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_qc8_gemm_xw_minmax_fp32_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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_xw_minmax_fp32_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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
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 int32_t*) w + 1);
59 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
60 const __m128i vbias0x2 = _mm_loadu_si32((const int32_t*) w + 2);
61 const __m128i vbias0x3 = _mm_loadu_si32((const int32_t*) w + 3);
62 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
63 const __m128i vbias0x4 = _mm_loadu_si32((const int32_t*) w + 4);
64 const __m128i vbias0x5 = _mm_loadu_si32((const int32_t*) w + 5);
65 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
66 const __m128i vbias0x6 = _mm_loadu_si32((const int32_t*) w + 6);
67 const __m128i vbias0x7 = _mm_loadu_si32((const int32_t*) 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 void*) ((const int32_t*) w + 8);
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*) ((const int16_t*) w + 16));
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*) ((const int16_t*) w + 32));
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*) ((const int16_t*) w + 48));
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*) ((const int16_t*) w + 64);
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 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
133 __m256 vscaled1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
134 __m256 vscaled2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
135
136 const __m256 vscale01234567 = _mm256_load_ps(w);
137 w = (const void*) ((const float*) w + 8);
138 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale01234567);
139 vscaled1x01234567 = _mm256_mul_ps(vscaled1x01234567, vscale01234567);
140 vscaled2x01234567 = _mm256_mul_ps(vscaled2x01234567, vscale01234567);
141
142 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx2.output_max_less_zero_point);
143 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
144 vscaled1x01234567 = _mm256_min_ps(vscaled1x01234567, voutput_max_less_zero_point);
145 vscaled2x01234567 = _mm256_min_ps(vscaled2x01234567, voutput_max_less_zero_point);
146
147 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
148 vacc1x01234567 = _mm256_cvtps_epi32(vscaled1x01234567);
149 vacc2x01234567 = _mm256_cvtps_epi32(vscaled2x01234567);
150
151 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point);
152 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
153 __m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), voutput_zero_point);
154
155 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
156 vacc22x01234567 = _mm256_permute4x64_epi64(vacc22x01234567, _MM_SHUFFLE(3, 1, 2, 0));
157
158 __m256i vout = _mm256_packs_epi16(vacc01x01234567, vacc22x01234567);
159
160 vout = _mm256_max_epi8(vout, _mm256_load_si256((const __m256i*) params->avx2.output_min));
161
162 __m128i vout_lo = _mm256_castsi256_si128(vout);
163 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
164
165 if (nc >= 8) {
166 _mm_storel_epi64((__m128i*) c0, vout_lo);
167 _mm_storel_epi64((__m128i*) c1, vout_hi);
168 _mm_storeh_pi((__m64*) c2, _mm_castsi128_ps(vout_lo));
169
170 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
171 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
172 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
173
174 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
175 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
176 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
177
178 nc -= 8;
179 } else {
180 if (nc & 4) {
181 _mm_storeu_si32(c0, vout_lo);
182 _mm_storeu_si32(c1, vout_hi);
183 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout_lo, 2);
184
185 c0 += 4;
186 c1 += 4;
187 c2 += 4;
188
189 vout_lo = _mm_srli_epi64(vout_lo, 32);
190 vout_hi = _mm_srli_epi64(vout_hi, 32);
191 }
192 if (nc & 2) {
193 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout_lo, 0);
194 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout_hi, 0);
195 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout_lo, 4);
196
197 c0 += 2;
198 c1 += 2;
199 c2 += 2;
200
201 vout_lo = _mm_srli_epi32(vout_lo, 16);
202 vout_hi = _mm_srli_epi32(vout_hi, 16);
203 }
204 if (nc & 1) {
205 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
206 *c1 = (int8_t) _mm_extract_epi8(vout_hi, 0);
207 *c2 = (int8_t) _mm_extract_epi8(vout_lo, 8);
208 }
209
210 nc = 0;
211 }
212 } while (nc != 0);
213 }
214