Skip to content

Commit

Permalink
multi-threaded scatterND and refactor perf
Browse files Browse the repository at this point in the history
  • Loading branch information
@fengyuentau
fengyuentau committed Jan 5, 2024
1 parent 2997b4c commit 2ed97b9
Show file tree
Hide file tree
Showing 2 changed files with 119 additions and 119 deletions.
158 changes: 74 additions & 84 deletions modules/dnn/perf/perf_layer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -324,103 +324,95 @@ INSTANTIATE_TEST_CASE_P(/**/, Layer_Scatter, Combine(
/* withCann= */ false) // only test on CPU
));

struct Layer_ScatterND : public TestBaseWithParam<tuple<Backend, Target> >
{
void test_layer(const std::vector<int>& shape, const String reduction = "none")
{
int backendId = get<0>(GetParam());
int targetId = get<1>(GetParam());
using Layer_ScatterND = TestBaseWithParam<tuple<std::vector<int>, std::string, tuple<Backend, Target>>>;
PERF_TEST_P_(Layer_ScatterND, scatterND) {
std::vector<int> shape = get<0>(GetParam());
std::string reduction = get<1>(GetParam());
int backend_id = get<0>(get<2>(GetParam()));
int target_id = get<1>(get<2>(GetParam()));

std::vector<int> indices_shape(shape);
indices_shape.push_back(int(shape.size()));
Mat data(shape, CV_32FC1);
Mat indices(indices_shape, CV_32FC1);
Mat updates(shape, CV_32FC1);
std::vector<int> indices_shape(shape);
indices_shape.push_back(int(shape.size()));
Mat data(shape, CV_32FC1);
Mat indices(indices_shape, CV_32FC1);
Mat updates(shape, CV_32FC1);

Scalar mean = 0.f;
Scalar std = 1.f;
randn(data, mean, std);
randn(updates, mean, std);

// initialize the indices with index tuples like [0...N, 0...C, 0...H, 0...W]
std::vector<int> current_index_tuple(shape.size());
int total = data.total();
std::vector<int> indices_step;
for (int i = 0; i < indices.dims; i++)
{
int step = indices.step.p[i] / sizeof(float);
indices_step.push_back(step);
}
int t, j, idx, offset_at_idx, offset;
for (int i = 0; i < total; i++)
randn(data, 0.f, 1.f);
randn(updates, 0.f, 1.f);

// initialize the indices with index tuples like [0...N, 0...C, 0...H, 0...W]
std::vector<int> current_index_tuple(shape.size());
int total = data.total();
std::vector<int> indices_step;
for (int i = 0; i < indices.dims; i++)
{
int step = indices.step.p[i] / sizeof(float);
indices_step.push_back(step);
}
int t, j, idx, offset_at_idx, offset;
for (int i = 0; i < total; i++)
{
t = i;
for (j = shape.size() - 1; j >= 0; j--)
{
t = i;
for (j = shape.size() - 1; j >= 0; j--)
{
idx = t / shape[j];
offset_at_idx = (int)(t - idx * shape[j]);
current_index_tuple[j] = offset_at_idx;
t = idx;
}

offset = 0;
for (j = 0; j < shape.size(); j++)
offset += current_index_tuple[j] * indices_step[j];

for (j = 0; j < shape.size(); j++)
indices.at<float>(offset + j) = current_index_tuple[j];
idx = t / shape[j];
offset_at_idx = (int)(t - idx * shape[j]);
current_index_tuple[j] = offset_at_idx;
t = idx;
}

Net net;
LayerParams lp;
lp.type = "ScatterND";
lp.name = "testLayer";
lp.set("reduction", reduction);
offset = 0;
for (j = 0; j < shape.size(); j++)
offset += current_index_tuple[j] * indices_step[j];

int id = net.addLayerToPrev(lp.name, lp.type, lp);
net.connect(0, 0, id, 0);
net.connect(0, 1, id, 1);
net.connect(0, 2, id, 2);
for (j = 0; j < shape.size(); j++)
indices.at<float>(offset + j) = current_index_tuple[j];
}

