[FFmpeg-devel] [PATCH 2/2] avfilter/dnn: unify the layer execution function in native mode

Guo, Yejun yejun.guo at intel.com
Tue Sep 24 03:56:02 EEST 2019


with this change, the code will be simpler when more layers supported.

Signed-off-by: Guo, Yejun <yejun.guo at intel.com>
---
 libavfilter/dnn/dnn_backend_native.c               | 38 ++++++++--------------
 libavfilter/dnn/dnn_backend_native_layer_conv2d.c  |  4 ++-
 libavfilter/dnn/dnn_backend_native_layer_conv2d.h  |  3 +-
 .../dnn/dnn_backend_native_layer_depth2space.c     |  5 ++-
 .../dnn/dnn_backend_native_layer_depth2space.h     |  3 +-
 libavfilter/dnn/dnn_backend_native_layer_maximum.c |  4 ++-
 libavfilter/dnn/dnn_backend_native_layer_maximum.h |  3 +-
 libavfilter/dnn/dnn_backend_native_layer_pad.c     |  5 +--
 libavfilter/dnn/dnn_backend_native_layer_pad.h     |  4 +--
 libavfilter/dnn/dnn_backend_native_layers.hxx      |  4 +++
 tests/dnn/dnn-layer-conv2d-test.c                  |  4 +--
 tests/dnn/dnn-layer-depth2space-test.c             |  4 ++-
 tests/dnn/dnn-layer-maximum-test.c                 |  2 +-
 tests/dnn/dnn-layer-pad-test.c                     |  6 ++--
 14 files changed, 48 insertions(+), 41 deletions(-)
 create mode 100644 libavfilter/dnn/dnn_backend_native_layers.hxx

diff --git a/libavfilter/dnn/dnn_backend_native.c b/libavfilter/dnn/dnn_backend_native.c
index 97549d3..8d04125 100644
--- a/libavfilter/dnn/dnn_backend_native.c
+++ b/libavfilter/dnn/dnn_backend_native.c
@@ -331,10 +331,6 @@ 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)
@@ -344,28 +340,22 @@ DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *output
 
     for (layer = 0; layer < network->layers_num; ++layer){
         switch (network->layers[layer].type){
-        case DLT_CONV2D:
-            conv_params = (ConvolutionalParams *)network->layers[layer].params;
-            convolve(network->operands, network->layers[layer].input_operand_indexes,
-                     network->layers[layer].output_operand_index, conv_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;
+
+#define SETUP_LAYER(layer)                                                          \
+        case layer:                                                                 \
+            dnn_execute_layer_##layer(network->operands,                            \
+                                     network->layers[layer].input_operand_indexes,  \
+                                     network->layers[layer].output_operand_index,   \
+                                     network->layers[layer].params);                \
+        break;
+#include "dnn_backend_native_layers.hxx"
+#undef SETUP_LAYER
+
         case DLT_INPUT:
             return DNN_ERROR;
+        default:
+            av_assert0(!"should not reach here");
+            break;
         }
     }
 
diff --git a/libavfilter/dnn/dnn_backend_native_layer_conv2d.c b/libavfilter/dnn/dnn_backend_native_layer_conv2d.c
index b13b431..d0b34b8 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_conv2d.c
+++ b/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_DLT_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;
diff --git a/libavfilter/dnn/dnn_backend_native_layer_conv2d.h b/libavfilter/dnn/dnn_backend_native_layer_conv2d.h
index 7ddfff3..84cb9b7 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_conv2d.h
+++ b/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_DLT_CONV2D(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                 int32_t output_operand_index, const void *parameters);
 #endif
diff --git a/libavfilter/dnn/dnn_backend_native_layer_depth2space.c b/libavfilter/dnn/dnn_backend_native_layer_depth2space.c
index a248764..9ac261b 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_depth2space.c
+++ b/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_DLT_DEPTH_TO_SPACE(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];
diff --git a/libavfilter/dnn/dnn_backend_native_layer_depth2space.h b/libavfilter/dnn/dnn_backend_native_layer_depth2space.h
index 8708be8..56eaf74 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_depth2space.h
+++ b/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_DLT_DEPTH_TO_SPACE(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                         int32_t output_operand_index, const void *parameters);
 
