xref: /aosp_15_r20/external/libaom/av1/encoder/x86/ml_sse3.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2018, Alliance for Open Media. All rights reserved.
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #include <stdbool.h>
13 #include <assert.h>
14 
15 #include "config/av1_rtcd.h"
16 #include "av1/encoder/ml.h"
17 #include "av1/encoder/x86/ml_sse3.h"
18 
19 // In order to avoid the high-latency of swapping between FPU and SIMD
20 // operations, we keep the result in a 128-bit register even though we only
21 // care about a single value.
nn_propagate_8to1(const float * const inputs,const float * const weights,__m128 * const output)22 static void nn_propagate_8to1(const float *const inputs,
23                               const float *const weights,
24                               __m128 *const output) {
25   const __m128 inputs_h = _mm_loadu_ps(&inputs[4]);
26   const __m128 inputs_l = _mm_loadu_ps(inputs);
27 
28   const __m128 weights_h = _mm_loadu_ps(&weights[4]);
29   const __m128 weights_l = _mm_loadu_ps(weights);
30 
31   const __m128 mul_h = _mm_mul_ps(inputs_h, weights_h);
32   const __m128 mul_l = _mm_mul_ps(inputs_l, weights_l);
33   // [7 6 5 4] [3 2 1 0] (weight and input indices)
34 
35   const __m128 vadd = _mm_add_ps(mul_l, mul_h);
36   // [7+3 6+2 5+1 4+0]
37   const __m128 hadd1 = _mm_hadd_ps(vadd, vadd);
38   // [7+6+3+2 5+4+1+0 7+6+3+2 5+4+1+0]
39   const __m128 hadd2 = _mm_hadd_ps(hadd1, hadd1);
40   // [7+6+5+4+3+2+1+0 7+6+5+4+3+2+1+0 7+6+5+4+3+2+1+0 7+6+5+4+3+2+1+0]
41   *output = _mm_add_ps(*output, hadd2);
42 }
43 
av1_nn_propagate_4to1_sse3(const float * const inputs,const float * const weights,__m128 * const output)44 void av1_nn_propagate_4to1_sse3(const float *const inputs,
45                                 const float *const weights,
46                                 __m128 *const output) {
47   const __m128 inputs128 = _mm_loadu_ps(inputs);
48 
49   const __m128 weights128 = _mm_loadu_ps(weights);
50 
51   const __m128 mul = _mm_mul_ps(inputs128, weights128);
52   // [3 2 1 0] (weight and input indices)
53 
54   const __m128 hadd1 = _mm_hadd_ps(mul, mul);
55   // [3+2 1+0 3+2 1+0]
56   const __m128 hadd2 = _mm_hadd_ps(hadd1, hadd1);
57   // [3+2+1+0 3+2+1+0 3+2+1+0 3+2+1+0]
58   *output = _mm_add_ps(*output, hadd2);
59 }
60 
av1_nn_propagate_4to4_sse3(const float * const inputs,const float * const weights,__m128 * const outputs,const int num_inputs)61 void av1_nn_propagate_4to4_sse3(const float *const inputs,
62                                 const float *const weights,
63                                 __m128 *const outputs, const int num_inputs) {
64   const __m128 inputs128 = _mm_loadu_ps(inputs);
65 
66   __m128 hadd[2];
67   for (int i = 0; i < 2; i++) {  // For each pair of outputs
68     const __m128 weight0 = _mm_loadu_ps(&weights[2 * i * num_inputs]);
69     const __m128 mul0 = _mm_mul_ps(weight0, inputs128);
70     const __m128 weight1 = _mm_loadu_ps(&weights[(2 * i + 1) * num_inputs]);
71     const __m128 mul1 = _mm_mul_ps(weight1, inputs128);
72     hadd[i] = _mm_hadd_ps(mul0, mul1);
73   }
74   // hadd[0] = [7+6 5+4 3+2 1+0] (weight indices)
75   // hadd[1] = [15+14 13+12 11+10 9+8]
76 
77   const __m128 hh = _mm_hadd_ps(hadd[0], hadd[1]);
78   // [15+14+13+12 11+10+9+8 7+6+5+4 3+2+1+0]
79 
80   *outputs = _mm_add_ps(*outputs, hh);
81 }
82 
av1_nn_propagate_4to8_sse3(const float * const inputs,const float * const weights,__m128 * const out_h,__m128 * const out_l,const int num_inputs)83 void av1_nn_propagate_4to8_sse3(const float *const inputs,
84                                 const float *const weights, __m128 *const out_h,