// warmup
{
std::vector<String> inpNames(3);
inpNames[0] = "data";
inpNames[1] = "indices";
inpNames[2] = "updates";
net.setInputsNames(inpNames);
net.setInput(data, inpNames[0]);
net.setInput(indices, inpNames[1]);
net.setInput(updates, inpNames[2]);
Net net;
LayerParams lp;
lp.type = "ScatterND";
lp.name = "testLayer";
lp.set("reduction", reduction);

net.setPreferableBackend(backendId);
net.setPreferableTarget(targetId);
Mat out = net.forward();
}
int id = net.addLayerToPrev(lp.name, lp.type, lp);
net.connect(0, 0, id, 0);
net.connect(0, 1, id, 1);
net.connect(0, 2, id, 2);

TEST_CYCLE()
{
Mat res = net.forward();
}
// warmup
{
std::vector<String> input_names{"data", "indices", "updates"};
net.setInputsNames(input_names);
net.setInput(data, input_names[0]);
net.setInput(indices, input_names[1]);
net.setInput(updates, input_names[2]);

SANITY_CHECK_NOTHING();
net.setPreferableBackend(backend_id);
net.setPreferableTarget(target_id);
Mat out = net.forward();
}

int N = 8;
int C = 256;
int H = 128;
int W = 100;
};
TEST_CYCLE()
{
Mat res = net.forward();
}

PERF_TEST_P_(Layer_ScatterND, DISABLED_ScatterND)
{
test_layer({N, C, H ,W});
SANITY_CHECK_NOTHING();
}

PERF_TEST_P_(Layer_ScatterND, DISABLED_ScatterND_add)
{
test_layer({N, C, H , W}, "add");
}
INSTANTIATE_TEST_CASE_P(/**/, Layer_ScatterND, Combine(
Values(std::vector<int>{2, 128, 64, 50}),
Values(std::string("none"), std::string("add")),
dnnBackendsAndTargets(/* withInferenceEngine= */ false,
/* withHalide= */ false,
/* withCpuOCV= */ true,
/* withVkCom= */ false,
/* withCUDA= */ false,
/* withNgraph= */ false,
/* withWebnn= */ false,
/* withCann= */ false) // only test on CPU
));

struct Layer_LayerNorm : public TestBaseWithParam<tuple<Backend, Target> >
{
Expand Down Expand Up @@ -795,8 +787,6 @@ INSTANTIATE_TEST_CASE_P(/**/, Layer_NaryEltwise, testing::Values(std::make_tuple
#ifdef HAVE_CUDA
INSTANTIATE_TEST_CASE_P(CUDA, Layer_NaryEltwise, testing::Values(std::make_tuple(DNN_BACKEND_CUDA, DNN_TARGET_CUDA)));
#endif
// INSTANTIATE_TEST_CASE_P(/**/, Layer_Scatter, testing::Values(std::make_tuple(DNN_BACKEND_OPENCV, DNN_TARGET_CPU)));
INSTANTIATE_TEST_CASE_P(/**/, Layer_ScatterND, testing::Values(std::make_tuple(DNN_BACKEND_OPENCV, DNN_TARGET_CPU)));
INSTANTIATE_TEST_CASE_P(/**/, Layer_LayerNorm, testing::Values(std::make_tuple(DNN_BACKEND_OPENCV, DNN_TARGET_CPU)));
INSTANTIATE_TEST_CASE_P(/**/, Layer_LayerNormExpanded, testing::Values(std::make_tuple(DNN_BACKEND_OPENCV, DNN_TARGET_CPU)));
INSTANTIATE_TEST_CASE_P(/**/, Layer_GatherElements, testing::Values(std::make_tuple(DNN_BACKEND_OPENCV, DNN_TARGET_CPU)));
Expand Down
80 changes: 45 additions & 35 deletions modules/dnn/src/layers/scatterND_layer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -89,49 +89,59 @@ class ScatterNDLayerImpl CV_FINAL : public ScatterNDLayer
// NOTE: This impl does not check whether indices have duplicate entries.
// The last duplicate entry will overwrite the previous.
template<typename T, typename Functor>
void forward_impl(const Functor& rd, const Mat& data, const Mat& indices, const Mat& updates, Mat& out)
{
data.copyTo(out);

const int* shape = data.size.p;
const size_t* step = data.step.p;
void forward_impl(const Functor &reduce_operation, const Mat &input_mat, const Mat &indices_mat, const Mat &updates_mat, Mat& output_mat) {
input_mat.copyTo(output_mat);

const int ind_ndims = indices.dims;
const int* ind_shape = indices.size.p;
const T* p_indices = indices.ptr<const T>();
const auto &input_mat_shape = shape(input_mat);
std::vector<size_t> input_mat_step(input_mat_shape.size());
for (int i = 0; i < input_mat.dims; i++) {
input_mat_step[i] = static_cast<size_t>(input_mat.step.p[i] / sizeof(T));
}

const int upd_ndims = updates.dims;
const int* upd_shape = updates.size.p;
const T* p_updates = updates.ptr<const T>();
const int indices_mat_ndims = indices_mat.dims;
const auto &indices_mat_shape = shape(indices_mat);

T* p_out = out.ptr<T>();
const int updates_mat_ndims = updates_mat.dims;
const auto &updates_mat_shape = shape(updates_mat);

int k = ind_shape[ind_ndims - 1]; // last dim of indices
size_t total = (size_t)(indices.total() / k);
int indices_last_dim = indices_mat_shape[indices_mat_ndims - 1]; // last dim of indices

size_t updates_size = 1;
for (int i = ind_ndims - 1; i < upd_ndims; i++)
updates_size *= upd_shape[i];

size_t inp_start_offset = 0;
size_t ind_start_offset = 0;
size_t upd_start_offset = 0;
for (size_t i = 0; i < total; i++, ind_start_offset += k, upd_start_offset += updates_size)
{
const T* tmp_p_indices = p_indices + ind_start_offset;
inp_start_offset = 0;
for (int j = 0; j < k; j++)
{
CV_Assert(tmp_p_indices[j] < shape[j] && tmp_p_indices[j] > -shape[j]);
inp_start_offset += (((int)tmp_p_indices[j] + shape[j]) % shape[j]) * step[j];
for (int i = indices_mat_ndims - 1; i < updates_mat_ndims; i++)
updates_size *= updates_mat_shape[i];

auto fn = [&](const Range &r) {
size_t input_offset = 0,
indices_offset = r.start * indices_last_dim,
updates_offset = r.start * updates_size;
for (int i = r.start; i < r.end; i++) {
const T* indices = indices_mat.ptr<const T>();
const T* updates = updates_mat.ptr<const T>();
T* output = output_mat.ptr<T>();

input_offset = 0;
indices += indices_offset;
for (int j = 0; j < indices_last_dim; j++) {
int index = static_cast<int>(*(indices + j));
index = (index + input_mat_shape[j]) % input_mat_shape[j];
CV_Assert(index < input_mat_shape[j] && index >= 0);
input_offset += index * input_mat_step[j];
}

updates += updates_offset;
output += input_offset;
for (int j = 0; j < updates_size; j++) {
output[j] = reduce_operation(output[j], updates[j]);
}

indices_offset += indices_last_dim;
updates_offset += updates_size;
}
inp_start_offset /= sizeof(T);
};

const T* tmp_p_updates = p_updates + upd_start_offset;
T* tmp_p_out = p_out + inp_start_offset;
for (int j = 0; j < updates_size; j++)
tmp_p_out[j] = rd(tmp_p_out[j], tmp_p_updates[j]);
}
size_t total = (size_t)(indices_mat.total() / indices_last_dim);
double nstripes = (size_t)total * (indices_last_dim + updates_size) * (1 / 1024.0);
parallel_for_(Range(0, total), fn, nstripes);
}

template<typename... Args>
Expand Down

0 comments on commit 2ed97b9

Please sign in to comment.