Skip to content

Refactor delay kernel for optimized memory layout #3515

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
manucian-official wants to merge 1 commit into from
+80 −66
Closed

Refactor delay kernel for optimized memory layout #3515

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
146 changes: 80 additions & 66 deletions signal/micro/kernels/delay.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,17 +1,5 @@
/* Copyright 2021 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.
==============================================================================*/
Licensed under the Apache License, Version 2.0 */

#include <stdint.h>

Expand All @@ -29,32 +17,28 @@ namespace {

constexpr int kInputTensor = 0;
constexpr int kOutputTensor = 0;

// Indices into the init flexbuffer's vector.
// The parameter's name is in the comment that follows.
// Elements in the vectors are ordered alphabetically by parameter name.
constexpr int kDelayLengthIndex = 0; // 'delay_length'
constexpr int kDelayLengthIndex = 0;

struct TFLMSignalFrontendDelayParams {
int32_t frame_size;
int32_t delay_length;
int32_t outer_dims;

int8_t** state_buffers;
// 🔥 optimized memory layout
int8_t* big_buffer;
tflm_signal::CircularBuffer** circular_buffers;
};

void* DelayInit(TfLiteContext* context, const char* buffer, size_t length) {
auto* params = static_cast<TFLMSignalFrontendDelayParams*>(
context->AllocatePersistentBuffer(context,
sizeof(TFLMSignalFrontendDelayParams)));
context->AllocatePersistentBuffer(
context, sizeof(TFLMSignalFrontendDelayParams)));

if (params == nullptr) {
return nullptr;
}
if (!params) return nullptr;

FlexbufferWrapper fbw(reinterpret_cast<const uint8_t*>(buffer), length);
params->delay_length = fbw.ElementAsInt32(kDelayLengthIndex);

return params;
}

Expand All @@ -63,92 +47,122 @@ TfLiteStatus DelayPrepare(TfLiteContext* context, TfLiteNode* node) {
TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);

MicroContext* micro_context = GetMicroContext(context);

TfLiteTensor* input =
micro_context->AllocateTempInputTensor(node, kInputTensor);
TF_LITE_ENSURE(context, input != nullptr);
TfLiteTensor* output =
micro_context->AllocateTempOutputTensor(node, kOutputTensor);
TF_LITE_ENSURE(context, output != nullptr);

TF_LITE_ENSURE(context, input && output);
TF_LITE_ENSURE_TYPES_EQ(context, input->type, kTfLiteInt16);
TF_LITE_ENSURE_TYPES_EQ(context, output->type, kTfLiteInt16);

auto* params =
reinterpret_cast<TFLMSignalFrontendDelayParams*>(node->user_data);

TF_LITE_ENSURE(context, params != nullptr);

RuntimeShape input_shape = GetTensorShape(input);
int innermost_dim = input_shape.Dims(input_shape.DimensionsCount() - 1);
params->outer_dims = input_shape.FlatSize() / innermost_dim;
params->frame_size = innermost_dim;
RuntimeShape shape = GetTensorShape(input);
int innermost = shape.Dims(shape.DimensionsCount() - 1);

params->state_buffers =
static_cast<int8_t**>(context->AllocatePersistentBuffer(
context, params->outer_dims * sizeof(int8_t*)));
params->circular_buffers = static_cast<tflm_signal::CircularBuffer**>(
context->AllocatePersistentBuffer(
context, params->outer_dims * sizeof(tflm_signal::CircularBuffer*)));

for (int i = 0; i < params->outer_dims; i++) {
size_t capacity = params->frame_size + params->delay_length;

size_t state_size = tflm_signal::CircularBufferGetNeededMemory(capacity);
params->state_buffers[i] =
static_cast<int8_t*>(context->AllocatePersistentBuffer(
context, state_size * sizeof(int8_t)));
params->circular_buffers[i] = tflm_signal::CircularBufferInit(
capacity, params->state_buffers[i], state_size);
tflm_signal::CircularBufferWriteZeros(params->circular_buffers[i],
params->delay_length);
params->frame_size = innermost;
params->outer_dims = shape.FlatSize() / innermost;

