1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
|
/*
* Copyright (c) 2023, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <stdbool.h>
#include <assert.h>
#include <immintrin.h>
#include "config/av1_rtcd.h"
#include "av1/encoder/ml.h"
#include "av1/encoder/x86/ml_sse3.h"
#define CALC_OUTPUT_FOR_2ROWS \
const int index = weight_idx + (2 * i * tot_num_inputs); \
const __m256 weight0 = _mm256_loadu_ps(&weights[index]); \
const __m256 weight1 = _mm256_loadu_ps(&weights[index + tot_num_inputs]); \
const __m256 mul0 = _mm256_mul_ps(inputs256, weight0); \
const __m256 mul1 = _mm256_mul_ps(inputs256, weight1); \
hadd[i] = _mm256_hadd_ps(mul0, mul1);
static INLINE void nn_propagate_8to1(
const float *const inputs, const float *const weights,
const float *const bias, int num_inputs_to_process, int tot_num_inputs,
int num_outputs, float *const output_nodes, int is_clip_required) {
// Process one output row at a time.
for (int out = 0; out < num_outputs; out++) {
__m256 in_result = _mm256_setzero_ps();
float bias_val = bias[out];
for (int in = 0; in < num_inputs_to_process; in += 8) {
const __m256 inputs256 = _mm256_loadu_ps(&inputs[in]);
const int weight_idx = in + (out * tot_num_inputs);
const __m256 weight0 = _mm256_loadu_ps(&weights[weight_idx]);
const __m256 mul0 = _mm256_mul_ps(inputs256, weight0);
in_result = _mm256_add_ps(in_result, mul0);
}
const __m128 low_128 = _mm256_castps256_ps128(in_result);
const __m128 high_128 = _mm256_extractf128_ps(in_result, 1);
const __m128 sum_par_0 = _mm_add_ps(low_128, high_128);
const __m128 sum_par_1 = _mm_hadd_ps(sum_par_0, sum_par_0);
const __m128 sum_tot =
_mm_add_ps(_mm_shuffle_ps(sum_par_1, sum_par_1, 0x99), sum_par_1);
bias_val += _mm_cvtss_f32(sum_tot);
if (is_clip_required) bias_val = AOMMAX(bias_val, 0);
output_nodes[out] = bias_val;
}
}
static INLINE void nn_propagate_8to4(
const float *const inputs, const float *const weights,
const float *const bias, int num_inputs_to_process, int tot_num_inputs,
int num_outputs, float *const output_nodes, int is_clip_required) {
__m256 hadd[2];
for (int out = 0; out < num_outputs; out += 4) {
__m128 bias_reg = _mm_loadu_ps(&bias[out]);
__m128 in_result = _mm_setzero_ps();
for (int in = 0; in < num_inputs_to_process; in += 8) {
const __m256 inputs256 = _mm256_loadu_ps(&inputs[in]);
const int weight_idx = in + (out * tot_num_inputs);
// Process two output row at a time.
for (int i = 0; i < 2; i++) {
CALC_OUTPUT_FOR_2ROWS
}
const __m256 sum_par = _mm256_hadd_ps(hadd[0], hadd[1]);
const __m128 low_128 = _mm256_castps256_ps128(sum_par);
const __m128 high_128 = _mm256_extractf128_ps(sum_par, 1);
const __m128 result = _mm_add_ps(low_128, high_128);
in_result = _mm_add_ps(in_result, result);
}
in_result = _mm_add_ps(in_result, bias_reg);
if (is_clip_required) in_result = _mm_max_ps(in_result, _mm_setzero_ps());
_mm_storeu_ps(&output_nodes[out], in_result);
}
}
static INLINE void nn_propagate_8to8(
const float *const inputs, const float *const weights,
const float *const bias, int num_inputs_to_process, int tot_num_inputs,
int num_outputs, float *const output_nodes, int is_clip_required) {
__m256 hadd[4];
for (int out = 0; out < num_outputs; out += 8) {
__m256 bias_reg = _mm256_loadu_ps(&bias[out]);
__m256 in_result = _mm256_setzero_ps();
for (int in = 0; in < num_inputs_to_process; in += 8) {
const __m256 inputs256 = _mm256_loadu_ps(&inputs[in]);
const int weight_idx = in + (out * tot_num_inputs);
// Process two output rows at a time.
for (int i = 0; i < 4; i++) {
CALC_OUTPUT_FOR_2ROWS
}
const __m256 hh0 = _mm256_hadd_ps(hadd[0], hadd[1]);
const __m256 hh1 = _mm256_hadd_ps(hadd[2], hadd[3]);
__m256 ht_0 = _mm256_permute2f128_ps(hh0, hh1, 0x20);
__m256 ht_1 = _mm256_permute2f128_ps(hh0, hh1, 0x31);
__m256 result = _mm256_add_ps(ht_0, ht_1);
in_result = _mm256_add_ps(in_result, result);
}
in_result = _mm256_add_ps(in_result, bias_reg);
if (is_clip_required)
in_result = _mm256_max_ps(in_result, _mm256_setzero_ps());
_mm256_storeu_ps(&output_nodes[out], in_result);
}
}
static INLINE void nn_propagate_input_multiple_of_8(
const float *const inputs, const float *const weights,
const float *const bias, int num_inputs_to_process, int tot_num_inputs,
bool is_output_layer, int num_outputs, float *const output_nodes) {
// The saturation of output is considered for hidden layer which is not equal
// to final hidden layer.
