Huawei_Technology
/
mindspore
Not watched
1
0
Fork 0
Code
Releases 13 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

test(custom_op): Delete conv2d custom op.

tags/v0.3.0-alpha
gongchen 5 years ago
parent
817b0e4a59
commit
36edbe411e
6 changed files with 0 additions and 1150 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +0
    -203
      tests/st/ops/custom_ops_tbe/conv2d.py
  2. +0
    -40
      tests/st/ops/custom_ops_tbe/conv2d_impl.py
  3. +0
    -520
      tests/st/ops/custom_ops_tbe/conv_layer.py
  4. +0
    -180
      tests/st/ops/custom_ops_tbe/conv_layer_fast.py
  5. +0
    -153
      tests/st/ops/custom_ops_tbe/cus_conv2d.py
  6. +0
    -54
      tests/st/ops/custom_ops_tbe/test_conv.py

+ 0
- 203
tests/st/ops/custom_ops_tbe/conv2d.py View File

@@ -1,203 +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.
# ============================================================================
from __future__ import absolute_import

import te.lang.cce
from te import platform as cce
from te.platform.fusion_manager import fusion_manager
from topi.cce import util

from .conv_layer import conv_layer_cce
from .conv_layer_fast import conv_layer_fast_cce

Nonetype = type(None)


# pylint: disable=unused-argument, no-value-for-parameter, too-many-branches
@fusion_manager.register("conv2d")
def conv2d_compute(inputs, weights, bias, outputs, strides, pad_list, dilations,
kernel_name="conv2d"):
"""
conv2d compute

Notice
------
only used by framework combine with IR

Parameters
----------
inputs: tvm placeholder
input 5hd feature map tensor
weights: tvm placeholder
input frac_z weight tensor
outputs: tvm placeholder
output tensor, dtype must be assigned
bias: tvm placeholder or None
input 1d bias tensor
strides: integers
stride on H/W, format sensitive
pads: tuple/list of 4 integers
[pad_top, pad_bottom, pad_left, pad_right]
dilations: integers
dilation on H/W, format sensitive
kernel_name: string
kernel name, default value is "conv2d"

Returns
-------
tvm compute
"""
shape_w = []
for i in weights.op.attrs['ori_shape']:
shape_w.append(i.value)

format_w = weights.op.attrs['ori_format']
if format_w == "NCHW":
weight_h = shape_w[2]
weight_w = shape_w[3]
elif format_w == "NHWC":
weight_h = shape_w[1]
weight_w = shape_w[2]
elif format_w == "HWCN":
weight_h = shape_w[0]
weight_w = shape_w[1]
else:
raise RuntimeError("weights ori_format should be NCHW, NHWC or HWCN")

format_x = inputs.op.attrs['ori_format']
if format_x == "NCHW":
strideh = strides[0]
stridew = strides[0]
dlt_h = dilations[0]
dlt_w = dilations[0]
elif format_x == "NHWC":
strideh = strides[0]
stridew = strides[0]
dlt_h = dilations[0]
dlt_w = dilations[0]
else:
raise RuntimeError("inputs ori_format should be NCHW or NHWC")

if len(pad_list) == 4:
padh = [pad_list[0], pad_list[1]]
padw = [pad_list[2], pad_list[3]]
else:
raise RuntimeError("pads shape should be 4d.")

para_dict = {"pad_h": padh, "pad_w": padw, "stride_h": strideh, "stride_w": stridew,
"filter_h": weight_h, "filter_w": weight_w, "bias_tensor": bias}

if cce.CceProductParams().cce_product == "5.10":
para_dict["mad_dtype"] = "float16"
res = te.lang.cce.conv(inputs, weights, para_dict)
else:
res = te.lang.cce.conv(inputs, weights, para_dict)

return res


@util.check_input_type(dict, dict, (dict, Nonetype), dict, (tuple, list), (tuple, list), (tuple, list),
str)
def conv2d(inputs, weights, bias, outputs, strides, pad_list, dilations,
kernel_name="conv2d"):
"""
algorithm: conv2d

Notice
------
only used by framework combine with IR

Parameters
----------
inputs: dict with keys(shape and dtype)
input 4d feature map tensor
weights: dict with keys(shape and dtype)
input 4d weight tensor
outputs: dict with keys(shape and dtype)
output tensor, dtype must be assigned
bias: dict with keys(shape and dtype) or None
input bias tensor
strides: integers
stride on H/W, format sensitive
pads: integers
[pad_top, pad_bottom, pad_left, pad_right]
dilations: tuple/list of 4 integers
dilation on H/W, format sensitive
kernel_name: str
kernel name, default value is "conv2d"

Returns
-------
None
"""
shape_x = inputs.get("ori_shape")
in_dtype = inputs.get("dtype")
shape_w = weights.get("ori_shape")
w_dtype = weights.get("dtype")
res_dtype = outputs.get("dtype")

if len(pad_list) == 4:
padh = [pad_list[0], pad_list[1]]
padw = [pad_list[2], pad_list[3]]
else:
raise RuntimeError("pads shape should be 4d.")

