saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-11 17:48:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

disable inner parallel for global avg pool as normally they are small (#9487)

Browse source 
* Using cost model's thread count rather than max number of threads when
parallel tasks.

* according to perf test result, decrease parallel on channels.

* Seems no use on parallel channels for qavg_pool according several models, remove it.

* Revert "Using cost model's thread count rather than max number of threads when"

This reverts commit 5fa47cd5b5ddbaa4e5ef97ccbc53200324379544.
This commit is contained in:
Zhang Lei 2021-10-26 15:35:49 -07:00 committed by GitHub
parent df7a5342a5
commit 33ef1d7700
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 13 additions and 34 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

47
onnxruntime/contrib_ops/cpu/qlinear_global_average_pool.cc
View file

@ -26,8 +26,6 @@ Status ComputeQLinearGlobalAvgPool(
int64_t image_size,
bool channels_last,
concurrency::ThreadPool* tp) {
static constexpr int64_t kMiniChannelGroup = 64;
if (!channels_last || C == 1) {
auto worker = [=](std::ptrdiff_t first, std::ptrdiff_t last) {
const uint8_t* input = (const uint8_t*)(x + (first * image_size));
@ -38,38 +36,19 @@ Status ComputeQLinearGlobalAvgPool(
concurrency::ThreadPool::TryParallelFor(
tp, static_cast<std::ptrdiff_t>(N * C), {1.0 * image_size, 1.0, 8.0 * image_size}, worker);
} else {
if (N == 1) {
int64_t channel_padded = (C + kMiniChannelGroup - 1) & (~(kMiniChannelGroup - 1));
int64_t channel_groups = channel_padded / kMiniChannelGroup;
auto worker = [=](std::ptrdiff_t first, std::ptrdiff_t last) {
std::vector<int32_t> acc_buffer(MlasQLinearSafePaddingElementCount(sizeof(int32_t), C));
std::vector<uint8_t> zero_buffer(MlasQLinearSafePaddingElementCount(sizeof(uint8_t), C), 0);
const uint8_t* input = x + first * kMiniChannelGroup;
uint8_t* output = y + first * kMiniChannelGroup;
int64_t channel_count = (last == channel_groups) ? (C - first * kMiniChannelGroup) : ((last - first) * kMiniChannelGroup);
MlasQLinearGlobalAveragePoolNhwc(
input, x_scale, x_zero_point, output, y_scale, y_zero_point,
N, image_size, C, channel_count, acc_buffer.data(), zero_buffer.data());
};
concurrency::ThreadPool::TryParallelFor(
tp, static_cast<std::ptrdiff_t>(channel_groups),
{1.0 * N * image_size * kMiniChannelGroup, 1.0 * N * kMiniChannelGroup, 8.0 * N * image_size * kMiniChannelGroup},
worker);
} else {
auto worker = [=](std::ptrdiff_t first, std::ptrdiff_t last) {
const uint8_t* input = x + first * C * image_size;
uint8_t* output = y + first * C;
std::vector<int32_t> acc_buffer(MlasQLinearSafePaddingElementCount(sizeof(int32_t), C));
std::vector<uint8_t> zero_buffer(MlasQLinearSafePaddingElementCount(sizeof(uint8_t), C), 0);
MlasQLinearGlobalAveragePoolNhwc(
input, x_scale, x_zero_point, output, y_scale, y_zero_point,
last - first, image_size, C, C, acc_buffer.data(), zero_buffer.data());
};
concurrency::ThreadPool::TryParallelFor(
tp, static_cast<std::ptrdiff_t>(N),
{1.0 * image_size * C, 1.0 * C, 8.0 *image_size * C},
worker);
}
auto worker = [=](std::ptrdiff_t first, std::ptrdiff_t last) {
const uint8_t* input = x + first * C * image_size;
uint8_t* output = y + first * C;
std::vector<int32_t> acc_buffer(MlasQLinearSafePaddingElementCount(sizeof(int32_t), C));
std::vector<uint8_t> zero_buffer(MlasQLinearSafePaddingElementCount(sizeof(uint8_t), C), 0);
MlasQLinearGlobalAveragePoolNhwc(
input, x_scale, x_zero_point, output, y_scale, y_zero_point,
last - first, image_size, C, C, acc_buffer.data(), zero_buffer.data());
};
concurrency::ThreadPool::TryParallelFor(
tp, static_cast<std::ptrdiff_t>(N),
{1.0 * image_size * C, 1.0 * C, 8.0 *image_size * C},
worker);
}
return Status::OK();
}