Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-2x8c8-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_2x8c8__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 <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (2 * 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
52 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);
53 const __m256i voutput_zero_point = _mm256_set1_epi16(params->fp32_scalar.output_zero_point);
54 const __m256i voutput_min = _mm256_set1_epi8(params->fp32_scalar.output_min);
55 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
56 // XNN_FORCE_REALIZATION(voutput_zero_point);
57 // XNN_FORCE_REALIZATION(voutput_min);
58
59 do {
60 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
61 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
62 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
63 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
64 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
65 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
66 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
67 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
68 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
69 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
70 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
71 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
72 __m256i vacc1x01 = vacc0x01;
73 __m256i vacc1x23 = vacc0x23;
74 __m256i vacc1x45 = vacc0x45;
75 __m256i vacc1x67 = vacc0x67;
76 w = (const int32_t*) w + 8;
77
78 size_t p = ks;
79 do {
80 const int8_t* restrict a0 = a[0];
81 if XNN_UNPREDICTABLE(a0 != zero) {
82 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
83 }
84 const int8_t* restrict a1 = a[1];
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
87 }
88 a += 2;
89
90 size_t k = 0;
91 while (k < kc) {
92 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
93 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
94 a0 += 8;
95 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
96 const __m256i vxa1 = _mm256_cvtepi8_epi16(va1);
97 a1 += 8;
98
99 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
100 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
101
102 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
103 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
104 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
105 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
106
107 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
108 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
109 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32));
110 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
111
112 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
113 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
114 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48));
115 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
116
117 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
118 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
119
120 w = (const void*) ((const int8_t*) w + 64);
121 k += 8 * sizeof(int8_t);
122 }
123 p -= 2 * sizeof(void*);
124 } while (p != 0);
125
126 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
127 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
128 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
129 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
130
131 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
132 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
133
134 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
135 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
136 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
137
138 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
139 __m256 vscaled1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
140
141 const __m256 vscale01234567 = _mm256_load_ps(w);
142 w = (const void*) ((const float*) w + 8);
143 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale01234567);
144 vscaled1x01234567 = _mm256_mul_ps(vscaled1x01234567, vscale01234567);
145
146 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
147 vscaled1x01234567 = _mm256_min_ps(vscaled1x01234567, voutput_max_less_zero_point);
148
149 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
150 vacc1x01234567 = _mm256_cvtps_epi32(vscaled1x01234567);
151
152 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
153
154 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
155
156 __m256i vout = _mm256_packs_epi16(vacc01x01234567, vacc01x01234567);
157
158 vout = _mm256_max_epi8(vout, voutput_min);
159
160 __m128i vout_lo = _mm256_castsi256_si128(vout);
161 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
162
163 if (nc >= 8) {
164 _mm_storel_epi64((__m128i*) c1, vout_hi);
165 _mm_storel_epi64((__m128i*) c0, vout_lo);
166
167 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
168 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
169
170 a = (const int8_t**restrict) ((uintptr_t) a - ks);
171
172 nc -= 8;
173 } else {
174 if (nc & 4) {
175 _mm_storeu_si32(c1, vout_hi);
176 _mm_storeu_si32(c0, vout_lo);
177
178 c1 += 4;
179 c0 += 4;
180
181 vout_lo = _mm_srli_epi64(vout_lo, 32);
182 vout_hi = _mm_srli_epi64(vout_hi, 32);
183 }
184 if (nc & 2) {
185 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout_hi, 0));
186 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
187
188 c1 += 2;
189 c0 += 2;
190
191 vout_lo = _mm_srli_epi32(vout_lo, 16);
192 vout_hi = _mm_srli_epi32(vout_hi, 16);
193 }
194 if (nc & 1) {
195 *c1 = (int8_t) _mm_extract_epi8(vout_hi, 0);
196 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
197 }
198
199 nc = 0;
200 }
201 } while (nc != 0);
202 }