python - GPU is not improving (cudamat) -

the following code calculates euclidean distance between x points of n dimensions , w fixed points (in same space, n dimensions)

there 10.000 x points, , 14 w points.

distance method calls made in batches (of 2.000 points, , 200 or 20 points)

distance operations involves matrixes multiplication of size (batchsizexn) x (nx14). involves element wise pow operations, , cols/rows sums.

using gpu has overcost of copying both matrixes pc ram nvidia ram. expected in space high dimensionality (n = 1.000 dimensions) reverted time execution improvement.

but attached code did bad results:

• with n=64 dimensions gpu 0.2 times faster (cpu 5 times faster gpu)
• with n=1000 dimensions gpu , cpu has same results
• with n=10000 dimensions gpu 1.2 times faster gpu
• with n=100000 dimensions gpu 0.85 times faster gpu

the cpu intel i7 , gpu nvidia 750ti. i'm using ubuntu dedicated server. think x windows not using graphic card. connect vía vnc / ssh

the best improvement of 1.2 versus 1 thread of 1 i7 core.

i'm newbie on gpus. have better results welcomed.

thanks in advance

import numpy np import cudamat cm  def distance(self, x, usegpu=false):     if usegpu:         print("using gpu...")         w=self.w         gpu_w = cm.cudamatrix(w)         gpu_x = cm.cudamatrix(x)         gpu_d = cm.empty ((w.shape[0] , x.shape[0]) )         gpu_x2 = cm.empty (x.shape)         gpu_w2 = cm.empty (w.shape)          cm.pow(gpu_x, 2, target=gpu_x2)         gpu_x2 = gpu_x2.sum(axis=1)         cm.pow(gpu_w, 2, target=gpu_w2)         gpu_w2 = gpu_w2.sum(axis=1)          gpu_d = cm.dot(gpu_w, gpu_x.t)         gpu_d = gpu_d.mult((-2))         gpu_d = gpu_d.add_col_vec(gpu_w2)         gpu_d = gpu_d.add_row_vec(gpu_x2.transpose())          d_t=gpu_d.transpose().asarray()          return d_t     else:         w = self.w           x2 = (x**2).sum(1)[:, none]         d = -2*np.dot(w, x.t) + (w**2).sum(1)[:, none] + x2.t         return d.t