Skip to content

optimize batch_norm & pool op for kunlun #30490

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 1 commit into from
Jan 18, 2021
Merged

optimize batch_norm & pool op for kunlun #30490

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
18 changes: 8 additions & 10 deletions paddle/fluid/operators/batch_norm_op_xpu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,16 +139,14 @@ class BatchNormGradXPUKernel : public framework::OpKernel<T> {
auto* dscale_data = dscale->mutable_data<T>(ctx.GetPlace());
auto* dbias_data = dbias->mutable_data<T>(ctx.GetPlace());
auto& dev_ctx = ctx.template device_context<DeviceContext>();
int r = xpu::batch_norm_backward(dev_ctx.x_context(), N, C, H, W, x_data,
dy_data, scale_data, saved_mean_data,
saved_inv_variance_data, dx_data,
dscale_data, dbias_data);
PADDLE_ENFORCE_EQ(
r, XPU_SUCCESS,
platform::errors::External("XPU API(batch_norm_infer_forward) return "
"wrong value[%d], please check whether "
"Baidu Kunlun Card is properly installed.",
r));
int r = xpu::batch_norm_grad<T>(dev_ctx.x_context(), x_data, dy_data,
dx_data, N, C, H, W, scale_data,
saved_mean_data, saved_inv_variance_data,
dscale_data, dbias_data, true);
PADDLE_ENFORCE_EQ(r, XPU_SUCCESS, platform::errors::External(
"XPU API(batch_norm_grad) return "
"wrong value[%d %s]",
r, XPUAPIErrorMsg[r]));
}
};

Expand Down
94 changes: 38 additions & 56 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 @@ -30,6 +30,7 @@ xpu::Pooling_t XPUPoolingType(const std::string& pooltype, bool exclusive,
"Pool op only supports 2D and 3D input."));
}
}

template <typename DeviceContext, typename T>
class PoolXPUKernel : public framework::OpKernel<T> {
public:
Expand All @@ -41,7 +42,6 @@ class PoolXPUKernel : public framework::OpKernel<T> {
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
bool exclusive = context.Attr<bool>("exclusive");
bool is_test = context.Attr<bool>("is_test");
bool adaptive = context.Attr<bool>("adaptive");
PADDLE_ENFORCE_EQ(
ksize.size(), 2,
Expand All @@ -60,36 +60,32 @@ class PoolXPUKernel : public framework::OpKernel<T> {
ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
}
}
const int c = in_x->dims()[0] * in_x->dims()[1];
const int n = in_x->dims()[0];
const int c = in_x->dims()[1];
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];
const int win_h = ksize[0];
const int win_w = ksize[1];
const int stride_h = strides[0];
const int stride_w = strides[1];
const int pad_up = paddings[0];
const int pad_down = paddings[0];
const int pad_left = paddings[1];
const int pad_right = paddings[1];
const float* input = in_x->data<float>();
out->mutable_data<T>(context.GetPlace());
float* output = out->data<float>();
xpu::Pooling_t pool_type = XPUPoolingType(pooling_type, exclusive, is_test);
auto& dev_ctx = context.template device_context<DeviceContext>();
int r = xpu::pooling_forward<float, float>(
dev_ctx.x_context(), input, output, index_data, pool_type, c, in_h,
in_w, pad_left, pad_right, pad_up, pad_down, win_h, win_w, stride_h,
stride_w, out_h, out_w);
PADDLE_ENFORCE_EQ(
r, xpu::Error_t::SUCCESS,
platform::errors::External(
"The pool2d XPU API return wrong value[%d], please check "
"where Baidu Kunlun Card is properly installed.",
r));
int r = xpu::Error_t::SUCCESS;
if (pooling_type == "max") {
r = xpu::max_pool2d(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(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));
}
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::External(
"The pool2d XPU API return wrong value[%d %s]", r,
XPUAPIErrorMsg[r]));
}
};

