1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/MRx8c8-avx2.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 <immintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 #include <xnnpack/unaligned.h>
18
19
xnn_qu8_igemm_minmax_fp32_ukernel_3x8c8__avx2(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qu8_igemm_minmax_fp32_ukernel_3x8c8__avx2(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const uint8_t** restrict a,
26 const void* restrict w,
27 uint8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const uint8_t* zero,
32 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(uint8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8);
46 uint8_t* c0 = c;
47 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55
56 do {
57 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
58 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
59 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
60 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
61 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
62 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
63 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
64 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
65 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
66 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
67 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
68 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
69 __m256i vacc1x01 = vacc0x01;
70 __m256i vacc1x23 = vacc0x23;
71 __m256i vacc1x45 = vacc0x45;
72 __m256i vacc1x67 = vacc0x67;
73 __m256i vacc2x01 = vacc0x01;
74 __m256i vacc2x23 = vacc0x23;
75 __m256i vacc2x45 = vacc0x45;
76 __m256i vacc2x67 = vacc0x67;
77 w = (const int32_t*) w + 8;
78
79 size_t p = ks;
80 const __m256i vb_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.kernel_zero_point);
81 do {
82 const uint8_t* restrict a0 = a[0];
83 if XNN_UNPREDICTABLE(a0 != zero) {
84 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
85 }
86 const uint8_t* restrict a1 = a[1];
87 if XNN_UNPREDICTABLE(a1 != zero) {
88 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
89 }
90 const uint8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
93 }
94 a += 3;
95
96 size_t k = 0;
97 while (k < kc) {
98 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
99 const __m256i vxa0 = _mm256_cvtepu8_epi16(va0);
100 a0 += 8;
101 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
102 const __m256i vxa1 = _mm256_cvtepu8_epi16(va1);
103 a1 += 8;
104 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
105 const __m256i vxa2 = _mm256_cvtepu8_epi16(va2);
106 a2 += 8;
107
108 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
109 const __m256i vxb01 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb01), vb_zero_point);
110
111 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
112 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
113 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
114 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 16));
115 const __m256i vxb23 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb23), vb_zero_point);
116
117 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
118 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
119 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
120 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 32));
121 const __m256i vxb45 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb45), vb_zero_point);
122
123 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
124 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
125 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
126 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 48));
127 const __m256i vxb67 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb67), vb_zero_point);
128
129 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
130 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
131 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
132
133 w = (const void*) ((const uint8_t*) w + 64);
134 k += 8 * sizeof(uint8_t);
135 }
136 p -= 3 * sizeof(void*);
137 } while (p != 0);
138
139 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
140 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
141 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
142 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
143 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
144 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
145
146 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
147 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
148 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
149
150 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
151 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
152 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
153 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
154
155 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
156 __m256 vscaled1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
157 __m256 vscaled2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
158
159 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
160 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale);
161 vscaled1x01234567 = _mm256_mul_ps(vscaled1x01234567, vscale);
162 vscaled2x01234567 = _mm256_mul_ps(vscaled2x01234567, vscale);
163
164 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
165 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
166 vscaled1x01234567 = _mm256_min_ps(vscaled1x01234567, voutput_max_less_zero_point);
167 vscaled2x01234567 = _mm256_min_ps(vscaled2x01234567, voutput_max_less_zero_point);
168
169 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
170 vacc1x01234567 = _mm256_cvtps_epi32(vscaled1x01234567);
171 vacc2x01234567 = _mm256_cvtps_epi32(vscaled2x01234567);
172
173 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
174 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
175 __m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), voutput_zero_point);
176
177 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
178 vacc22x01234567 = _mm256_permute4x64_epi64(vacc22x01234567, _MM_SHUFFLE(3, 1, 2, 0));
179
180 __m256i vout = _mm256_packus_epi16(vacc01x01234567, vacc22x01234567);
181
182 vout = _mm256_max_epu8(vout, _mm256_load_si256((const __m256i*) params->fp32_avx2.output_min));
183
184 __m128i vout_lo = _mm256_castsi256_si128(vout);
185 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
186
187 if (nc >= 8) {
188 _mm_storeh_pi((__m64*) c2, _mm_castsi128_ps(vout_lo));
189 _mm_storel_epi64((__m128i*) c1, vout_hi);
190 _mm_storel_epi64((__m128i*) c0, vout_lo);
191
192 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
193 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
194 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
195
196 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
197
198 nc -= 8;
199 } else {
200 if (nc & 4) {
201 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout_lo, 2));
202 _mm_storeu_si32(c1, vout_hi);
203 _mm_storeu_si32(c0, vout_lo);
204
205 c2 += 4;
206 c1 += 4;
207 c0 += 4;
208
209 vout_lo = _mm_srli_epi64(vout_lo, 32);
210 vout_hi = _mm_srli_epi64(vout_hi, 32);
211 }
212 if (nc & 2) {
213 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout_lo, 4));
214 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout_hi, 0));
215 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
216
217 c2 += 2;
218 c1 += 2;
219 c0 += 2;
220
221 vout_lo = _mm_srli_epi32(vout_lo, 16);
222 vout_hi = _mm_srli_epi32(vout_hi, 16);
223 }
224 if (nc & 1) {
225 *c2 = (uint8_t) _mm_extract_epi8(vout_lo, 8);
226 *c1 = (uint8_t) _mm_extract_epi8(vout_hi, 0);
227 *c0 = (uint8_t) _mm_extract_epi8(vout_lo, 0);
228 }
229
230 nc = 0;
231 }
232 } while (nc != 0);
233 }
234