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_qu8_gemm_minmax_fp32_ukernel_2x8c8__avx2(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qu8_gemm_minmax_fp32_ukernel_2x8c8__avx2(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const uint8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 uint8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(uint8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 8);
42 const uint8_t* a0 = a;
43 uint8_t* c0 = c;
44 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50
51 do {
52 const __m128i vbias0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
53 const __m128i vbias0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
54 __m256i vacc0x01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x0), vbias0x1, 1);
55 const __m128i vbias0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
56 const __m128i vbias0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
57 __m256i vacc0x23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x2), vbias0x3, 1);
58 const __m128i vbias0x4 = _mm_cvtsi32_si128(((const int*) w)[4]);
59 const __m128i vbias0x5 = _mm_cvtsi32_si128(((const int*) w)[5]);
60 __m256i vacc0x45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x4), vbias0x5, 1);
61 const __m128i vbias0x6 = _mm_cvtsi32_si128(((const int*) w)[6]);
62 const __m128i vbias0x7 = _mm_cvtsi32_si128(((const int*) w)[7]);
63 __m256i vacc0x67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vbias0x6), vbias0x7, 1);
64 __m256i vacc1x01 = vacc0x01;
65 __m256i vacc1x23 = vacc0x23;
66 __m256i vacc1x45 = vacc0x45;
67 __m256i vacc1x67 = vacc0x67;
68 w = (const int32_t*) w + 8;
69
70 size_t k = 0;
71 const __m256i vb_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.kernel_zero_point);
72 while (k < kc) {
73 const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0));
74 const __m256i vxa0 = _mm256_cvtepu8_epi16(va0);
75 a0 += 8;
76 const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1));
77 const __m256i vxa1 = _mm256_cvtepu8_epi16(va1);
78 a1 += 8;
79
80 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
81 const __m256i vxb01 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb01), vb_zero_point);
82
83 vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01));
84 vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01));
85 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 16));
86 const __m256i vxb23 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb23), vb_zero_point);
87
88 vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23));
89 vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23));
90 const __m128i vb45 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 32));
91 const __m256i vxb45 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb45), vb_zero_point);
92
93 vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45));
94 vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45));
95 const __m128i vb67 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 48));
96 const __m256i vxb67 = _mm256_sub_epi16(_mm256_cvtepu8_epi16(vb67), vb_zero_point);
97
98 vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67));
99 vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67));
100
101 w = (const void*) ((const uint8_t*) w + 64);
102 k += 8 * sizeof(uint8_t);
103 }
104
105 const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23);
106 const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67);
107 const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23);
108 const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67);
109
110 const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657);
111 const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657);
112
113 const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0);
114 __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask);
115 __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask);
116
117 __m256 vscaled0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567);
118 __m256 vscaled1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567);
119
120 const __m256 vscale = _mm256_load_ps(params->fp32_avx2.scale);
121 vscaled0x01234567 = _mm256_mul_ps(vscaled0x01234567, vscale);
122 vscaled1x01234567 = _mm256_mul_ps(vscaled1x01234567, vscale);
123
124 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->fp32_avx2.output_max_less_zero_point);
125 vscaled0x01234567 = _mm256_min_ps(vscaled0x01234567, voutput_max_less_zero_point);
126 vscaled1x01234567 = _mm256_min_ps(vscaled1x01234567, voutput_max_less_zero_point);
127
128 vacc0x01234567 = _mm256_cvtps_epi32(vscaled0x01234567);
129 vacc1x01234567 = _mm256_cvtps_epi32(vscaled1x01234567);
130
131 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->fp32_avx2.output_zero_point);
132 __m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), voutput_zero_point);
133
134 vacc01x01234567 = _mm256_permute4x64_epi64(vacc01x01234567, _MM_SHUFFLE(3, 1, 2, 0));
135
136 __m256i vout = _mm256_packus_epi16(vacc01x01234567, vacc01x01234567);
137
138 vout = _mm256_max_epu8(vout, _mm256_load_si256((const __m256i*) params->fp32_avx2.output_min));
139
140 __m128i vout_lo = _mm256_castsi256_si128(vout);
141 __m128i vout_hi = _mm256_extracti128_si256(vout, 1);
142
143 if (nc >= 8) {
144 _mm_storel_epi64((__m128i*) c0, vout_lo);
145 _mm_storel_epi64((__m128i*) c1, vout_hi);
146
147 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
148 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
149
150 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
151 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
152
153 nc -= 8;
154 } else {
155 if (nc & 4) {
156 _mm_storeu_si32(c0, vout_lo);
157 _mm_storeu_si32(c1, vout_hi);
158
159 c0 += 4;
160 c1 += 4;
161
162 vout_lo = _mm_srli_epi64(vout_lo, 32);
163 vout_hi = _mm_srli_epi64(vout_hi, 32);
164 }
165 if (nc & 2) {
166 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout_lo, 0));
167 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout_hi, 0));
168
169 c0 += 2;
170 c1 += 2;
171
172 vout_lo = _mm_srli_epi32(vout_lo, 16);
173 vout_hi = _mm_srli_epi32(vout_hi, 16);
174 }
175 if (nc & 1) {
176 *c0 = (uint8_t) _mm_extract_epi8(vout_lo, 0);
177 *c1 = (uint8_t) _mm_extract_epi8(vout_hi, 0);
178 }
179
180 nc = 0;
181 }
182 } while (nc != 0);
183 }
184