saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-05-30 23:18:20 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Optimize upsample (#1008)

Browse source 
Extract the index and scale computation out of the loop to improve the perf
This commit is contained in:
Yufeng Li 2019-05-13 16:30:02 -07:00 committed by GitHub
parent 30f87c7526
commit 9330a42fae
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 57 additions and 34 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

91
onnxruntime/core/providers/cpu/tensor/upsample.cc
View file

@ -158,47 +158,68 @@ void upsampleBilinear(
float height_scale,
float width_scale,
const T* Xdata,
T* Ydata) {
T* Ydata,
AllocatorPtr& alloc) {
int64_t output_width = static_cast<int64_t>(input_width * width_scale);
int64_t output_height = static_cast<int64_t>(input_height * height_scale);
size_t inx_buffer_size = 2 * sizeof(int64_t) * (output_height + output_width);
size_t scale_buffer_size = 2 * sizeof(float_t) * (output_height + output_width);
auto inx_scale_data_buffer = alloc->Alloc(inx_buffer_size + scale_buffer_size);
BufferUniquePtr inx_scale_data_buffer_holder(inx_scale_data_buffer, BufferDeleter(alloc));
int64_t* inx_data = static_cast<int64_t*>(inx_scale_data_buffer_holder.get());
int64_t* input_width_mul_y1 = inx_data;
int64_t* input_width_mul_y2 = inx_data + output_height;
int64_t* in_x1 = inx_data + 2 * output_height;
int64_t* in_x2 = inx_data + 2 * output_height + output_width;
float* scale_data = reinterpret_cast<float*>( in_x2 + output_width );
float* dy1 = scale_data;
float* dy2 = scale_data + output_height;
float* dx1 = scale_data + 2 * output_height;
float* dx2 = scale_data + 2 * output_height + output_width;
for (int64_t y = 0; y < output_height; ++y) {
float in_y = std::min(y / height_scale, static_cast<float>(input_height - 1));
const int64_t in_y1 = std::min(static_cast<int64_t>(in_y), input_height - 1);
const int64_t in_y2 = std::min(in_y1 + 1, input_height - 1);
dy1[y] = fabs(in_y - in_y1);
dy2[y] = fabs(in_y - in_y2);
if (in_y1 == in_y2) {
dy1[y] = 0.5f;
dy2[y] = 0.5f;
}
input_width_mul_y1[y] = input_width * in_y1;
input_width_mul_y2[y] = input_width * in_y2;
}
for (int64_t x = 0; x < output_width; ++x) {
float in_x = std::min(x / width_scale, static_cast<float>(input_width - 1));
in_x1[x] = std::min(static_cast<int64_t>(in_x), input_width - 1);
in_x2[x] = std::min(in_x1[x] + 1, input_width - 1);
dx1[x] = std::abs(in_x - in_x1[x]);
dx2[x] = std::abs(in_x - in_x2[x]);
if (in_x1[x] == in_x2[x]) {
dx1[x] = 0.5f;
dx2[x] = 0.5f;
}
}
for (int64_t n = 0; n < batch_size; ++n) {
for (int64_t c = 0; c < num_channels; ++c) {
for (int64_t y = 0; y < output_height; ++y) {
float in_y = std::min(y / height_scale, static_cast<float>(input_height - 1));
const int64_t in_y1 = std::min(static_cast<int64_t>(in_y), input_height - 1);
const int64_t in_y2 = std::min(in_y1 + 1, input_height - 1);
float dy1 = fabs(in_y - in_y1);
float dy2 = fabs(in_y - in_y2);
if (in_y1 == in_y2) {
dy1 = 0.5f;
dy2 = 0.5f;
}
const int64_t input_width_mul_y1 = input_width * in_y1;
const int64_t input_width_mul_y2 = input_width * in_y2;
for (int64_t x = 0; x < output_width; ++x) {
float in_x = std::min(x / width_scale, static_cast<float>(input_width - 1));
const int64_t in_x1 = std::min(static_cast<int64_t>(in_x), input_width - 1);
const int64_t in_x2 = std::min(in_x1 + 1, input_width - 1);
T X11 = Xdata[input_width_mul_y1[y] + in_x1[x]];
T X21 = Xdata[input_width_mul_y1[y] + in_x2[x]];
T X12 = Xdata[input_width_mul_y2[y] + in_x1[x]];
T X22 = Xdata[input_width_mul_y2[y] + in_x2[x]];
float dx1 = std::abs(in_x - in_x1);
float dx2 = std::abs(in_x - in_x2);
if (in_x1 == in_x2) {
dx1 = 0.5f;
dx2 = 0.5f;
}
T X11 = Xdata[input_width_mul_y1 + in_x1];
T X21 = Xdata[input_width_mul_y1 + in_x2];
T X12 = Xdata[input_width_mul_y2 + in_x1];
T X22 = Xdata[input_width_mul_y2 + in_x2];
Ydata[output_width * y + x] = static_cast<T>(dx2 * dy2 * X11 +
dx1 * dy2 * X21 +
dx2 * dy1 * X12 +
dx1 * dy1 * X22);
Ydata[output_width * y + x] = static_cast<T>(dx2[x] * dy2[y] * X11 +
dx1[x] * dy2[y] * X21 +
dx2[x] * dy1[y] * X12 +
dx1[x] * dy1[y] * X22);
}
}
Xdata += input_height * input_width;
@ -235,8 +256,10 @@ Status Upsample<T>::BaseCompute(OpKernelContext* context, const std::vector<floa
const int64_t batch_size = dims[0], num_channels = dims[1];
const int64_t input_height = dims[2], input_width = dims[3];
AllocatorPtr alloc;
ORT_RETURN_IF_ERROR(context->GetTempSpaceAllocator(&alloc));
upsampleBilinear(batch_size, num_channels, input_height, input_width,
scales[2], scales[3], X->template Data<T>(), Y->template MutableData<T>());
scales[2], scales[3], X->template Data<T>(), Y->template MutableData<T>(), alloc);
return Status::OK();
}
default: