2022-08-26 17:42:09 +00:00

259 lines
9.2 KiB
Python

# ----------------------------------------------------------------------
# Project: TinyEngine
# Title: conv2d.py
#
# Reference papers:
# - MCUNet: Tiny Deep Learning on IoT Device, NeurIPS 2020
# - MCUNetV2: Memory-Efficient Patch-based Inference for Tiny Deep Learning, NeurIPS 2021
# - MCUNetV3: On-Device Training Under 256KB Memory, arXiv:2206.15472
# Contact authors:
# - Wei-Ming Chen, wmchen@mit.edu
# - Wei-Chen Wang, wweichen@mit.edu
# - Ji Lin, jilin@mit.edu
# - Ligeng Zhu, ligeng@mit.edu
# - Song Han, songhan@mit.edu
#
# Target ISA: ARMv7E-M
# ----------------------------------------------------------------------
import warnings
from .basic_utils import basicOperator, deep_copy_dicts, overwrite_dicts
__all__ = ["Conv2d"]
default_params = {
# op related
"op": "CONV_2D",
"stride_h": None,
"stride_w": None,
"input_idx": None,
"output_idx": None,
# tensor related
"input_dim": None,
"input_h": None,
"input_w": None,
"input_c": None,
"output_dim": None,
"output_h": None,
"output_w": None,
"output_c": None,
"kernel_h": None,
"kernel_w": None,
"input_dtype": "int8",
"output_dtype": "int8",
"padding": None,
# quantization related
"weight_value": None,
"weight_name": None,
"bias": None,
"bias_name": None,
"effective_scale": None,
"input_zero_point": None,
"output_zero_point": None,
"input_scale": None,
"output_scale": None,
"weight_scale": None,
"multiplier": None,
"shift": None,
}
class Conv2d(basicOperator):
def __init__(self, params: dict) -> None:
self.params = deep_copy_dicts(default_params)
overwrite_dicts(self.params, params)
super().__init__()
# handle input/output tensors in HWC format
self._add_input(
self.params["input_idx"],
self.params["input_dtype"],
self.params["input_c"],
self.params["input_w"],
self.params["input_h"],
)
self._add_output(
self.params["output_idx"],
self.params["output_dtype"],
self.params["output_c"],
self.params["output_w"],
self.params["output_h"],
)
if None in default_params:
warnings.warn(f"parameters are not all set for op {self.params['op']}")
def _op_hparam_info(self) -> str:
string = f" k{self.params['kernel_h']}x{self.params['kernel_w']}_r{self.params['input_h']}"
string += f"x{self.params['input_w']}x{self.params['input_c']}_{self.params['output_h']}"
string += f"x{self.params['output_w']}x{self.params['output_c']}"
return string
def set_input_zero_point(self, zero_x):
p = self.params
p["input_zero_point"] = zero_x
def set_output_zero_point(self, zero_y):
p = self.params
p["input_zero_point"] = zero_y
def get_macs(self):
p = self.params
return p["output_h"] * p["output_w"] * p["kernel_h"] * p["kernel_w"] * p["input_c"] * p["output_c"]
def get_weights_size(self) -> int:
p = self.params
if p["input_dtype"] in {"float32", "fp32"}:
size = 4
else:
size = 1
return p["kernel_h"] * p["kernel_w"] * p["input_c"] * p["output_c"] * size
def get_bias_size(self) -> int:
p = self.params
return 4 * p["output_c"]
def get_scale_size(self) -> int:
p = self.params
return 4 * p["output_c"]
def get_sbuf_size(self) -> int:
params = self.params
if params["input_dtype"] == params["output_dtype"] == "int8":
p = self.params
return p["kernel_h"] * p["kernel_w"] * p["input_c"] * 2 * 2 # 2 col and 16 bit
else:
return 0
def get_kbuf_size(self) -> int:
p = self.params
if p["kernel_h"] == 1:
return 0
else:
return self.get_weights_size() * 2 # 16bit
def generate_inference_str(
self,
unsigned_input: bool = False,
FP_output: bool = False,
use_aggressive_unroll: bool = False,
use_hard_switsh: bool = False,
fp_requantize: bool = False,
tflite_op: bool = False,
dummy_address: bool = False,
):
string = ""
params = self.params
# floating point implmenetation
kernel_h = params["kernel_h"]
