1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qs8_gemm_xw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8);
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55
56 do {
57 const __m128i vbias0x0 = _mm_loadu_si32(w);
58 const __m128i vbias0x1 = _mm_loadu_si32((const void*) ((uintptr_t) w + sizeof(int32_t)));
59 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
60 const __m128i vbias0x2 = _mm_loadu_si32((const void*) ((uintptr_t) w + 2 * sizeof(int32_t)));
61 const __m128i vbias0x3 = _mm_loadu_si32((const void*) ((uintptr_t) w + 3 * sizeof(int32_t)));
62 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
63 const __m128i vbias0x4 = _mm_loadu_si32((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
64 const __m128i vbias0x5 = _mm_loadu_si32((const void*) ((uintptr_t) w + 5 * sizeof(int32_t)));
65 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
66 const __m128i vbias0x6 = _mm_loadu_si32((const void*) ((uintptr_t) w + 6 * sizeof(int32_t)));
67 const __m128i vbias0x7 = _mm_loadu_si32((const void*) ((uintptr_t) w + 7 * sizeof(int32_t)));
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 void*) ((uintptr_t) w + 8 * sizeof(int32_t));
78
79 size_t k = 0;
80 while (k < kc) {
81 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
82 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
83 a0 += 8;
84 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
85 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
86 a1 += 8;
87 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
88 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
89 a2 += 8;
90
91 const __m256i vxb01 = _mm256_load_si256((const __m256i*) w);
92
93 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
94 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
95 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
96 const __m256i vxb23 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 16 * sizeof(int16_t)));
97
98 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
99 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
100 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
101 const __m256i vxb45 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 32 * sizeof(int16_t)));
102
103 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
104 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
105 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
106 const __m256i vxb67 = _mm256_load_si256((const __m256i*) ((uintptr_t) w + 48 * sizeof(int16_t)));
107
108 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
109 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
110 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
111
112 w = (const void*) ((uintptr_t) w + 64 * sizeof(int16_t));
113 k += 8 * sizeof(int8_t);
114 }
115
116 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
117 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
118 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
119 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
120 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
121 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
122
123 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
124 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
125 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
126
127 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
128 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
129 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
130 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
131
132 const __m256i vmultiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.multiplier));
133 const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));
134
135 const __m256i vacc0x11335577 = _mm256_shuffle_epi32(vacc0x01234567, _MM_SHUFFLE(3, 3, 1, 1));
136 const __m256i vacc1x11335577 = _mm256_shuffle_epi32(vacc1x01234567, _MM_SHUFFLE(3, 3, 1, 1));
137 const __m256i vacc2x11335577 = _mm256_shuffle_epi32(vacc2x01234567, _MM_SHUFFLE(3, 3, 1, 1));
138
139 const __m256i vprod0x0246 = _mm256_add_epi64(_mm256_mul_epi32(vacc0x01234567, vmultiplier), vrounding);
140 const __m256i vprod1x0246 = _mm256_add_epi64(_mm256_mul_epi32(vacc1x01234567, vmultiplier), vrounding);
141 const __m256i vprod2x0246 = _mm256_add_epi64(_mm256_mul_epi32(vacc2x01234567, vmultiplier), vrounding);
142
143 const __m256i vprod0x1357 = _mm256_add_epi64(_mm256_mul_epi32(vacc0x11335577, vmultiplier), vrounding);
