Skip to content

【Hackathon 9th Sprint No.06】NO.06 paddle.nn.MaxPool2D兼容性增强 #76457

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
WintersMontagne10335 wants to merge 16 commits into from
Closed

【Hackathon 9th Sprint No.06】NO.06 paddle.nn.MaxPool2D兼容性增强 #76457

Show file tree
Hide file tree
Changes from 9 commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
e88abac
Enhanced compatibility for paddle.nn.MaxPool2D
WintersMontagne10335 Nov 18, 2025
95fdadc
Add code for the CPU part
WintersMontagne10335 Nov 21, 2025
c85f6c9
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
WintersMontagne10335 Nov 21, 2025
b45adb9
Add the code for the impl section
WintersMontagne10335 Nov 21, 2025
c9a223a
Add the code for the GPU part
WintersMontagne10335 Nov 23, 2025
b6098cd
Add the code for the YAML part
WintersMontagne10335 Nov 23, 2025
0c6a3c9
Update the Python part
WintersMontagne10335 Nov 23, 2025
66d181e
Fix bugs related to compilation
WintersMontagne10335 Nov 23, 2025
5304609
Fix bugs related to codestyle
WintersMontagne10335 Nov 23, 2025
e811447
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
WintersMontagne10335 Nov 24, 2025
0f2d670
update
WintersMontagne10335 Nov 26, 2025
313ccfb
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
WintersMontagne10335 Nov 26, 2025
bf675b4
update
WintersMontagne10335 Nov 27, 2025
9eb6533
Merge remote-tracking branch 'upstream/develop' into hackathon9th06
WintersMontagne10335 Nov 27, 2025
12609f9
update
WintersMontagne10335 Nov 28, 2025
9b7bff1
Merge remote-tracking branch 'upstream/develop' into hackathon9th06
WintersMontagne10335 Nov 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions paddle/fluid/pir/drr/src/ir_operation_factory.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -448,6 +448,11 @@ void OperationFactory::RegisterManualOpCreator() {
common::errors::InvalidArgument(
"'paddings' Attribute is expected for "
"Pool2dOp. "));
PADDLE_ENFORCE_EQ(attrs.find("dilations") != attrs.end(),
true,
common::errors::InvalidArgument(
"'dilations' Attribute is expected for "
"Pool2dOp. "));
std::vector<int64_t> paddings;
for (size_t i = 0;
i < attrs.at("paddings").dyn_cast<pir::ArrayAttribute>().size();
Expand All @@ -459,6 +464,17 @@ void OperationFactory::RegisterManualOpCreator() {
.data());
}

std::vector<int64_t> dilations;
for (size_t i = 0;
i < attrs.at("dilations").dyn_cast<pir::ArrayAttribute>().size();
i++) {
dilations.push_back(attrs.at("dilations")
.dyn_cast<pir::ArrayAttribute>()
.at(i)
.dyn_cast<pir::Int64Attribute>()
.data());
}

PADDLE_ENFORCE_EQ(attrs.find("ceil_mode") != attrs.end(),
true,
common::errors::InvalidArgument(
Expand Down Expand Up @@ -554,6 +570,7 @@ void OperationFactory::RegisterManualOpCreator() {
inputs[1],
strides,
paddings,
dilations,
ceil_mode,
exclusive,
data_format,
Expand Down
188 changes: 158 additions & 30 deletions paddle/phi/kernels/funcs/pooling.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,7 @@ class Pool2dFunctor<CPUContext, PoolProcess, T> {
const std::vector<int64_t>& ksize,
const std::vector<int64_t>& strides,
const std::vector<int64_t>& paddings,
const std::vector<int64_t>& dilations,
Copy link
Contributor

@zhwesky2010 zhwesky2010 Nov 24, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

infermeta、infersy_mbolic 没有添加。单测应该是过不了的。

Copy link
Contributor Author

@WintersMontagne10335 WintersMontagne10335 Nov 28, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

研发大大您好，新增的 dilations 不影响 shape，infermeta、infer_symbolic 用之前就可以了

const std::string data_format,
bool exclusive,
bool adaptive,
Expand Down Expand Up @@ -65,6 +66,9 @@ class Pool2dFunctor<CPUContext, PoolProcess, T> {
const int64_t padding_height = paddings[0];
const int64_t padding_width = paddings[1];

const int64_t dilation_height = dilations[0];
const int64_t dilation_width = dilations[1];

const T* input_data = input.data<T>();
T* output_data = context.template Alloc<T>(output);