85                                 __m128 *const out_l, const int num_inputs) {
86   const __m128 inputs128 = _mm_loadu_ps(inputs);
87 
88   __m128 hadd[4];
89   for (int i = 0; i < 4; i++) {  // For each pair of outputs
90     const __m128 weight0 = _mm_loadu_ps(&weights[2 * i * num_inputs]);
91     const __m128 weight1 = _mm_loadu_ps(&weights[(2 * i + 1) * num_inputs]);
92     const __m128 mul0 = _mm_mul_ps(inputs128, weight0);
93     const __m128 mul1 = _mm_mul_ps(inputs128, weight1);
94     hadd[i] = _mm_hadd_ps(mul0, mul1);
95   }
96   // hadd[0] = [7+6 5+4 3+2 1+0] (weight indices)
97   // hadd[1] = [15+14 13+12 11+10 9+8]
98   // hadd[2] = [23+22 21+20 19+18 17+16]
99   // hadd[3] = [31+30 29+28 27+26 25+24]
100 
101   const __m128 hh0 = _mm_hadd_ps(hadd[0], hadd[1]);
102   // [15+14+13+12 11+10+9+8 7+6+5+4 3+2+1+0]
103   const __m128 hh1 = _mm_hadd_ps(hadd[2], hadd[3]);
104   // [31+30+29+28 27+26+25+24 23+22+21+20 19+18+17+16]
105 
106   *out_h = _mm_add_ps(*out_h, hh1);
107   *out_l = _mm_add_ps(*out_l, hh0);
108 }
109 
nn_propagate_8to4(const float * const inputs,const float * const weights,__m128 * const outputs,const int num_inputs)110 static void nn_propagate_8to4(const float *const inputs,
111                               const float *const weights, __m128 *const outputs,
112                               const int num_inputs) {
113   const __m128 inputs_h = _mm_loadu_ps(inputs + 4);
114   const __m128 inputs_l = _mm_loadu_ps(inputs);
115   // [7 6 5 4] [3 2 1 0] (input indices)
116 
117   __m128 add[4];
118   for (int i = 0; i < 4; i++) {  // For each output:
119     const __m128 weight_h = _mm_loadu_ps(&weights[i * num_inputs + 4]);
120     const __m128 weight_l = _mm_loadu_ps(&weights[i * num_inputs]);
121     const __m128 mul_h = _mm_mul_ps(inputs_h, weight_h);
122     const __m128 mul_l = _mm_mul_ps(inputs_l, weight_l);
123     add[i] = _mm_add_ps(mul_l, mul_h);
124   }
125   // add[0] = [7+3 6+2 5+1 4+0]
126   // add[1] = [15+11 14+10 13+9 12+8]
127   // add[2] = [23+19 22+18 21+17 20+16]
128   // add[3] = [31+27 30+26 29+25 28+24]
129 
130   const __m128 hadd_h = _mm_hadd_ps(add[2], add[3]);
131   // [31+30+27+26 29+28+25+24 23+22+19+18 21+20+17+16]
132   const __m128 hadd_l = _mm_hadd_ps(add[0], add[1]);
133   // [15+14+11+10 13+12+9+8 7+6+3+2 5+4+1+0]
134 
135   const __m128 haddhadd = _mm_hadd_ps(hadd_l, hadd_h);
136   // [31+30+29+28+27+26+25+24 23+22+21+20+19+18+17+16
137   //  15+14+13+12+11+10+9+8 7+6+5+4+3+2+1+0]
138 
139   *outputs = _mm_add_ps(*outputs, haddhadd);
140 }
141 
nn_activate8(__m128 * out_h,__m128 * out_l)142 static void nn_activate8(__m128 *out_h, __m128 *out_l) {
143   const __m128 zero = _mm_setzero_ps();
144   *out_h = _mm_max_ps(*out_h, zero);
145   *out_l = _mm_max_ps(*out_l, zero);
146 }
147 
nn_activate4(__m128 * x)148 static void nn_activate4(__m128 *x) { *x = _mm_max_ps(*x, _mm_setzero_ps()); }
149 
150 // Calculate prediction based on the given input features and neural net config.
151 // Assume there are no more than NN_MAX_NODES_PER_LAYER nodes in each hidden
152 // layer.
av1_nn_predict_sse3(const float * input_nodes,const NN_CONFIG * const nn_config,int reduce_prec,float * const output)153 void av1_nn_predict_sse3(const float *input_nodes,
154                          const NN_CONFIG *const nn_config, int reduce_prec,
155                          float *const output) {
156   float buf[2][NN_MAX_NODES_PER_LAYER];
157   int buf_index = 0;
158   int num_inputs = nn_config->num_inputs;
159 
160   // Hidden layers, except the final iteration is the output layer.