if (not isinstance(shape_x, (tuple, list))) or len(shape_x) != 4:
raise RuntimeError("inputs should be 4d list.")

if (not isinstance(shape_w, (tuple, list))) or len(shape_w) != 4:
raise RuntimeError("weights should be 4d list.")

format_x = inputs.get("ori_format")
if format_x == "NCHW":
shape_fm = shape_x
strideh = strides[0]
stridew = strides[0]
dlt_h = dilations[0]
dlt_w = dilations[0]
elif format_x == "NHWC":
shape_fm = [shape_x[0], shape_x[3], shape_x[1], shape_x[2]]
strideh = strides[0]
stridew = strides[0]
dlt_h = dilations[0]
dlt_w = dilations[0]
else:
raise RuntimeError("inputs ori_format should be NCHW or NHWC.")

format_w = weights.get("ori_format")
if format_w == "NCHW":
shape_filter = shape_w
elif format_w == "NHWC":
shape_filter = [shape_w[0], shape_w[3], shape_w[1], shape_w[2]]
elif format_w == "HWCN":
shape_filter = [shape_w[3], shape_w[2], shape_w[0], shape_w[1]]
else:
raise RuntimeError("weights ori_format should be NCHW, NHWC or HWCN.")

if bias is None:
use_bias = False
else:
use_bias = True

if cce.CceProductParams().cce_product == "5.10":
conv_layer_fast_cce(shape_fm, shape_filter, in_dtype, w_dtype, res_dtype,
padh, padw, strideh, stridew, bias=use_bias,
kernel_name=kernel_name, need_build=True, need_print=False)
else:
conv_layer_cce(shape_fm, shape_filter, in_dtype, w_dtype, res_dtype,
padh, padw, strideh, stridew,
quantize_config=[0, 0, 0], scale_sqrt=[0, 0, 0],
bias=use_bias, kernel_name=kernel_name,
need_build=True, need_print=False)

+ 0
- 40
tests/st/ops/custom_ops_tbe/conv2d_impl.py View File

@@ -1,40 +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.
# ============================================================================
from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
from tests.st.ops.custom_ops_tbe.conv2d import conv2d

cus_conv2D_op_info = TBERegOp("Cus_Conv2D") \
.fusion_type("CONVLUTION") \
.async_flag(False) \
.binfile_name("conv2d.so") \
.compute_cost(10) \
.kernel_name("Cus_Conv2D") \
.partial_flag(True) \
.attr("stride", "required", "listInt", "all") \
.attr("pad_list", "required", "listInt", "all") \
.attr("dilation", "required", "listInt", "all") \
.input(0, "x", False, "required", "all") \
.input(1, "filter", False, "required", "all") \
.input(2, "bias", False, "optional", "all") \
.output(0, "y", True, "required", "all") \
.dtype_format(DataType.F16_5HD, DataType.F16_FracZ, DataType.F32_Default, DataType.F16_5HD) \
.get_op_info()


@op_info_register(cus_conv2D_op_info)
def Cus_Conv2D(inputs, weights, bias, outputs, strides, pads, dilations,
kernel_name="conv2d"):
conv2d(inputs, weights, bias, outputs, strides, pads, dilations,
kernel_name)

+ 0
- 520
tests/st/ops/custom_ops_tbe/conv_layer.py View File

@@ -1,520 +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.
# ============================================================================
import te.lang.cce
from te import tvm
from te.platform import CUBE_MKN
from topi import generic
from topi.cce import util
from topi.cce.util import is_v200_version

# pylint: disable=R0912,R0913,R0914,R0915,E1101
# the dim of shape in conv must be 4
PAD_SHAPE_DIM = 2

NONETYPE = type(None)


@util.check_input_type((list, tuple), (list, tuple), str, str, str, (list, int), (list, int),
int, int, (list, tuple), (list, tuple),
str, str, str,
str, str, str,
str, bool, str)
def conv_layer_cce_para_check(shape_in, shape_w, in_dtype, w_dtype, res_dtype, padh, padw,
strideh, stridew, quantize_config, scale_sqrt,
scale_q_dtype, offset_q_dtype, scale_dq_dtype,
scale_rq_dtype, offset_rq_dtype, offset_w_dtype,
offset_pad_dtype, bias, kernel_name):
# conv shape check
util.check_kernel_name(kernel_name)

# conv data type check
util.check_dtype_rule(in_dtype, ['float16', 'int8', 'uint8'])
util.check_dtype_rule(w_dtype, ['float16', 'int8', 'uint8'])
res_dtype_list = ['float16', 'int8', 'uint8']
if is_v200_version():
res_dtype_list.append('int32')
util.check_dtype_rule(res_dtype, res_dtype_list)
util.check_dtype_rule(scale_q_dtype, ['float16'])
util.check_dtype_rule(offset_q_dtype, ['float16'])
util.check_dtype_rule(scale_dq_dtype, ['float16'])
util.check_dtype_rule(scale_rq_dtype, ['float16'])
util.check_dtype_rule(offset_rq_dtype, ['float16'])
util.check_dtype_rule(offset_w_dtype, ['int32'])
util.check_dtype_rule(offset_pad_dtype, ['uint8'])

if not isinstance(bias, bool):
raise RuntimeError("bias dtype should be bool.")

