Simulation

1    import numpy as np
2    import os
3    
4    def generate_binary_file(filename, nrecords):
5        n_records = np.random.poisson(lam = nrecords)
6        record_lengths = np.random.poisson(lam = 10, size = n_records).astype(dtype = 'i4')
7        x = np.random.normal(size = record_lengths.sum()).astype(dtype = 'd')
8        with open(filename, 'wb') as f:
9            s = 0
10            for i in range(n_records):
11                f.write(record_lengths[i].tobytes())
12                f.write(x[s:s+record_lengths[i]].tobytes())
13                s += record_lengths[i]
14            # Trick for testing: make sum of records equal to 0
15            f.write(np.array([1], dtype = 'i4').tobytes())
16            f.write(np.array([-x.sum()], dtype = 'd').tobytes())
17        return os.path.getsize(filename)
18    
19    def read_binary_npfromfile(filename):
20        checksum = 0.0
21        with open(filename, 'rb') as f:
22            while True:
23                try:
24                    record_length = np.fromfile(f, 'i4', 1)[0]
25                    x = np.fromfile(f, 'd', record_length)
26                    checksum += x.sum()
27                except:
28                    break
29        assert(np.abs(checksum) < 1e-6)
30
31    def read_binary_npfrombuffer(filename):
32        checksum = 0.0
33        with open(filename, 'rb') as f:
34            while True:
35                try:
36                    record_length = np.frombuffer(f.read(np.dtype('i4').itemsize), dtype = 'i4', count = 1)[0]
37                    x = np.frombuffer(f.read(np.dtype('d').itemsize * record_length), dtype = 'd', count = record_length)
38                    checksum += x.sum()
39                except:
40                    break
41        assert(np.abs(checksum) < 1e-6)
42    
43    
44    if __name__ == '__main__':
45        from timeit import Timer
46        from functools import partial
47    
48        fname = 'testfile.tmp'
49        print("# File size[MB], Timings and errors [s]: fromfile, frombuffer")
50        for i in [10**3, 3*10**3, 10**4, 3*10**4, 10**5, 3*10**5, 10**6, 3*10**6]:
51            fsize = generate_binary_file(fname, i)
52            t1 = Timer(partial(read_binary_npfromfile, fname))
53            t2 = Timer(partial(read_binary_npfrombuffer, fname))
54            a1 = np.array(t1.repeat(5, 1))
55            a2 = np.array(t2.repeat(5, 1))
56            print('%8.3f %12.6f %12.6f %12.6f %12.6f' % (1.0 * fsize / (2**20), a1.mean(), a1.std(), a2.mean(), a2.std()))
57def read_binary_faster_numpy(filename):
58    buff = np.fromfile(filename, dtype=np.uint8)
59    buff_int32 = buff.view(np.int32)
60    buff_double_1 = buff[0:len(buff)//8*8].view(np.float64)
61    buff_double_2 = buff[4:4+(len(buff)-4)//8*8].view(np.float64)
62    nblocks = buff.size // 4      # Number of 4-byte blocks
63    pos = 0                       # Displacement by block of 4 bytes
64    lst = []
65    while pos < nblocks:
66        record_length = buff_int32[pos]
67        pos += 1
68        if pos + record_length * 2 > nblocks:
69            break
70        offset = pos // 2
71        if pos % 2 == 0:          # Aligned with buff_double_1
72            x = buff_double_1[offset:offset+record_length]
73        else:                     # Aligned with buff_double_2
74            x = buff_double_2[offset:offset+record_length]
75        lst.append(x)             # np.sum is too expensive here
76        pos += record_length * 2
77    checksum = np.sum(np.concatenate(lst))
78    assert(np.abs(checksum) < 1e-6)
79

1    import numpy as np
2    import os
3    
4    def generate_binary_file(filename, nrecords):
5        n_records = np.random.poisson(lam = nrecords)
6        record_lengths = np.random.poisson(lam = 10, size = n_records).astype(dtype = 'i4')
7        x = np.random.normal(size = record_lengths.sum()).astype(dtype = 'd')
8        with open(filename, 'wb') as f:
9            s = 0
10            for i in range(n_records):
11                f.write(record_lengths[i].tobytes())
12                f.write(x[s:s+record_lengths[i]].tobytes())
13                s += record_lengths[i]
14            # Trick for testing: make sum of records equal to 0
15            f.write(np.array([1], dtype = 'i4').tobytes())
16            f.write(np.array([-x.sum()], dtype = 'd').tobytes())
17        return os.path.getsize(filename)
18    
19    def read_binary_npfromfile(filename):
20        checksum = 0.0
21        with open(filename, 'rb') as f:
22            while True:
23                try:
24                    record_length = np.fromfile(f, 'i4', 1)[0]
25                    x = np.fromfile(f, 'd', record_length)
26                    checksum += x.sum()
27                except:
28                    break
29        assert(np.abs(checksum) < 1e-6)
30
31    def read_binary_npfrombuffer(filename):
32        checksum = 0.0
33        with open(filename, 'rb') as f:
34            while True:
35                try:
36                    record_length = np.frombuffer(f.read(np.dtype('i4').itemsize), dtype = 'i4', count = 1)[0]
37                    x = np.frombuffer(f.read(np.dtype('d').itemsize * record_length), dtype = 'd', count = record_length)
38                    checksum += x.sum()
39                except:
40                    break
41        assert(np.abs(checksum) < 1e-6)
42    
43    
44    if __name__ == '__main__':
45        from timeit import Timer
46        from functools import partial
47    
48        fname = 'testfile.tmp'
49        print("# File size[MB], Timings and errors [s]: fromfile, frombuffer")
50        for i in [10**3, 3*10**3, 10**4, 3*10**4, 10**5, 3*10**5, 10**6, 3*10**6]:
51            fsize = generate_binary_file(fname, i)
52            t1 = Timer(partial(read_binary_npfromfile, fname))
53            t2 = Timer(partial(read_binary_npfrombuffer, fname))
54            a1 = np.array(t1.repeat(5, 1))
55            a2 = np.array(t2.repeat(5, 1))
56            print('%8.3f %12.6f %12.6f %12.6f %12.6f' % (1.0 * fsize / (2**20), a1.mean(), a1.std(), a2.mean(), a2.std()))
57def read_binary_faster_numpy(filename):
58    buff = np.fromfile(filename, dtype=np.uint8)
59    buff_int32 = buff.view(np.int32)
60    buff_double_1 = buff[0:len(buff)//8*8].view(np.float64)
61    buff_double_2 = buff[4:4+(len(buff)-4)//8*8].view(np.float64)
62    nblocks = buff.size // 4      # Number of 4-byte blocks
63    pos = 0                       # Displacement by block of 4 bytes
64    lst = []
65    while pos < nblocks:
66        record_length = buff_int32[pos]
67        pos += 1
68        if pos + record_length * 2 > nblocks:
69            break
70        offset = pos // 2
71        if pos % 2 == 0:          # Aligned with buff_double_1
72            x = buff_double_1[offset:offset+record_length]
73        else:                     # Aligned with buff_double_2
74            x = buff_double_2[offset:offset+record_length]
75        lst.append(x)             # np.sum is too expensive here
76        pos += record_length * 2
77    checksum = np.sum(np.concatenate(lst))
78    assert(np.abs(checksum) < 1e-6)
79from struct import unpack_from
80
81def read_binary_python_struct(filename):
82    checksum = 0.0
83    with open(filename, 'rb') as f:
84        data = f.read()
85        offset = 0
86        while offset < len(data):
87            record_length = unpack_from('@i', data, offset)[0]
88            checksum += sum(unpack_from(f'{record_length}d', data, offset + 4))
89            offset += 4 + record_length * 8
90    assert(np.abs(checksum) < 1e-6)
91

1    import numpy as np
2    import os
3    
4    def generate_binary_file(filename, nrecords):
5        n_records = np.random.poisson(lam = nrecords)
6        record_lengths = np.random.poisson(lam = 10, size = n_records).astype(dtype = 'i4')
7        x = np.random.normal(size = record_lengths.sum()).astype(dtype = 'd')
8        with open(filename, 'wb') as f:
9            s = 0
10            for i in range(n_records):
11                f.write(record_lengths[i].tobytes())
12                f.write(x[s:s+record_lengths[i]].tobytes())
13                s += record_lengths[i]
14            # Trick for testing: make sum of records equal to 0
15            f.write(np.array([1], dtype = 'i4').tobytes())
16            f.write(np.array([-x.sum()], dtype = 'd').tobytes())
17        return os.path.getsize(filename)
18    
19    def read_binary_npfromfile(filename):
20        checksum = 0.0
21        with open(filename, 'rb') as f:
22            while True:
23                try:
24                    record_length = np.fromfile(f, 'i4', 1)[0]
25                    x = np.fromfile(f, 'd', record_length)
26                    checksum += x.sum()
27                except:
28                    break
29        assert(np.abs(checksum) < 1e-6)
30
31    def read_binary_npfrombuffer(filename):
32        checksum = 0.0
33        with open(filename, 'rb') as f:
34            while True:
35                try:
36                    record_length = np.frombuffer(f.read(np.dtype('i4').itemsize), dtype = 'i4', count = 1)[0]
37                    x = np.frombuffer(f.read(np.dtype('d').itemsize * record_length), dtype = 'd', count = record_length)
38                    checksum += x.sum()
39                except:
40                    break
41        assert(np.abs(checksum) < 1e-6)
42    
43    
44    if __name__ == '__main__':
45        from timeit import Timer
46        from functools import partial
47    
48        fname = 'testfile.tmp'
49        print("# File size[MB], Timings and errors [s]: fromfile, frombuffer")
50        for i in [10**3, 3*10**3, 10**4, 3*10**4, 10**5, 3*10**5, 10**6, 3*10**6]:
51            fsize = generate_binary_file(fname, i)
52            t1 = Timer(partial(read_binary_npfromfile, fname))
53            t2 = Timer(partial(read_binary_npfrombuffer, fname))
54            a1 = np.array(t1.repeat(5, 1))
55            a2 = np.array(t2.repeat(5, 1))
56            print('%8.3f %12.6f %12.6f %12.6f %12.6f' % (1.0 * fsize / (2**20), a1.mean(), a1.std(), a2.mean(), a2.std()))
57def read_binary_faster_numpy(filename):
58    buff = np.fromfile(filename, dtype=np.uint8)
59    buff_int32 = buff.view(np.int32)
60    buff_double_1 = buff[0:len(buff)//8*8].view(np.float64)
61    buff_double_2 = buff[4:4+(len(buff)-4)//8*8].view(np.float64)
62    nblocks = buff.size // 4      # Number of 4-byte blocks
63    pos = 0                       # Displacement by block of 4 bytes
64    lst = []
65    while pos < nblocks:
66        record_length = buff_int32[pos]
67        pos += 1
68        if pos + record_length * 2 > nblocks:
69            break
70        offset = pos // 2
71        if pos % 2 == 0:          # Aligned with buff_double_1
72            x = buff_double_1[offset:offset+record_length]
73        else:                     # Aligned with buff_double_2
74            x = buff_double_2[offset:offset+record_length]
75        lst.append(x)             # np.sum is too expensive here
76        pos += record_length * 2
77    checksum = np.sum(np.concatenate(lst))
78    assert(np.abs(checksum) < 1e-6)
79from struct import unpack_from
80
81def read_binary_python_struct(filename):
82    checksum = 0.0
83    with open(filename, 'rb') as f:
84        data = f.read()
85        offset = 0
86        while offset < len(data):
87            record_length = unpack_from('@i', data, offset)[0]
88            checksum += sum(unpack_from(f'{record_length}d', data, offset + 4))
89            offset += 4 + record_length * 8
90    assert(np.abs(checksum) < 1e-6)
91@nb.njit('float64(uint8[::1])')
92def decode_buffer(buff):
93    checksum = 0.0
94    offset = 0
95    while offset + 4 < buff.size:
96        record_length = buff[offset:offset+4].view(np.int32)[0]
97        start = offset + 4
98        end = start + record_length * 8
99        if end > buff.size:
100            break
101        x = buff[start:end].view(np.float64)
102        checksum += x.sum()
103        offset = end
104    return checksum
105
106def read_binary_numba(filename):
107    buff = np.fromfile(filename, dtype=np.uint8)
108    checksum = decode_buffer(buff)
109    assert(np.abs(checksum) < 1e-6)
110

1    import numpy as np
2    import os
3    
4    def generate_binary_file(filename, nrecords):
5        n_records = np.random.poisson(lam = nrecords)
6        record_lengths = np.random.poisson(lam = 10, size = n_records).astype(dtype = 'i4')
7        x = np.random.normal(size = record_lengths.sum()).astype(dtype = 'd')
8        with open(filename, 'wb') as f:
9            s = 0
10            for i in range(n_records):
11                f.write(record_lengths[i].tobytes())
12                f.write(x[s:s+record_lengths[i]].tobytes())
13                s += record_lengths[i]
14            # Trick for testing: make sum of records equal to 0
15            f.write(np.array([1], dtype = 'i4').tobytes())
16            f.write(np.array([-x.sum()], dtype = 'd').tobytes())
17        return os.path.getsize(filename)
18    
19    def read_binary_npfromfile(filename):
20        checksum = 0.0
21        with open(filename, 'rb') as f:
22            while True:
23                try:
24                    record_length = np.fromfile(f, 'i4', 1)[0]
25                    x = np.fromfile(f, 'd', record_length)
26                    checksum += x.sum()
27                except:
28                    break
29        assert(np.abs(checksum) < 1e-6)
30
31    def read_binary_npfrombuffer(filename):
32        checksum = 0.0
33        with open(filename, 'rb') as f:
34            while True:
35                try:
36                    record_length = np.frombuffer(f.read(np.dtype('i4').itemsize), dtype = 'i4', count = 1)[0]
37                    x = np.frombuffer(f.read(np.dtype('d').itemsize * record_length), dtype = 'd', count = record_length)
38                    checksum += x.sum()
39                except:
40                    break
41        assert(np.abs(checksum) < 1e-6)
42    
43    
44    if __name__ == '__main__':
45        from timeit import Timer
46        from functools import partial
47    
48        fname = 'testfile.tmp'
49        print("# File size[MB], Timings and errors [s]: fromfile, frombuffer")
50        for i in [10**3, 3*10**3, 10**4, 3*10**4, 10**5, 3*10**5, 10**6, 3*10**6]:
51            fsize = generate_binary_file(fname, i)
52            t1 = Timer(partial(read_binary_npfromfile, fname))
53            t2 = Timer(partial(read_binary_npfrombuffer, fname))
54            a1 = np.array(t1.repeat(5, 1))
55            a2 = np.array(t2.repeat(5, 1))
56            print('%8.3f %12.6f %12.6f %12.6f %12.6f' % (1.0 * fsize / (2**20), a1.mean(), a1.std(), a2.mean(), a2.std()))
57def read_binary_faster_numpy(filename):
58    buff = np.fromfile(filename, dtype=np.uint8)
59    buff_int32 = buff.view(np.int32)
60    buff_double_1 = buff[0:len(buff)//8*8].view(np.float64)
61    buff_double_2 = buff[4:4+(len(buff)-4)//8*8].view(np.float64)
62    nblocks = buff.size // 4      # Number of 4-byte blocks
63    pos = 0                       # Displacement by block of 4 bytes
64    lst = []
65    while pos < nblocks:
66        record_length = buff_int32[pos]
67        pos += 1
68        if pos + record_length * 2 > nblocks:
69            break
70        offset = pos // 2
71        if pos % 2 == 0:          # Aligned with buff_double_1
72            x = buff_double_1[offset:offset+record_length]
73        else:                     # Aligned with buff_double_2
74            x = buff_double_2[offset:offset+record_length]
75        lst.append(x)             # np.sum is too expensive here
76        pos += record_length * 2
77    checksum = np.sum(np.concatenate(lst))
78    assert(np.abs(checksum) < 1e-6)
79from struct import unpack_from
80
81def read_binary_python_struct(filename):
82    checksum = 0.0
83    with open(filename, 'rb') as f:
84        data = f.read()
85        offset = 0
86        while offset < len(data):
87            record_length = unpack_from('@i', data, offset)[0]
88            checksum += sum(unpack_from(f'{record_length}d', data, offset + 4))
89            offset += 4 + record_length * 8
90    assert(np.abs(checksum) < 1e-6)
91@nb.njit('float64(uint8[::1])')
92def decode_buffer(buff):
93    checksum = 0.0
94    offset = 0
95    while offset + 4 < buff.size:
96        record_length = buff[offset:offset+4].view(np.int32)[0]
97        start = offset + 4
98        end = start + record_length * 8
99        if end > buff.size:
100            break
101        x = buff[start:end].view(np.float64)
102        checksum += x.sum()
103        offset = end
104    return checksum
105
106def read_binary_numba(filename):
107    buff = np.fromfile(filename, dtype=np.uint8)
108    checksum = decode_buffer(buff)
109    assert(np.abs(checksum) < 1e-6)
110read_binary_npfromfile:      10616 ms    (   x1)
111read_binary_npfrombuffer:     1132 ms    (   x9)
112read_binary_faster_numpy:      509 ms    (  x21)
113read_binary_python_struct:     222 ms    (  x48)
114read_binary_numba:              12 ms    ( x885)
115Optimal time:                    7 ms    (x1517)
116

1@tf.function
2def create_tf_dataset_for_client_fn(dataset_path):
3   return tf.data.experimental.CsvDataset(dataset_path, 
4                                          record_defaults=record_defaults,
5                                          header=True)
6   
7@tf.function
8def add_parsing(dataset):
9  def parse_dataset(*x):
10    return OrderedDict([('y', x[-1]), ('x', x[1:-1])])
11  return dataset.map(parse_dataset, num_parallel_calls=tf.data.AUTOTUNE)
12source = tff.simulation.datasets.FilePerUserClientData(
13  dataset_paths, create_tf_dataset_for_client_fn) 
14
15client_ids = sorted(source.client_ids)
16
17# Make sure the client ids are tensor strings when splitting data.
18source._client_ids = [tf.cast(c, tf.string) for c in source.client_ids] 
19source = source.preprocess(add_parsing)
20
21train, test = source.train_test_client_split(source, 1)
22
23train_client_ids = train.client_ids
24
25train_data = train.create_tf_dataset_for_client(train_client_ids[0])
26
27def create_keras_model():
28  initializer = tf.keras.initializers.GlorotNormal(seed=0)
29  return tf.keras.models.Sequential([
30      tf.keras.layers.Input(shape=(32,)),
31      tf.keras.layers.Dense(10, kernel_initializer=initializer),
32      tf.keras.layers.Softmax(),
33  ])
34def model_fn():
35  keras_model = create_keras_model()
36  return tff.learning.from_keras_model(
37      keras_model,
38      input_spec=train_data.element_spec,
39      loss=tf.keras.losses.SparseCategoricalCrossentropy(),
40      metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
41ValueError: in user code:
42
43   File "<ipython-input-14-cada45ffae0f>", line 12, in client_update  *
44       for batch in dataset:
45   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 455, in forward_pass  *
46       return self._forward_pass(batch_input, training=training)
47   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 408, in _forward_pass  *
48       predictions = self.predict_on_batch(inputs, training)
49   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 398, in predict_on_batch  *
50       return self._keras_model(x, training=training)
51   File "/usr/local/lib/python3.7/dist-packages/keras/engine/base_layer_v1.py", line 740, in __call__  **
52       self.name)
53   File "/usr/local/lib/python3.7/dist-packages/keras/engine/input_spec.py", line 200, in assert_input_compatibility
54       raise ValueError(f'Layer "{layer_name}" expects {len(input_spec)} input(s),'
55
56ValueError: Layer "sequential" expects 1 input(s), but it received 10 input tensors. Inputs received: [<tf.Tensor 'x:0' shape=() dtype=int32>, <tf.Tensor 'x_1:0' shape=() dtype=int32>, <tf.Tensor 'x_2:0' shape=() dtype=int32>, <tf.Tensor 'x_3:0' shape=() dtype=float32>, <tf.Tensor 'x_4:0' shape=() dtype=float32>, <tf.Tensor 'x_5:0' shape=() dtype=float32>, <tf.Tensor 'x_6:0' shape=() dtype=float32>, <tf.Tensor 'x_7:0' shape=() dtype=float32>, <tf.Tensor 'x_8:0' shape=() dtype=float32>, <tf.Tensor 'x_9:0' shape=() dtype=int32>]
57import tensorflow as tf
58import tensorflow_federated as tff
59import pandas as pd
60from collections import OrderedDict
61import nest_asyncio
62nest_asyncio.apply()
63
64# Dummy data
65samples = 5
66data = [[tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
67        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
68        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
69        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
70        tf.random.normal((samples,)).numpy().tolist(),
71        tf.random.normal((samples,)).numpy().tolist(),
72        tf.random.normal((samples,)).numpy().tolist(),
73        tf.random.normal((samples,)).numpy().tolist(),
74        tf.random.normal((samples,)).numpy().tolist(),
75        tf.random.normal((samples,)).numpy().tolist(),
76        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
77        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist()]]
78df = pd.DataFrame(data)
79df = df.explode(list(df.columns))
80df.to_csv('client1.csv', index= False)
81df.to_csv('client2.csv', index= False)
82

