add flow.rand

[VertexC]

[chengtbf]

mark。暂时移除 automerge，两个小时以后加回来。

[chengtbf]

有冲突，需要解决一下~

[github-actions]

Speed stats:

GPU Name: GeForce GTX 1080 

PyTorch resnet50 time: 140.4ms (= 7018.6ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 128.2ms (= 6412.3ms / 50, input_shape=[16, 3, 224, 224], backward is enabled)
Relative speed: 1.09 (= 140.4ms / 128.2ms)

PyTorch resnet50 time: 83.4ms (= 4167.7ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 74.5ms (= 3723.8ms / 50, input_shape=[8, 3, 224, 224], backward is enabled)
Relative speed: 1.12 (= 83.4ms / 74.5ms)

PyTorch resnet50 time: 57.4ms (= 2870.6ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 48.5ms (= 2423.1ms / 50, input_shape=[4, 3, 224, 224], backward is enabled)
Relative speed: 1.18 (= 57.4ms / 48.5ms)

PyTorch resnet50 time: 50.0ms (= 2500.8ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 39.4ms (= 1969.6ms / 50, input_shape=[2, 3, 224, 224], backward is enabled)
Relative speed: 1.27 (= 50.0ms / 39.4ms)

PyTorch resnet50 time: 37.9ms (= 1892.7ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
OneFlow resnet50 time: 38.8ms (= 1941.1ms / 50, input_shape=[1, 3, 224, 224], backward is enabled)
Relative speed: 0.98 (= 37.9ms / 38.8ms)

[liujuncheng]

这个kernel的block_num和thread_num是根据设备上的硬件相关信息计算的，这里并不能保证block_num*thread_num一定 > elem_cnt，所以这个kernel可能会只初始化tensor的前一部分，后一部分保持不变吧，这里应该用 CUDA_1D_KERNEL_LOOP吧

[hjchen2]

嗯，是的，这里确实有问题。@博文 你参考 oneflow/user/kernels/partial_fc_sample_kernel.cu 文件修改一下。