if quantize_config[0] == 0:
if is_v200_version():
util.check_dtype_rule(in_dtype, ('int8',))
util.check_dtype_rule(w_dtype, ('int8',))
util.check_dtype_rule(res_dtype, ('int32',))
else:
util.check_dtype_rule(in_dtype, ['float16'])
util.check_dtype_rule(w_dtype, ['float16'])
util.check_dtype_rule(res_dtype, ['float16'])

if quantize_config[0] == 1:
util.check_dtype_rule(w_dtype, ['int8'])
if quantize_config[1] == 0:
util.check_dtype_rule(in_dtype, ['int8', 'float16'])
util.check_dtype_rule(res_dtype, ['int8', 'float16'])
elif quantize_config[1] == 1:
util.check_dtype_rule(in_dtype, ['uint8', 'float16'])
util.check_dtype_rule(res_dtype, ['uint8', 'float16'])
elif quantize_config[1] == 2:
raise RuntimeError("All Offset mode quantize not support.")
else:
raise RuntimeError("Invalid quantize algorithm.")

# quantize switch on
if quantize_config[0] == 1:
quantize_turn_on = True
# quantize -> DeQuantize dataflow
if in_dtype == 'float16' and w_dtype == 'int8' and res_dtype == 'float16':
pass
# DeQuantize dataflow
elif (in_dtype in ['int8', 'uint8'] and w_dtype == 'int8' and
res_dtype == 'float16'):
pass
# quantize -> ReQuantize dataflow
elif (in_dtype == 'float16' and w_dtype == 'int8' and res_dtype in
['int8', 'uint8']):
pass
# ReQuantize dataflow
elif (in_dtype in ['int8', 'uint8'] and w_dtype == 'int8' and res_dtype in
['int8', 'uint8']):
pass
else:
raise RuntimeError("Not support in/out data type for quantize.")

if quantize_config not in ([1, 0, 0], [1, 1, 0], [1, 0, 1], [1, 1, 1]):
raise RuntimeError("Invalid Quantize Config.")

if scale_sqrt not in ([0, 0, 0], [1, 0, 0], [0, 1, 0], [1, 1, 0], [0, 0, 1],
[1, 0, 1], [0, 1, 1], [1, 1, 1]):
raise RuntimeError("Invalid Quantize Config.")

# quantize switch off
elif quantize_config[0] == 0:
if quantize_config != [0, 0, 0]:
raise RuntimeError("Invalid Quantize Config.")
if scale_sqrt != [0, 0, 0]:
raise RuntimeError("Invalid Quantize Config.")
else:
raise RuntimeError("Invalid Quantize Config.")

if isinstance(padh, list):
if len(padh) != PAD_SHAPE_DIM:
raise RuntimeError("Dimension must be %d when padh is a list." % PAD_SHAPE_DIM)
pad_top = padh[0]
pad_bottom = padh[1]
else:
pad_top = padh
pad_bottom = padh

if isinstance(padw, list):
if len(padw) != PAD_SHAPE_DIM:
raise RuntimeError("Dimension must be %d when padw is a list." % PAD_SHAPE_DIM)
pad_left = padw[0]
pad_right = padw[1]
else:
pad_left = padw
pad_right = padw

shape_in, shape_w = te.lang.cce.check_conv_shape(shape_in, shape_w, pad_top, pad_bottom, \
pad_left, pad_right, strideh, \
stridew, in_dtype, w_dtype, res_dtype)

return shape_in, shape_w


@util.check_input_type((list, tuple), (list, tuple), str, str, str, \
(list, int), (list, int), int, int,
(list, NONETYPE), (list, NONETYPE),
str, str, str,
str, str, str, str,
bool, str, bool, bool)
def conv_layer_cce(shape_in, shape_w, in_dtype, w_dtype, res_dtype, padh, padw, strideh, stridew,
quantize_config=None, scale_sqrt=None,
scale_q_dtype='float16', offset_q_dtype='float16', scale_dq_dtype='float16',
scale_rq_dtype='float16', offset_rq_dtype='float16', offset_w_dtype='int32',
offset_pad_dtype='uint8', bias=False, kernel_name="cce_conv", need_build=False,
need_print=False):
"""

Parameters
----------
shape_in : shape of data_in

shape_w : shape of filter

in_dtype : the feature map data type

w_dtype : the weight data type

res_dtype : the result data type

padh: the padding shape in H

padw: the padding shape in weight

strideh: the stride value in H

stridew: the stride value in weight

quantize_config: quantize config table, default [0, 0, 0]
quantize_config[0] - quantize function switch
0: quantize off
1: quantize on
quantize_config[1] - quantize_algorithm
0: non offset
1: half offset
2: all offset ( Not supported now )
quantize_config[2] - QuantizeScaleType (for Dequantize/Requantize, quantize always scalar)
0: scalar
1: vector

scale_sqrt: scale mode
scale_sqrt[0] - Quantize scale mode
0: non sqrt
1: sqrt
scale_sqrt[1] - DeQuantize scale mode
0: non sqrt
1: sqrt
scale_sqrt[2] - ReQuantize scale mode
0: non sqrt
1: sqrt

