dnn_backend_native_layer_mathbinary: add sub support

more math binary operations will be added here

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

libavfilter/dnn/Makefile

@@ -5,6 +5,7 @@ OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_pad
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_conv2d.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_depth2space.o
 OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_maximum.o
+OBJS-$(CONFIG_DNN)                           += dnn/dnn_backend_native_layer_mathbinary.o
 
 DNN-OBJS-$(CONFIG_LIBTENSORFLOW)             += dnn/dnn_backend_tf.o
 

libavfilter/dnn/dnn_backend_native.h

@@ -41,6 +41,7 @@ typedef enum {
     DLT_DEPTH_TO_SPACE = 2,
     DLT_MIRROR_PAD = 3,
     DLT_MAXIMUM = 4,
+    DLT_MATH_BINARY = 5,
     DLT_COUNT
 } DNNLayerType;
 

libavfilter/dnn/dnn_backend_native_layer_mathbinary.c

@@ -0,0 +1,113 @@
+/*
+ * Copyright (c) 2020
+ *
+ * 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
+ */
+
+/**
+ * @file
+ * DNN native backend implementation.
+ */
+
+#include "dnn_backend_native.h"
+#include "libavutil/avassert.h"
+#include "dnn_backend_native_layer_mathbinary.h"
+
+int dnn_load_layer_math_binary(Layer *layer, AVIOContext *model_file_context, int file_size)
+{
+    DnnLayerMathBinaryParams *params;
+    int dnn_size = 0;
+    int input_index = 0;
+    params = av_malloc(sizeof(*params));
+    if (!params)
+        return 0;
+
+    params->bin_op = (int32_t)avio_rl32(model_file_context);
+    dnn_size += 4;
+
+    params->input0_broadcast = (int32_t)avio_rl32(model_file_context);
+    dnn_size += 4;
+    if (params->input0_broadcast) {
+        params->v = av_int2float(avio_rl32(model_file_context));
+    } else {
+        layer->input_operand_indexes[input_index] = (int32_t)avio_rl32(model_file_context);
+        input_index++;
+    }
+    dnn_size += 4;
+
+    params->input1_broadcast = (int32_t)avio_rl32(model_file_context);
+    dnn_size += 4;
+    if (params->input1_broadcast) {
+        params->v = av_int2float(avio_rl32(model_file_context));
+    } else {
+        layer->input_operand_indexes[input_index] = (int32_t)avio_rl32(model_file_context);
+        input_index++;
+    }
+    dnn_size += 4;
+
+    layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
+    dnn_size += 4;
+    layer->params = params;
+
+    return dnn_size;
+}
+
+int dnn_execute_layer_math_binary(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 DnnLayerMathBinaryParams *params = (const DnnLayerMathBinaryParams *)parameters;
+    int dims_count;
+    const float *src;
+    float *dst;
+
+    for (int i = 0; i < 4; ++i)
+        output->dims[i] = input->dims[i];
+
+    output->data_type = input->data_type;
+    output->length = calculate_operand_data_length(output);
+    output->data = av_realloc(output->data, output->length);
+    if (!output->data)
+        return DNN_ERROR;
+
+    dims_count = calculate_operand_dims_count(output);
+    src = input->data;
+    dst = output->data;
+
+    switch (params->bin_op) {
+    case DMBO_SUB:
+        if (params->input0_broadcast) {
+            for (int i = 0; i < dims_count; ++i) {
+                dst[i] = params->v - src[i];
+            }
+        } else if (params->input1_broadcast) {
+            for (int i = 0; i < dims_count; ++i) {
+                dst[i] = src[i] - params->v;
+            }
+        } else {
+            const DnnOperand *input1 = &operands[input_operand_indexes[1]];
+            const float *src1 = input1->data;
+            for (int i = 0; i < dims_count; ++i) {
+                dst[i] = src[i] - src1[i];
+            }
+        }
+        return 0;
+    default:
+        return -1;
+    }
+}

libavfilter/dnn/dnn_backend_native_layer_mathbinary.h

@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2020
+ *
+ * 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
+ */
+
+/**
+ * @file
+ * DNN inference functions interface for native backend.
+ */
+
+
+#ifndef AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_MATHBINARY_H
+#define AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_MATHBINARY_H
+
+#include "libavformat/avio.h"
+#include "dnn_backend_native.h"
+
+typedef enum {
+    DMBO_SUB = 0,
+    DMBO_COUNT
+} DNNMathBinaryOperation;
+
+typedef struct DnnLayerMathBinaryParams{
+    DNNMathBinaryOperation bin_op;
+    int input0_broadcast;
+    int input1_broadcast;
+    float v;
+} DnnLayerMathBinaryParams;
+
+int dnn_load_layer_math_binary(Layer *layer, AVIOContext *model_file_context, int file_size);
+int dnn_execute_layer_math_binary(DnnOperand *operands, const int32_t *input_operand_indexes,
+                                 int32_t output_operand_index, const void *parameters);
+
+#endif

libavfilter/dnn/dnn_backend_native_layers.c

@@ -24,6 +24,7 @@
 #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_layer_mathbinary.h"
 
