mindspore-ci-bot
5 years ago
committed by
Gitee
commit
aeffccb7f8
@ -1,126 +0,0 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""learning rate generator"""
|
||||
import math
|
||||
|
||||
import numpy as np
|
||||
|
||||
|
||||
def linear_warmup_lr(current_step, warmup_steps, base_lr, init_lr):
|
||||
"""linear_warmup_lr"""
|
||||
lr_inc = (float(base_lr) - float(init_lr)) / float(warmup_steps)
|
||||
lr = float(init_lr) + lr_inc * current_step
|
||||
return lr
|
||||
|
||||
|
||||
def cosine_annealing_lr(lr, steps_per_epoch, warmup_epochs, max_epoch, T_max, eta_min=0, num_periods=0.5):
|
||||
"""linear_warmup_lr"""
|
||||
base_lr = lr
|
||||
warmup_init_lr = 0
|
||||
total_steps = int(max_epoch * steps_per_epoch)
|
||||
warmup_steps = int(warmup_epochs * steps_per_epoch)
|
||||
decay_steps = total_steps - warmup_steps
|
||||
lr_each_step = []
|
||||
for i in range(total_steps):
|
||||
if i < warmup_steps:
|
||||
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
||||
else:
|
||||
# linear_decay = (total_steps - i) / decay_steps
|
||||
cosine_decay = 0.5 * (1 + math.cos(math.pi * i / decay_steps))
|
||||
decayed = cosine_decay
|
||||
lr = base_lr * decayed
|
||||
lr_each_step.append(lr)
|
||||
return np.array(lr_each_step).astype(np.float32)
|
||||
|
||||
|
||||
def warmup_cosine_annealing_lr(lr, steps_per_epoch, warmup_epochs, max_epoch, T_max, eta_min=0, num_periods=0.5):
|
||||
"""warmup_cosine_annealing_lr"""
|
||||
base_lr = lr
|
||||
warmup_init_lr = 0
|
||||
total_steps = int(max_epoch * steps_per_epoch * 0.99)
|
||||
warmup_steps = int(warmup_epochs * steps_per_epoch)
|
||||
decay_steps = total_steps - warmup_steps
|
||||
lr_each_step = []
|
||||
for i in range(total_steps):
|
||||
if i < warmup_steps:
|
||||
lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
|
||||
else:
|
||||
linear_decay = (total_steps - i) / decay_steps
|
||||
cosine_decay = 0.5 * (1 + math.cos(math.pi * 2 * num_periods * i / decay_steps))
|
||||
decayed = linear_decay * cosine_decay
|
||||
lr = base_lr * decayed + 0.000005
|
||||
lr_each_step.append(lr)
|
||||
return np.array(lr_each_step).astype(np.float32)
|
||||
|
||||
|
||||
def get_lr(global_step, lr_init, lr_end, lr_max, warmup_epochs, total_epochs, steps_per_epoch, lr_decay_mode):
|
||||
"""
|
||||
generate learning rate array
|
||||
|
||||
Args:
|
||||
global_step(int): total steps of the training
|
||||
lr_init(float): init learning rate
|
||||
lr_end(float): end learning rate
|
||||
lr_max(float): max learning rate
|
||||
warmup_epochs(int): number of warmup epochs
|
||||
total_epochs(int): total epoch of training
|
||||
steps_per_epoch(int): steps of one epoch
|
||||
lr_decay_mode(string): learning rate decay mode, including steps, poly or default
|
||||
|
||||
Returns:
|
||||
np.array, learning rate array
|
||||
"""
|
||||
lr_each_step = []
|
||||
total_steps = steps_per_epoch * total_epochs
|
||||
warmup_steps = steps_per_epoch * warmup_epochs
|
||||
if lr_decay_mode == 'steps':
|
||||
decay_epoch_index = [0.3 * total_steps, 0.6 * total_steps, 0.8 * total_steps]
|
||||
for i in range(total_steps):
|
||||
if i < decay_epoch_index[0]:
|
||||
lr = lr_max
|
||||
elif i < decay_epoch_index[1]:
|
||||
lr = lr_max * 0.1
|
||||
elif i < decay_epoch_index[2]:
|
||||
lr = lr_max * 0.01
|
||||
else:
|
||||
lr = lr_max * 0.001
|
||||
lr_each_step.append(lr)
|
||||
elif lr_decay_mode == 'poly':
|
||||
if warmup_steps != 0:
|
||||
inc_each_step = (float(lr_max) - float(lr_init)) / float(warmup_steps)
|
||||
else:
|
||||
inc_each_step = 0
|
||||
for i in range(total_steps):
|
||||
if i < warmup_steps:
|
||||
lr = float(lr_init) + inc_each_step * float(i)
|
||||
else:
|
||||
base = (1.0 - (float(i) - float(warmup_steps)) / (float(total_steps) - float(warmup_steps)))
|
||||
lr = float(lr_max) * base * base
|
||||
if lr < 0.0:
|
||||
lr = 0.0
|
||||
lr_each_step.append(lr)
|
||||
else:
|
||||
for i in range(total_steps):
|
||||
if i < warmup_steps:
|
||||
lr = lr_init + (lr_max - lr_init) * i / warmup_steps
|
||||
else:
|
||||
lr = lr_max - (lr_max - lr_end) * (i - warmup_steps) / (total_steps - warmup_steps)
|
||||
lr_each_step.append(lr)
|
||||
|
||||
current_step = global_step
|
||||
lr_each_step = np.array(lr_each_step).astype(np.float32)
|
||||
learning_rate = lr_each_step[current_step:]
|
||||
|
||||
return learning_rate
|
||||