1@tf.function
2def create_tf_dataset_for_client_fn(dataset_path):
3   return tf.data.experimental.CsvDataset(dataset_path, 
4                                          record_defaults=record_defaults,
5                                          header=True)
6   
7@tf.function
8def add_parsing(dataset):
9  def parse_dataset(*x):
10    return OrderedDict([('y', x[-1]), ('x', x[1:-1])])
11  return dataset.map(parse_dataset, num_parallel_calls=tf.data.AUTOTUNE)
12source = tff.simulation.datasets.FilePerUserClientData(
13  dataset_paths, create_tf_dataset_for_client_fn) 
14
15client_ids = sorted(source.client_ids)
16
17# Make sure the client ids are tensor strings when splitting data.
18source._client_ids = [tf.cast(c, tf.string) for c in source.client_ids] 
19source = source.preprocess(add_parsing)
20
21train, test = source.train_test_client_split(source, 1)
22
23train_client_ids = train.client_ids
24
25train_data = train.create_tf_dataset_for_client(train_client_ids[0])
26
27def create_keras_model():
28  initializer = tf.keras.initializers.GlorotNormal(seed=0)
29  return tf.keras.models.Sequential([
30      tf.keras.layers.Input(shape=(32,)),
31      tf.keras.layers.Dense(10, kernel_initializer=initializer),
32      tf.keras.layers.Softmax(),
33  ])
34def model_fn():
35  keras_model = create_keras_model()
36  return tff.learning.from_keras_model(
37      keras_model,
38      input_spec=train_data.element_spec,
39      loss=tf.keras.losses.SparseCategoricalCrossentropy(),
40      metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
41ValueError: in user code:
42
43   File "<ipython-input-14-cada45ffae0f>", line 12, in client_update  *
44       for batch in dataset:
45   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 455, in forward_pass  *
46       return self._forward_pass(batch_input, training=training)
47   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 408, in _forward_pass  *
48       predictions = self.predict_on_batch(inputs, training)
49   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 398, in predict_on_batch  *
50       return self._keras_model(x, training=training)
51   File "/usr/local/lib/python3.7/dist-packages/keras/engine/base_layer_v1.py", line 740, in __call__  **
52       self.name)
53   File "/usr/local/lib/python3.7/dist-packages/keras/engine/input_spec.py", line 200, in assert_input_compatibility
54       raise ValueError(f'Layer "{layer_name}" expects {len(input_spec)} input(s),'
55
56ValueError: Layer "sequential" expects 1 input(s), but it received 10 input tensors. Inputs received: [<tf.Tensor 'x:0' shape=() dtype=int32>, <tf.Tensor 'x_1:0' shape=() dtype=int32>, <tf.Tensor 'x_2:0' shape=() dtype=int32>, <tf.Tensor 'x_3:0' shape=() dtype=float32>, <tf.Tensor 'x_4:0' shape=() dtype=float32>, <tf.Tensor 'x_5:0' shape=() dtype=float32>, <tf.Tensor 'x_6:0' shape=() dtype=float32>, <tf.Tensor 'x_7:0' shape=() dtype=float32>, <tf.Tensor 'x_8:0' shape=() dtype=float32>, <tf.Tensor 'x_9:0' shape=() dtype=int32>]
57import tensorflow as tf
58import tensorflow_federated as tff
59import pandas as pd
60from collections import OrderedDict
61import nest_asyncio
62nest_asyncio.apply()
63
64# Dummy data
65samples = 5
66data = [[tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
67        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
68        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
69        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
70        tf.random.normal((samples,)).numpy().tolist(),
71        tf.random.normal((samples,)).numpy().tolist(),
72        tf.random.normal((samples,)).numpy().tolist(),
73        tf.random.normal((samples,)).numpy().tolist(),
74        tf.random.normal((samples,)).numpy().tolist(),
75        tf.random.normal((samples,)).numpy().tolist(),
76        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
77        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist()]]
78df = pd.DataFrame(data)
79df = df.explode(list(df.columns))
80df.to_csv('client1.csv', index= False)
81df.to_csv('client2.csv', index= False)
82import tensorflow as tf
83
84record_defaults = [int(), int(), int(), int(), float(),float(),float(),float(),float(),float(), int(), int()]
85
86@tf.function
87def create_tf_dataset_for_client_fn(dataset_path):
88   return tf.data.experimental.CsvDataset(dataset_path, 
89                                          record_defaults=record_defaults,
90                                          header=True)
91@tf.function
92def add_parsing(dataset):
93  def parse_dataset(*x):
94    return OrderedDict([('y', x[-1]), ('x', x[1:-1])])
95  return dataset.map(parse_dataset, num_parallel_calls=tf.data.AUTOTUNE)
96
97dataset_paths = {'client1': '/content/client1.csv', 'client2': '/content/client2.csv'}
98
99source = tff.simulation.datasets.FilePerUserClientData(
100  dataset_paths, create_tf_dataset_for_client_fn) 
101
102client_ids = sorted(source.client_ids)
103
104# Make sure the client ids are tensor strings when splitting data.
105source._client_ids = [tf.cast(c, tf.string) for c in source.client_ids] 
106source = source.preprocess(add_parsing)
107
108train, test = source.train_test_client_split(source, 1)
109
110train_client_ids = train.client_ids
111
112def reshape_data(d):
113  d['x'] = tf.stack([tf.cast(x, dtype=tf.float32) for x in d['x']])
114  return d
115
116train_data = [train.create_tf_dataset_for_client(c).map(reshape_data).batch(1) for c in train_client_ids]
117

1@tf.function
2def create_tf_dataset_for_client_fn(dataset_path):
3   return tf.data.experimental.CsvDataset(dataset_path, 
4                                          record_defaults=record_defaults,
5                                          header=True)
6   
7@tf.function
8def add_parsing(dataset):
9  def parse_dataset(*x):
10    return OrderedDict([('y', x[-1]), ('x', x[1:-1])])
11  return dataset.map(parse_dataset, num_parallel_calls=tf.data.AUTOTUNE)
12source = tff.simulation.datasets.FilePerUserClientData(
13  dataset_paths, create_tf_dataset_for_client_fn) 
14
15client_ids = sorted(source.client_ids)
16
17# Make sure the client ids are tensor strings when splitting data.
18source._client_ids = [tf.cast(c, tf.string) for c in source.client_ids] 
19source = source.preprocess(add_parsing)
20
21train, test = source.train_test_client_split(source, 1)
22
23train_client_ids = train.client_ids
24
25train_data = train.create_tf_dataset_for_client(train_client_ids[0])
26
27def create_keras_model():
28  initializer = tf.keras.initializers.GlorotNormal(seed=0)
29  return tf.keras.models.Sequential([
30      tf.keras.layers.Input(shape=(32,)),
31      tf.keras.layers.Dense(10, kernel_initializer=initializer),
32      tf.keras.layers.Softmax(),
33  ])
34def model_fn():
35  keras_model = create_keras_model()
36  return tff.learning.from_keras_model(
37      keras_model,
38      input_spec=train_data.element_spec,
39      loss=tf.keras.losses.SparseCategoricalCrossentropy(),
40      metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
41ValueError: in user code:
42
43   File "<ipython-input-14-cada45ffae0f>", line 12, in client_update  *
44       for batch in dataset:
45   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 455, in forward_pass  *
46       return self._forward_pass(batch_input, training=training)
47   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 408, in _forward_pass  *
48       predictions = self.predict_on_batch(inputs, training)
49   File "/usr/local/lib/python3.7/dist-packages/tensorflow_federated/python/learning/keras_utils.py", line 398, in predict_on_batch  *
50       return self._keras_model(x, training=training)
51   File "/usr/local/lib/python3.7/dist-packages/keras/engine/base_layer_v1.py", line 740, in __call__  **
52       self.name)
53   File "/usr/local/lib/python3.7/dist-packages/keras/engine/input_spec.py", line 200, in assert_input_compatibility
54       raise ValueError(f'Layer "{layer_name}" expects {len(input_spec)} input(s),'
55
56ValueError: Layer "sequential" expects 1 input(s), but it received 10 input tensors. Inputs received: [<tf.Tensor 'x:0' shape=() dtype=int32>, <tf.Tensor 'x_1:0' shape=() dtype=int32>, <tf.Tensor 'x_2:0' shape=() dtype=int32>, <tf.Tensor 'x_3:0' shape=() dtype=float32>, <tf.Tensor 'x_4:0' shape=() dtype=float32>, <tf.Tensor 'x_5:0' shape=() dtype=float32>, <tf.Tensor 'x_6:0' shape=() dtype=float32>, <tf.Tensor 'x_7:0' shape=() dtype=float32>, <tf.Tensor 'x_8:0' shape=() dtype=float32>, <tf.Tensor 'x_9:0' shape=() dtype=int32>]
57import tensorflow as tf
58import tensorflow_federated as tff
59import pandas as pd
60from collections import OrderedDict
61import nest_asyncio
62nest_asyncio.apply()
63
64# Dummy data
65samples = 5
66data = [[tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
67        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
68        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
69        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
70        tf.random.normal((samples,)).numpy().tolist(),
71        tf.random.normal((samples,)).numpy().tolist(),
72        tf.random.normal((samples,)).numpy().tolist(),
73        tf.random.normal((samples,)).numpy().tolist(),
74        tf.random.normal((samples,)).numpy().tolist(),
75        tf.random.normal((samples,)).numpy().tolist(),
76        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist(),
77        tf.random.uniform((samples,), maxval=50, dtype=tf.int32).numpy().tolist()]]
78df = pd.DataFrame(data)
79df = df.explode(list(df.columns))
80df.to_csv('client1.csv', index= False)
81df.to_csv('client2.csv', index= False)
82import tensorflow as tf
83
84record_defaults = [int(), int(), int(), int(), float(),float(),float(),float(),float(),float(), int(), int()]
85
86@tf.function
87def create_tf_dataset_for_client_fn(dataset_path):
88   return tf.data.experimental.CsvDataset(dataset_path, 
89                                          record_defaults=record_defaults,
90                                          header=True)
91@tf.function
92def add_parsing(dataset):
93  def parse_dataset(*x):
94    return OrderedDict([('y', x[-1]), ('x', x[1:-1])])
95  return dataset.map(parse_dataset, num_parallel_calls=tf.data.AUTOTUNE)
96
97dataset_paths = {'client1': '/content/client1.csv', 'client2': '/content/client2.csv'}
98
99source = tff.simulation.datasets.FilePerUserClientData(
100  dataset_paths, create_tf_dataset_for_client_fn) 
101
102client_ids = sorted(source.client_ids)
103
104# Make sure the client ids are tensor strings when splitting data.
105source._client_ids = [tf.cast(c, tf.string) for c in source.client_ids] 
106source = source.preprocess(add_parsing)
107
108train, test = source.train_test_client_split(source, 1)
109
110train_client_ids = train.client_ids
111
112def reshape_data(d):
113  d['x'] = tf.stack([tf.cast(x, dtype=tf.float32) for x in d['x']])
114  return d
115
116train_data = [train.create_tf_dataset_for_client(c).map(reshape_data).batch(1) for c in train_client_ids]
117def create_keras_model():
118  initializer = tf.keras.initializers.GlorotNormal(seed=0)
119  return tf.keras.models.Sequential([
120      tf.keras.layers.Input(shape=(10,)),
121      tf.keras.layers.Dense(75, kernel_initializer=initializer),
122      tf.keras.layers.Dense(50, kernel_initializer=initializer),
123      tf.keras.layers.Softmax(),
124  ])
125def model_fn():
126  keras_model = create_keras_model()
127  return tff.learning.from_keras_model(
128      keras_model,
129      input_spec=train_data[0].element_spec,
130      loss=tf.keras.losses.SparseCategoricalCrossentropy(),
131      metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])
132
133def initialize_fn():
134  model = model_fn()
135  return model.trainable_variables
136
137@tf.function
138def client_update(model, dataset, server_weights, client_optimizer):
139  """Performs training (using the server model weights) on the client's dataset."""
140  client_weights = model.trainable_variables
141  tf.nest.map_structure(lambda x, y: x.assign(y),
142                        client_weights, server_weights)
143
144  for batch in dataset:
145    with tf.GradientTape() as tape:
146      outputs = model.forward_pass(batch)
147
148    grads = tape.gradient(outputs.loss, client_weights)
149    grads_and_vars = zip(grads, client_weights)
150    client_optimizer.apply_gradients(grads_and_vars)
151
152  return client_weights
153
154@tf.function
155def server_update(model, mean_client_weights):
156  """Updates the server model weights as the average of the client model weights."""
157  model_weights = model.trainable_variables
158  tf.nest.map_structure(lambda x, y: x.assign(y),
159                        model_weights, mean_client_weights)
160  return model_weights
161
162federated_float_on_clients = tff.FederatedType(tf.float32, tff.CLIENTS)
163
164@tff.federated_computation(tff.FederatedType(tf.float32, tff.CLIENTS))
165def get_average_temperature(client_temperatures):
166  return tff.federated_mean(client_temperatures)
167str(get_average_temperature.type_signature)
168get_average_temperature([68.5, 70.3, 69.8])
169
170@tff.tf_computation
171def server_init():
172  model = model_fn()
173  return model.trainable_variables
174
175@tff.federated_computation
176def initialize_fn():
177  return tff.federated_value(server_init(), tff.SERVER)
178
179whimsy_model = model_fn()
180tf_dataset_type = tff.SequenceType(whimsy_model.input_spec)
181model_weights_type = server_init.type_signature.result
182
183@tff.tf_computation(tf_dataset_type, model_weights_type)
184def client_update_fn(tf_dataset, server_weights):
185  model = model_fn()
186  client_optimizer = tf.keras.optimizers.SGD(learning_rate=0.01)
187  return client_update(model, tf_dataset, server_weights, client_optimizer)
188
189@tff.tf_computation(model_weights_type)
190def server_update_fn(mean_client_weights):
191  model = model_fn()
192  return server_update(model, mean_client_weights)
193
194federated_server_type = tff.FederatedType(model_weights_type, tff.SERVER)
195federated_dataset_type = tff.FederatedType(tf_dataset_type, tff.CLIENTS)
196
197@tff.federated_computation(federated_server_type, federated_dataset_type)
198def next_fn(server_weights, federated_dataset):
199  server_weights_at_client = tff.federated_broadcast(server_weights)
200  client_weights = tff.federated_map(
201      client_update_fn, (federated_dataset, server_weights_at_client))
202  mean_client_weights = tff.federated_mean(client_weights)
203
204  server_weights = tff.federated_map(server_update_fn, mean_client_weights)
205  return server_weights
206
207federated_algorithm = tff.templates.IterativeProcess(
208    initialize_fn=initialize_fn,
209    next_fn=next_fn
210)
211
212server_state = federated_algorithm.initialize()
213for round in range(15):
214  server_state = federated_algorithm.next(server_state, train_data)
215

1void convolution2D(int isize,   // width/height of input
2        int osize,              // width/height of output
3        int ksize,              // width/height of kernel
4        int stride,             // shift between input pixels, between consecutive outputs
5        int pad,                // offset between (0,0) pixels between input and output
6        int idepth, int odepth, // number of input and output channels
7        float idata[isize][isize][idepth],
8        float odata[osize][osize][odepth],
9        float kdata[odepth][ksize][ksize][idepth])
10{
11    // iterate over the output
12    for (int oy = 0; oy < osize; ++oy) {
13    for (int ox = 0; ox < osize; ++ox) {
14    for (int od = 0; od < odepth; ++od) {
15        odata[oy][ox][od] = 0;  // When you iterate multiple times without closing the program, this number would stack up to infinity, so we have to zero it out every time.
16        for (int ky = 0; ky < ksize; ++ky) {
17        for (int kx = 0; kx < ksize; ++kx) {
18            // map position in output and kernel to the input
19            int iy = stride * oy + ky - pad;
20            int ix = stride * ox + kx - pad;
21            // use only valid inputs
22            if (iy >= 0 && iy < isize && ix >= 0 && ix < isize)
23                for (int id = 0; id < idepth; ++id)
24                    odata[oy][ox][od] += kdata[od][ky][kx][id] * idata[iy][ix][id];
25        }}
26    }}}
27
28}
29
300000000000007880 <convolution2D>:
31    7880:   f3 0f 1e fa             endbr64 
32    7884:   55                      push   rbp
33    7885:   48 89 e5                mov    rbp,rsp
34    7888:   41 57                   push   r15
35    788a:   41 56                   push   r14
36    788c:   41 55                   push   r13
37    788e:   41 54                   push   r12
38    7890:   53                      push   rbx
39    7891:   48 81 ec b0 00 00 00    sub    rsp,0xb0
40    7898:   ff 15 4a a7 00 00       call   QWORD PTR [rip+0xa74a]        # 11fe8 <mcount@GLIBC_2.2.5>
41    789e:   89 d3                   mov    ebx,edx
42    78a0:   89 55 a8                mov    DWORD PTR [rbp-0x58],edx
43    78a3:   89 8d 74 ff ff ff       mov    DWORD PTR [rbp-0x8c],ecx
44    78a9:   49 63 d1                movsxd rdx,r9d
45    78ac:   48 63 cf                movsxd rcx,edi
46    78af:   41 89 f2                mov    r10d,esi
47    78b2:   89 b5 38 ff ff ff       mov    DWORD PTR [rbp-0xc8],esi
48    78b8:   49 63 c0                movsxd rax,r8d
49    78bb:   48 0f af ca             imul   rcx,rdx
50    78bf:   48 63 75 10             movsxd rsi,DWORD PTR [rbp+0x10]
51    78c3:   49 89 d6                mov    r14,rdx
52    78c6:   4c 8d 24 95 00 00 00 00     lea    r12,[rdx*4+0x0]
53    78ce:   41 89 fd                mov    r13d,edi
54    78d1:   49 89 cb                mov    r11,rcx
55    78d4:   48 89 8d 60 ff ff ff    mov    QWORD PTR [rbp-0xa0],rcx
56    78db:   49 63 ca                movsxd rcx,r10d
57    78de:   4c 8d 0c b5 00 00 00 00     lea    r9,[rsi*4+0x0]
58    78e6:   49 89 f0                mov    r8,rsi
59    78e9:   48 0f af f1             imul   rsi,rcx
60    78ed:   48 63 cb                movsxd rcx,ebx
61    78f0:   4c 89 8d 48 ff ff ff    mov    QWORD PTR [rbp-0xb8],r9
62    78f7:   48 0f af d1             imul   rdx,rcx
63    78fb:   48 8d 3c 95 00 00 00 00     lea    rdi,[rdx*4+0x0]
64    7903:   45 85 d2                test   r10d,r10d
65    7906:   0f 8e 73 02 00 00       jle    7b7f <convolution2D+0x2ff>
66    790c:   48 c1 ef 02             shr    rdi,0x2
67    7910:   49 c1 e9 02             shr    r9,0x2
68    7914:   48 89 7d c8             mov    QWORD PTR [rbp-0x38],rdi
69    7918:   4c 89 e7                mov    rdi,r12
70    791b:   4c 89 8d 58 ff ff ff    mov    QWORD PTR [rbp-0xa8],r9
71    7922:   48 c1 ef 02             shr    rdi,0x2
72    7926:   48 89 bd 50 ff ff ff    mov    QWORD PTR [rbp-0xb0],rdi
73    792d:   45 85 c0                test   r8d,r8d
74    7930:   0f 8e 49 02 00 00       jle    7b7f <convolution2D+0x2ff>
75    7936:   48 c1 e6 02             shl    rsi,0x2
76    793a:   48 0f af d1             imul   rdx,rcx
77    793e:   29 c3                   sub    ebx,eax
78    7940:   89 c7                   mov    edi,eax
79    7942:   48 89 b5 30 ff ff ff    mov    QWORD PTR [rbp-0xd0],rsi
80    7949:   48 8b 75 20             mov    rsi,QWORD PTR [rbp+0x20]
81    794d:   48 89 85 68 ff ff ff    mov    QWORD PTR [rbp-0x98],rax
82    7954:   f7 df                   neg    edi
83    7956:   45 8d 7e ff             lea    r15d,[r14-0x1]
84    795a:   89 9d 70 ff ff ff       mov    DWORD PTR [rbp-0x90],ebx
85    7960:   89 bd 3c ff ff ff       mov    DWORD PTR [rbp-0xc4],edi
86    7966:   48 8d 0c 95 00 00 00 00     lea    rcx,[rdx*4+0x0]
87    796e:   89 7d ac                mov    DWORD PTR [rbp-0x54],edi
88    7971:   89 5d d4                mov    DWORD PTR [rbp-0x2c],ebx
89    7974:   48 89 4d 98             mov    QWORD PTR [rbp-0x68],rcx
90    7978:   4a 8d 0c 9d 00 00 00 00     lea    rcx,[r11*4+0x0]
91    7980:   c7 45 80 00 00 00 00    mov    DWORD PTR [rbp-0x80],0x0
92    7987:   48 89 75 88             mov    QWORD PTR [rbp-0x78],rsi
93    798b:   41 8d 70 ff             lea    esi,[r8-0x1]
94    798f:   48 89 4d c0             mov    QWORD PTR [rbp-0x40],rcx
95    7993:   48 8d 04 b5 04 00 00 00     lea    rax,[rsi*4+0x4]
96    799b:   c7 45 90 00 00 00 00    mov    DWORD PTR [rbp-0x70],0x0
97    79a2:   48 89 85 28 ff ff ff    mov    QWORD PTR [rbp-0xd8],rax
98    79a9:   44 89 f0                mov    eax,r14d
99    79ac:   45 89 ee                mov    r14d,r13d
100    79af:   41 89 c5                mov    r13d,eax
101    79b2:   48 8b 85 28 ff ff ff    mov    rax,QWORD PTR [rbp-0xd8]
102    79b9:   48 03 45 88             add    rax,QWORD PTR [rbp-0x78]
103    79bd:   48 c7 85 78 ff ff ff 00 00 00 00    mov    QWORD PTR [rbp-0x88],0x0
104    79c8:   c7 45 84 00 00 00 00    mov    DWORD PTR [rbp-0x7c],0x0
105    79cf:   c7 45 94 00 00 00 00    mov    DWORD PTR [rbp-0x6c],0x0
106    79d6:   44 8b 95 70 ff ff ff    mov    r10d,DWORD PTR [rbp-0x90]
107    79dd:   48 89 45 b0             mov    QWORD PTR [rbp-0x50],rax
108    79e1:   48 63 45 80             movsxd rax,DWORD PTR [rbp-0x80]
109    79e5:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
110    79ec:   48 0f af 85 60 ff ff ff     imul   rax,QWORD PTR [rbp-0xa0]
111    79f4:   48 89 85 40 ff ff ff    mov    QWORD PTR [rbp-0xc0],rax
112    79fb:   8b 85 3c ff ff ff       mov    eax,DWORD PTR [rbp-0xc4]
113    7a01:   89 45 d0                mov    DWORD PTR [rbp-0x30],eax
114    7a04:   48 8b 45 88             mov    rax,QWORD PTR [rbp-0x78]
115    7a08:   48 8b 9d 78 ff ff ff    mov    rbx,QWORD PTR [rbp-0x88]
116    7a0f:   4c 8d 04 98             lea    r8,[rax+rbx*4]
117    7a13:   48 8b 45 28             mov    rax,QWORD PTR [rbp+0x28]
118    7a17:   48 8b 5d 18             mov    rbx,QWORD PTR [rbp+0x18]
119    7a1b:   48 89 45 b8             mov    QWORD PTR [rbp-0x48],rax
120    7a1f:   48 63 45 84             movsxd rax,DWORD PTR [rbp-0x7c]
121    7a23:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
122    7a2a:   48 0f af 85 50 ff ff ff     imul   rax,QWORD PTR [rbp-0xb0]
123    7a32:   48 03 85 40 ff ff ff    add    rax,QWORD PTR [rbp-0xc0]
124    7a39:   48 8d 04 83             lea    rax,[rbx+rax*4]
125    7a3d:   48 89 45 a0             mov    QWORD PTR [rbp-0x60],rax
126    7a41:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
127    7a4c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
128    7a50:   8b 45 a8                mov    eax,DWORD PTR [rbp-0x58]
129    7a53:   41 c7 00 00 00 00 00    mov    DWORD PTR [r8],0x0
130    7a5a:   45 31 db                xor    r11d,r11d
131    7a5d:   48 8b 5d a0             mov    rbx,QWORD PTR [rbp-0x60]
132    7a61:   44 8b 4d ac             mov    r9d,DWORD PTR [rbp-0x54]
133    7a65:   85 c0                   test   eax,eax
134    7a67:   0f 8e 98 00 00 00       jle    7b05 <convolution2D+0x285>
135    7a6d:   0f 1f 00                nop    DWORD PTR [rax]
136    7a70:   45 85 c9                test   r9d,r9d
137    7a73:   78 7b                   js     7af0 <convolution2D+0x270>
138    7a75:   45 39 ce                cmp    r14d,r9d
139    7a78:   7e 76                   jle    7af0 <convolution2D+0x270>
140    7a7a:   48 8b 45 b8             mov    rax,QWORD PTR [rbp-0x48]
141    7a7e:   8b 55 d0                mov    edx,DWORD PTR [rbp-0x30]
142    7a81:   48 89 de                mov    rsi,rbx
143    7a84:   4a 8d 3c 98             lea    rdi,[rax+r11*4]
144    7a88:   eb 13                   jmp    7a9d <convolution2D+0x21d>
145    7a8a:   66 0f 1f 44 00 00       nop    WORD PTR [rax+rax*1+0x0]
146    7a90:   ff c2                   inc    edx
147    7a92:   4c 01 e7                add    rdi,r12
148    7a95:   4c 01 e6                add    rsi,r12
149    7a98:   44 39 d2                cmp    edx,r10d
150    7a9b:   74 53                   je     7af0 <convolution2D+0x270>
151    7a9d:   85 d2                   test   edx,edx
152    7a9f:   78 ef                   js     7a90 <convolution2D+0x210>
153    7aa1:   41 39 d6                cmp    r14d,edx
154    7aa4:   7e ea                   jle    7a90 <convolution2D+0x210>
155    7aa6:   45 85 ed                test   r13d,r13d
156    7aa9:   7e e5                   jle    7a90 <convolution2D+0x210>
157    7aab:   c4 c1 7a 10 08          vmovss xmm1,DWORD PTR [r8]
158    7ab0:   31 c0                   xor    eax,eax
159    7ab2:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
160    7abd:   0f 1f 00                nop    DWORD PTR [rax]
161    7ac0:   c5 fa 10 04 87          vmovss xmm0,DWORD PTR [rdi+rax*4]
162    7ac5:   48 89 c1                mov    rcx,rax
163    7ac8:   c5 fa 59 04 86          vmulss xmm0,xmm0,DWORD PTR [rsi+rax*4]
164    7acd:   48 ff c0                inc    rax
165    7ad0:   c5 f2 58 c8             vaddss xmm1,xmm1,xmm0
166    7ad4:   c4 c1 7a 11 08          vmovss DWORD PTR [r8],xmm1
167    7ad9:   49 39 cf                cmp    r15,rcx
168    7adc:   75 e2                   jne    7ac0 <convolution2D+0x240>
169    7ade:   ff c2                   inc    edx
170    7ae0:   4c 01 e7                add    rdi,r12
171    7ae3:   4c 01 e6                add    rsi,r12
172    7ae6:   44 39 d2                cmp    edx,r10d
173    7ae9:   75 b2                   jne    7a9d <convolution2D+0x21d>
174    7aeb:   0f 1f 44 00 00          nop    DWORD PTR [rax+rax*1+0x0]
175    7af0:   4c 03 5d c8             add    r11,QWORD PTR [rbp-0x38]
176    7af4:   48 03 5d c0             add    rbx,QWORD PTR [rbp-0x40]
177    7af8:   41 ff c1                inc    r9d
178    7afb:   44 3b 4d d4             cmp    r9d,DWORD PTR [rbp-0x2c]
179    7aff:   0f 85 6b ff ff ff       jne    7a70 <convolution2D+0x1f0>
180    7b05:   48 8b 5d 98             mov    rbx,QWORD PTR [rbp-0x68]
181    7b09:   49 83 c0 04             add    r8,0x4
182    7b0d:   48 01 5d b8             add    QWORD PTR [rbp-0x48],rbx
183    7b11:   4c 3b 45 b0             cmp    r8,QWORD PTR [rbp-0x50]
184    7b15:   0f 85 35 ff ff ff       jne    7a50 <convolution2D+0x1d0>
185    7b1b:   8b 9d 74 ff ff ff       mov    ebx,DWORD PTR [rbp-0x8c]
186    7b21:   8b 45 94                mov    eax,DWORD PTR [rbp-0x6c]
187    7b24:   48 8b 8d 48 ff ff ff    mov    rcx,QWORD PTR [rbp-0xb8]
188    7b2b:   01 5d d0                add    DWORD PTR [rbp-0x30],ebx
189    7b2e:   48 01 4d b0             add    QWORD PTR [rbp-0x50],rcx
190    7b32:   01 5d 84                add    DWORD PTR [rbp-0x7c],ebx
191    7b35:   48 8b 8d 58 ff ff ff    mov    rcx,QWORD PTR [rbp-0xa8]
192    7b3c:   41 01 da                add    r10d,ebx
193    7b3f:   48 01 8d 78 ff ff ff    add    QWORD PTR [rbp-0x88],rcx
194    7b46:   ff c0                   inc    eax
195    7b48:   39 85 38 ff ff ff       cmp    DWORD PTR [rbp-0xc8],eax
196    7b4e:   74 08                   je     7b58 <convolution2D+0x2d8>
197    7b50:   89 45 94                mov    DWORD PTR [rbp-0x6c],eax
198    7b53:   e9 ac fe ff ff          jmp    7a04 <convolution2D+0x184>
199    7b58:   8b 4d 90                mov    ecx,DWORD PTR [rbp-0x70]
200    7b5b:   48 8b b5 30 ff ff ff    mov    rsi,QWORD PTR [rbp-0xd0]
201    7b62:   01 5d d4                add    DWORD PTR [rbp-0x2c],ebx
202    7b65:   01 5d ac                add    DWORD PTR [rbp-0x54],ebx
203    7b68:   01 5d 80                add    DWORD PTR [rbp-0x80],ebx
204    7b6b:   48 01 75 88             add    QWORD PTR [rbp-0x78],rsi
205    7b6f:   8d 41 01                lea    eax,[rcx+0x1]
206    7b72:   39 4d 94                cmp    DWORD PTR [rbp-0x6c],ecx
207    7b75:   74 08                   je     7b7f <convolution2D+0x2ff>
208    7b77:   89 45 90                mov    DWORD PTR [rbp-0x70],eax
209    7b7a:   e9 33 fe ff ff          jmp    79b2 <convolution2D+0x132>
210    7b7f:   48 81 c4 b0 00 00 00    add    rsp,0xb0
211    7b86:   5b                      pop    rbx
212    7b87:   41 5c                   pop    r12
213    7b89:   41 5d                   pop    r13
214    7b8b:   41 5e                   pop    r14
215    7b8d:   41 5f                   pop    r15
216    7b8f:   5d                      pop    rbp
217    7b90:   c3                      ret    
218    7b91:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
219    7b9c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
220
221float (*arr)[x][y] = calloc(z, sizeof *arr);
222==430300== Cachegrind, a cache and branch-prediction profiler
223==430300== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
224==430300== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
225==430300== Command: ./EmbeddedNet test 1
226==430300== Parent PID: 170008
227==430300== 
228--430300-- warning: L3 cache found, using its data for the LL simulation.
229==430300== 
230==430300== I   refs:      6,369,594,192
231==430300== I1  misses:            4,271
232==430300== LLi misses:            2,442
233==430300== I1  miss rate:          0.00%
234==430300== LLi miss rate:          0.00%
235==430300== 
236==430300== D   refs:      2,064,233,110  (1,359,003,131 rd   + 705,229,979 wr)
237==430300== D1  misses:       34,476,969  (   19,010,839 rd   +  15,466,130 wr)
238==430300== LLd misses:        5,311,277  (    1,603,955 rd   +   3,707,322 wr)
239==430300== D1  miss rate:           1.7% (          1.4%     +         2.2%  )
240==430300== LLd miss rate:           0.3% (          0.1%     +         0.5%  )
241==430300== 
242==430300== LL refs:          34,481,240  (   19,015,110 rd   +  15,466,130 wr)
243==430300== LL misses:         5,313,719  (    1,606,397 rd   +   3,707,322 wr)
244==430300== LL miss rate:            0.1% (          0.0%     +         0.5%  )
245float fma_scl(float a, float b, float c) {
246    return a * b + c;
247}
248
249fma_scl:
250        vfmadd132ss     xmm0, xmm2, xmm1
251        ret
252

