[英]Numba CUDA shared memory matrix multiplication
I'm running a shared memory numba code for matrix multiplication, but I think the algorithm to solve it is incorrect as I'm getting incorrect results我正在运行用于矩阵乘法的共享内存 numba 代码,但我认为解决它的算法不正确,因为我得到的结果不正确
I saw another thread for this code but there the answer remained undisclosed and the code did not work我看到了此代码的另一个线程,但答案仍未公开并且代码不起作用
The code is below:代码如下:
# This part is for initializing everything
M = 128
N = 32
a = np.arange(M*N).reshape(M,N).astype(np.int32)
b = np.arange(M*N).reshape(N,M).astype(np.int32)
c = np.zeros((M, M)).astype(np.int32)
d_a = cuda.to_device(a)
d_b = cuda.to_device(b)
d_c = cuda.to_device(c)
block_size = (N,N)
grid_size = (int(M/N),int(M/N))
And this is my kernel definition:这是我的内核定义:
import numpy as np
from numba import cuda, types
@cuda.jit
def fast_matmul(A, B, C):
# Define an array in the shared memory
# The size and type of the arrays must be known at compile time
TPB = N
sA = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
sB = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
x, y = cuda.grid(2)
tx = cuda.threadIdx.x
ty = cuda.threadIdx.y
bpg = cuda.gridDim.x # blocks per grid
if x >= C.shape[0] and y >= C.shape[1]:
# Quit if (x, y) is outside of valid C boundary
return
# Each thread computes one element in the result matrix.
# The dot product is chunked into dot products of TPB-long vectors.
tmp = 0.
for i in range(bpg):
# Preload data into shared memory
sA[tx, ty] = A[x, ty + i * TPB]
sB[tx, ty] = B[tx + i * TPB, y]
# Wait until all threads finish preloading
cuda.syncthreads()
# Computes partial product on the shared memory
for j in range(TPB):
tmp += sA[tx, j] * sB[j, ty]
# Wait until all threads finish computing
cuda.syncthreads()
C[x, y] = tmp
Which I followed from here: https://numba.pydata.org/numba-doc/dev/cuda/examples.html我从这里遵循: https : //numba.pydata.org/numba-doc/dev/cuda/examples.html
However running the code gives me odd results such as然而,运行代码给了我奇怪的结果,比如
x: array([[2147483647, 2147483647, 2147483647, ..., 2147483647, 2147483647,
2147483647],
[2147483647, 2147483647, 2147483647, ..., 2147483647, 2147483647,...
when it should be something like:什么时候应该是这样的:
y: array([[ 1333248, 1333744, 1334240, ..., 1395248, 1395744,
1396240],
[ 3364864, 3366384, 3367904, ..., 3554864, 3556384,...
Could anyone please point out where I am going wrong?谁能指出我哪里出错了?
int32
arrays to a numba kernel that is expecting float32
data.您将int32
数组传递给需要float32
数据的 numba 内核。 Don't do that.不要那样做。x
and y
are.不清楚x
和y
是什么。 Your code produces only one result, and it is in d_c
.您的代码只产生一个结果,它在d_c
。int32
type.我也怀疑你的输入数据会溢出int32
类型。 You should probably convert to float32
throughout.您可能应该始终转换为float32
。 Using arange
makes it hard to verify numerical correctness quickly/visually.使用arange
很难快速/直观地验证数值的正确性。 While you're trying to get things working just use ones
instead.当您试图让事情正常工作时ones
改用它们。Here's a version of what you've show, with the above ideas taken into account:这是您展示的内容的一个版本,并考虑了上述想法:
$ cat t35.py
import numpy as np
from numba import cuda, types, float32
@cuda.jit
def fast_matmul(A, B, C):
# Define an array in the shared memory
# The size and type of the arrays must be known at compile time
TPB = N
sA = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
sB = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
x, y = cuda.grid(2)
tx = cuda.threadIdx.x
ty = cuda.threadIdx.y
bpg = cuda.gridDim.x # blocks per grid
if x >= C.shape[0] and y >= C.shape[1]:
# Quit if (x, y) is outside of valid C boundary
return
# Each thread computes one element in the result matrix.
# The dot product is chunked into dot products of TPB-long vectors.
tmp = 0.
for i in range(bpg):
# Preload data into shared memory
sA[tx, ty] = A[x, ty + i * TPB]
sB[tx, ty] = B[tx + i * TPB, y]
# Wait until all threads finish preloading
cuda.syncthreads()
# Computes partial product on the shared memory
for j in range(TPB):
tmp += sA[tx, j] * sB[j, ty]
# Wait until all threads finish computing
cuda.syncthreads()
C[x, y] = tmp
# This part is for initializing everything
M = 128
N = 32
#a = np.arange(M*N).reshape(M,N).astype(np.float32)
#b = np.arange(M*N).reshape(N,M).astype(np.float32)
a = np.ones(M*N).reshape(M,N).astype(np.float32)
b = np.ones(M*N).reshape(N,M).astype(np.float32)
c = np.zeros((M, M)).astype(np.float32)
d_a = cuda.to_device(a)
d_b = cuda.to_device(b)
d_c = cuda.to_device(c)
block_size = (N,N)
grid_size = (int(M/N),int(M/N))
fast_matmul[grid_size,block_size](d_a, d_b, d_c)
c = d_c.copy_to_host()
print(c)
$ python t35.py
[[32. 32. 32. ... 32. 32. 32.]
[32. 32. 32. ... 32. 32. 32.]
[32. 32. 32. ... 32. 32. 32.]
...
[32. 32. 32. ... 32. 32. 32.]
[32. 32. 32. ... 32. 32. 32.]
[32. 32. 32. ... 32. 32. 32.]]
$
I believe 32 is the correct answer here.我相信 32 是这里的正确答案。
Also note the posted example from which this is derived probably has some errors, see here还要注意从中得出的已发布示例可能有一些错误,请参见此处
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.