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,76 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""batch_matmul_impl"""
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusBatchMatMul",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "batchmatmul.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusBatchMatMul",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 1,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x2",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
def CusBatchMatMul(input_x1, input_x2, output, transpose_a=False, transpose_b=True, kernel_name="batchmatmul"):
|
||||
"""CusBatchMatMul"""
|
||||
return
|
||||
@ -0,0 +1,64 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""CusCholeskyTrsm"""
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusCholeskyTrsm",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "choleskytrsm.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusCholeskyTrsm",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
def CusCholeskyTrsm(input_x, output, kernel_name):
|
||||
"""CusCholeskyTrsm"""
|
||||
return
|
||||
@ -0,0 +1,69 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""CusFusedAbsMax1"""
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusFusedAbsMax1",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "fusedabsmax1.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusFusedAbsMax1",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
{
|
||||
"name": "origin_shape",
|
||||
"param_type": "required",
|
||||
"type": "listInt",
|
||||
"value": "all"
|
||||
}
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
def CusFusedAbsMax1(input_x, output, origin_shape=None, kernel_name="fused_abs_max1"):
|
||||
"""CusFusedAbsMax1"""
|
||||
return
|
||||
@ -0,0 +1,87 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""CusImg2ColNC1HWC0"""
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusImg2ColNC1HWC0",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "img2colnc1hwc0.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusImg2ColNC1HWC0",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
{
|
||||
"name": "ksizes",
|
||||
"param_type": "required",
|
||||
"type": "listInt",
|
||||
"value": "all"
|
||||
},
|
||||
{
|
||||
"name": "strides",
|
||||
"param_type": "required",
|
||||
"type": "listInt",
|
||||
"value": "all"
|
||||
},
|
||||
{
|
||||
"name": "dilates",
|
||||
"param_type": "required",
|
||||
"type": "listInt",
|
||||
"value": "all"
|
||||
},
|
||||
{
|
||||
"name": "padding",
|
||||
"param_type": "required",
|
||||
"type": "str",
|
||||
"value": "all"
|
||||
}
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"NC1HWC0"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"FRACTAL_NZ"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
def CusImg2ColNC1HWC0(input_x, output, ksizes, strides, dilates, padding, kernel_name="img2col"):
|
||||
"""CusImg2ColNC1HWC0"""
|
||||
return
|
||||
@ -0,0 +1,101 @@
|
||||
# -*- coding:utf-8 -*-
|
||||
"""
|
||||
copyright 2020 Huawei Technologies Co., Ltd
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License == distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
|
||||
matmul
|
||||
"""
|
||||
from __future__ import absolute_import
|
||||
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
from topi.cce import util
|
||||
|
||||
# General limitation of the size for input shape: 2**31
|
||||
SHAPE_SIZE_LIMIT = 2147483648
|
||||
NoneType = type(None)
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusMatMulCubeDenseLeft",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "matmulcubedenseleft.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusMatMulCubeDenseLeft",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 1,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"FRACTAL_NZ"
|
||||
],
|
||||
"name": "x2",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 2,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x3",
|
||||
"need_compile": false,
|
||||
"param_type": "optional",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"FRACTAL_NZ"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
@util.check_input_type(dict, dict, (dict, NoneType), dict, bool, bool, str)
|
||||
def CusMatMulCubeDenseLeft(input_x1, input_x2, bias=None, output_y={}, trans_a=False, trans_b=False,
|
||||
kernel_name="matmulcube"):
|
||||
"""CusMatMulCubeDenseLeft"""
|
||||
return
|
||||
@ -0,0 +1,102 @@
|
||||
# -*- coding:utf-8 -*-
|
||||
"""
|