1void convolution2D(int isize,   // width/height of input
2        int osize,              // width/height of output
3        int ksize,              // width/height of kernel
4        int stride,             // shift between input pixels, between consecutive outputs
5        int pad,                // offset between (0,0) pixels between input and output
6        int idepth, int odepth, // number of input and output channels
7        float idata[isize][isize][idepth],
8        float odata[osize][osize][odepth],
9        float kdata[odepth][ksize][ksize][idepth])
10{
11    // iterate over the output
12    for (int oy = 0; oy < osize; ++oy) {
13    for (int ox = 0; ox < osize; ++ox) {
14    for (int od = 0; od < odepth; ++od) {
15        odata[oy][ox][od] = 0;  // When you iterate multiple times without closing the program, this number would stack up to infinity, so we have to zero it out every time.
16        for (int ky = 0; ky < ksize; ++ky) {
17        for (int kx = 0; kx < ksize; ++kx) {
18            // map position in output and kernel to the input
19            int iy = stride * oy + ky - pad;
20            int ix = stride * ox + kx - pad;
21            // use only valid inputs
22            if (iy >= 0 && iy < isize && ix >= 0 && ix < isize)
23                for (int id = 0; id < idepth; ++id)
24                    odata[oy][ox][od] += kdata[od][ky][kx][id] * idata[iy][ix][id];
25        }}
26    }}}
27
28}
29
300000000000007880 <convolution2D>:
31    7880:   f3 0f 1e fa             endbr64 
32    7884:   55                      push   rbp
33    7885:   48 89 e5                mov    rbp,rsp
34    7888:   41 57                   push   r15
35    788a:   41 56                   push   r14
36    788c:   41 55                   push   r13
37    788e:   41 54                   push   r12
38    7890:   53                      push   rbx
39    7891:   48 81 ec b0 00 00 00    sub    rsp,0xb0
40    7898:   ff 15 4a a7 00 00       call   QWORD PTR [rip+0xa74a]        # 11fe8 <mcount@GLIBC_2.2.5>
41    789e:   89 d3                   mov    ebx,edx
42    78a0:   89 55 a8                mov    DWORD PTR [rbp-0x58],edx
43    78a3:   89 8d 74 ff ff ff       mov    DWORD PTR [rbp-0x8c],ecx
44    78a9:   49 63 d1                movsxd rdx,r9d
45    78ac:   48 63 cf                movsxd rcx,edi
46    78af:   41 89 f2                mov    r10d,esi
47    78b2:   89 b5 38 ff ff ff       mov    DWORD PTR [rbp-0xc8],esi
48    78b8:   49 63 c0                movsxd rax,r8d
49    78bb:   48 0f af ca             imul   rcx,rdx
50    78bf:   48 63 75 10             movsxd rsi,DWORD PTR [rbp+0x10]
51    78c3:   49 89 d6                mov    r14,rdx
52    78c6:   4c 8d 24 95 00 00 00 00     lea    r12,[rdx*4+0x0]
53    78ce:   41 89 fd                mov    r13d,edi
54    78d1:   49 89 cb                mov    r11,rcx
55    78d4:   48 89 8d 60 ff ff ff    mov    QWORD PTR [rbp-0xa0],rcx
56    78db:   49 63 ca                movsxd rcx,r10d
57    78de:   4c 8d 0c b5 00 00 00 00     lea    r9,[rsi*4+0x0]
58    78e6:   49 89 f0                mov    r8,rsi
59    78e9:   48 0f af f1             imul   rsi,rcx
60    78ed:   48 63 cb                movsxd rcx,ebx
61    78f0:   4c 89 8d 48 ff ff ff    mov    QWORD PTR [rbp-0xb8],r9
62    78f7:   48 0f af d1             imul   rdx,rcx
63    78fb:   48 8d 3c 95 00 00 00 00     lea    rdi,[rdx*4+0x0]
64    7903:   45 85 d2                test   r10d,r10d
65    7906:   0f 8e 73 02 00 00       jle    7b7f <convolution2D+0x2ff>
66    790c:   48 c1 ef 02             shr    rdi,0x2
67    7910:   49 c1 e9 02             shr    r9,0x2
68    7914:   48 89 7d c8             mov    QWORD PTR [rbp-0x38],rdi
69    7918:   4c 89 e7                mov    rdi,r12
70    791b:   4c 89 8d 58 ff ff ff    mov    QWORD PTR [rbp-0xa8],r9
71    7922:   48 c1 ef 02             shr    rdi,0x2
72    7926:   48 89 bd 50 ff ff ff    mov    QWORD PTR [rbp-0xb0],rdi
73    792d:   45 85 c0                test   r8d,r8d
74    7930:   0f 8e 49 02 00 00       jle    7b7f <convolution2D+0x2ff>
75    7936:   48 c1 e6 02             shl    rsi,0x2
76    793a:   48 0f af d1             imul   rdx,rcx
77    793e:   29 c3                   sub    ebx,eax
78    7940:   89 c7                   mov    edi,eax
79    7942:   48 89 b5 30 ff ff ff    mov    QWORD PTR [rbp-0xd0],rsi
80    7949:   48 8b 75 20             mov    rsi,QWORD PTR [rbp+0x20]
81    794d:   48 89 85 68 ff ff ff    mov    QWORD PTR [rbp-0x98],rax
82    7954:   f7 df                   neg    edi
83    7956:   45 8d 7e ff             lea    r15d,[r14-0x1]
84    795a:   89 9d 70 ff ff ff       mov    DWORD PTR [rbp-0x90],ebx
85    7960:   89 bd 3c ff ff ff       mov    DWORD PTR [rbp-0xc4],edi
86    7966:   48 8d 0c 95 00 00 00 00     lea    rcx,[rdx*4+0x0]
87    796e:   89 7d ac                mov    DWORD PTR [rbp-0x54],edi
88    7971:   89 5d d4                mov    DWORD PTR [rbp-0x2c],ebx
89    7974:   48 89 4d 98             mov    QWORD PTR [rbp-0x68],rcx
90    7978:   4a 8d 0c 9d 00 00 00 00     lea    rcx,[r11*4+0x0]
91    7980:   c7 45 80 00 00 00 00    mov    DWORD PTR [rbp-0x80],0x0
92    7987:   48 89 75 88             mov    QWORD PTR [rbp-0x78],rsi
93    798b:   41 8d 70 ff             lea    esi,[r8-0x1]
94    798f:   48 89 4d c0             mov    QWORD PTR [rbp-0x40],rcx
95    7993:   48 8d 04 b5 04 00 00 00     lea    rax,[rsi*4+0x4]
96    799b:   c7 45 90 00 00 00 00    mov    DWORD PTR [rbp-0x70],0x0
97    79a2:   48 89 85 28 ff ff ff    mov    QWORD PTR [rbp-0xd8],rax
98    79a9:   44 89 f0                mov    eax,r14d
99    79ac:   45 89 ee                mov    r14d,r13d
100    79af:   41 89 c5                mov    r13d,eax
101    79b2:   48 8b 85 28 ff ff ff    mov    rax,QWORD PTR [rbp-0xd8]
102    79b9:   48 03 45 88             add    rax,QWORD PTR [rbp-0x78]
103    79bd:   48 c7 85 78 ff ff ff 00 00 00 00    mov    QWORD PTR [rbp-0x88],0x0
104    79c8:   c7 45 84 00 00 00 00    mov    DWORD PTR [rbp-0x7c],0x0
105    79cf:   c7 45 94 00 00 00 00    mov    DWORD PTR [rbp-0x6c],0x0
106    79d6:   44 8b 95 70 ff ff ff    mov    r10d,DWORD PTR [rbp-0x90]
107    79dd:   48 89 45 b0             mov    QWORD PTR [rbp-0x50],rax
108    79e1:   48 63 45 80             movsxd rax,DWORD PTR [rbp-0x80]
109    79e5:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
110    79ec:   48 0f af 85 60 ff ff ff     imul   rax,QWORD PTR [rbp-0xa0]
111    79f4:   48 89 85 40 ff ff ff    mov    QWORD PTR [rbp-0xc0],rax
112    79fb:   8b 85 3c ff ff ff       mov    eax,DWORD PTR [rbp-0xc4]
113    7a01:   89 45 d0                mov    DWORD PTR [rbp-0x30],eax
114    7a04:   48 8b 45 88             mov    rax,QWORD PTR [rbp-0x78]
115    7a08:   48 8b 9d 78 ff ff ff    mov    rbx,QWORD PTR [rbp-0x88]
116    7a0f:   4c 8d 04 98             lea    r8,[rax+rbx*4]
117    7a13:   48 8b 45 28             mov    rax,QWORD PTR [rbp+0x28]
118    7a17:   48 8b 5d 18             mov    rbx,QWORD PTR [rbp+0x18]
119    7a1b:   48 89 45 b8             mov    QWORD PTR [rbp-0x48],rax
120    7a1f:   48 63 45 84             movsxd rax,DWORD PTR [rbp-0x7c]
121    7a23:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
122    7a2a:   48 0f af 85 50 ff ff ff     imul   rax,QWORD PTR [rbp-0xb0]
123    7a32:   48 03 85 40 ff ff ff    add    rax,QWORD PTR [rbp-0xc0]
124    7a39:   48 8d 04 83             lea    rax,[rbx+rax*4]
125    7a3d:   48 89 45 a0             mov    QWORD PTR [rbp-0x60],rax
126    7a41:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
127    7a4c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
128    7a50:   8b 45 a8                mov    eax,DWORD PTR [rbp-0x58]
129    7a53:   41 c7 00 00 00 00 00    mov    DWORD PTR [r8],0x0
130    7a5a:   45 31 db                xor    r11d,r11d
131    7a5d:   48 8b 5d a0             mov    rbx,QWORD PTR [rbp-0x60]
132    7a61:   44 8b 4d ac             mov    r9d,DWORD PTR [rbp-0x54]
133    7a65:   85 c0                   test   eax,eax
134    7a67:   0f 8e 98 00 00 00       jle    7b05 <convolution2D+0x285>
135    7a6d:   0f 1f 00                nop    DWORD PTR [rax]
136    7a70:   45 85 c9                test   r9d,r9d
137    7a73:   78 7b                   js     7af0 <convolution2D+0x270>
138    7a75:   45 39 ce                cmp    r14d,r9d
139    7a78:   7e 76                   jle    7af0 <convolution2D+0x270>
140    7a7a:   48 8b 45 b8             mov    rax,QWORD PTR [rbp-0x48]
141    7a7e:   8b 55 d0                mov    edx,DWORD PTR [rbp-0x30]
142    7a81:   48 89 de                mov    rsi,rbx
143    7a84:   4a 8d 3c 98             lea    rdi,[rax+r11*4]
144    7a88:   eb 13                   jmp    7a9d <convolution2D+0x21d>
145    7a8a:   66 0f 1f 44 00 00       nop    WORD PTR [rax+rax*1+0x0]
146    7a90:   ff c2                   inc    edx
147    7a92:   4c 01 e7                add    rdi,r12
148    7a95:   4c 01 e6                add    rsi,r12
149    7a98:   44 39 d2                cmp    edx,r10d
150    7a9b:   74 53                   je     7af0 <convolution2D+0x270>
151    7a9d:   85 d2                   test   edx,edx
152    7a9f:   78 ef                   js     7a90 <convolution2D+0x210>
153    7aa1:   41 39 d6                cmp    r14d,edx
154    7aa4:   7e ea                   jle    7a90 <convolution2D+0x210>
155    7aa6:   45 85 ed                test   r13d,r13d
156    7aa9:   7e e5                   jle    7a90 <convolution2D+0x210>
157    7aab:   c4 c1 7a 10 08          vmovss xmm1,DWORD PTR [r8]
158    7ab0:   31 c0                   xor    eax,eax
159    7ab2:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
160    7abd:   0f 1f 00                nop    DWORD PTR [rax]
161    7ac0:   c5 fa 10 04 87          vmovss xmm0,DWORD PTR [rdi+rax*4]
162    7ac5:   48 89 c1                mov    rcx,rax
163    7ac8:   c5 fa 59 04 86          vmulss xmm0,xmm0,DWORD PTR [rsi+rax*4]
164    7acd:   48 ff c0                inc    rax
165    7ad0:   c5 f2 58 c8             vaddss xmm1,xmm1,xmm0
166    7ad4:   c4 c1 7a 11 08          vmovss DWORD PTR [r8],xmm1
167    7ad9:   49 39 cf                cmp    r15,rcx
168    7adc:   75 e2                   jne    7ac0 <convolution2D+0x240>
169    7ade:   ff c2                   inc    edx
170    7ae0:   4c 01 e7                add    rdi,r12
171    7ae3:   4c 01 e6                add    rsi,r12
172    7ae6:   44 39 d2                cmp    edx,r10d
173    7ae9:   75 b2                   jne    7a9d <convolution2D+0x21d>
174    7aeb:   0f 1f 44 00 00          nop    DWORD PTR [rax+rax*1+0x0]
175    7af0:   4c 03 5d c8             add    r11,QWORD PTR [rbp-0x38]
176    7af4:   48 03 5d c0             add    rbx,QWORD PTR [rbp-0x40]
177    7af8:   41 ff c1                inc    r9d
178    7afb:   44 3b 4d d4             cmp    r9d,DWORD PTR [rbp-0x2c]
179    7aff:   0f 85 6b ff ff ff       jne    7a70 <convolution2D+0x1f0>
180    7b05:   48 8b 5d 98             mov    rbx,QWORD PTR [rbp-0x68]
181    7b09:   49 83 c0 04             add    r8,0x4
182    7b0d:   48 01 5d b8             add    QWORD PTR [rbp-0x48],rbx
183    7b11:   4c 3b 45 b0             cmp    r8,QWORD PTR [rbp-0x50]
184    7b15:   0f 85 35 ff ff ff       jne    7a50 <convolution2D+0x1d0>
185    7b1b:   8b 9d 74 ff ff ff       mov    ebx,DWORD PTR [rbp-0x8c]
186    7b21:   8b 45 94                mov    eax,DWORD PTR [rbp-0x6c]
187    7b24:   48 8b 8d 48 ff ff ff    mov    rcx,QWORD PTR [rbp-0xb8]
188    7b2b:   01 5d d0                add    DWORD PTR [rbp-0x30],ebx
189    7b2e:   48 01 4d b0             add    QWORD PTR [rbp-0x50],rcx
190    7b32:   01 5d 84                add    DWORD PTR [rbp-0x7c],ebx
191    7b35:   48 8b 8d 58 ff ff ff    mov    rcx,QWORD PTR [rbp-0xa8]
192    7b3c:   41 01 da                add    r10d,ebx
193    7b3f:   48 01 8d 78 ff ff ff    add    QWORD PTR [rbp-0x88],rcx
194    7b46:   ff c0                   inc    eax
195    7b48:   39 85 38 ff ff ff       cmp    DWORD PTR [rbp-0xc8],eax
196    7b4e:   74 08                   je     7b58 <convolution2D+0x2d8>
197    7b50:   89 45 94                mov    DWORD PTR [rbp-0x6c],eax
198    7b53:   e9 ac fe ff ff          jmp    7a04 <convolution2D+0x184>
199    7b58:   8b 4d 90                mov    ecx,DWORD PTR [rbp-0x70]
200    7b5b:   48 8b b5 30 ff ff ff    mov    rsi,QWORD PTR [rbp-0xd0]
201    7b62:   01 5d d4                add    DWORD PTR [rbp-0x2c],ebx
202    7b65:   01 5d ac                add    DWORD PTR [rbp-0x54],ebx
203    7b68:   01 5d 80                add    DWORD PTR [rbp-0x80],ebx
204    7b6b:   48 01 75 88             add    QWORD PTR [rbp-0x78],rsi
205    7b6f:   8d 41 01                lea    eax,[rcx+0x1]
206    7b72:   39 4d 94                cmp    DWORD PTR [rbp-0x6c],ecx
207    7b75:   74 08                   je     7b7f <convolution2D+0x2ff>
208    7b77:   89 45 90                mov    DWORD PTR [rbp-0x70],eax
209    7b7a:   e9 33 fe ff ff          jmp    79b2 <convolution2D+0x132>
210    7b7f:   48 81 c4 b0 00 00 00    add    rsp,0xb0
211    7b86:   5b                      pop    rbx
212    7b87:   41 5c                   pop    r12
213    7b89:   41 5d                   pop    r13
214    7b8b:   41 5e                   pop    r14
215    7b8d:   41 5f                   pop    r15
216    7b8f:   5d                      pop    rbp
217    7b90:   c3                      ret    
218    7b91:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
219    7b9c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
220
221float (*arr)[x][y] = calloc(z, sizeof *arr);
222==430300== Cachegrind, a cache and branch-prediction profiler
223==430300== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
224==430300== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
225==430300== Command: ./EmbeddedNet test 1
226==430300== Parent PID: 170008
227==430300== 
228--430300-- warning: L3 cache found, using its data for the LL simulation.
229==430300== 
230==430300== I   refs:      6,369,594,192
231==430300== I1  misses:            4,271
232==430300== LLi misses:            2,442
233==430300== I1  miss rate:          0.00%
234==430300== LLi miss rate:          0.00%
235==430300== 
236==430300== D   refs:      2,064,233,110  (1,359,003,131 rd   + 705,229,979 wr)
237==430300== D1  misses:       34,476,969  (   19,010,839 rd   +  15,466,130 wr)
238==430300== LLd misses:        5,311,277  (    1,603,955 rd   +   3,707,322 wr)
239==430300== D1  miss rate:           1.7% (          1.4%     +         2.2%  )
240==430300== LLd miss rate:           0.3% (          0.1%     +         0.5%  )
241==430300== 
242==430300== LL refs:          34,481,240  (   19,015,110 rd   +  15,466,130 wr)
243==430300== LL misses:         5,313,719  (    1,606,397 rd   +   3,707,322 wr)
244==430300== LL miss rate:            0.1% (          0.0%     +         0.5%  )
245float fma_scl(float a, float b, float c) {
246    return a * b + c;
247}
248
249fma_scl:
250        vfmadd132ss     xmm0, xmm2, xmm1
251        ret
252typedef float Vec __attribute__((vector_size(32), aligned(32)));
253

