Skip to content

FusionG3 operators. #7315

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
merged 5 commits into from
Dec 12, 2024
Merged

FusionG3 operators. #7315

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
b354f46
Allow backends/cadence to use TestUtil.h. (#7304)
hsharma35 Dec 12, 2024
4145467
Xtensa ISS PAL layer for logging/timing. (#7311)
hsharma35 Dec 12, 2024
9d2157e
Use macro `XT_KERNEL_CHECK` to handle errors returned by nnlib. (#7312)
hsharma35 Dec 12, 2024
0a2e2f3
Separate buck targets per operator. (#7314)
hsharma35 Dec 12, 2024
9d9f728
FusionG3 operators. (#7315)
hsharma35 Dec 12, 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
124 changes: 99 additions & 25 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 @@ -6,25 +6,37 @@
* LICENSE file in the root directory of this source tree.
*/

#include <executorch/backends/cadence/fusion_g3/operators/operators.h>

#include <xa_nnlib_kernels_api.h>

#include <executorch/kernels/portable/cpu/scalar_utils.h>
#include <executorch/kernels/portable/cpu/util/elementwise_util.h>
#include <executorch/kernels/portable/cpu/util/kernel_ops_util.h>
#include <executorch/runtime/kernel/kernel_includes.h>
#include <executorch/runtime/platform/assert.h>
#include <xa_nnlib_kernels_api.h>

using exec_aten::Scalar;
using exec_aten::ScalarType;
using exec_aten::Tensor;
using executorch::runtime::canCast;
using torch::executor::Error;
using torch::executor::KernelRuntimeContext;
using ::executorch::aten::Scalar;
using ::executorch::aten::ScalarType;
using ::executorch::aten::Tensor;
using ::executorch::runtime::canCast;
using ::executorch::runtime::Error;
using ::executorch::runtime::KernelRuntimeContext;

namespace cadence {
namespace impl {
namespace G3 {
namespace native {

#define XT_KERNEL_CHECK(ctx, out, kernel, ...) \
const auto ret = kernel(__VA_ARGS__); \
ET_KERNEL_CHECK_MSG( \
ctx, \
ret == 0, \
InvalidArgument, \
out, \
"Failed to run kernel: " #kernel "(" #__VA_ARGS__ ")");

Tensor& add_out(
KernelRuntimeContext& ctx,
const Tensor& a,
Expand Down Expand Up @@ -121,13 +133,30 @@ Tensor& add_out(
torch::executor::native::utils::extract_scalar(alpha, &alpha_val);

if ((a.numel() == 1) && (alpha_val == 1)) {
xa_nn_elm_add_scalar_32x32_32(
out_data, inp2_data, inp1_data[0], alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_scalar_32x32_32,
out_data,
inp2_data,
inp1_data[0],
alpha_val,
out.numel());
} else if (b.numel() == 1) {
xa_nn_elm_add_scalar_32x32_32(
out_data, inp1_data, inp2_data[0], alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_scalar_32x32_32,
out_data,
inp1_data,
inp2_data[0],
alpha_val,
out.numel());
} else if (broadcast) {
xa_nn_elm_add_broadcast_5D_32x32_32(
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_broadcast_5D_32x32_32,
out_data,
out_shape,
inp1_data,
Expand All @@ -137,8 +166,15 @@ Tensor& add_out(
max_dim,
alpha_val);
} else {
xa_nn_elm_add_32x32_32(
out_data, inp1_data, inp2_data, alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_32x32_32,
out_data,
inp1_data,
inp2_data,
alpha_val,
out.numel());
}
} else if ((compute_type == ScalarType::Float) && (optimized)) {
const float* const inp1_data = a.const_data_ptr<float>();
Expand All @@ -149,13 +185,30 @@ Tensor& add_out(
torch::executor::native::utils::extract_scalar(alpha, &alpha_val);