scale_q_dtype: Quantize scale data type, default 'float16'

offset_q_dtype: Quantize offset data type, default 'float16'

scale_dq_dtype: DeQuantize scale data type, default 'float16'

scale_rq_dtype: ReQuantize scale data type, default 'float16'

offset_rq_dtype: ReQuantize offset data type, default 'float16'

offset_w_dtype: weight offset data type, default 'int32'

offset_pad_dtype: Quantize Cube offset data type, default 'uint8'

bias: the tag for bias or not

kernel_name : cce kernel name, default value is "cce_conv"

need_build : if need to build CCEC kernel, default value is False

need_print : if need to print the ir, default value is False

Returns
-------
wrapped_tensor

"""
# for pylint, otherwise "Dangerous default value [] as argument"
if quantize_config is None:
quantize_config = [0, 0, 0]
if scale_sqrt is None:
scale_sqrt = [0, 0, 0]

in_dtype = in_dtype.lower()
w_dtype = w_dtype.lower()
res_dtype = res_dtype.lower()
scale_q_dtype = scale_q_dtype.lower()
offset_q_dtype = offset_q_dtype.lower()
scale_dq_dtype = scale_dq_dtype.lower()
scale_rq_dtype = scale_rq_dtype.lower()
offset_rq_dtype = offset_rq_dtype.lower()
offset_w_dtype = offset_w_dtype.lower()
offset_pad_dtype = offset_pad_dtype.lower()

mad_dtype = 'float32'
if w_dtype == 'int8':
mad_dtype = 'int32'

shape_in = list(shape_in)
shape_w = list(shape_w)

shape_in, shape_w = conv_layer_cce_para_check(shape_in, shape_w, in_dtype, w_dtype, res_dtype, padh, padw, strideh,
stridew,
quantize_config, scale_sqrt, scale_q_dtype, offset_q_dtype,
scale_dq_dtype,
scale_rq_dtype, offset_rq_dtype, offset_w_dtype, offset_pad_dtype,
bias, kernel_name)

# quantize switch on
if quantize_config[0] == 1:
quantize_turn_on = True
# quantize -> DeQuantize dataflow
if in_dtype == 'float16' and w_dtype == 'int8' and res_dtype == 'float16':
is_quantize = True
is_dequantize = True
is_requantize = False
# DeQuantize dataflow
elif (in_dtype in ['int8', 'uint8'] and w_dtype == 'int8' and
res_dtype == 'float16'):
is_quantize = False
is_dequantize = True
is_requantize = False
# quantize -> ReQuantize dataflow
elif (in_dtype == 'float16' and w_dtype == 'int8' and res_dtype in
['int8', 'uint8']):
is_quantize = True
is_dequantize = False
is_requantize = True
# ReQuantize dataflow
elif (in_dtype in ['int8', 'uint8'] and w_dtype == 'int8' and res_dtype in
['int8', 'uint8']):
is_quantize = False
is_dequantize = False
is_requantize = True
else:
raise RuntimeError("Not support in/out data type for quantize.")

# quantize switch off
elif quantize_config[0] == 0:
quantize_turn_on = False
is_quantize = False
is_dequantize = False
is_requantize = False

if quantize_config != [0, 0, 0]:
raise RuntimeError("Invalid Quantize Config.")
if scale_sqrt != [0, 0, 0]:
raise RuntimeError("Invalid Quantize Config.")
else:
raise RuntimeError("Invalid Quantize Config.")

batch_size = shape_in[0]
in_channel = shape_in[1]
feature_map_h = shape_in[2]
feature_map_w = shape_in[3]
block_size_k = CUBE_MKN[in_dtype]['mac'][1]
fmap_shape_nc1hwc0 = (batch_size, (in_channel + block_size_k - 1) // block_size_k,
feature_map_h, feature_map_w, block_size_k)

out_channel = shape_w[0]
in_channel_weight = shape_w[1]
filter_h = shape_w[2]
filter_w = shape_w[3]
block_size_k = CUBE_MKN[w_dtype]['mac'][1]
block_size_n = CUBE_MKN[w_dtype]['mac'][2]
filter_shape_frac_z = (in_channel_weight * filter_h * filter_w // block_size_k,
out_channel // block_size_n, block_size_n, block_size_k)

with tvm.target.cce():
data = tvm.placeholder(
fmap_shape_nc1hwc0, name='Fmap', dtype=in_dtype)
weight = tvm.placeholder(
filter_shape_frac_z, name='Filter', dtype=w_dtype)
bias_tensor = None
scale_q = None
scale_dq = None
scale_rq = None
offset_pad = None
offset_rq = None
offset_q = None
scale_drq = None

# bias or fusion_bias(half offset)
if bias or (quantize_config[1] == 1 and quantize_turn_on):
bias_tensor = tvm.placeholder(
(out_channel,), name='bias_tensor', \
dtype="int32" if quantize_turn_on else res_dtype)

