avfilter/dnn: unify the layer execution function in native mode

Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
Signed-off-by: Pedro Arthur <bygrandao@gmail.com>

libavfilter/dnn/Makefile

@@ -1,5 +1,6 @@
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_interface.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native.o
+OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layers.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_pad.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_conv2d.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_depth2space.o

libavfilter/dnn/dnn_backend_native.c

@@ -29,6 +29,7 @@
 #include "dnn_backend_native_layer_conv2d.h"
 #include "dnn_backend_native_layer_depth2space.h"
 #include "dnn_backend_native_layer_maximum.h"
+#include "dnn_backend_native_layers.h"
 
 static DNNReturnType set_input_output_native(void *model, DNNInputData *input, const char *input_name, const char **output_names, uint32_t nb_output)
 {
@@ -331,10 +332,6 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output
 {
     ConvolutionalNetwork *network = (ConvolutionalNetwork *)model->model;
     int32_t layer;
-    ConvolutionalParams *conv_params;
-    DepthToSpaceParams *depth_to_space_params;
-    LayerPadParams *pad_params;
-    DnnLayerMaximumParams *maximum_params;
     uint32_t nb = FFMIN(nb_output, network->nb_output);
 
     if (network->layers_num <= 0 || network->operands_num <= 0)
@@ -343,30 +340,11 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output
         return DNN_ERROR;
 
     for (layer = 0; layer < network->layers_num; ++layer){
-        switch (network->layers[layer].type){
+        DNNLayerType layer_type = network->layers[layer].type;
-        case DLT_CONV2D:
+        layer_funcs[layer_type](network->operands,
-            conv_params = (ConvolutionalParams *)network->layers[layer].params;
+                                  network->layers[layer].input_operand_indexes,
-            convolve(network->operands, network->layers[layer].input_operand_indexes,
+                                  network->layers[layer].output_operand_index,
-                     network->layers[layer].output_operand_index, conv_params);
+                                  network->layers[layer].params);
-            break;
-        case DLT_DEPTH_TO_SPACE:
-            depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params;
-            depth_to_space(network->operands, network->layers[layer].input_operand_indexes,
-                           network->layers[layer].output_operand_index, depth_to_space_params->block_size);
-            break;
-        case DLT_MIRROR_PAD:
-            pad_params = (LayerPadParams *)network->layers[layer].params;
-            dnn_execute_layer_pad(network->operands, network->layers[layer].input_operand_indexes,
-                                  network->layers[layer].output_operand_index, pad_params);
-            break;
-        case DLT_MAXIMUM:
-            maximum_params = (DnnLayerMaximumParams *)network->layers[layer].params;
-            dnn_execute_layer_maximum(network->operands, network->layers[layer].input_operand_indexes,
-                                  network->layers[layer].output_operand_index, maximum_params);
-            break;
-        case DLT_INPUT:
-            return DNN_ERROR;
-        }
     }
 
     for (uint32_t i = 0; i < nb; ++i) {

libavfilter/dnn/dnn_backend_native.h

@@ -33,13 +33,15 @@
 /**
  * the enum value of DNNLayerType should not be changed,
  * the same values are used in convert_from_tensorflow.py
+ * and, it is used to index the layer execution function pointer.
  */
 typedef enum {
     DLT_INPUT = 0,
     DLT_CONV2D = 1,
     DLT_DEPTH_TO_SPACE = 2,
     DLT_MIRROR_PAD = 3,
-    DLT_MAXIMUM = 4
+    DLT_MAXIMUM = 4,
+    DLT_COUNT
 } DNNLayerType;
 
 typedef enum {DOT_INPUT = 1, DOT_OUTPUT = 2, DOT_INTERMEDIATE = DOT_INPUT | DOT_INPUT} DNNOperandType;

libavfilter/dnn/dnn_backend_native_layer_conv2d.c

@@ -23,7 +23,8 @@
 
 #define CLAMP_TO_EDGE(x, w) ((x) < 0 ? 0 : ((x) >= (w) ? (w - 1) : (x)))
 
-int convolve(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index, const ConvolutionalParams *conv_params)
+int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_indexes,
+                             int32_t output_operand_index, const void *parameters)
 {
     float *output;
     int32_t input_operand_index = input_operand_indexes[0];
@@ -32,6 +33,7 @@ int convolve(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t
     int width = operands[input_operand_index].dims[2];
     int channel = operands[input_operand_index].dims[3];
     const float *input = operands[input_operand_index].data;
+    const ConvolutionalParams *conv_params = (const ConvolutionalParams *)parameters;
 
     int radius = conv_params->kernel_size >> 1;
     int src_linesize = width * conv_params->input_num;

libavfilter/dnn/dnn_backend_native_layer_conv2d.h

@@ -35,5 +35,6 @@ typedef struct ConvolutionalParams{
     float *biases;
 } ConvolutionalParams;
 
-int convolve(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index, const ConvolutionalParams *conv_params);
+int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_indexes,
+                             int32_t output_operand_index, const void *parameters);
 #endif

libavfilter/dnn/dnn_backend_native_layer_depth2space.c

@@ -27,9 +27,12 @@
 #include "libavutil/avassert.h"
 #include "dnn_backend_native_layer_depth2space.h"
 
