Skip to content

[ET][Portable][Build Size] Reduce build size of op_addmm #6018

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

Closed
wants to merge 5 commits into from
Closed

[ET][Portable][Build Size] Reduce build size of op_addmm #6018

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
bea830d
[ET][Portable][Build Size] Reduce build size of op_addmm
manuelcandales Oct 8, 2024
ec2d190
Update on "[ET][Portable][Build Size] Reduce build size of op_addmm"
manuelcandales Oct 8, 2024
d0886ef
Update on "[ET][Portable][Build Size] Reduce build size of op_addmm"
manuelcandales Oct 8, 2024
8d2beae
Update on "[ET][Portable][Build Size] Reduce build size of op_addmm"
manuelcandales Oct 9, 2024
53082ba
Update on "[ET][Portable][Build Size] Reduce build size of op_addmm"
manuelcandales Oct 9, 2024
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
99 changes: 45 additions & 54 deletions kernels/portable/cpu/op_addmm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
*/

#include <executorch/kernels/portable/cpu/scalar_utils.h>
#include <executorch/kernels/portable/cpu/util/broadcast_util.h>
#include <executorch/kernels/portable/cpu/util/elementwise_util.h>
#include <executorch/kernels/portable/cpu/util/matmul_ops_util.h>
#include <executorch/kernels/portable/cpu/vec_ops.h>
#include <executorch/runtime/kernel/kernel_includes.h>
Expand Down Expand Up @@ -53,62 +53,53 @@ Tensor& addmm_out(

ET_KERNEL_CHECK(ctx, tensor_is_default_dim_order(in), InvalidArgument, out);

ScalarType alpha_dtype = utils::get_scalar_dtype(alpha);
ScalarType beta_dtype = utils::get_scalar_dtype(beta);
ET_SWITCH_REAL_TYPES_AND(
Half, in.scalar_type(), ctx, "addmm.out", CTYPE, [&]() {
ET_SWITCH_SCALAR_OBJ_TYPES(
alpha_dtype, ctx, "addmm.out", ALPHA_T, [&]() {
ET_SWITCH_SCALAR_OBJ_TYPES(
beta_dtype, ctx, "addmm.out", BETA_T, [&]() {
size_t m = mat1.size(0);
size_t n = mat1.size(1);
size_t p = mat2.size(1);
static constexpr const char op_name[] = "addmm.out";

if (out.sizes() == in.sizes()) {
// vec_addmm assumes that no broadcasting is required.
vec_addmm<CTYPE, CTYPE>(
out.mutable_data_ptr<CTYPE>(),
in.const_data_ptr<CTYPE>(),
mat1.const_data_ptr<CTYPE>(),
mat2.const_data_ptr<CTYPE>(),
m,
n,
p,
convert<CTYPE>(beta.to<BETA_T>()),
convert<CTYPE>(alpha.to<ALPHA_T>()));
} else {
// If broadcasting is required, them compute the matmul
// and addition separately, using
// apply_binary_elementwise_fn to perform the addition
// while applying broadcasting
vec_matmul<CTYPE, CTYPE>(
out.mutable_data_ptr<CTYPE>(),
mat1.const_data_ptr<CTYPE>(),
mat2.const_data_ptr<CTYPE>(),
m,
n,
p);
ET_SWITCH_REALHBF16_TYPES(in.scalar_type(), ctx, op_name, CTYPE, [&]() {
CTYPE alpha_val = utils::scalar_to<CTYPE>(alpha);
CTYPE beta_val = utils::scalar_to<CTYPE>(beta);
size_t m = mat1.size(0);
size_t n = mat1.size(1);
size_t p = mat2.size(1);

CTYPE alpha_val = convert<CTYPE>(alpha.to<ALPHA_T>());
CTYPE beta_val = convert<CTYPE>(beta.to<BETA_T>());
apply_binary_elementwise_fn<CTYPE, CTYPE, CTYPE>(
[alpha_val, beta_val](
const CTYPE val_a, const CTYPE val_b) {
CTYPE a_casted = static_cast<CTYPE>(val_a);
CTYPE b_casted = static_cast<CTYPE>(val_b);
CTYPE value =
a_casted * alpha_val + b_casted * beta_val;
if (out.sizes() == in.sizes()) {
// vec_addmm assumes that no broadcasting is required.
vec_addmm<CTYPE, CTYPE>(
out.mutable_data_ptr<CTYPE>(),
in.const_data_ptr<CTYPE>(),
mat1.const_data_ptr<CTYPE>(),
mat2.const_data_ptr<CTYPE>(),
m,
n,
p,
beta_val,
alpha_val);
} else {
// If broadcasting is required, them compute the matmul
// and addition separately, using
// apply_binary_elementwise_fn to perform the addition
// while applying broadcasting
vec_matmul<CTYPE, CTYPE>(
out.mutable_data_ptr<CTYPE>(),
mat1.const_data_ptr<CTYPE>(),
mat2.const_data_ptr<CTYPE>(),
m,
n,
p);

return value;
},
out,
in,
out);
}
});
});
});
utils::apply_bitensor_elementwise_fn<CTYPE, op_name>(
[alpha_val, beta_val](const CTYPE val_a, const CTYPE val_b) {
return val_a * alpha_val + val_b * beta_val;
},
ctx,
out,
utils::SupportedTensorDtypes::REALHBF16,
in,
utils::SupportedTensorDtypes::REALHBF16,
out,
utils::SupportedTensorDtypes::REALHBF16);
}
});

return out;
}
Expand Down
2 changes: 2 additions & 0 deletions shim/xplat/executorch/kernels/portable/op_registration_util.bzl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -224,6 +224,8 @@ ATEN_OPS = (
name = "op_addmm",
deps = [
"//executorch/kernels/portable/cpu/util:broadcast_util",
"//executorch/kernels/portable/cpu/util:dtype_util",
"//executorch/kernels/portable/cpu/util:elementwise_util",
"//executorch/kernels/portable/cpu/util:matmul_ops_util",
":scalar_utils",
":vec_ops",
Expand Down
Loading