if ((a.numel() == 1) && (alpha_val == 1.0)) {
xa_nn_elm_add_scalar_f32xf32_f32(
out_data, inp2_data, inp1_data[0], alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_scalar_f32xf32_f32,
out_data,
inp2_data,
inp1_data[0],
alpha_val,
out.numel());
} else if (b.numel() == 1) {
xa_nn_elm_add_scalar_f32xf32_f32(
out_data, inp1_data, inp2_data[0], alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_scalar_f32xf32_f32,
out_data,
inp1_data,
inp2_data[0],
alpha_val,
out.numel());
} else if (broadcast) {
xa_nn_elm_add_broadcast_5D_f32xf32_f32(
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_broadcast_5D_f32xf32_f32,
out_data,
out_shape,
inp1_data,
Expand All @@ -165,8 +218,15 @@ Tensor& add_out(
max_dim,
alpha_val);
} else {
xa_nn_elm_add_f32xf32_f32(
out_data, inp1_data, inp2_data, alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_f32xf32_f32,
out_data,
inp1_data,
inp2_data,
alpha_val,
out.numel());
}
} else {
ET_SWITCH_REALB_TYPES(compute_type, ctx, op_name, CTYPE_COMPUTE, [&]() {
Expand Down Expand Up @@ -242,8 +302,15 @@ Tensor& add_scalar_out(

int* const out_data = out.mutable_data_ptr<int>();

xa_nn_elm_add_scalar_32x32_32(
out_data, inp1_data, inp2_val, alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_scalar_32x32_32,
out_data,
inp1_data,
inp2_val,
alpha_val,
out.numel());

} else if (compute_type == ScalarType::Float) {
const float* const inp1_data = a.const_data_ptr<float>();
Expand All @@ -255,8 +322,15 @@ Tensor& add_scalar_out(

float* const out_data = out.mutable_data_ptr<float>();

xa_nn_elm_add_scalar_f32xf32_f32(
out_data, inp1_data, inp2_val, alpha_val, out.numel());
XT_KERNEL_CHECK(
ctx,
out,
xa_nn_elm_add_scalar_f32xf32_f32,
out_data,
inp1_data,
inp2_val,
alpha_val,
out.numel());

} else {
ET_SWITCH_REALB_TYPES(compute_type, ctx, op_name, CTYPE_COMPUTE, [&]() {
Expand Down
17 changes: 9 additions & 8 deletions backends/cadence/fusion_g3/operators/op_cat.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,16 +6,17 @@
* LICENSE file in the root directory of this source tree.
*/

#include <cstring>

#include <xa_nnlib_kernels_api.h>

#include <executorch/kernels/portable/cpu/util/copy_ops_util.h>
#include <executorch/runtime/kernel/kernel_includes.h>
#include <xa_nnlib_kernels_api.h>
#include <cstring>

using exec_aten::Scalar;
using exec_aten::ScalarType;
using exec_aten::Tensor;
using torch::executor::Error;
using torch::executor::KernelRuntimeContext;
using ::executorch::aten::ScalarType;
using ::executorch::aten::Tensor;
using ::executorch::runtime::Error;
using ::executorch::runtime::KernelRuntimeContext;

/* ScalarType in Executorch do not have support for below data types.
* So, creating a placeholder for these data types. Once, ScalarTypes is
Expand Down Expand Up @@ -194,4 +195,4 @@ Tensor& cat_out(
} // namespace native
} // namespace G3
} // namespace impl
} // namespace cadence
} // namespace cadence
24 changes: 13 additions & 11 deletions backends/cadence/fusion_g3/operators/op_dequantize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,18 +6,20 @@
* LICENSE file in the root directory of this source tree.
*/

#include <executorch/kernels/portable/cpu/util/reduce_util.h>
#include <executorch/runtime/kernel/kernel_includes.h>
#include <xa_nnlib_kernels_api.h>
#include <algorithm>
#include <cinttypes>
#include <cmath>

using exec_aten::Scalar;
using exec_aten::ScalarType;
using exec_aten::Tensor;
using torch::executor::Error;
using torch::executor::KernelRuntimeContext;
#include <xa_nnlib_kernels_api.h>