Expand Down Expand Up @@ -94,15 +98,19 @@ class Pool2dFunctor<CPUContext, PoolProcess, T> {
std::min(wstart + ksize_width, input_width + padding_width);
pool_size = (hend - hstart) * (wend - wstart);

wstart = std::max(wstart, static_cast<int64_t>(0));
hstart = std::max(hstart, static_cast<int64_t>(0));
hend = std::min(hend, input_height);
wend = std::min(wend, input_width);
hend = hstart + (ksize_height - 1) * dilation_height + 1;
wend = wstart + (ksize_width - 1) * dilation_width + 1;
while (hstart < static_cast<int64_t>(0))
hstart += dilation_height;
while (hend > input_height) hend -= dilation_height;
while (wstart < static_cast<int64_t>(0))
wstart += dilation_width;
while (wend > input_width) wend -= dilation_width;
}

T ele = pool_process.initial();
for (int64_t h = hstart; h < hend; ++h) {
for (int64_t w = wstart; w < wend; ++w) {
for (int64_t h = hstart; h < hend; h += dilation_height) {
for (int64_t w = wstart; w < wend; w += dilation_width) {
pool_process.compute(input_data[h * input_width + w], &ele);
}
}
Expand Down Expand Up @@ -142,14 +150,18 @@ class Pool2dFunctor<CPUContext, PoolProcess, T> {
std::min(wstart + ksize_width, input_width + padding_width);
pool_size = (hend - hstart) * (wend - wstart);

wstart = std::max(wstart, static_cast<int64_t>(0));
hstart = std::max(hstart, static_cast<int64_t>(0));
hend = std::min(hend, input_height);
wend = std::min(wend, input_width);
hend = hstart + (ksize_height - 1) * dilation_height + 1;
wend = wstart + (ksize_width - 1) * dilation_width + 1;
while (hstart < static_cast<int64_t>(0))
hstart += dilation_height;
while (hend > input_height) hend -= dilation_height;
while (wstart < static_cast<int64_t>(0))
wstart += dilation_width;
while (wend > input_width) wend -= dilation_width;
}
T ele = pool_process.initial();
for (int64_t h = hstart; h < hend; ++h) {
for (int64_t w = wstart; w < wend; ++w) {
for (int64_t h = hstart; h < hend; h += dilation_height) {
for (int64_t w = wstart; w < wend; w += dilation_width) {
pool_process.compute(
input_data[h * input_width * input_channels +
w * input_channels + c],
Expand Down Expand Up @@ -189,6 +201,7 @@ class Pool2dGradFunctor<CPUContext, PoolProcess, T> {
const std::vector<int64_t>& ksize,
const std::vector<int64_t>& strides,
const std::vector<int64_t>& paddings,
const std::vector<int64_t>& dilations UNUSED,
const std::string data_format,
bool exclusive,
bool adaptive,
Expand Down Expand Up @@ -355,6 +368,7 @@ class MaxPool2dGradFunctor<CPUContext, T> {
const std::vector<int64_t>& ksize,
const std::vector<int64_t>& strides,
const std::vector<int64_t>& paddings,
const std::vector<int64_t>& dilations,
const std::string data_format,
DenseTensor* input_grad) {
bool channel_last = (data_format == "NHWC");
Expand Down Expand Up @@ -384,6 +398,9 @@ class MaxPool2dGradFunctor<CPUContext, T> {
const int64_t padding_height = paddings[0];
const int64_t padding_width = paddings[1];

const int64_t dilation_height = dilations[0];
const int64_t dilation_width = dilations[1];

const T* input_data = input.data<T>();
const T* output_data = output.data<T>();
const T* output_grad_data = output_grad.data<T>();
Expand All @@ -396,16 +413,21 @@ class MaxPool2dGradFunctor<CPUContext, T> {
for (int64_t c = 0; c < output_channels; ++c) {
for (int64_t ph = 0; ph < output_height; ++ph) {
int64_t hstart = ph * stride_height - padding_height;
int64_t hend = std::min(hstart + ksize_height, input_height);
hstart = std::max(hstart, static_cast<int64_t>(0));
int64_t hend = hstart + (ksize_height - 1) * dilation_height + 1;
while (hstart < static_cast<int64_t>(0)) hstart += dilation_height;
while (hend > input_height) hend -= dilation_height;

for (int64_t pw = 0; pw < output_width; ++pw) {
int64_t wstart = pw * stride_width - padding_width;
int64_t wend = std::min(wstart + ksize_width, input_width);
wstart = std::max(wstart, static_cast<int64_t>(0));
int64_t wend = wstart + (ksize_width - 1) * dilation_width + 1;
while (wstart < static_cast<int64_t>(0)) wstart += dilation_width;
while (wend > input_width) wend -= dilation_width;