# quantize on
if quantize_turn_on:
quantize_algorithm = quantize_config[1]
if is_quantize:
scale_q = tvm.placeholder(
(CUBE_MKN[scale_q_dtype]['mac'][1],), name='scaleQ', dtype=scale_q_dtype)
if quantize_algorithm == 1:
offset_q = tvm.placeholder(
(CUBE_MKN[offset_q_dtype]['mac'][1],), name='offsetQ', dtype=offset_q_dtype)

if is_dequantize:
scale_dq_shape = (CUBE_MKN[scale_dq_dtype]['mac'][1],) if quantize_config[2] == 0 \
else (out_channel,)
scale_dq = tvm.placeholder(
scale_dq_shape, name='scaleDq', dtype=scale_dq_dtype)

if is_requantize:
scale_rq_shape = (CUBE_MKN[scale_rq_dtype]['mac'][1],) if quantize_config[2] == 0 \
else (out_channel,)
scale_rq = tvm.placeholder(
scale_rq_shape, name='scaleRq', dtype=scale_rq_dtype)
if quantize_algorithm == 1:
offset_rq_shape = (CUBE_MKN[offset_rq_dtype]['mac'][1],)
offset_rq = tvm.placeholder(
offset_rq_shape, name='offsetRq', dtype=offset_rq_dtype)

# need offset_pad , for half offset
if quantize_algorithm == 1:
offset_pad = tvm.placeholder(
(CUBE_MKN[offset_pad_dtype]['mac'][1],), name='offset_pad',
dtype=offset_pad_dtype)

if quantize_algorithm == 0:
if is_quantize:
if is_dequantize:
scale_drq = scale_dq
else:
scale_drq = scale_rq

conv_res = te.lang.cce.conv(
data, weight, {"bias_tensor": bias_tensor,
"scale_q": scale_q,
"offset_q": offset_q,
"scale_drq": scale_drq,
"offset_pad": offset_pad,
"offset_rq": offset_rq,
"quantize_config": quantize_config,
"is_quantize": is_quantize,
"is_dequantize": is_dequantize,
"is_requantize": is_requantize,
"scale_sqrt": scale_sqrt,
"pad_h": padh, "pad_w": padw,
"stride_h": strideh, "stride_w": stridew,
"filter_h": filter_h, "filter_w": filter_w,
"res_dtype": res_dtype, "mad_dtype": mad_dtype},
dsl_flag=False)
if bias:
tensor_list = [data, weight, bias_tensor, scale_q,
scale_drq, conv_res]
else:
tensor_list = [data, weight, scale_q,
scale_drq, conv_res]
else:
if is_dequantize:
scale_drq = scale_dq
else:
scale_drq = scale_rq
conv_res = te.lang.cce.conv(
data, weight, {"bias_tensor": bias_tensor,
"scale_q": scale_q,
"offset_q": offset_q,
"scale_drq": scale_drq,
"offset_pad": offset_pad,
"offset_rq": offset_rq,
"quantize_config": quantize_config,
"is_quantize": is_quantize,
"is_dequantize": is_dequantize,
"is_requantize": is_requantize,
"scale_sqrt": scale_sqrt,
"pad_h": padh, "pad_w": padw,
"stride_h": strideh, "stride_w": stridew,
"filter_h": filter_h, "filter_w": filter_w,
"res_dtype": res_dtype, "mad_dtype": mad_dtype},
dsl_flag=False)
if bias:
tensor_list = [data, weight, bias_tensor,
scale_drq, conv_res]
else:
tensor_list = [data, weight,
scale_drq, conv_res]

# half offset
else:
if is_quantize:
if is_dequantize:
scale_drq = scale_dq
else:
scale_drq = scale_rq
conv_res = te.lang.cce.conv(
data, weight, {"bias_tensor": bias_tensor,
"scale_q": scale_q,
"offset_q": offset_q,
"scale_drq": scale_drq,
"offset_pad": offset_pad,
"offset_rq": offset_rq,
"quantize_config": quantize_config,
"is_quantize": is_quantize,
"is_dequantize": is_dequantize,
"is_requantize": is_requantize,
"scale_sqrt": scale_sqrt,
"pad_h": padh, "pad_w": padw,
"stride_h": strideh, "stride_w": stridew,
"filter_h": filter_h, "filter_w": filter_w,
"res_dtype": res_dtype, "mad_dtype": mad_dtype},
dsl_flag=False)
if is_dequantize:
tensor_list = [data, weight, bias_tensor, scale_q, offset_q,
scale_drq, offset_pad, conv_res]
else:
tensor_list = [data, weight, bias_tensor, scale_q, offset_q,
scale_drq, offset_rq, offset_pad, conv_res]
else:
if is_dequantize:
scale_drq = scale_dq
else:
scale_drq = scale_rq
conv_res = te.lang.cce.conv(
data, weight, {"bias_tensor": bias_tensor,
"scale_q": scale_q,
"offset_q": offset_q,
"scale_drq": scale_drq,
"offset_pad": offset_pad,
"offset_rq": offset_rq,
"quantize_config": quantize_config,
"is_quantize": is_quantize,
"is_dequantize": is_dequantize,
"is_requantize": is_requantize,
"scale_sqrt": scale_sqrt,
"pad_h": padh, "pad_w": padw,
"stride_h": strideh, "stride_w": stridew,
"filter_h": filter_h, "filter_w": filter_w,
"res_dtype": res_dtype, "mad_dtype": mad_dtype},
dsl_flag=False)
if is_dequantize:
tensor_list = [data, weight, bias_tensor,
scale_drq, offset_pad, conv_res]
else:
tensor_list = [data, weight, bias_tensor,
scale_drq, offset_rq, offset_pad, conv_res]
else:
conv_res = te.lang.cce.conv(
data, weight, {"bias_tensor": bias_tensor,
"scale_q": scale_q,
"offset_q": offset_q,
"scale_drq": scale_drq,
"offset_pad": offset_pad,
"offset_rq": offset_rq,
"quantize_config": quantize_config,
"is_quantize": is_quantize,
"is_dequantize": is_dequantize,
"is_requantize": is_requantize,
"scale_sqrt": scale_sqrt,
"pad_h": padh, "pad_w": padw,
"stride_h": strideh, "stride_w": stridew,
"filter_h": filter_h, "filter_w": filter_w,
"res_dtype": res_dtype, "mad_dtype": mad_dtype},
dsl_flag=False)
if bias:
tensor_list = [data, weight, bias_tensor, conv_res]
else:
tensor_list = [data, weight, conv_res]
sch = generic.auto_schedule(conv_res)

