xref: /aosp_15_r20/external/XNNPACK/src/qu8-avgpool/9x-minmax-sse2-c8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
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 #include <xnnpack/unaligned.h>
15 
16 
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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8(
18     size_t output_pixels,
19     size_t kernel_elements,
20     size_t channels,
21     const uint8_t** input,
22     size_t input_offset,
23     const uint8_t* zero,
24     uint8_t* output,
25     size_t input_increment,
26     size_t output_increment,
27     const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(output_pixels != 0);
30   assert(kernel_elements != 0);
31   assert(kernel_elements <= 9);
32   assert(channels != 0);
33 
34   const __m128i vbias = _mm_load_si128((const __m128i*) &params->sse2.bias);
35   const __m128i vzero = _mm_setzero_si128();
36   const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
37   const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
38   const __m128i vright_shift = _mm_loadl_epi64((const __m128i*) params->sse2.right_shift);
39 
40   do {
41     const uint8_t* i0 = input[0];
42     assert(i0 != NULL);
43     const uint8_t* i1 = input[1];
44     const uint8_t* i2 = input[2];
45     const uint8_t* i3 = input[3];
46     const uint8_t* i4 = input[4];
47     const uint8_t* i5 = input[5];
48     const uint8_t* i6 = input[6];
49     const uint8_t* i7 = input[7];
50     const uint8_t* i8 = input[8];
51     input = (const uint8_t**) ((uintptr_t) input + input_increment);
52     if (kernel_elements < 2) {
53       i1 = zero;
54     }
55     assert(i1 != NULL);
56     if (kernel_elements <= 2) {
57       i2 = zero;
58     }
59     assert(i2 != NULL);
60     if (kernel_elements < 4) {
61       i3 = zero;
62     }
63     assert(i3 != NULL);
64     if (kernel_elements <= 4) {
65       i4 = zero;
66     }
67     assert(i4 != NULL);
68     if (kernel_elements < 6) {
69       i5 = zero;
70     }
71     assert(i5 != NULL);
72     if (kernel_elements <= 6) {
73       i6 = zero;
74     }
75     assert(i6 != NULL);
76     if (kernel_elements < 8) {
77       i7 = zero;
78     }
79     assert(i7 != NULL);
80     if (kernel_elements <= 8) {
81       i8 = zero;
82     }
83     assert(i8 != NULL);
84     if XNN_UNPREDICTABLE(i0 != zero) {
85       i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
86     }
87     if XNN_UNPREDICTABLE(i1 != zero) {
88       i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
89     }
90     if XNN_UNPREDICTABLE(i2 != zero) {
91       i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
92     }
93     if XNN_UNPREDICTABLE(i3 != zero) {
94       i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
95     }
96     if XNN_UNPREDICTABLE(i4 != zero) {
97       i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
98     }
99     if XNN_UNPREDICTABLE(i5 != zero) {
100       i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
101     }
102     if XNN_UNPREDICTABLE(i6 != zero) {
103       i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
104     }
105     if XNN_UNPREDICTABLE(i7 != zero) {
106       i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
107     }
108     if XNN_UNPREDICTABLE(i8 != zero) {
109       i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
110     }
111 
112     size_t c = channels;
113     while (c >= 8) {
114       const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0); i0 += 8;
115       const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1); i1 += 8;
116       const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2); i2 += 8;
117       const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3); i3 += 8;
118       const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4); i4 += 8;
119       const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5); i5 += 8;
120       const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6); i6 += 8;
121       const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7); i7 += 8;
122       const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8); i8 += 8;
123 
124       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
125       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
126       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
127       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
128       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
129       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
130       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
131       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
132       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);
133 
134       const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8);
135       const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
136       const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
137       const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
138 
139       const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45);
140       const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67);
141       const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678);
142 
143       const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero));
144       const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero));
145 
146       const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);
147       const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);
148 
149       const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo);
150       const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi);
151 
152       const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1));
153       const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1));
154 
155       const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier);
156       const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier);
157 
158       const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier);
159       const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier);
160 
161       const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift);
162       const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift);