||||
copyright 2020 Huawei Technologies Co., Ltd
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License == distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
|
||||
matmul
|
||||
"""
|
||||
from __future__ import absolute_import
|
||||
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
from topi.cce import util
|
||||
|
||||
# General limitation of the size for input shape: 2**31
|
||||
SHAPE_SIZE_LIMIT = 2147483648
|
||||
NoneType = type(None)
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusMatMulCubeFraczLeftCast",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "matmulcubefraczleftcast.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusMatMulCubeFraczLeftCast",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 1,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"FracZ"
|
||||
],
|
||||
"name": "x2",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 2,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x3",
|
||||
"need_compile": false,
|
||||
"param_type": "optional",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"FracZ"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
# pylint: disable=locally-disabled,too-many-arguments, too-many-locals, too-many-statements
|
||||
@util.check_input_type(dict, dict, (dict, NoneType), dict, bool, bool, str)
|
||||
def CusMatMulCubeFraczLeftCast(input_x1, input_x2, bias=None, output_y={}, trans_a=False, trans_b=False,
|
||||
kernel_name="CusMatMulCubeFraczLeftCast"):
|
||||
"""CusMatMulCubeFraczLeftCast"""
|
||||
return
|
||||
@ -0,0 +1,113 @@
|
||||
#!/usr/bin/env python
|
||||
# -*- coding:utf-8 -*-
|
||||
"""
|
||||
copyright 2020 Huawei Technologies Co., Ltd
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License == distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
|
||||
matmul
|
||||
"""
|
||||
from __future__ import absolute_import
|
||||
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
|
||||
# General limitation of the size for input shape: 2**31
|
||||
SHAPE_SIZE_LIMIT = 2147483648
|
||||
NoneType = type(None)
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusMatMulCubeFraczRightMul",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "matmulcubefraczrightmul.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusMatMulCubeFraczRightMul",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"FracZ"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 1,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x2",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 2,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x3",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 3,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x4",
|
||||
"need_compile": false,
|
||||
"param_type": "optional",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"FracZ"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
def CusMatMulCubeFraczRightMul(input_x1, input_x2, input_x3, bias=None, output_y={}, trans_a=False, trans_b=False,
|
||||
kernel_name="matmulcube"):
|
||||
"""CusMatMulCubeFraczRightMul"""
|
||||
return
|
||||
@ -0,0 +1,114 @@
|
||||
#!/usr/bin/env python
|
||||
# -*- coding:utf-8 -*-
|
||||
"""
|
||||
copyright 2020 Huawei Technologies Co., Ltd
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License == distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
|
||||
matmul
|
||||
"""
|
||||
from __future__ import absolute_import
|
||||
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
from topi.cce import util
|
||||
|
||||
# General limitation of the size for input shape: 2**31
|
||||
SHAPE_SIZE_LIMIT = 2147483648
|
||||
NoneType = type(None)
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusMatMulCube",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "matmulcube.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusMatMulCube",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
{
|
||||
"name": "transpose_a",
|
||||
"param_type": "required",
|
||||
"type": "bool",
|
||||
"value": "all"
|
||||
},
|
||||
{
|
||||
"name": "transpose_b",
|
||||
"param_type": "required",
|
||||
"type": "bool",
|
||||
"value": "all"
|
||||
}
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"FRACTAL_NZ"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 1,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"FRACTAL_NZ"
|
||||
],
|
||||
"name": "x2",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
},
|
||||
