From 3aa9809f85027dea106140b39933e37e4ba9f34e Mon Sep 17 00:00:00 2001
From: AddSalt8227 <595311669@qq.com>
Date: Mon, 27 Dec 2021 16:55:21 +0800
Subject: [PATCH] fix op bugs with split and batchnorm (#1236)

* fix retinaface_timvx memory leak

* fix rgb error in tm_yolact.cpp

* 1.fix eltwise uint8 bug
2.add hardsigmoid uint8

* fix op bugs with split and batchnorm
---
 .../batchnorm/cortex-a/batchnorm_kernel_arm.c  | 18 +++++++++++++++++-
 source/device/cpu/op/split/split_ref.c         | 16 +++++++++++++++-
 2 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/source/device/cpu/op/batchnorm/cortex-a/batchnorm_kernel_arm.c b/source/device/cpu/op/batchnorm/cortex-a/batchnorm_kernel_arm.c
index 181648a0..65ef1528 100644
--- a/source/device/cpu/op/batchnorm/cortex-a/batchnorm_kernel_arm.c
+++ b/source/device/cpu/op/batchnorm/cortex-a/batchnorm_kernel_arm.c
@@ -64,7 +64,23 @@ int batchnorm_run(struct tensor* output_tensor, struct tensor* input_tensor, flo
 {
     int batch_number = input_tensor->dims[0];
     int channel_num = input_tensor->dims[1];
-    int channel_size = (input_tensor->dims[2]) * (input_tensor->dims[3]);
+    int channel_size;
+    if (4 == input_tensor->dim_num)
+    {
+        channel_size = (input_tensor->dims[2]) * (input_tensor->dims[3]);
+    }
+    else if (3 == input_tensor->dim_num)
+    {
+        channel_size = (input_tensor->dims[2]);
+    }
+    else if (2 == input_tensor->dim_num)
+    {
+        channel_size = 1;
+    }
+    else
+    {
+        return -1;
+    }
     int img_size = channel_num * channel_size;
 
     const float* input = (const float*)input_tensor->data;
diff --git a/source/device/cpu/op/split/split_ref.c b/source/device/cpu/op/split/split_ref.c
index b6db65f0..bb0c2359 100644
--- a/source/device/cpu/op/split/split_ref.c
+++ b/source/device/cpu/op/split/split_ref.c
@@ -69,6 +69,11 @@ int ref_split_uint8(struct tensor* input_tensor, struct tensor* output_tensor, s
 {
     uint8_t* input_data = (uint8_t*)input_tensor->data;
     uint8_t* output_data = (uint8_t*)output_tensor->data;
+    float input_scale = input_tensor->scale;
+    float output_scale = output_tensor->scale;
+    int32_t input_zero = input_tensor->zero_point;
+    int32_t output_zero = output_tensor->zero_point;
+    float rescale = input_scale / output_scale;
 
     if (split_param->is_caffe)
     {
@@ -84,7 +89,15 @@ int ref_split_uint8(struct tensor* input_tensor, struct tensor* output_tensor, s
         {
             int in_offset = (n * in_slice + *slice_index) * slice_size;
             int out_offset = n * out_slice * slice_size;
-            memcpy(output_data + out_offset, input_data + in_offset, (size_t)slice_size * out_slice * sizeof(uint8_t));
+            for (size_t i = 0; i < slice_size * out_slice * sizeof(uint8_t); i++)
+            {
+                int udata = roundf((input_data[in_offset + i] - input_zero) * rescale + output_zero);
+                if (udata > 255)
+                    udata = 255;
+                else if (udata < 0)
+                    udata = 0;
+                output_data[i] = udata;
+            }
         }
 
         *slice_index += out_slice;
@@ -148,6 +161,7 @@ static int run(struct node_ops* node_ops, struct exec_node* exec_node, struct ex
     struct split_param* split_param = (struct split_param*)ir_node->op.param_mem;
 
     /* the follow codes need to be checked ! */
+    /* maybe int8 need dequant and quant */
     int slice_axis = split_param->axis;
     int num_slices = 1;
     int slice_size = 1;
-- 
GitLab