1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8 
9 #include <assert.h>
10 
11 #include <emmintrin.h>
12 
13 #include <xnnpack/avgpool.h>
14 
15 
xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8(size_t output_pixels,size_t kernel_elements,size_t channels,const uint8_t ** input,size_t input_offset,const uint8_t * zero,uint8_t * output,size_t input_increment,size_t output_increment,const union xnn_qu8_avgpool_params params[restrict XNN_MIN_ELEMENTS (1)])16 void xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8(
17     size_t output_pixels,
18     size_t kernel_elements,
19     size_t channels,
20     const uint8_t** input,
21     size_t input_offset,
22     const uint8_t* zero,
23     uint8_t* output,
24     size_t input_increment,
25     size_t output_increment,
26     const union xnn_qu8_avgpool_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
27 {
28   assert(output_pixels != 0);
29   assert(kernel_elements != 0);
30   assert(kernel_elements <= 9);
31   assert(channels != 0);
32 
33   const __m128i vbias = _mm_load_si128((const __m128i*) &params->sse2.bias);
34   const __m128i vzero = _mm_setzero_si128();
35   const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
36   const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
37   const __m128i vright_shift = _mm_loadl_epi64((const __m128i*) params->sse2.right_shift);
38 
39   do {
40     const uint8_t* i0 = input[0];
41     assert(i0 != NULL);
42     const uint8_t* i1 = input[1];
43     const uint8_t* i2 = input[2];
44     const uint8_t* i3 = input[3];
45     const uint8_t* i4 = input[4];
46     const uint8_t* i5 = input[5];
47     const uint8_t* i6 = input[6];
48     const uint8_t* i7 = input[7];
49     const uint8_t* i8 = input[8];
50     input = (const uint8_t**) ((uintptr_t) input + input_increment);
51     if (kernel_elements < 2) {
52       i1 = zero;
53     }
54     assert(i1 != NULL);
55     if (kernel_elements <= 2) {
56       i2 = zero;
57     }
58     assert(i2 != NULL);
59     if (kernel_elements < 4) {
60       i3 = zero;
61     }
62     assert(i3 != NULL);
63     if (kernel_elements <= 4) {
64       i4 = zero;
65     }
66     assert(i4 != NULL);
67     if (kernel_elements < 6) {
68       i5 = zero;
69     }
70     assert(i5 != NULL);
71     if (kernel_elements <= 6) {
72       i6 = zero;
73     }
74     assert(i6 != NULL);
75     if (kernel_elements < 8) {
76       i7 = zero;
77     }
78     assert(i7 != NULL);
79     if (kernel_elements <= 8) {
80       i8 = zero;
81     }
82     assert(i8 != NULL);
83     if XNN_UNPREDICTABLE(i0 != zero) {
84       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
85     }
86     if XNN_UNPREDICTABLE(i1 != zero) {
87       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
88     }
89     if XNN_UNPREDICTABLE(i2 != zero) {
90       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
91     }
92     if XNN_UNPREDICTABLE(i3 != zero) {
93       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
94     }
95     if XNN_UNPREDICTABLE(i4 != zero) {
96       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
97     }
98     if XNN_UNPREDICTABLE(i5 != zero) {
99       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
100     }
101     if XNN_UNPREDICTABLE(i6 != zero) {
102       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
103     }
104     if XNN_UNPREDICTABLE(i7 != zero) {
105       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
106     }
107     if XNN_UNPREDICTABLE(i8 != zero) {
108       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
109     }
110 
111     size_t c = channels;
112     while (c >= 8) {
113       const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8;
114       const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8;
115       const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8;
116       const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8;
117       const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8;
118       const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8;
119       const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8;
120       const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8;
121       const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); i8 += 8;
122 
123       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
124       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
125       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
126       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
127       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
128       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
129       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
130       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
131       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);
132 
133       const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8);
134       const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
135       const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
136       const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
137 
138       const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45);
139       const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67);
140       const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678);
141 
142       const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero));
143       const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero));
144 
145       const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);
146       const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);
147 
148       const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo);
149       const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi);
150 
151       const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1));
152       const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1));
153 
154       const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier);
155       const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier);
156 
157       const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier);
158       const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier);
159 
160       const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift);
161       const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift);
162       const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift);
163       const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift);
164 
165       const __m128i vabs_scaled_lo0213 = _mm_castps_si128(
166           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0)));
167       const __m128i vabs_scaled_hi0213 = _mm_castps_si128(
168           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0)));
169 
170       const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0));
171       const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0));
172 
173       const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo);
174       const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi);
175 
176       __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);
177       vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) &params->sse2.output_zero_point));
178       vout = _mm_packus_epi16(vout, vout);
179       vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_max));
180       vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_min));
181 
182       _mm_storel_epi64((__m128i*) output, vout);
183       output += 8;
184 
185       c -= 8;
186     }
187     if (c != 0) {
188       const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0);
189       const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1);
190       const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2);
191       const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3);
192       const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4);
193       const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5);
194       const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6);
195       const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7);
196       const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8);
197 
198       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
199       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
200       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
201       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
202       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
203       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
204       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
205       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
206       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);
207 
208       const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8);
209       const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
210       const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
211       const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
212 
213       const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45);
214       const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67);
215       const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678);
216 
217       const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero));
218       const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero));
219 
220       const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);
221       const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);
222 
223       const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo);
224       const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi);
225 
226       const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1));
227       const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1));
228 
229       const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier);
230       const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier);
231 
232       const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier);
233       const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier);
234 
235       const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift);
236       const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift);
237       const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift);
238       const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift);
239 
240       const __m128i vabs_scaled_lo0213 = _mm_castps_si128(
241           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0)));
242       const __m128i vabs_scaled_hi0213 = _mm_castps_si128(
243           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0)));
244 
245       const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0));
246       const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0));
247 
248       const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo);
249       const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi);
250 
251       __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);
252       vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) &params->sse2.output_zero_point));
253       vout = _mm_packus_epi16(vout, vout);
254       vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_max));
255       vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_min));
256 
257       if (c & 4) {
258         *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout);
259         output += 4;
260         vout = _mm_srli_epi64(vout, 32);
261       }
262       if (c & 2) {
263         *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout, 0);
264         output += 2;
265         vout = _mm_srli_epi32(vout, 16);
266       }
267       if (c & 1) {
268         *((uint8_t*) output) = (uint8_t) _mm_cvtsi128_si32(vout);
269         output += 1;
270       }
271     }
272     output = (uint8_t*) ((uintptr_t) output + output_increment);
273   } while (--output_pixels != 0);
274 }
275