TF_LITE_ENSURE(context, params->frame_size > 0);
TF_LITE_ENSURE(context, params->delay_length >= 0);

size_t capacity =
static_cast<size_t>(params->frame_size) +
static_cast<size_t>(params->delay_length);

TF_LITE_ENSURE(context, capacity > params->frame_size); // overflow guard

// allocate pointer array
params->circular_buffers =
static_cast<tflm_signal::CircularBuffer**>(
context->AllocatePersistentBuffer(
context, params->outer_dims * sizeof(void*)));

TF_LITE_ENSURE(context, params->circular_buffers != nullptr);

// compute total memory
size_t single_size =
tflm_signal::CircularBufferGetNeededMemory(capacity);

size_t total_size = single_size * params->outer_dims;

params->big_buffer =
static_cast<int8_t*>(context->AllocatePersistentBuffer(
context, total_size));

TF_LITE_ENSURE(context, params->big_buffer != nullptr);

// init buffers
for (int i = 0; i < params->outer_dims; ++i) {
int8_t* slice = params->big_buffer + i * single_size;

params->circular_buffers[i] =
tflm_signal::CircularBufferInit(capacity, slice, single_size);

TF_LITE_ENSURE(context, params->circular_buffers[i] != nullptr);

tflm_signal::CircularBufferWriteZeros(
params->circular_buffers[i], params->delay_length);
}

micro_context->DeallocateTempTfLiteTensor(input);
micro_context->DeallocateTempTfLiteTensor(output);

return kTfLiteOk;
}

TfLiteStatus DelayEval(TfLiteContext* context, TfLiteNode* node) {
auto* params =
reinterpret_cast<TFLMSignalFrontendDelayParams*>(node->user_data);

const TfLiteEvalTensor* input =
micro::GetEvalInput(context, node, kInputTensor);
TfLiteEvalTensor* output = micro::GetEvalOutput(context, node, kOutputTensor);
TfLiteEvalTensor* output =
micro::GetEvalOutput(context, node, kOutputTensor);

const int16_t* input_data = micro::GetTensorData<int16_t>(input);
int16_t* output_data = micro::GetTensorData<int16_t>(output);

for (int dim_index = 0, sample_index = 0; dim_index < params->outer_dims;
dim_index++, sample_index += params->frame_size) {
tflm_signal::CircularBufferWrite(params->circular_buffers[dim_index],
&input_data[sample_index],
params->frame_size);
tflm_signal::CircularBufferGet(params->circular_buffers[dim_index],
params->frame_size,
&output_data[sample_index]);
tflm_signal::CircularBufferDiscard(params->circular_buffers[dim_index],
params->frame_size);
const int frame = params->frame_size;

for (int i = 0; i < params->outer_dims; ++i) {
auto* cb = params->circular_buffers[i];

const int16_t* in = input_data + i * frame;
int16_t* out = output_data + i * frame;

tflm_signal::CircularBufferWrite(cb, in, frame);
tflm_signal::CircularBufferGet(cb, frame, out);
tflm_signal::CircularBufferDiscard(cb, frame);
}

return kTfLiteOk;
}

void DelayReset(TfLiteContext* context, void* buffer) {
auto* params = static_cast<TFLMSignalFrontendDelayParams*>(buffer);

for (int i = 0; i < params->outer_dims; ++i) {
tflm_signal::CircularBufferReset(params->circular_buffers[i]);
tflm_signal::CircularBufferWriteZeros(params->circular_buffers[i],
params->delay_length);
auto* cb = params->circular_buffers[i];
tflm_signal::CircularBufferReset(cb);
tflm_signal::CircularBufferWriteZeros(cb, params->delay_length);
}
}

} // namespace

namespace tflm_signal {

TFLMRegistration* Register_DELAY() {
static TFLMRegistration r = micro::RegisterOp(DelayInit, DelayPrepare,
DelayEval, nullptr, DelayReset);
static TFLMRegistration r =
micro::RegisterOp(DelayInit, DelayPrepare, DelayEval, nullptr,
DelayReset);
return &r;
}
} // namespace tflm_signal

} // namespace tflm_signal
} // namespace tflite
Loading