1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gavgpool/unipass-sse4.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <smmintrin.h>
13
14 #include <xnnpack/gavgpool.h>
15 #include <xnnpack/unaligned.h>
16
17
xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16(size_t rows,size_t channels,const uint8_t * input,size_t input_stride,const uint8_t * zero,uint8_t * output,const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c16(
19 size_t rows,
20 size_t channels,
21 const uint8_t* input,
22 size_t input_stride,
23 const uint8_t* zero,
24 uint8_t* output,
25 const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
26 {
27 assert(rows != 0);
28 assert(rows <= 7);
29 assert(channels != 0);
30
31 const uint8_t* i0 = input;
32 const uint8_t* i1 = (const uint8_t*) ((uintptr_t) i0 + input_stride);
33 if XNN_UNPREDICTABLE(rows < 2) {
34 i1 = zero;
35 }
36 const uint8_t* i2 = (const uint8_t*) ((uintptr_t) i1 + input_stride);
37 if XNN_UNPREDICTABLE(rows <= 2) {
38 i2 = zero;
39 }
40 const uint8_t* i3 = (const uint8_t*) ((uintptr_t) i2 + input_stride);
41 if XNN_UNPREDICTABLE(rows < 4) {
42 i3 = zero;
43 }
44 const uint8_t* i4 = (const uint8_t*) ((uintptr_t) i3 + input_stride);
45 if XNN_UNPREDICTABLE(rows <= 4) {
46 i4 = zero;
47 }
48 const uint8_t* i5 = (const uint8_t*) ((uintptr_t) i4 + input_stride);
49 if XNN_UNPREDICTABLE(rows < 6) {
50 i5 = zero;
51 }
52 const uint8_t* i6 = (const uint8_t*) ((uintptr_t) i5 + input_stride);
53 if XNN_UNPREDICTABLE(rows <= 6) {
54 i6 = zero;
55 }
56
57 const __m128i vinit_bias = _mm_load_si128((const __m128i*) params->fp32_sse4.init_bias);
58 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
59 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
60 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
61 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse4.output_min);
62 for (; channels >= 16; channels -= 16) {
63 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
64 const __m128i vxi0x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i0 + 8)));
65 i0 += 16;
66 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
67 const __m128i vxi1x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i1 + 8)));
68 i1 += 16;
69
70 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
71 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
72 __m128i vacc89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF);
73 const __m128i vxi2x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i2 + 8)));
74 i2 += 16;
75
76 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
77 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
78 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi2x89ABCDEF);
79 const __m128i vxi3x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i3 + 8)));
80 i3 += 16;
81 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
82 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
83 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi3x89ABCDEF);
84 const __m128i vxi4x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i4 + 8)));
85 i4 += 16;
86 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
87 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
88 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi4x89ABCDEF);
89 const __m128i vxi5x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i5 + 8)));
90 i5 += 16;
91 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
92 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
93 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi5x89ABCDEF);
94 const __m128i vxi6x89ABCDEF = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (i6 + 8)));
95 i6 += 16;
96
97 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
98 vacc89ABCDEF = _mm_add_epi16(vacc89ABCDEF, vxi6x89ABCDEF);
99
100 const __m128i vzero = _mm_setzero_si128();
101 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
102 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, vzero);
103 __m128i vacc89AB = _mm_cvtepu16_epi32(vacc89ABCDEF);
104 __m128i vaccCDEF = _mm_unpackhi_epi16(vacc89ABCDEF, vzero);
105
106 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
107 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
108 vacc89AB = _mm_add_epi32(vacc89AB, vinit_bias);
109 vaccCDEF = _mm_add_epi32(vaccCDEF, vinit_bias);
110
111 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
112 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
113 __m128 vfpacc89AB = _mm_cvtepi32_ps(vacc89AB);
114 __m128 vfpaccCDEF = _mm_cvtepi32_ps(vaccCDEF);
115
116 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
117 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
118 vfpacc89AB = _mm_mul_ps(vfpacc89AB, vscale);
119 vfpaccCDEF = _mm_mul_ps(vfpaccCDEF, vscale);
120
121 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
122 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
123 vfpacc89AB = _mm_min_ps(vfpacc89AB, voutput_max_less_zero_point);
124 vfpaccCDEF = _mm_min_ps(vfpaccCDEF, voutput_max_less_zero_point);
125
126 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
127 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
128 vacc89AB = _mm_cvtps_epi32(vfpacc89AB);
129 vaccCDEF = _mm_cvtps_epi32(vfpaccCDEF);
130
131 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
132 __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
133
134 __m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF);
135
136 vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min);
137
138 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
139 output += 16;
140 }
141 if XNN_UNLIKELY(channels != 0) {
142 do {
143 const __m128i vxi0x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i0));
144 i0 += 8;
145 const __m128i vxi1x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i1));
146 i1 += 8;
147
148 __m128i vacc01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567);
149 const __m128i vxi2x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i2));
150 i2 += 8;
151
152 vacc01234567 = _mm_add_epi16(vacc01234567, vxi2x01234567);
153 const __m128i vxi3x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i3));
154 i3 += 8;
155 vacc01234567 = _mm_add_epi16(vacc01234567, vxi3x01234567);
156 const __m128i vxi4x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i4));
157 i4 += 8;
158 vacc01234567 = _mm_add_epi16(vacc01234567, vxi4x01234567);
159 const __m128i vxi5x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i5));
160 i5 += 8;
161 vacc01234567 = _mm_add_epi16(vacc01234567, vxi5x01234567);
162 const __m128i vxi6x01234567 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) i6));
163 i6 += 8;
164
165 vacc01234567 = _mm_add_epi16(vacc01234567, vxi6x01234567);
166
167 __m128i vacc0123 = _mm_cvtepu16_epi32(vacc01234567);
168 __m128i vacc4567 = _mm_unpackhi_epi16(vacc01234567, _mm_setzero_si128());
169
170 vacc0123 = _mm_add_epi32(vacc0123, vinit_bias);
171 vacc4567 = _mm_add_epi32(vacc4567, vinit_bias);
172
173 __m128 vfpacc0123 = _mm_cvtepi32_ps(vacc0123);
174 __m128 vfpacc4567 = _mm_cvtepi32_ps(vacc4567);
175
176 vfpacc0123 = _mm_mul_ps(vfpacc0123, vscale);
177 vfpacc4567 = _mm_mul_ps(vfpacc4567, vscale);
178
179 vfpacc0123 = _mm_min_ps(vfpacc0123, voutput_max_less_zero_point);
180 vfpacc4567 = _mm_min_ps(vfpacc4567, voutput_max_less_zero_point);
181
182 vacc0123 = _mm_cvtps_epi32(vfpacc0123);
183 vacc4567 = _mm_cvtps_epi32(vfpacc4567);
184
185 __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
186
187 __m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567);
188 vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min);
189
190 if XNN_LIKELY(channels >= 8) {
191 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
192 output += 8;
193 channels -= 8;
194 } else {
195 if (channels & 4) {
196 unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567));
197 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
198 output += 4;
199 }
200 if (channels & 2) {
201 unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0));
202 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
203 output += 2;
204 }
205 if (channels & 1) {
206 *output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0);
207 output += 1;
208 }
209 channels = 0;
210 }
211 } while (channels != 0);
212 }
213 }
214