144 const __m256i vprod1x1357 = _mm256_add_epi64(_mm256_mul_epi32(vacc1x11335577, vmultiplier), vrounding);
145 const __m256i vprod2x1357 = _mm256_add_epi64(_mm256_mul_epi32(vacc2x11335577, vmultiplier), vrounding);
146
147 const __m256i vq31prod0x0246 = _mm256_srli_epi64(vprod0x0246, 31);
148 const __m256i vq31prod0x1357 = _mm256_add_epi64(vprod0x1357, vprod0x1357);
149 const __m256i vq31prod1x0246 = _mm256_srli_epi64(vprod1x0246, 31);
150 const __m256i vq31prod1x1357 = _mm256_add_epi64(vprod1x1357, vprod1x1357);
151 const __m256i vq31prod2x0246 = _mm256_srli_epi64(vprod2x0246, 31);
152 const __m256i vq31prod2x1357 = _mm256_add_epi64(vprod2x1357, vprod2x1357);
153
154 const __m256i vq31prod0x01234567 = _mm256_blend_epi16(vq31prod0x0246, vq31prod0x1357, 0xCC);
155 const __m256i vq31prod1x01234567 = _mm256_blend_epi16(vq31prod1x0246, vq31prod1x1357, 0xCC);
156 const __m256i vq31prod2x01234567 = _mm256_blend_epi16(vq31prod2x0246, vq31prod2x1357, 0xCC);
157
158 const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
159 const __m256i vrem0x01234567 =
160 _mm256_add_epi32(_mm256_and_si256(vq31prod0x01234567, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod0x01234567));
161 const __m256i vrem1x01234567 =
162 _mm256_add_epi32(_mm256_and_si256(vq31prod1x01234567, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod1x01234567));
163 const __m256i vrem2x01234567 =
164 _mm256_add_epi32(_mm256_and_si256(vq31prod2x01234567, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod2x01234567));
165
166 const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
167 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
168 vacc0x01234567 =
169 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod0x01234567, vshift), _mm256_cmpgt_epi32(vrem0x01234567, vremainder_threshold));
170 vacc1x01234567 =
171 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod1x01234567, vshift), _mm256_cmpgt_epi32(vrem1x01234567, vremainder_threshold));
172 vacc2x01234567 =
173 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod2x01234567, vshift), _mm256_cmpgt_epi32(vrem2x01234567, vremainder_threshold));
174
175 const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
176 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
177 __m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), voutput_zero_point);
178
179 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
180 vacc22x01234567 = _mm256_permute4x64_epi64(vacc22x01234567, _MM_SHUFFLE(3, 1, 2, 0));
181
182 const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
183 const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
184 vacc01x01234567 = _mm256_min_epi16(_mm256_max_epi16(vacc01x01234567, voutput_min), voutput_max);
185 vacc22x01234567 = _mm256_min_epi16(_mm256_max_epi16(vacc22x01234567, voutput_min), voutput_max);
186
187 __m256i vout = _mm256_packs_epi16(vacc01x01234567, vacc22x01234567);
188 __m128i vout_lo = _mm256_castsi256_si128(vout);
189 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
190
191 if (nc >= 8) {
192 _mm_storel_epi64((__m128i*) c0, vout_lo);
193 _mm_storel_epi64((__m128i*) c1, vout_hi);
194 _mm_storeh_pi((__m64*) c2, _mm_castsi128_ps(vout_lo));
195
196 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
197 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
198 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
199
200 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
201 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
202 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
203
204 nc -= 8;
205 } else {
206 if (nc & 4) {
207 _mm_storeu_si32(c0, vout_lo);
208 _mm_storeu_si32(c1, vout_hi);
209 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout_lo, 2);
210
211 c0 += 4;
212 c1 += 4;
213 c2 += 4;
214
215 vout_lo = _mm_srli_epi64(vout_lo, 32);
216 vout_hi = _mm_srli_epi64(vout_hi, 32);
217 }
218 if (nc & 2) {
219 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout_lo, 0);
220 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout_hi, 0);
221 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout_lo, 4);
222
223 c0 += 2;
224 c1 += 2;
225 c2 += 2;
226
227 vout_lo = _mm_srli_epi32(vout_lo, 16);
228 vout_hi = _mm_srli_epi32(vout_hi, 16);
229 }
230 if (nc & 1) {
231 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
232 *c1 = (uint8_t) _mm_extract_epi8(vout_hi, 0);
233 *c2 = (uint8_t) _mm_extract_epi8(vout_lo, 8);
234 }
235
236 nc = 0;
237 }
238 } while (nc != 0);
239 }
240