1void convolution2D(int isize,   // width/height of input
2        int osize,              // width/height of output
3        int ksize,              // width/height of kernel
4        int stride,             // shift between input pixels, between consecutive outputs
5        int pad,                // offset between (0,0) pixels between input and output
6        int idepth, int odepth, // number of input and output channels
7        float idata[isize][isize][idepth],
8        float odata[osize][osize][odepth],
9        float kdata[odepth][ksize][ksize][idepth])
10{
11    // iterate over the output
12    for (int oy = 0; oy < osize; ++oy) {
13    for (int ox = 0; ox < osize; ++ox) {
14    for (int od = 0; od < odepth; ++od) {
15        odata[oy][ox][od] = 0;  // When you iterate multiple times without closing the program, this number would stack up to infinity, so we have to zero it out every time.
16        for (int ky = 0; ky < ksize; ++ky) {
17        for (int kx = 0; kx < ksize; ++kx) {
18            // map position in output and kernel to the input
19            int iy = stride * oy + ky - pad;
20            int ix = stride * ox + kx - pad;
21            // use only valid inputs
22            if (iy >= 0 && iy < isize && ix >= 0 && ix < isize)
23                for (int id = 0; id < idepth; ++id)
24                    odata[oy][ox][od] += kdata[od][ky][kx][id] * idata[iy][ix][id];
25        }}
26    }}}
27
28}
29
300000000000007880 <convolution2D>:
31    7880:   f3 0f 1e fa             endbr64 
32    7884:   55                      push   rbp
33    7885:   48 89 e5                mov    rbp,rsp
34    7888:   41 57                   push   r15
35    788a:   41 56                   push   r14
36    788c:   41 55                   push   r13
37    788e:   41 54                   push   r12
38    7890:   53                      push   rbx
39    7891:   48 81 ec b0 00 00 00    sub    rsp,0xb0
40    7898:   ff 15 4a a7 00 00       call   QWORD PTR [rip+0xa74a]        # 11fe8 <mcount@GLIBC_2.2.5>
41    789e:   89 d3                   mov    ebx,edx
42    78a0:   89 55 a8                mov    DWORD PTR [rbp-0x58],edx
43    78a3:   89 8d 74 ff ff ff       mov    DWORD PTR [rbp-0x8c],ecx
44    78a9:   49 63 d1                movsxd rdx,r9d
45    78ac:   48 63 cf                movsxd rcx,edi
46    78af:   41 89 f2                mov    r10d,esi
47    78b2:   89 b5 38 ff ff ff       mov    DWORD PTR [rbp-0xc8],esi
48    78b8:   49 63 c0                movsxd rax,r8d
49    78bb:   48 0f af ca             imul   rcx,rdx
50    78bf:   48 63 75 10             movsxd rsi,DWORD PTR [rbp+0x10]
51    78c3:   49 89 d6                mov    r14,rdx
52    78c6:   4c 8d 24 95 00 00 00 00     lea    r12,[rdx*4+0x0]
53    78ce:   41 89 fd                mov    r13d,edi
54    78d1:   49 89 cb                mov    r11,rcx
55    78d4:   48 89 8d 60 ff ff ff    mov    QWORD PTR [rbp-0xa0],rcx
56    78db:   49 63 ca                movsxd rcx,r10d
57    78de:   4c 8d 0c b5 00 00 00 00     lea    r9,[rsi*4+0x0]
58    78e6:   49 89 f0                mov    r8,rsi
59    78e9:   48 0f af f1             imul   rsi,rcx
60    78ed:   48 63 cb                movsxd rcx,ebx
61    78f0:   4c 89 8d 48 ff ff ff    mov    QWORD PTR [rbp-0xb8],r9
62    78f7:   48 0f af d1             imul   rdx,rcx
63    78fb:   48 8d 3c 95 00 00 00 00     lea    rdi,[rdx*4+0x0]
64    7903:   45 85 d2                test   r10d,r10d
65    7906:   0f 8e 73 02 00 00       jle    7b7f <convolution2D+0x2ff>
66    790c:   48 c1 ef 02             shr    rdi,0x2
67    7910:   49 c1 e9 02             shr    r9,0x2
68    7914:   48 89 7d c8             mov    QWORD PTR [rbp-0x38],rdi
69    7918:   4c 89 e7                mov    rdi,r12
70    791b:   4c 89 8d 58 ff ff ff    mov    QWORD PTR [rbp-0xa8],r9
71    7922:   48 c1 ef 02             shr    rdi,0x2
72    7926:   48 89 bd 50 ff ff ff    mov    QWORD PTR [rbp-0xb0],rdi
73    792d:   45 85 c0                test   r8d,r8d
74    7930:   0f 8e 49 02 00 00       jle    7b7f <convolution2D+0x2ff>
75    7936:   48 c1 e6 02             shl    rsi,0x2
76    793a:   48 0f af d1             imul   rdx,rcx
77    793e:   29 c3                   sub    ebx,eax
78    7940:   89 c7                   mov    edi,eax
79    7942:   48 89 b5 30 ff ff ff    mov    QWORD PTR [rbp-0xd0],rsi
80    7949:   48 8b 75 20             mov    rsi,QWORD PTR [rbp+0x20]
81    794d:   48 89 85 68 ff ff ff    mov    QWORD PTR [rbp-0x98],rax
82    7954:   f7 df                   neg    edi
83    7956:   45 8d 7e ff             lea    r15d,[r14-0x1]
84    795a:   89 9d 70 ff ff ff       mov    DWORD PTR [rbp-0x90],ebx
85    7960:   89 bd 3c ff ff ff       mov    DWORD PTR [rbp-0xc4],edi
86    7966:   48 8d 0c 95 00 00 00 00     lea    rcx,[rdx*4+0x0]
87    796e:   89 7d ac                mov    DWORD PTR [rbp-0x54],edi
88    7971:   89 5d d4                mov    DWORD PTR [rbp-0x2c],ebx
89    7974:   48 89 4d 98             mov    QWORD PTR [rbp-0x68],rcx
90    7978:   4a 8d 0c 9d 00 00 00 00     lea    rcx,[r11*4+0x0]
91    7980:   c7 45 80 00 00 00 00    mov    DWORD PTR [rbp-0x80],0x0
92    7987:   48 89 75 88             mov    QWORD PTR [rbp-0x78],rsi
93    798b:   41 8d 70 ff             lea    esi,[r8-0x1]
94    798f:   48 89 4d c0             mov    QWORD PTR [rbp-0x40],rcx
95    7993:   48 8d 04 b5 04 00 00 00     lea    rax,[rsi*4+0x4]
96    799b:   c7 45 90 00 00 00 00    mov    DWORD PTR [rbp-0x70],0x0
97    79a2:   48 89 85 28 ff ff ff    mov    QWORD PTR [rbp-0xd8],rax
98    79a9:   44 89 f0                mov    eax,r14d
99    79ac:   45 89 ee                mov    r14d,r13d
100    79af:   41 89 c5                mov    r13d,eax
101    79b2:   48 8b 85 28 ff ff ff    mov    rax,QWORD PTR [rbp-0xd8]
102    79b9:   48 03 45 88             add    rax,QWORD PTR [rbp-0x78]
103    79bd:   48 c7 85 78 ff ff ff 00 00 00 00    mov    QWORD PTR [rbp-0x88],0x0
104    79c8:   c7 45 84 00 00 00 00    mov    DWORD PTR [rbp-0x7c],0x0
105    79cf:   c7 45 94 00 00 00 00    mov    DWORD PTR [rbp-0x6c],0x0
106    79d6:   44 8b 95 70 ff ff ff    mov    r10d,DWORD PTR [rbp-0x90]
107    79dd:   48 89 45 b0             mov    QWORD PTR [rbp-0x50],rax
108    79e1:   48 63 45 80             movsxd rax,DWORD PTR [rbp-0x80]
109    79e5:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
110    79ec:   48 0f af 85 60 ff ff ff     imul   rax,QWORD PTR [rbp-0xa0]
111    79f4:   48 89 85 40 ff ff ff    mov    QWORD PTR [rbp-0xc0],rax
112    79fb:   8b 85 3c ff ff ff       mov    eax,DWORD PTR [rbp-0xc4]
113    7a01:   89 45 d0                mov    DWORD PTR [rbp-0x30],eax
114    7a04:   48 8b 45 88             mov    rax,QWORD PTR [rbp-0x78]
115    7a08:   48 8b 9d 78 ff ff ff    mov    rbx,QWORD PTR [rbp-0x88]
116    7a0f:   4c 8d 04 98             lea    r8,[rax+rbx*4]
117    7a13:   48 8b 45 28             mov    rax,QWORD PTR [rbp+0x28]
118    7a17:   48 8b 5d 18             mov    rbx,QWORD PTR [rbp+0x18]
119    7a1b:   48 89 45 b8             mov    QWORD PTR [rbp-0x48],rax
120    7a1f:   48 63 45 84             movsxd rax,DWORD PTR [rbp-0x7c]
121    7a23:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
122    7a2a:   48 0f af 85 50 ff ff ff     imul   rax,QWORD PTR [rbp-0xb0]
123    7a32:   48 03 85 40 ff ff ff    add    rax,QWORD PTR [rbp-0xc0]
124    7a39:   48 8d 04 83             lea    rax,[rbx+rax*4]
125    7a3d:   48 89 45 a0             mov    QWORD PTR [rbp-0x60],rax
126    7a41:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
127    7a4c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
128    7a50:   8b 45 a8                mov    eax,DWORD PTR [rbp-0x58]
129    7a53:   41 c7 00 00 00 00 00    mov    DWORD PTR [r8],0x0
130    7a5a:   45 31 db                xor    r11d,r11d
131    7a5d:   48 8b 5d a0             mov    rbx,QWORD PTR [rbp-0x60]
132    7a61:   44 8b 4d ac             mov    r9d,DWORD PTR [rbp-0x54]
133    7a65:   85 c0                   test   eax,eax
134    7a67:   0f 8e 98 00 00 00       jle    7b05 <convolution2D+0x285>
135    7a6d:   0f 1f 00                nop    DWORD PTR [rax]
136    7a70:   45 85 c9                test   r9d,r9d
137    7a73:   78 7b                   js     7af0 <convolution2D+0x270>
138    7a75:   45 39 ce                cmp    r14d,r9d
139    7a78:   7e 76                   jle    7af0 <convolution2D+0x270>
140    7a7a:   48 8b 45 b8             mov    rax,QWORD PTR [rbp-0x48]
141    7a7e:   8b 55 d0                mov    edx,DWORD PTR [rbp-0x30]
142    7a81:   48 89 de                mov    rsi,rbx
143    7a84:   4a 8d 3c 98             lea    rdi,[rax+r11*4]
144    7a88:   eb 13                   jmp    7a9d <convolution2D+0x21d>
145    7a8a:   66 0f 1f 44 00 00       nop    WORD PTR [rax+rax*1+0x0]
146    7a90:   ff c2                   inc    edx
147    7a92:   4c 01 e7                add    rdi,r12
148    7a95:   4c 01 e6                add    rsi,r12
149    7a98:   44 39 d2                cmp    edx,r10d
150    7a9b:   74 53                   je     7af0 <convolution2D+0x270>
151    7a9d:   85 d2                   test   edx,edx
152    7a9f:   78 ef                   js     7a90 <convolution2D+0x210>
153    7aa1:   41 39 d6                cmp    r14d,edx
154    7aa4:   7e ea                   jle    7a90 <convolution2D+0x210>
155    7aa6:   45 85 ed                test   r13d,r13d
156    7aa9:   7e e5                   jle    7a90 <convolution2D+0x210>
157    7aab:   c4 c1 7a 10 08          vmovss xmm1,DWORD PTR [r8]
158    7ab0:   31 c0                   xor    eax,eax
159    7ab2:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
160    7abd:   0f 1f 00                nop    DWORD PTR [rax]
161    7ac0:   c5 fa 10 04 87          vmovss xmm0,DWORD PTR [rdi+rax*4]
162    7ac5:   48 89 c1                mov    rcx,rax
163    7ac8:   c5 fa 59 04 86          vmulss xmm0,xmm0,DWORD PTR [rsi+rax*4]
164    7acd:   48 ff c0                inc    rax
165    7ad0:   c5 f2 58 c8             vaddss xmm1,xmm1,xmm0
166    7ad4:   c4 c1 7a 11 08          vmovss DWORD PTR [r8],xmm1
167    7ad9:   49 39 cf                cmp    r15,rcx
168    7adc:   75 e2                   jne    7ac0 <convolution2D+0x240>
169    7ade:   ff c2                   inc    edx
170    7ae0:   4c 01 e7                add    rdi,r12
171    7ae3:   4c 01 e6                add    rsi,r12
172    7ae6:   44 39 d2                cmp    edx,r10d
173    7ae9:   75 b2                   jne    7a9d <convolution2D+0x21d>
174    7aeb:   0f 1f 44 00 00          nop    DWORD PTR [rax+rax*1+0x0]
175    7af0:   4c 03 5d c8             add    r11,QWORD PTR [rbp-0x38]
176    7af4:   48 03 5d c0             add    rbx,QWORD PTR [rbp-0x40]
177    7af8:   41 ff c1                inc    r9d
178    7afb:   44 3b 4d d4             cmp    r9d,DWORD PTR [rbp-0x2c]
179    7aff:   0f 85 6b ff ff ff       jne    7a70 <convolution2D+0x1f0>
180    7b05:   48 8b 5d 98             mov    rbx,QWORD PTR [rbp-0x68]
181    7b09:   49 83 c0 04             add    r8,0x4
182    7b0d:   48 01 5d b8             add    QWORD PTR [rbp-0x48],rbx
183    7b11:   4c 3b 45 b0             cmp    r8,QWORD PTR [rbp-0x50]
184    7b15:   0f 85 35 ff ff ff       jne    7a50 <convolution2D+0x1d0>
185    7b1b:   8b 9d 74 ff ff ff       mov    ebx,DWORD PTR [rbp-0x8c]
186    7b21:   8b 45 94                mov    eax,DWORD PTR [rbp-0x6c]
187    7b24:   48 8b 8d 48 ff ff ff    mov    rcx,QWORD PTR [rbp-0xb8]
188    7b2b:   01 5d d0                add    DWORD PTR [rbp-0x30],ebx
189    7b2e:   48 01 4d b0             add    QWORD PTR [rbp-0x50],rcx
190    7b32:   01 5d 84                add    DWORD PTR [rbp-0x7c],ebx
191    7b35:   48 8b 8d 58 ff ff ff    mov    rcx,QWORD PTR [rbp-0xa8]
192    7b3c:   41 01 da                add    r10d,ebx
193    7b3f:   48 01 8d 78 ff ff ff    add    QWORD PTR [rbp-0x88],rcx
194    7b46:   ff c0                   inc    eax
195    7b48:   39 85 38 ff ff ff       cmp    DWORD PTR [rbp-0xc8],eax
196    7b4e:   74 08                   je     7b58 <convolution2D+0x2d8>
197    7b50:   89 45 94                mov    DWORD PTR [rbp-0x6c],eax
198    7b53:   e9 ac fe ff ff          jmp    7a04 <convolution2D+0x184>
199    7b58:   8b 4d 90                mov    ecx,DWORD PTR [rbp-0x70]
200    7b5b:   48 8b b5 30 ff ff ff    mov    rsi,QWORD PTR [rbp-0xd0]
201    7b62:   01 5d d4                add    DWORD PTR [rbp-0x2c],ebx
202    7b65:   01 5d ac                add    DWORD PTR [rbp-0x54],ebx
203    7b68:   01 5d 80                add    DWORD PTR [rbp-0x80],ebx
204    7b6b:   48 01 75 88             add    QWORD PTR [rbp-0x78],rsi
205    7b6f:   8d 41 01                lea    eax,[rcx+0x1]
206    7b72:   39 4d 94                cmp    DWORD PTR [rbp-0x6c],ecx
207    7b75:   74 08                   je     7b7f <convolution2D+0x2ff>
208    7b77:   89 45 90                mov    DWORD PTR [rbp-0x70],eax
209    7b7a:   e9 33 fe ff ff          jmp    79b2 <convolution2D+0x132>
210    7b7f:   48 81 c4 b0 00 00 00    add    rsp,0xb0
211    7b86:   5b                      pop    rbx
212    7b87:   41 5c                   pop    r12
213    7b89:   41 5d                   pop    r13
214    7b8b:   41 5e                   pop    r14
215    7b8d:   41 5f                   pop    r15
216    7b8f:   5d                      pop    rbp
217    7b90:   c3                      ret    
218    7b91:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
219    7b9c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
220
221float (*arr)[x][y] = calloc(z, sizeof *arr);
222==430300== Cachegrind, a cache and branch-prediction profiler
223==430300== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
224==430300== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
225==430300== Command: ./EmbeddedNet test 1
226==430300== Parent PID: 170008
227==430300== 
228--430300-- warning: L3 cache found, using its data for the LL simulation.
229==430300== 
230==430300== I   refs:      6,369,594,192
231==430300== I1  misses:            4,271
232==430300== LLi misses:            2,442
233==430300== I1  miss rate:          0.00%
234==430300== LLi miss rate:          0.00%
235==430300== 
236==430300== D   refs:      2,064,233,110  (1,359,003,131 rd   + 705,229,979 wr)
237==430300== D1  misses:       34,476,969  (   19,010,839 rd   +  15,466,130 wr)
238==430300== LLd misses:        5,311,277  (    1,603,955 rd   +   3,707,322 wr)
239==430300== D1  miss rate:           1.7% (          1.4%     +         2.2%  )
240==430300== LLd miss rate:           0.3% (          0.1%     +         0.5%  )
241==430300== 
242==430300== LL refs:          34,481,240  (   19,015,110 rd   +  15,466,130 wr)
243==430300== LL misses:         5,313,719  (    1,606,397 rd   +   3,707,322 wr)
244==430300== LL miss rate:            0.1% (          0.0%     +         0.5%  )
245float fma_scl(float a, float b, float c) {
246    return a * b + c;
247}
248
249fma_scl:
250        vfmadd132ss     xmm0, xmm2, xmm1
251        ret
252typedef float Vec __attribute__((vector_size(32), aligned(32)));
253Vec fma_vec(Vec a, Vec b, Vec c) {
254    return a * b + c;
255}
256
257fma_vec:
258        vfmadd132ps     ymm0, ymm2, ymm1
259        ret
260

