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_minmax_ukernel_1x8c8__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_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gemm_minmax_ukernel_1x8c8__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_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
30 {
31 assert(mr != 0);
32 assert(mr <= 1);
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
44 do {
45 const __m128i vbias0x0 = _mm_loadu_si32(w);
46 const __m128i vbias0x1 = _mm_loadu_si32((const void*) ((uintptr_t) w + sizeof(int32_t)));
47 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
48 const __m128i vbias0x2 = _mm_loadu_si32((const void*) ((uintptr_t) w + 2 * sizeof(int32_t)));
49 const __m128i vbias0x3 = _mm_loadu_si32((const void*) ((uintptr_t) w + 3 * sizeof(int32_t)));
50 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
51 const __m128i vbias0x4 = _mm_loadu_si32((const void*) ((uintptr_t) w + 4 * sizeof(int32_t)));
52 const __m128i vbias0x5 = _mm_loadu_si32((const void*) ((uintptr_t) w + 5 * sizeof(int32_t)));
53 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
54 const __m128i vbias0x6 = _mm_loadu_si32((const void*) ((uintptr_t) w + 6 * sizeof(int32_t)));
55 const __m128i vbias0x7 = _mm_loadu_si32((const void*) ((uintptr_t) w + 7 * sizeof(int32_t)));
56 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
57 w = (const void*) ((uintptr_t) w + 8 * sizeof(int32_t));
58
59 size_t k = 0;
60 while (k < kc) {
61 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
62 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
63 a0 += 8;
64
65 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
66 const __m256i vxb01 = _mm256_cvtepi8_epi16(vb01);
67
68 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
69 const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
70 const __m256i vxb23 = _mm256_cvtepi8_epi16(vb23);
71
72 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
73 const __m128i vb45 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int8_t)));
74 const __m256i vxb45 = _mm256_cvtepi8_epi16(vb45);
75
76 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
77 const __m128i vb67 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 48 * sizeof(int8_t)));
78 const __m256i vxb67 = _mm256_cvtepi8_epi16(vb67);
79
80 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
81
82 w = (const void*) ((uintptr_t) w + 64 * sizeof(int8_t));
83 k += 8 * sizeof(int8_t);
84 }
85
86 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
87 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
88
89 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
90
91 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
92 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
93
94 const __m256i vmultiplier = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.multiplier));
95 const __m256i vrounding = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.rounding));
96
97 const __m256i vacc0x11335577 = _mm256_shuffle_epi32(vacc0x01234567, _MM_SHUFFLE(3, 3, 1, 1));
98
99 const __m256i vprod0x0246 = _mm256_add_epi64(_mm256_mul_epi32(vacc0x01234567, vmultiplier), vrounding);
100
101 const __m256i vprod0x1357 = _mm256_add_epi64(_mm256_mul_epi32(vacc0x11335577, vmultiplier), vrounding);
102
103 const __m256i vq31prod0x0246 = _mm256_srli_epi64(vprod0x0246, 31);
104 const __m256i vq31prod0x1357 = _mm256_add_epi64(vprod0x1357, vprod0x1357);
105
106 const __m256i vq31prod0x01234567 = _mm256_blend_epi16(vq31prod0x0246, vq31prod0x1357, 0xCC);
107
108 const __m256i vremainder_mask = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_mask));
109 const __m256i vrem0x01234567 =
110 _mm256_add_epi32(_mm256_and_si256(vq31prod0x01234567, vremainder_mask), _mm256_cmpgt_epi32(_mm256_setzero_si256(), vq31prod0x01234567));
111
112 const __m256i vremainder_threshold = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.remainder_threshold));
113 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
114 vacc0x01234567 =
115 _mm256_sub_epi32(_mm256_sra_epi32(vq31prod0x01234567, vshift), _mm256_cmpgt_epi32(vrem0x01234567, vremainder_threshold));
116
117 const __m256i voutput_zero_point = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_zero_point));
118 __m256i vacc00x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc0x01234567), voutput_zero_point);
119
120 vacc00x01234567 = _mm256_permute4x64_epi64(vacc00x01234567, _MM_SHUFFLE(3, 1, 2, 0));
121
122 const __m256i voutput_min = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_min));
123 const __m256i voutput_max = _mm256_broadcastsi128_si256(_mm_load_si128((const __m128i*) params->sse2.output_max));
124 vacc00x01234567 = _mm256_min_epi16(_mm256_max_epi16(vacc00x01234567, voutput_min), voutput_max);
125
126 __m256i vout = _mm256_packs_epi16(vacc00x01234567, vacc00x01234567);
127 __m128i vout_lo = _mm256_castsi256_si128(vout);
128 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
129
130 if (nc >= 8) {
131 _mm_storel_epi64((__m128i*) c0, vout_lo);
132
133 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
134
135 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
136
137 nc -= 8;
138 } else {
139 if (nc & 4) {
140 _mm_storeu_si32(c0, vout_lo);
141
142 c0 += 4;
143
144 vout_lo = _mm_srli_epi64(vout_lo, 32);
145 vout_hi = _mm_srli_epi64(vout_hi, 32);
146 }
147 if (nc & 2) {
148 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout_lo, 0);
149
150 c0 += 2;
151
152 vout_lo = _mm_srli_epi32(vout_lo, 16);
153 vout_hi = _mm_srli_epi32(vout_hi, 16);
154 }
155 if (nc & 1) {
156 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
157 }
158
159 nc = 0;
160 }
161 } while (nc != 0);
162 }
163