const int is_clip_required =
!is_output_layer && num_inputs_to_process == tot_num_inputs;
if (num_outputs % 8 == 0) {
nn_propagate_8to8(inputs, weights, bias, num_inputs_to_process,
tot_num_inputs, num_outputs, output_nodes,
is_clip_required);
} else if (num_outputs % 4 == 0) {
nn_propagate_8to4(inputs, weights, bias, num_inputs_to_process,
tot_num_inputs, num_outputs, output_nodes,
is_clip_required);
} else {
nn_propagate_8to1(inputs, weights, bias, num_inputs_to_process,
tot_num_inputs, num_outputs, output_nodes,
is_clip_required);
}
}
void av1_nn_predict_avx2(const float *input_nodes,
const NN_CONFIG *const nn_config, int reduce_prec,
float *const output) {
float buf[2][NN_MAX_NODES_PER_LAYER];
int buf_index = 0;
int num_inputs = nn_config->num_inputs;
assert(num_inputs > 0 && num_inputs <= NN_MAX_NODES_PER_LAYER);
for (int layer = 0; layer <= nn_config->num_hidden_layers; layer++) {
const float *layer_weights = nn_config->weights[layer];
const float *layer_bias = nn_config->bias[layer];
bool is_output_layer = layer == nn_config->num_hidden_layers;
float *const output_nodes = is_output_layer ? output : &buf[buf_index][0];
const int num_outputs = is_output_layer
? nn_config->num_outputs
: nn_config->num_hidden_nodes[layer];
assert(num_outputs > 0 && num_outputs <= NN_MAX_NODES_PER_LAYER);
// Process input multiple of 8 using AVX2 intrinsic.
if (num_inputs % 8 == 0) {
nn_propagate_input_multiple_of_8(input_nodes, layer_weights, layer_bias,
num_inputs, num_inputs, is_output_layer,
num_outputs, output_nodes);
} else {
// When number of inputs is not multiple of 8, use hybrid approach of AVX2
// and SSE3 based on the need.
const int in_mul_8 = num_inputs / 8;
const int num_inputs_to_process = in_mul_8 * 8;
int bias_is_considered = 0;
if (in_mul_8) {
nn_propagate_input_multiple_of_8(
input_nodes, layer_weights, layer_bias, num_inputs_to_process,
num_inputs, is_output_layer, num_outputs, output_nodes);
bias_is_considered = 1;
}
const float *out_temp = bias_is_considered ? output_nodes : layer_bias;
const int input_remaining = num_inputs % 8;
if (input_remaining % 4 == 0 && num_outputs % 8 == 0) {
for (int out = 0; out < num_outputs; out += 8) {
__m128 out_h = _mm_loadu_ps(&out_temp[out + 4]);
__m128 out_l = _mm_loadu_ps(&out_temp[out]);
for (int in = in_mul_8 * 8; in < num_inputs; in += 4) {
av1_nn_propagate_4to8_sse3(&input_nodes[in],
&layer_weights[out * num_inputs + in],
&out_h, &out_l, num_inputs);
}
if (!is_output_layer) {
const __m128 zero = _mm_setzero_ps();
out_h = _mm_max_ps(out_h, zero);
out_l = _mm_max_ps(out_l, zero);
}
_mm_storeu_ps(&output_nodes[out + 4], out_h);
_mm_storeu_ps(&output_nodes[out], out_l);
}
} else if (input_remaining % 4 == 0 && num_outputs % 4 == 0) {
for (int out = 0; out < num_outputs; out += 4) {
__m128 outputs = _mm_loadu_ps(&out_temp[out]);
for (int in = in_mul_8 * 8; in < num_inputs; in += 4) {
av1_nn_propagate_4to4_sse3(&input_nodes[in],
&layer_weights[out * num_inputs + in],
&outputs, num_inputs);
}
if (!is_output_layer) outputs = _mm_max_ps(outputs, _mm_setzero_ps());
_mm_storeu_ps(&output_nodes[out], outputs);
}
} else if (input_remaining % 4 == 0) {
for (int out = 0; out < num_outputs; out++) {
__m128 outputs = _mm_load1_ps(&out_temp[out]);
for (int in = in_mul_8 * 8; in < num_inputs; in += 4) {
av1_nn_propagate_4to1_sse3(&input_nodes[in],
&layer_weights[out * num_inputs + in],
&outputs);
}
if (!is_output_layer) outputs = _mm_max_ps(outputs, _mm_setzero_ps());
output_nodes[out] = _mm_cvtss_f32(outputs);
}
} else {
// Use SSE instructions for scalar operations to avoid the latency
// of swapping between SIMD and FPU modes.
for (int out = 0; out < num_outputs; out++) {
__m128 outputs = _mm_load1_ps(&out_temp[out]);
for (int in_node = in_mul_8 * 8; in_node < num_inputs; in_node++) {
__m128 input = _mm_load1_ps(&input_nodes[in_node]);
__m128 weight =
_mm_load1_ps(&layer_weights[num_inputs * out + in_node]);
outputs = _mm_add_ps(outputs, _mm_mul_ps(input, weight));
}
if (!is_output_layer) outputs = _mm_max_ps(outputs, _mm_setzero_ps());
output_nodes[out] = _mm_cvtss_f32(outputs);
}
}
}
// Before processing the next layer, treat the output of current layer as
// input to next layer.
input_nodes = output_nodes;
num_inputs = num_outputs;
buf_index = 1 - buf_index;
}
if (reduce_prec) av1_nn_output_prec_reduce(output, nn_config->num_outputs);
}
|
