Skip to content

speed up CPU EmbeddingBag (indexSelectAdd op) #5433

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
soumith merged 5 commits into from
Mar 15, 2018
Merged

speed up CPU EmbeddingBag (indexSelectAdd op) #5433

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
71d38cb
speed up CPU EmbeddingBag (indexSelectAdd op)
Feb 26, 2018
05f20b8
keep operator inside EmbeddingBag + speedup
Mar 14, 2018
c97baa5
comment
Mar 14, 2018
4a6d97a
update checkScalarTypes signature
Mar 14, 2018
daa1c07
enforce type in embedding_bag_backward_cpu
Mar 14, 2018
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
20 changes: 20 additions & 0 deletions aten/src/ATen/TensorUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,6 +162,26 @@ void checkScalarType(CheckedFrom c, const TensorArg& t, ScalarType ty) {
}
}

void checkScalarTypes(CheckedFrom c, const TensorArg& t,
at::ArrayRef<ScalarType> l) {
if (std::find(l.begin(), l.end(), t->type().scalarType()) == l.end()) {
std::ostringstream oss;
oss << "Expected tensor for " << t << " to have one of the following "
<< "scalar types: ";
size_t i = 0;
for (auto ty : l) {
if (i != 0) {
oss << ", ";
}
oss << toString(ty);
i++;
}
oss << "; but got " << t->toString()
<< " instead (while checking arguments for " << c << ")";
throw std::runtime_error(oss.str());
}
}

void checkAllSameType(CheckedFrom c, ArrayRef<TensorArg> tensors) {
checkAllSame(c, tensors, checkSameType);
}
Expand Down
2 changes: 1 addition & 1 deletion aten/src/ATen/TensorUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,7 @@ void checkNumel(CheckedFrom c, const TensorGeometryArg& t, int64_t numel);
void checkSameNumel(CheckedFrom c, const TensorGeometryArg& t1, const TensorGeometryArg& t2);
void checkAllSameNumel(CheckedFrom c, ArrayRef<TensorArg> tensors);
void checkScalarType(CheckedFrom c, const TensorArg& t, ScalarType s);
void checkScalarTypes(CheckedFrom c, const TensorArg& t, at::ArrayRef<ScalarType> l);
void checkSameGPU(CheckedFrom c, const TensorArg& t1, const TensorArg& t2);
void checkAllSameGPU(CheckedFrom c, ArrayRef<TensorArg> tensors);
void checkSameType(CheckedFrom c, const TensorArg& t1, const TensorArg& t2);
Expand All @@ -78,4 +79,3 @@ void * maybe_data_ptr(const Tensor& tensor);
void * maybe_data_ptr(const TensorArg& tensor);

}

62 changes: 50 additions & 12 deletions aten/src/ATen/native/EmbeddingBag.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,39 @@ static void make_offset2bag(const Tensor &offsets, const Tensor &indices,
offset2bag = offset2bag.cumsum(0); // offset2bag = [0 0 1 1 2]
}

template<typename T>
static void axpy(int64_t n, T a, T *x, int64_t incx, T *y, int64_t incy);
template<>
void axpy<float>(int64_t n, float a, float *x, int64_t incx,
float *y, int64_t incy) {
THFloatBlas_axpy(n, a, x, incx, y, incy);
}
template<>
void axpy<double>(int64_t n, double a, double *x, int64_t incx,
double *y, int64_t incy) {
THDoubleBlas_axpy(n, a, x, incx, y, incy);
}

