
-        return get_inp_gen_f32_range<ctype>(-2.95, 2.95);
-    }
-};
 /* ======================= ternary config ======================= */
 template <>
 struct CheckerConfig<COND_LEQ_MOV> : public BinaryInputMinGap<false> {};
 template <>
 struct CheckerConfig<COND_LT_MOV> : public BinaryInputMinGap<false> {};
-struct CheckerConfig<PRELU_GRAD> : public NoGradCheckerConfig {};
-template <>
-struct CheckerConfig<CLIP> : public CheckerConfig<void> {
-    template <typename ctype, class Checker>
-    static void do_update_checker(Checker& checker) {
-        auto icoord = [](const typename Checker::NumInpArray& inp) {
-            auto p0 = inp[0]->template ptr<ctype>(), p1 = inp[1]->template ptr<ctype>(),
-                 p2 = inp[2]->template ptr<ctype>();
-            for (size_t i = 0, it = inp[0]->shape().total_nr_elems(); i < it; ++i) {
-                if (p1[i] > p2[i]) {
-                    std::swap(p1[i], p2[i]);
-                }
-                if (p1[i] + 1 > p2[i]) {
-                    p2[i] = p1[i] + 1;
-                }
-                if (std::abs(p1[i] - p0[i]) < 1) {
-                    if (p1[i] < p0[i])
-                        p0[i] += 1;
-                    else
-                        p0[i] -= 1;
-                }
-                if (std::abs(p2[i] - p0[i]) < 1) {
-                    if (p2[i] < p0[i])
-                        p0[i] += 1;
-                    else
-                        p0[i] -= 1;
-                }
-            }
-        };
-        checker.set_input_coordinator(icoord);
-    }
-
-    template <class Checker>
-    static void update_checker(Checker& checker) {
-        using ctype = typename Checker::ctype;
-        return do_update_checker<ctype>(checker);
-    }
 
-    template <class Opt>
-    static void update_opt(Opt& opt) {
-        opt.numdiff_eps = 1e-3;
-        opt.numdiff_max_err = 0.1;
-    }
-};
 /* ======================= test runner ======================= */
 namespace detail {
 template <typename dtype, class Trait>
diff --git a/src/opr/test/basic_arith/elemwise_binary_trait_def.inl b/src/opr/test/basic_arith/elemwise_binary_trait_def.inl
index 1ed742db53dfdb7c5afa28a70c91de33f07a8927..0663986a88b6f8fe8ab369715faf34fcf137879c 100644
--- a/src/opr/test/basic_arith/elemwise_binary_trait_def.inl
+++ b/src/opr/test/basic_arith/elemwise_binary_trait_def.inl
@@ -41,7 +41,6 @@ DEF_TRAIT(SWITCH_GT0, x > 0 ? y : 0)
 DEF_TRAIT(TANH_GRAD, (1 - x * x) * y)
 
 DEF_TRAIT(FUSE_ADD_RELU, std::max<ctype>(x + y, 0))
-DEF_TRAIT(PRELU, (x > 0) ? x : (x* y))
 #undef _ALLOW_INT
 
 #define _ALLOW_INT false
@@ -58,12 +57,6 @@ DEF_TRAIT(
         SILU_GRAD, y*(1 + std::exp(-x) + x * std::exp(-x)) / (1 + std::exp(-x)) /
                            (1 + std::exp(-x)))
 DEF_TRAIT(GELU_GRAD, do_gelu_grad(x, y))
-DEF_TRAIT(ASINH_GRAD, y / std::sqrt(x * x + 1))
-DEF_TRAIT(ACOSH_GRAD, y / std::sqrt(x * x - 1))
-DEF_TRAIT(ATANH_GRAD, y / (1 - x * x))
-DEF_TRAIT(SOFTPLUS_GRAD, y* std::exp(x) / (1.f + std::exp(x)))
-DEF_TRAIT(RELU6_GRAD, x <= 0.f ? 0.f : (x >= 6.f ? 0.f : y))
-DEF_TRAIT(HSIGMOID_GRAD, x <= -3.f ? 0.f : (x >= 3.f ? 0.f : (y / 6.f)))
 