config = {
"print_ir": need_print,
"need_build": need_build,
"name": kernel_name,
"tensor_list": tensor_list
}

te.lang.cce.cce_build_code(sch, config)

+ 0
- 180
tests/st/ops/custom_ops_tbe/conv_layer_fast.py View File

@@ -1,180 +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.
# ============================================================================
import te.lang.cce
from te import tvm
from te.platform import CUBE_MKN
from topi import generic
from topi.cce import util

# pylint: disable=R0913,R0914,R0915,E1101
# the dim of shape in conv must be 4
PAD_SHAPE_DIM = 2

NoneType = type(None)


@util.check_input_type((list, tuple), (list, tuple), str, str, str,
(list, int), (list, int), int, int, bool, str)
def conv_layer_fast_cce_para_check(shape_in, shape_w, in_dtype, w_dtype, res_dtype,
padh, padw, strideh, stridew, bias, kernel_name):
# conv shape check
util.check_kernel_name(kernel_name)

# conv data type check
util.check_dtype_rule(in_dtype, ['float16'])
util.check_dtype_rule(w_dtype, ['float16'])
util.check_dtype_rule(res_dtype, ['float16'])

if not isinstance(bias, bool):
raise RuntimeError("bias dtype should be bool.")

if isinstance(padh, list):
if len(padh) != PAD_SHAPE_DIM:
raise RuntimeError("Dimension must be %d when padh is a list." % PAD_SHAPE_DIM)
pad_top = padh[0]
pad_bottom = padh[1]
else:
pad_top = padh
pad_bottom = padh

if isinstance(padw, list):
if len(padw) != PAD_SHAPE_DIM:
raise RuntimeError("Dimension must be %d when padw is a list." % PAD_SHAPE_DIM)
pad_left = padw[0]
pad_right = padw[1]
else:
pad_left = padw
pad_right = padw

shape_in, shape_w = te.lang.cce.check_conv_shape(shape_in, shape_w, pad_top, pad_bottom,
pad_left, pad_right, strideh, stridew,
in_dtype, w_dtype, res_dtype)

return shape_in, shape_w


@util.check_input_type((list, tuple), (list, tuple), str, str, str,
(list, int), (list, int), int, int,
bool, str, bool, bool)
def conv_layer_fast_cce(shape_in, shape_w, in_dtype, w_dtype, res_dtype,
padh, padw, strideh, stridew, bias=False,
kernel_name="cce_conv",
need_build=False, need_print=False):
"""

Parameters
----------
shape_in : shape of data_in

shape_w : shape of filter

in_dtype : the feature map data type

w_dtype : the weight data type

res_dtype : the result data type

padh: the padding shape in H

padw: the padding shape in weight

strideh: the stride value in H

stridew: the stride value in weight

bias: the tag for bias or not

kernel_name : cce kernel name, default value is "cce_conv"

need_buid : if need to build CCEC kernel, default value is False

need_print : if need to print the ir, default value is False

Returns
-------
None

"""
in_dtype = in_dtype.lower()
w_dtype = w_dtype.lower()
res_dtype = res_dtype.lower()

shape_in = list(shape_in)
shape_w = list(shape_w)

shape_in, shape_w = conv_layer_fast_cce_para_check(shape_in, shape_w, in_dtype, w_dtype, res_dtype,
padh, padw, strideh, stridew, bias, kernel_name)

batch_size = shape_in[0]
in_channel = shape_in[1]
feature_map_h = shape_in[2]
feature_map_w = shape_in[3]
block_size_k = CUBE_MKN[in_dtype]['mac'][1]
fmap_shape_nc1hwc0 = (batch_size, (in_channel + block_size_k - 1) // block_size_k,
feature_map_h, feature_map_w, block_size_k)