-int depth_to_space(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index, int block_size)
+int dnn_execute_layer_depth2space(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                  int32_t output_operand_index, const void *parameters)
 {
     float *output;
+    const DepthToSpaceParams *params = (const DepthToSpaceParams *)parameters;
+    int block_size = params->block_size;
     int32_t input_operand_index = input_operand_indexes[0];
     int number = operands[input_operand_index].dims[0];
     int height = operands[input_operand_index].dims[1];

libavfilter/dnn/dnn_backend_native_layer_depth2space.h

@@ -34,6 +34,7 @@ typedef struct DepthToSpaceParams{
     int block_size;
 } DepthToSpaceParams;
 
-int depth_to_space(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index, int block_size);
+int dnn_execute_layer_depth2space(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                  int32_t output_operand_index, const void *parameters);
 
 #endif

libavfilter/dnn/dnn_backend_native_layer_maximum.c

@@ -27,10 +27,12 @@
 #include "libavutil/avassert.h"
 #include "dnn_backend_native_layer_maximum.h"
 
-int dnn_execute_layer_maximum(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index, const DnnLayerMaximumParams *params)
+int dnn_execute_layer_maximum(DnnOperand *operands, const int32_t *input_operand_indexes,
+                              int32_t output_operand_index, const void *parameters)
 {
     const DnnOperand *input = &operands[input_operand_indexes[0]];
     DnnOperand *output = &operands[output_operand_index];
+    const DnnLayerMaximumParams *params = (const DnnLayerMaximumParams *)parameters;
     int dims_count;
     const float *src;
     float *dst;

libavfilter/dnn/dnn_backend_native_layer_maximum.h

@@ -37,6 +37,7 @@ typedef struct DnnLayerMaximumParams{
     }val;
 } DnnLayerMaximumParams;
 
-int dnn_execute_layer_maximum(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index, const DnnLayerMaximumParams *params);
+int dnn_execute_layer_maximum(DnnOperand *operands, const int32_t *input_operand_indexes,
+                              int32_t output_operand_index, const void *parameters);
 
 #endif

libavfilter/dnn/dnn_backend_native_layer_pad.c

@@ -48,12 +48,13 @@ static int after_get_buddy(int given, int border, LayerPadModeParam mode)
     }
 }
 
-int dnn_execute_layer_pad(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index,
+int dnn_execute_layer_pad(DnnOperand *operands, const int32_t *input_operand_indexes,
-                           const LayerPadParams *params)
+                          int32_t output_operand_index, const void *parameters)
 {
     int32_t before_paddings;
     int32_t after_paddings;
     float* output;
+    const LayerPadParams *params = (const LayerPadParams *)parameters;
 
     // suppose format is <N, H, W, C>
     int32_t input_operand_index = input_operand_indexes[0];

libavfilter/dnn/dnn_backend_native_layer_pad.h

@@ -36,7 +36,7 @@ typedef struct LayerPadParams{
     float constant_values;
 } LayerPadParams;
 
-int dnn_execute_layer_pad(DnnOperand *operands, const int32_t *input_operand_indexes, int32_t output_operand_index,
+int dnn_execute_layer_pad(DnnOperand *operands, const int32_t *input_operand_indexes,
-                           const LayerPadParams *params);
+                          int32_t output_operand_index, const void *parameters);
 
 #endif

libavfilter/dnn/dnn_backend_native_layers.c

@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2019 Guo Yejun
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <string.h>
+#include "dnn_backend_native_layers.h"
+#include "dnn_backend_native_layer_pad.h"
+#include "dnn_backend_native_layer_conv2d.h"
+#include "dnn_backend_native_layer_depth2space.h"
+#include "dnn_backend_native_layer_maximum.h"
+
+LAYER_EXEC_FUNC layer_funcs[DLT_COUNT] = {
+    NULL,
+    dnn_execute_layer_conv2d,
+    dnn_execute_layer_depth2space,
+    dnn_execute_layer_pad,
+    dnn_execute_layer_maximum,
+};

libavfilter/dnn/dnn_backend_native_layers.h

@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2019 Guo Yejun
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYERS_H
+#define AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYERS_H
+
+#include <stdint.h>
+#include "dnn_backend_native.h"
+
+typedef int (*LAYER_EXEC_FUNC)(DnnOperand *operands, const int32_t *input_operand_indexes,
+                               int32_t output_operand_index, const void *parameters);
+
+extern LAYER_EXEC_FUNC layer_funcs[DLT_COUNT];
+
+#endif

tests/dnn/dnn-layer-conv2d-test.c

@@ -113,7 +113,7 @@ static int test_with_same_dilate(void)
     operands[1].data = NULL;
 
     input_indexes[0] = 0;
-    convolve(operands, input_indexes, 1, &params);
+    dnn_execute_layer_conv2d(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {
@@ -212,7 +212,7 @@ static int test_with_valid(void)
     operands[1].data = NULL;
 
     input_indexes[0] = 0;
-    convolve(operands, input_indexes, 1, &params);
+    dnn_execute_layer_conv2d(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {

tests/dnn/dnn-layer-depth2space-test.c

@@ -48,6 +48,7 @@ static int test(void)
     print(list(output.flatten()))
     */
 
+    DepthToSpaceParams params;
     DnnOperand operands[2];
     int32_t input_indexes[1];
     float input[1*5*3*4] = {
@@ -79,7 +80,8 @@ static int test(void)
     operands[1].data = NULL;
 
     input_indexes[0] = 0;
-    depth_to_space(operands, input_indexes, 1, 2);
+    params.block_size = 2;
+    dnn_execute_layer_depth2space(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {