2020-03-20 12:55:38 +00:00
|
|
|
/*
|
|
|
|
* 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"
|
|
|
|
|
2020-08-23 15:12:12 +00:00
|
|
|
typedef float (*FunType)(float src0, float src1);
|
|
|
|
FunType pfun;
|
|
|
|
|
|
|
|
static float sub(float src0, float src1)
|
|
|
|
{
|
|
|
|
return src0 - src1;
|
|
|
|
}
|
|
|
|
static float add(float src0, float src1)
|
|
|
|
{
|
|
|
|
return src0 + src1;
|
|
|
|
}
|
|
|
|
static float mul(float src0, float src1)
|
|
|
|
{
|
|
|
|
return src0 * src1;
|
|
|
|
}
|
|
|
|
static float realdiv(float src0, float src1)
|
|
|
|
{
|
|
|
|
return src0 / src1;
|
|
|
|
}
|
|
|
|
static float minimum(float src0, float src1)
|
|
|
|
{
|
|
|
|
return FFMIN(src0, src1);
|
|
|
|
}
|
2020-08-23 15:12:13 +00:00
|
|
|
static float floormod(float src0, float src1)
|
|
|
|
{
|
|
|
|
return (float)((int)(src0) % (int)(src1));
|
|
|
|
}
|
2020-08-23 15:12:12 +00:00
|
|
|
|
|
|
|
static void math_binary_commutative(FunType pfun, const DnnLayerMathBinaryParams *params, const DnnOperand *input, DnnOperand *output, DnnOperand *operands, const int32_t *input_operand_indexes)
|
|
|
|
{
|
|
|
|
int dims_count;
|
|
|
|
const float *src;
|
|
|
|
float *dst;
|
|
|
|
dims_count = calculate_operand_dims_count(output);
|
|
|
|
src = input->data;
|
|
|
|
dst = output->data;
|
|
|
|
if (params->input0_broadcast || params->input1_broadcast) {
|
|
|
|
for (int i = 0; i < dims_count; ++i) {
|
|
|
|
dst[i] = pfun(params->v, src[i]);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
const DnnOperand *input1 = &operands[input_operand_indexes[1]];
|
|
|
|
const float *src1 = input1->data;
|
|
|
|
for (int i = 0; i < dims_count; ++i) {
|
|
|
|
dst[i] = pfun(src[i], src1[i]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
static void math_binary_not_commutative(FunType pfun, const DnnLayerMathBinaryParams *params, const DnnOperand *input, DnnOperand *output, DnnOperand *operands, const int32_t *input_operand_indexes)
|
|
|
|
{
|
|
|
|
int dims_count;
|
|
|
|
const float *src;
|
|
|
|
float *dst;
|
|
|
|
dims_count = calculate_operand_dims_count(output);
|
|
|
|
src = input->data;
|
|
|
|
dst = output->data;
|
|
|
|
if (params->input0_broadcast) {
|
|
|
|
for (int i = 0; i < dims_count; ++i) {
|
|
|
|
dst[i] = pfun(params->v, src[i]);
|
|
|
|
}
|
|
|
|
} else if (params->input1_broadcast) {
|
|
|
|
for (int i = 0; i < dims_count; ++i) {
|
|
|
|
dst[i] = pfun(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] = pfun(src[i], src1[i]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2020-06-10 05:36:11 +00:00
|
|
|
int dnn_load_layer_math_binary(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
|
2020-03-20 12:55:38 +00:00
|
|
|
{
|
|
|
|
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);
|
2020-06-10 05:36:11 +00:00
|
|
|
if (layer->input_operand_indexes[input_index] >= operands_num) {
|
|
|
|
return 0;
|
|
|
|
}
|
2020-03-20 12:55:38 +00:00
|
|
|
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);
|
2020-06-10 05:36:11 +00:00
|
|
|
if (layer->input_operand_indexes[input_index] >= operands_num) {
|
|
|
|
return 0;
|
|
|
|
}
|
2020-03-20 12:55:38 +00:00
|
|
|
input_index++;
|
|
|
|
}
|
|
|
|
dnn_size += 4;
|
|
|
|
|
|
|
|
layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
|
|
|
|
dnn_size += 4;
|
|
|
|
layer->params = params;
|
|
|
|
|
2020-06-10 05:36:11 +00:00
|
|
|
if (layer->output_operand_index >= operands_num) {
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2020-03-20 12:55:38 +00:00
|
|
|
return dnn_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
int dnn_execute_layer_math_binary(DnnOperand *operands, const int32_t *input_operand_indexes,
|
2020-08-25 03:47:50 +00:00
|
|
|
int32_t output_operand_index, const void *parameters, NativeContext *ctx)
|
2020-03-20 12:55:38 +00:00
|
|
|
{
|
|
|
|
const DnnOperand *input = &operands[input_operand_indexes[0]];
|
|
|
|
DnnOperand *output = &operands[output_operand_index];
|
|
|
|
const DnnLayerMathBinaryParams *params = (const DnnLayerMathBinaryParams *)parameters;
|
|
|
|
|
|
|
|
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);
|
2020-08-25 03:47:50 +00:00
|
|
|
if (output->length <= 0) {
|
|
|
|
av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
|
2020-07-06 07:32:17 +00:00
|
|
|
return DNN_ERROR;
|
2020-08-25 03:47:50 +00:00
|
|
|
}
|
2020-03-20 12:55:38 +00:00
|
|
|
output->data = av_realloc(output->data, output->length);
|
2020-08-25 03:47:50 +00:00
|
|
|
if (!output->data) {
|
|
|
|
av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