out_channel = shape_w[0]
in_channel_weight = shape_w[1]
filter_h = shape_w[2]
filter_w = shape_w[3]
block_size_k = CUBE_MKN[w_dtype]['mac'][1]
block_size_n = CUBE_MKN[w_dtype]['mac'][2]
filter_shape_frac_z = (in_channel_weight * filter_h * filter_w // block_size_k,
out_channel // block_size_n, block_size_n, block_size_k)

with tvm.target.cce():
data = tvm.placeholder(
fmap_shape_nc1hwc0, name='Fmap', dtype=in_dtype)
weight = tvm.placeholder(
filter_shape_frac_z, name='Filter', dtype=w_dtype)
bias_tensor = None

if bias:
bias_tensor = tvm.placeholder(
(out_channel,), name='bias_tensor', dtype=res_dtype)

mad_dtype = "float16"

conv_res = te.lang.cce.conv(
data, weight, {"bias_tensor": bias_tensor,
"scale_q": None,
"offset_q": None,
"scale_drq": None,
"offset_pad": None,
"offset_rq": None,
"quantize_config": [0, 0, 0],
"is_quantize": False,
"is_dequantize": False,
"is_requantize": False,
"scale_sqrt": [0, 0, 0],
"pad_h": padh, "pad_w": padw,
"stride_h": strideh, "stride_w": stridew,
"filter_h": filter_h, "filter_w": filter_w,
"res_dtype": res_dtype, "mad_dtype": mad_dtype},
dsl_flag=False)
if bias:
tensor_list = [data, weight, bias_tensor, conv_res]
else:
tensor_list = [data, weight, conv_res]
sch = generic.auto_schedule(conv_res)

config = {
"print_ir": need_print,
"need_build": need_build,
"name": kernel_name,
"tensor_list": tensor_list
}

te.lang.cce.cce_build_code(sch, config)

+ 0
- 153
tests/st/ops/custom_ops_tbe/cus_conv2d.py View File

@@ -1,153 +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.
# ============================================================================
import math
import numpy as np
from functools import reduce

from mindspore import Tensor
from mindspore._checkparam import ParamValidator as validator
from mindspore._checkparam import Rel, check_bool, check_int_positive, twice
from mindspore.common import dtype as mstype
from mindspore.ops import prim_attr_register, PrimitiveWithInfer


class Cus_Conv2D(PrimitiveWithInfer):
r"""
Applies 2D convolution for the input.

Input is typically of shape :math:`(N, C, H, W)`, where :math:`N` is batch size and :math:`C` is channel number.
For each batch of shape :math:`(C, H, W)` the formula (given mode 1) is defined as:

.. math::
out_j = \sum_{i=0}^{C_{in} - 1} ccor(W_{ij}, X_i) + b_j,

where :math:`ccor` is cross correlation operator, :math:`C_{in}` is the input channel number, :math:`j` ranges
from :math:`0` to :math:`C_{out} - 1`, :math:`W_{ij}` corresponds to i-th channel of the j-th filter and
:math:`out_{j}` corresponds to the j-th channel of the output.

The first introduction can be found in paper `Gradient Based Learning Applied to Document Recognition
<http://vision.stanford.edu/cs598_spring07/papers/Lecun98.pdf>`_.

More detailed introduction can be found here: http://cs231n.github.io/convolutional-networks/.

Args:
out_channel (int): The dimensionality of the output space.
kernel_size (Union[int, tuple[int]]): The kernel size of the 2D convolution.
mode (int): 0 Math convolutiuon, 1 cross-correlation convolution ,
2 deconvolution, 3 depthwise convolution. Default: 1.
pad_mode (str): "valid", "same", "pad" the mode to fill padding. Default: "valid".
pad (int): The pad value to fill. Default: 0.
stride (int): The stride to apply conv filter. Default: 1.
dilation (int): Specifying the dilation rate to use for dilated convolution. Default: 1.
group (int): Split input into groups. Default: 1.

Returns:
Tensor, the value that applied 2D convolution.

Inputs:
- **input** (Tensor) - Tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})`.

Outputs:
Tensor of shape :math:`(N, C_{out}, H_{out}, W_{out})`.
"""

@prim_attr_register
def __init__(self,
out_channel,
kernel_size,
mode=1,
pad_mode="valid",
pad=0,
stride=1,
dilation=1,
group=1):
"""init Conv2D"""
self.init_prim_io_names(inputs=['x', 'w'], outputs=['output'])
self.kernel_size = kernel_size
self.kernel_size = validator.check_type('kernel_size', kernel_size, (int, tuple))
if isinstance(self.kernel_size, int):
self.kernel_size = (self.kernel_size, self.kernel_size)
validator.check_integer('length of kernel_size', len(self.kernel_size), 2, Rel.GE)
validator.equal('type of pad', type(pad), 'not bool', not isinstance(pad, bool))
validator.equal('type of pad', type(pad), 'int', isinstance(pad, int))
self.pad_mode = validator.check_string('pad_mode', pad_mode, ['valid', 'same', 'pad'])
self.pad = validator.check_pad_value_by_mode(self.__class__.__name__, pad_mode, pad)
if self.pad_mode == 'pad':
validator.check_integer('pad', self.pad, 0, Rel.GE)

