Skip to content

add adaptive pool and softmax with cross entropy supports different axis, * test = kunlun #44428

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

Merged
merged 3 commits into from
Jul 21, 2022
Merged

add adaptive pool and softmax with cross entropy supports different axis, * test = kunlun #44428

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
931c75a
add xpu pnorm op and fix pool op, *test=kunlun
ykkk2333 Jul 11, 2022
cb23aed
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
ykkk2333 Jul 19, 2022
afbf1a8
add adaptive pool, and softmax with cross entropy supports different …
ykkk2333 Jul 19, 2022
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
1 change: 1 addition & 0 deletions paddle/fluid/operators/p_norm_op_xpu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,7 @@ class P_NormXPUKernel : public framework::OpKernel<T> {
XPUType* zeros = RAII_GUARD.alloc_l3_or_gm<XPUType>(1);
PADDLE_ENFORCE_XDNN_NOT_NULL(zeros);
r = xpu::constant(dev_ctx.x_context(), zeros, 1, 0.0f);
PADDLE_ENFORCE_XDNN_SUCCESS(r, "constant");
std::vector<int> zeros_dim(1, 1);

bool* tmp2_x = RAII_GUARD.alloc_l3_or_gm<bool>(m * t * n);
Expand Down
133 changes: 91 additions & 42 deletions paddle/fluid/operators/pool_op_xpu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,11 +60,7 @@ class PoolXPUKernel : public framework::OpKernel<T> {
2,
platform::errors::InvalidArgument(
"The Pool2d XPU OP only support 2 dimension pooling!"));
PADDLE_ENFORCE_EQ(!adaptive || (ksize[0] * ksize[1] == 1),
true,
platform::errors::InvalidArgument(
"The Pool2d XPU OP does not support (adaptive == "
"true && output_size != 1)"));

int* index_data = nullptr;
bool global_pooling = context.Attr<bool>("global_pooling") ||
(adaptive && (ksize[0] * ksize[1] == 1));
Expand All @@ -80,6 +76,9 @@ class PoolXPUKernel : public framework::OpKernel<T> {
const int in_h = in_x->dims()[2];
const int in_w = in_x->dims()[3];

const int out_h = out->dims()[2];
const int out_w = out->dims()[3];

framework::DDim data_dims;

data_dims = phi::slice_ddim(in_x->dims(), 2, in_x->dims().size());
Expand All @@ -90,45 +89,79 @@ class PoolXPUKernel : public framework::OpKernel<T> {
data_dims,
strides,
ksize);

if (ceil_mode) {
paddings[1] += (strides[0] - 1);
paddings[3] += (strides[1] - 1);
int in_h_ceil = (out_h - 1) * strides[0] + ksize[0] - 2 * paddings[0];
int in_w_ceil = (out_w - 1) * strides[1] + ksize[1] - 2 * paddings[2];

paddings[1] += (in_h_ceil - in_h);
paddings[3] += (in_w_ceil - in_w);
}

auto input = reinterpret_cast<const XPUType*>(in_x->data<T>());
out->mutable_data<T>(context.GetPlace());
auto output = reinterpret_cast<XPUType*>(out->data<T>());
auto& dev_ctx = context.template device_context<DeviceContext>();
int r = xpu::Error_t::SUCCESS;
if (pooling_type == "max") {
r = xpu::max_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
index_data,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
true);
} else if (pooling_type == "avg") {
r = xpu::avg_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
!exclusive,
true);
if (!adaptive) {
if (pooling_type == "max") {
r = xpu::max_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
index_data,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
true);
} else if (pooling_type == "avg") {
r = xpu::avg_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
n,
c,
in_h,
in_w,
ksize,
strides,
paddings,
!exclusive,
true);
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
if (pooling_type == "max") {
r = xpu::adaptive_max_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
index_data,
n,
c,
in_h,
in_w,
out_h,
out_w,
true);
} else if (pooling_type == "avg") {
r = xpu::adaptive_avg_pool2d<XPUType>(dev_ctx.x_context(),
input,
output,
n,
c,
in_h,
in_w,
out_h,
out_w,
true);
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
}
PADDLE_ENFORCE_EQ(r,
xpu::Error_t::SUCCESS,
Expand Down Expand Up @@ -167,11 +200,6 @@ class PoolGradXPUKernel : public framework::OpKernel<T> {
"dimension pooling!, but received "
"%d-dimension pool kernel size",
ksize.size()));
PADDLE_ENFORCE_EQ(!adaptive || (ksize[0] * ksize[1] == 1),
true,
platform::errors::InvalidArgument(
"The Pool2d XPU OP does not support (adaptive == "
"true && output_size != 1)"));
bool global_pooling = context.Attr<bool>("global_pooling") ||
(adaptive && (ksize[0] * ksize[1] == 1));
if (global_pooling) {
Expand All @@ -188,6 +216,16 @@ class PoolGradXPUKernel : public framework::OpKernel<T> {
const int in_h = in_x->dims()[2];
const int in_w = in_x->dims()[3];

const int out_h = out->dims()[2];
const int out_w = out->dims()[3];

PADDLE_ENFORCE_EQ(!adaptive || (ksize[0] * ksize[1] == 1) ||
(in_h % out_h == 0 && in_w % out_w == 0),
true,
platform::errors::InvalidArgument(
"The Pool2d XPU OP does not support (adaptive == "
"true && output_size != 1)"));

framework::DDim data_dims;

data_dims = phi::slice_ddim(in_x->dims(), 2, in_x->dims().size());
Expand All @@ -199,8 +237,11 @@ class PoolGradXPUKernel : public framework::OpKernel<T> {
strides,
ksize);
if (ceil_mode) {
paddings[1] += (strides[0] - 1);
paddings[3] += (strides[1] - 1);
int in_h_ceil = (out_h - 1) * strides[0] + ksize[0] - 2 * paddings[0];
int in_w_ceil = (out_w - 1) * strides[1] + ksize[1] - 2 * paddings[2];

paddings[1] += (in_h_ceil - in_h);
paddings[3] += (in_w_ceil - in_w);
}

auto input = reinterpret_cast<const XPUType*>(in_x->data<T>());
Expand All @@ -210,6 +251,14 @@ class PoolGradXPUKernel : public framework::OpKernel<T> {
auto input_grad = reinterpret_cast<XPUType*>(in_x_grad->data<T>());
auto& dev_ctx = context.template device_context<DeviceContext>();
int r = xpu::Error_t::SUCCESS;
if (adaptive && in_h % out_h == 0 && in_w % out_w == 0) {
strides = {in_h / out_h, in_w / out_w};
int kh = in_h - (out_h - 1) * strides[0];
int kw = in_w - (out_w - 1) * strides[1];
ksize = {kh, kw};
paddings = {0, 0, 0, 0};
}

if (pooling_type == "max") {
r = xpu::max_pool2d_grad<XPUType>(dev_ctx.x_context(),
input,
Expand Down
Loading