161   for (int layer = 0; layer <= nn_config->num_hidden_layers; layer++) {
162     const float *layer_weights = nn_config->weights[layer];
163     const float *layer_bias = nn_config->bias[layer];
164     bool output_layer = (layer == nn_config->num_hidden_layers);
165     float *const output_nodes = output_layer ? output : &buf[buf_index][0];
166     const int num_outputs = output_layer ? nn_config->num_outputs
167                                          : nn_config->num_hidden_nodes[layer];
168 
169     if (num_inputs % 4 == 0 && num_outputs % 8 == 0) {
170       for (int out = 0; out < num_outputs; out += 8) {
171         __m128 out_h = _mm_loadu_ps(&layer_bias[out + 4]);
172         __m128 out_l = _mm_loadu_ps(&layer_bias[out]);
173         for (int in = 0; in < num_inputs; in += 4) {
174           av1_nn_propagate_4to8_sse3(&input_nodes[in],
175                                      &layer_weights[out * num_inputs + in],
176                                      &out_h, &out_l, num_inputs);
177         }
178         if (!output_layer) nn_activate8(&out_h, &out_l);
179         _mm_storeu_ps(&output_nodes[out + 4], out_h);
180         _mm_storeu_ps(&output_nodes[out], out_l);
181       }
182     } else if (num_inputs % 8 == 0 && num_outputs % 4 == 0) {
183       for (int out = 0; out < num_outputs; out += 4) {
184         __m128 outputs = _mm_loadu_ps(&layer_bias[out]);
185         for (int in = 0; in < num_inputs; in += 8) {
186           nn_propagate_8to4(&input_nodes[in],
187                             &layer_weights[out * num_inputs + in], &outputs,
188                             num_inputs);
189         }
190         if (!output_layer) nn_activate4(&outputs);
191         _mm_storeu_ps(&output_nodes[out], outputs);
192       }
193     } else if (num_inputs % 4 == 0 && num_outputs % 4 == 0) {
194       for (int out = 0; out < num_outputs; out += 4) {
195         __m128 outputs = _mm_loadu_ps(&layer_bias[out]);
196         for (int in = 0; in < num_inputs; in += 4) {
197           av1_nn_propagate_4to4_sse3(&input_nodes[in],
198                                      &layer_weights[out * num_inputs + in],
199                                      &outputs, num_inputs);
200         }
201         if (!output_layer) nn_activate4(&outputs);
202         _mm_storeu_ps(&output_nodes[out], outputs);
203       }
204     } else if (num_inputs % 8 == 0) {
205       for (int out = 0; out < num_outputs; out++) {
206         __m128 total = _mm_load1_ps(&layer_bias[out]);
207         for (int in = 0; in < num_inputs; in += 8) {
208           nn_propagate_8to1(&input_nodes[in],
209                             &layer_weights[out * num_inputs + in], &total);
210         }
211         if (!output_layer) nn_activate4(&total);
212         output_nodes[out] = _mm_cvtss_f32(total);
213       }
214     } else if (num_inputs % 4 == 0) {
215       for (int out = 0; out < num_outputs; out++) {
216         __m128 total = _mm_load1_ps(&layer_bias[out]);
217         for (int in = 0; in < num_inputs; in += 4) {
218           av1_nn_propagate_4to1_sse3(
219               &input_nodes[in], &layer_weights[out * num_inputs + in], &total);
220         }
221         if (!output_layer) nn_activate4(&total);
222         output_nodes[out] = _mm_cvtss_f32(total);
223       }
224     } else {
225       // Use SSE instructions for scalar operations to avoid the latency of
226       // swapping between SIMD and FPU modes.
227       for (int out = 0; out < num_outputs; out++) {
228         __m128 total = _mm_load1_ps(&layer_bias[out]);
229         for (int in_node = 0; in_node < num_inputs; in_node++) {
230           __m128 input = _mm_load1_ps(&input_nodes[in_node]);
231           __m128 weight =
232               _mm_load1_ps(&layer_weights[num_inputs * out + in_node]);
233           total = _mm_add_ps(total, _mm_mul_ps(input, weight));
234         }
235         if (!output_layer) nn_activate4(&total);
236         output_nodes[out] = _mm_cvtss_f32(total);
237       }
238     }
239     input_nodes = output_nodes;
240     num_inputs = num_outputs;
241     buf_index = 1 - buf_index;
242   }
243   if (reduce_prec) av1_nn_output_prec_reduce(output, nn_config->num_outputs);
244 }
245 
246 // Based on N. N. Schraudolph. A Fast, Compact Approximation of the Exponential
247 // Function. Neural Computation, 11(4):853–862, 1999.
approx_exp(__m128 y)248 static inline __m128 approx_exp(__m128 y) {
249 #define A ((1 << 23) / 0.69314718056f)  // (1 << 23) / ln(2)
250 #define B \
251   127  // Offset for the exponent according to IEEE floating point standard.