#include <executorch/kernels/portable/cpu/util/reduce_util.h>
#include <executorch/runtime/kernel/kernel_includes.h>

using ::executorch::aten::Scalar;
using ::executorch::aten::ScalarType;
using ::executorch::aten::Tensor;
using ::executorch::runtime::Error;
using ::executorch::runtime::KernelRuntimeContext;

template <typename T>
using optional = exec_aten::optional<T>;
Expand Down Expand Up @@ -185,7 +187,7 @@ void dequantize_impl(
if (axis == NULL) {
// calculate the dequantized output, cast scale to float to match fbgemm
// behavior
#define ASYM_DEQUANTIZE_IMPL_TESNOR(IN_CTYPE, OUT_CTYPE, out_dtype) \
#define ASYM_DEQUANTIZE_IMPL_TENSOR(IN_CTYPE, OUT_CTYPE, out_dtype) \
case ScalarType::out_dtype: { \
/* Hoist these function calls out of our inner loop because they might not \
* get inlined without LTO, particularly in ATen mode. */ \
Expand All @@ -201,7 +203,7 @@ void dequantize_impl(
#define ASYM_CALCULATE_INT_TYPE_TENSOR(IN_CTYPE, in_dtype) \
case ScalarType::in_dtype: \
switch (out.scalar_type()) { \
ET_FORALL_FLOAT_TYPES_WITH(IN_CTYPE, ASYM_DEQUANTIZE_IMPL_TESNOR); \
ET_FORALL_FLOAT_TYPES_WITH(IN_CTYPE, ASYM_DEQUANTIZE_IMPL_TENSOR); \
default: \
ET_CHECK_MSG( \
false, \
Expand All @@ -219,7 +221,7 @@ void dequantize_impl(
static_cast<int8_t>(input.scalar_type()));
}
#undef ASYM_CALCULATE_INT_TYPE_TENSOR
#undef ASYM_DEQUANTIZE_IMPL_TESNOR
#undef ASYM_DEQUANTIZE_IMPL_TENSOR
} else {
// a list contains all dimensions except axis
int64_t dims[input.dim() - 1];
Expand Down
17 changes: 9 additions & 8 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 @@ -6,18 +6,19 @@
* LICENSE file in the root directory of this source tree.
*/

#include <xa_nnlib_kernels_api.h>

#include <executorch/kernels/portable/cpu/scalar_utils.h>
#include <executorch/kernels/portable/cpu/util/elementwise_util.h>
#include <executorch/runtime/kernel/kernel_includes.h>
#include <executorch/runtime/platform/assert.h>
#include <xa_nnlib_kernels_api.h>

using exec_aten::Scalar;
using exec_aten::ScalarType;
using exec_aten::Tensor;
using executorch::runtime::canCast;
using torch::executor::Error;
using torch::executor::KernelRuntimeContext;
using ::executorch::aten::Scalar;
using ::executorch::aten::ScalarType;
using ::executorch::aten::Tensor;
using ::executorch::runtime::canCast;
using ::executorch::runtime::Error;
using ::executorch::runtime::KernelRuntimeContext;

namespace cadence {
namespace impl {
Expand Down Expand Up @@ -238,4 +239,4 @@ Tensor& mul_scalar_out(
} // namespace native
} // namespace G3
} // namespace impl
} // namespace cadence
} // namespace cadence
20 changes: 11 additions & 9 deletions backends/cadence/fusion_g3/operators/op_native_layer_norm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,18 +6,20 @@
* LICENSE file in the root directory of this source tree.
*/

#include <cmath>
#include <tuple>

#include <xa_nnlib_kernels_api.h>

#include <executorch/kernels/portable/cpu/util/normalization_ops_util.h>
#include <executorch/kernels/portable/cpu/vec_ops.h>
#include <executorch/runtime/kernel/kernel_includes.h>
#include <xa_nnlib_kernels_api.h>
#include <cmath>
#include <tuple>

using Tensor = exec_aten::Tensor;
using ScalarType = exec_aten::ScalarType;
using IntArrayRef = exec_aten::ArrayRef<int64_t>;
using torch::executor::Error;
using torch::executor::KernelRuntimeContext;
using ::executorch::aten::IntArrayRef;
using ::executorch::aten::ScalarType;
using ::executorch::aten::Tensor;
using ::executorch::runtime::Error;
using ::executorch::runtime::KernelRuntimeContext;

namespace cadence {
namespace impl {
Expand Down Expand Up @@ -255,4 +257,4 @@ std::tuple<Tensor&, Tensor&, Tensor&> native_layer_norm_out(
} // namespace native
} // namespace G3
} // namespace impl
} // namespace cadence
} // namespace cadence
Loading
Loading