Aksal/nuttx-apps
1
0
Fork 1
mirror of https://github.com/apache/nuttx-apps.git synced 2026-08-01 20:29:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

ml: tflm mean optimization patch

Browse source 
Separate the Int8 implementation of the 'mean' operator to reduce the code size.

Signed-off-by: xinhaiteng <xinhaiteng@xiaomi.com>
This commit is contained in:
xinhaiteng 2024-04-25 11:38:19 +08:00 • committed by Xiang Xiao
parent 068b2b16bb
commit dda8cab335
2 changed files with 113 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

112
mlearning/tflite-micro/0003-mean-int8.patch Normal file
View file

@ -0,0 +1,112 @@
From 9507345939722d4bc2f2791394217680fc5c987b Mon Sep 17 00:00:00 2001
From: xinhaiteng <xinhaiteng@xiaomi.com>
Date: Wed, 24 Apr 2024 09:26:54 +0800
Subject: [PATCH] tflite: add mean-int8 operator
VELAPLATFO-30990
Separate the Int8 implementation of the 'mean' operator to reduce the code size.
Change-Id: I5eb9e3cc939668b6531ef9e56f7c32090f1de141
Signed-off-by: xinhaiteng <xinhaiteng@xiaomi.com>
---
tensorflow/lite/micro/kernels/reduce.cc | 9 +++++++++
tensorflow/lite/micro/kernels/reduce.h | 3 +++
tensorflow/lite/micro/kernels/reduce_common.cc | 16 ++++++++++++++++
.../lite/micro/micro_mutable_op_resolver.h | 5 +++--
4 files changed, 31 insertions(+), 2 deletions(-)
diff --git a/tensorflow/lite/micro/kernels/reduce.cc b/tensorflow/lite/micro/kernels/reduce.cc
index a689d3b93..090bafc3c 100644
--- a/tensorflow/lite/micro/kernels/reduce.cc
+++ b/tensorflow/lite/micro/kernels/reduce.cc
@@ -49,6 +49,11 @@ TfLiteStatus EvalMean(TfLiteContext* context, TfLiteNode* node) {
static_cast<OpDataReduce*>(node->user_data));
}
+TfLiteStatus EvalMeanInt8(TfLiteContext* context, TfLiteNode* node) {
+ return EvalMeanHelperInt8(context, node,
+ static_cast<OpDataReduce*>(node->user_data));
+}
+
TfLiteStatus EvalMax(TfLiteContext* context, TfLiteNode* node) {
OpDataReduce* op_data = static_cast<OpDataReduce*>(node->user_data);
return EvalMaxHelper(context, node, op_data);
@@ -63,6 +68,10 @@ TFLMRegistration Register_MEAN() {
return tflite::micro::RegisterOp(InitReduce, PrepareMeanOrSum, EvalMean);
}
+TFLMRegistration Register_MEAN_INT8() {
+ return tflite::micro::RegisterOp(InitReduce, PrepareMeanOrSum, EvalMeanInt8);
+}
+
TFLMRegistration Register_REDUCE_MAX() {
return tflite::micro::RegisterOp(InitReduce, PrepareMax, EvalMax);
}
diff --git a/tensorflow/lite/micro/kernels/reduce.h b/tensorflow/lite/micro/kernels/reduce.h
index 2daeef5fe..0e29b3af8 100644
--- a/tensorflow/lite/micro/kernels/reduce.h
+++ b/tensorflow/lite/micro/kernels/reduce.h
@@ -50,6 +50,8 @@ TfLiteStatus EvalMaxHelper(TfLiteContext* context, TfLiteNode* node,
OpDataReduce* op_data);
TfLiteStatus EvalMeanHelper(TfLiteContext* context, TfLiteNode* node,
OpDataReduce* op_data);
+TfLiteStatus EvalMeanHelperInt8(TfLiteContext* context, TfLiteNode* node,
+ OpDataReduce* op_data);
TfLiteStatus EvalSumHelper(TfLiteContext* context, TfLiteNode* node,
OpDataReduce* op_data);
@@ -57,6 +59,7 @@ void ReduceResolveAxis(const int* axis_data, int axis_count,
MeanParams* op_params);
TFLMRegistration Register_MEAN();
+TFLMRegistration Register_MEAN_INT8();
TFLMRegistration Register_REDUCE_MAX();
TFLMRegistration Register_SUM();
diff --git a/tensorflow/lite/micro/kernels/reduce_common.cc b/tensorflow/lite/micro/kernels/reduce_common.cc
index 2c1a92a50..6e4ab1cc9 100644
--- a/tensorflow/lite/micro/kernels/reduce_common.cc
+++ b/tensorflow/lite/micro/kernels/reduce_common.cc
@@ -236,6 +236,22 @@ TfLiteStatus EvalMeanHelper(TfLiteContext* context, TfLiteNode* node,
return kTfLiteOk;
}
+TfLiteStatus EvalMeanHelperInt8(TfLiteContext* context, TfLiteNode* node,
+ OpDataReduce* op_data) {
+ const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0);
+ const TfLiteEvalTensor* axis = tflite::micro::GetEvalInput(context, node, 1);
+
+ int num_axis = static_cast<int>(ElementCount(*axis->dims));
+ int temp_index[kMaxNumberOfAxis];
+ int resolved_axis[kMaxNumberOfReducedAxis];
+
+ TFLITE_DCHECK(input->type == kTfLiteInt8);
+ TF_LITE_ENSURE_OK(
+ context, EvalIntegerMean<int8_t>(context, node, num_axis, op_data,
+ temp_index, resolved_axis));
+ return kTfLiteOk;
+}
+
TfLiteStatus EvalMaxHelper(TfLiteContext* context, TfLiteNode* node,
OpDataReduce* op_data) {
const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0);
diff --git a/tensorflow/lite/micro/micro_mutable_op_resolver.h b/tensorflow/lite/micro/micro_mutable_op_resolver.h
index ec8979c4e..253f83de1 100644
--- a/tensorflow/lite/micro/micro_mutable_op_resolver.h
+++ b/tensorflow/lite/micro/micro_mutable_op_resolver.h
@@ -432,8 +432,9 @@ class MicroMutableOpResolver : public MicroOpResolver {
ParseMirrorPad);
}
- TfLiteStatus AddMean() {
- return AddBuiltin(BuiltinOperator_MEAN, Register_MEAN(), ParseReducer);
+ TfLiteStatus AddMean(
+ const TFLMRegistration& registration = Register_MEAN()) {
+ return AddBuiltin(BuiltinOperator_MEAN, registration, ParseReducer);
}
TfLiteStatus AddMinimum() {
--
2.25.1

1
mlearning/tflite-micro/Makefile
View file

@ -30,6 +30,7 @@ tflite-micro.zip:
$(Q) patch -d $(TFLITE_MICRO_UNPACK) -p1 < tflite-micro.patch
$(Q) patch -d $(TFLITE_MICRO_UNPACK) -p1 < 0001-dequantize-int8.patch
$(Q) patch -d $(TFLITE_MICRO_UNPACK) -p1 < 0002-quantize-int8.patch
$(Q) patch -d $(TFLITE_MICRO_UNPACK) -p1 < 0003-mean-int8.patch
# Download and unpack tarball if no git repo found
ifeq ($(wildcard tflite-micro/.git),)