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 #include <xnnpack/unaligned.h>
18
19
xnn_qs8_gemm_xw_minmax_fp32_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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_1x8c8__avx2(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 1);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8);
42 const int8_t* a0 = a;
43 int8_t* c0 = c;
44
45 do {
46 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
47 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
48 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
49 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
50 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
51 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
52 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
53 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
54 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
55 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
56 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
57 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
58 w = (const int32_t*) w + 8;
59
60 size_t k = 0;
61 while (k < kc) {
62 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
63 const __m256i vxa0 = _mm256_cvtepi8_epi16(va0);
64 a0 += 8;
65
66 const __m256i vxb01 = _mm256_load_si256((const __m256i*) w);
67
68 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
69 const __m256i vxb23 = _mm256_load_si256((const __m256i*) ((const int16_t*) w + 16));
70
71 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
72 const __m256i vxb45 = _mm256_load_si256((const __m256i*) ((const int16_t*) w + 32));
73
74 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
75 const __m256i vxb67 = _mm256_load_si256((const __m256i*) ((const int16_t*) w + 48));
76
77 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
78
79 w = (const void*) ((const int16_t*) w + 64);
80 k += 8 * sizeof(int8_t);
81 }
82
83 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
84 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
85
86 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
87
88 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
89 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
90
91 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
92
93 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
94 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale);
95
96 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
97 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
98
99 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
100
101 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
102 __m256i vacc00x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc0x01234567), voutput_zero_point);
103
104 vacc00x01234567 = _mm256_permute4x64_epi64(vacc00x01234567, _MM_SHUFFLE(3, 1, 2, 0));
105
106 __m256i vout = _mm256_packs_epi16(vacc00x01234567, vacc00x01234567);
107
108 vout = _mm256_max_epi8(vout, _mm256_load_si256((const __m256i*) params->fp32_avx2.output_min));
109
110 __m128i vout_lo = _mm256_castsi256_si128(vout);
111 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
112
113 if (nc >= 8) {
114 _mm_storel_epi64((__m128i*) c0, vout_lo);
115
116 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
117
118 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
119
120 nc -= 8;
121 } else {
122 if (nc & 4) {
123 _mm_storeu_si32(c0, vout_lo);
124
125 c0 += 4;
126
127 vout_lo = _mm_srli_epi64(vout_lo, 32);
128 vout_hi = _mm_srli_epi64(vout_hi, 32);
129 }
130 if (nc & 2) {
131 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
132
133 c0 += 2;
134
135 vout_lo = _mm_srli_epi32(vout_lo, 16);
136 vout_hi = _mm_srli_epi32(vout_hi, 16);
137 }
138 if (nc & 1) {
139 *c0 = (int8_t) _mm_extract_epi8(vout_lo, 0);
140 }
141
142 nc = 0;
143 }
144 } while (nc != 0);
145 }
146