fix out_of_range bug of multinomial op's cuda kernel

[pangyoki]

PR types

Bug fixes

PR changes

OPs

Describe

A bug in sample method (multinomial op) of Categorical API is mentioned in issue #36401.

issue

• test case

import paddle
with paddle.no_grad():
    actor=paddle.nn.Sequential(paddle.nn.Linear(20, 2800))
    logits=actor(paddle.rand([20]))
    cat=paddle.distribution.Categorical(logits.exp())
    print(cat.sample([1]))

• error message

Error: /paddle/paddle/fluid/operators/multinomial_op.cu:42 Assertion `in_data[id] >= 0.0` failed. The input of multinomial distribution should be >= 0, but got -0.038249.
（省略类似错误）
Error: /paddle/paddle/fluid/operators/multinomial_op.cu:42 Assertion `in_data[id] >= 0.0` failed. The input of multinomial distribution should be >= 0, but got -0.044257.
Traceback (most recent call last):
  File "bug.py", line 6, in <module>
    print(cat.sample([1]))
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/distribution.py", line 771, in sample
    sample_index = multinomial(logits, num_samples, True)
  File "/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/tensor/random.py", line 133, in multinomial
    'replacement', replacement)
SystemError: (Fatal) Operator multinomial raises an thrust::system::system_error exception.
The exception content is
:transform: failed to synchronize: cudaErrorLaunchFailure: unspecified launch failure. (at /paddle/paddle/fluid/imperative/tracer.cc:192)

Reason for error

In the cuda kernel implementation, the number of threads exceeding the size of the input array will be used to perform calculations (the reason is that the block size is limited, when setting the grid size, more threads will be set for rounding).

However, when the cuda kernel calculates, it does not limit the array subscripts. As a result, when calculating in the thread, the space exceeding the size of the array is accessed, causing an error.

bug fix

Restrictions on the subscripts of the accessed arrays.

[paddle-bot-old]

✅ This PR's description meets the template requirements!
Please wait for other CI results.

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhiqiu]

LGTM

[paddle-bot-old]

Sorry to inform you that 0641bbc's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[chenwhql]

LGTM for PADDLE_ENFORCE

[Avin0323]

LGTM for PR-CI-OP-benchmark