self.mode = validator.check_integer('mode', mode, 1, Rel.EQ)
self.add_prim_attr('data_format', "NCHW")
self.out_channel = validator.check_integer('out_channel', out_channel, 0, Rel.GT)
self.group = validator.check_integer('group', group, 0, Rel.GT)
self.dilation = validator.check_integer('dilation', dilation, 1, Rel.GE)
validator.check_type('kernel_size', kernel_size, [int, tuple])
if isinstance(kernel_size, int) and kernel_size < 1:
raise ValueError('Attr \'kernel_size\' of \'Conv2D\' Op passed '
+ str(self.kernel_size) + ', should be a int or tuple and equal to or greater than 1.')
if isinstance(kernel_size, tuple) and (len(kernel_size) != 2 or
(not isinstance(kernel_size[0], int)) or
(not isinstance(kernel_size[1], int)) or
kernel_size[0] < 1 or kernel_size[1] < 1):
raise ValueError('Attr \'kernel_size\' of \'Conv2D\' Op passed '
+ str(self.kernel_size) + ', should be a int or tuple and equal to or greater than 1.')
self.stride = validator.check_integer('stride', stride, 1, Rel.GE)
from conv2d_impl import Cus_Conv2D

def infer_shape(self, x_shape, w_shape):
validator.check_integer("weight_shape", len(w_shape), 4, Rel.EQ)
validator.check_integer("x_shape", len(x_shape), 4, Rel.EQ)
validator.check_param_equal("x_shape[1]", x_shape[1] // self.group, "w_shape[1]", w_shape[1])
validator.check_param_equal('out_channel', self.out_channel, 'w_shape[0]', w_shape[0])
validator.check_param_equal('kernel_size', self.kernel_size, 'w_shape[2:4]', tuple(w_shape[2:4]))

kernel_size_h = w_shape[2]
kernel_size_w = w_shape[3]

if self.pad_mode == "valid":
h_out = math.ceil((x_shape[2] - kernel_size_h + 1) / self.stride)
w_out = math.ceil((x_shape[3] - kernel_size_w + 1) / self.stride)
pad_top, pad_bottom, pad_left, pad_right = 0, 0, 0, 0
elif self.pad_mode == "same":
h_out = math.ceil(x_shape[2] / self.stride)
w_out = math.ceil(x_shape[3] / self.stride)

pad_needed_h = max(0, (h_out - 1) * self.stride + kernel_size_h - x_shape[2])
pad_top = math.floor(pad_needed_h / 2)
pad_bottom = pad_needed_h - pad_top

pad_needed_w = max(0, (w_out - 1) * self.stride + kernel_size_w - x_shape[3])
pad_left = math.floor(pad_needed_w / 2)
pad_right = pad_needed_w - pad_left
elif self.pad_mode == 'pad':
pad_top, pad_bottom, pad_left, pad_right = self.pad, self.pad, self.pad, self.pad

h_out = 1 + (x_shape[2] + 2 * self.pad - kernel_size_h - (kernel_size_h - 1) * (self.dilation - 1)) \
/ self.stride
w_out = 1 + (x_shape[3] + 2 * self.pad - kernel_size_w - (kernel_size_w - 1) * (self.dilation - 1)) \
/ self.stride
h_out = math.floor(h_out)
w_out = math.floor(w_out)

self.pad_list = [pad_top, pad_bottom, pad_left, pad_right]
self.add_prim_attr('pad_list', (pad_top, pad_bottom, pad_left, pad_right))

out_channel = self.out_channel
out_shape = [x_shape[0], out_channel, h_out, w_out]
return out_shape

def infer_dtype(self, x_dtype, w_dtype):
args = {'x_dtype': x_dtype, 'w_dtype': w_dtype}
validator.check_type_same(args, [mstype.int8, mstype.int32, mstype.float16, mstype.float32])
return x_dtype

+ 0
- 54
tests/st/ops/custom_ops_tbe/test_conv.py View File

@@ -1,54 +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.
# ============================================================================
import numpy as np

import mindspore.context as context
import mindspore.nn as nn
from mindspore import Tensor
from mindspore.common.api import ms_function
from mindspore.common.initializer import initializer
from mindspore.common.parameter import Parameter
from .cus_conv2d import Cus_Conv2D

context.set_context(device_target="Ascend")


class Net(nn.Cell):
def __init__(self):
super(Net, self).__init__()
out_channel = 64
kernel_size = 7
self.conv = Cus_Conv2D(out_channel,
kernel_size,
mode=1,
pad_mode="valid",
pad=0,
stride=1,
dilation=1,
group=1)
self.w = Parameter(initializer(
'normal', [64, 3, 7, 7]), name='w')

@ms_function
def construct(self, x):
return self.conv(x, self.w)


def test_net():
np.random.seed(3800)
x = np.random.randn(32, 3, 224, 224).astype(np.float32)
conv = Net()
output = conv(Tensor(x))
print(output.asnumpy())

Write Preview
Loading…
Cancel
Save