{
|
||||
"index": 2,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x3",
|
||||
"need_compile": false,
|
||||
"param_type": "optional",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"FRACTAL_NZ"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
# pylint: disable=locally-disabled,too-many-arguments, too-many-locals, too-many-statements
|
||||
@util.check_input_type(dict, dict, (dict, NoneType), dict, bool, bool, str)
|
||||
def CusMatMulCube(input_x1, input_x2, bias=None, output_y={}, trans_a=False, trans_b=False, kernel_name="matmulcube"):
|
||||
"""CusMatMulCube"""
|
||||
return
|
||||
@ -0,0 +1,63 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""CusMatrixCombine"""
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusMatrixCombine",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "matrixcombine.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusMatrixCombine",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float32"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
def CusMatrixCombine(input_x, output, kernel_name="matrix_combine"):
|
||||
"""CusMatrixCombine"""
|
||||
return
|
||||
@ -0,0 +1,63 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""CusTranspose02314"""
|
||||
from mindspore.ops.op_info_register import op_info_register
|
||||
|
||||
|
||||
@op_info_register("""{
|
||||
"op_name": "CusTranspose02314",
|
||||
"imply_type": "TBE",
|
||||
"fusion_type": "OPAQUE",
|
||||
"async_flag": false,
|
||||
"binfile_name": "transpose02314.so",
|
||||
"compute_cost": 10,
|
||||
"kernel_name": "CusTranspose02314",
|
||||
"partial_flag": true,
|
||||
"attr": [
|
||||
],
|
||||
"inputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"NC1HWC0"
|
||||
],
|
||||
"name": "x1",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
],
|
||||
"outputs": [
|
||||
{
|
||||
"index": 0,
|
||||
"dtype": [
|
||||
"float16"
|
||||
],
|
||||
"format": [
|
||||
"DefaultFormat"
|
||||
],
|
||||
"name": "y",
|
||||
"need_compile": false,
|
||||
"param_type": "required",
|
||||
"shape": "all"
|
||||
}
|
||||
]
|
||||
}""")
|
||||
def CusTranspose02314(input_x, output, kernel_name="transpose021354"):
|
||||
"""CusTranspose02314"""
|
||||
return
|
||||
@ -0,0 +1,248 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
"""thor_ops"""
|
||||
import mindspore as ms
|
||||
from mindspore.ops import prim_attr_register, PrimitiveWithInfer
|
||||
from mindspore.ops.composite import multitype_ops as C
|
||||
|
||||
|
||||
class CusBatchMatMul(PrimitiveWithInfer):
|
||||
"""CusMatMulCube definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self):
|
||||
"""init CusMatMulCube"""
|
||||
self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['y'])
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x1, x2, out, dout):
|
||||
return (C.zeros_like(x1), C.zeros_like(x2))
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data1_shape, data2_shape):
|
||||
return data1_shape
|
||||
|
||||
def infer_dtype(self, data1_dtype, data2_dtype):
|
||||
return data1_dtype
|
||||
|
||||
|
||||
class CusCholeskyTrsm(PrimitiveWithInfer):
|
||||
"""CusCholeskyTrsm definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self):
|
||||
"""init CusCholeskyTrsm"""
|
||||
self.init_prim_io_names(inputs=['x1'], outputs=['y'])
|
||||
|
||||
def infer_shape(self, data1_shape):
|
||||
ll = []
|
||||
m, _ = data1_shape
|
||||
if m >= 128:
|
||||
ll = [m // 128, 128, 128]
|
||||
else:
|
||||
ll = [1, 64, 64]
|
||||
return ll
|
||||
|
||||
def infer_dtype(self, data1_dtype):
|
||||
return data1_dtype
|
||||
|
||||
|
||||
class CusFusedAbsMax1(PrimitiveWithInfer):
|
||||
"""CusCholeskyTrsm definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self, origin_shape=[-1, -1]):
|
||||
"""init CusCholeskyTrsm"""
|
||||
self.init_prim_io_names(inputs=['x1'], outputs=['y'])
|
||||
self.origin_shape = origin_shape
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x, out, dout):
|
||||
return (C.zeros_like(x),)
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data1_shape):
|
||||
ll = []
|
||||
if len(data1_shape) == 2:
|
||||
ll = [1,]
|
||||
else:
|
||||
ll = [32, 64]
|
||||
return ll
|
||||
|
||||
def infer_dtype(self, data1_dtype):
|
||||
return data1_dtype
|
||||
|
||||
|
||||
class CusImg2Col(PrimitiveWithInfer):
|
||||
"""CusImg2Col definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self, ksizes, strides, dilates=(1, 1, 1, 1), mode="NC1HWC0"):