 #undef _ALLOW_INT
 #undef _ALLOW_FLOAT
diff --git a/src/opr/test/basic_arith/elemwise_ternary_trait_def.inl b/src/opr/test/basic_arith/elemwise_ternary_trait_def.inl
index a9bf8af143e43911567cad2d68735f6d49cc87bc..e9e5cd8ecb2605ba8869f55326364a7c12f56e7a 100644
--- a/src/opr/test/basic_arith/elemwise_ternary_trait_def.inl
+++ b/src/opr/test/basic_arith/elemwise_ternary_trait_def.inl
@@ -15,10 +15,6 @@
 DEF_TRAIT(COND_LEQ_MOV, x <= y ? z : 0)
 DEF_TRAIT(COND_LT_MOV, x < y ? z : 0)
 DEF_TRAIT(FUSE_MUL_ADD3, x* y + z)
-DEF_TRAIT(CLIP, x < y ? y : (x < z ? x : z))
-#undef _ALLOW_INT
-#define _ALLOW_INT false
-DEF_TRAIT(PRELU_GRAD, x > 0 ? y : (y * z))
 #undef _ALLOW_INT
 #undef _ALLOW_FLOAT
 
diff --git a/src/opr/test/basic_arith/elemwise_unary_trait_def.inl b/src/opr/test/basic_arith/elemwise_unary_trait_def.inl
index edacc03581adc5d62b6c3486afeb8e3620caa256..a6e64ecd3e75af5e505e0a22eea96419373dff8d 100644
--- a/src/opr/test/basic_arith/elemwise_unary_trait_def.inl
+++ b/src/opr/test/basic_arith/elemwise_unary_trait_def.inl
@@ -22,9 +22,6 @@ DEF_TRAIT(NOT, !x)
 DEF_TRAIT(ABS, std::abs(x))
 DEF_TRAIT(NEGATE, -x)
 DEF_TRAIT(RELU, std::max<ctype>(x, 0))
-DEF_TRAIT(RELU6, std::min<ctype>(std::max<ctype>(x, 0), 6))
-DEF_TRAIT(SQUARE, x* x)
-DEF_TRAIT(SIGN, x < 0 ? -1 : (x > 0 ? 1 : 0))
 #undef _ALLOW_INT
 
 #define _ALLOW_INT false
@@ -49,16 +46,6 @@ DEF_TRAIT(ERFCINV, do_erfcinv(x))
 DEF_TRAIT(H_SWISH, do_h_swish(x))
 DEF_TRAIT(SILU, x / (1 + std::exp(-x)))
 DEF_TRAIT(GELU, x*(0.5f * (1.f + std::erf(x / std::sqrt(2.f)))))
-DEF_TRAIT(SINH, std::sinh(x))
-DEF_TRAIT(COSH, std::cosh(x))
-DEF_TRAIT(ASINH, std::asinh(x))
-DEF_TRAIT(ACOSH, std::acosh(x))
-DEF_TRAIT(ATANH, std::atanh(x))
-DEF_TRAIT(TAN, std::tan(x))
-DEF_TRAIT(SOFTPLUS, std::log1p(std::exp(-std::abs(x))) + std::max<ctype>(x, 0))
-DEF_TRAIT(HSIGMOID, x <= -3.f ? 0.f : (x >= 3.f ? 1.f : ((x + 3.f) / 6.f)))
-DEF_TRAIT(SQRT, std::sqrt(x))
-DEF_TRAIT(LOGSIGMOID, -std::log1p(std::exp(-std::abs(x))) - std::max<ctype>(-x, 0))
 #undef _ALLOW_INT
 
 #undef _ALLOW_FLOAT