Skip to content

Enable HiFi SIMD for FULLY_CONNECTED operator #3086

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Enable HiFi SIMD for FULLY_CONNECTED operator #3086

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
12 changes: 9 additions & 3 deletions tensorflow/lite/micro/kernels/fully_connected.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,14 +117,20 @@ TFLMRegistration Register_FULLY_CONNECTED_INT4();
// define fallback implementation that allow reference kernels to still be used
// from applications that call a more specific kernel variant.

inline TFLMRegistration Register_FULLY_CONNECTED_INT16() {
inline TFLMRegistration Register_FULLY_CONNECTED_INT4() {
return Register_FULLY_CONNECTED();
}

inline TFLMRegistration Register_FULLY_CONNECTED_INT4() {
#endif

#if defined(XTENSA)
// Returns a TFLMRegistration struct for kernel variant that only supports
// float32.
TFLMRegistration Register_FULLY_CONNECTED_FLOAT32();
#else
inline TFLMRegistration Register_FULLY_CONNECTED_FLOAT32() {
return Register_FULLY_CONNECTED();
}

#endif

} // namespace tflite
Expand Down
56 changes: 4 additions & 52 deletions tensorflow/lite/micro/kernels/xtensa/fully_connected.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,8 +33,6 @@ namespace {

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
TFLITE_DCHECK(node->builtin_data != nullptr);
const auto* params =
static_cast<const TfLiteFullyConnectedParams*>(node->builtin_data);

const TfLiteEvalTensor* input =
tflite::micro::GetEvalInput(context, node, kFullyConnectedInputTensor);
Expand All @@ -45,18 +43,6 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
TfLiteEvalTensor* output =
tflite::micro::GetEvalOutput(context, node, kFullyConnectedOutputTensor);

#ifdef USE_TFLM_COMPRESSION

MicroContext* micro_context = GetMicroContext(context);

const CompressionTensorData* weights_comp_td =
micro_context->GetTensorCompressionData(node,
kFullyConnectedWeightsTensor);
const CompressionTensorData* bias_comp_td =
micro_context->GetTensorCompressionData(node, kFullyConnectedBiasTensor);

#endif // USE_TFLM_COMPRESSION

TFLITE_DCHECK(node->user_data != nullptr);

const auto& data =
Expand All @@ -65,25 +51,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
// Checks in Prepare ensure input, output and filter types are all the same.
switch (input->type) {
case kTfLiteFloat32: {
tflite::reference_ops::FullyConnected(
FullyConnectedParamsFloat(params->activation),
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<float>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<float>(micro_context, filter,
weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<float>(
micro_context, bias, bias_comp_td, data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<float>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<float>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<float>(output));
return XtensaEvalFullyConnectedF32(
context, node, data, input, filter, bias, output);
break;
}

Expand All @@ -109,25 +78,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
case kTfLiteInt16: {
switch (filter->type) {
case kTfLiteInt8: {
tflite::reference_integer_ops::FullyConnected(
FullyConnectedParamsQuantized(data),
tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter),
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(micro_context, filter,
weights_comp_td,
data.weights_scratch_index),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(
micro_context, bias, bias_comp_td, data.bias_scratch_index),
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorShape(bias),
tflite::micro::GetOptionalTensorData<int64_t>(bias),
#endif // USE_TFLM_COMPRESSION
tflite::micro::GetTensorShape(output),
tflite::micro::GetTensorData<int16_t>(output));
return XtensaEvalFullyConnectedQuantizedInt16(
context, node, data, input, filter, bias, output);
break;
}
default: {
Expand Down
142 changes: 142 additions & 0 deletions tensorflow/lite/micro/kernels/xtensa/fully_connected_f32.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,142 @@
/* Copyright 2025 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "tensorflow/lite/c/builtin_op_data.h"
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/portable_tensor_utils.h"
#include "tensorflow/lite/kernels/internal/quantization_util.h"
#include "tensorflow/lite/kernels/internal/reference/fully_connected.h"
#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "tensorflow/lite/kernels/kernel_util.h"
#include "tensorflow/lite/micro/kernels/kernel_util.h"
#include "tensorflow/lite/micro/kernels/xtensa/xtensa.h"
#include "tensorflow/lite/micro/kernels/xtensa/xtensa_fully_connected.h"

namespace tflite {

TfLiteStatus XtensaEvalFullyConnectedF32(
TfLiteContext* context, TfLiteNode* node, const OpDataFullyConnected& data,
const TfLiteEvalTensor* input, const TfLiteEvalTensor* filter,
const TfLiteEvalTensor* bias, TfLiteEvalTensor* output) {
#if !defined(VISION_P6)