1void convolution2D(int isize,   // width/height of input
2        int osize,              // width/height of output
3        int ksize,              // width/height of kernel
4        int stride,             // shift between input pixels, between consecutive outputs
5        int pad,                // offset between (0,0) pixels between input and output
6        int idepth, int odepth, // number of input and output channels
7        float idata[isize][isize][idepth],
8        float odata[osize][osize][odepth],
9        float kdata[odepth][ksize][ksize][idepth])
10{
11    // iterate over the output
12    for (int oy = 0; oy < osize; ++oy) {
13    for (int ox = 0; ox < osize; ++ox) {
14    for (int od = 0; od < odepth; ++od) {
15        odata[oy][ox][od] = 0;  // When you iterate multiple times without closing the program, this number would stack up to infinity, so we have to zero it out every time.
16        for (int ky = 0; ky < ksize; ++ky) {
17        for (int kx = 0; kx < ksize; ++kx) {
18            // map position in output and kernel to the input
19            int iy = stride * oy + ky - pad;
20            int ix = stride * ox + kx - pad;
21            // use only valid inputs
22            if (iy >= 0 && iy < isize && ix >= 0 && ix < isize)
23                for (int id = 0; id < idepth; ++id)
24                    odata[oy][ox][od] += kdata[od][ky][kx][id] * idata[iy][ix][id];
25        }}
26    }}}
27
28}
29
300000000000007880 <convolution2D>:
31    7880:   f3 0f 1e fa             endbr64 
32    7884:   55                      push   rbp
33    7885:   48 89 e5                mov    rbp,rsp
34    7888:   41 57                   push   r15
35    788a:   41 56                   push   r14
36    788c:   41 55                   push   r13
37    788e:   41 54                   push   r12
38    7890:   53                      push   rbx
39    7891:   48 81 ec b0 00 00 00    sub    rsp,0xb0
40    7898:   ff 15 4a a7 00 00       call   QWORD PTR [rip+0xa74a]        # 11fe8 <mcount@GLIBC_2.2.5>
41    789e:   89 d3                   mov    ebx,edx
42    78a0:   89 55 a8                mov    DWORD PTR [rbp-0x58],edx
43    78a3:   89 8d 74 ff ff ff       mov    DWORD PTR [rbp-0x8c],ecx
44    78a9:   49 63 d1                movsxd rdx,r9d
45    78ac:   48 63 cf                movsxd rcx,edi
46    78af:   41 89 f2                mov    r10d,esi
47    78b2:   89 b5 38 ff ff ff       mov    DWORD PTR [rbp-0xc8],esi
48    78b8:   49 63 c0                movsxd rax,r8d
49    78bb:   48 0f af ca             imul   rcx,rdx
50    78bf:   48 63 75 10             movsxd rsi,DWORD PTR [rbp+0x10]
51    78c3:   49 89 d6                mov    r14,rdx
52    78c6:   4c 8d 24 95 00 00 00 00     lea    r12,[rdx*4+0x0]
53    78ce:   41 89 fd                mov    r13d,edi
54    78d1:   49 89 cb                mov    r11,rcx
55    78d4:   48 89 8d 60 ff ff ff    mov    QWORD PTR [rbp-0xa0],rcx
56    78db:   49 63 ca                movsxd rcx,r10d
57    78de:   4c 8d 0c b5 00 00 00 00     lea    r9,[rsi*4+0x0]
58    78e6:   49 89 f0                mov    r8,rsi
59    78e9:   48 0f af f1             imul   rsi,rcx
60    78ed:   48 63 cb                movsxd rcx,ebx
61    78f0:   4c 89 8d 48 ff ff ff    mov    QWORD PTR [rbp-0xb8],r9
62    78f7:   48 0f af d1             imul   rdx,rcx
63    78fb:   48 8d 3c 95 00 00 00 00     lea    rdi,[rdx*4+0x0]
64    7903:   45 85 d2                test   r10d,r10d
65    7906:   0f 8e 73 02 00 00       jle    7b7f <convolution2D+0x2ff>
66    790c:   48 c1 ef 02             shr    rdi,0x2
67    7910:   49 c1 e9 02             shr    r9,0x2
68    7914:   48 89 7d c8             mov    QWORD PTR [rbp-0x38],rdi
69    7918:   4c 89 e7                mov    rdi,r12
70    791b:   4c 89 8d 58 ff ff ff    mov    QWORD PTR [rbp-0xa8],r9
71    7922:   48 c1 ef 02             shr    rdi,0x2
72    7926:   48 89 bd 50 ff ff ff    mov    QWORD PTR [rbp-0xb0],rdi
73    792d:   45 85 c0                test   r8d,r8d
74    7930:   0f 8e 49 02 00 00       jle    7b7f <convolution2D+0x2ff>
75    7936:   48 c1 e6 02             shl    rsi,0x2
76    793a:   48 0f af d1             imul   rdx,rcx
77    793e:   29 c3                   sub    ebx,eax
78    7940:   89 c7                   mov    edi,eax
79    7942:   48 89 b5 30 ff ff ff    mov    QWORD PTR [rbp-0xd0],rsi
80    7949:   48 8b 75 20             mov    rsi,QWORD PTR [rbp+0x20]
81    794d:   48 89 85 68 ff ff ff    mov    QWORD PTR [rbp-0x98],rax
82    7954:   f7 df                   neg    edi
83    7956:   45 8d 7e ff             lea    r15d,[r14-0x1]
84    795a:   89 9d 70 ff ff ff       mov    DWORD PTR [rbp-0x90],ebx
85    7960:   89 bd 3c ff ff ff       mov    DWORD PTR [rbp-0xc4],edi
86    7966:   48 8d 0c 95 00 00 00 00     lea    rcx,[rdx*4+0x0]
87    796e:   89 7d ac                mov    DWORD PTR [rbp-0x54],edi
88    7971:   89 5d d4                mov    DWORD PTR [rbp-0x2c],ebx
89    7974:   48 89 4d 98             mov    QWORD PTR [rbp-0x68],rcx
90    7978:   4a 8d 0c 9d 00 00 00 00     lea    rcx,[r11*4+0x0]
91    7980:   c7 45 80 00 00 00 00    mov    DWORD PTR [rbp-0x80],0x0
92    7987:   48 89 75 88             mov    QWORD PTR [rbp-0x78],rsi
93    798b:   41 8d 70 ff             lea    esi,[r8-0x1]
94    798f:   48 89 4d c0             mov    QWORD PTR [rbp-0x40],rcx
95    7993:   48 8d 04 b5 04 00 00 00     lea    rax,[rsi*4+0x4]
96    799b:   c7 45 90 00 00 00 00    mov    DWORD PTR [rbp-0x70],0x0
97    79a2:   48 89 85 28 ff ff ff    mov    QWORD PTR [rbp-0xd8],rax
98    79a9:   44 89 f0                mov    eax,r14d
99    79ac:   45 89 ee                mov    r14d,r13d
100    79af:   41 89 c5                mov    r13d,eax
101    79b2:   48 8b 85 28 ff ff ff    mov    rax,QWORD PTR [rbp-0xd8]
102    79b9:   48 03 45 88             add    rax,QWORD PTR [rbp-0x78]
103    79bd:   48 c7 85 78 ff ff ff 00 00 00 00    mov    QWORD PTR [rbp-0x88],0x0
104    79c8:   c7 45 84 00 00 00 00    mov    DWORD PTR [rbp-0x7c],0x0
105    79cf:   c7 45 94 00 00 00 00    mov    DWORD PTR [rbp-0x6c],0x0
106    79d6:   44 8b 95 70 ff ff ff    mov    r10d,DWORD PTR [rbp-0x90]
107    79dd:   48 89 45 b0             mov    QWORD PTR [rbp-0x50],rax
108    79e1:   48 63 45 80             movsxd rax,DWORD PTR [rbp-0x80]
109    79e5:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
110    79ec:   48 0f af 85 60 ff ff ff     imul   rax,QWORD PTR [rbp-0xa0]
111    79f4:   48 89 85 40 ff ff ff    mov    QWORD PTR [rbp-0xc0],rax
112    79fb:   8b 85 3c ff ff ff       mov    eax,DWORD PTR [rbp-0xc4]
113    7a01:   89 45 d0                mov    DWORD PTR [rbp-0x30],eax
114    7a04:   48 8b 45 88             mov    rax,QWORD PTR [rbp-0x78]
115    7a08:   48 8b 9d 78 ff ff ff    mov    rbx,QWORD PTR [rbp-0x88]
116    7a0f:   4c 8d 04 98             lea    r8,[rax+rbx*4]
117    7a13:   48 8b 45 28             mov    rax,QWORD PTR [rbp+0x28]
118    7a17:   48 8b 5d 18             mov    rbx,QWORD PTR [rbp+0x18]
119    7a1b:   48 89 45 b8             mov    QWORD PTR [rbp-0x48],rax
120    7a1f:   48 63 45 84             movsxd rax,DWORD PTR [rbp-0x7c]
121    7a23:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
122    7a2a:   48 0f af 85 50 ff ff ff     imul   rax,QWORD PTR [rbp-0xb0]
123    7a32:   48 03 85 40 ff ff ff    add    rax,QWORD PTR [rbp-0xc0]
124    7a39:   48 8d 04 83             lea    rax,[rbx+rax*4]
125    7a3d:   48 89 45 a0             mov    QWORD PTR [rbp-0x60],rax
126    7a41:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
127    7a4c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
128    7a50:   8b 45 a8                mov    eax,DWORD PTR [rbp-0x58]
129    7a53:   41 c7 00 00 00 00 00    mov    DWORD PTR [r8],0x0
130    7a5a:   45 31 db                xor    r11d,r11d
131    7a5d:   48 8b 5d a0             mov    rbx,QWORD PTR [rbp-0x60]
132    7a61:   44 8b 4d ac             mov    r9d,DWORD PTR [rbp-0x54]
133    7a65:   85 c0                   test   eax,eax
134    7a67:   0f 8e 98 00 00 00       jle    7b05 <convolution2D+0x285>
135    7a6d:   0f 1f 00                nop    DWORD PTR [rax]
136    7a70:   45 85 c9                test   r9d,r9d
137    7a73:   78 7b                   js     7af0 <convolution2D+0x270>
138    7a75:   45 39 ce                cmp    r14d,r9d
139    7a78:   7e 76                   jle    7af0 <convolution2D+0x270>
140    7a7a:   48 8b 45 b8             mov    rax,QWORD PTR [rbp-0x48]
141    7a7e:   8b 55 d0                mov    edx,DWORD PTR [rbp-0x30]
142    7a81:   48 89 de                mov    rsi,rbx
143    7a84:   4a 8d 3c 98             lea    rdi,[rax+r11*4]
144    7a88:   eb 13                   jmp    7a9d <convolution2D+0x21d>
145    7a8a:   66 0f 1f 44 00 00       nop    WORD PTR [rax+rax*1+0x0]
146    7a90:   ff c2                   inc    edx
147    7a92:   4c 01 e7                add    rdi,r12
148    7a95:   4c 01 e6                add    rsi,r12
149    7a98:   44 39 d2                cmp    edx,r10d
150    7a9b:   74 53                   je     7af0 <convolution2D+0x270>
151    7a9d:   85 d2                   test   edx,edx
152    7a9f:   78 ef                   js     7a90 <convolution2D+0x210>
153    7aa1:   41 39 d6                cmp    r14d,edx
154    7aa4:   7e ea                   jle    7a90 <convolution2D+0x210>
155    7aa6:   45 85 ed                test   r13d,r13d
156    7aa9:   7e e5                   jle    7a90 <convolution2D+0x210>
157    7aab:   c4 c1 7a 10 08          vmovss xmm1,DWORD PTR [r8]
158    7ab0:   31 c0                   xor    eax,eax
159    7ab2:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
160    7abd:   0f 1f 00                nop    DWORD PTR [rax]
161    7ac0:   c5 fa 10 04 87          vmovss xmm0,DWORD PTR [rdi+rax*4]
162    7ac5:   48 89 c1                mov    rcx,rax
163    7ac8:   c5 fa 59 04 86          vmulss xmm0,xmm0,DWORD PTR [rsi+rax*4]
164    7acd:   48 ff c0                inc    rax
165    7ad0:   c5 f2 58 c8             vaddss xmm1,xmm1,xmm0
166    7ad4:   c4 c1 7a 11 08          vmovss DWORD PTR [r8],xmm1
167    7ad9:   49 39 cf                cmp    r15,rcx
168    7adc:   75 e2                   jne    7ac0 <convolution2D+0x240>
169    7ade:   ff c2                   inc    edx
170    7ae0:   4c 01 e7                add    rdi,r12
171    7ae3:   4c 01 e6                add    rsi,r12
172    7ae6:   44 39 d2                cmp    edx,r10d
173    7ae9:   75 b2                   jne    7a9d <convolution2D+0x21d>
174    7aeb:   0f 1f 44 00 00          nop    DWORD PTR [rax+rax*1+0x0]
175    7af0:   4c 03 5d c8             add    r11,QWORD PTR [rbp-0x38]
176    7af4:   48 03 5d c0             add    rbx,QWORD PTR [rbp-0x40]
177    7af8:   41 ff c1                inc    r9d
178    7afb:   44 3b 4d d4             cmp    r9d,DWORD PTR [rbp-0x2c]
179    7aff:   0f 85 6b ff ff ff       jne    7a70 <convolution2D+0x1f0>
180    7b05:   48 8b 5d 98             mov    rbx,QWORD PTR [rbp-0x68]
181    7b09:   49 83 c0 04             add    r8,0x4
182    7b0d:   48 01 5d b8             add    QWORD PTR [rbp-0x48],rbx
183    7b11:   4c 3b 45 b0             cmp    r8,QWORD PTR [rbp-0x50]
184    7b15:   0f 85 35 ff ff ff       jne    7a50 <convolution2D+0x1d0>
185    7b1b:   8b 9d 74 ff ff ff       mov    ebx,DWORD PTR [rbp-0x8c]
186    7b21:   8b 45 94                mov    eax,DWORD PTR [rbp-0x6c]
187    7b24:   48 8b 8d 48 ff ff ff    mov    rcx,QWORD PTR [rbp-0xb8]
188    7b2b:   01 5d d0                add    DWORD PTR [rbp-0x30],ebx
189    7b2e:   48 01 4d b0             add    QWORD PTR [rbp-0x50],rcx
190    7b32:   01 5d 84                add    DWORD PTR [rbp-0x7c],ebx
191    7b35:   48 8b 8d 58 ff ff ff    mov    rcx,QWORD PTR [rbp-0xa8]
192    7b3c:   41 01 da                add    r10d,ebx
193    7b3f:   48 01 8d 78 ff ff ff    add    QWORD PTR [rbp-0x88],rcx
194    7b46:   ff c0                   inc    eax
195    7b48:   39 85 38 ff ff ff       cmp    DWORD PTR [rbp-0xc8],eax
196    7b4e:   74 08                   je     7b58 <convolution2D+0x2d8>
197    7b50:   89 45 94                mov    DWORD PTR [rbp-0x6c],eax
198    7b53:   e9 ac fe ff ff          jmp    7a04 <convolution2D+0x184>
199    7b58:   8b 4d 90                mov    ecx,DWORD PTR [rbp-0x70]
200    7b5b:   48 8b b5 30 ff ff ff    mov    rsi,QWORD PTR [rbp-0xd0]
201    7b62:   01 5d d4                add    DWORD PTR [rbp-0x2c],ebx
202    7b65:   01 5d ac                add    DWORD PTR [rbp-0x54],ebx
203    7b68:   01 5d 80                add    DWORD PTR [rbp-0x80],ebx
204    7b6b:   48 01 75 88             add    QWORD PTR [rbp-0x78],rsi
205    7b6f:   8d 41 01                lea    eax,[rcx+0x1]
206    7b72:   39 4d 94                cmp    DWORD PTR [rbp-0x6c],ecx
207    7b75:   74 08                   je     7b7f <convolution2D+0x2ff>
208    7b77:   89 45 90                mov    DWORD PTR [rbp-0x70],eax
209    7b7a:   e9 33 fe ff ff          jmp    79b2 <convolution2D+0x132>
210    7b7f:   48 81 c4 b0 00 00 00    add    rsp,0xb0
211    7b86:   5b                      pop    rbx
212    7b87:   41 5c                   pop    r12
213    7b89:   41 5d                   pop    r13
214    7b8b:   41 5e                   pop    r14
215    7b8d:   41 5f                   pop    r15
216    7b8f:   5d                      pop    rbp
217    7b90:   c3                      ret    
218    7b91:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
219    7b9c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
220
221float (*arr)[x][y] = calloc(z, sizeof *arr);
222==430300== Cachegrind, a cache and branch-prediction profiler
223==430300== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
224==430300== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
225==430300== Command: ./EmbeddedNet test 1
226==430300== Parent PID: 170008
227==430300== 
228--430300-- warning: L3 cache found, using its data for the LL simulation.
229==430300== 
230==430300== I   refs:      6,369,594,192
231==430300== I1  misses:            4,271
232==430300== LLi misses:            2,442
233==430300== I1  miss rate:          0.00%
234==430300== LLi miss rate:          0.00%
235==430300== 
236==430300== D   refs:      2,064,233,110  (1,359,003,131 rd   + 705,229,979 wr)
237==430300== D1  misses:       34,476,969  (   19,010,839 rd   +  15,466,130 wr)
238==430300== LLd misses:        5,311,277  (    1,603,955 rd   +   3,707,322 wr)
239==430300== D1  miss rate:           1.7% (          1.4%     +         2.2%  )
240==430300== LLd miss rate:           0.3% (          0.1%     +         0.5%  )
241==430300== 
242==430300== LL refs:          34,481,240  (   19,015,110 rd   +  15,466,130 wr)
243==430300== LL misses:         5,313,719  (    1,606,397 rd   +   3,707,322 wr)
244==430300== LL miss rate:            0.1% (          0.0%     +         0.5%  )
245float fma_scl(float a, float b, float c) {
246    return a * b + c;
247}
248
249fma_scl:
250        vfmadd132ss     xmm0, xmm2, xmm1
251        ret
252typedef float Vec __attribute__((vector_size(32), aligned(32)));
253Vec fma_vec(Vec a, Vec b, Vec c) {
254    return a * b + c;
255}
256
257fma_vec:
258        vfmadd132ps     ymm0, ymm2, ymm1
259        ret
260typedef struct {
261    float f[8];
262} Vec_;
263
264Vec_ fma_vec_(Vec_ a, Vec_ b, Vec_ c) {
265    Vec_ r;
266    for (unsigned i = 0; i < 8; ++i) {
267        r.f[i] = a.f[i] * b.f[i] + c.f[i];
268    }
269    return r;
270}
271

