You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
mindspore/tests/st/ops/custom_ops_tbe/conv_layer.py

520 lines
22 KiB

# 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 -> 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)