template <typename DeviceContext, typename T>
class PoolGradXPUKernel : public framework::OpKernel<T> {
public:
Expand Down Expand Up @@ -126,47 +122,33 @@ class PoolGradXPUKernel : public framework::OpKernel<T> {
if (!in_x_grad) {
return;
}
const int c = in_x->dims()[0] * in_x->dims()[1];
const int n = in_x->dims()[0];
const int c = in_x->dims()[1];
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];
const int win_h = ksize[0];
const int win_w = ksize[1];
const int stride_h = strides[0];
const int stride_w = strides[1];
const int pad_up = paddings[0];
const int pad_down = paddings[0];
const int pad_left = paddings[1];
const int pad_right = paddings[1];
const float* input = in_x->data<float>();
const float* output = out->data<float>();
const float* output_grad = out_grad->data<float>();
in_x_grad->mutable_data<T>(context.GetPlace());
float* input_grad = in_x_grad->data<float>();
xpu::Pooling_t pool_type = XPUPoolingType(pooling_type, exclusive, false);
auto& dev_ctx = context.template device_context<DeviceContext>();
// Need to init memory in the first place
const int zero = 0;
int r =
xpu::memset(dev_ctx.x_context(), reinterpret_cast<void**>(input_grad),
zero, in_x_grad->numel() * sizeof(float));
PADDLE_ENFORCE_EQ(
r, xpu::Error_t::SUCCESS,
platform::errors::External(
"The Pool2d XPU OP return wrong value[%d], please check "
"where Baidu Kunlun Card is properly installed.",
r));
r = xpu::pooling_backward(dev_ctx.x_context(), input, output, index_data,
output_grad, input_grad, pool_type, c, in_h, in_w,
pad_left, pad_right, pad_up, pad_down, win_h,
win_w, stride_h, stride_w, out_h, out_w);
PADDLE_ENFORCE_EQ(
r, xpu::Error_t::SUCCESS,
platform::errors::External(
"The Pool2d XPU OP return wrong value[%d], please check "
"where Baidu Kunlun Card is properly installed.",
r));
int r = xpu::Error_t::SUCCESS;
if (pooling_type == "max") {
r = xpu::max_pool2d_grad(dev_ctx.x_context(), input, output, index_data,
output_grad, input_grad, n, c, in_h, in_w, ksize,
strides, paddings, true);
} else if (pooling_type == "avg") {
r = xpu::avg_pool2d_grad(dev_ctx.x_context(), input, output, output_grad,
input_grad, n, c, in_h, in_w, ksize, strides,
paddings, !exclusive, true);
} else {
PADDLE_THROW(platform::errors::InvalidArgument(
"Unsupported pooling type for kunlun ", pooling_type));
}
PADDLE_ENFORCE_EQ(r, xpu::Error_t::SUCCESS,
platform::errors::External(
"The Pool2dGrad XPU OP return wrong value[%d %s]", r,
XPUAPIErrorMsg[r]));
}
};

Expand Down
11 changes: 1 addition & 10 deletions paddle/fluid/platform/device_context.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,16 +172,7 @@ Place CPUDeviceContext::GetPlace() const { return place_; }
#ifdef PADDLE_WITH_XPU
XPUDeviceContext::XPUDeviceContext() { context_ = xpu::create_context(); }

XPUDeviceContext::~XPUDeviceContext() {
xpu::destroy_context(context_);
void* l3ptr = nullptr;
int l3_size = 13.5 * 1024 * 1024;
xpu_malloc(static_cast<void**>(&l3ptr), l3_size, XPU_MEM_L3);
if (l3ptr != nullptr) {
context_->_l3_mgr.set(l3ptr, l3_size);
std::cout << "set l3 size " << l3_size << std::endl;
}
}
XPUDeviceContext::~XPUDeviceContext() {}

XPUDeviceContext::XPUDeviceContext(XPUPlace place) : place_(place) {
int dev_id = -1;
Expand Down
Loading