1void convolution2D(int isize,   // width/height of input
2        int osize,              // width/height of output
3        int ksize,              // width/height of kernel
4        int stride,             // shift between input pixels, between consecutive outputs
5        int pad,                // offset between (0,0) pixels between input and output
6        int idepth, int odepth, // number of input and output channels
7        float idata[isize][isize][idepth],
8        float odata[osize][osize][odepth],
9        float kdata[odepth][ksize][ksize][idepth])
10{
11    // iterate over the output
12    for (int oy = 0; oy < osize; ++oy) {
13    for (int ox = 0; ox < osize; ++ox) {
14    for (int od = 0; od < odepth; ++od) {
15        odata[oy][ox][od] = 0;  // When you iterate multiple times without closing the program, this number would stack up to infinity, so we have to zero it out every time.
16        for (int ky = 0; ky < ksize; ++ky) {
17        for (int kx = 0; kx < ksize; ++kx) {
18            // map position in output and kernel to the input
19            int iy = stride * oy + ky - pad;
20            int ix = stride * ox + kx - pad;
21            // use only valid inputs
22            if (iy >= 0 && iy < isize && ix >= 0 && ix < isize)
23                for (int id = 0; id < idepth; ++id)
24                    odata[oy][ox][od] += kdata[od][ky][kx][id] * idata[iy][ix][id];
25        }}
26    }}}
27
28}
29
300000000000007880 <convolution2D>:
31    7880:   f3 0f 1e fa             endbr64 
32    7884:   55                      push   rbp
33    7885:   48 89 e5                mov    rbp,rsp
34    7888:   41 57                   push   r15
35    788a:   41 56                   push   r14
36    788c:   41 55                   push   r13
37    788e:   41 54                   push   r12
38    7890:   53                      push   rbx
39    7891:   48 81 ec b0 00 00 00    sub    rsp,0xb0
40    7898:   ff 15 4a a7 00 00       call   QWORD PTR [rip+0xa74a]        # 11fe8 <mcount@GLIBC_2.2.5>
41    789e:   89 d3                   mov    ebx,edx
42    78a0:   89 55 a8                mov    DWORD PTR [rbp-0x58],edx
43    78a3:   89 8d 74 ff ff ff       mov    DWORD PTR [rbp-0x8c],ecx
44    78a9:   49 63 d1                movsxd rdx,r9d
45    78ac:   48 63 cf                movsxd rcx,edi
46    78af:   41 89 f2                mov    r10d,esi
47    78b2:   89 b5 38 ff ff ff       mov    DWORD PTR [rbp-0xc8],esi
48    78b8:   49 63 c0                movsxd rax,r8d
49    78bb:   48 0f af ca             imul   rcx,rdx
50    78bf:   48 63 75 10             movsxd rsi,DWORD PTR [rbp+0x10]
51    78c3:   49 89 d6                mov    r14,rdx
52    78c6:   4c 8d 24 95 00 00 00 00     lea    r12,[rdx*4+0x0]
53    78ce:   41 89 fd                mov    r13d,edi
54    78d1:   49 89 cb                mov    r11,rcx
55    78d4:   48 89 8d 60 ff ff ff    mov    QWORD PTR [rbp-0xa0],rcx
56    78db:   49 63 ca                movsxd rcx,r10d
57    78de:   4c 8d 0c b5 00 00 00 00     lea    r9,[rsi*4+0x0]
58    78e6:   49 89 f0                mov    r8,rsi
59    78e9:   48 0f af f1             imul   rsi,rcx
60    78ed:   48 63 cb                movsxd rcx,ebx
61    78f0:   4c 89 8d 48 ff ff ff    mov    QWORD PTR [rbp-0xb8],r9
62    78f7:   48 0f af d1             imul   rdx,rcx
63    78fb:   48 8d 3c 95 00 00 00 00     lea    rdi,[rdx*4+0x0]
64    7903:   45 85 d2                test   r10d,r10d
65    7906:   0f 8e 73 02 00 00       jle    7b7f <convolution2D+0x2ff>
66    790c:   48 c1 ef 02             shr    rdi,0x2
67    7910:   49 c1 e9 02             shr    r9,0x2
68    7914:   48 89 7d c8             mov    QWORD PTR [rbp-0x38],rdi
69    7918:   4c 89 e7                mov    rdi,r12
70    791b:   4c 89 8d 58 ff ff ff    mov    QWORD PTR [rbp-0xa8],r9
71    7922:   48 c1 ef 02             shr    rdi,0x2
72    7926:   48 89 bd 50 ff ff ff    mov    QWORD PTR [rbp-0xb0],rdi
73    792d:   45 85 c0                test   r8d,r8d
74    7930:   0f 8e 49 02 00 00       jle    7b7f <convolution2D+0x2ff>
75    7936:   48 c1 e6 02             shl    rsi,0x2
76    793a:   48 0f af d1             imul   rdx,rcx
77    793e:   29 c3                   sub    ebx,eax
78    7940:   89 c7                   mov    edi,eax
79    7942:   48 89 b5 30 ff ff ff    mov    QWORD PTR [rbp-0xd0],rsi
80    7949:   48 8b 75 20             mov    rsi,QWORD PTR [rbp+0x20]
81    794d:   48 89 85 68 ff ff ff    mov    QWORD PTR [rbp-0x98],rax
82    7954:   f7 df                   neg    edi
83    7956:   45 8d 7e ff             lea    r15d,[r14-0x1]
84    795a:   89 9d 70 ff ff ff       mov    DWORD PTR [rbp-0x90],ebx
85    7960:   89 bd 3c ff ff ff       mov    DWORD PTR [rbp-0xc4],edi
86    7966:   48 8d 0c 95 00 00 00 00     lea    rcx,[rdx*4+0x0]
87    796e:   89 7d ac                mov    DWORD PTR [rbp-0x54],edi
88    7971:   89 5d d4                mov    DWORD PTR [rbp-0x2c],ebx
89    7974:   48 89 4d 98             mov    QWORD PTR [rbp-0x68],rcx
90    7978:   4a 8d 0c 9d 00 00 00 00     lea    rcx,[r11*4+0x0]
91    7980:   c7 45 80 00 00 00 00    mov    DWORD PTR [rbp-0x80],0x0
92    7987:   48 89 75 88             mov    QWORD PTR [rbp-0x78],rsi
93    798b:   41 8d 70 ff             lea    esi,[r8-0x1]
94    798f:   48 89 4d c0             mov    QWORD PTR [rbp-0x40],rcx
95    7993:   48 8d 04 b5 04 00 00 00     lea    rax,[rsi*4+0x4]
96    799b:   c7 45 90 00 00 00 00    mov    DWORD PTR [rbp-0x70],0x0
97    79a2:   48 89 85 28 ff ff ff    mov    QWORD PTR [rbp-0xd8],rax
98    79a9:   44 89 f0                mov    eax,r14d
99    79ac:   45 89 ee                mov    r14d,r13d
100    79af:   41 89 c5                mov    r13d,eax
101    79b2:   48 8b 85 28 ff ff ff    mov    rax,QWORD PTR [rbp-0xd8]
102    79b9:   48 03 45 88             add    rax,QWORD PTR [rbp-0x78]
103    79bd:   48 c7 85 78 ff ff ff 00 00 00 00    mov    QWORD PTR [rbp-0x88],0x0
104    79c8:   c7 45 84 00 00 00 00    mov    DWORD PTR [rbp-0x7c],0x0
105    79cf:   c7 45 94 00 00 00 00    mov    DWORD PTR [rbp-0x6c],0x0
106    79d6:   44 8b 95 70 ff ff ff    mov    r10d,DWORD PTR [rbp-0x90]
107    79dd:   48 89 45 b0             mov    QWORD PTR [rbp-0x50],rax
108    79e1:   48 63 45 80             movsxd rax,DWORD PTR [rbp-0x80]
109    79e5:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
110    79ec:   48 0f af 85 60 ff ff ff     imul   rax,QWORD PTR [rbp-0xa0]
111    79f4:   48 89 85 40 ff ff ff    mov    QWORD PTR [rbp-0xc0],rax
112    79fb:   8b 85 3c ff ff ff       mov    eax,DWORD PTR [rbp-0xc4]
113    7a01:   89 45 d0                mov    DWORD PTR [rbp-0x30],eax
114    7a04:   48 8b 45 88             mov    rax,QWORD PTR [rbp-0x78]
115    7a08:   48 8b 9d 78 ff ff ff    mov    rbx,QWORD PTR [rbp-0x88]
116    7a0f:   4c 8d 04 98             lea    r8,[rax+rbx*4]
117    7a13:   48 8b 45 28             mov    rax,QWORD PTR [rbp+0x28]
118    7a17:   48 8b 5d 18             mov    rbx,QWORD PTR [rbp+0x18]
119    7a1b:   48 89 45 b8             mov    QWORD PTR [rbp-0x48],rax
120    7a1f:   48 63 45 84             movsxd rax,DWORD PTR [rbp-0x7c]
121    7a23:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
122    7a2a:   48 0f af 85 50 ff ff ff     imul   rax,QWORD PTR [rbp-0xb0]
123    7a32:   48 03 85 40 ff ff ff    add    rax,QWORD PTR [rbp-0xc0]
124    7a39:   48 8d 04 83             lea    rax,[rbx+rax*4]
125    7a3d:   48 89 45 a0             mov    QWORD PTR [rbp-0x60],rax
126    7a41:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
127    7a4c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
128    7a50:   8b 45 a8                mov    eax,DWORD PTR [rbp-0x58]
129    7a53:   41 c7 00 00 00 00 00    mov    DWORD PTR [r8],0x0
130    7a5a:   45 31 db                xor    r11d,r11d
131    7a5d:   48 8b 5d a0             mov    rbx,QWORD PTR [rbp-0x60]
132    7a61:   44 8b 4d ac             mov    r9d,DWORD PTR [rbp-0x54]
133    7a65:   85 c0                   test   eax,eax
134    7a67:   0f 8e 98 00 00 00       jle    7b05 <convolution2D+0x285>
135    7a6d:   0f 1f 00                nop    DWORD PTR [rax]
136    7a70:   45 85 c9                test   r9d,r9d
137    7a73:   78 7b                   js     7af0 <convolution2D+0x270>
138    7a75:   45 39 ce                cmp    r14d,r9d
139    7a78:   7e 76                   jle    7af0 <convolution2D+0x270>
140    7a7a:   48 8b 45 b8             mov    rax,QWORD PTR [rbp-0x48]
141    7a7e:   8b 55 d0                mov    edx,DWORD PTR [rbp-0x30]
142    7a81:   48 89 de                mov    rsi,rbx
143    7a84:   4a 8d 3c 98             lea    rdi,[rax+r11*4]
144    7a88:   eb 13                   jmp    7a9d <convolution2D+0x21d>
145    7a8a:   66 0f 1f 44 00 00       nop    WORD PTR [rax+rax*1+0x0]
146    7a90:   ff c2                   inc    edx
147    7a92:   4c 01 e7                add    rdi,r12
148    7a95:   4c 01 e6                add    rsi,r12
149    7a98:   44 39 d2                cmp    edx,r10d
150    7a9b:   74 53                   je     7af0 <convolution2D+0x270>
151    7a9d:   85 d2                   test   edx,edx
152    7a9f:   78 ef                   js     7a90 <convolution2D+0x210>
153    7aa1:   41 39 d6                cmp    r14d,edx
154    7aa4:   7e ea                   jle    7a90 <convolution2D+0x210>
155    7aa6:   45 85 ed                test   r13d,r13d
156    7aa9:   7e e5                   jle    7a90 <convolution2D+0x210>
157    7aab:   c4 c1 7a 10 08          vmovss xmm1,DWORD PTR [r8]
158    7ab0:   31 c0                   xor    eax,eax
159    7ab2:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
160    7abd:   0f 1f 00                nop    DWORD PTR [rax]
161    7ac0:   c5 fa 10 04 87          vmovss xmm0,DWORD PTR [rdi+rax*4]
162    7ac5:   48 89 c1                mov    rcx,rax
163    7ac8:   c5 fa 59 04 86          vmulss xmm0,xmm0,DWORD PTR [rsi+rax*4]
164    7acd:   48 ff c0                inc    rax
165    7ad0:   c5 f2 58 c8             vaddss xmm1,xmm1,xmm0
166    7ad4:   c4 c1 7a 11 08          vmovss DWORD PTR [r8],xmm1
167    7ad9:   49 39 cf                cmp    r15,rcx
168    7adc:   75 e2                   jne    7ac0 <convolution2D+0x240>
169    7ade:   ff c2                   inc    edx
170    7ae0:   4c 01 e7                add    rdi,r12
171    7ae3:   4c 01 e6                add    rsi,r12
172    7ae6:   44 39 d2                cmp    edx,r10d
173    7ae9:   75 b2                   jne    7a9d <convolution2D+0x21d>
174    7aeb:   0f 1f 44 00 00          nop    DWORD PTR [rax+rax*1+0x0]
175    7af0:   4c 03 5d c8             add    r11,QWORD PTR [rbp-0x38]
176    7af4:   48 03 5d c0             add    rbx,QWORD PTR [rbp-0x40]
177    7af8:   41 ff c1                inc    r9d
178    7afb:   44 3b 4d d4             cmp    r9d,DWORD PTR [rbp-0x2c]
179    7aff:   0f 85 6b ff ff ff       jne    7a70 <convolution2D+0x1f0>
180    7b05:   48 8b 5d 98             mov    rbx,QWORD PTR [rbp-0x68]
181    7b09:   49 83 c0 04             add    r8,0x4
182    7b0d:   48 01 5d b8             add    QWORD PTR [rbp-0x48],rbx
183    7b11:   4c 3b 45 b0             cmp    r8,QWORD PTR [rbp-0x50]
184    7b15:   0f 85 35 ff ff ff       jne    7a50 <convolution2D+0x1d0>
185    7b1b:   8b 9d 74 ff ff ff       mov    ebx,DWORD PTR [rbp-0x8c]
186    7b21:   8b 45 94                mov    eax,DWORD PTR [rbp-0x6c]
187    7b24:   48 8b 8d 48 ff ff ff    mov    rcx,QWORD PTR [rbp-0xb8]
188    7b2b:   01 5d d0                add    DWORD PTR [rbp-0x30],ebx
189    7b2e:   48 01 4d b0             add    QWORD PTR [rbp-0x50],rcx
190    7b32:   01 5d 84                add    DWORD PTR [rbp-0x7c],ebx
191    7b35:   48 8b 8d 58 ff ff ff    mov    rcx,QWORD PTR [rbp-0xa8]
192    7b3c:   41 01 da                add    r10d,ebx
193    7b3f:   48 01 8d 78 ff ff ff    add    QWORD PTR [rbp-0x88],rcx
194    7b46:   ff c0                   inc    eax
195    7b48:   39 85 38 ff ff ff       cmp    DWORD PTR [rbp-0xc8],eax
196    7b4e:   74 08                   je     7b58 <convolution2D+0x2d8>
197    7b50:   89 45 94                mov    DWORD PTR [rbp-0x6c],eax
198    7b53:   e9 ac fe ff ff          jmp    7a04 <convolution2D+0x184>
199    7b58:   8b 4d 90                mov    ecx,DWORD PTR [rbp-0x70]
200    7b5b:   48 8b b5 30 ff ff ff    mov    rsi,QWORD PTR [rbp-0xd0]
201    7b62:   01 5d d4                add    DWORD PTR [rbp-0x2c],ebx
202    7b65:   01 5d ac                add    DWORD PTR [rbp-0x54],ebx
203    7b68:   01 5d 80                add    DWORD PTR [rbp-0x80],ebx
204    7b6b:   48 01 75 88             add    QWORD PTR [rbp-0x78],rsi
205    7b6f:   8d 41 01                lea    eax,[rcx+0x1]
206    7b72:   39 4d 94                cmp    DWORD PTR [rbp-0x6c],ecx
207    7b75:   74 08                   je     7b7f <convolution2D+0x2ff>
208    7b77:   89 45 90                mov    DWORD PTR [rbp-0x70],eax
209    7b7a:   e9 33 fe ff ff          jmp    79b2 <convolution2D+0x132>
210    7b7f:   48 81 c4 b0 00 00 00    add    rsp,0xb0
211    7b86:   5b                      pop    rbx
212    7b87:   41 5c                   pop    r12
213    7b89:   41 5d                   pop    r13
214    7b8b:   41 5e                   pop    r14
215    7b8d:   41 5f                   pop    r15
216    7b8f:   5d                      pop    rbp
217    7b90:   c3                      ret    
218    7b91:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
219    7b9c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
220
221float (*arr)[x][y] = calloc(z, sizeof *arr);
222==430300== Cachegrind, a cache and branch-prediction profiler
223==430300== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
224==430300== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
225==430300== Command: ./EmbeddedNet test 1
226==430300== Parent PID: 170008
227==430300== 
228--430300-- warning: L3 cache found, using its data for the LL simulation.
229==430300== 
230==430300== I   refs:      6,369,594,192
231==430300== I1  misses:            4,271
232==430300== LLi misses:            2,442
233==430300== I1  miss rate:          0.00%
234==430300== LLi miss rate:          0.00%
235==430300== 
236==430300== D   refs:      2,064,233,110  (1,359,003,131 rd   + 705,229,979 wr)
237==430300== D1  misses:       34,476,969  (   19,010,839 rd   +  15,466,130 wr)
238==430300== LLd misses:        5,311,277  (    1,603,955 rd   +   3,707,322 wr)
239==430300== D1  miss rate:           1.7% (          1.4%     +         2.2%  )
240==430300== LLd miss rate:           0.3% (          0.1%     +         0.5%  )
241==430300== 
242==430300== LL refs:          34,481,240  (   19,015,110 rd   +  15,466,130 wr)
243==430300== LL misses:         5,313,719  (    1,606,397 rd   +   3,707,322 wr)
244==430300== LL miss rate:            0.1% (          0.0%     +         0.5%  )
245float fma_scl(float a, float b, float c) {
246    return a * b + c;
247}
248
249fma_scl:
250        vfmadd132ss     xmm0, xmm2, xmm1
251        ret
252typedef float Vec __attribute__((vector_size(32), aligned(32)));
253Vec fma_vec(Vec a, Vec b, Vec c) {
254    return a * b + c;
255}
256
257fma_vec:
258        vfmadd132ps     ymm0, ymm2, ymm1
259        ret
260typedef struct {
261    float f[8];
262} Vec_;
263
264Vec_ fma_vec_(Vec_ a, Vec_ b, Vec_ c) {
265    Vec_ r;
266    for (unsigned i = 0; i < 8; ++i) {
267        r.f[i] = a.f[i] * b.f[i] + c.f[i];
268    }
269    return r;
270}
271void loopadd_scl(float *restrict a, float *restrict b, float *restrict c, unsigned n) {
272    for (unsigned i = 0; i < n; ++i) {
273        a[i] = fma_scl(b[i], c[i], a[i]);
274    }
275}
276

1void convolution2D(int isize,   // width/height of input
2        int osize,              // width/height of output
3        int ksize,              // width/height of kernel
4        int stride,             // shift between input pixels, between consecutive outputs
5        int pad,                // offset between (0,0) pixels between input and output
6        int idepth, int odepth, // number of input and output channels
7        float idata[isize][isize][idepth],
8        float odata[osize][osize][odepth],
9        float kdata[odepth][ksize][ksize][idepth])
10{
11    // iterate over the output
12    for (int oy = 0; oy < osize; ++oy) {
13    for (int ox = 0; ox < osize; ++ox) {
14    for (int od = 0; od < odepth; ++od) {
15        odata[oy][ox][od] = 0;  // When you iterate multiple times without closing the program, this number would stack up to infinity, so we have to zero it out every time.
16        for (int ky = 0; ky < ksize; ++ky) {
17        for (int kx = 0; kx < ksize; ++kx) {
18            // map position in output and kernel to the input
19            int iy = stride * oy + ky - pad;
20            int ix = stride * ox + kx - pad;
21            // use only valid inputs
22            if (iy >= 0 && iy < isize && ix >= 0 && ix < isize)
23                for (int id = 0; id < idepth; ++id)
24                    odata[oy][ox][od] += kdata[od][ky][kx][id] * idata[iy][ix][id];
25        }}
26    }}}
27
28}
29
300000000000007880 <convolution2D>:
31    7880:   f3 0f 1e fa             endbr64 
32    7884:   55                      push   rbp
33    7885:   48 89 e5                mov    rbp,rsp
34    7888:   41 57                   push   r15
35    788a:   41 56                   push   r14
36    788c:   41 55                   push   r13
37    788e:   41 54                   push   r12
38    7890:   53                      push   rbx
39    7891:   48 81 ec b0 00 00 00    sub    rsp,0xb0
40    7898:   ff 15 4a a7 00 00       call   QWORD PTR [rip+0xa74a]        # 11fe8 <mcount@GLIBC_2.2.5>
41    789e:   89 d3                   mov    ebx,edx
42    78a0:   89 55 a8                mov    DWORD PTR [rbp-0x58],edx
43    78a3:   89 8d 74 ff ff ff       mov    DWORD PTR [rbp-0x8c],ecx
44    78a9:   49 63 d1                movsxd rdx,r9d
45    78ac:   48 63 cf                movsxd rcx,edi
46    78af:   41 89 f2                mov    r10d,esi
47    78b2:   89 b5 38 ff ff ff       mov    DWORD PTR [rbp-0xc8],esi
48    78b8:   49 63 c0                movsxd rax,r8d
49    78bb:   48 0f af ca             imul   rcx,rdx
50    78bf:   48 63 75 10             movsxd rsi,DWORD PTR [rbp+0x10]
51    78c3:   49 89 d6                mov    r14,rdx
52    78c6:   4c 8d 24 95 00 00 00 00     lea    r12,[rdx*4+0x0]
53    78ce:   41 89 fd                mov    r13d,edi
54    78d1:   49 89 cb                mov    r11,rcx
55    78d4:   48 89 8d 60 ff ff ff    mov    QWORD PTR [rbp-0xa0],rcx
56    78db:   49 63 ca                movsxd rcx,r10d
57    78de:   4c 8d 0c b5 00 00 00 00     lea    r9,[rsi*4+0x0]
58    78e6:   49 89 f0                mov    r8,rsi
59    78e9:   48 0f af f1             imul   rsi,rcx
60    78ed:   48 63 cb                movsxd rcx,ebx
61    78f0:   4c 89 8d 48 ff ff ff    mov    QWORD PTR [rbp-0xb8],r9
62    78f7:   48 0f af d1             imul   rdx,rcx
63    78fb:   48 8d 3c 95 00 00 00 00     lea    rdi,[rdx*4+0x0]
64    7903:   45 85 d2                test   r10d,r10d
65    7906:   0f 8e 73 02 00 00       jle    7b7f <convolution2D+0x2ff>
66    790c:   48 c1 ef 02             shr    rdi,0x2
67    7910:   49 c1 e9 02             shr    r9,0x2
68    7914:   48 89 7d c8             mov    QWORD PTR [rbp-0x38],rdi
69    7918:   4c 89 e7                mov    rdi,r12
70    791b:   4c 89 8d 58 ff ff ff    mov    QWORD PTR [rbp-0xa8],r9
71    7922:   48 c1 ef 02             shr    rdi,0x2
72    7926:   48 89 bd 50 ff ff ff    mov    QWORD PTR [rbp-0xb0],rdi
73    792d:   45 85 c0                test   r8d,r8d
74    7930:   0f 8e 49 02 00 00       jle    7b7f <convolution2D+0x2ff>
75    7936:   48 c1 e6 02             shl    rsi,0x2
76    793a:   48 0f af d1             imul   rdx,rcx
77    793e:   29 c3                   sub    ebx,eax
78    7940:   89 c7                   mov    edi,eax
79    7942:   48 89 b5 30 ff ff ff    mov    QWORD PTR [rbp-0xd0],rsi
80    7949:   48 8b 75 20             mov    rsi,QWORD PTR [rbp+0x20]
81    794d:   48 89 85 68 ff ff ff    mov    QWORD PTR [rbp-0x98],rax
82    7954:   f7 df                   neg    edi
83    7956:   45 8d 7e ff             lea    r15d,[r14-0x1]
84    795a:   89 9d 70 ff ff ff       mov    DWORD PTR [rbp-0x90],ebx
85    7960:   89 bd 3c ff ff ff       mov    DWORD PTR [rbp-0xc4],edi
86    7966:   48 8d 0c 95 00 00 00 00     lea    rcx,[rdx*4+0x0]
87    796e:   89 7d ac                mov    DWORD PTR [rbp-0x54],edi
88    7971:   89 5d d4                mov    DWORD PTR [rbp-0x2c],ebx
89    7974:   48 89 4d 98             mov    QWORD PTR [rbp-0x68],rcx
90    7978:   4a 8d 0c 9d 00 00 00 00     lea    rcx,[r11*4+0x0]
91    7980:   c7 45 80 00 00 00 00    mov    DWORD PTR [rbp-0x80],0x0
92    7987:   48 89 75 88             mov    QWORD PTR [rbp-0x78],rsi
93    798b:   41 8d 70 ff             lea    esi,[r8-0x1]
94    798f:   48 89 4d c0             mov    QWORD PTR [rbp-0x40],rcx
95    7993:   48 8d 04 b5 04 00 00 00     lea    rax,[rsi*4+0x4]
96    799b:   c7 45 90 00 00 00 00    mov    DWORD PTR [rbp-0x70],0x0
97    79a2:   48 89 85 28 ff ff ff    mov    QWORD PTR [rbp-0xd8],rax
98    79a9:   44 89 f0                mov    eax,r14d
99    79ac:   45 89 ee                mov    r14d,r13d
100    79af:   41 89 c5                mov    r13d,eax
101    79b2:   48 8b 85 28 ff ff ff    mov    rax,QWORD PTR [rbp-0xd8]
102    79b9:   48 03 45 88             add    rax,QWORD PTR [rbp-0x78]
103    79bd:   48 c7 85 78 ff ff ff 00 00 00 00    mov    QWORD PTR [rbp-0x88],0x0
104    79c8:   c7 45 84 00 00 00 00    mov    DWORD PTR [rbp-0x7c],0x0
105    79cf:   c7 45 94 00 00 00 00    mov    DWORD PTR [rbp-0x6c],0x0
106    79d6:   44 8b 95 70 ff ff ff    mov    r10d,DWORD PTR [rbp-0x90]
107    79dd:   48 89 45 b0             mov    QWORD PTR [rbp-0x50],rax
108    79e1:   48 63 45 80             movsxd rax,DWORD PTR [rbp-0x80]
109    79e5:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
110    79ec:   48 0f af 85 60 ff ff ff     imul   rax,QWORD PTR [rbp-0xa0]
111    79f4:   48 89 85 40 ff ff ff    mov    QWORD PTR [rbp-0xc0],rax
112    79fb:   8b 85 3c ff ff ff       mov    eax,DWORD PTR [rbp-0xc4]
113    7a01:   89 45 d0                mov    DWORD PTR [rbp-0x30],eax
114    7a04:   48 8b 45 88             mov    rax,QWORD PTR [rbp-0x78]
115    7a08:   48 8b 9d 78 ff ff ff    mov    rbx,QWORD PTR [rbp-0x88]
116    7a0f:   4c 8d 04 98             lea    r8,[rax+rbx*4]
117    7a13:   48 8b 45 28             mov    rax,QWORD PTR [rbp+0x28]
118    7a17:   48 8b 5d 18             mov    rbx,QWORD PTR [rbp+0x18]
119    7a1b:   48 89 45 b8             mov    QWORD PTR [rbp-0x48],rax
120    7a1f:   48 63 45 84             movsxd rax,DWORD PTR [rbp-0x7c]
121    7a23:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
122    7a2a:   48 0f af 85 50 ff ff ff     imul   rax,QWORD PTR [rbp-0xb0]
123    7a32:   48 03 85 40 ff ff ff    add    rax,QWORD PTR [rbp-0xc0]
124    7a39:   48 8d 04 83             lea    rax,[rbx+rax*4]
125    7a3d:   48 89 45 a0             mov    QWORD PTR [rbp-0x60],rax
126    7a41:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
127    7a4c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
128    7a50:   8b 45 a8                mov    eax,DWORD PTR [rbp-0x58]
129    7a53:   41 c7 00 00 00 00 00    mov    DWORD PTR [r8],0x0
130    7a5a:   45 31 db                xor    r11d,r11d
131    7a5d:   48 8b 5d a0             mov    rbx,QWORD PTR [rbp-0x60]
132    7a61:   44 8b 4d ac             mov    r9d,DWORD PTR [rbp-0x54]
133    7a65:   85 c0                   test   eax,eax
134    7a67:   0f 8e 98 00 00 00       jle    7b05 <convolution2D+0x285>
135    7a6d:   0f 1f 00                nop    DWORD PTR [rax]
136    7a70:   45 85 c9                test   r9d,r9d
137    7a73:   78 7b                   js     7af0 <convolution2D+0x270>
138    7a75:   45 39 ce                cmp    r14d,r9d
139    7a78:   7e 76                   jle    7af0 <convolution2D+0x270>
140    7a7a:   48 8b 45 b8             mov    rax,QWORD PTR [rbp-0x48]
141    7a7e:   8b 55 d0                mov    edx,DWORD PTR [rbp-0x30]
142    7a81:   48 89 de                mov    rsi,rbx
143    7a84:   4a 8d 3c 98             lea    rdi,[rax+r11*4]
144    7a88:   eb 13                   jmp    7a9d <convolution2D+0x21d>
145    7a8a:   66 0f 1f 44 00 00       nop    WORD PTR [rax+rax*1+0x0]
146    7a90:   ff c2                   inc    edx
147    7a92:   4c 01 e7                add    rdi,r12
148    7a95:   4c 01 e6                add    rsi,r12
149    7a98:   44 39 d2                cmp    edx,r10d
150    7a9b:   74 53                   je     7af0 <convolution2D+0x270>
151    7a9d:   85 d2                   test   edx,edx
152    7a9f:   78 ef                   js     7a90 <convolution2D+0x210>
153    7aa1:   41 39 d6                cmp    r14d,edx
154    7aa4:   7e ea                   jle    7a90 <convolution2D+0x210>
155    7aa6:   45 85 ed                test   r13d,r13d
156    7aa9:   7e e5                   jle    7a90 <convolution2D+0x210>
157    7aab:   c4 c1 7a 10 08          vmovss xmm1,DWORD PTR [r8]
158    7ab0:   31 c0                   xor    eax,eax
159    7ab2:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
160    7abd:   0f 1f 00                nop    DWORD PTR [rax]
161    7ac0:   c5 fa 10 04 87          vmovss xmm0,DWORD PTR [rdi+rax*4]
162    7ac5:   48 89 c1                mov    rcx,rax
163    7ac8:   c5 fa 59 04 86          vmulss xmm0,xmm0,DWORD PTR [rsi+rax*4]
164    7acd:   48 ff c0                inc    rax
165    7ad0:   c5 f2 58 c8             vaddss xmm1,xmm1,xmm0
166    7ad4:   c4 c1 7a 11 08          vmovss DWORD PTR [r8],xmm1
167    7ad9:   49 39 cf                cmp    r15,rcx
168    7adc:   75 e2                   jne    7ac0 <convolution2D+0x240>
169    7ade:   ff c2                   inc    edx
170    7ae0:   4c 01 e7                add    rdi,r12
171    7ae3:   4c 01 e6                add    rsi,r12
172    7ae6:   44 39 d2                cmp    edx,r10d
173    7ae9:   75 b2                   jne    7a9d <convolution2D+0x21d>
174    7aeb:   0f 1f 44 00 00          nop    DWORD PTR [rax+rax*1+0x0]
175    7af0:   4c 03 5d c8             add    r11,QWORD PTR [rbp-0x38]
176    7af4:   48 03 5d c0             add    rbx,QWORD PTR [rbp-0x40]
177    7af8:   41 ff c1                inc    r9d
178    7afb:   44 3b 4d d4             cmp    r9d,DWORD PTR [rbp-0x2c]
179    7aff:   0f 85 6b ff ff ff       jne    7a70 <convolution2D+0x1f0>
180    7b05:   48 8b 5d 98             mov    rbx,QWORD PTR [rbp-0x68]
181    7b09:   49 83 c0 04             add    r8,0x4
182    7b0d:   48 01 5d b8             add    QWORD PTR [rbp-0x48],rbx
183    7b11:   4c 3b 45 b0             cmp    r8,QWORD PTR [rbp-0x50]
184    7b15:   0f 85 35 ff ff ff       jne    7a50 <convolution2D+0x1d0>
185    7b1b:   8b 9d 74 ff ff ff       mov    ebx,DWORD PTR [rbp-0x8c]
186    7b21:   8b 45 94                mov    eax,DWORD PTR [rbp-0x6c]
187    7b24:   48 8b 8d 48 ff ff ff    mov    rcx,QWORD PTR [rbp-0xb8]
188    7b2b:   01 5d d0                add    DWORD PTR [rbp-0x30],ebx
189    7b2e:   48 01 4d b0             add    QWORD PTR [rbp-0x50],rcx
190    7b32:   01 5d 84                add    DWORD PTR [rbp-0x7c],ebx
191    7b35:   48 8b 8d 58 ff ff ff    mov    rcx,QWORD PTR [rbp-0xa8]
192    7b3c:   41 01 da                add    r10d,ebx
193    7b3f:   48 01 8d 78 ff ff ff    add    QWORD PTR [rbp-0x88],rcx
194    7b46:   ff c0                   inc    eax
195    7b48:   39 85 38 ff ff ff       cmp    DWORD PTR [rbp-0xc8],eax
196    7b4e:   74 08                   je     7b58 <convolution2D+0x2d8>
197    7b50:   89 45 94                mov    DWORD PTR [rbp-0x6c],eax
198    7b53:   e9 ac fe ff ff          jmp    7a04 <convolution2D+0x184>
199    7b58:   8b 4d 90                mov    ecx,DWORD PTR [rbp-0x70]
200    7b5b:   48 8b b5 30 ff ff ff    mov    rsi,QWORD PTR [rbp-0xd0]
201    7b62:   01 5d d4                add    DWORD PTR [rbp-0x2c],ebx
202    7b65:   01 5d ac                add    DWORD PTR [rbp-0x54],ebx
203    7b68:   01 5d 80                add    DWORD PTR [rbp-0x80],ebx
204    7b6b:   48 01 75 88             add    QWORD PTR [rbp-0x78],rsi
205    7b6f:   8d 41 01                lea    eax,[rcx+0x1]
206    7b72:   39 4d 94                cmp    DWORD PTR [rbp-0x6c],ecx
207    7b75:   74 08                   je     7b7f <convolution2D+0x2ff>
208    7b77:   89 45 90                mov    DWORD PTR [rbp-0x70],eax
209    7b7a:   e9 33 fe ff ff          jmp    79b2 <convolution2D+0x132>
210    7b7f:   48 81 c4 b0 00 00 00    add    rsp,0xb0
211    7b86:   5b                      pop    rbx
212    7b87:   41 5c                   pop    r12
213    7b89:   41 5d                   pop    r13
214    7b8b:   41 5e                   pop    r14
215    7b8d:   41 5f                   pop    r15
216    7b8f:   5d                      pop    rbp
217    7b90:   c3                      ret    
218    7b91:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
219    7b9c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
220
221float (*arr)[x][y] = calloc(z, sizeof *arr);
222==430300== Cachegrind, a cache and branch-prediction profiler
223==430300== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
224==430300== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
225==430300== Command: ./EmbeddedNet test 1
226==430300== Parent PID: 170008
227==430300== 
228--430300-- warning: L3 cache found, using its data for the LL simulation.
229==430300== 
230==430300== I   refs:      6,369,594,192
231==430300== I1  misses:            4,271
232==430300== LLi misses:            2,442
233==430300== I1  miss rate:          0.00%
234==430300== LLi miss rate:          0.00%
235==430300== 
236==430300== D   refs:      2,064,233,110  (1,359,003,131 rd   + 705,229,979 wr)
237==430300== D1  misses:       34,476,969  (   19,010,839 rd   +  15,466,130 wr)
238==430300== LLd misses:        5,311,277  (    1,603,955 rd   +   3,707,322 wr)
239==430300== D1  miss rate:           1.7% (          1.4%     +         2.2%  )
240==430300== LLd miss rate:           0.3% (          0.1%     +         0.5%  )
241==430300== 
242==430300== LL refs:          34,481,240  (   19,015,110 rd   +  15,466,130 wr)
243==430300== LL misses:         5,313,719  (    1,606,397 rd   +   3,707,322 wr)
244==430300== LL miss rate:            0.1% (          0.0%     +         0.5%  )
245float fma_scl(float a, float b, float c) {
246    return a * b + c;
247}
248
249fma_scl:
250        vfmadd132ss     xmm0, xmm2, xmm1
251        ret
252typedef float Vec __attribute__((vector_size(32), aligned(32)));
253Vec fma_vec(Vec a, Vec b, Vec c) {
254    return a * b + c;
255}
256
257fma_vec:
258        vfmadd132ps     ymm0, ymm2, ymm1
259        ret
260typedef struct {
261    float f[8];
262} Vec_;
263
264Vec_ fma_vec_(Vec_ a, Vec_ b, Vec_ c) {
265    Vec_ r;
266    for (unsigned i = 0; i < 8; ++i) {
267        r.f[i] = a.f[i] * b.f[i] + c.f[i];
268    }
269    return r;
270}
271void loopadd_scl(float *restrict a, float *restrict b, float *restrict c, unsigned n) {
272    for (unsigned i = 0; i < n; ++i) {
273        a[i] = fma_scl(b[i], c[i], a[i]);
274    }
275}
276loopadd_scl:
277        test    ecx, ecx
278        je      .L25
279        lea     eax, [rcx-1]
280        cmp     eax, 6
281        jbe     .L13
282        mov     r8d, ecx
283        xor     eax, eax
284        shr     r8d, 3
285        sal     r8, 5
286.L9:
287        vmovups ymm1, YMMWORD PTR [rdi+rax]
288        vmovups ymm0, YMMWORD PTR [rdx+rax]
289        vfmadd132ps     ymm0, ymm1, YMMWORD PTR [rsi+rax]
290        vmovups YMMWORD PTR [rdi+rax], ymm0
291        add     rax, 32
292        cmp     r8, rax
293        jne     .L9
294        mov     eax, ecx
295        and     eax, -8
296        test    cl, 7
297        je      .L26
298        vzeroupper
299.L8:
300        mov     r9d, ecx
301        sub     r9d, eax
302        lea     r8d, [r9-1]
303        cmp     r8d, 2
304        jbe     .L11
305        mov     r8d, eax
306        sal     r8, 2
307        lea     r10, [rdi+r8]
308        vmovups xmm0, XMMWORD PTR [rdx+r8]
309        vmovups xmm2, XMMWORD PTR [r10]
310        vfmadd132ps     xmm0, xmm2, XMMWORD PTR [rsi+r8]
311        mov     r8d, r9d
312        and     r8d, -4
313        add     eax, r8d
314        and     r9d, 3
315        vmovups XMMWORD PTR [r10], xmm0
316        je      .L25
317.L11:
318        mov     r8d, eax
319        sal     r8, 2
320        lea     r9, [rdi+r8]
321        vmovss  xmm0, DWORD PTR [rdx+r8]
322        vmovss  xmm3, DWORD PTR [r9]
323        vfmadd132ss     xmm0, xmm3, DWORD PTR [rsi+r8]
324        lea     r8d, [rax+1]
325        vmovss  DWORD PTR [r9], xmm0
326        cmp     r8d, ecx
327        jnb     .L25
328        sal     r8, 2
329        add     eax, 2
330        lea     r9, [rdi+r8]
331        vmovss  xmm0, DWORD PTR [rsi+r8]
332        vmovss  xmm4, DWORD PTR [r9]
333        vfmadd132ss     xmm0, xmm4, DWORD PTR [rdx+r8]
334        vmovss  DWORD PTR [r9], xmm0
335        cmp     eax, ecx
336        jnb     .L25
337        sal     rax, 2
338        add     rdi, rax
339        vmovss  xmm0, DWORD PTR [rdx+rax]
340        vmovss  xmm5, DWORD PTR [rdi]
341        vfmadd132ss     xmm0, xmm5, DWORD PTR [rsi+rax]
342        vmovss  DWORD PTR [rdi], xmm0
343.L25:
344        ret
345.L26:
346        vzeroupper
347        ret
348.L13:
349        xor     eax, eax
350        jmp     .L8
351