// This function combines index_select (using select_indices as the index) and
// index_add (using add_indices as the index), without creating an intermediary
// tensor to hold the selected embeddings
template<typename T>
static void index_select_add(const Tensor &select_indices,
const Tensor &add_indices,
const Tensor &src,
Tensor &output) {
auto add_indices_data = add_indices.data<int64_t>();
auto select_indices_data = select_indices.data<int64_t>();
auto src_data = src.data<T>();
auto output_data = output.data<T>();
auto numel = add_indices.numel();
int64_t ddim = src.sizes()[1];
for (int64_t i = 0; i < numel; i++) {
axpy<T>(ddim, 1, src_data + ddim * select_indices_data[i], 1,
output_data + ddim * add_indices_data[i], 1);
}
}

static void make_bag_size(const Tensor &offsets, const Tensor &indices,
const int64_t mode, Tensor &bag_size) {
if (mode == 1) { // MODE_MEAN
Expand Down Expand Up @@ -61,10 +94,10 @@ static Tensor apply_bag_size_backward(const Tensor &offsets,
auto bag_size_ = indices.sizes()[0];
output /= bag_size_;
} else {
auto bag_size_ = bag_size.toType(output.type())
.unsqueeze(1)
.index_select(0, offset2bag);
output /= bag_size_;
auto inv_bag_size_ = (1 / bag_size.toType(output.type()))
.unsqueeze(1)
.index_select(0, offset2bag);
output *= inv_bag_size_;
}
}
return output;
Expand All @@ -80,14 +113,19 @@ embedding_bag_cpu(const Tensor &weight, const Tensor &indices__,
checkScalarType("embedding_bag", offsets_arg, kLong);
Tensor indices = indices__.contiguous();
Tensor offsets = offsets__.contiguous();
auto weight_arg = TensorArg(weight, "weight", 1);
checkScalarTypes("embedding_bag", weight_arg, {kFloat, kDouble});

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view


auto bag_size = at::zeros(indices.type(), offsets.sizes());
auto offset2bag =
at::zeros(indices__.type(), {indices.sizes()[0]}); // offset2bag = [0 0 0 0 0]
make_offset2bag(offsets, indices, offset2bag);
auto output = at::zeros(weight.type(), {offsets.sizes()[0], weight.sizes()[1]});
auto index_output = weight.index_select(0, indices);
output.index_add_(0, offset2bag, index_output);
if (weight.type().scalarType() == kFloat) {
index_select_add<float>(indices, offset2bag, weight, output);
} else if (weight.type().scalarType() == kDouble) {
index_select_add<double>(indices, offset2bag, weight, output);
}
make_bag_size(offsets, indices, mode, bag_size);
auto ret = apply_bag_size(offsets, indices, mode, output, bag_size);
return std::tuple<Tensor, Tensor, Tensor>(ret, offset2bag, bag_size);
Expand Down Expand Up @@ -126,6 +164,8 @@ Tensor embedding_bag_backward_cpu(const Tensor &grad_, const Tensor &indices__,
const Tensor &bag_size_, int64_t num_weights,
bool scale_grad_by_freq, int64_t mode) {
auto grad = grad_.contiguous();
auto grad_arg = TensorArg(grad, "grad_", 1);
checkScalarTypes("embedding_bag", grad_arg, {kFloat, kDouble});
auto indices_arg = TensorArg(indices__, "indices__", 1);
checkScalarType("embedding_bag", indices_arg, kLong);
auto offsets_arg = TensorArg(offsets__, "offsets__", 1);
Expand Down Expand Up @@ -196,15 +236,13 @@ Tensor embedding_bag_backward_cpu(const Tensor &grad_, const Tensor &indices__,
if (grad.type().scalarType() == kFloat) {
auto igwd = index_grad_weight.data<float>();
auto gd = grad.data<float>();
THFloatBlas_axpy(ddim, (float)scale, gd + ddim * source, 1,
igwd + ddim * index, 1);
axpy<float>(ddim, (float)scale, gd + ddim * source, 1,
igwd + ddim * index, 1);
} else if (grad.type().scalarType() == kDouble) {
auto igwd = index_grad_weight.data<double>();
auto gd = grad.data<double>();
THDoubleBlas_axpy(ddim, (double)scale, gd + ddim * source, 1,
igwd + ddim * index, 1);
} else {
index_grad_weight[index].add_(grad[source], scale);
axpy<double>(ddim, (double)scale, gd + ddim * source, 1,
igwd + ddim * index, 1);
}
}
}
Expand Down