cuBLAS Dgemm product with python

1
votes

I have 2 simple matrices A and B and I'm calculating their multiplication. The arrays looks like this (using numpy as as mockup)

A=np.array(([1,2,3],[4,5,6])).astype(np.float64)
B=np.array(([7,8],[9,10],[11,12])).astype(np.float64)

Here are the shapes of the Matrix

A: (2, 3)

B: (3, 2)

Now, I am trying to do this using cublasDgemmBatched to get the product.

I am confused on what my m,n,and k values should be when applying cublasDgemmBatched. Also, I'm not sure what my leading dimension (lda, ldb, ldc) of the array would be.

There is a nice 3d example here but I can't seem to get this function to work on 2d matrices.

Ideally, i would like to get the same results as np.dot.

2
pythonmatrixcudapycudacublas
It might be because the batched gemm is for an array of two dimensional matrices (or what you have called three dimensional matrices). If you really insist on using gemmbatched for a two dimensional array then for your example: m = 2, n = 2, k = 3, lda = 3, ldb = 2, ldc = 2. Although I cannot say with confidence whether or not this will work. Or, referencing the example you linked to, just set l=1. - inJeans
You might want to be a little more explicit about what you mean by the dot product of 2 matrices. matlab's definition indicates a set of dot products of the matrix columns. The implication is the matrices would have the same shape. If you indeed want a set of individual dot products of matrix columns, using batched gemm is probably not the way to go, since it is fairly compilicated to set up and using a matrix product for a vector dot-product is overkill I think - Robert Crovella
I would like to get the same results as np.dot() - NinjaGaiden
Then you probably want to use cublasDgemm. Adjusting the 3D example to fit your 2D one might look like this cublasDgemm(self.cublas_handle, 'n','n', 2, 2, 3, alpha, b_arr.gpudata, 2, a_arr.gpudata, 3, beta, c_arr.gpudata, 2) - inJeans
Thankyou. I would like to mark your response as an answer but I am not sure how. - NinjaGaiden

2 Answers

1
votes

I don't have skcuda.blas to confirm this. But a more complete example might look like

A = np.array(([1, 2, 3], [4, 5, 6])).astype(np.float64)
B = np.array(([7, 8], [9, 10], [11, 12])).astype(np.float64)

m, k = A.shape
k, n = B.shape

a_gpu = gpuarray.to_gpu(A)
b_gpu = gpuarray.to_gpu(B)
c_gpu = gpuarray.empty((m, n), np.float64)

alpha = np.float64(1.0)
beta = np.float64(0.0)

a_arr = bptrs(a_gpu)
b_arr = bptrs(b_gpu)
c_arr = bptrs(c_gpu)

cublas_handle = cublas.cublasCreate()

cublas.cublasDgemm(cublas_handle, 'n','n',
                   n, m, k, alpha,
                   b_arr.gpudata, m,
                   a_arr.gpudata, k,
                   beta, c_arr.gpudata, m)
0
votes

A very simple way to imitate np.dot() is using culinalg.dot() which uses cuBLAS behind, see skcuda.linalg.dot. Below, a simple example:

import pycuda.autoinit
import pycuda.gpuarray as gpuarray
import pycuda.driver as drv
import numpy as np

import skcuda.linalg as culinalg
import skcuda.misc as cumisc
culinalg.init()

A = np.array(([1, 2, 3], [4, 5, 6])).astype(np.float64)
B = np.array(([7, 8, 1, 5], [9, 10, 0, 9], [11, 12, 5, 5])).astype(np.float64)

A_gpu = gpuarray.to_gpu(A)
B_gpu = gpuarray.to_gpu(B)

C_gpu = culinalg.dot(A_gpu, B_gpu)

print(np.dot(A, B))
print(C_gpu)