1void convolution2D(int isize,   // width/height of input
2        int osize,              // width/height of output
3        int ksize,              // width/height of kernel
4        int stride,             // shift between input pixels, between consecutive outputs
5        int pad,                // offset between (0,0) pixels between input and output
6        int idepth, int odepth, // number of input and output channels
7        float idata[isize][isize][idepth],
8        float odata[osize][osize][odepth],
9        float kdata[odepth][ksize][ksize][idepth])
10{
11    // iterate over the output
12    for (int oy = 0; oy < osize; ++oy) {
13    for (int ox = 0; ox < osize; ++ox) {
14    for (int od = 0; od < odepth; ++od) {
15        odata[oy][ox][od] = 0;  // When you iterate multiple times without closing the program, this number would stack up to infinity, so we have to zero it out every time.
16        for (int ky = 0; ky < ksize; ++ky) {
17        for (int kx = 0; kx < ksize; ++kx) {
18            // map position in output and kernel to the input
19            int iy = stride * oy + ky - pad;
20            int ix = stride * ox + kx - pad;
21            // use only valid inputs
22            if (iy >= 0 && iy < isize && ix >= 0 && ix < isize)
23                for (int id = 0; id < idepth; ++id)
24                    odata[oy][ox][od] += kdata[od][ky][kx][id] * idata[iy][ix][id];
25        }}
26    }}}
27
28}
29
300000000000007880 <convolution2D>:
31    7880:   f3 0f 1e fa             endbr64 
32    7884:   55                      push   rbp
33    7885:   48 89 e5                mov    rbp,rsp
34    7888:   41 57                   push   r15
35    788a:   41 56                   push   r14
36    788c:   41 55                   push   r13
37    788e:   41 54                   push   r12
38    7890:   53                      push   rbx
39    7891:   48 81 ec b0 00 00 00    sub    rsp,0xb0
40    7898:   ff 15 4a a7 00 00       call   QWORD PTR [rip+0xa74a]        # 11fe8 <mcount@GLIBC_2.2.5>
41    789e:   89 d3                   mov    ebx,edx
42    78a0:   89 55 a8                mov    DWORD PTR [rbp-0x58],edx
43    78a3:   89 8d 74 ff ff ff       mov    DWORD PTR [rbp-0x8c],ecx
44    78a9:   49 63 d1                movsxd rdx,r9d
45    78ac:   48 63 cf                movsxd rcx,edi
46    78af:   41 89 f2                mov    r10d,esi
47    78b2:   89 b5 38 ff ff ff       mov    DWORD PTR [rbp-0xc8],esi
48    78b8:   49 63 c0                movsxd rax,r8d
49    78bb:   48 0f af ca             imul   rcx,rdx
50    78bf:   48 63 75 10             movsxd rsi,DWORD PTR [rbp+0x10]
51    78c3:   49 89 d6                mov    r14,rdx
52    78c6:   4c 8d 24 95 00 00 00 00     lea    r12,[rdx*4+0x0]
53    78ce:   41 89 fd                mov    r13d,edi
54    78d1:   49 89 cb                mov    r11,rcx
55    78d4:   48 89 8d 60 ff ff ff    mov    QWORD PTR [rbp-0xa0],rcx
56    78db:   49 63 ca                movsxd rcx,r10d
57    78de:   4c 8d 0c b5 00 00 00 00     lea    r9,[rsi*4+0x0]
58    78e6:   49 89 f0                mov    r8,rsi
59    78e9:   48 0f af f1             imul   rsi,rcx
60    78ed:   48 63 cb                movsxd rcx,ebx
61    78f0:   4c 89 8d 48 ff ff ff    mov    QWORD PTR [rbp-0xb8],r9
62    78f7:   48 0f af d1             imul   rdx,rcx
63    78fb:   48 8d 3c 95 00 00 00 00     lea    rdi,[rdx*4+0x0]
64    7903:   45 85 d2                test   r10d,r10d
65    7906:   0f 8e 73 02 00 00       jle    7b7f <convolution2D+0x2ff>
66    790c:   48 c1 ef 02             shr    rdi,0x2
67    7910:   49 c1 e9 02             shr    r9,0x2
68    7914:   48 89 7d c8             mov    QWORD PTR [rbp-0x38],rdi
69    7918:   4c 89 e7                mov    rdi,r12
70    791b:   4c 89 8d 58 ff ff ff    mov    QWORD PTR [rbp-0xa8],r9
71    7922:   48 c1 ef 02             shr    rdi,0x2
72    7926:   48 89 bd 50 ff ff ff    mov    QWORD PTR [rbp-0xb0],rdi
73    792d:   45 85 c0                test   r8d,r8d
74    7930:   0f 8e 49 02 00 00       jle    7b7f <convolution2D+0x2ff>
75    7936:   48 c1 e6 02             shl    rsi,0x2
76    793a:   48 0f af d1             imul   rdx,rcx
77    793e:   29 c3                   sub    ebx,eax
78    7940:   89 c7                   mov    edi,eax
79    7942:   48 89 b5 30 ff ff ff    mov    QWORD PTR [rbp-0xd0],rsi
80    7949:   48 8b 75 20             mov    rsi,QWORD PTR [rbp+0x20]
81    794d:   48 89 85 68 ff ff ff    mov    QWORD PTR [rbp-0x98],rax
82    7954:   f7 df                   neg    edi
83    7956:   45 8d 7e ff             lea    r15d,[r14-0x1]
84    795a:   89 9d 70 ff ff ff       mov    DWORD PTR [rbp-0x90],ebx
85    7960:   89 bd 3c ff ff ff       mov    DWORD PTR [rbp-0xc4],edi
86    7966:   48 8d 0c 95 00 00 00 00     lea    rcx,[rdx*4+0x0]
87    796e:   89 7d ac                mov    DWORD PTR [rbp-0x54],edi
88    7971:   89 5d d4                mov    DWORD PTR [rbp-0x2c],ebx
89    7974:   48 89 4d 98             mov    QWORD PTR [rbp-0x68],rcx
90    7978:   4a 8d 0c 9d 00 00 00 00     lea    rcx,[r11*4+0x0]
91    7980:   c7 45 80 00 00 00 00    mov    DWORD PTR [rbp-0x80],0x0
92    7987:   48 89 75 88             mov    QWORD PTR [rbp-0x78],rsi
93    798b:   41 8d 70 ff             lea    esi,[r8-0x1]
94    798f:   48 89 4d c0             mov    QWORD PTR [rbp-0x40],rcx
95    7993:   48 8d 04 b5 04 00 00 00     lea    rax,[rsi*4+0x4]
96    799b:   c7 45 90 00 00 00 00    mov    DWORD PTR [rbp-0x70],0x0
97    79a2:   48 89 85 28 ff ff ff    mov    QWORD PTR [rbp-0xd8],rax
98    79a9:   44 89 f0                mov    eax,r14d
99    79ac:   45 89 ee                mov    r14d,r13d
100    79af:   41 89 c5                mov    r13d,eax
101    79b2:   48 8b 85 28 ff ff ff    mov    rax,QWORD PTR [rbp-0xd8]
102    79b9:   48 03 45 88             add    rax,QWORD PTR [rbp-0x78]
103    79bd:   48 c7 85 78 ff ff ff 00 00 00 00    mov    QWORD PTR [rbp-0x88],0x0
104    79c8:   c7 45 84 00 00 00 00    mov    DWORD PTR [rbp-0x7c],0x0
105    79cf:   c7 45 94 00 00 00 00    mov    DWORD PTR [rbp-0x6c],0x0
106    79d6:   44 8b 95 70 ff ff ff    mov    r10d,DWORD PTR [rbp-0x90]
107    79dd:   48 89 45 b0             mov    QWORD PTR [rbp-0x50],rax
108    79e1:   48 63 45 80             movsxd rax,DWORD PTR [rbp-0x80]
109    79e5:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
110    79ec:   48 0f af 85 60 ff ff ff     imul   rax,QWORD PTR [rbp-0xa0]
111    79f4:   48 89 85 40 ff ff ff    mov    QWORD PTR [rbp-0xc0],rax
112    79fb:   8b 85 3c ff ff ff       mov    eax,DWORD PTR [rbp-0xc4]
113    7a01:   89 45 d0                mov    DWORD PTR [rbp-0x30],eax
114    7a04:   48 8b 45 88             mov    rax,QWORD PTR [rbp-0x78]
115    7a08:   48 8b 9d 78 ff ff ff    mov    rbx,QWORD PTR [rbp-0x88]
116    7a0f:   4c 8d 04 98             lea    r8,[rax+rbx*4]
117    7a13:   48 8b 45 28             mov    rax,QWORD PTR [rbp+0x28]
118    7a17:   48 8b 5d 18             mov    rbx,QWORD PTR [rbp+0x18]
119    7a1b:   48 89 45 b8             mov    QWORD PTR [rbp-0x48],rax
120    7a1f:   48 63 45 84             movsxd rax,DWORD PTR [rbp-0x7c]
121    7a23:   48 2b 85 68 ff ff ff    sub    rax,QWORD PTR [rbp-0x98]
122    7a2a:   48 0f af 85 50 ff ff ff     imul   rax,QWORD PTR [rbp-0xb0]
123    7a32:   48 03 85 40 ff ff ff    add    rax,QWORD PTR [rbp-0xc0]
124    7a39:   48 8d 04 83             lea    rax,[rbx+rax*4]
125    7a3d:   48 89 45 a0             mov    QWORD PTR [rbp-0x60],rax
126    7a41:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
127    7a4c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
128    7a50:   8b 45 a8                mov    eax,DWORD PTR [rbp-0x58]
129    7a53:   41 c7 00 00 00 00 00    mov    DWORD PTR [r8],0x0
130    7a5a:   45 31 db                xor    r11d,r11d
131    7a5d:   48 8b 5d a0             mov    rbx,QWORD PTR [rbp-0x60]
132    7a61:   44 8b 4d ac             mov    r9d,DWORD PTR [rbp-0x54]
133    7a65:   85 c0                   test   eax,eax
134    7a67:   0f 8e 98 00 00 00       jle    7b05 <convolution2D+0x285>
135    7a6d:   0f 1f 00                nop    DWORD PTR [rax]
136    7a70:   45 85 c9                test   r9d,r9d
137    7a73:   78 7b                   js     7af0 <convolution2D+0x270>
138    7a75:   45 39 ce                cmp    r14d,r9d
139    7a78:   7e 76                   jle    7af0 <convolution2D+0x270>
140    7a7a:   48 8b 45 b8             mov    rax,QWORD PTR [rbp-0x48]
141    7a7e:   8b 55 d0                mov    edx,DWORD PTR [rbp-0x30]
142    7a81:   48 89 de                mov    rsi,rbx
143    7a84:   4a 8d 3c 98             lea    rdi,[rax+r11*4]
144    7a88:   eb 13                   jmp    7a9d <convolution2D+0x21d>
145    7a8a:   66 0f 1f 44 00 00       nop    WORD PTR [rax+rax*1+0x0]
146    7a90:   ff c2                   inc    edx
147    7a92:   4c 01 e7                add    rdi,r12
148    7a95:   4c 01 e6                add    rsi,r12
149    7a98:   44 39 d2                cmp    edx,r10d
150    7a9b:   74 53                   je     7af0 <convolution2D+0x270>
151    7a9d:   85 d2                   test   edx,edx
152    7a9f:   78 ef                   js     7a90 <convolution2D+0x210>
153    7aa1:   41 39 d6                cmp    r14d,edx
154    7aa4:   7e ea                   jle    7a90 <convolution2D+0x210>
155    7aa6:   45 85 ed                test   r13d,r13d
156    7aa9:   7e e5                   jle    7a90 <convolution2D+0x210>
157    7aab:   c4 c1 7a 10 08          vmovss xmm1,DWORD PTR [r8]
158    7ab0:   31 c0                   xor    eax,eax
159    7ab2:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
160    7abd:   0f 1f 00                nop    DWORD PTR [rax]
161    7ac0:   c5 fa 10 04 87          vmovss xmm0,DWORD PTR [rdi+rax*4]
162    7ac5:   48 89 c1                mov    rcx,rax
163    7ac8:   c5 fa 59 04 86          vmulss xmm0,xmm0,DWORD PTR [rsi+rax*4]
164    7acd:   48 ff c0                inc    rax
165    7ad0:   c5 f2 58 c8             vaddss xmm1,xmm1,xmm0
166    7ad4:   c4 c1 7a 11 08          vmovss DWORD PTR [r8],xmm1
167    7ad9:   49 39 cf                cmp    r15,rcx
168    7adc:   75 e2                   jne    7ac0 <convolution2D+0x240>
169    7ade:   ff c2                   inc    edx
170    7ae0:   4c 01 e7                add    rdi,r12
171    7ae3:   4c 01 e6                add    rsi,r12
172    7ae6:   44 39 d2                cmp    edx,r10d
173    7ae9:   75 b2                   jne    7a9d <convolution2D+0x21d>
174    7aeb:   0f 1f 44 00 00          nop    DWORD PTR [rax+rax*1+0x0]
175    7af0:   4c 03 5d c8             add    r11,QWORD PTR [rbp-0x38]
176    7af4:   48 03 5d c0             add    rbx,QWORD PTR [rbp-0x40]
177    7af8:   41 ff c1                inc    r9d
178    7afb:   44 3b 4d d4             cmp    r9d,DWORD PTR [rbp-0x2c]
179    7aff:   0f 85 6b ff ff ff       jne    7a70 <convolution2D+0x1f0>
180    7b05:   48 8b 5d 98             mov    rbx,QWORD PTR [rbp-0x68]
181    7b09:   49 83 c0 04             add    r8,0x4
182    7b0d:   48 01 5d b8             add    QWORD PTR [rbp-0x48],rbx
183    7b11:   4c 3b 45 b0             cmp    r8,QWORD PTR [rbp-0x50]
184    7b15:   0f 85 35 ff ff ff       jne    7a50 <convolution2D+0x1d0>
185    7b1b:   8b 9d 74 ff ff ff       mov    ebx,DWORD PTR [rbp-0x8c]
186    7b21:   8b 45 94                mov    eax,DWORD PTR [rbp-0x6c]
187    7b24:   48 8b 8d 48 ff ff ff    mov    rcx,QWORD PTR [rbp-0xb8]
188    7b2b:   01 5d d0                add    DWORD PTR [rbp-0x30],ebx
189    7b2e:   48 01 4d b0             add    QWORD PTR [rbp-0x50],rcx
190    7b32:   01 5d 84                add    DWORD PTR [rbp-0x7c],ebx
191    7b35:   48 8b 8d 58 ff ff ff    mov    rcx,QWORD PTR [rbp-0xa8]
192    7b3c:   41 01 da                add    r10d,ebx
193    7b3f:   48 01 8d 78 ff ff ff    add    QWORD PTR [rbp-0x88],rcx
194    7b46:   ff c0                   inc    eax
195    7b48:   39 85 38 ff ff ff       cmp    DWORD PTR [rbp-0xc8],eax
196    7b4e:   74 08                   je     7b58 <convolution2D+0x2d8>
197    7b50:   89 45 94                mov    DWORD PTR [rbp-0x6c],eax
198    7b53:   e9 ac fe ff ff          jmp    7a04 <convolution2D+0x184>
199    7b58:   8b 4d 90                mov    ecx,DWORD PTR [rbp-0x70]
200    7b5b:   48 8b b5 30 ff ff ff    mov    rsi,QWORD PTR [rbp-0xd0]
201    7b62:   01 5d d4                add    DWORD PTR [rbp-0x2c],ebx
202    7b65:   01 5d ac                add    DWORD PTR [rbp-0x54],ebx
203    7b68:   01 5d 80                add    DWORD PTR [rbp-0x80],ebx
204    7b6b:   48 01 75 88             add    QWORD PTR [rbp-0x78],rsi
205    7b6f:   8d 41 01                lea    eax,[rcx+0x1]
206    7b72:   39 4d 94                cmp    DWORD PTR [rbp-0x6c],ecx
207    7b75:   74 08                   je     7b7f <convolution2D+0x2ff>
208    7b77:   89 45 90                mov    DWORD PTR [rbp-0x70],eax
209    7b7a:   e9 33 fe ff ff          jmp    79b2 <convolution2D+0x132>
210    7b7f:   48 81 c4 b0 00 00 00    add    rsp,0xb0
211    7b86:   5b                      pop    rbx
212    7b87:   41 5c                   pop    r12
213    7b89:   41 5d                   pop    r13
214    7b8b:   41 5e                   pop    r14
215    7b8d:   41 5f                   pop    r15
216    7b8f:   5d                      pop    rbp
217    7b90:   c3                      ret    
218    7b91:   66 66 2e 0f 1f 84 00 00 00 00 00    data16 nop WORD PTR cs:[rax+rax*1+0x0]
219    7b9c:   0f 1f 40 00             nop    DWORD PTR [rax+0x0]
220
221float (*arr)[x][y] = calloc(z, sizeof *arr);
222==430300== Cachegrind, a cache and branch-prediction profiler
223==430300== Copyright (C) 2002-2017, and GNU GPL'd, by Nicholas Nethercote et al.
224==430300== Using Valgrind-3.15.0 and LibVEX; rerun with -h for copyright info
225==430300== Command: ./EmbeddedNet test 1
226==430300== Parent PID: 170008
227==430300== 
228--430300-- warning: L3 cache found, using its data for the LL simulation.
229==430300== 
230==430300== I   refs:      6,369,594,192
231==430300== I1  misses:            4,271
232==430300== LLi misses:            2,442
233==430300== I1  miss rate:          0.00%
234==430300== LLi miss rate:          0.00%
235==430300== 
236==430300== D   refs:      2,064,233,110  (1,359,003,131 rd   + 705,229,979 wr)
237==430300== D1  misses:       34,476,969  (   19,010,839 rd   +  15,466,130 wr)
238==430300== LLd misses:        5,311,277  (    1,603,955 rd   +   3,707,322 wr)
239==430300== D1  miss rate:           1.7% (          1.4%     +         2.2%  )
240==430300== LLd miss rate:           0.3% (          0.1%     +         0.5%  )
241==430300== 
242==430300== LL refs:          34,481,240  (   19,015,110 rd   +  15,466,130 wr)
243==430300== LL misses:         5,313,719  (    1,606,397 rd   +   3,707,322 wr)
244==430300== LL miss rate:            0.1% (          0.0%     +         0.5%  )
245float fma_scl(float a, float b, float c) {
246    return a * b + c;
247}
248
249fma_scl:
250        vfmadd132ss     xmm0, xmm2, xmm1
251        ret
252typedef float Vec __attribute__((vector_size(32), aligned(32)));
253Vec fma_vec(Vec a, Vec b, Vec c) {
254    return a * b + c;
255}
256
257fma_vec:
258        vfmadd132ps     ymm0, ymm2, ymm1
259        ret
260typedef struct {
261    float f[8];
262} Vec_;
263
264Vec_ fma_vec_(Vec_ a, Vec_ b, Vec_ c) {
265    Vec_ r;
266    for (unsigned i = 0; i < 8; ++i) {
267        r.f[i] = a.f[i] * b.f[i] + c.f[i];
268    }
269    return r;
270}
271void loopadd_scl(float *restrict a, float *restrict b, float *restrict c, unsigned n) {
272    for (unsigned i = 0; i < n; ++i) {
273        a[i] = fma_scl(b[i], c[i], a[i]);
274    }
275}
276loopadd_scl:
277        test    ecx, ecx
278        je      .L25
279        lea     eax, [rcx-1]
280        cmp     eax, 6
281        jbe     .L13
282        mov     r8d, ecx
283        xor     eax, eax
284        shr     r8d, 3
285        sal     r8, 5
286.L9:
287        vmovups ymm1, YMMWORD PTR [rdi+rax]
288        vmovups ymm0, YMMWORD PTR [rdx+rax]
289        vfmadd132ps     ymm0, ymm1, YMMWORD PTR [rsi+rax]
290        vmovups YMMWORD PTR [rdi+rax], ymm0
291        add     rax, 32
292        cmp     r8, rax
293        jne     .L9
294        mov     eax, ecx
295        and     eax, -8
296        test    cl, 7
297        je      .L26
298        vzeroupper
299.L8:
300        mov     r9d, ecx
301        sub     r9d, eax
302        lea     r8d, [r9-1]
303        cmp     r8d, 2
304        jbe     .L11
305        mov     r8d, eax
306        sal     r8, 2
307        lea     r10, [rdi+r8]
308        vmovups xmm0, XMMWORD PTR [rdx+r8]
309        vmovups xmm2, XMMWORD PTR [r10]
310        vfmadd132ps     xmm0, xmm2, XMMWORD PTR [rsi+r8]
311        mov     r8d, r9d
312        and     r8d, -4
313        add     eax, r8d
314        and     r9d, 3
315        vmovups XMMWORD PTR [r10], xmm0
316        je      .L25
317.L11:
318        mov     r8d, eax
319        sal     r8, 2
320        lea     r9, [rdi+r8]
321        vmovss  xmm0, DWORD PTR [rdx+r8]
322        vmovss  xmm3, DWORD PTR [r9]
323        vfmadd132ss     xmm0, xmm3, DWORD PTR [rsi+r8]
324        lea     r8d, [rax+1]
325        vmovss  DWORD PTR [r9], xmm0
326        cmp     r8d, ecx
327        jnb     .L25
328        sal     r8, 2
329        add     eax, 2
330        lea     r9, [rdi+r8]
331        vmovss  xmm0, DWORD PTR [rsi+r8]
332        vmovss  xmm4, DWORD PTR [r9]
333        vfmadd132ss     xmm0, xmm4, DWORD PTR [rdx+r8]
334        vmovss  DWORD PTR [r9], xmm0
335        cmp     eax, ecx
336        jnb     .L25
337        sal     rax, 2
338        add     rdi, rax
339        vmovss  xmm0, DWORD PTR [rdx+rax]
340        vmovss  xmm5, DWORD PTR [rdi]
341        vfmadd132ss     xmm0, xmm5, DWORD PTR [rsi+rax]
342        vmovss  DWORD PTR [rdi], xmm0
343.L25:
344        ret
345.L26:
346        vzeroupper
347        ret
348.L13:
349        xor     eax, eax
350        jmp     .L8
351void loopadd_vec(Vec *restrict a, Vec *restrict b, Vec *restrict c, unsigned n) {
352    n /= 8;
353    for (unsigned i = 0; i < n; ++i) {
354        a[i] = fma_vec(b[i], c[i], a[i]);
355    }
356}
357
358loopadd_vec:
359        shr     ecx, 3
360        je      .L34
361        mov     ecx, ecx
362        xor     eax, eax
363        sal     rcx, 5
364.L29:
365        vmovaps ymm1, YMMWORD PTR [rdi+rax]
366        vmovaps ymm0, YMMWORD PTR [rdx+rax]
367        vfmadd132ps     ymm0, ymm1, YMMWORD PTR [rsi+rax]
368        vmovaps YMMWORD PTR [rdi+rax], ymm0
369        add     rax, 32
370        cmp     rcx, rax
371        jne     .L29
372        vzeroupper
373.L34:
374        ret
375}
376

