Fix paddle.vision.ops.deform_conv2d API big Tensor

[xingmingyyj]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

问题修复

1. 修复int溢出导致的访存越界问题
2. 修复 CUDA error(9)问题
3. batch较大时存在精度diff

paddle.vision.ops.deform_conv2d(Tensor([5070448, 3, 5, 5],"float32"), Tensor([5070448, 18, 5, 5],"float32"), Tensor([6, 3, 3, 3],"float32"), None, list[1,1,], list[1,1,], list[1,1,], 1, 1, Tensor([5070448, 9, 5, 5],"float32"), ) Not equal to tolerance rtol=0.01, atol=0.01 Tensor-likes are not close! Mismatched elements: 3 / 2281701600 (0.0%) Greatest absolute difference: 0.04030437394976616 at index (1852051, 0, 0, 2) (up to 0.01 allowed) Greatest relative difference: 1.0 at index (910808, 0, 0, 3) (up to 0.01 allowed) ACTUAL: (shape=torch.Size([5070448, 18, 5, 5]), dtype=torch.float32) tensor([[[[ 0.0000e+00, 0.0000e+00, -3.4506e-02, -8.3038e-04, 0.0000e+00], [-8.7802e-04, 3.7523e-03, 1.5622e-02, -8.7787e-04, 5.6141e-03], 

offset的反向传播公式为：
$$\frac{\partial L}{\partial \Delta p_n} = \sum_{p_0} \frac{\partial L}{\partial y(p_0)} \cdot w(p_n) \cdot \frac{\partial x (p_0 + p_n + \Delta p_n)}{\partial \Delta p_n}$$
其中：

• $p_0$:表示采样点。
• $p_n$:表示相对位置。
• $\Delta p_n$:表示偏移量。
• $\frac{\partial x (p_0 + p_n + \Delta p_n)}{\partial \Delta p_n}$:表示双线性插值函数的梯度。
  paddle和torch在计算双线性插值反向时有diff。当点$p_0 + p_n + \Delta p_n$越界时，例如，$p_0 + p_n + \Delta p_n=(-1,0)$, paddle前向直接返回0，反向直接返回0。torch前向直接返回0，但是会给周围没有越界的点回传梯度。对齐torch实现后，精度可以对齐。这里保持了paddle的实现。paddle代码实现位置：https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/phi/kernels/impl/deformable_conv_grad_kernel_impl.h#L62
  ；torch代码实现位置：https://github.com/pytorch/vision/blob/main/torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu#L494。

4. 排除上述问题后，image_shape较大时仍然存在精度diff

test begin: paddle.vision.ops.deform_conv2d(x=Tensor([4, 22817014, 5, 5],"float32"), offset=Tensor([4, 18, 3, 3],"float32"), weight=Tensor([5, 22817014, 3, 3],"float32"), bias=None, stride=list[1,1,], padding=list[0,0,], dilation=list[1,1,], deformable_groups=1, groups=1, mask=Tensor([4, 9, 3, 3],"float32"), ) [accuracy error] backward paddle.vision.ops.deform_conv2d(x=Tensor([4, 22817014, 5, 5],"float32"), offset=Tensor([4, 18, 3, 3],"float32"), weight=Tensor([5, 22817014, 3, 3],"float32"), bias=None, stride=list[1,1,], padding=list[0,0,], dilation=list[1,1,], deformable_groups=1, groups=1, mask=Tensor([4, 9, 3, 3],"float32"), ) Not equal to tolerance rtol=0.01, atol=0.01 Tensor-likes are not close! Mismatched elements: 1833583996 / 2281701400 (80.4%) Greatest absolute difference: 1.451427698135376 at index (3, 22517627, 2, 1) (up to 0.01 allowed) Greatest relative difference: inf at index (1, 13256072, 1, 4) (up to 0.01 allowed) ACTUAL: (shape=torch.Size([4, 22817014, 5, 5]), dtype=torch.float32) tensor([[[[-1.1867e-02, -9.4617e-02, -6.7302e-02, 4.5707e-02, 1.0042e-01], [-1.4820e-03, 5.5286e-03, 2.1573e-02, -7.3028e-03, 7.3889e-02], 

怀疑是Matmul导致，Matmul修复后进行验证。

性能统计

PR #74058 中的修复全部使用int64_t作为index,在小shape下会造成性能降低，已revert。

Input Config	Stage	修复前 (s)	修复后 (s)	Pre Fix (PR 74058)(s)
Config1	Forward	1.1488	1.1348	1.1714
	Backward	2.5604	2.5957	2.8967
Config2	Forward	0.0890	0.0805	0.0888
	Backward	0.2128	0.2000	0.2144
Config3	Forward	0.0530	0.0531	0.0603
	Backward	0.1793	0.1802	0.2098

Config1: paddle.vision.ops.deform_conv2d(Tensor([79220, 3, 4, 4],"float32"), Tensor([79220, 18, 4, 4],"float32"), Tensor([6, 3, 3, 3],"float32"), None, list[2,2,], list[3,3,], list[1,1,], 1, 1, None, )
Config2: paddle.vision.ops.deform_conv2d(Tensor([5070, 3, 5, 5],"float32"), Tensor([5070, 18, 5, 5],"float32"), Tensor([6, 3, 3, 3],"float32"), None, list[2,2,], list[3,3,], list[1,1,], 1, 1, None, )
Config3: paddle.vision.ops.deform_conv2d(x=Tensor([446, 128, 200, 200],"float32"), offset=Tensor([446, 36, 100, 100],"float32"), weight=Tensor([128, 128, 3, 3],"float32"), bias=None, stride=list[2,2,], padding=list[1,1,], dilation=list[1,1,], deformable_groups=2, groups=1, mask=Tensor([446, 18, 100, 100],"float32"), )
Pcard-73263

[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.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 16.66667% with 85 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@deddb2b). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...dle/phi/kernels/cpu/deformable_conv_grad_kernel.cc	0.00%	49 Missing ⚠️
...hi/kernels/impl/deformable_conv_grad_kernel_impl.h	0.00%	36 Missing ⚠️

❌ Your patch status has failed because the patch coverage (16.66%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@ Coverage Diff @@ ## develop #74324 +/- ## ========================================== Coverage ? 16.66% ========================================== Files ? 3 Lines ? 102 Branches ? 0 ========================================== Hits ? 17 Misses ? 85 Partials ? 0 

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[xingmingyyj]

/re-run all-failed

[wanghuancoder]

LGTM