163       const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift);
164       const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift);
165 
166       const __m128i vabs_scaled_lo0213 = _mm_castps_si128(
167           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0)));
168       const __m128i vabs_scaled_hi0213 = _mm_castps_si128(
169           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0)));
170 
171       const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0));
172       const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0));
173 
174       const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo);
175       const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi);
176 
177       __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);
178       vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) &params->sse2.output_zero_point));
179       vout = _mm_packus_epi16(vout, vout);
180       vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_max));
181       vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_min));
182 
183       _mm_storel_epi64((__m128i*) output, vout);
184       output += 8;
185 
186       c -= 8;
187     }
188     if (c != 0) {
189       const __m128i vi0 = _mm_loadl_epi64((const __m128i*) i0);
190       const __m128i vi1 = _mm_loadl_epi64((const __m128i*) i1);
191       const __m128i vi2 = _mm_loadl_epi64((const __m128i*) i2);
192       const __m128i vi3 = _mm_loadl_epi64((const __m128i*) i3);
193       const __m128i vi4 = _mm_loadl_epi64((const __m128i*) i4);
194       const __m128i vi5 = _mm_loadl_epi64((const __m128i*) i5);
195       const __m128i vi6 = _mm_loadl_epi64((const __m128i*) i6);
196       const __m128i vi7 = _mm_loadl_epi64((const __m128i*) i7);
197       const __m128i vi8 = _mm_loadl_epi64((const __m128i*) i8);
198 
199       const __m128i vxi0 = _mm_unpacklo_epi8(vi0, vzero);
200       const __m128i vxi1 = _mm_unpacklo_epi8(vi1, vzero);
201       const __m128i vxi2 = _mm_unpacklo_epi8(vi2, vzero);
202       const __m128i vxi3 = _mm_unpacklo_epi8(vi3, vzero);
203       const __m128i vxi4 = _mm_unpacklo_epi8(vi4, vzero);
204       const __m128i vxi5 = _mm_unpacklo_epi8(vi5, vzero);
205       const __m128i vxi6 = _mm_unpacklo_epi8(vi6, vzero);
206       const __m128i vxi7 = _mm_unpacklo_epi8(vi7, vzero);
207       const __m128i vxi8 = _mm_unpacklo_epi8(vi8, vzero);
208 
209       const __m128i vsum018 = _mm_add_epi16(_mm_add_epi16(vxi0, vxi1), vxi8);
210       const __m128i vsum23 = _mm_add_epi16(vxi2, vxi3);
211       const __m128i vsum45 = _mm_add_epi16(vxi4, vxi5);
212       const __m128i vsum67 = _mm_add_epi16(vxi6, vxi7);
213 
214       const __m128i vsum2345 = _mm_add_epi16(vsum23, vsum45);
215       const __m128i vsum01678 = _mm_add_epi16(vsum018, vsum67);
216       const __m128i vsum = _mm_add_epi16(vsum2345, vsum01678);
217 
218       const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero));
219       const __m128i vacc_hi = _mm_add_epi32(vbias, _mm_unpackhi_epi16(vsum, vzero));
220 
221       const __m128i vneg_mask_lo = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_lo);
222       const __m128i vneg_mask_hi = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc_hi);
223 
224       const __m128i vabs_lo0123 = _mm_sub_epi32(_mm_xor_si128(vacc_lo, vneg_mask_lo), vneg_mask_lo);
225       const __m128i vabs_hi0123 = _mm_sub_epi32(_mm_xor_si128(vacc_hi, vneg_mask_hi), vneg_mask_hi);
226 
227       const __m128i vabs_lo1032 = _mm_shuffle_epi32(vabs_lo0123, _MM_SHUFFLE(2, 3, 0, 1));
228       const __m128i vabs_hi1032 = _mm_shuffle_epi32(vabs_hi0123, _MM_SHUFFLE(2, 3, 0, 1));
229 
230       const __m128i vabsmul_lo02 = _mm_mul_epu32(vabs_lo0123, vmultiplier);
231       const __m128i vabsmul_hi02 = _mm_mul_epu32(vabs_hi0123, vmultiplier);
232 
233       const __m128i vabsmul_lo13 = _mm_mul_epu32(vabs_lo1032, vmultiplier);
234       const __m128i vabsmul_hi13 = _mm_mul_epu32(vabs_hi1032, vmultiplier);
235 
236       const __m128i vabs_scaled_lo02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo02, vrounding), vright_shift);
237       const __m128i vabs_scaled_lo13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_lo13, vrounding), vright_shift);
238       const __m128i vabs_scaled_hi02 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi02, vrounding), vright_shift);
239       const __m128i vabs_scaled_hi13 = _mm_srl_epi64(_mm_add_epi64(vabsmul_hi13, vrounding), vright_shift);
240 
241       const __m128i vabs_scaled_lo0213 = _mm_castps_si128(
242           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_lo02), _mm_castsi128_ps(vabs_scaled_lo13), _MM_SHUFFLE(2, 0, 2, 0)));
243       const __m128i vabs_scaled_hi0213 = _mm_castps_si128(
244           _mm_shuffle_ps(_mm_castsi128_ps(vabs_scaled_hi02), _mm_castsi128_ps(vabs_scaled_hi13), _MM_SHUFFLE(2, 0, 2, 0)));
245 
246       const __m128i vabs_scaled_lo = _mm_shuffle_epi32(vabs_scaled_lo0213, _MM_SHUFFLE(3, 1, 2, 0));
247       const __m128i vabs_scaled_hi = _mm_shuffle_epi32(vabs_scaled_hi0213, _MM_SHUFFLE(3, 1, 2, 0));
248 
249       const __m128i vscaled_lo = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_lo, vneg_mask_lo), vneg_mask_lo);
250       const __m128i vscaled_hi = _mm_sub_epi32(_mm_xor_si128(vabs_scaled_hi, vneg_mask_hi), vneg_mask_hi);
251 
252       __m128i vout = _mm_packs_epi32(vscaled_lo, vscaled_hi);
253       vout = _mm_adds_epi16(vout, _mm_load_si128((const __m128i*) &params->sse2.output_zero_point));
254       vout = _mm_packus_epi16(vout, vout);
255       vout = _mm_min_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_max));
256       vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) &params->sse2.output_min));
257 
258       if (c & 4) {
259         unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout));
260         output += 4;
261         vout = _mm_srli_epi64(vout, 32);
262       }
263       if (c & 2) {
264         unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout, 0));
265         output += 2;
266         vout = _mm_srli_epi32(vout, 16);
267       }
268       if (c & 1) {
269         *output = (uint8_t) _mm_cvtsi128_si32(vout);
270         output += 1;
271       }
272     }
273     output = (uint8_t*) ((uintptr_t) output + output_increment);
274   } while (--output_pixels != 0);
275 }
276