Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-3x8c8-minmax-fp32-avx2.c

raw
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
20 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_3x8c8__avx2(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_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(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 int8_t* c0 = c;
47 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55
56 const __m256 voutput_max_less_zero_point = _mm256_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
57 const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point);
58 const __m256i voutput_min = _mm256_set1_epi8(params->fp32_scalar.output_min);
59 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
60 // XNN_FORCE_REALIZATION(voutput_zero_point);
61 // XNN_FORCE_REALIZATION(voutput_min);
62
63 do {
64 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
65 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
66 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
67 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
68 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
69 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
70 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
71 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
72 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
73 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
74 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
75 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
76 __m256i vacc1x01 = vacc0x01;
77 __m256i vacc1x23 = vacc0x23;
78 __m256i vacc1x45 = vacc0x45;
79 __m256i vacc1x67 = vacc0x67;
80 __m256i vacc2x01 = vacc0x01;
81 __m256i vacc2x23 = vacc0x23;
82 __m256i vacc2x45 = vacc0x45;
83 __m256i vacc2x67 = vacc0x67;
84 w = (const int32_t*) w + 8;
85
86 size_t p = ks;
87 do {
88 const int8_t* restrict a0 = a[0];
89 if XNN_UNPREDICTABLE(a0 != zero) {
90 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
91 }
92 const int8_t* restrict a1 = a[1];
93 if XNN_UNPREDICTABLE(a1 != zero) {
94 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
95 }
96 const int8_t* restrict a2 = a[2];
97 if XNN_UNPREDICTABLE(a2 != zero) {
98 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
99 }
100 a += 3;
101
102 size_t k = 0;
103 while (k < kc) {
104 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
105 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
106 a0 += 8;
107 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
108 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
109 a1 += 8;
110 const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2));
111 const __m256i vxa2 = _mm256_cvtepi8_epi16(va2);
112 a2 += 8;
113
114 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
115 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
116
117 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
118 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
119 vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01));
120 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
121 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
122
123 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
124 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
125 vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23));
126 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
127 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
128
129 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
130 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
131 vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45));
132 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
133 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
134
135 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
136 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
137 vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67));
138
139 w = (const void*) ((const int8_t*) w + 64);
140 k += 8 * sizeof(int8_t);
141 }
142 p -= 3 * sizeof(void*);
143 } while (p != 0);
144
145 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
146 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
147 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
148 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
149 const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23);
150 const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67);
151
152 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
153 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
154 const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657);
155
156 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
157 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
158 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
159 __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask);
160
161 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
162 __m256 vscaled1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
163 __m256 vscaled2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567);
164
165 const __m256 vscale01234567 = _mm256_load_ps(w);
166 w = (const void*) ((const float*) w + 8);
167 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale01234567);
168 vscaled1x01234567 = _mm256_mul_ps(vscaled1x01234567, vscale01234567);
169 vscaled2x01234567 = _mm256_mul_ps(vscaled2x01234567, vscale01234567);
170
171 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
172 vscaled1x01234567 = _mm256_min_ps(vscaled1x01234567, voutput_max_less_zero_point);
173 vscaled2x01234567 = _mm256_min_ps(vscaled2x01234567, voutput_max_less_zero_point);
174
175 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
176 vacc1x01234567 = _mm256_cvtps_epi32(vscaled1x01234567);
177 vacc2x01234567 = _mm256_cvtps_epi32(vscaled2x01234567);
178
179 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
180 __m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), voutput_zero_point);
181
182 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
183 vacc22x01234567 = _mm256_permute4x64_epi64(vacc22x01234567, _MM_SHUFFLE(3, 1, 2, 0));
184
185 __m256i vout = _mm256_packs_epi16(vacc01x01234567, vacc22x01234567);
186
187 vout = _mm256_max_epi8(vout, voutput_min);
188
189 __m128i vout_lo = _mm256_castsi256_si128(vout);
190 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
191
192 if (nc >= 8) {
193 _mm_storeh_pi((__m64*) c2, _mm_castsi128_ps(vout_lo));
194 _mm_storel_epi64((__m128i*) c1, vout_hi);
195 _mm_storel_epi64((__m128i*) c0, vout_lo);
196
197 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
198 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
199 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
200
201 a = (const int8_t**restrict) ((uintptr_t) a - ks);
202
203 nc -= 8;
204 } else {
205 if (nc & 4) {
206 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout_lo, 2));
207 _mm_storeu_si32(c1, vout_hi);
208 _mm_storeu_si32(c0, vout_lo);
209
210 c2 += 4;
211 c1 += 4;
212 c0 += 4;
213
214 vout_lo = _mm_srli_epi64(vout_lo, 32);
215 vout_hi = _mm_srli_epi64(vout_hi, 32);
216 }
217 if (nc & 2) {
218 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout_lo, 4));
219 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout_hi, 0));
220 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
221
222 c2 += 2;
223 c1 += 2;
224 c0 += 2;
225
226 vout_lo = _mm_srli_epi32(vout_lo, 16);
227 vout_hi = _mm_srli_epi32(vout_hi, 16);
228 }
229 if (nc & 1) {
230 *c2 = (int8_t) _mm_extract_epi8(vout_lo, 8);
231 *c1 = (int8_t) _mm_extract_epi8(vout_hi, 0);
232 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
233 }
234
235 nc = 0;
236 }
237 } while (nc != 0);
238 }