def test_cublasSdot(self):
x = np.random.rand(5).astype(np.float32)
x_gpu = gpuarray.to_gpu(x)
y = np.random.rand(5).astype(np.float32)
y_gpu = gpuarray.to_gpu(y)
result = cublas.cublasSdot(self.cublas_handle, x_gpu.size, x_gpu.gpudata, 1,
y_gpu.gpudata, 1)
assert np.allclose(result, np.dot(x, y))
def forward(self, vector1, vector2):
with torch.cuda.device_of(vector1):
output = vector1.new(1)
handle = torch.cuda.current_blas_handle()
stream = torch.cuda.current_stream()
cublas.cublasSetStream(handle, stream)
if isinstance(vector1, torch.cuda.FloatTensor):
result = cublas.cublasSdot(handle, vector1.numel(),
vector1.data_ptr(), 1,
vector2.data_ptr(), 1)
elif isinstance(vector1, torch.cuda.DoubleTensor):
result = cublas.cublasDdot(handle, vector1.numel(),
vector1.data_ptr(), 1,
vector2.data_ptr(), 1)
output = output.fill_(float(result))
self.save_for_backward(vector1, vector2)
return output
# (In contrast, if you were using a vector from a column in a row-wise matrix, you would set the stride to the width of the matrix.)
# We then put in the pointer to the y_gpu array, and set its stride to 1 as well
#We can now use the cublasSaxpy function. The S stands for single precision, which is what we will need since we are working with 32-bit floating point arrays:
cublas.cublasSaxpy(cublas_context_h, x_gpu.size, a, x_gpu.gpudata, 1,
y_gpu.gpudata, 1)
print(y_gpu.get())
print 'This is close to the NumPy approximation: %s' % np.allclose(
a * x + y, y_gpu.get())
w_gpu = gpuarray.to_gpu(x)
v_gpu = gpuarray.to_gpu(y)
#perform a dot product
dot_output = cublas.cublasSdot(cublas_context_h, v_gpu.size, v_gpu.gpudata, 1,
w_gpu.gpudata, 1)
print(dot_output)
l2_output = cublas.cublasSnrm2(cublas_context_h, v_gpu.size, v_gpu.gpudata, 1)
print(l2_output)
cublas.cublasDestroy(cublas_context_h)
#(f we want to operate on arrays of 64-bit real floating point values, (float64 in NumPy and PyCUDA), then we would use the cublasDaxpy)
"""Level-2 GEMV (general matrix-vector)"""
# m and n are the number of rows and columns
m = 10
n = 100