 #endif
diff --git a/libavfilter/dnn/dnn_backend_native_layer_maximum.c b/libavfilter/dnn/dnn_backend_native_layer_maximum.c
index a2669af..42d0439 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_maximum.c
+++ b/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_DLT_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;
diff --git a/libavfilter/dnn/dnn_backend_native_layer_maximum.h b/libavfilter/dnn/dnn_backend_native_layer_maximum.h
index 6396e58..2f087b2 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_maximum.h
+++ b/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_DLT_MAXIMUM(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                  int32_t output_operand_index, const void *parameters);
 
 #endif
diff --git a/libavfilter/dnn/dnn_backend_native_layer_pad.c b/libavfilter/dnn/dnn_backend_native_layer_pad.c
index c2905a7..99cf5f4 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_pad.c
+++ b/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,
-                           const LayerPadParams *params)
+int dnn_execute_layer_DLT_MIRROR_PAD(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                     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];
diff --git a/libavfilter/dnn/dnn_backend_native_layer_pad.h b/libavfilter/dnn/dnn_backend_native_layer_pad.h
index 7cc8213..bf9cf8d 100644
--- a/libavfilter/dnn/dnn_backend_native_layer_pad.h
+++ b/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,
-                           const LayerPadParams *params);
+int dnn_execute_layer_DLT_MIRROR_PAD(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                     int32_t output_operand_index, const void *parameters);
 
 #endif
diff --git a/libavfilter/dnn/dnn_backend_native_layers.hxx b/libavfilter/dnn/dnn_backend_native_layers.hxx
new file mode 100644
index 0000000..87f7ba9
--- /dev/null
+++ b/libavfilter/dnn/dnn_backend_native_layers.hxx
@@ -0,0 +1,4 @@
+SETUP_LAYER(DLT_CONV2D)
+SETUP_LAYER(DLT_DEPTH_TO_SPACE)
+SETUP_LAYER(DLT_MAXIMUM)
+SETUP_LAYER(DLT_MIRROR_PAD)
diff --git a/tests/dnn/dnn-layer-conv2d-test.c b/tests/dnn/dnn-layer-conv2d-test.c
index afc5391..8849fa6 100644
--- a/tests/dnn/dnn-layer-conv2d-test.c
+++ b/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_DLT_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_DLT_CONV2D(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {
diff --git a/tests/dnn/dnn-layer-depth2space-test.c b/tests/dnn/dnn-layer-depth2space-test.c
index 87118de..3bb1612 100644
--- a/tests/dnn/dnn-layer-depth2space-test.c
+++ b/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_DLT_DEPTH_TO_SPACE(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {
diff --git a/tests/dnn/dnn-layer-maximum-test.c b/tests/dnn/dnn-layer-maximum-test.c
index 06daf64..27b7be6 100644
--- a/tests/dnn/dnn-layer-maximum-test.c
+++ b/tests/dnn/dnn-layer-maximum-test.c
@@ -45,7 +45,7 @@ static int test(void)
     operands[1].data = NULL;
 
     input_indexes[0] = 0;
-    dnn_execute_layer_maximum(operands, input_indexes, 1, &params);
+    dnn_execute_layer_DLT_MAXIMUM(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(input) / sizeof(float); i++) {
diff --git a/tests/dnn/dnn-layer-pad-test.c b/tests/dnn/dnn-layer-pad-test.c
index 1fb2be1..7271ceb 100644
--- a/tests/dnn/dnn-layer-pad-test.c
+++ b/tests/dnn/dnn-layer-pad-test.c
@@ -79,7 +79,7 @@ static int test_with_mode_symmetric(void)
     operands[1].data = NULL;
 
     input_indexes[0] = 0;
-    dnn_execute_layer_pad(operands, input_indexes, 1, &params);
+    dnn_execute_layer_DLT_MIRROR_PAD(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {
@@ -144,7 +144,7 @@ static int test_with_mode_reflect(void)
     operands[1].data = NULL;
 
     input_indexes[0] = 0;
-    dnn_execute_layer_pad(operands, input_indexes, 1, &params);
+    dnn_execute_layer_DLT_MIRROR_PAD(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {
@@ -210,7 +210,7 @@ static int test_with_mode_constant(void)
     operands[1].data = NULL;
 
     input_indexes[0] = 0;
-    dnn_execute_layer_pad(operands, input_indexes, 1, &params);
+    dnn_execute_layer_DLT_MIRROR_PAD(operands, input_indexes, 1, &params);
 
     output = operands[1].data;
     for (int i = 0; i < sizeof(expected_output) / sizeof(float); i++) {
-- 
2.7.4



More information about the ffmpeg-devel mailing list