Chiebot-Mirror
/
opencv
8
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

multi-threaded scatter and refactor perf

Browse Source
pull/24813/head
fengyuentau 1 year ago
parent a8ec658611
commit 63cde0b90d
2 changed files with 103 additions and 104 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 115
      modules/dnn/perf/perf_layer.cpp
  2. 92
      modules/dnn/src/layers/scatter_layer.cpp

115
modules/dnn/perf/perf_layer.cpp
Unescape View File

@ -258,76 +258,71 @@ PERF_TEST_P_(Layer_Slice, FastNeuralStyle_eccv16)
test_slice<4>(inputShape, begin, end); test_slice<4>(inputShape, begin, end);
} }
struct Layer_Scatter : public TestBaseWithParam<tuple<Backend, Target> > using Layer_Scatter = TestBaseWithParam<tuple<std::vector<int>, std::string, int, tuple<Backend, Target>>>;
{ PERF_TEST_P_(Layer_Scatter, scatter) {
void test_layer(const std::vector<int>& shape, const String reduction = "none", int axis = 0) std::vector<int> shape = get<0>(GetParam());
{ std::string reduction = get<1>(GetParam());
int backendId = get<0>(GetParam()); int axis = get<2>(GetParam());
int targetId = get<1>(GetParam()); int backend_id = get<0>(get<3>(GetParam()));
int target_id = get<1>(get<3>(GetParam()));
Mat data(shape, CV_32FC1);
Mat indices(shape, CV_32FC1);
Mat updates(shape, CV_32FC1);
Scalar mean = 0.f;
Scalar std = 1.f;
randn(data, mean, std);
randu(indices, 0, shape[axis]);
randn(updates, mean, std);
indices.convertTo(indices, CV_32SC1, 1, -1);
Net net; Mat data(shape, CV_32FC1);
LayerParams lp; Mat indices(shape, CV_32FC1);
lp.type = "Scatter"; Mat updates(shape, CV_32FC1);
lp.name = "testLayer";
lp.set("reduction", reduction);
lp.set("axis", axis);
int id = net.addLayerToPrev(lp.name, lp.type, lp); randn(data, 0.f, 1.f);
net.connect(0, 0, id, 0); randu(indices, 0, shape[axis]);
net.connect(0, 1, id, 1); randn(updates, 0.f, 1.f);
net.connect(0, 2, id, 2);
// warmup indices.convertTo(indices, CV_32SC1, 1, -1);
{
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.setPreferableBackend(backendId); Net net;
net.setPreferableTarget(targetId); LayerParams lp;
Mat out = net.forward(); lp.type = "Scatter";
} lp.name = "testLayer";
lp.set("reduction", reduction);
lp.set("axis", axis);
TEST_CYCLE() int id = net.addLayerToPrev(lp.name, lp.type, lp);
{ net.connect(0, 0, id, 0);
Mat res = net.forward(); net.connect(0, 1, id, 1);
} net.connect(0, 2, id, 2);
SANITY_CHECK_NOTHING(); // 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]);
net.setPreferableBackend(backend_id);
net.setPreferableTarget(target_id);
Mat out = net.forward();
} }
int N = 8; // perf
int C = 256; TEST_CYCLE()
int H = 128; {
int W = 100; Mat res = net.forward();
}; }
PERF_TEST_P_(Layer_Scatter, DISABLED_Scatter) SANITY_CHECK_NOTHING();
{
test_layer({N, C, H, W});
} }
PERF_TEST_P_(Layer_Scatter, DISABLED_Scatter_add) INSTANTIATE_TEST_CASE_P(/**/, Layer_Scatter, Combine(
{ Values(std::vector<int>{2, 128, 64, 50}),
test_layer({N, C, H, W}, "add"); Values(std::string("none"), std::string("add")),
} Values(0), // use Values(0, 1, 2, 3) for more details
dnnBackendsAndTargets(/* withInferenceEngine= */ false,
/* withHalide= */ false,
/* withCpuOCV= */ true,
/* withVkCom= */ false,
/* withCUDA= */ false,
/* withNgraph= */ false,
/* withWebnn= */ false,
/* withCann= */ false) // only test on CPU
));
struct Layer_ScatterND : public TestBaseWithParam<tuple<Backend, Target> > struct Layer_ScatterND : public TestBaseWithParam<tuple<Backend, Target> >
{ {
@ -800,7 +795,7 @@ INSTANTIATE_TEST_CASE_P(/**/, Layer_NaryEltwise, testing::Values(std::make_tuple
#ifdef HAVE_CUDA #ifdef HAVE_CUDA
INSTANTIATE_TEST_CASE_P(CUDA, Layer_NaryEltwise, testing::Values(std::make_tuple(DNN_BACKEND_CUDA, DNN_TARGET_CUDA))); INSTANTIATE_TEST_CASE_P(CUDA, Layer_NaryEltwise, testing::Values(std::make_tuple(DNN_BACKEND_CUDA, DNN_TARGET_CUDA)));
#endif #endif
INSTANTIATE_TEST_CASE_P(/**/, Layer_Scatter, testing::Values(std::make_tuple(DNN_BACKEND_OPENCV, DNN_TARGET_CPU))); // 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_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_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_LayerNormExpanded, testing::Values(std::make_tuple(DNN_BACKEND_OPENCV, DNN_TARGET_CPU)));

92
modules/dnn/src/layers/scatter_layer.cpp
Unescape View File

@ -81,59 +81,63 @@ public:
} }
template<typename T, typename Functor> template<typename T, typename Functor>
void forward_impl(const Functor& rd, const Mat& data, const Mat& indices, const Mat& updates, Mat& out) void forward_impl(const Functor& reduce_operation, const Mat& input_mat, const Mat& indices_mat, const Mat& updates_mat, Mat& output_mat)
{ {
data.copyTo(out); input_mat.copyTo(output_mat);
const int ndims = data.dims; const int ndims = input_mat.dims;
const int* shape = data.size.p; const auto &input_mat_shape = shape(input_mat);
const size_t* step = data.step.p; std::vector<size_t> input_mat_step(ndims);
const int* ind_shape = indices.size.p; const auto &indices_mat_shape = shape(indices_mat);
const size_t* ind_step = indices.step.p; // const auto &indices_mat_step = indices_mat.step;
std::vector<size_t> indices_mat_step(ndims);
size_t inp_offset = 0; for (int i = 0; i < ndims; i++) {
size_t ind_offset = 0; input_mat_step[i] = static_cast<size_t>(input_mat.step.p[i] / sizeof(T));
const T* p_index = indices.ptr<const T>(); indices_mat_step[i] = static_cast<size_t>(indices_mat.step.p[i] / sizeof(T));
const T* p_update = updates.ptr<const T>(); }
T* p_out = out.ptr<T>();
size_t total = indices.total();
int j, offset_at_idx, index; const T* indices = indices_mat.ptr<const T>();
size_t t, idx; const T* updates = updates_mat.ptr<const T>();
for (size_t i = 0; i < total; i++) T* output = output_mat.ptr<T>();
{
t = i; auto fn = [&](const Range &r) {
inp_offset = 0; size_t input_offset = 0, indices_offset = 0;
ind_offset = 0;
int offset_at_axis = 0; int indices_index, index;
for (j = ndims - 1; j >= 0; j--) size_t axis_offset, tmp_index, j_index;
{ for (int i = r.start; i < r.end; i++) {
idx = t / ind_shape[j]; input_offset = 0;
offset_at_idx = (int)(t - idx * ind_shape[j]); indices_offset = 0;
ind_offset += offset_at_idx * ind_step[j]; indices_index = i;
inp_offset += offset_at_idx * step[j]; axis_offset = 0;
t = idx; for (int j = ndims - 1; j >= 0; j--) {
if (j == axis) tmp_index = indices_index / indices_mat_shape[j];
{ j_index = (size_t)(indices_index - tmp_index * indices_mat_shape[j]);
offset_at_axis = offset_at_idx * step[j]; input_offset += j_index * input_mat_step[j];
indices_offset += j_index * indices_mat_step[j];
indices_index = tmp_index;
if (j == axis) {
axis_offset = j_index * input_mat_step[j];
}
} }
}
ind_offset /= sizeof(T);
// get index and overwrite current indices // get index and overwrite current indices
const T* tmp_p_index = p_index + ind_offset; index = static_cast<int>(*(indices + indices_offset));
index = (int)(*tmp_p_index); index = (index + input_mat_shape[axis]) % input_mat_shape[axis];
CV_Assert(index < shape[axis] && index > -shape[axis]); CV_Assert(index < input_mat_shape[axis] && index >= 0);
input_offset = input_offset - axis_offset + index * input_mat_step[axis];
inp_offset = inp_offset - offset_at_axis + ((index + shape[axis]) % shape[axis]) * step[axis]; const T* update = updates + indices_offset;
inp_offset /= sizeof(T); T* y = output + input_offset;
*y = reduce_operation(*y, *update);
}
};
const T* tmp_p_update = p_update + ind_offset; size_t total = indices_mat.total();
T* tmp_p_out = p_out + inp_offset; double nstripes = (size_t)total * ndims * (1 / 1024.0);
*tmp_p_out = rd(*tmp_p_out, *tmp_p_update); parallel_for_(Range(0, total), fn, nstripes);
}
} }
template<typename... Args> template<typename... Args>

Write Preview
Loading…
Cancel
Save