bool stop = false;
for (int64_t h = hstart; h < hend && !stop; ++h) {
for (int64_t w = wstart; w < wend && !stop; ++w) {
for (int64_t h = hstart; h < hend && !stop;
h += dilation_height) {
for (int64_t w = wstart; w < wend && !stop;
w += dilation_width) {
int64_t input_idx = h * input_width + w;
int64_t output_idx = ph * output_width + pw;
if (input_data[input_idx] == output_data[output_idx]) {
Expand All @@ -430,16 +452,20 @@ class MaxPool2dGradFunctor<CPUContext, T> {
for (int64_t c = 0; c < output_channels; ++c) {
for (int64_t ph = 0; ph < output_height; ++ph) {
int64_t hstart = ph * stride_height - padding_height;
int64_t hend = std::min(hstart + ksize_height, input_height);
hstart = std::max(hstart, static_cast<int64_t>(0));
int64_t hend = hstart + (ksize_height - 1) * dilation_height + 1;
while (hstart < static_cast<int64_t>(0)) hstart += dilation_height;
while (hend > input_height) hend -= dilation_height;
for (int64_t pw = 0; pw < output_width; ++pw) {
int64_t wstart = pw * stride_width - padding_width;
int64_t wend = std::min(wstart + ksize_width, input_width);
wstart = std::max(wstart, static_cast<int64_t>(0));
int64_t wend = wstart + (ksize_width - 1) * dilation_width + 1;
while (wstart < static_cast<int64_t>(0)) wstart += dilation_width;
while (wend > input_width) wend -= dilation_width;

bool stop = false;
for (int64_t h = hstart; h < hend && !stop; ++h) {
for (int64_t w = wstart; w < wend && !stop; ++w) {
for (int64_t h = hstart; h < hend && !stop;
h += dilation_height) {
for (int64_t w = wstart; w < wend && !stop;
w += dilation_width) {
int64_t input_idx =
h * input_width * input_channels + w * input_channels + c;
int64_t output_idx = ph * output_width * output_channels +
Expand Down Expand Up @@ -1103,6 +1129,102 @@ template class Pool3dGradFunctor<phi::CPUContext,
template class Pool3dGradFunctor<phi::CPUContext,
AvgPoolGrad<dtype::float16>,
dtype::float16>;

// native版本，先不删
// /*
// * All tensors are in NCHW format.
// * Ksize, strides, paddings, dilations are two elements. These two elements
// represent
// * height and width, respectively.
// */
// template <typename T1, typename T2>
// class MaxPool2dWithIndexFunctor<CPUContext, T1, T2> {
// public:
// void operator()(const CPUContext& context,
// const DenseTensor& input,
// const std::vector<int64_t>& ksize,
// const std::vector<int64_t>& strides,
// const std::vector<int64_t>& paddings,
// const std::vector<int64_t>& dilations,
// bool adaptive,
// DenseTensor* output,
// DenseTensor* mask) {
// const int64_t batch_size = input.dims()[0];
// const int64_t input_height = input.dims()[2];
// const int64_t input_width = input.dims()[3];
// const int64_t output_channels = output->dims()[1];
// const int64_t output_height = output->dims()[2];
// const int64_t output_width = output->dims()[3];
// const int64_t ksize_height = ksize[0];
// const int64_t ksize_width = ksize[1];
// const int64_t stride_height = strides[0];
// const int64_t stride_width = strides[1];
// const int64_t padding_height = paddings[0];
// const int64_t padding_width = paddings[1];
// const int64_t dilation_height = dilations[0];
// const int64_t dilation_width = dilations[1];
// const int64_t input_stride = input_height * input_width;
// const int64_t output_stride = output_height * output_width;

// const T1* input_data = input.data<T1>();
// T1* output_data = context.template Alloc<T1>(output);
// T2* mask_data = context.template Alloc<T2>(mask);

