Skip to content

For broadcast, added support to process distinct input dimensions #7107

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
mcremon-meta merged 1 commit into from
Nov 27, 2024
Merged

For broadcast, added support to process distinct input dimensions #7107

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
48 changes: 33 additions & 15 deletions backends/cadence/fusion_g3/operators/op_add.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,27 +76,45 @@ Tensor& add_out(
int inp2_shape[kTensorDimensionLimit];
int out_shape[kTensorDimensionLimit];

/* input shapes and output shapes */
for (auto i = 0; i < a_size.size(); i++) {
inp1_shape[i] = a_size[i];
}

for (auto i = 0; i < b_size.size(); i++) {
inp2_shape[i] = b_size[i];
}

for (auto i = 0; i < out_size.size(); i++) {
out_shape[i] = out_size[i];
}

/*find broadcast*/
const bool a_is_broadcasted = !out.sizes().equals(a.sizes());
const bool b_is_broadcasted = !out.sizes().equals(b.sizes());
const bool broadcast = (a_is_broadcasted || b_is_broadcasted);

int max_dim = a.dim() > b.dim() ? a.dim() : b.dim();
max_dim = out.dim() > max_dim ? out.dim() : max_dim;

if (compute_type == ScalarType::Int) {
bool optimized = 1;

if ((a.dim() == 0) || (b.dim() == 0)) {
optimized = 0;
}

if ((broadcast == 1) && (max_dim > kTensorDimensionLimit)) {
optimized = 0;
}

for (int i = 0; i < max_dim; i++) {
out_shape[i] = 1;
inp1_shape[i] = 1;
inp2_shape[i] = 1;
}

int offset_out = max_dim - out.dim();
int offset_inp1 = max_dim - a.dim();
int offset_inp2 = max_dim - b.dim();

for (int i = 0; i < out.dim(); i++) {
out_shape[i + offset_out] = out.size(i);
}
for (int i = 0; i < a.dim(); i++) {
inp1_shape[i + offset_inp1] = a.size(i);
}
for (int i = 0; i < b.dim(); i++) {
inp2_shape[i + offset_inp2] = b.size(i);
}

if ((compute_type == ScalarType::Int) && (optimized)){
const int* const inp1_data = a.const_data_ptr<int>();
const int* const inp2_data = b.const_data_ptr<int>();
int* const out_data = out.mutable_data_ptr<int>();
Expand All @@ -117,7 +135,7 @@ Tensor& add_out(
xa_nn_elm_add_32x32_32(
out_data, inp1_data, inp2_data, alpha_val, out.numel());
}
} else if (compute_type == ScalarType::Float) {
} else if ((compute_type == ScalarType::Float) && (optimized)) {
const float* const inp1_data = a.const_data_ptr<float>();
const float* const inp2_data = b.const_data_ptr<float>();
float* const out_data = out.mutable_data_ptr<float>();
Expand Down
48 changes: 33 additions & 15 deletions backends/cadence/fusion_g3/operators/op_mul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,27 +68,45 @@ Tensor& mul_out(
int inp2_shape[kTensorDimensionLimit];
int out_shape[kTensorDimensionLimit];

/* input shapes and output shapes */
for (auto i = 0; i < a_size.size(); i++) {
inp1_shape[i] = a_size[i];
}

for (auto i = 0; i < b_size.size(); i++) {
inp2_shape[i] = b_size[i];
}

for (auto i = 0; i < out_size.size(); i++) {
out_shape[i] = out_size[i];
}

/*find broadcast*/
const bool a_is_broadcasted = !out.sizes().equals(a.sizes());
const bool b_is_broadcasted = !out.sizes().equals(b.sizes());
const bool broadcast = (a_is_broadcasted || b_is_broadcasted);

int max_dim = a.dim() > b.dim() ? a.dim() : b.dim();
max_dim = out.dim() > max_dim ? out.dim() : max_dim;

if (compute_type == ScalarType::Int) {
bool optimized = 1;

if ((a.dim() == 0) || (b.dim() == 0)) {
optimized = 0;
}

if ((broadcast == 1) && (max_dim > kTensorDimensionLimit)) {
optimized = 0;
}

for (int i = 0; i < max_dim; i++) {
out_shape[i] = 1;
inp1_shape[i] = 1;
inp2_shape[i] = 1;
}

int offset_out = max_dim - out.dim();
int offset_inp1 = max_dim - a.dim();
int offset_inp2 = max_dim - b.dim();

for (int i = 0; i < out.dim(); i++) {
out_shape[i + offset_out] = out.size(i);
}
for (int i = 0; i < a.dim(); i++) {
inp1_shape[i + offset_inp1] = a.size(i);
}
for (int i = 0; i < b.dim(); i++) {
inp2_shape[i + offset_inp2] = b.size(i);
}

if ((compute_type == ScalarType::Int) && (optimized)) {
const int* const inp1_data = a.const_data_ptr<int>();
const int* const inp2_data = b.const_data_ptr<int>();
int* const out_data = out.mutable_data_ptr<int>();
Expand All @@ -105,7 +123,7 @@ Tensor& mul_out(
} else {
xa_nn_elm_mul_32x32_32(out_data, inp1_data, inp2_data, out.numel());
}
} else if (compute_type == ScalarType::Float) {
} else if ((compute_type == ScalarType::Float) && (optimized)) {
const float* const inp1_data = a.const_data_ptr<float>();
const float* const inp2_data = b.const_data_ptr<float>();
float* const out_data = out.mutable_data_ptr<float>();
Expand Down
Loading