|
||||
"""init CusImg2Col"""
|
||||
self.init_prim_io_names(inputs=['x1'], outputs=['y'])
|
||||
self.ksizes = ksizes
|
||||
self.strides = strides
|
||||
self.dilates = dilates
|
||||
self.mode = mode
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x, out, dout):
|
||||
return (C.zeros_like(x),)
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data1_shape):
|
||||
bs, c, h, w = data1_shape
|
||||
_, stride_h, stride_w, _ = self.strides
|
||||
_, k_w, k_h, _ = self.ksizes
|
||||
# assert m == n
|
||||
c0 = 16
|
||||
c1 = c // 16
|
||||
if c1 == 0:
|
||||
c1 = 1
|
||||
shape = [bs * int(h // stride_h) * int(w // stride_w), k_w * k_h * c1 * c0]
|
||||
return shape
|
||||
|
||||
def infer_dtype(self, data1_dtype):
|
||||
return data1_dtype
|
||||
|
||||
|
||||
class CusMatMulCubeDenseLeft(PrimitiveWithInfer):
|
||||
"""CusMatMulCube definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self):
|
||||
"""init CusMatMulCube"""
|
||||
self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['y'])
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x1, x2, out, dout):
|
||||
return (C.zeros_like(x1), C.zeros_like(x2))
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data1_shape, data2_shape):
|
||||
return data2_shape
|
||||
|
||||
def infer_dtype(self, data1_dtype, data2_dtype):
|
||||
return ms.common.dtype.tensor_type(getattr(ms, "float16"))
|
||||
|
||||
|
||||
class CusMatMulCubeFraczRightMul(PrimitiveWithInfer):
|
||||
"""CusMatMulCubeFraczRightMul definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self):
|
||||
"""init CusMatMulCubeFraczRightMul"""
|
||||
self.init_prim_io_names(inputs=['x1', 'x2', 'x3'], outputs=['y'])
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x1, x2, x3, out, dout):
|
||||
return (C.zeros_like(x1), C.zeros_like(x2), C.zeros_like(x3))
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data1_shape, data2_shape, data3_shape):
|
||||
return data1_shape
|
||||
|
||||
def infer_dtype(self, data1_dtype, data2_dtype, data3_dtype):
|
||||
return ms.common.dtype.tensor_type(getattr(ms, "float32"))
|
||||
|
||||
|
||||
class CusMatMulCube(PrimitiveWithInfer):
|
||||
"""CusMatMulCube definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self, transpose_a=False, transpose_b=False):
|
||||
"""init CusMatMulCube"""
|
||||
self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['y'])
|
||||
self.transpose_a = transpose_a
|
||||
self.transpose_b = transpose_b
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x1, x2, out, dout):
|
||||
return (C.zeros_like(x1), C.zeros_like(x2))
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data1_shape, data2_shape):
|
||||
# shape = [1, data1_shape[1], data2_shape[2], 16, 16]
|
||||
# return shape
|
||||
if self.transpose_a:
|
||||
k1, m = data1_shape
|
||||
else:
|
||||
m, k1 = data1_shape
|
||||
if self.transpose_b:
|
||||
n, k2 = data2_shape
|
||||
else:
|
||||
k2, n = data2_shape
|
||||
assert k1 == k2
|
||||
shape = [m, n]
|
||||
return shape
|
||||
|
||||
def infer_dtype(self, data1_dtype, data2_dtype):
|
||||
return ms.common.dtype.tensor_type(getattr(ms, "float32"))
|
||||
|
||||
|
||||
class CusMatrixCombine(PrimitiveWithInfer):
|
||||
"""CusMatMulCube definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self):
|
||||
"""init CusMatMulCube"""
|
||||
self.init_prim_io_names(inputs=['x'], outputs=['y'])
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x, out, dout):
|
||||
return (C.zeros_like(x),)
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data_shape):
|
||||
a, b, c = data_shape
|
||||
shape = [a * b, a * c]
|
||||
|
||||
return shape
|
||||
|
||||
def infer_dtype(self, data_dtype):
|
||||
return data_dtype
|
||||
|
||||
|
||||
class CusTranspose02314(PrimitiveWithInfer):
|
||||
"""CusTranspose02314 definition"""
|
||||
|
||||
@prim_attr_register
|
||||
def __init__(self):
|
||||
"""init CusTranspose02314"""
|
||||
self.init_prim_io_names(inputs=['x1'], outputs=['y'])
|
||||
|
||||
def get_bprop(self):
|
||||
def bprop(x, out, dout):
|
||||
return (C.zeros_like(x),)
|
||||
|
||||
return bprop
|
||||
|
||||
def infer_shape(self, data1_shape):
|
||||
assert len(data1_shape) == 4
|
||||
n, c, h, w = data1_shape
|
||||
c0 = 16
|
||||
c1 = c // 16
|
||||
shape = (n * h * w, c1 * c0)
|
||||
return shape
|
||||
|
||||
def infer_dtype(self, data1_dtype):
|
||||
return data1_dtype
|
||||
Loading…
Reference in new issue