#ifdef USE_TFLM_COMPRESSION

MicroContext* micro_context = GetMicroContext(context);

const CompressionTensorData* weights_comp_td =
micro_context->GetTensorCompressionData(node,
kFullyConnectedWeightsTensor);
const CompressionTensorData* bias_comp_td =
micro_context->GetTensorCompressionData(node, kFullyConnectedBiasTensor);

#endif // USE_TFLM_COMPRESSION

const float* bias_data =
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetOptionalTensorData<float>(
micro_context, bias, bias_comp_td, data.bias_scratch_index);
#else // USE_TFLM_COMPRESSION
tflite::micro::GetOptionalTensorData<float>(bias);
#endif // USE_TFLM_COMPRESSION

const float* filter_data =
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<float>(
micro_context, filter, weights_comp_td, data.weights_scratch_index);
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<float>(filter);
#endif // USE_TFLM_COMPRESSION

#endif // !defined(VISION_P6)

const auto* node_data =
static_cast<const TfLiteFullyConnectedParams*>(node->builtin_data);
const FullyConnectedParams params =
FullyConnectedParamsFloat(node_data->activation);
const float* input_data = tflite::micro::GetTensorData<float>(input);
float* output_data = tflite::micro::GetTensorData<float>(output);

#if HIFI_VFPU && (defined(HIFI3) || defined(HIFI4) || defined(HIFI5))
const RuntimeShape& output_shape = tflite::micro::GetTensorShape(output);
const int num_batches =
FlatSizeSkipDim(output_shape, output_shape.DimensionsCount() - 1);
const int output_depth =
output_shape.Dims(output_shape.DimensionsCount() - 1);

const RuntimeShape& filter_shape = tflite::micro::GetTensorShape(filter);
const int filter_dim_count = filter_shape.DimensionsCount();
const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
if (num_batches == 1) {
TF_LITE_ENSURE_EQ(
context,
xa_nn_fully_connected_f32(output_data, filter_data, input_data,
bias_data, accum_depth, output_depth),
0);
} else {
TF_LITE_ENSURE_EQ(context,
xa_nn_matmul_f32xf32_f32(
output_data, filter_data, input_data, bias_data,
output_depth, accum_depth, accum_depth, num_batches,
accum_depth, output_depth, 1),
0);
}
TF_LITE_ENSURE_EQ(
context,
xa_nn_vec_activation_min_max_f32_f32(
output_data, output_data, params.float_activation_min,
params.float_activation_max, num_batches * output_depth),
0);
#else
tflite::reference_ops::FullyConnected(
params, tflite::micro::GetTensorShape(input), input_data,
tflite::micro::GetTensorShape(filter), filter_data,
tflite::micro::GetTensorShape(bias), bias_data,
tflite::micro::GetTensorShape(output), output_data);
#endif // HIFI_VFPU && (defined(HIFI3) || defined(HIFI4) || defined(HIFI5))

return kTfLiteOk;
}

namespace {

TfLiteStatus EvalFloat32(TfLiteContext* context, TfLiteNode* node) {
TFLITE_DCHECK(node->user_data != nullptr);
const auto& data =
*(static_cast<const OpDataFullyConnected*>(node->user_data));

const TfLiteEvalTensor* input =
tflite::micro::GetEvalInput(context, node, kFullyConnectedInputTensor);
const TfLiteEvalTensor* filter =
tflite::micro::GetEvalInput(context, node, kFullyConnectedWeightsTensor);
const TfLiteEvalTensor* bias =
tflite::micro::GetEvalInput(context, node, kFullyConnectedBiasTensor);

TfLiteEvalTensor* output =
tflite::micro::GetEvalOutput(context, node, kFullyConnectedOutputTensor);

return XtensaEvalFullyConnectedF32(context, node, data, input, filter, bias,
output);
}

} // namespace

TFLMRegistration Register_FULLY_CONNECTED_FLOAT32() {
return tflite::micro::RegisterOp(XtensaInitFullyConnected,
XtensaPrepareFullyConnected, EvalFloat32);
}

} // namespace tflite
146 changes: 146 additions & 0 deletions tensorflow/lite/micro/kernels/xtensa/fully_connected_int16.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,146 @@
/* Copyright 2025 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include "tensorflow/lite/c/builtin_op_data.h"
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/kernels/internal/common.h"
#include "tensorflow/lite/kernels/internal/portable_tensor_utils.h"
#include "tensorflow/lite/kernels/internal/quantization_util.h"
#include "tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h"
#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "tensorflow/lite/kernels/kernel_util.h"
#include "tensorflow/lite/micro/kernels/kernel_util.h"
#include "tensorflow/lite/micro/kernels/xtensa/xtensa.h"
#include "tensorflow/lite/micro/kernels/xtensa/xtensa_fully_connected.h"

namespace tflite {

TfLiteStatus XtensaEvalFullyConnectedQuantizedInt16(
TfLiteContext* context, TfLiteNode* node, const OpDataFullyConnected& data,
const TfLiteEvalTensor* input, const TfLiteEvalTensor* filter,
const TfLiteEvalTensor* bias, TfLiteEvalTensor* output) {
#if !defined(VISION_P6)