1$ dotnet --version
2$ dotnet --list-sdks
3$ dotnet --list-runtimes
4
5$ dotnet new globaljson                                     # Create a "global.json" file in your current directory folder.  
6$ dotnet new globaljson --sdk-version 3.1.416 --force       # Change the current SDK version of your project.
7
8$ dotnet new -i Microsoft.DotNet.Web.Spa.ProjectTemplates
9$ dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
10
11# First, create a Q# application and a .NET host, and then make a call to Q# from the host.
12# Create a project for your Q# library and for the .NET host that will call
13# into the operations and functions defined in your Q# library.
14
15$ dotnet new classlib -lang Q# -o quantum                         # Create a new Q# class-library project.
16$ dotnet new console -lang C# -o host                             # Create a new C# console project.
17
18$ cd host                                                         # Navigate into the C# host directory.
19$ dotnet add reference ../quantum/quantum.csproj                  # Add your Q# class-library project as a reference to your C# console project.
20
21$ cd ..                                                           # Exit the C# host directory.
22$ dotnet new sln -n quantum-dotnet                                # Create a new solution for both projects.
23$ dotnet sln quantum-dotnet.sln add ./quantum/quantum.csproj      # Add the Q# class-library project to the solution.
24$ dotnet sln quantum-dotnet.sln add ./host/host.csproj            # Add the C# host project to the solution.
25
26# The -o or --output command specifies the location to place the generated output.
27$ dotnet --list-sdks
283.1.416 [C:\Program Files\dotnet\sdk]
295.0.404 [C:\Program Files\dotnet\sdk]
30
31$ dotnet --version  
323.1.416
33
34$ dotnet restore
35  Determining projects to restore...
36  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\quantum\quantum.csproj (in 484 ms).
37  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\host\host.csproj (in 782 ms).
38dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
39

1$ dotnet --version
2$ dotnet --list-sdks
3$ dotnet --list-runtimes
4
5$ dotnet new globaljson                                     # Create a "global.json" file in your current directory folder.  
6$ dotnet new globaljson --sdk-version 3.1.416 --force       # Change the current SDK version of your project.
7
8$ dotnet new -i Microsoft.DotNet.Web.Spa.ProjectTemplates
9$ dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
10
11# First, create a Q# application and a .NET host, and then make a call to Q# from the host.
12# Create a project for your Q# library and for the .NET host that will call
13# into the operations and functions defined in your Q# library.
14
15$ dotnet new classlib -lang Q# -o quantum                         # Create a new Q# class-library project.
16$ dotnet new console -lang C# -o host                             # Create a new C# console project.
17
18$ cd host                                                         # Navigate into the C# host directory.
19$ dotnet add reference ../quantum/quantum.csproj                  # Add your Q# class-library project as a reference to your C# console project.
20
21$ cd ..                                                           # Exit the C# host directory.
22$ dotnet new sln -n quantum-dotnet                                # Create a new solution for both projects.
23$ dotnet sln quantum-dotnet.sln add ./quantum/quantum.csproj      # Add the Q# class-library project to the solution.
24$ dotnet sln quantum-dotnet.sln add ./host/host.csproj            # Add the C# host project to the solution.
25
26# The -o or --output command specifies the location to place the generated output.
27$ dotnet --list-sdks
283.1.416 [C:\Program Files\dotnet\sdk]
295.0.404 [C:\Program Files\dotnet\sdk]
30
31$ dotnet --version  
323.1.416
33
34$ dotnet restore
35  Determining projects to restore...
36  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\quantum\quantum.csproj (in 484 ms).
37  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\host\host.csproj (in 782 ms).
38dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
39dotnet new -i "Microsoft.Quantum.ProjectTemplates"
40

1$ dotnet --version
2$ dotnet --list-sdks
3$ dotnet --list-runtimes
4
5$ dotnet new globaljson                                     # Create a "global.json" file in your current directory folder.  
6$ dotnet new globaljson --sdk-version 3.1.416 --force       # Change the current SDK version of your project.
7
8$ dotnet new -i Microsoft.DotNet.Web.Spa.ProjectTemplates
9$ dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
10
11# First, create a Q# application and a .NET host, and then make a call to Q# from the host.
12# Create a project for your Q# library and for the .NET host that will call
13# into the operations and functions defined in your Q# library.
14
15$ dotnet new classlib -lang Q# -o quantum                         # Create a new Q# class-library project.
16$ dotnet new console -lang C# -o host                             # Create a new C# console project.
17
18$ cd host                                                         # Navigate into the C# host directory.
19$ dotnet add reference ../quantum/quantum.csproj                  # Add your Q# class-library project as a reference to your C# console project.
20
21$ cd ..                                                           # Exit the C# host directory.
22$ dotnet new sln -n quantum-dotnet                                # Create a new solution for both projects.
23$ dotnet sln quantum-dotnet.sln add ./quantum/quantum.csproj      # Add the Q# class-library project to the solution.
24$ dotnet sln quantum-dotnet.sln add ./host/host.csproj            # Add the C# host project to the solution.
25
26# The -o or --output command specifies the location to place the generated output.
27$ dotnet --list-sdks
283.1.416 [C:\Program Files\dotnet\sdk]
295.0.404 [C:\Program Files\dotnet\sdk]
30
31$ dotnet --version  
323.1.416
33
34$ dotnet restore
35  Determining projects to restore...
36  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\quantum\quantum.csproj (in 484 ms).
37  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\host\host.csproj (in 782 ms).
38dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
39dotnet new -i "Microsoft.Quantum.ProjectTemplates"
40dotnet new --update-apply
41

1$ dotnet --version
2$ dotnet --list-sdks
3$ dotnet --list-runtimes
4
5$ dotnet new globaljson                                     # Create a "global.json" file in your current directory folder.  
6$ dotnet new globaljson --sdk-version 3.1.416 --force       # Change the current SDK version of your project.
7
8$ dotnet new -i Microsoft.DotNet.Web.Spa.ProjectTemplates
9$ dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
10
11# First, create a Q# application and a .NET host, and then make a call to Q# from the host.
12# Create a project for your Q# library and for the .NET host that will call
13# into the operations and functions defined in your Q# library.
14
15$ dotnet new classlib -lang Q# -o quantum                         # Create a new Q# class-library project.
16$ dotnet new console -lang C# -o host                             # Create a new C# console project.
17
18$ cd host                                                         # Navigate into the C# host directory.
19$ dotnet add reference ../quantum/quantum.csproj                  # Add your Q# class-library project as a reference to your C# console project.
20
21$ cd ..                                                           # Exit the C# host directory.
22$ dotnet new sln -n quantum-dotnet                                # Create a new solution for both projects.
23$ dotnet sln quantum-dotnet.sln add ./quantum/quantum.csproj      # Add the Q# class-library project to the solution.
24$ dotnet sln quantum-dotnet.sln add ./host/host.csproj            # Add the C# host project to the solution.
25
26# The -o or --output command specifies the location to place the generated output.
27$ dotnet --list-sdks
283.1.416 [C:\Program Files\dotnet\sdk]
295.0.404 [C:\Program Files\dotnet\sdk]
30
31$ dotnet --version  
323.1.416
33
34$ dotnet restore
35  Determining projects to restore...
36  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\quantum\quantum.csproj (in 484 ms).
37  Restored C:\Users\Muhy\Dropbox\Q\Q#-programs\Qtest\host\host.csproj (in 782 ms).
38dotnet new -i "Microsoft.Quantum.ProjectTemplates::0.2-*"
39dotnet new -i "Microsoft.Quantum.ProjectTemplates"
40dotnet new --update-apply
41# Install the .NET Framework and Core as well as the Quantum Development Kit (QDK)
42# plus the Q# Interoperability Tools and also C# and Python extensions for VS Code.
43
44$ dotnet --version
45$ dotnet --list-sdks
46$ dotnet --list-runtimes
47
48$ dotnet new globaljson                                             # Create a "global.json" file in your current directory folder.  
49$ dotnet new globaljson --sdk-version 3.1.416 --force               # Change the current SDK version of your project to .NET 3.1.416
50
51$ dotnet new -i Microsoft.DotNet.Web.Spa.ProjectTemplates
52$ dotnet new -i "Microsoft.Quantum.ProjectTemplates"
53$ dotnet new --update-apply
54
55# First, create a Q# application and a .NET host, and then make a call to Q# from the host.
56# Create a project for your Q# library and for the .NET host that will call
57# into the operations and functions defined in your Q# library.
58
59$ dotnet new classlib -lang Q# -o quantum                           # Create a new Q# class-library project.
60$ dotnet new console -lang C# -o host                               # Create a new C# console project.
61                                                                    # The -o or --output command specifies the location to place the generated output.
62
63$ cd host                                                           # Navigate into the C# host directory.
64$ dotnet add reference ../quantum/quantum.csproj                    # Add your Q# class-library project as a reference to your C# console project.
65
66$ cd ..                                                             # Exit the C# host directory.
67$ dotnet new sln -n quantum-dotnet                                  # Create a new solution for both projects.
68$ dotnet sln quantum-dotnet.sln add ./quantum/quantum.csproj        # Add the Q# class-library project to the solution.
69$ dotnet sln quantum-dotnet.sln add ./host/host.csproj              # Add the C# host project to the solution.
70
71$ cd host
72$ dotnet build
73$ dotnet run
74

1Rows: 42
2Columns: 5
3$ district       <dbl> 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 1~
4$ party          <chr> "REPUBLICAN", "DEMOCRAT", "INDEPENDENT", "REPUBLICAN", "DEMOCRAT", "DEMOCRAT", ~
5$ candidatevotes <dbl> 41708, 235394, 4829, 33381, 318176, 123933, 165826, 12755, 6210, 181603, 11524,~
6$ totalvotes     <dbl> 277102, 277102, 356386, 356386, 356386, 302514, 302514, 302514, 303980, 303980,~
7$ pct_votes      <dbl> 15.051497, 84.948503, 1.354991, 9.366530, 89.278479, 40.967691, 54.815975, 4.21~
8PA2012_house_dem <- PA2012_house %>% filter(party == "DEMOCRAT") 
9obs_median <- median(PA2012_house_dem$pct_votes)
10obs_mean <- mean(PA2012_house_dem$pct_votes)
11obs_median - obs_mean
12Rows: 18
13Columns: 5
14$ district       <dbl> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18
15$ party          <chr> "DEMOCRAT", "DEMOCRAT", "DEMOCRAT", "DEMOCRAT", "DEMOCRAT", "DEMOCRAT", "DEMOCRAT", "DEMOCRAT", "DEMOCR~
16$ candidatevotes <dbl> 235394, 318176, 123933, 104643, 104725, 143803, 143509, 152859, 105128, 94227, 118231, 163589, 209901, ~
17$ totalvotes     <dbl> 277102, 356386, 302514, 303980, 282465, 335528, 353451, 352238, 274305, 273790, 285198, 338941, 303819,~
18$ pct_votes      <dbl> 84.94850, 89.27848, 40.96769, 34.42430, 37.07539, 42.85872, 40.60223, 43.39651, 38.32522, 34.41579, 41.~
19require(mosaic) # for the tally() function
20data <- PA2012_reduced_dem
21B <- 100
22samples_diff <- vector("numeric", B)
23samples_mean <- vector("numeric", B)
24samples_median <- vector("numeric", B)
25
26for(samp in 1:B) {
27data_w_sample <- mutate(data, sample_vote = tally(sample(district, sum(candidatevotes),replace=T, prob = totalvotes)))
28  data_w_sample <- mutate(data_w_sample, sample_vote_pct = (sample_vote / totalvotes *100))
29  mean_sample <- weighted.mean(data_w_sample$sample_vote_pct, w = data_w_sample$totalvotes)
30  median_sample <- median(data_w_sample$sample_vote_pct)
31  diff_mean_median <- mean_sample - median_sample
32  samples_diff[samp] <- diff_mean_median
33  samples_mean[samp] <- mean_sample
34  samples_median[samp] <- median_sample
35}
36
37samples <- data.frame(samples_mean,samples_median,samples_diff)
38data <- house_2012_reduced 
39# created with dplyr, contains total and percentage of votes
40# for Democrats and Republicans.
41B <- 100000
42del_districts <- 18 # 18 districts in PA
43samples_diff <- vector("numeric", B)
44samples_mean <- vector("numeric", B)
45samples_median <- vector("numeric", B)
46
47for(samp in 1:B) {
48  sample_delegation <- sample_n(data, del_districts)
49  sample_delegation_pct_dem_mean <- weighted.mean(sample_delegation$pct_dem_votes, w = sample_delegation$total_votes)
50  sample_delegation_pct_dem_median <- median(sample_delegation$pct_dem_votes)
51  if(near(mean_dem_pct_PA, sample_delegation_pct_dem_mean, 1)){
52    samples_mean[samp] <- sample_delegation_pct_dem_mean
53    samples_median[samp] <- sample_delegation_pct_dem_median
54    samples_diff[samp] <- (sample_delegation_pct_dem_mean - sample_delegation_pct_dem_median)
55  }
56}
57
58samples <- data.frame(samples_mean,samples_median,samples_diff)
59samples <- filter_all(samples, any_vars(. != 0))
60quantile(samples$samples_median, c(0.025,0.975))
61