据说tvm会支持训练,然后沐神他们为了推广tvm,专门对tvm重写了文档d2l-tvm,具体详见资料[1],下面就是对着沐神他们写的文档做的学习记录:
1.tvm的数据类型
我们在声明placeholder的时候,可以显式的指定数据的类型,如:'float16', 'float64', 'int8','int16', 'int64'等
import tvm
A = tvm.placeholder((n,), dtype='float32')
print(A.dtype)
B = tvm.placeholder((n,), dtype='int8')
print(B.dtype)
结果如下:
下面来看一下,如何使用tvm来定义两个向量相加的运算:
首先声明两个占位符A和B,并指定他们的shape:
A = tvm.placeholder((n,), dtype=dtype)
A = tvm.placeholder((n,), dtype=dtype)
然后使用使用tvm.compute定义C的计算,并使用lambda函数来指定输出C中每个元素的计算方式:
C = tvm.compute(A.shape, lambda i: A[i] + B[i])
其次,我们需要指定如何来执行程序,如,数据访问顺序及如何进行多线程并行计算,这样一个执行设计称之为一个schedule,C为输出,下面使用tvm.create_schedule就对C的操作创建一个默认的schedule:
s = tvm.create_schedule(C.op)
后面可以看到如何通过改变执行设计来更好的利用硬件资源,进而提升执行效率。这里,我们可以使用tvm.lower操作,打印一下当前默认执行设计:
tvm.lower(s, [A, B, C], simple_mode=True)
可以得到:
最后,使用tvm.build操作将定义好的计算和设计编译成可执行模块:
module = tvm.build(s, [A, B, C])
将上面的代码连在一起,就是使用tvm来创建两个向量相加的函数:
import tvm
import numpy as np
n = 100
# Defined in file: ./chapter_expression/vector_add.md
def eval_mod(mod, *inputs, out):
"""Evaluate a TVM module, and save results in out.
"""
# Convert all numpy arrays to tvm arrays
tvm_args = [tvm.nd.array(x) if isinstance(x, np.ndarray)
else x for x in inputs + (out,)]
mod(*tvm_args)
# If out is a tvm array, then its value has already been inplaced.
# Otherwise, explicitly copy the results.
if isinstance(out, np.ndarray):
np.copyto(out, tvm_args[-1].asnumpy())
def tvm_vector_add(dtype):
A = tvm.placeholder((n,), dtype=dtype)
B = tvm.placeholder((n,), dtype=dtype)
C = tvm.compute(A.shape, lambda i: A[i] + B[i])
s = tvm.create_schedule(C.op)
return tvm.build(s, [A, B, C])
mod = tvm_vector_add('int32')
def test_mod(mod, dtype):
# you can use astype to convert data type
a, b = (np.random.normal(size=100).astype(dtype) for _ in range(2))
c = np.empty(100, dtype=dtype)
eval_mod(mod, a, b, out=c)
print("data type of c: {}".format(c.dtype))
np.testing.assert_equal(c, a + b)
test_mod(mod, 'int32')
for dtype in ['float16', 'float64', 'int8', 'int16', 'int64']:
mod = tvm_vector_add(dtype)
test_mod(mod, dtype)
运行效果如下:
2.变化形状
在定义计算时,可能对于输入的形状是未知的,可以通过tvm.var定义一个变量来指定形状,然后在具体调用时,传入具体值即可,如对于定义A、B、C三个占位符时,如果不知道输入的维度,可以使用变量n来创建任意长度数组:
n = tvm.var(name='n')
print(type(n), n.dtype)
A = tvm.placeholder((n,), name='a')
B = tvm.placeholder((n,), name='b')
C = tvm.compute(A.shape, lambda i: A[i] + B[i])
s = tvm.create_schedule(C.op)
mod = tvm.build(s, [A, B, C])
在调用时,传入数组长度:
def test_mod(mod, size):
a, b = (np.random.normal(size=size).astype('float32') for _ in range(2))
c = np.empty(size, dtype='float32')
print("c shape: {}".format(c.shape))
eval_mod(mod, a, b, out=c)
np.testing.assert_equal(c, a + b)
完整代码如下:
import numpy as np
import tvm
# Defined in file: ./chapter_expression/vector_add.md
def eval_mod(mod, *inputs, out):
"""Evaluate a TVM module, and save results in out.
"""
# Convert all numpy arrays to tvm arrays
tvm_args = [tvm.nd.array(x) if isinstance(x, np.ndarray)
else x for x in inputs + (out,)]
mod(*tvm_args)
# If out is a tvm array, then its value has already been inplaced.
# Otherwise, explicitly copy the results.
if isinstance(out, np.ndarray):
np.copyto(out, tvm_args[-1].asnumpy())
n = tvm.var(name='n')
print(type(n), n.dtype)
A = tvm.placeholder((n,), name='a')
B = tvm.placeholder((n,), name='b')
C = tvm.compute(A.shape, lambda i: A[i] + B[i])
s = tvm.create_schedule(C.op)
tvm.lower(s, [A, B, C], simple_mode=True)
def test_mod(mod, size):
a, b = (np.random.normal(size=size).astype('float32') for _ in range(2))
c = np.empty(size, dtype='float32')
print("c shape: {}".format(c.shape))
eval_mod(mod, a, b, out=c)
np.testing.assert_equal(c, a + b)
mod = tvm.build(s, [A, B, C])
test_mod(mod, 5)
test_mod(mod, 1000)
def tvm_vector_add(ndim):
A = tvm.placeholder([tvm.var() for _ in range(ndim)])
B = tvm.placeholder(A.shape)
C = tvm.compute(A.shape, lambda *i: A[i] + B[i])
s = tvm.create_schedule(C.op)
return tvm.build(s, [A, B, C])
mod = tvm_vector_add(2)
test_mod(mod, (2, 2))
mod = tvm_vector_add(4)
test_mod(mod, (2, 3, 4, 5))
运行结果如下:
3.矩阵转置