#ifdef USE_TFLM_COMPRESSION

MicroContext* micro_context = GetMicroContext(context);

const CompressionTensorData* weights_comp_td =
micro_context->GetTensorCompressionData(node,
kFullyConnectedWeightsTensor);
const CompressionTensorData* bias_comp_td =
micro_context->GetTensorCompressionData(node, kFullyConnectedBiasTensor);

#endif // USE_TFLM_COMPRESSION

const int64_t* bias_data =
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetOptionalTensorData<int64_t>(
micro_context, bias, bias_comp_td, data.bias_scratch_index);
#else // USE_TFLM_COMPRESSION
tflite::micro::GetOptionalTensorData<int64_t>(bias);
#endif // USE_TFLM_COMPRESSION

const int8_t* filter_data =
#ifdef USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(
micro_context, filter, weights_comp_td, data.weights_scratch_index);
#else // USE_TFLM_COMPRESSION
tflite::micro::GetTensorData<int8_t>(filter);
#endif // USE_TFLM_COMPRESSION

#endif // !defined(VISION_P6)

int16_t* output_data = tflite::micro::GetTensorData<int16_t>(output);

#if defined(HIFI3) || defined(HIFI4) || defined(HIFI5)
const RuntimeShape& output_shape = tflite::micro::GetTensorShape(output);
const int num_batches =
FlatSizeSkipDim(output_shape, output_shape.DimensionsCount() - 1);
const int output_depth =
output_shape.Dims(output_shape.DimensionsCount() - 1);

const RuntimeShape& filter_shape = tflite::micro::GetTensorShape(filter);
const int filter_dim_count = filter_shape.DimensionsCount();
const int accum_depth = filter_shape.Dims(filter_dim_count - 1);

FullyConnectedParams op_params = FullyConnectedParamsQuantized(data);
if (num_batches == 1) {
TF_LITE_ENSURE_EQ(context,
xa_nn_fully_connected_sym8sxsym16s_sym16s(
output_data, filter_data,
tflite::micro::GetTensorData<int16_t>(input),
bias_data, accum_depth, output_depth,
op_params.output_multiplier, op_params.output_shift),
0);
} else {
TF_LITE_ENSURE_EQ(context,
xa_nn_matmul_sym8sxsym16s_sym16s(
tflite::micro::GetTensorData<int16_t>(output),
tflite::micro::GetTensorData<int8_t>(filter),
tflite::micro::GetTensorData<int16_t>(input),
bias_data, output_depth, accum_depth, accum_depth,
num_batches, accum_depth, output_depth, 1,
op_params.input_offset, op_params.output_multiplier,
op_params.output_shift, op_params.output_offset),
0);
}

TF_LITE_ENSURE_EQ(context,
xa_nn_vec_activation_min_max_16_16(
output_data, output_data, data.output_activation_min,
data.output_activation_max, num_batches * output_depth),
0);
#else
tflite::reference_integer_ops::FullyConnected(
FullyConnectedParamsQuantized(data), tflite::micro::GetTensorShape(input),
tflite::micro::GetTensorData<int16_t>(input),
tflite::micro::GetTensorShape(filter), filter_data,
tflite::micro::GetTensorShape(bias), bias_data,
tflite::micro::GetTensorShape(output), output_data);
#endif // defined(HIFI3) || defined(HIFI4) || defined(HIFI5)

return kTfLiteOk;
}

namespace {

TfLiteStatus EvalInt16(TfLiteContext* context, TfLiteNode* node) {
TFLITE_DCHECK(node->user_data != nullptr);
const auto& data =
*(static_cast<const OpDataFullyConnected*>(node->user_data));

const TfLiteEvalTensor* input =
tflite::micro::GetEvalInput(context, node, kFullyConnectedInputTensor);
const TfLiteEvalTensor* filter =
tflite::micro::GetEvalInput(context, node, kFullyConnectedWeightsTensor);
const TfLiteEvalTensor* bias =
tflite::micro::GetEvalInput(context, node, kFullyConnectedBiasTensor);

TfLiteEvalTensor* output =
tflite::micro::GetEvalOutput(context, node, kFullyConnectedOutputTensor);

return XtensaEvalFullyConnectedQuantizedInt16(context, node, data, input,
filter, bias, output);
}

} // namespace

TFLMRegistration Register_FULLY_CONNECTED_INT16() {
return tflite::micro::RegisterOp(XtensaInitFullyConnected,
XtensaPrepareFullyConnected, EvalInt16);
}

} // namespace tflite
Loading