// int64_t hstart = 0, hend = 0;
// int64_t wstart = 0, wend = 0;
// for (int64_t i = 0; i < batch_size; i++) {
// for (int64_t c = 0; c < output_channels; ++c) {
// for (int64_t ph = 0; ph < output_height; ++ph) {
// if(adaptive){
// hstart = AdaptStartIndex(ph, input_height, output_height);
// hend = AdaptEndIndex(ph, input_height, output_height);
// }else{
// hstart = ph * stride_height - padding_height;
// }
// for(int64_t pw=0;pw<output_width;++pw){
// T1 ele = static_cast<T1>(-FLT_MAX);
// int64_t index = -1;
// if(adaptive){
// wstart = AdaptStartIndex(pw, input_width, output_width);
// wend = AdaptEndIndex(pw, input_width, output_width);
// for (int64_t h = hstart; h < hend; ++h) {
// for (int64_t w = wstart; w < wend; ++w) {
// if (ele < input_data[h * input_width + w]) {
// ele = input_data[h * input_width + w];
// index = h * input_width + w;
// }
// }
// }
// }else{
// wstart=pw*stride_width-padding_width;
// for (int64_t kh=0;kh<ksize_height;++kh){
// int64_t h=hstart+kh*dilation_height;
// if (h < 0 || h >= input_height)
// continue;
// for(int64_t kw=0;kw<ksize_width;++kw){
// int64_t w=wstart+kw*dilation_width;
// if(w<0||w>=input_width)
// continue;
// if (ele < input_data[h * input_width + w]) {
// ele = input_data[h * input_width + w];
// index = h * input_width + w;
// }
// }
// }
// }
// output_data[ph * output_width + pw] = ele;
// mask_data[ph * output_width + pw] = index;
// }
// }
// // offset
// input_data += input_stride;
// output_data += output_stride;
// mask_data += output_stride;
// }
// }
// }
// };

/*
* All tensors are in NCHW format.
* Ksize, strides, paddings are two elements. These two elements represent
Expand All @@ -1116,6 +1238,7 @@ class MaxPool2dWithIndexFunctor<CPUContext, T1, T2> {
const std::vector<int64_t>& ksize,
const std::vector<int64_t>& strides,
const std::vector<int64_t>& paddings,
const std::vector<int64_t>& dilations,
bool adaptive,
DenseTensor* output,
DenseTensor* mask) {
Expand All @@ -1131,6 +1254,8 @@ class MaxPool2dWithIndexFunctor<CPUContext, T1, T2> {
const int64_t stride_width = strides[1];
const int64_t padding_height = paddings[0];
const int64_t padding_width = paddings[1];
const int64_t dilation_height = dilations[0];
const int64_t dilation_width = dilations[1];
const int64_t input_stride = input_height * input_width;
const int64_t output_stride = output_height * output_width;

Expand All @@ -1148,23 +1273,25 @@ class MaxPool2dWithIndexFunctor<CPUContext, T1, T2> {
hend = AdaptEndIndex(ph, input_height, output_height);
} else {
hstart = ph * stride_height - padding_height;
hend = std::min(hstart + ksize_height, input_height);
hstart = std::max(hstart, static_cast<int64_t>(0));
hend = hstart + (ksize_height - 1) * dilation_height + 1;
while (hstart < static_cast<int64_t>(0)) hstart += dilation_height;
while (hend > input_height) hend -= dilation_height;
}
for (int64_t pw = 0; pw < output_width; ++pw) {
if (adaptive) {
wstart = AdaptStartIndex(pw, input_width, output_width);
wend = AdaptEndIndex(pw, input_width, output_width);
} else {
wstart = pw * stride_width - padding_width;
wend = std::min(wstart + ksize_width, input_width);
wstart = std::max(wstart, static_cast<int64_t>(0));
wend = wstart + (ksize_width - 1) * dilation_width + 1;
while (wstart < static_cast<int64_t>(0)) wstart += dilation_width;
while (wend > input_width) wend -= dilation_width;
}

T1 ele = static_cast<T1>(-FLT_MAX);
int64_t index = -1;
for (int64_t h = hstart; h < hend; ++h) {
for (int64_t w = wstart; w < wend; ++w) {
for (int64_t h = hstart; h < hend; h += dilation_height) {
for (int64_t w = wstart; w < wend; w += dilation_width) {
if (ele < input_data[h * input_width + w]) {
ele = input_data[h * input_width + w];
index = h * input_width + w;
Expand Down Expand Up @@ -1198,6 +1325,7 @@ class MaxPool2dWithIndexGradFunctor<CPUContext, T1, T2> {
const std::vector<int64_t>& ksize UNUSED,
const std::vector<int64_t>& strides UNUSED,
const std::vector<int64_t>& paddings UNUSED,
const std::vector<int64_t>& dilations UNUSED,
bool adaptive UNUSED,
DenseTensor* input_grad) {
const int64_t batch_size = input_grad->dims()[0];
Expand Down
Loading
Loading