# function name
if params["kernel_h"] == 1:
# find the proper function
if params["output_c"] % 2 != 0:
if (
FP_output
and "effective_scale" in params
and params["output_scale"] is not None
and params["effective_scale"] is not None
):
function_name = "convolve_1x1_s8_oddch_fp"
else:
function_name = "convolve_1x1_s8_oddch"
else:
if use_aggressive_unroll and params["input_c"] in [8, 16, 24, 48]:
function_name = f"convolve_1x1_s8_ch{str(params['input_c'])}"
else:
if (
FP_output
and "effective_scale" in params
and params["output_scale"] is not None
and params["effective_scale"] is not None
):
function_name = "convolve_1x1_s8_fp"
else:
function_name = "convolve_1x1_s8"
elif kernel_h == 3 and params["stride_h"] == 2 and params["padding"] == 1:
if unsigned_input:
function_name = "convolve_u8_kernel3_inputch3_stride2_pad1"
else:
if "is_patch" in params and params["is_patch"]:
function_name = "patchpadding_convolve_s8_kernel3_inputch3_stride2"
else:
function_name = "convolve_s8_kernel3_inputch3_stride2_pad1"
elif kernel_h == 3 and params["stride_h"] == 1 and params["padding"] == 1:
function_name = "convolve_s8_kernel3_stride1_pad1"
else:
raise NotImplementedError
if fp_requantize:
function_name += "_fpreq"
# input tensor, shape, weight, bias
if unsigned_input:
string += f"{function_name}((const q8_t *)"
string += f"{self._getBufferstr(params['input_buf_add'], params['input_buf_add_offset'])},"
else:
string += f"{function_name}({self._getBufferstr(params['input_buf_add'], params['input_buf_add_offset'])},"
string += f"{str(params['input_w'])},{str(params['input_h'])},{str(params['input_c'])},"
parsed_idx = str(params["parsed_trainable"])
string += f"(const q7_t*) weight{parsed_idx},bias{parsed_idx},"
# scales or multiplier and shift
if (
fp_requantize
or FP_output
and "effective_scale" in params
and params["output_scale"] is not None
and params["effective_scale"] is not None
):
string += f"scales{parsed_idx},"
else:
string += f"shift{parsed_idx},multiplier{parsed_idx},"
# output: zero point, min, max, output tensor, shape
string += f"{str(params['output_zero_point'])},{str(params['input_zero_point'] * -1)},-128,127,"
string += f"{self._getBufferstr(params['output_buf_add'], params['output_buf_add_offset'])},"
string += f"{str(params['output_w'])},{str(params['output_h'])},{str(params['output_c'])},"
# intemediate buffers
string += "sbuf"
if (
kernel_h == 3
and params["stride_h"] == 2
and params["padding"] == 1
and not ("is_patch" in params and params["is_patch"])
):
string += ",kbuf"
# pad value for kernel size > 1
if kernel_h > 1:
string += f",{str(params['input_zero_point'])}"
# patch-based parameters
if "is_patch" in params and params["is_patch"] and (params["kernel_h"] > 1 or params["kernel_w"] > 1):
string += ", pad_t, pad_b, pad_l, pad_r);\n"
stride_string = str(params["stride_h"])
string += f"pad_t /= {stride_string};pad_b /= {stride_string};pad_l /= {stride_string};"
string += f"pad_r /= {stride_string};\n"
else:
string += ");\n"
if use_hard_switsh:
string = (
"input_shape.Resize(4); ptr = input_shape.DimsData();\n"
+ f"ptr[0] = 1;ptr[1] = {str(params['output_h'])};ptr[2] = {str(params['output_w'])};"
+ f"ptr[3] = {str(params['output_c'])};\n"
)
string += (
"output_shape.Resize(4); ptr = output_shape.DimsData();\n"
+ f"ptr[0] = 1;ptr[1] = {str(params['output_h'])};ptr[2] = {str(params['output_w'])};"
+ f"ptr[3] = {str(params['output_c'])};\n"
)
string += (
"tflite::reference_ops::HardSwish<int8_t>(hs_op_params, input_shape, "
+ f"{self._getBufferstr(params['output_buf_add'], params['output_buf_add_offset'])},"
+ f" output_shape, {self._getBufferstr(params['output_buf_add'], params['output_buf_add_offset'])});\n"
)
return string