diff --git a/src/fpga/api.cpp b/src/fpga/api.cpp
index d6994dc7443150ea2c8e4d05499a3b2fac3db579..47acd275fa644f7c6d51c34a547c814531fd88c5 100644
--- a/src/fpga/api.cpp
+++ b/src/fpga/api.cpp
@@ -67,6 +67,20 @@ void fpga_copy(void *dest, const void *src, size_t num) {
   memcpy(dest, src, num);
 }
 
+int fpga_flush(void *address, size_t size) {
+  struct MemoryCacheArgs args;
+  args.address = address;
+  args.size = size;
+  return do_ioctl(IOCTL_MEMCACHE_FLUSH, &args);
+}
+
+int fpga_invalidate(void *address, size_t size) {
+  struct MemoryCacheArgs args;
+  args.address = address;
+  args.size = size;
+  return do_ioctl(IOCTL_MEMCACHE_INVAL, &args);
+}
+
 int ComputeFpgaConv(const struct WrapperConvArgs &args) {
 #ifdef FPGA_TEST_MODE
 /*DLOG << "   relu_enabled:" << args.relu_enabled
diff --git a/src/fpga/api.h b/src/fpga/api.h
index ea50c54b7f6da746aea24f13c66324aea1d7f5b3..9d17e05d6cbfeeb8abac1e06c731510fed2ee65d 100644
--- a/src/fpga/api.h
+++ b/src/fpga/api.h
@@ -139,6 +139,11 @@ struct FpgaRegReadArgs {
   uint64_t value;
 };
 
+struct MemoryCacheArgs {
+  void* address;
+  size_t size;
+};
+
 #define IOCTL_FPGA_MAGIC 'FPGA'
 
 #define IOCTL_VERSION _IOW(IOCTL_FPGA_MAGIC, 01, struct VersionArgs)
@@ -146,6 +151,8 @@ struct FpgaRegReadArgs {
 #define IOCTL_SEPARATOR_0 10
 
 #define IOCTL_MEM_COPY _IOW(IOCTL_FPGA_MAGIC, 11, struct MemoryCopyArgs)
+#define IOCTL_MEMCACHE_INVAL _IOW(IOCTL_FPGA_MAGIC, 12, struct MemoryCacheArgs)
+#define IOCTL_MEMCACHE_FLUSH _IOW(IOCTL_FPGA_MAGIC, 13, struct MemoryCacheArgs)
 
 #define IOCTL_SEPARATOR_1 20
 
@@ -188,6 +195,8 @@ int close_device();
 void* fpga_malloc(size_t size);
 void fpga_free(void* ptr);
 void fpga_copy(void* dst, const void* src, size_t num);
+int fpga_flush(void* address, size_t size);
+int fpga_invalidate(void* address, size_t size);
 
 int PerformBypass(const struct BypassArgs& args);
 int ComputeFpgaConv(const struct WrapperConvArgs& args);
diff --git a/src/fpga/bias_scale.cpp b/src/fpga/bias_scale.cpp
index a1b0c8577b9100f69f823a39e9e136c46b7e09ff..3e5c3419a0c35b5c7c81b0ee1fd89a58838b5a26 100644
--- a/src/fpga/bias_scale.cpp
+++ b/src/fpga/bias_scale.cpp
@@ -79,6 +79,7 @@ void format_bias_scale_array(float **bias_scale_array,
   int element_num_after_division =
       align_to_x(element_num_per_division, BS_NUM_ALIGNMENT);
   interleave(bias_scale_array, div_num * element_num_after_division);
+  fpga_flush(*bias_scale_array, 2 * element_num_after_division * sizeof(float));
 }
 
 }  // namespace bias_scale
diff --git a/src/fpga/filter.cpp b/src/fpga/filter.cpp
index f3672182959ff21bff5c4cb264bc53ab6de53738..3b09ede10d10f605e69d06df2e148dd463e94d5b 100644
--- a/src/fpga/filter.cpp
+++ b/src/fpga/filter.cpp
@@ -206,6 +206,8 @@ void format_filter(float **data_in, int num, int channel, int height, int width,
   align_num(quantize_data, num_per_div_before_alignment, num, chw);
   reorder(quantize_data, num_after_alignment, chw);
   interleave(quantize_data, num_after_alignment, chw);
+  fpga_flush(*quantize_data, align_to_x(chw, FILTER_ELEMENT_ALIGNMENT) *
+                                 num_after_alignment * sizeof(char));
 }
 
 }  // namespace filter
diff --git a/src/fpga/image.cpp b/src/fpga/image.cpp
index 872abcd7c2dd6b16ab8ec8077e9afa6ec60c10d4..0603d164dfa88eb5620ebf588c610ea25a78be5f 100644
--- a/src/fpga/image.cpp
+++ b/src/fpga/image.cpp
@@ -38,7 +38,6 @@ void convert_to_hwc(float **data_in, int channel, int height, int width) {
 }
 
 void align_element_conv(float **data_in, int height, int cw) {
-  int i = 0;
   int h = 0;
   int align_cw = align_to_x(cw, IMAGE_ALIGNMENT);
   if (align_cw != cw) {
@@ -60,6 +59,8 @@ void align_element_conv(float **data_in, int height, int cw) {
 void format_image(float **data_in, int channel, int height, int width) {
   convert_to_hwc(data_in, channel, height, width);
   align_element_conv(data_in, height, channel * width);
+  fpga_flush(*data_in, align_to_x(channel * width, IMAGE_ALIGNMENT) * height *
+                           sizeof(float));
 }
 
 void concat_images(int16_t **images_in, float **scales_in, void *image_out,
@@ -77,6 +78,10 @@ void concat_images(int16_t **images_in, float **scales_in, void *image_out,
   for (i = 0; i < image_num; i++) {
     each_out_line_channel += channel_num[i];
     *scale_out = std::max(*scale_out, scales_in[i][0]);
+    fpga_invalidate(images_in[i],
+                    height *
+                        align_to_x(channel_num[i] * width, IMAGE_ALIGNMENT) *
+                        sizeof(int16_t));
   }
   align_each_out_area_cw =
       align_to_x(each_out_line_channel * width, IMAGE_ALIGNMENT);
@@ -97,6 +102,8 @@ void concat_images(int16_t **images_in, float **scales_in, void *image_out,
       }
     }
   }
+
+  fpga_flush(image_out, height * align_each_out_area_cw * sizeof(int16_t));
 }
 
 }  // namespace image