 LayerFunc layer_funcs[DLT_COUNT] = {
     {NULL, NULL},
@@ -31,4 +32,5 @@ LayerFunc layer_funcs[DLT_COUNT] = {
     {dnn_execute_layer_depth2space, dnn_load_layer_depth2space},
     {dnn_execute_layer_pad,         dnn_load_layer_pad},
     {dnn_execute_layer_maximum,     dnn_load_layer_maximum},
+    {dnn_execute_layer_math_binary, dnn_load_layer_math_binary},
 };

tools/python/convert_from_tensorflow.py

@@ -70,7 +70,8 @@ class TFConverter:
         self.converted_nodes = set()
         self.conv2d_scope_names = set()
         self.conv2d_scopename_inputname_dict = {}
-        self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4}
+        self.op2code = {'Conv2D':1, 'DepthToSpace':2, 'MirrorPad':3, 'Maximum':4, 'MathBinary':5}
+        self.mathbin2code = {'Sub':0}
         self.mirrorpad_mode = {'CONSTANT':0, 'REFLECT':1, 'SYMMETRIC':2}
         self.name_operand_dict = {}
 
@@ -113,6 +114,8 @@ class TFConverter:
         # if activation is None, and BiasAdd.next is the last op which is Identity
         if conv2d_scope_name + '/BiasAdd' in self.edges:
             anode = self.edges[conv2d_scope_name + '/BiasAdd'][0]
+            if anode.op not in self.conv_activations:
+                anode = None
         else:
             anode = None
         return knode, bnode, dnode, anode
@@ -252,14 +255,47 @@ class TFConverter:
         np.array([input_operand_index, output_operand_index], dtype=np.uint32).tofile(f)
 
 
+    def dump_sub_to_file(self, node, f):
+        assert(node.op == 'Sub')
+        self.layer_number = self.layer_number + 1
+        self.converted_nodes.add(node.name)
+        i0_node = self.name_node_dict[node.input[0]]
+        i1_node = self.name_node_dict[node.input[1]]
+        np.array([self.op2code['MathBinary'], self.mathbin2code[node.op]], dtype=np.uint32).tofile(f)
+        if i0_node.op == 'Const':
+            scalar = i0_node.attr['value'].tensor.float_val[0]
+            assert(i0_node.name.find('sub/x'))
+            np.array([1], dtype=np.uint32).tofile(f)
+            np.array([scalar], dtype=np.float32).tofile(f)
+            np.array([0], dtype=np.uint32).tofile(f)
+            input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
+            np.array([input_operand_index], dtype=np.uint32).tofile(f)
+        elif i1_node.op == 'Const':
+            scalar = i1_node.attr['value'].tensor.float_val[0]
+            assert(i1_node.name.find('sub/y'))
+            np.array([0], dtype=np.uint32).tofile(f)
+            input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
+            np.array([input_operand_index], dtype=np.uint32).tofile(f)
+            np.array([1], dtype=np.uint32).tofile(f)
+            np.array([scalar], dtype=np.float32).tofile(f)
+        else:
+            np.array([0], dtype=np.uint32).tofile(f)
+            input_operand_index = self.add_operand(i0_node.name, Operand.IOTYPE_INPUT)
+            np.array([input_operand_index], dtype=np.uint32).tofile(f)
+            np.array([0], dtype=np.uint32).tofile(f)
+            input_operand_index = self.add_operand(i1_node.name, Operand.IOTYPE_INPUT)
+            np.array([input_operand_index], dtype=np.uint32).tofile(f)
+        output_operand_index = self.add_operand(node.name, Operand.IOTYPE_OUTPUT)
+        np.array([output_operand_index], dtype=np.uint32).tofile(f)
+
+
     def dump_layers_to_file(self, f):
         for node in self.nodes:
             if node.name in self.converted_nodes:
                 continue
 
             # conv2d with dilation generates very complex nodes, so handle it in special
-            scope_name = TFConverter.get_scope_name(node.name)
-            if scope_name in self.conv2d_scope_names:
+            if self.in_conv2d_scope(node.name):
                 if node.op == 'Conv2D':
                     self.dump_complex_conv2d_to_file(node, f)
                 continue
@@ -272,6 +308,8 @@ class TFConverter:
                 self.dump_mirrorpad_to_file(node, f)
             elif node.op == 'Maximum':
                 self.dump_maximum_to_file(node, f)
+            elif node.op == 'Sub':
+                self.dump_sub_to_file(node, f)
 
 
     def dump_operands_to_file(self, f):
@@ -352,6 +390,17 @@ class TFConverter:
         return name[0:index]
 
 
+    def in_conv2d_scope(self, name):
+        inner_scope = TFConverter.get_scope_name(name)
+        if inner_scope == "":
+            return False;
+        for scope in self.conv2d_scope_names:
+            index = inner_scope.find(scope)
+            if index == 0:
+                return True
+        return False
+
+
     def generate_conv2d_scope_info(self):
         # mostly, conv2d is a sub block in graph, get the scope name
         for node in self.nodes:

tools/python/convert_header.py

@@ -23,4 +23,4 @@ str = 'FFMPEGDNNNATIVE'
 major = 1
 
 # increase minor when we don't have to re-convert the model file
-minor = 0
+minor = 1