本节设计了一个张量加法的Sample，通过综合应用文档介绍的TIK优化机制，增强用户对TIK编程与优化的理解。
为保证张量加法Sample的结构清晰性和易读性，本章采用类的形式对Sample进行组织和实现，类的定义如下所示。
class Vadd():
# 接收数据，并完成相关初始化计算
def __init__(self, input_x, input_y, output_z, kernel_name="vadd_sample"):
# 完成算子计算与编译
def vadd_compute(self):
# 定义每个AI Core上的运算
def vadd_compute_each_core(self, move_offset, move_num):
# 定义AI Core上的分片计算
def vadd_compute_each_loop(self, move_offset, move_num):
# 用于功能和性能测试
def vadd_sample(input_x, input_y, output_z, kernel_name):
完整Sample如下所示。
import math
from functools import reduce as functools_reduce
import numpy as np
from tbe import tik
import tbe.common.platform as tbe_platform
from tbe.common.utils import para_check
# 计算每种数据类型所占bit数
def get_bit_len(dtype):
index = 0
for i in dtype:
if i.isdigit():
break
index += 1
return int(dtype[index:])
class Vadd():
def __init__(self, input_x, input_y, output_z, kernel_name="vadd_sample"):
self.shape_x = input_x.get("shape")
self.dtype_x = input_x.get("dtype")
self.shape_y = input_y.get("shape")
self.dtype_y = input_y.get("dtype")
self.shape_z = output_z.get("shape")
self.dtype_z = output_z.get("dtype")
self.kernel_name = kernel_name
# 构造TIK容器并开启debug调试功能
self.tik_instance = tik.Tik(disable_debug=False)
# 设置昇腾AI处理器的版本，若不设置，默认使用“Ascend310”
soc_version="Ascend310"
tbe_platform.set_current_compile_soc_info(soc_version,core_type="AiCore")
# 获取AI Core的个数
self.aicore_num = tbe_platform.get_soc_spec("CORE_NUM")
# Unified Buffer上数据读取和写入必须32B对齐，此参数用来计算tensor划分和数据搬运指令参数
block_byte_size = 32
# 获取Unified Buffer空间大小，单位为bytes
ub_size_bytes = tbe_platform.get_soc_spec("UB_SIZE")
# 根据输入的数据类型计算一个block可以存放多少个对应的元素
dtype_bytes_size = get_bit_len(self.dtype_x) // 8
self.data_each_block = block_byte_size // dtype_bytes_size
# 计算在Unified Buffer上给两个输入和计算结果分别分配多少空间（地址重叠），并进行32B对齐
self.ub_tensor_size = (
ub_size_bytes // dtype_bytes_size // 2 // self.data_each_block *
self.data_each_block)
# 计算输入的元素个数
self.input_num = functools_reduce(lambda x, y: x * y, self.shape_x)
# 计算每个AI Core需要处理的数据量，当前只考虑均分场景，且均分后32 Bytes对齐
self.data_num_each_core = self.input_num // self.aicore_num
# vector指令每个repeat最多计算8个block，该参数为mask的最大值
self.vector_mask_max = 8 * self.data_each_block
self.input_x_gm = self.tik_instance.Tensor(
self.dtype_x, self.shape_x, name="input_x_gm", scope=tik.scope_gm)
self.input_y_gm = self.tik_instance.Tensor(
self.dtype_y, self.shape_y, name="input_y_gm", scope=tik.scope_gm)
self.output_z_gm = self.tik_instance.Tensor(
self.dtype_z, self.shape_z, name="output_z_gm", scope=tik.scope_gm)
def vadd_compute(self):
with self.tik_instance.for_range(
0, self.aicore_num, block_num=self.aicore_num) as index:
# 创建两个输入在Unified Buffer上的tensor
self.input_x_ub = self.tik_instance.Tensor(
self.dtype_x, (self.ub_tensor_size,),
name="input_x_ub",
scope=tik.scope_ubuf)
self.input_y_ub = self.tik_instance.Tensor(
self.dtype_y, (self.ub_tensor_size,),
name="input_y_ub",
scope=tik.scope_ubuf)
# 将对应的GM上的数据搬运到Unified Buffer，每次搬运的偏移量为已经处理过的数据个数
move_offset = index * self.data_num_each_core
# 每个aicore计算自己负责的数据分片
self.vadd_compute_each_core(move_offset, self.data_num_each_core)
self.tik_instance.BuildCCE(
kernel_name=self.kernel_name,
inputs=[self.input_x_gm, self.input_y_gm],
outputs=[self.output_z_gm])
return self.tik_instance
def vadd_compute_each_core(self, move_offset, move_num):
loop_time = move_num // self.ub_tensor_size
move_offset_init = move_offset
if loop_time > 0:
with self.tik_instance.for_range(0, loop_time) as loop_index:
move_offset += loop_index * self.ub_tensor_size
self.vadd_compute_each_loop(move_offset, self.ub_tensor_size)
move_offset = move_offset_init + loop_time * self.ub_tensor_size
last_num = move_num % self.ub_tensor_size
if last_num > 0:
self.vadd_compute_each_loop(move_offset, last_num)
def vadd_compute_each_loop(self, move_offset, move_num):
# 计算每次搬运的burst_len
burst_len = math.ceil(move_num / self.data_each_block)
self.tik_instance.data_move(self.input_x_ub,
self.input_x_gm[move_offset], 0, 1,
burst_len, 0, 0)
self.tik_instance.data_move(self.input_y_ub,
self.input_y_gm[move_offset], 0, 1,
burst_len, 0, 0)
vadd_loop = move_num // (self.vector_mask_max * 255)
add_offset = 0
if vadd_loop > 0:
with self.tik_instance.for_range(0, vadd_loop) as add_index:
add_offset = add_index * self.vector_mask_max * 255
self.tik_instance.vec_add(self.vector_mask_max,
self.input_x_ub[add_offset],
self.input_x_ub[add_offset],
self.input_y_ub[add_offset],
255, 8, 8, 8)
add_offset = vadd_loop * self.vector_mask_max * 255
repeat_time = (
move_num % (self.vector_mask_max * 255) // self.vector_mask_max)
if repeat_time > 0:
self.tik_instance.vec_add(self.vector_mask_max,
self.input_x_ub[add_offset],
self.input_x_ub[add_offset],
self.input_y_ub[add_offset],
repeat_time, 8, 8, 8)
add_offset += repeat_time * self.vector_mask_max
last_num = move_num % self.vector_mask_max
if last_num > 0:
self.tik_instance.vec_add(last_num,
self.input_x_ub[add_offset],
self.input_x_ub[add_offset],
self.input_y_ub[add_offset],
1, 8, 8, 8)
self.tik_instance.data_move(self.output_z_gm[move_offset],
self.input_x_ub, 0, 1, burst_len, 0, 0)
@para_check.check_input_type(dict, dict, dict, str)
def vadd_sample(input_x, input_y, output_z, kernel_name):
"""
calculating data
Parameters
----------
input_x : dict
shape and dtype of input
input_y : dict
shape and dtype of input
output_z : dict
shape and dtype of output, should be same shape and type as input
kernel_name : str
kernel name, default value is "vadd_sample"
Returns
-------
None
"""
vadd_instance = Vadd(input_x, input_y, output_z, kernel_name)
tik_instance = vadd_instance.vadd_compute()
return tik_instance
if __name__ == "__main__":
tik_instance = vadd_sample({"shape":(32, 16384), "dtype":"float16"},
{"shape":(32, 16384), "dtype":"float16"},
{"shape":(32, 16384), "dtype":"float16"},
"vadd_32_16384_float16")
input_x = np.random.uniform(-10, 5, (32, 16384)).astype(np.float16)
input_y = np.random.uniform(-10, 5, (32, 16384)).astype(np.float16)
feed_dict = {
'input_x_gm': input_x,
'input_y_gm': input_y
}
result = tik_instance.tikdb.start_debug(feed_dict, interactive=True)
print(result)