[Sparse conv] Implement implicit gemm algo for SubmConv3D

[Wong4j]

PR Category

Performance Optimization

PR Types

Performance

Description

Support implicit GEMM algorithm for SubmConv3D.

Usage:

• nn.functional.subm_conv3d_igemm

# 3D
y = paddle.sparse.nn.functional.subm_conv3d(x, weight, key='key1', padding=[0, 1, 1])   # original
y = paddle.sparse.nn.functional.subm_conv3d_igemm(x, weight, key='key2', padding=[0, 1, 1])   # use implicit gemm

# 2D
y = paddle.sparse.nn.functional.subm_conv2d(x, weight, key='key1', padding=[1, 1])   # original
y = paddle.sparse.nn.functional.subm_conv2d_igemm(x, weight, key='key2', padding=[1, 1])   # use implicit gemm

• nn.SubmConv3D

# 3D
paddle.nn.SubmConv3D(32, 32, kernel_size=[1, 3, 3], key="key1")  # original
paddle.nn.SubmConv3D(32, 32, kernel_size=[1, 3, 3], key="key2", backend="igemm")  # use implicit gemm

# 2D
paddle.nn.SubmConv2D(32, 32, kernel_size=[3, 3], key="key1")  # original
paddle.nn.SubmConv2D(32, 32, kernel_size=[3, 3], key="key2", backend="igemm")  # use implicit gemm

Perf:
GPU: 3080
Prec: FP16
case: single SubmConv
nnz=214202 dense_shape=[1, 1, 4608, 4608, 32] kernel_size=[1, 3, 3] stride=1 in_channel=out_channel=32
(This perf numbers do not include the overhead of hashmap/rulebook creation, which I assume has been cached.)

---	cutlass	igemm
time(us)	703	194

Note:

• Implicit gemm only supports forward now.
• I have only verified the correctness for SubmConv, so I'm asserting subm==True and stride==1 and dilation==1 in the code.
• The cuda kernels are modified based on the torchsparse's implementation.
• The input must be 3D (NDHWC), and kernel must has dims=3. For 2D case, please insert zeros to to the D dimention of indices and set kernel sizes to (1, 3, 3).
• The input can be 3D (NDHWC) or 2D (NHWC)

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[paddle-ci-bot]

Sorry to inform you that ee0d63d's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[Wangzheee]

ci的代码没有覆盖到这个OP，可以针对这个OP增加单测

[Wong4j]

Done

[Wangzheee]

单测中没有执行这个析构，可以增加一下

[Wong4j]

我增加了单测，本地跑单测会跑到这个析构，但是CI仍然显示没有跑到。

[paddle-ci-bot]

Sorry to inform you that 4851677's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[qingqing01]

后续需要更新中文文档

[qingqing01]

对用户暴露的接口需增加注释

[jzhang533]

the newly introduced arg backend should be documented in the docstring.

[jzhang533]

LGTM