252 #define C 60801  // Magic number controls the accuracy of approximation
253   const __m128 multiplier = _mm_set1_ps(A);
254   const __m128i offset = _mm_set1_epi32(B * (1 << 23) - C);
255 
256   y = _mm_mul_ps(y, multiplier);
257   y = _mm_castsi128_ps(_mm_add_epi32(_mm_cvtps_epi32(y), offset));
258   return y;
259 #undef A
260 #undef B
261 #undef C
262 }
263 
reduce_max(__m128 reg)264 static inline __m128 reduce_max(__m128 reg) {
265   __m128 tmp_reg;
266 
267   tmp_reg = _mm_shuffle_ps(reg, reg, 0x4e);  // 01 00 11 10
268   reg = _mm_max_ps(reg, tmp_reg);
269 
270   tmp_reg = _mm_shuffle_ps(reg, reg, 0xb1);  // 10 11 00 01
271   reg = _mm_max_ps(reg, tmp_reg);
272 
273   return reg;
274 }
275 
reduce_sum(__m128 reg)276 static inline __m128 reduce_sum(__m128 reg) {
277   __m128 tmp_reg;
278 
279   tmp_reg = _mm_shuffle_ps(reg, reg, 0x4e);  // 01 00 11 10
280   reg = _mm_add_ps(reg, tmp_reg);
281 
282   tmp_reg = _mm_shuffle_ps(reg, reg, 0xb1);  // 10 11 00 01
283   reg = _mm_add_ps(reg, tmp_reg);
284 
285   return reg;
286 }
287 
av1_nn_fast_softmax_16_sse3(const float * input,float * output)288 void av1_nn_fast_softmax_16_sse3(const float *input, float *output) {
289   // Clips at -10 to avoid underflowing
290   const __m128 clipper = _mm_set1_ps(-10.0f);
291 
292   // Load in 16 values
293   __m128 in_0 = _mm_loadu_ps(&input[0]);
294   __m128 in_1 = _mm_loadu_ps(&input[4]);
295   __m128 in_2 = _mm_loadu_ps(&input[8]);
296   __m128 in_3 = _mm_loadu_ps(&input[12]);
297 
298   // Get the max
299   __m128 max_0 = _mm_max_ps(in_0, in_1);
300   __m128 max_1 = _mm_max_ps(in_2, in_3);
301 
302   max_0 = _mm_max_ps(max_0, max_1);
303   max_0 = reduce_max(max_0);
304 
305   // Subtract the max off and clip
306   in_0 = _mm_sub_ps(in_0, max_0);
307   in_1 = _mm_sub_ps(in_1, max_0);
308   in_2 = _mm_sub_ps(in_2, max_0);
309   in_3 = _mm_sub_ps(in_3, max_0);
310 
311   in_0 = _mm_max_ps(in_0, clipper);
312   in_1 = _mm_max_ps(in_1, clipper);
313   in_2 = _mm_max_ps(in_2, clipper);
314   in_3 = _mm_max_ps(in_3, clipper);
315 
316   // Exponentiate and compute the denominator
317   __m128 sum = in_0 = approx_exp(in_0);
318   in_1 = approx_exp(in_1);
319   sum = _mm_add_ps(sum, in_1);
320   in_2 = approx_exp(in_2);
321   sum = _mm_add_ps(sum, in_2);
322   in_3 = approx_exp(in_3);
323   sum = _mm_add_ps(sum, in_3);
324   sum = reduce_sum(sum);
325 
326   // Divide to get the probability
327   in_0 = _mm_div_ps(in_0, sum);
328   in_1 = _mm_div_ps(in_1, sum);
329   in_2 = _mm_div_ps(in_2, sum);
330   in_3 = _mm_div_ps(in_3, sum);
331 
332   _mm_storeu_ps(&output[0], in_0);
333   _mm_storeu_ps(&output[4], in_1);
334   _mm_storeu_ps(&output[8], in_2);
335   _mm_storeu_ps(&output[12], in_3);
336 }
337