Aksal/nuttx-apps
1
0
Fork 1
mirror of https://github.com/apache/nuttx-apps.git synced 2026-08-09 06:25:16 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

ml: tflm quantize optimization patch

Browse source 
Separate the 'Float32 to Int8' implementation of the 'quantize' operator to reduce the code size.

Signed-off-by: xinhaiteng <xinhaiteng@xiaomi.com>
This commit is contained in:
xinhaiteng 2024-04-25 11:34:26 +08:00 • committed by Xiang Xiao
parent 570102c501
commit 068b2b16bb
2 changed files with 109 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

108
mlearning/tflite-micro/0002-quantize-int8.patch Normal file
View file

@ -0,0 +1,108 @@
From b7c954c7594b9a39df5604dbadf6e59b6f98b1e9 Mon Sep 17 00:00:00 2001
From: xinhaiteng <xinhaiteng@xiaomi.com>
Date: Wed, 24 Apr 2024 09:18:52 +0800
Subject: [PATCH] tflite: add quantize float32 to int8 operator
VELAPLATFO-30990
Separate the 'Float32 to Int8' implementation of the 'quantize' operator to reduce the code size.
Change-Id: I83dc809d36587741f3087b25e83e454b30f9626b
Signed-off-by: xinhaiteng <xinhaiteng@xiaomi.com>
---
tensorflow/lite/micro/kernels/quantize.cc | 5 +++++
tensorflow/lite/micro/kernels/quantize.h | 4 ++++
tensorflow/lite/micro/kernels/quantize_common.cc | 16 ++++++++++++++++
.../lite/micro/micro_mutable_op_resolver.h | 7 ++++---
4 files changed, 29 insertions(+), 3 deletions(-)
diff --git a/tensorflow/lite/micro/kernels/quantize.cc b/tensorflow/lite/micro/kernels/quantize.cc
index ba11f19b6..a6ff4f055 100644
--- a/tensorflow/lite/micro/kernels/quantize.cc
+++ b/tensorflow/lite/micro/kernels/quantize.cc
@@ -39,4 +39,9 @@ TFLMRegistration Register_QUANTIZE() {
InitQuantizeReference, PrepareQuantizeReference, EvalQuantizeReference);
}
+TFLMRegistration Register_QUANTIZE_FLOAT32_INT8() {
+ return tflite::micro::RegisterOp(
+ InitQuantizeReference, PrepareQuantizeReference, EvalQuantizeReferenceFloat32ToInt8);
+}
+
} // namespace tflite
diff --git a/tensorflow/lite/micro/kernels/quantize.h b/tensorflow/lite/micro/kernels/quantize.h
index ba93809a2..b97e4a2ab 100644
--- a/tensorflow/lite/micro/kernels/quantize.h
+++ b/tensorflow/lite/micro/kernels/quantize.h
@@ -17,6 +17,7 @@ limitations under the License.
#include "tensorflow/lite/c/common.h"
#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/micro/micro_common.h"
namespace tflite {
@@ -31,7 +32,10 @@ struct OpDataQuantizeReference {
};
TfLiteStatus EvalQuantizeReference(TfLiteContext* context, TfLiteNode* node);
+TfLiteStatus EvalQuantizeReferenceFloat32ToInt8(TfLiteContext* context, TfLiteNode* node);
TfLiteStatus PrepareQuantizeReference(TfLiteContext* context, TfLiteNode* node);
+
+TFLMRegistration Register_QUANTIZE_FLOAT32_INT8();
} // namespace tflite
#endif // TENSORFLOW_LITE_MICRO_KERNELS_QUANTIZE_H_
diff --git a/tensorflow/lite/micro/kernels/quantize_common.cc b/tensorflow/lite/micro/kernels/quantize_common.cc
index cb04eafce..6bdd0df39 100644
--- a/tensorflow/lite/micro/kernels/quantize_common.cc
+++ b/tensorflow/lite/micro/kernels/quantize_common.cc
@@ -236,4 +236,20 @@ TfLiteStatus EvalQuantizeReference(TfLiteContext* context, TfLiteNode* node) {
return kTfLiteOk;
}
+TfLiteStatus EvalQuantizeReferenceFloat32ToInt8(TfLiteContext* context, TfLiteNode* node) {
+ TFLITE_DCHECK(node->user_data != nullptr);
+ auto* data = static_cast<OpDataQuantizeReference*>(node->user_data);
+
+ const TfLiteEvalTensor* input = tflite::micro::GetEvalInput(context, node, 0);
+ TfLiteEvalTensor* output = tflite::micro::GetEvalOutput(context, node, 0);
+
+ TFLITE_DCHECK(input->type == kTfLiteFloat32 && output->type == kTfLiteInt8);
+ reference_ops::AffineQuantize(
+ data->quantization_params, tflite::micro::GetTensorShape(input),
+ tflite::micro::GetTensorData<float>(input),
+ tflite::micro::GetTensorShape(output),
+ tflite::micro::GetTensorData<int8_t>(output));
+ return kTfLiteOk;
+}
+
} // namespace tflite
diff --git a/tensorflow/lite/micro/micro_mutable_op_resolver.h b/tensorflow/lite/micro/micro_mutable_op_resolver.h
index b3e3cbcf8..ec8979c4e 100644
--- a/tensorflow/lite/micro/micro_mutable_op_resolver.h
+++ b/tensorflow/lite/micro/micro_mutable_op_resolver.h
@@ -32,6 +32,7 @@ limitations under the License.
#include "tensorflow/lite/micro/kernels/micro_ops.h"
#include "tensorflow/lite/micro/kernels/mul.h"
#include "tensorflow/lite/micro/kernels/pooling.h"
+#include "tensorflow/lite/micro/kernels/quantize.h"
#include "tensorflow/lite/micro/kernels/reduce.h"
#include "tensorflow/lite/micro/kernels/softmax.h"
#include "tensorflow/lite/micro/kernels/transpose_conv.h"
@@ -481,9 +482,9 @@ class MicroMutableOpResolver : public MicroOpResolver {
ParsePrelu);
}
- TfLiteStatus AddQuantize() {
- return AddBuiltin(BuiltinOperator_QUANTIZE, Register_QUANTIZE(),
- ParseQuantize);
+ TfLiteStatus AddQuantize(
+ const TFLMRegistration& registration = Register_QUANTIZE()) {
+ return AddBuiltin(BuiltinOperator_QUANTIZE, registration, ParseQuantize);
}
TfLiteStatus AddReadVariable() {
--
2.25.1

1
mlearning/tflite-micro/Makefile
View file

@ -29,6 +29,7 @@ tflite-micro.zip:
$(Q) mv tflite-micro-$(TFLITE_MICRO_VER) tflite-micro
$(Q) patch -d $(TFLITE_MICRO_UNPACK) -p1 < tflite-micro.patch
$(Q) patch -d $(TFLITE_MICRO_UNPACK) -p1 < 0001-dequantize-int8.patch
$(Q) patch -d $(TFLITE_MICRO_UNPACK) -p1 < 0002-quantize-int8.patch
# Download and unpack tarball if no git repo found
ifeq ($(wildcard tflite-micro/.git),)