|
2020-03-20 12:55:38 +00:00
|
|
|
return DNN_ERROR;
|
2020-08-25 03:47:50 +00:00
|
|
|
}
|
2020-03-20 12:55:38 +00:00
|
|
|
|
|
|
|
switch (params->bin_op) {
|
|
|
|
case DMBO_SUB:
|
2020-08-23 15:12:12 +00:00
|
|
|
math_binary_not_commutative(sub, params, input, output, operands, input_operand_indexes);
|
2020-03-20 12:55:38 +00:00
|
|
|
return 0;
|
dnn/native: add native support for 'add'
It can be tested with the model file generated with below python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
z1 = 0.039 + x
z2 = x + 0.042
z3 = z1 + z2
z4 = z3 - 0.381
z5 = z4 - x
y = tf.math.maximum(z5, 0.0, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-10 13:35:11 +00:00
|
|
|
case DMBO_ADD:
|
2020-08-23 15:12:12 +00:00
|
|
|
math_binary_commutative(add, params, input, output, operands, input_operand_indexes);
|
dnn/native: add native support for 'add'
It can be tested with the model file generated with below python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
z1 = 0.039 + x
z2 = x + 0.042
z3 = z1 + z2
z4 = z3 - 0.381
z5 = z4 - x
y = tf.math.maximum(z5, 0.0, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-10 13:35:11 +00:00
|
|
|
return 0;
|
dnn/native: add native support for 'mul'
it can be tested with model file generated from above python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
z1 = 0.5 + 0.3 * x
z2 = z1 * 4
z3 = z2 - x - 2.0
y = tf.identity(z3, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-11 05:22:24 +00:00
|
|
|
case DMBO_MUL:
|
2020-08-23 15:12:12 +00:00
|
|
|
math_binary_commutative(mul, params, input, output, operands, input_operand_indexes);
|
dnn/native: add native support for 'mul'
it can be tested with model file generated from above python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
z1 = 0.5 + 0.3 * x
z2 = z1 * 4
z3 = z2 - x - 2.0
y = tf.identity(z3, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-11 05:22:24 +00:00
|
|
|
return 0;
|
dnn/native: add native support for divide
it can be tested with model file generated with below python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
z1 = 2 / x
z2 = 1 / z1
z3 = z2 / 0.25 + 0.3
z4 = z3 - x * 1.5 - 0.3
y = tf.identity(z4, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-11 05:46:47 +00:00
|
|
|
case DMBO_REALDIV:
|
2020-08-23 15:12:12 +00:00
|
|
|
math_binary_not_commutative(realdiv, params, input, output, operands, input_operand_indexes);
|
dnn/native: add native support for divide
it can be tested with model file generated with below python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
z1 = 2 / x
z2 = 1 / z1
z3 = z2 / 0.25 + 0.3
z4 = z3 - x * 1.5 - 0.3
y = tf.identity(z4, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-11 05:46:47 +00:00
|
|
|
return 0;
|
dnn/native: add native support for minimum
it can be tested with model file generated with below python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
x1 = tf.minimum(0.7, x)
x2 = tf.maximum(x1, 0.4)
y = tf.identity(x2, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-26 07:46:38 +00:00
|
|
|
case DMBO_MINIMUM:
|
2020-08-23 15:12:12 +00:00
|
|
|
math_binary_commutative(minimum, params, input, output, operands, input_operand_indexes);
|
dnn/native: add native support for minimum
it can be tested with model file generated with below python script:
import tensorflow as tf
import numpy as np
import imageio
in_img = imageio.imread('input.jpg')
in_img = in_img.astype(np.float32)/255.0
in_data = in_img[np.newaxis, :]
x = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='dnn_in')
x1 = tf.minimum(0.7, x)
x2 = tf.maximum(x1, 0.4)
y = tf.identity(x2, name='dnn_out')
sess=tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ['dnn_out'])
tf.train.write_graph(graph_def, '.', 'image_process.pb', as_text=False)
print("image_process.pb generated, please use \
path_to_ffmpeg/tools/python/convert.py to generate image_process.model\n")
output = sess.run(y, feed_dict={x: in_data})
imageio.imsave("out.jpg", np.squeeze(output))
Signed-off-by: Guo, Yejun <yejun.guo@intel.com>
2020-04-26 07:46:38 +00:00
|
|
|
return 0;
|
2020-08-23 15:12:13 +00:00
|
|
|
case DMBO_FLOORMOD:
|
|
|
|
math_binary_not_commutative(floormod, params, input, output, operands, input_operand_indexes);
|
|
|
|
return 0;
|
2020-03-20 12:55:38 +00:00
|
|
|
default:
|
2020-08-25 03:47:50 +00:00
|
|
|
av_log(ctx, AV_LOG_ERROR, "Unmatch math binary operator\n");
|
2020-08-25 03:47:49 +00:00
|
|
|
return DNN_ERROR;
|
2020-03-20 12:55:38 +00:00
|
|
|
}
|
|
|
|
}
|